Re4b en PDF

This is my own bulletin board
This book is probably outdated already. (Unless it was just downloaded from https://beginners.
re/.)
The book is changing too often, content being added, bugs are (hopefully) being fixed. The latest version
is always at https://beginners.re/.
This PDF you currently reading was compiled at March 28, 2020.
If you have printed this book on paper, can you please send me a picture of it, for collection?
<first_name @ last_name . com> or <first_name . last_name @ gmail . com>. The collection, so far:
https://yurichev.com/news/20200222_printed_RE4B/.
My dear readers! From time to time, I have questions, I don’t know who (or where) to ask. Or I’m just
lazy... Can you please help me?
I have a 2TB HDD that is constantly busy/noisy. Even if no OS booted! That drives me crazy. A strange
thing – when I turn on ”long” SMART test (smartctl --test=long /dev/sda), it is almost silent, but for a
5 hours. When test finish, the HDD is constantly noisy again. While everything is fine in SMART info. The
HDD is new. Maybe this is a kind of maintenance I should just wait for the end of? Nothing suspicious in
SMART info: https://pastebin.com/5eFePXgG.
Can anyone help me with Low Fragmentation Heap in Windows?
Survey: how do you use DLL injection aside of API interception?
I’ve got a huge graph, for example, 106 nodes (vertices). I want to visualize it somehow, so that a user
would navigate it using mouse. Clicked on link (edge) and moved to another node (vertex). Like in IDA.
Maybe, using JavaScript. Are there any mature open-source solutions?
Do you remember “The Incredible Machine” videogame for DOS? Do you know about Rube Goldberg
machines? What can be used today for simulating them? Some kind of physics engine, maybe?
The ERGO BT-590 bluetooth headphones have a touch control that is too sensitive and easy to hit with
clothing. How to force Android ignore messages from headphones about buttons?
If you know something, please help me: <first_name @ last_name . com> or <first_name . last_name @
gmail . com>.
i
____
| _ \ _____ _____ _ __ ___ ___
| |_) / _ \ \ / / _ \ '__/ __|/ _ \
| _ < __/\ V / __/ | \__ \ __/
|_| \_\___| \_/ \___|_| |___/\___|
_____ _ _
| ____|_ __ __ _(_)_ __ ___ ___ _ __(_)_ __ __ _
| _| | '_ \ / _` | | '_ \ / _ \/ _ \ '__| | '_ \ / _` |
| |___| | | | (_| | | | | | __/ __/ | | | | | | (_| |
|_____|_| |_|\__, |_|_| |_|\___|\___|_| |_|_| |_|\__, |
|___/ |___/
__
/ _| ___ _ __
| |_ / _ \| '__|
| _| (_) | |
|_| \___/|_|
____ _
| __ ) ___ __ _(_)_ __ _ __ ___ _ __ ___
| _ \ / _ \/ _` | | '_ \| '_ \ / _ \ '__/ __|
| |_) | __/ (_| | | | | | | | | __/ | \__ \
|____/ \___|\__, |_|_| |_|_| |_|\___|_| |___/
|___/
Reverse Engineering for Beginners
(Understanding Assembly Language)
Why two titles? Read here: on page xiii.
Dennis Yurichev
<first_name @ last_name . com>
cba
©2013-2020, Dennis Yurichev.
This work is licensed under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)
license. To view a copy of this license, visit https://creativecommons.org/licenses/by-sa/4.0/.
Text version (March 28, 2020).
The latest version (and Russian edition) of this text is accessible at beginners.re.
Call for translators!
You may want to help me with translating this work into languages other than English and Russian. Just
send me any piece of translated text (no matter how short) and I’ll put it into my LaTeX source code.
Read here.
The language statistics is available right here: https://beginners.re/.
Speed isn’t important, because this is an open-source project, after all. Your name will be mentioned as
a project contributor. Korean, Chinese, and Persian languages are reserved by publishers. English and
Russian versions I do by myself, but my English is still that horrible, so I’m very grateful for any notes
about grammar, etc. Even my Russian is flawed, so I’m grateful for notes about Russian text as well!
So do not hesitate to contact me: <first_name @ last_name . com> or <first_name . last_name @
gmail . com>.
i
Abridged contents
1 Code Patterns 1
2 Important fundamentals 443
3 Slightly more advanced examples 464
4 Java 654
5 Finding important/interesting stuff in the code 692
6 OS-specific 725
7 Tools 780
8 Case studies 784
9 Examples of reversing proprietary file formats 908
10 Dynamic binary instrumentation 972
11 Other things 980
12 Books/blogs worth reading 998
13 Communities 1001
Afterword 1003
Appendix 1005
Acronyms Used 1033
Glossary 1038
Index 1040
ii
Contents
1 Code Patterns 1
1.1 The method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Some basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 A short introduction to the CPU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.2 Numeral Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.3 Converting From One Radix To Another . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 An Empty Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.1 x86 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.2 ARM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.3 MIPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.4 Empty Functions in Practice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Returning Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4.1 x86 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4.2 ARM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4.3 MIPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Hello, world! . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5.1 x86 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5.2 x86-64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.5.3 ARM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5.4 MIPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.5.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.6 Function prologue and epilogue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.6.1 Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.7 An Empty Function: redux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.8 Returning Values: redux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.9 Stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.9.1 Why does the stack grow backwards? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.9.2 What is the stack used for? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.9.3 A typical stack layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
1.9.4 Noise in stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
1.9.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
1.10 Almost empty function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
1.11 printf() with several arguments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.11.1 x86 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.11.2 ARM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
1.11.3 MIPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
1.11.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
1.11.5 By the way . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
1.12 scanf() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
1.12.1 Simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
1.12.2 The classic mistake . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
1.12.3 Global variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
1.12.4 scanf() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
1.12.5 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
1.13 Worth noting: global vs. local variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
1.14 Accessing passed arguments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
1.14.1 x86 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
1.14.2 x64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
1.14.3 ARM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
1.14.4 MIPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
1.15 More about results returning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
1.15.1 Attempt to use the result of a function returning void . . . . . . . . . . . . . . . . . . . . . 106
iii
1.15.2 What if we do not use the function result? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
1.15.3 Returning a structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
1.16 Pointers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
1.16.1 Returning values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
1.16.2 Swap input values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
1.17 GOTO operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
1.17.1 Dead code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
1.17.2 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
1.18 Conditional jumps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
1.18.1 Simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
1.18.2 Calculating absolute value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
1.18.3 Ternary conditional operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
1.18.4 Getting minimal and maximal values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
1.18.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
1.18.6 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
1.19 Software cracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
1.20 Impossible shutdown practical joke (Windows 7) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
1.21 switch()/case/default . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
1.21.1 Small number of cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
1.21.2 A lot of cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
1.21.3 When there are several case statements in one block . . . . . . . . . . . . . . . . . . . . . 178
1.21.4 Fall-through . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
1.21.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
1.22 Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
1.22.1 Simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
1.22.2 Memory blocks copying routine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
1.22.3 Condition check . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
1.22.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
1.22.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
1.23 More about strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
1.23.1 strlen() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
1.23.2 Boundaries of strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
1.24 Replacing arithmetic instructions to other ones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
1.24.1 Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
1.24.2 Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
1.24.3 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
1.25 Floating-point unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
1.25.1 IEEE 754 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
1.25.2 x86 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
1.25.3 ARM, MIPS, x86/x64 SIMD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
1.25.4 C/C++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
1.25.5 Simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
1.25.6 Passing floating point numbers via arguments . . . . . . . . . . . . . . . . . . . . . . . . . . 228
1.25.7 Comparison example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
1.25.8 Some constants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
1.25.9 Copying . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
1.25.10 Stack, calculators and reverse Polish notation . . . . . . . . . . . . . . . . . . . . . . . . . 264
1.25.11 80 bits? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
1.25.12 x64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
1.25.13 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
1.26 Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
1.26.1 Simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
1.26.2 Buffer overflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
1.26.3 Buffer overflow protection methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
1.26.4 One more word about arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
1.26.5 Array of pointers to strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
1.26.6 Multidimensional arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
1.26.7 Pack of strings as a two-dimensional array . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
1.26.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
1.26.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
1.27 Example: a bug in Angband . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
1.28 Manipulating specific bit(s) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
1.28.1 Specific bit checking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
1.28.2 Setting and clearing specific bits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
iv
1.28.3 Shifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
1.28.4 Setting and clearing specific bits: FPU1 example . . . . . . . . . . . . . . . . . . . . . . . . 315
1.28.5 Counting bits set to 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
1.28.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
1.28.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
1.29 Linear congruential generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
1.29.1 x86 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
1.29.2 x64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
1.29.3 32-bit ARM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
1.29.4 MIPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
1.29.5 Thread-safe version of the example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
1.30 Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
1.30.1 MSVC: SYSTEMTIME example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
1.30.2 Let’s allocate space for a structure using malloc() . . . . . . . . . . . . . . . . . . . . . . . 344
1.30.3 UNIX: struct tm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
1.30.4 Fields packing in structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
1.30.5 Nested structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
1.30.6 Bit fields in a structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
1.30.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
1.31 The classic struct bug . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
1.32 Unions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374
1.32.1 Pseudo-random number generator example . . . . . . . . . . . . . . . . . . . . . . . . . . . 374
1.32.2 Calculating machine epsilon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
1.32.3 FSCALE instruction replacement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
1.32.4 Fast square root calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
1.33 Pointers to functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
1.33.1 MSVC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
1.33.2 GCC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
1.33.3 Danger of pointers to functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
1.34 64-bit values in 32-bit environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
1.34.1 Returning of 64-bit value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
1.34.2 Arguments passing, addition, subtraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
1.34.3 Multiplication, division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
1.34.4 Shifting right . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
1.34.5 Converting 32-bit value into 64-bit one . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
1.35 LARGE_INTEGER structure case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402
1.36 SIMD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
1.36.1 Vectorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
1.36.2 SIMD strlen() implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
1.37 64 bits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 418
1.37.1 x86-64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 418
1.37.2 ARM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
1.37.3 Float point numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
1.37.4 64-bit architecture criticism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
1.38 Working with floating point numbers using SIMD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
1.38.1 Simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
1.38.2 Passing floating point number via arguments . . . . . . . . . . . . . . . . . . . . . . . . . . 433
1.38.3 Comparison example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 434
1.38.4 Calculating machine epsilon: x64 and SIMD . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
1.38.5 Pseudo-random number generator example revisited . . . . . . . . . . . . . . . . . . . . . 436
1.38.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
1.39 ARM-specific details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
1.39.1 Number sign (#) before number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
1.39.2 Addressing modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
1.39.3 Loading a constant into a register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
1.39.4 Relocs in ARM64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
1.40 MIPS-specific details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
1.40.1 Loading a 32-bit constant into register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
1.40.2 Further reading about MIPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
2 Important fundamentals 443

2.1 Integral datatypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
2.1.1 Bit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
1 Floating-Point Unit
v
2.1.2 Nibble AKA nybble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
2.1.3 Byte . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
2.1.4 Wide char . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
2.1.5 Signed integer vs unsigned . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
2.1.6 Word . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
2.1.7 Address register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
2.1.8 Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
2.2 Signed number representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
2.2.1 Using IMUL over MUL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
2.2.2 Couple of additions about two’s complement form . . . . . . . . . . . . . . . . . . . . . . . . 450
2.2.3 -1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
2.3 Integer overflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
2.4 AND . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
2.4.1 Checking if a value is on 2n boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
2.4.2 KOI-8R Cyrillic encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
2.5 AND and OR as subtraction and addition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
2.5.1 ZX Spectrum ROM text strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
2.6 XOR (exclusive OR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
2.6.1 Logical difference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
2.6.2 Everyday speech . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
2.6.3 Encryption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
2.6.4 RAID2 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
2.6.5 XOR swap algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
2.6.6 XOR linked list . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458
2.6.7 Switching value trick . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458
2.6.8 Zobrist hashing / tabulation hashing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
2.6.9 By the way . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
2.6.10 AND/OR/XOR as MOV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
2.7 Population count . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
2.8 Endianness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
2.8.1 Big-endian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
2.8.2 Little-endian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
2.8.3 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
2.8.4 Bi-endian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
2.8.5 Converting data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
2.9 Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
2.10 CPU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
2.10.1 Branch predictors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
2.10.2 Data dependencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
2.11 Hash functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
2.11.1 How do one-way functions work? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
3 Slightly more advanced examples 464

3.1 Zero register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
3.2 Double negation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
3.3 const correctness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
3.3.1 Overlapping const strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
3.4 strstr() example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
3.5 qsort() revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
3.6 Temperature converting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
3.6.1 Integer values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
3.6.2 Floating-point values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
3.7 Fibonacci numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475
3.7.1 Example #1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475
3.7.2 Example #2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 478
3.7.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481
3.8 CRC32 calculation example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482
3.9 Network address calculation example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484
3.9.1 calc_network_address() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486
3.9.2 form_IP() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486
3.9.3 print_as_IP() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
3.9.4 form_netmask() and set_bit() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
3.9.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 490
2 Redundant Array of Independent Disks
vi
3.10 Loops: several iterators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 490
3.10.1 Three iterators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 490
3.10.2 Two iterators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
3.10.3 Intel C++ 2011 case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
3.11 Duff’s device . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 494
3.11.1 Should one use unrolled loops? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496
3.12 Division using multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496
3.12.1 x86 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496
3.12.2 How it works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 498
3.12.3 ARM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 498
3.12.4 MIPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
3.12.5 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
3.13 String to number conversion (atoi()) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
3.13.1 Simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
3.13.2 A slightly advanced example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
3.13.3 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506
3.14 Inline functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506
3.14.1 Strings and memory functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
3.15 C99 restrict . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514
3.16 Branchless abs() function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
3.16.1 Optimizing GCC 4.9.1 x64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
3.16.2 Optimizing GCC 4.9 ARM64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518
3.17 Variadic functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518
3.17.1 Computing arithmetic mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518
3.17.2 vprintf() function case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522
3.17.3 Pin case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523
3.17.4 Format string exploit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523
3.18 Strings trimming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524
3.18.1 x64: Optimizing MSVC 2013 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
3.18.2 x64: Non-optimizing GCC 4.9.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
3.18.3 x64: Optimizing GCC 4.9.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
3.18.4 ARM64: Non-optimizing GCC (Linaro) 4.9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529
3.18.5 ARM64: Optimizing GCC (Linaro) 4.9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530
3.18.6 ARM: Optimizing Keil 6/2013 (ARM mode) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530
3.18.7 ARM: Optimizing Keil 6/2013 (Thumb mode) . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
3.18.8 MIPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
3.19 toupper() function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533
3.19.1 x64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533
3.19.2 ARM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535
3.19.3 Using bit operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
3.19.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
3.20 Obfuscation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
3.20.1 Text strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
3.20.2 Executable code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 538
3.20.3 Virtual machine / pseudo-code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
3.20.4 Other things to mention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
3.20.5 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
3.21 C++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
3.21.1 Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
3.21.2 ostream . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
3.21.3 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 558
3.21.4 STL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559
3.21.5 Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592
3.22 Negative array indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593
3.22.1 Addressing string from the end . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593
3.22.2 Addressing some kind of block from the end . . . . . . . . . . . . . . . . . . . . . . . . . . . 593
3.22.3 Arrays started at 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593
3.23 More about pointers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 596
3.23.1 Working with addresses instead of pointers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 596
3.23.2 Passing values as pointers; tagged unions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 598
3.23.3 Pointers abuse in Windows kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
3.23.4 Null pointers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603
3.23.5 Array as function argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 608
3.23.6 Pointer to a function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 608
vii
3.23.7 Pointer to a function: copy protection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 609
3.23.8 Pointer to a function: a common bug (or typo) . . . . . . . . . . . . . . . . . . . . . . . . . . 610
3.23.9 Pointer as object identificator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610
3.23.10 Oracle RDBMS and a simple garbage collector for C/C++ . . . . . . . . . . . . . . . . . . 611
3.24 Loop optimizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612
3.24.1 Weird loop optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612
3.24.2 Another loop optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614
3.25 More about structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 616
3.25.1 Sometimes a C structure can be used instead of array . . . . . . . . . . . . . . . . . . . . 616
3.25.2 Unsized array in C structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 617
3.25.3 Version of C structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 618
3.25.4 High-score file in “Block out” game and primitive serialization . . . . . . . . . . . . . . . 620
3.26 memmove() and memcpy() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624
3.26.1 Anti-debugging trick . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624
3.27 setjmp/longjmp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 625
3.28 Other weird stack hacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 627
3.28.1 Accessing arguments/local variables of caller . . . . . . . . . . . . . . . . . . . . . . . . . . 627
3.28.2 Returning string . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 629
3.29 OpenMP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 630
3.29.1 MSVC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 632
3.29.2 GCC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634
3.30 Signed division using shifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 636
3.31 Another heisenbug . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 637
3.32 The case of forgotten return . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638
3.33 Homework: more about function pointers and unions . . . . . . . . . . . . . . . . . . . . . . . . . 642
3.34 Windows 16-bit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
3.34.1 Example#1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
3.34.2 Example #2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644
3.34.3 Example #3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644
3.34.4 Example #4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 645
3.34.5 Example #5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 648
3.34.6 Example #6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 651
4 Java 654
4.1 Java . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
4.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
4.1.2 Returning a value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
4.1.3 Simple calculating functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 659
4.1.4 JVM3 memory model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 661
4.1.5 Simple function calling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 662
4.1.6 Calling beep() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 663
4.1.7 Linear congruential PRNG4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 664
4.1.8 Conditional jumps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 665
4.1.9 Passing arguments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 667
4.1.10 Bitfields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 668
4.1.11 Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 669
4.1.12 switch() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 671
4.1.13 Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 672
4.1.14 Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 680
4.1.15 Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 681
4.1.16 Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684
4.1.17 Simple patching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 686
4.1.18 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691
5 Finding important/interesting stuff in the code 692

5.1 Identification of executable files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 692
5.1.1 Microsoft Visual C++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 692
5.1.2 GCC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693
5.1.3 Intel Fortran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693
5.1.4 Watcom, OpenWatcom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693
5.1.5 Borland . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694
5.1.6 Other known DLLs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 695
3 Java Virtual Machine
4 Pseudorandom Number Generator
viii
5.2 Communication with outer world (function level) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 695
5.3 Communication with the outer world (win32) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 695
5.3.1 Often used functions in the Windows API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 696
5.3.2 Extending trial period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 696
5.3.3 Removing nag dialog box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 696
5.3.4 tracer: Intercepting all functions in specific module . . . . . . . . . . . . . . . . . . . . . . . 696
5.4 Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 697
5.4.1 Text strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 697
5.4.2 Finding strings in binary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 701
5.4.3 Error/debug messages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 702
5.4.4 Suspicious magic strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 703
5.5 Calls to assert() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 703
5.6 Constants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704
5.6.1 Magic numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704
5.6.2 Specific constants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 706
5.6.3 Searching for constants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 706
5.7 Finding the right instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 706
5.8 Suspicious code patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 707
5.8.1 XOR instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 707
5.8.2 Hand-written assembly code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 708
5.9 Using magic numbers while tracing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 709
5.10 Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 709
5.10.1 Some binary file patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 710
5.10.2 Memory “snapshots” comparing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 717
5.11 ISA5 detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 719
5.11.1 Incorrectly disassembled code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 719
5.11.2 Correctly disassembled code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
5.12 Other things . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
5.12.1 General idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
5.12.2 Order of functions in binary code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
5.12.3 Tiny functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
5.12.4 C++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
5.12.5 Crash on purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
6 OS-specific 725
6.1 Arguments passing methods (calling conventions) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725
6.1.1 cdecl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725
6.1.2 stdcall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725
6.1.3 fastcall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
6.1.4 thiscall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 727
6.1.5 x86-64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 728
6.1.6 Return values of float and double type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 730
6.1.7 Modifying arguments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 730
6.1.8 Taking a pointer to function argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
6.1.9 Python ctypes problem (x86 assembly homework) . . . . . . . . . . . . . . . . . . . . . . . . 733
6.1.10 Cdecl example: a DLL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733
6.2 Thread Local Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 734
6.2.1 Linear congruential generator revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 734
6.3 System calls (syscall-s) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 738
6.3.1 Linux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 739
6.3.2 Windows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 739
6.4 Linux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 739
6.4.1 Position-independent code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 739
6.4.2 LD_PRELOAD hack in Linux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 742
6.5 Windows NT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 744
6.5.1 CRT (win32) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 744
6.5.2 Win32 PE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 748
6.5.3 Windows SEH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 755
6.5.4 Windows NT: Critical section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 777
7 Tools 780
7.1 Binary analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 780
7.1.1 Disassemblers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 780
5 Instruction Set Architecture
ix
7.1.2 Decompilers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781
7.1.3 Patch comparison/diffing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781
7.2 Live analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781
7.2.1 Debuggers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781
7.2.2 Library calls tracing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781
7.2.3 System calls tracing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781
7.2.4 Network sniffing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782
7.2.5 Sysinternals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782
7.2.6 Valgrind . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782
7.2.7 Emulators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782
7.3 Other tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782
7.3.1 SMT solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782
7.3.2 Calculators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 783
7.4 Do You Think Something Is Missing Here? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 783
8 Case studies 784

8.1 Mahjong solitaire prank (Windows 7) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 784
8.2 Task manager practical joke (Windows Vista) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 786
8.2.1 Using LEA to load values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 789
8.3 Color Lines game practical joke . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 791
8.4 Minesweeper (Windows XP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 794
8.4.1 Finding grid automatically . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 799
8.4.2 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 800
8.5 Hacking Windows clock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 800
8.6 (Windows 7) Solitaire: practical jokes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 807
8.6.1 51 cards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 807
8.6.2 53 cards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 812
8.7 FreeCell prank (Windows 7) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813
8.7.1 Part I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813
8.7.2 Part II: breaking the Select Game submenu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 817
8.8 Dongles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 818
8.8.1 Example #1: MacOS Classic and PowerPC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 818
8.8.2 Example #2: SCO OpenServer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 826
8.8.3 Example #3: MS-DOS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 835
8.9 Encrypted database case #1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 840
8.9.1 Base64 and entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 840
8.9.2 Is data compressed? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 842
8.9.3 Is data encrypted? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 843
8.9.4 CryptoPP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 843
8.9.5 Cipher Feedback mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845
8.9.6 Initializing Vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 847
8.9.7 Structure of the buffer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 848
8.9.8 Noise at the end . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 850
8.9.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 850
8.9.10 Post Scriptum: brute-forcing IV6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 851
8.10 Overclocking Cointerra Bitcoin miner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 851
8.11 Breaking simple executable cryptor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 855
8.11.1 Other ideas to consider . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 860
8.12 SAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 860
8.12.1 About SAP client network traffic compression . . . . . . . . . . . . . . . . . . . . . . . . . . 860
8.12.2 SAP 6.0 password checking functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 871
8.13 Oracle RDBMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 874
8.13.1 V$VERSION table in the Oracle RDBMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 874
8.13.2 X$KSMLRU table in Oracle RDBMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 882
8.13.3 V$TIMER table in Oracle RDBMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 883
8.14 Handwritten assembly code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 887
8.14.1 EICAR test file . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 887
8.15 Demos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 888
8.15.1 10 PRINT CHR$(205.5+RND(1)); : GOTO 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . 888
8.15.2 Mandelbrot set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 891
8.16 A nasty bug in MSVCRT.DLL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 901
8.17 Other examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 907
6 Initialization Vector
x
9 Examples of reversing proprietary file formats 908
9.1 Primitive XOR-encryption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 908
9.1.1 Simplest ever XOR encryption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 908
9.1.2 Norton Guide: simplest possible 1-byte XOR encryption . . . . . . . . . . . . . . . . . . . . 910
9.1.3 Simplest possible 4-byte XOR encryption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 913
9.1.4 Simple encryption using XOR mask . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 917
9.1.5 Simple encryption using XOR mask, case II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 924
9.1.6 Homework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 929
9.2 Information entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 929
9.2.1 Analyzing entropy in Mathematica . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 930
9.2.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 939
9.2.3 Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 939
9.2.4 A word about primitive encryption like XORing . . . . . . . . . . . . . . . . . . . . . . . . . . 940
9.2.5 More about entropy of executable code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 940
9.2.6 PRNG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 940
9.2.7 More examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 940
9.2.8 Entropy of various files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 940
9.2.9 Making lower level of entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 942
9.3 Millenium game save file . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 942
9.4 fortune program indexing file . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 949
9.4.1 Hacking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 953
9.4.2 The files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 954
9.5 Oracle RDBMS: .SYM-files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 954
9.6 Oracle RDBMS: .MSB-files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 964
9.6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 971
9.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 971
9.8 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 971
10 Dynamic binary instrumentation 972

10.1 Using PIN DBI for XOR interception . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 972
10.2 Cracking Minesweeper with PIN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 975
10.2.1 Intercepting all rand() calls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 975
10.2.2 Replacing rand() calls with our function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 975
10.2.3 Peeking into placement of mines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 977
10.2.4 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 978
10.3 Building Pin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 978
10.4 Why “instrumentation”? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 979
11 Other things 980

11.1 Executable files patching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 980
11.1.1 x86 code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 980
11.2 Function arguments number statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 980
11.3 Compiler intrinsic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 981
11.4 Compiler’s anomalies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 981
11.4.1 Oracle RDBMS 11.2 and Intel C++ 10.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 981
11.4.2 MSVC 6.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 982
11.4.3 ftol2() in MSVC 2012 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 982
11.4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 983
11.5 Itanium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 984
11.6 8086 memory model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 986
11.7 Basic blocks reordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 987
11.7.1 Profile-guided optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 987
11.8 My experience with Hex-Rays 2.2.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 988
11.8.1 Bugs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 988
11.8.2 Odd peculiarities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 990
11.8.3 Silence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 991
11.8.4 Comma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 992
11.8.5 Data types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 993
11.8.6 Long and messed expressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 993
11.8.7 De Morgan’s laws and decompilation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 994
11.8.8 My plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 995
11.8.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 995
11.9 Cyclomatic complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 995
12 Books/blogs worth reading 998
xi
12.1 Books and other materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 998
12.1.1 Reverse Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 998
12.1.2 Windows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 998
12.1.3 C/C++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 998
12.1.4 x86 / x86-64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 999
12.1.5 ARM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 999
12.1.6 Assembly language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 999
12.1.7 Java . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 999
12.1.8 UNIX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 999
12.1.9 Programming in general . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 999
12.1.10 Cryptography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1000
13 Communities 1001
Afterword 1003
13.1 Questions? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1003
Appendix 1005
.1 x86 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1005
.1.1 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1005
.1.2 General purpose registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1005
.1.3 FPU registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1009
.1.4 SIMD registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1011
.1.5 Debugging registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1011
.1.6 Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1012
.1.7 npad . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1024
.2 ARM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1025
.2.1 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1025
.2.2 Versions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1026
.2.3 32-bit ARM (AArch32) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1026
.2.4 64-bit ARM (AArch64) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1027
.2.5 Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1027
.3 MIPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1028
.3.1 Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1028
.3.2 Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1029
.4 Some GCC library functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1029
.5 Some MSVC library functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1029
.6 Cheatsheets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1029
.6.1 IDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1029
.6.2 OllyDbg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1030
.6.3 MSVC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1030
.6.4 GCC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1030
.6.5 GDB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1031
Acronyms Used 1033

Glossary 1038
Index 1040
xii
Preface
What is with two titles?
The book was named “Reverse Engineering for Beginners” in 2014-2018, but I always suspected this
makes readership too narrow.
Infosec people know about “reverse engineering”, but I’ve rarely hear the “assembler” word from them.
Likewise, the “reverse engineering” term is somewhat cryptic to a general audience of programmers, but
they know about “assembler”.
In July 2018, as an experiment, I’ve changed the title to “Assembly Language for Beginners” and posted
the link to Hacker News website7 , and the book was received generally well.
So let it be, the book now has two titles.
However, I’ve changed the second title to “Understanding Assembly Language”, because someone had
already written “Assembly Language for Beginners” book. Also, people say “for Beginners” sounds a bit
sarcastic for a book of ~1000 pages.
The two books differ only by title, filename (UAL-XX.pdf versus RE4B-XX.pdf), URL and a couple of the first
pages.
About reverse engineering

There are several popular meanings of the term “reverse engineering”:
1) The reverse engineering of software; researching compiled programs
2) The scanning of 3D structures and the subsequent digital manipulation required in order to duplicate
them
3) Recreating DBMS8 structure
This book is about the first meaning.
Prerequisites
Basic knowledge of the C PL9 . Recommended reading: 12.1.3 on page 998.
Exercises and tasks

…can be found at: http://challenges.re.
Praise for this book

https://beginners.re/#praise.
Universities
The book is recommended at least at these universities: https://beginners.re/#uni.
Thanks
For patiently answering all my questions: Slava “Avid” Kazakov, SkullC0DEr.
For sending me notes about mistakes and inaccuracies: Stanislav “Beaver” Bobrytskyy, Alexander Ly-
senko, Alexander “Solar Designer” Peslyak, Federico Ramondino, Mark Wilson, Xenia Galinskaya, Razikhova
Meiramgul Kayratovna, Anatoly Prokofiev, Kostya Begunets, Valentin “netch” Nechayev, Aleksandr Plakhov,
Artem Metla, Alexander Yastrebov, Vlad Golovkin10 , Evgeny Proshin, Alexander Myasnikov, Zhu Ruijin,
Changmin Heo, Vitor Vidal, Stijn Crevits, Jean-Gregoire Foulon11 , Ben L., Etienne Khan, Norbert Szetei12 ,
7 https://news.ycombinator.com/item?id=17549050
8 Database Management Systems
9 Programming Language
10 goto-vlad@github
11 https://github.com/pixjuan
12 https://github.com/73696e65
xiii
Marc Remy, Michael Hansen, Derk Barten, The Renaissance13 , Hugo Chan, Emil Mursalimov, Tanner
Hoke, Tan90909090@GitHub, Ole Petter Orhagen, Sourav Punoriyar, Vitor Oliveira, Alexis Ehret, Maxim
Shlochiski, Greg Paton, Pierrick Lebourgeois.
For helping me in other ways: Andrew Zubinski, Arnaud Patard (rtp on #debian-arm IRC), noshadow on
#gcc IRC, Aliaksandr Autayeu, Mohsen Mostafa Jokar, Peter Sovietov, Misha “tiphareth” Verbitsky.
For translating the book into Simplified Chinese: Antiy Labs (antiy.cn), Archer.
For translating the book into Korean: Byungho Min.
For translating the book into Dutch: Cedric Sambre (AKA Midas).
For translating the book into Spanish: Diego Boy, Luis Alberto Espinosa Calvo, Fernando Guida, Diogo
Mussi, Patricio Galdames.
For translating the book into Portuguese: Thales Stevan de A. Gois, Diogo Mussi, Luiz Filipe.
For translating the book into Italian: Federico Ramondino14 , Paolo Stivanin15 , twyK, Fabrizio Bertone, Mat-
teo Sticco, Marco Negro16 .
For translating the book into French: Florent Besnard17 , Marc Remy18 , Baudouin Landais, Téo Dacquet19 ,
BlueSkeye@GitHub20 .
For translating the book into German: Dennis Siekmeier21 , Julius Angres22 , Dirk Loser23 , Clemens Tamme.
For translating the book into Polish: Kateryna Rozanova, Aleksander Mistewicz, Wiktoria Lewicka.
For translating the book into Japanese: shmz@github24 .
For proofreading: Alexander “Lstar” Chernenkiy, Vladimir Botov, Andrei Brazhuk, Mark “Logxen” Cooper,
Yuan Jochen Kang, Mal Malakov, Lewis Porter, Jarle Thorsen, Hong Xie.
Vasil Kolev25 did a great amount of work in proofreading and correcting many mistakes.
Thanks also to all the folks on github.com who have contributed notes and corrections.
Many LATEX packages were used: I would like to thank the authors as well.
Donors
Those who supported me during the time when I wrote significant part of the book:
2 * Oleg Vygovsky (50+100 UAH), Daniel Bilar ($50), James Truscott ($4.5), Luis Rocha ($63), Joris van
de Vis ($127), Richard S Shultz ($20), Jang Minchang ($20), Shade Atlas (5 AUD), Yao Xiao ($10), Pawel
Szczur (40 CHF), Justin Simms ($20), Shawn the R0ck ($27), Ki Chan Ahn ($50), Triop AB (100 SEK), Ange
Albertini (e10+50), Sergey Lukianov (300 RUR), Ludvig Gislason (200 SEK), Gérard Labadie (e40), Sergey
Volchkov (10 AUD), Vankayala Vigneswararao ($50), Philippe Teuwen ($4), Martin Haeberli ($10), Victor
Cazacov (e5), Tobias Sturzenegger (10 CHF), Sonny Thai ($15), Bayna AlZaabi ($75), Redfive B.V. (e25),
Joona Oskari Heikkilä (e5), Marshall Bishop ($50), Nicolas Werner (e12), Jeremy Brown ($100), Alexan-
dre Borges ($25), Vladimir Dikovski (e50), Jiarui Hong (100.00 SEK), Jim Di (500 RUR), Tan Vincent ($30),
Sri Harsha Kandrakota (10 AUD), Pillay Harish (10 SGD), Timur Valiev (230 RUR), Carlos Garcia Prado
(e10), Salikov Alexander (500 RUR), Oliver Whitehouse (30 GBP), Katy Moe ($14), Maxim Dyakonov ($3),
Sebastian Aguilera (e20), Hans-Martin Münch (e15), Jarle Thorsen (100 NOK), Vitaly Osipov ($100), Yuri
Romanov (1000 RUR), Aliaksandr Autayeu (e10), Tudor Azoitei ($40), Z0vsky (e10), Yu Dai ($10), Anony-
mous ($15), Vladislav Chelnokov ($25), Nenad Noveljic ($50), Ryan Smith ($25), Andreas Schommer (e5),
Nikolay Gavrilov ($300).
Thanks a lot to every donor!
13 https://github.com/TheRenaissance
14 https://github.com/pinkrab
15 https://github.com/paolostivanin
16 https://github.com/Internaut401
17 https://github.com/besnardf
18 https://github.com/mremy
19 https://github.com/T30rix
20 https://github.com/BlueSkeye
21 https://github.com/DSiekmeier
22 https://github.com/JAngres
23 https://github.com/PolymathMonkey
24 https://github.com/shmz
25 https://vasil.ludost.net/
xiv
mini-FAQ
Q: Is this book simpler/easier than others?
A: No, it is at about the same level as other books of this subject.
Q: I’m too frightened to start reading this book, there are more than 1000 pages. ”...for Beginners” in the
name sounds a bit sarcastic.
A: All sorts of listings are the bulk of the book. The book is indeed for beginners, there is a lot missing
(yet).
Q: What are the prerequisites for reading this book?
A: A basic understanding of C/C++ is desirable.
Q: Should I really learn x86/x64/ARM and MIPS at once? Isn’t it too much?
A: Starters can read about just x86/x64, while skipping or skimming the ARM and MIPS parts.
Q: Can I buy a Russian or English hard copy/paper book?
A: Unfortunately, no. No publisher got interested in publishing a Russian or English version so far. Mean-
while, you can ask your favorite copy shop to print and bind it. https://yurichev.com/news/20200222_
printed_RE4B/.
Q: Is there an epub or mobi version?
A: No. The book is highly dependent on TeX/LaTeX-specific hacks, so converting to HTML (epub/mobi are
a set of HTMLs) would not be easy.
Q: Why should one learn assembly language these days?
A: Unless you are an OS26 developer, you probably don’t need to code in assembly—the latest compilers
(2010s) are much better at performing optimizations than humans 27 .
Also, the latest CPU28 s are very complex devices, and assembly knowledge doesn’t really help towards
understand their internals.
That being said, there are at least two areas where a good understanding of assembly can be helpful:
First and foremost, for security/malware research. It is also a good way to gain a better understanding of
your compiled code while debugging. This book is therefore intended for those who want to understand
assembly language rather than to code in it, which is why there are many examples of compiler output
contained within.
Q: I clicked on a hyperlink inside a PDF-document, how do I go back?
A: In Adobe Acrobat Reader click Alt+LeftArrow. In Evince click “<” button.
Q: May I print this book / use it for teaching?
A: Of course! That’s why the book is licensed under the Creative Commons license (CC BY-SA 4.0).
Q: Why is this book free? You’ve done great job. This is suspicious, as with many other free things.
A: In my own experience, authors of technical literature write mostly for self-advertisement purposes. It’s
not possible to make any decent money from such work.
Q: How does one get a job in reverse engineering?
A: There are hiring threads that appear from time to time on reddit, devoted to RE29 . Try looking there.
A somewhat related hiring thread can be found in the “netsec” subreddit.
Q: Compilers’ versions in the book are outdated already...
A: No need to follow all steps precisely. Use the compilers you already have installed on your OS. Also,
there is: Compiler Explorer.
Q: I have a question...
A: Send it to me by email (<first_name @ last_name . com> or <first_name . last_name @ gmail . com>).
26 Operating System
27 A very good text on this topic: [Agner Fog, The microarchitecture of Intel, AMD and VIA CPUs, (2016)]
28 Central Processing Unit
29 reddit.com/r/ReverseEngineering/
xv
About the Korean translation
In January 2015, the Acorn publishing company (www.acornpub.co.kr) in South Korea did a huge amount
of work in translating and publishing this book (as it was in August 2014) into Korean.
It’s available now at their website.
The translator is Byungho Min (twitter/tais9). The cover art was done by the artistic Andy Nechaevsky, a
friend of the author: facebook/andydinka. Acorn also holds the copyright to the Korean translation.
So, if you want to have a real book on your shelf in Korean and want to support this work, it is now available
for purchase.
About the Persian/Farsi translation

In 2016 the book was translated by Mohsen Mostafa Jokar (who is also known to Iranian community for
his translation of Radare manual30 ). It is available on the publisher’s website31 (Pendare Pars).
Here is a link to a 40-page excerpt: https://beginners.re/farsi.pdf.
National Library of Iran registration information: http://opac.nlai.ir/opac-prod/bibliographic/4473995.
About the Chinese translation

In April 2017, translation to Chinese was completed by Chinese PTPress. They are also the Chinese trans-
lation copyright holders.
The Chinese version is available for order here: http://www.epubit.com.cn/book/details/4174. A par-
tial review and history behind the translation can be found here: http://www.cptoday.cn/news/detail/
3155.
The principal translator is Archer, to whom the author owes very much. He was extremely meticulous (in
a good sense) and reported most of the known mistakes and bugs, which is very important in literature
such as this book. The author would recommend his services to any other author!
The guys from Antiy Labs has also helped with translation. Here is preface written by them.
30 http://rada.re/get/radare2book-persian.pdf
31 http://goo.gl/2Tzx0H
xvi
Chapter 1
Code Patterns
1.1 The method

When the author of this book first started learning C and, later, C++, he used to write small pieces of
code, compile them, and then look at the assembly language output. This made it very easy for him to
understand what was going on in the code that he had written. 1 . He did this so many times that the
relationship between the C/C++ code and what the compiler produced was imprinted deeply in his mind.
It’s now easy for him to imagine instantly a rough outline of a C code’s appearance and function. Perhaps
this technique could be helpful for others.
By the way, there is a great website where you can do the same, with various compilers, instead of
installing them on your box. You can use it as well: https://godbolt.org/.
Exercises
When the author of this book studied assembly language, he also often compiled small C functions and
then rewrote them gradually to assembly, trying to make their code as short as possible. This probably is
not worth doing in real-world scenarios today, because it’s hard to compete with the latest compilers in
terms of efficiency. It is, however, a very good way to gain a better understanding of assembly. Feel free,
therefore, to take any assembly code from this book and try to make it shorter. However, don’t forget to
test what you have written.
Optimization levels and debug information

Source code can be compiled by different compilers with various optimization levels. A typical compiler
has about three such levels, where level zero means that optimization is completely disabled. Optimization
can also be targeted towards code size or code speed. A non-optimizing compiler is faster and produces
more understandable (albeit verbose) code, whereas an optimizing compiler is slower and tries to produce
code that runs faster (but is not necessarily more compact). In addition to optimization levels, a compiler
can include some debug information in the resulting file, producing code that is easy to debug. One of the
important features of the ´debug’ code is that it might contain links between each line of the source code
and its respective machine code address. Optimizing compilers, on the other hand, tend to produce output
where entire lines of source code can be optimized away and thus not even be present in the resulting
machine code. Reverse engineers can encounter either version, simply because some developers turn
on the compiler’s optimization flags and others do not. Because of this, we’ll try to work on examples of
both debug and release versions of the code featured in this book, wherever possible.
Sometimes some pretty ancient compilers are used in this book, in order to get the shortest (or simplest)
possible code snippet.
1 In
fact, he still does this when he can’t understand what a particular bit of code does. A recent example from the year 2019:
p += p+(i&1)+2; from the “SAT0W” SAT-solver by D.Knuth.
1
1.2 Some basics
1.2.1 A short introduction to the CPU
The CPU is the device that executes the machine code a program consists of.
A short glossary:
Instruction : A primitive CPU command. The simplest examples include: moving data between registers,
working with memory, primitive arithmetic operations. As a rule, each CPU has its own instruction
set architecture (ISA).
Machine code : Code that the CPU directly processes. Each instruction is usually encoded by several
bytes.
Assembly language : Mnemonic code and some extensions, like macros, that are intended to make a
programmer’s life easier.
CPU register : Each CPU has a fixed set of general purpose registers (GPR2 ). ≈ 8 in x86, ≈ 16 in x86-
64, and also ≈ 16 in ARM. The easiest way to understand a register is to think of it as an untyped
temporary variable. Imagine if you were working with a high-level PL and could only use eight 32-bit
(or 64-bit) variables. Yet a lot can be done using just these!
One might wonder why there needs to be a difference between machine code and a PL. The answer lies
in the fact that humans and CPUs are not alike—it is much easier for humans to use a high-level PL like
C/C++, Java, or Python, but it is easier for a CPU to use a much lower level of abstraction. Perhaps it
would be possible to invent a CPU that can execute high-level PL code, but it would be many times more
complex than the CPUs we know of today. In a similar fashion, it is very inconvenient for humans to write
in assembly language, due to it being so low-level and difficult to write in without making a huge number
of annoying mistakes. The program that converts the high-level PL code into assembly is called a compiler.
3
.
A couple of words about different ISAs
The x86 ISA has always had variable-length instructions, so when the 64-bit era came, the x64 extensions
did not impact the ISA very significantly. In fact, the x86 ISA still contains a lot of instructions that first
appeared in 16-bit 8086 CPU, yet are still found in the CPUs of today. ARM is a RISC4 CPU designed with
constant-length instructions in mind, which had some advantages in the past. In the very beginning, all
ARM instructions were encoded in 4 bytes5 . This is now referred to as “ARM mode”. Then they realized it
wasn’t as frugal as they first imagined. In fact, the most common CPU instructions6 in real world applica-
tions can be encoded using less information. They therefore added another ISA, called Thumb, in which
each instruction was encoded in just 2 bytes. This is now referred to as “Thumb mode”. However, not all
ARM instructions can be encoded in just 2 bytes, so the Thumb instruction set is somewhat limited. It is
worth noting that code compiled for ARM mode and Thumb mode can coexist within one single program.
The ARM creators thought Thumb could be extended, giving rise to Thumb-2, which appeared in ARMv7.
Thumb-2 still uses 2-byte instructions, but has some new instructions which have the size of 4 bytes. There
is a common misconception that Thumb-2 is a mix of ARM and Thumb. This is incorrect. Rather, Thumb-2
was extended to fully support all processor features so it could compete with ARM mode—a goal that
was clearly achieved, as the majority of applications for iPod/iPhone/iPad are compiled for the Thumb-2
instruction set. (Though, admittedly, this is largely due to the fact that Xcode does this by default). Later
the 64-bit ARM came out. This ISA has 4-byte instructions, and lacked the need of any additional Thumb
mode. However, the 64-bit requirements affected the ISA, resulting in us now having three ARM instruc-
tion sets: ARM mode, Thumb mode (including Thumb-2) and ARM64. These ISAs intersect partially, but
it can be said that they are different ISAs, rather than variations of the same one. Therefore, we will try
to add fragments of code in all three ARM ISAs in this book. There are, by the way, many other RISC ISAs
with fixed length 32-bit instructions, such as MIPS, PowerPC and Alpha AXP.
2 General Purpose Registers
3 Old-school Russian literature also uses the term “translator”.
4 Reduced Instruction Set Computing
5 Fixed-length instructions are handy because one can calculate the next (or previous) instruction address without effort. This
feature will be discussed in the switch() operator ( 1.21.2 on page 173) section.
6 e.g. MOV/PUSH/CALL/Jcc
2
1.2.2 Numeral Systems
Nowadays octal numbers seem to be used

for exactly one purpose—file permissions on
POSIX systems—but hexadecimal numbers
are widely used to emphasize the bit pattern
of a number over its numeric value.
Alan A. A. Donovan, Brian W. Kernighan —

The Go Programming Language
Humans have become accustomed to a decimal numeral system, probably because almost everyone has
10 fingers. Nevertheless, the number “10” has no significant meaning in science and mathematics. The
natural numeral system in digital electronics is binary: 0 is for an absence of current in the wire, and 1 for
presence. 10 in binary is 2 in decimal, 100 in binary is 4 in decimal, and so on.
If the numeral system has 10 digits, it has a radix (or base) of 10. The binary numeral system has a radix
of 2.
Important things to recall:
1) A number is a number, while a digit is a term from writing systems, and is usually one character
2) The value of a number does not change when converted to another radix; only the writing notation for
that value has changed (and therefore the way of representing it in RAM7 ).
1.2.3 Converting From One Radix To Another

Positional notation is used in almost every numerical system. This means that a digit has weight relative
to where it is placed inside of the larger number. If 2 is placed at the rightmost place, it’s 2, but if it’s
placed one digit before rightmost, it’s 20.
What does 1234 stand for?
103 ⋅ 1 + 102 ⋅ 2 + 101 ⋅ 3 + 1 ⋅ 4 = 1234 or 1000 ⋅ 1 + 100 ⋅ 2 + 10 ⋅ 3 + 4 = 1234
It’s the same story for binary numbers, but the base is 2 instead of 10. What does 0b101011 stand for?
25 ⋅ 1 + 24 ⋅ 0 + 23 ⋅ 1 + 22 ⋅ 0 + 21 ⋅ 1 + 20 ⋅ 1 = 43 or 32 ⋅ 1 + 16 ⋅ 0 + 8 ⋅ 1 + 4 ⋅ 0 + 2 ⋅ 1 + 1 = 43
There is such a thing as non-positional notation, such as the Roman numeral system. 8 . Perhaps, hu-
mankind switched to positional notation because it’s easier to do basic operations (addition, multiplication,
etc.) on paper by hand.
Binary numbers can be added, subtracted and so on in the very same as taught in schools, but only 2
digits are available.
Binary numbers are bulky when represented in source code and dumps, so that is where the hexadecimal
numeral system can be useful. A hexadecimal radix uses the digits 0..9, and also 6 Latin characters: A..F.
Each hexadecimal digit takes 4 bits or 4 binary digits, so it’s very easy to convert from binary number to
hexadecimal and back, even manually, in one’s mind.
hexadecimal binary decimal

0 0000 0
1 0001 1
2 0010 2
3 0011 3
4 0100 4
5 0101 5
6 0110 6
7 0111 7
8 1000 8
9 1001 9
A 1010 10
B 1011 11
C 1100 12
D 1101 13
E 1110 14
7 Random-Access Memory
8 About numeric system evolution, see [Donald E. Knuth, The Art of Computer Programming, Volume 2, 3rd ed., (1997), 195–213.]
3
F 1111 15
How can one tell which radix is being used in a specific instance?
Decimal numbers are usually written as is, i.e., 1234. Some assemblers allow an identifier on decimal
radix numbers, in which the number would be written with a ”d” suffix: 1234d.
Binary numbers are sometimes prepended with the ”0b” prefix: 0b100110111 (GCC9 has a non-standard
language extension for this10 ). There is also another way: using a ”b” suffix, for example: 100110111b.
This book tries to use the ”0b” prefix consistently throughout the book for binary numbers.
Hexadecimal numbers are prepended with ”0x” prefix in C/C++ and other PLs: 0x1234ABCD. Alternatively,
they are given a ”h” suffix: 1234ABCDh. This is common way of representing them in assemblers and
debuggers. In this convention, if the number is started with a Latin (A..F) digit, a 0 is added at the
beginning: 0ABCDEFh. There was also convention that was popular in 8-bit home computers era, using $
prefix, like $ABCD. The book will try to stick to ”0x” prefix throughout the book for hexadecimal numbers.
Should one learn to convert numbers mentally? A table of 1-digit hexadecimal numbers can easily be
memorized. As for larger numbers, it’s probably not worth tormenting yourself.
Perhaps the most visible hexadecimal numbers are in URL11 s. This is the way that non-Latin characters are
encoded. For example: https://en.wiktionary.org/wiki/na%C3%AFvet%C3%A9 is the URL of Wiktionary
article about “naïveté” word.
Octal Radix
Another numeral system heavily used in the past of computer programming is octal. In octal there are 8
digits (0..7), and each is mapped to 3 bits, so it’s easy to convert numbers back and forth. It has been
superseded by the hexadecimal system almost everywhere, but, surprisingly, there is a *NIX utility, used
often by many people, which takes octal numbers as argument: chmod.
As many *NIX users know, chmod argument can be a number of 3 digits. The first digit represents the
rights of the owner of the file (read, write and/or execute), the second is the rights for the group to which
the file belongs, and the third is for everyone else. Each digit that chmod takes can be represented in
binary form:
decimal binary meaning

7 111 rwx
6 110 rw-
5 101 r-x
4 100 r--
3 011 -wx
2 010 -w-
1 001 --x
0 000 ---
So each bit is mapped to a flag: read/write/execute.

The importance of chmod here is that the whole number in argument can be represented as octal number.
Let’s take, for example, 644. When you run chmod 644 file, you set read/write permissions for owner,
read permissions for group and again, read permissions for everyone else. If we convert the octal number
644 to binary, it would be 110100100, or, in groups of 3 bits, 110 100 100.
Now we see that each triplet describe permissions for owner/group/others: first is rw-, second is r-- and
third is r--.
The octal numeral system was also popular on old computers like PDP-8, because word there could be 12,
24 or 36 bits, and these numbers are all divisible by 3, so the octal system was natural in that environment.
Nowadays, all popular computers employ word/address sizes of 16, 32 or 64 bits, and these numbers are
all divisible by 4, so the hexadecimal system is more natural there.
9 GNU Compiler Collection
10 https://gcc.gnu.org/onlinedocs/gcc/Binary-constants.html
11 Uniform Resource Locator
4
The octal numeral system is supported by all standard C/C++ compilers. This is a source of confusion
sometimes, because octal numbers are encoded with a zero prepended, for example, 0377 is 255. Some-
times, you might make a typo and write ”09” instead of 9, and the compiler would report an error. GCC
might report something like this:
error: invalid digit "9" in octal constant.
Also, the octal system is somewhat popular in Java. When the IDA shows Java strings with non-printable
characters, they are encoded in the octal system instead of hexadecimal. The JAD Java decompiler be-
haves the same way.
Divisibility
When you see a decimal number like 120, you can quickly deduce that it’s divisible by 10, because the
last digit is zero. In the same way, 123400 is divisible by 100, because the two last digits are zeros.
Likewise, the hexadecimal number 0x1230 is divisible by 0x10 (or 16), 0x123000 is divisible by 0x1000
(or 4096), etc.
The binary number 0b1000101000 is divisible by 0b1000 (8), etc.
This property can often be used to quickly realize if an address or a size of some block in memory is padded
to some boundary. For example, sections in PE12 files are almost always started at addresses ending with
3 hexadecimal zeros: 0x41000, 0x10001000, etc. The reason behind this is the fact that almost all PE
sections are padded to a boundary of 0x1000 (4096) bytes.
Multi-Precision Arithmetic and Radix
Multi-precision arithmetic can use huge numbers, and each one may be stored in several bytes. For
example, RSA keys, both public and private, span up to 4096 bits, and maybe even more.
In [Donald E. Knuth, The Art of Computer Programming, Volume 2, 3rd ed., (1997), 265] we find the
following idea: when you store a multi-precision number in several bytes, the whole number can be
represented as having a radix of 28 = 256, and each digit goes to the corresponding byte. Likewise, if you
store a multi-precision number in several 32-bit integer values, each digit goes to each 32-bit slot, and
you may think about this number as stored in radix of 232 .
How to Pronounce Non-Decimal Numbers
Numbers in a non-decimal base are usually pronounced by digit by digit: “one-zero-zero-one-one-...”.

Words like “ten” and “thousand” are usually not pronounced, to prevent confusion with the decimal base
system.
Floating point numbers
To distinguish floating point numbers from integers, they are usually written with “.0” at the end, like 0.0,
123.0, etc.
1.3 An Empty Function

The simplest possible function is arguably one that does nothing:
Listing 1.1: C/C++ Code

void f()
{
return;
};
Let’s compile it!

12 Portable Executable
5
1.3.1 x86
Here’s what both the GCC and MSVC compilers produce on the x86 platform:
Listing 1.2: Optimizing GCC/MSVC (assembly output)

f:
ret
There is just one instruction: RET, which returns execution to the caller.
1.3.2 ARM
Listing 1.3: Optimizing Keil 6/2013 (ARM mode) assembly output
f PROC
BX lr
ENDP
The return address is not saved on the local stack in the ARM ISA, but rather in the link register, so the BX
LR instruction causes execution to jump to that address—effectively returning execution to the caller.
1.3.3 MIPS
There are two naming conventions used in the world of MIPS when naming registers: by number (from $0
to $31) or by pseudo name ($V0, $A0, etc.).
The GCC assembly output below lists registers by number:
Listing 1.4: Optimizing GCC 4.4.5 (assembly output)

j $31
nop
…while IDA13 does it by pseudo name:
Listing 1.5: Optimizing GCC 4.4.5 (IDA)

j $ra
nop
The first instruction is the jump instruction (J or JR) which returns the execution flow to the caller, jumping
to the address in the $31 (or $RA) register.
This is the register analogous to LR14 in ARM.
The second instruction is NOP15 , which does nothing. We can ignore it for now.
A Note About MIPS Instructions and Register Names
Register and instruction names in the world of MIPS are traditionally written in lowercase. However, for
the sake of consistency, this book will stick to using uppercase letters, as it is the convention followed by
all the other ISAs featured in this book.
1.3.4 Empty Functions in Practice

Despite the fact empty functions seem useless, they are quite frequent in low-level code.
First of all, they are quite popular in debugging functions, like this one:
Listing 1.6: C/C++ code

void dbg_print (const char *fmt, ...)
{
#ifdef _DEBUG
// open log file
13 Interactive Disassembler and Debugger developed by Hex-Rays
14 LinkRegister
15 No Operation
6
// write to log file
// close log file
#endif
};
void some_function()
{
...
dbg_print ("we did something\n");
...
};
In a non-debug build (as in a “release”), _DEBUG is not defined, so the dbg_print() function, despite still
being called during execution, will be empty.
Similarly, a popular method of software protection is to make one build for legal customers, and another
demo build. A demo build can lack some important functions, as with this example:
Listing 1.7: C/C++ code

void save_file ()
{
#ifndef DEMO
// a real saving code
#endif
};
The save_file() function can be called when the user clicks File->Save on the menu. The demo version
may be delivered with this menu item disabled, but even if a software cracker will enable it, only an empty
function with no useful code will be called.
IDA marks such functions with names like nullsub_00, nullsub_01, etc.
1.4 Returning Values

Another simple function is the one that simply returns a constant value:

int f()
{
return 123;
};
Let’s compile it.
1.4.1 x86
Here’s what both the GCC and MSVC compilers produce (with optimization) on the x86 platform:
Listing 1.9: Optimizing GCC/MSVC (assembly output)

f:
mov eax, 123
ret
There are just two instructions: the first places the value 123 into the EAX register, which is used by
convention for storing the return value, and the second one is RET, which returns execution to the caller.
The caller will take the result from the EAX register.
1.4.2 ARM
There are a few differences on the ARM platform:
7
Listing 1.10: Optimizing Keil 6/2013 (ARM mode) ASM Output
f PROC
MOV r0,#0x7b ; 123
BX lr
ENDP
ARM uses the register R0 for returning the results of functions, so 123 is copied into R0.
It is worth noting that MOV is a misleading name for the instruction in both the x86 and ARM ISAs.
The data is not in fact moved, but copied.
1.4.3 MIPS
The GCC assembly output below lists registers by number:

j $31
li $2,123 # 0x7b
…while IDA does it by their pseudo names:

jr $ra
li $v0, 0x7B
The $2 (or $V0) register is used to store the function’s return value. LI stands for “Load Immediate” and
is the MIPS equivalent to MOV.
The other instruction is the jump instruction (J or JR) which returns the execution flow to the caller.
You might be wondering why the positions of the load instruction (LI) and the jump instruction (J or JR) are
swapped. This is due to a RISC feature called “branch delay slot”.
The reason this happens is a quirk in the architecture of some RISC ISAs and isn’t important for our
purposes—we must simply keep in mind that in MIPS, the instruction following a jump or branch instruction
is executed before the jump/branch instruction itself.
As a consequence, branch instructions always swap places with the instruction executed immediately
beforehand.
In practice, functions which merely return 1 (true) or 0 (false) are very frequent.
The smallest ever of the standard UNIX utilities, /bin/true and /bin/false return 0 and 1 respectively, as an
exit code. (Zero as an exit code usually means success, non-zero means error.)
1.5 Hello, world!

Let’s use the famous example from the book [Brian W. Kernighan, Dennis M. Ritchie, The C Programming
Language, 2ed, (1988)]:

#include <stdio.h>
int main()
{
printf("hello, world\n");
return 0;
}
1.5.1 x86
MSVC
Let’s compile it in MSVC 2010:
8
cl 1.cpp /Fa1.asm
(The /Fa option instructs the compiler to generate an assembly listing file)
Listing 1.14: MSVC 2010

CONST SEGMENT
$SG3830 DB 'hello, world', 0AH, 00H
CONST ENDS
PUBLIC _main
EXTRN _printf:PROC
; Function compile flags: /Odtp
_TEXT SEGMENT
_main PROC
push ebp
mov ebp, esp
push OFFSET $SG3830
call _printf
add esp, 4
xor eax, eax
pop ebp
ret 0
_main ENDP
_TEXT ENDS
MSVC produces assembly listings in Intel-syntax. The differences between Intel-syntax and AT&T-syntax
will be discussed in 1.5.1 on page 11.
The compiler generated the file, 1.obj, which is to be linked into 1.exe. In our case, the file contains two
segments: CONST (for data constants) and _TEXT (for code).
The string hello, world in C/C++ has type const char[][Bjarne Stroustrup, The C++ Programming
Language, 4th Edition, (2013)p176, 7.3.2], but it does not have its own name. The compiler needs to deal
with the string somehow, so it defines the internal name $SG3830 for it.
That is why the example may be rewritten as follows:
#include <stdio.h>
const char $SG3830[]="hello, world\n";
int main()
{
printf($SG3830);
return 0;
}
Let’s go back to the assembly listing. As we can see, the string is terminated by a zero byte, which is
standard for C/C++ strings. More about C/C++ strings: 5.4.1 on page 697.
In the code segment, _TEXT, there is only one function so far: main(). The function main() starts with
prologue code and ends with epilogue code (like almost any function) 16 .
After the function prologue we see the call to the printf() function:
CALL _printf. Before the call, a string address (or a pointer to it) containing our greeting is placed on
the stack with the help of the PUSH instruction.
When the printf() function returns the control to the main() function, the string address (or a pointer
to it) is still on the stack. Since we do not need it anymore, the stack pointer (the ESP register) needs to
be corrected.
ADD ESP, 4 means add 4 to the ESP register value.
Why 4? Since this is a 32-bit program, we need exactly 4 bytes for address passing through the stack. If it
was x64 code we would need 8 bytes. ADD ESP, 4 is effectively equivalent to POP register but without
using any register17 .
For the same purpose, some compilers (like the Intel C++ Compiler) may emit POP ECX instead of ADD
(e.g., such a pattern can be observed in the Oracle RDBMS code as it is compiled with the Intel C++
16 You can read more about it in the section about function prologues and epilogues ( 1.6 on page 28).
17 CPU flags, however, are modified
9
compiler). This instruction has almost the same effect but the ECX register contents will be overwritten.
The Intel C++ compiler supposedly uses POP ECX since this instruction’s opcode is shorter than ADD ESP,
x (1 byte for POP against 3 for ADD).
Here is an example of using POP instead of ADD from Oracle RDBMS:
Listing 1.15: Oracle RDBMS 10.2 Linux (app.o file)

.text:0800029A push ebx
.text:0800029B call qksfroChild
.text:080002A0 pop ecx
However, MSVC can do the same.
Listing 1.16: MineSweeper from Windows 7 32-bit

.text:0102106F push 0
.text:01021071 call ds:time
.text:01021077 pop ecx
After calling printf(), the original C/C++ code contains the statement return 0 —return 0 as the result
of the main() function.
In the generated code this is implemented by the instruction XOR EAX, EAX.
XOR is in fact just “eXclusive OR”18 but the compilers often use it instead of MOV EAX, 0—again because
it is a slightly shorter opcode (2 bytes for XOR against 5 for MOV).
Some compilers emit SUB EAX, EAX, which means SUBtract the value in the EAX from the value in EAX.
That in any case will results in zero.
The last instruction RET returns the control to the caller. Usually, this is C/C++ CRT19 code which in turn
returns control to the OS.
GCC
Now let’s try to compile the same C/C++ code in the GCC 4.4.1 compiler in Linux: gcc 1.c -o 1. Next,
with the assistance of the IDA disassembler, let’s see how the main() function was created. IDA, like
MSVC, uses Intel-syntax20 .
Listing 1.17: code in IDA

main proc near
var_10 = dword ptr -10h
push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
sub esp, 10h
mov eax, offset aHelloWorld ; "hello, world\n"
mov [esp+10h+var_10], eax
call _printf
mov eax, 0
leave
retn
main endp
The result is almost the same. The address of the hello, world string (stored in the data segment) is
loaded in the EAX register first, and then saved onto the stack.
In addition, the function prologue has AND ESP, 0FFFFFFF0h —this instruction aligns the ESP register
value on a 16-byte boundary. This results in all values in the stack being aligned the same way (The CPU
performs better if the values it is dealing with are located in memory at addresses aligned on a 4-byte or
16-byte boundary).
SUB ESP, 10h allocates 16 bytes on the stack. Although, as we can see hereafter, only 4 are necessary
here.
18 Wikipedia
19 C Runtime library
20 We could also have GCC produce assembly listings in Intel-syntax by applying the options -S -masm=intel.
10
This is because the size of the allocated stack is also aligned on a 16-byte boundary.
The string address (or a pointer to the string) is then stored directly onto the stack without using the PUSH
instruction. var_10 —is a local variable and is also an argument for printf(). Read about it below.
Then the printf() function is called.
Unlike MSVC, when GCC is compiling without optimization turned on, it emits MOV EAX, 0 instead of a
shorter opcode.
The last instruction, LEAVE —is the equivalent of the MOV ESP, EBP and POP EBP instruction pair —in other
words, this instruction sets the stack pointer (ESP) back and restores the EBP register to its initial state.
This is necessary since we modified these register values (ESP and EBP) at the beginning of the function
(by executing MOV EBP, ESP / AND ESP, …).
GCC: AT&T syntax
Let’s see how this can be represented in assembly language AT&T syntax. This syntax is much more
popular in the UNIX-world.
Listing 1.18: let’s compile in GCC 4.7.3

gcc -S 1_1.c
We get this:
Listing 1.19: GCC 4.7.3

.file "1_1.c"
.section .rodata
.LC0:
.string "hello, world\n"
.text
.globl main
.type main, @function
main:
.LFB0:
.cfi_startproc
pushl %ebp
.cfi_def_cfa_offset 8
.cfi_offset 5, -8
movl %esp, %ebp
.cfi_def_cfa_register 5
andl $-16, %esp
subl $16, %esp
movl $.LC0, (%esp)
call printf
movl $0, %eax
leave
.cfi_restore 5
.cfi_def_cfa 4, 4
ret
.cfi_endproc
.LFE0:
.size main, .-main
.ident "GCC: (Ubuntu/Linaro 4.7.3-1ubuntu1) 4.7.3"
.section .note.GNU-stack,"",@progbits
The listing contains many macros (the parts that begin with a dot). These are not interesting for us at the
moment.
For now, for the sake of simplicity, we can ignore them (except the .string macro which encodes a null-
terminated character sequence just like a C-string). Then we’ll see this 21 :

.LC0:
main:
pushl %ebp
21 This GCC option can be used to eliminate “unnecessary” macros: -fno-asynchronous-unwind-tables
11
movl %esp, %ebp
andl $-16, %esp
subl $16, %esp
movl $.LC0, (%esp)
call printf
movl $0, %eax
leave
ret
Some of the major differences between Intel and AT&T syntax are:
• Source and destination operands are written in opposite order.
In Intel-syntax: <instruction> <destination operand> <source operand>.
In AT&T syntax: <instruction> <source operand> <destination operand>.
Here is an easy way to memorize the difference: when you deal with Intel-syntax, you can imagine
that there is an equality sign (=) between operands and when you deal with AT&T-syntax imagine
there is a right arrow (→) 22 .
• AT&T: Before register names, a percent sign must be written (%) and before numbers a dollar sign
($). Parentheses are used instead of brackets.
• AT&T: A suffix is added to instructions to define the operand size:
– q — quad (64 bits)
– l — long (32 bits)
– w — word (16 bits)
– b — byte (8 bits)
To go back to the compiled result: it is almost identical to what was displayed by IDA. There is one subtle
difference: 0FFFFFFF0h is presented as $-16. It’s the same thing: 16 in the decimal system is 0x10 in
hexadecimal. -0x10 is equal to 0xFFFFFFF0 (for a 32-bit data type).
One more thing: the return value is set to 0 by using the usual MOV, not XOR. MOV just loads a value to a
register. Its name is a misnomer (as the data is not moved but rather copied). In other architectures, this
instruction is named “LOAD” or “STORE” or something similar.
String patching (Win32)
We can easily find the “hello, world” string in the executable file using Hiew:
Figure 1.1: Hiew
And we can try to translate our message into Spanish:

22 By the way, in some C standard functions (e.g., memcpy(), strcpy()) the arguments are listed in the same way as in Intel-syntax:
first the pointer to the destination memory block, and then the pointer to the source memory block.
12
Figure 1.2: Hiew
The Spanish text is one byte shorter than English, so we also added the 0x0A byte at the end (\n) with a
zero byte.
It works.
What if we want to insert a longer message? There are some zero bytes after original English text. It’s
hard to say if they can be overwritten: they may be used somewhere in CRT code, or maybe not. Anyway,
only overwrite them if you really know what you’re doing.
String patching (Linux x64)
Let’s try to patch a Linux x64 executable using rada.re:
Listing 1.21: rada.re session

dennis@bigbox ~/tmp % gcc hw.c
dennis@bigbox ~/tmp % radare2 a.out

-- SHALL WE PLAY A GAME?
[0x00400430]> / hello
Searching 5 bytes from 0x00400000 to 0x00601040: 68 65 6c 6c 6f
Searching 5 bytes in [0x400000-0x601040]
hits: 1
0x004005c4 hit0_0 .HHhello, world;0.
[0x00400430]> s 0x004005c4
[0x004005c4]> px
- offset - 0 1 2 3 4 5 6 7 8 9 A B C D E F 0123456789ABCDEF
0x004005c4 6865 6c6c 6f2c 2077 6f72 6c64 0000 0000 hello, world....
0x004005d4 011b 033b 3000 0000 0500 0000 1cfe ffff ...;0...........
0x004005e4 7c00 0000 5cfe ffff 4c00 0000 52ff ffff |...\...L...R...
0x004005f4 a400 0000 6cff ffff c400 0000 dcff ffff ....l...........
0x00400604 0c01 0000 1400 0000 0000 0000 017a 5200 .............zR.
0x00400614 0178 1001 1b0c 0708 9001 0710 1400 0000 .x..............
0x00400624 1c00 0000 08fe ffff 2a00 0000 0000 0000 ........*.......
0x00400634 0000 0000 1400 0000 0000 0000 017a 5200 .............zR.
0x00400644 0178 1001 1b0c 0708 9001 0000 2400 0000 .x..........$...
0x00400654 1c00 0000 98fd ffff 3000 0000 000e 1046 ........0......F
0x00400664 0e18 4a0f 0b77 0880 003f 1a3b 2a33 2422 ..J..w...?.;*3$"
0x00400674 0000 0000 1c00 0000 4400 0000 a6fe ffff ........D.......
0x00400684 1500 0000 0041 0e10 8602 430d 0650 0c07 .....A....C..P..
0x00400694 0800 0000 4400 0000 6400 0000 a0fe ffff ....D...d.......
0x004006a4 6500 0000 0042 0e10 8f02 420e 188e 0345 e....B....B....E
0x004006b4 0e20 8d04 420e 288c 0548 0e30 8606 480e . ..B.(..H.0..H.
[0x004005c4]> oo+
File a.out reopened in read-write mode
[0x004005c4]> w hola, mundo\x00
[0x004005c4]> q
13
dennis@bigbox ~/tmp % ./a.out
hola, mundo
Here’s what’s going on: I searched for the “hello” string using the / command, then I set the cursor (seek,
in rada.re terms) to that address. Then I want to be sure that this is really that place: px dumps bytes
there. oo+ switches rada.re to read-write mode. w writes an ASCII string at the current seek. Note the \00
at the end—this is a zero byte. q quits.
This is a real story of software cracking
An image processing software, when not registered, added watermarks, like “This image was processed
by evaluation version of [software name]”, across a picture. We tried at random: we found that string
in the executable file and put spaces instead of it. Watermarks disappeared. Technically speaking, they
continued to appear. With the help of Qt functions, the watermark was still added to the resulting image.
But adding spaces didn’t alter the image itself...
Software localization of MS-DOS era
This method was a common way to translate MS-DOS software to Russian language back to 1980’s and
1990’s. This technique is available even for those who are not aware of machine code and executable file
formats. The new string shouldn’t be bigger than the old one, because there’s a risk of overwriting another
value or code there. Russian words and sentences are usually slightly longer than its English counterparts,
so that is why localized software has a lot of weird acronyms and hardly readable abbreviations.
Figure 1.3: Localized Norton Commander 5.51
Perhaps this also happened to other languages during that era, in other countries.
As for Delphi strings, the string’s size must also be corrected, if needed.
1.5.2 x86-64
MSVC: x86-64
Let’s also try 64-bit MSVC:
Listing 1.22: MSVC 2012 x64

$SG2989 DB 'hello, world', 0AH, 00H
main PROC
sub rsp, 40
lea rcx, OFFSET FLAT:$SG2989
call printf
xor eax, eax
add rsp, 40
ret 0
main ENDP
In x86-64, all registers were extended to 64-bit, and now their names have an R- prefix. In order to use the
stack less often (in other words, to access external memory/cache less often), there is a popular way to
pass function arguments via registers (fastcall) 6.1.3 on page 726. I.e., a part of the function’s arguments
are passed in registers, and the rest—via the stack. In Win64, 4 function arguments are passed in the RCX,
RDX, R8, and R9 registers. That is what we see here: a pointer to the string for printf() is now passed not
in the stack, but rather in the RCX register. The pointers are 64-bit now, so they are passed in the 64-bit
registers (which have the R- prefix). However, for backward compatibility, it is still possible to access the
32-bit parts, using the E- prefix. This is how the RAX/EAX/AX/AL register looks like in x86-64:
14
Byte number:
7th 6th 5th 4th 3rd 2nd 1st 0th
RAXx64
EAX
AX
AH AL
The main() function returns an int-typed value, which in C/C++ is still 32-bit, for better backward com-
patibility and portability, so that is why the EAX register is cleared at the function end (i.e., the 32-bit part
of the register) instead of with RAX. There are also 40 bytes allocated in the local stack. This is called the
“shadow space”, which we’ll talk about later: 1.14.2 on page 100.
GCC: x86-64
Let’s also try GCC in 64-bit Linux:
Listing 1.23: GCC 4.4.6 x64

main:
sub rsp, 8
mov edi, OFFSET FLAT:.LC0 ; "hello, world\n"
xor eax, eax ; number of vector registers passed
call printf
xor eax, eax
add rsp, 8
ret
Linux, *BSD and Mac OS X also use a method to pass function arguments in registers. [Michael Matz,
Jan Hubicka, Andreas Jaeger, Mark Mitchell, System V Application Binary Interface. AMD64 Architecture
Processor Supplement, (2013)] 23 .
The first 6 arguments are passed in the RDI, RSI, RDX, RCX, R8, and R9 registers, and the rest—via the
stack.
So the pointer to the string is passed in EDI (the 32-bit part of the register). Why doesn’t it use the 64-bit
part, RDI?
It is important to keep in mind that all MOV instructions in 64-bit mode that write something into the lower
32-bit register part also clear the higher 32-bits (as stated in Intel manuals: 12.1.4 on page 999).
I.e., the MOV EAX, 011223344h writes a value into RAX correctly, since the higher bits will be cleared.
24
If we open the compiled object file (.o), we can also see all the instructions’ opcodes :

.text:00000000004004D0 main proc near
.text:00000000004004D0 48 83 EC 08 sub rsp, 8
.text:00000000004004D4 BF E8 05 40 00 mov edi, offset format ; "hello, world\n"
.text:00000000004004D9 31 C0 xor eax, eax
.text:00000000004004DB E8 D8 FE FF FF call _printf
.text:00000000004004E0 31 C0 xor eax, eax
.text:00000000004004E2 48 83 C4 08 add rsp, 8
.text:00000000004004E6 C3 retn
.text:00000000004004E6 main endp
As we can see, the instruction that writes into EDI at 0x4004D4 occupies 5 bytes. The same instruction
writing a 64-bit value into RDI occupies 7 bytes. Apparently, GCC is trying to save some space. Besides,
it can be sure that the data segment containing the string will not be allocated at the addresses higher
than 4GiB.
We also see that the EAX register has been cleared before the printf() function call. This is done because
according to ABI25 standard mentioned above, the number of used vector registers is to be passed in EAX
in *NIX systems on x86-64.
23 Also available as https://software.intel.com/sites/default/files/article/402129/mpx-linux64-abi.pdf
24 Thismust be enabled in Options → Disassembly → Number of opcode bytes
25 Application Binary Interface
15
Address patching (Win64)
If our example was compiled in MSVC 2013 using /MD switch (meaning a smaller executable due to
MSVCR*.DLL file linkage), the main() function comes first, and can be easily found:
Figure 1.4: Hiew
As an experiment, we can increment address by 1:
16
Figure 1.5: Hiew
Hiew shows “ello, world”. And when we run the patched executable, this very string is printed.
Pick another string from binary image (Linux x64)
The binary file I’ve got when I compile our example using GCC 5.4.0 on Linux x64 box has many other text
strings. They are mostly imported function names and library names.
Run objdump to get the contents of all sections of the compiled file:
$ objdump -s a.out
a.out: file format elf64-x86-64
Contents of section .interp:

400238 2f6c6962 36342f6c 642d6c69 6e75782d /lib64/ld-linux-
400248 7838362d 36342e73 6f2e3200 x86-64.so.2.
Contents of section .note.ABI-tag:
400254 04000000 10000000 01000000 474e5500 ............GNU.
400264 00000000 02000000 06000000 20000000 ............ ...
Contents of section .note.gnu.build-id:
400274 04000000 14000000 03000000 474e5500 ............GNU.
400284 fe461178 5bb710b4 bbf2aca8 5ec1ec10 .F.x[.......^...
400294 cf3f7ae4 .?z.
...
It’s not a problem to pass address of the text string “/lib64/ld-linux-x86-64.so.2” to printf():
#include <stdio.h>
int main()
17
{
printf(0x400238);
return 0;
}
It’s hard to believe, but this code prints the aforementioned string.
If you would change the address to 0x400260, the “GNU” string would be printed. This address is true
for my specific GCC version, GNU toolset, etc. On your system, the executable may be slightly different,
and all addresses will also be different. Also, adding/removing code to/from this source code will probably
shift all addresses back or forward.
1.5.3 ARM
For my experiments with ARM processors, several compilers were used:
• Popular in the embedded area: Keil Release 6/2013.
26
• Apple Xcode 4.6.3 IDE with the LLVM-GCC 4.2 compiler .
• GCC 4.9 (Linaro) (for ARM64), available as win32-executables at http://go.yurichev.com/17325.
32-bit ARM code is used (including Thumb and Thumb-2 modes) in all cases in this book, if not mentioned
otherwise. When we talk about 64-bit ARM here, we call it ARM64.
Non-optimizing Keil 6/2013 (ARM mode)
Let’s start by compiling our example in Keil:

armcc.exe --arm --c90 -O0 1.c
The armcc compiler produces assembly listings in Intel-syntax, but it has high-level ARM-processor related
macros 27 , but it is more important for us to see the instructions “as is” so let’s see the compiled result in
IDA.
Listing 1.25: Non-optimizing Keil 6/2013 (ARM mode) IDA

.text:00000000 main
.text:00000000 10 40 2D E9 STMFD SP!, {R4,LR}
.text:00000004 1E 0E 8F E2 ADR R0, aHelloWorld ; "hello, world"
.text:00000008 15 19 00 EB BL __2printf
.text:0000000C 00 00 A0 E3 MOV R0, #0
.text:00000010 10 80 BD E8 LDMFD SP!, {R4,PC}
.text:000001EC 68 65 6C 6C+aHelloWorld DCB "hello, world",0 ; DATA XREF: main+4
In the example, we can easily see each instruction has a size of 4 bytes. Indeed, we compiled our code
for ARM mode, not for Thumb.
The very first instruction, STMFD SP!, {R4,LR}28 , works as an x86 PUSH instruction, writing the values of
two registers (R4 and LR) into the stack.
Indeed, in the output listing from the armcc compiler, for the sake of simplification, actually shows the
PUSH {r4,lr} instruction. But that is not quite precise. The PUSH instruction is only available in Thumb
mode. So, to make things less confusing, we’re doing this in IDA.
This instruction first decrements the SP30 so it points to the place in the stack that is free for new entries,
then it saves the values of the R4 and LR registers at the address stored in the modified SP.
This instruction (like the PUSH instruction in Thumb mode) is able to save several register values at once
which can be very useful. By the way, this has no equivalent in x86. It can also be noted that the STMFD
instruction is a generalization of the PUSH instruction (extending its features), since it can work with any
register, not just with SP. In other words, STMFD may be used for storing a set of registers at the specified
memory address.
26 Itis indeed so: Apple Xcode 4.6.3 uses open-source GCC as front-end compiler and LLVM code generator
27 e.g. ARM mode lacks PUSH/POP instructions
28 STMFD29
30 stack pointer. SP/ESP/RSP in x86/x64. SP in ARM.
18
The ADR R0, aHelloWorld instruction adds or subtracts the value in the PC31 register to the offset where
the hello, world string is located. How is the PC register used here, one might ask? This is called
“position-independent code”32 .
Such code can be executed at a non-fixed address in memory. In other words, this is PC-relative addressing.
The ADR instruction takes into account the difference between the address of this instruction and the
address where the string is located. This difference (offset) is always to be the same, no matter at what
address our code is loaded by the OS. That’s why all we need is to add the address of the current instruction
(from PC) in order to get the absolute memory address of our C-string.
BL __2printf33 instruction calls the printf() function. Here’s how this instruction works:
• store the address following the BL instruction (0xC) into the LR;
• then pass the control to printf() by writing its address into the PC register.
When printf() finishes its execution it must have information about where it needs to return the control
to. That’s why each function passes control to the address stored in the LR register.
That is a difference between “pure” RISC-processors like ARM and CISC34 -processors like x86, where the
return address is usually stored on the stack. Read more about this in next section ( 1.9 on page 29).
By the way, an absolute 32-bit address or offset cannot be encoded in the 32-bit BL instruction because
it only has space for 24 bits. As we may recall, all ARM-mode instructions have a size of 4 bytes (32 bits).
Hence, they can only be located on 4-byte boundary addresses. This implies that the last 2 bits of the
instruction address (which are always zero bits) may be omitted. In summary, we have 26 bits for offset
encoding. This is enough to encode current_P C ± ≈ 32M .
Next, the MOV R0, #035 instruction just writes 0 into the R0 register. That’s because our C-function returns
0 and the return value is to be placed in the R0 register.
The last instruction LDMFD SP!, R4,PC36 . It loads values from the stack (or any other memory place) in
order to save them into R4 and PC, and increments the stack pointer SP. It works like POP here.
N.B. The very first instruction STMFD saved the R4 and LR registers pair on the stack, but R4 and PC are
restored during the LDMFD execution.
As we already know, the address of the place where each function must return control to is usually saved
in the LR register. The very first instruction saves its value in the stack because the same register will
be used by our main() function when calling printf(). In the function’s end, this value can be written
directly to the PC register, thus passing control to where our function has been called.
Since main() is usually the primary function in C/C++, the control will be returned to the OS loader or to
a point in a CRT, or something like that.
All that allows omitting the BX LR instruction at the end of the function.
DCB is an assembly language directive defining an array of bytes or ASCII strings, akin to the DB directive
in the x86-assembly language.
Non-optimizing Keil 6/2013 (Thumb mode)
Let’s compile the same example using Keil in Thumb mode:

armcc.exe --thumb --c90 -O0 1.c
We are getting (in IDA):

Listing 1.26: Non-optimizing Keil 6/2013 (Thumb mode) + IDA
.text:00000000 main
.text:00000000 10 B5 PUSH {R4,LR}
.text:00000002 C0 A0 ADR R0, aHelloWorld ; "hello, world"
.text:00000004 06 F0 2E F9 BL __2printf
.text:00000008 00 20 MOVS R0, #0
.text:0000000A 10 BD POP {R4,PC}
.text:00000304 68 65 6C 6C+aHelloWorld DCB "hello, world",0 ; DATA XREF: main+2
31 Program Counter. IP/EIP/RIP in x86/64. PC in ARM.

32 Read more about it in relevant section ( 6.4.1 on page 739)
33 Branch with Link
34 Complex Instruction Set Computing
35 Meaning MOVe
36 LDMFD37 is an inverse instruction of STMFD
19
We can easily spot the 2-byte (16-bit) opcodes. This is, as was already noted, Thumb. The BL instruction,
however, consists of two 16-bit instructions. This is because it is impossible to load an offset for the
printf() function while using the small space in one 16-bit opcode. Therefore, the first 16-bit instruction
loads the higher 10 bits of the offset and the second instruction loads the lower 11 bits of the offset.
As was noted, all instructions in Thumb mode have a size of 2 bytes (or 16 bits). This implies it is impossible
for a Thumb-instruction to be at an odd address whatsoever. Given the above, the last address bit may
be omitted while encoding instructions.
In summary, the BL Thumb-instruction can encode an address in current_P C ± ≈ 2M .
As for the other instructions in the function: PUSH and POP work here just like the described STMFD/LDMFD
only the SP register is not mentioned explicitly here. ADR works just like in the previous example. MOVS
writes 0 into the R0 register in order to return zero.
Optimizing Xcode 4.6.3 (LLVM) (ARM mode)
Xcode 4.6.3 without optimization turned on produces a lot of redundant code so we’ll study optimized
output, where the instruction count is as small as possible, setting the compiler switch -O3.
Listing 1.27: Optimizing Xcode 4.6.3 (LLVM) (ARM mode)

__text:000028C4 _hello_world
__text:000028C4 80 40 2D E9 STMFD SP!, {R7,LR}
__text:000028C8 86 06 01 E3 MOV R0, #0x1686
__text:000028CC 0D 70 A0 E1 MOV R7, SP
__text:000028D0 00 00 40 E3 MOVT R0, #0
__text:000028D4 00 00 8F E0 ADD R0, PC, R0
__text:000028D8 C3 05 00 EB BL _puts
__text:000028DC 00 00 A0 E3 MOV R0, #0
__text:000028E0 80 80 BD E8 LDMFD SP!, {R7,PC}
__cstring:00003F62 48 65 6C 6C+aHelloWorld_0 DCB "Hello world!",0
The instructions STMFD and LDMFD are already familiar to us.

The MOV instruction just writes the number 0x1686 into the R0 register. This is the offset pointing to the
“Hello world!” string.
The R7 register (as it is standardized in [iOS ABI Function Call Guide, (2010)]38 ) is a frame pointer. More
on that below.
The MOVT R0, #0 (MOVe Top) instruction writes 0 into higher 16 bits of the register. The issue here is that
the generic MOV instruction in ARM mode may write only the lower 16 bits of the register.
Keep in mind, all instruction opcodes in ARM mode are limited in size to 32 bits. Of course, this limitation is
not related to moving data between registers. That’s why an additional instruction MOVT exists for writing
into the higher bits (from 16 to 31 inclusive). Its usage here, however, is redundant because the MOV R0,
#0x1686 instruction above cleared the higher part of the register. This is supposedly a shortcoming of the
compiler.
The ADD R0, PC, R0 instruction adds the value in the PC to the value in the R0, to calculate the abso-
lute address of the “Hello world!” string. As we already know, it is “position-independent code” so this
correction is essential here.
The BL instruction calls the puts() function instead of printf().
LLVM has replaced the first printf() call with puts(). Indeed: printf() with a sole argument is almost
analogous to puts().
Almost, because the two functions are producing the same result only in case the string does not contain
printf format identifiers starting with %. In case it does, the effect of these two functions would be different
39
.
40
Why did the compiler replace the printf() with puts()? Presumably because puts() is faster .
Because it just passes characters to stdout without comparing every one of them with the % symbol.
Next, we see the familiar MOV R0, #0 instruction intended to set the R0 register to 0.
38 Also available as http://go.yurichev.com/17276
39 Ithas also to be noted the puts() does not require a ‘\n’ new line symbol at the end of a string, so we do not see it here.
40 ciselant.de/projects/gcc_printf/gcc_printf.html
20
Optimizing Xcode 4.6.3 (LLVM) (Thumb-2 mode)
By default Xcode 4.6.3 generates code for Thumb-2 in this manner:
Listing 1.28: Optimizing Xcode 4.6.3 (LLVM) (Thumb-2 mode)

__text:00002B6C _hello_world
__text:00002B6C 80 B5 PUSH {R7,LR}
__text:00002B6E 41 F2 D8 30 MOVW R0, #0x13D8
__text:00002B72 6F 46 MOV R7, SP
__text:00002B74 C0 F2 00 00 MOVT.W R0, #0
__text:00002B78 78 44 ADD R0, PC
__text:00002B7A 01 F0 38 EA BLX _puts
__text:00002B7E 00 20 MOVS R0, #0
__text:00002B80 80 BD POP {R7,PC}
...
__cstring:00003E70 48 65 6C 6C 6F 20+aHelloWorld DCB "Hello world!",0xA,0
The BL and BLX instructions in Thumb mode, as we recall, are encoded as a pair of 16-bit instructions. In
Thumb-2 these surrogate opcodes are extended in such a way so that new instructions may be encoded
here as 32-bit instructions.
That is obvious considering that the opcodes of the Thumb-2 instructions always begin with 0xFx or 0xEx.
But in the IDA listing the opcode bytes are swapped because for ARM processor the instructions are
encoded as follows: last byte comes first and after that comes the first one (for Thumb and Thumb-2
modes) or for instructions in ARM mode the fourth byte comes first, then the third, then the second and
finally the first (due to different endianness).
So that is how bytes are located in IDA listings:
• for ARM and ARM64 modes: 4-3-2-1;
• for Thumb mode: 2-1;
• for 16-bit instructions pair in Thumb-2 mode: 2-1-4-3.
So as we can see, the MOVW, MOVT.W and BLX instructions begin with 0xFx.
One of the Thumb-2 instructions is MOVW R0, #0x13D8 —it stores a 16-bit value into the lower part of the
R0 register, clearing the higher bits.
Also, MOVT.W R0, #0 works just like MOVT from the previous example only it works in Thumb-2.
Among the other differences, the BLX instruction is used in this case instead of the BL.
The difference is that, besides saving the RA41 in the LR register and passing control to the puts() function,
the processor is also switching from Thumb/Thumb-2 mode to ARM mode (or back).
This instruction is placed here since the instruction to which control is passed looks like (it is encoded in
ARM mode):
__symbolstub1:00003FEC _puts ; CODE XREF: _hello_world+E
__symbolstub1:00003FEC 44 F0 9F E5 LDR PC, =__imp__puts
This is essentially a jump to the place where the address of puts() is written in the imports’ section.
So, the observant reader may ask: why not call puts() right at the point in the code where it is needed?
Because it is not very space-efficient.
Almost any program uses external dynamic libraries (like DLL in Windows, .so in *NIX or .dylib in Mac OS X).
The dynamic libraries contain frequently used library functions, including the standard C-function puts().
In an executable binary file (Windows PE .exe, ELF or Mach-O) an import section is present. This is a list
of symbols (functions or global variables) imported from external modules along with the names of the
modules themselves.
The OS loader loads all modules it needs and, while enumerating import symbols in the primary module,
determines the correct addresses of each symbol.
41 Return Address
21
In our case, __imp__puts is a 32-bit variable used by the OS loader to store the correct address of the
function in an external library. Then the LDR instruction just reads the 32-bit value from this variable and
writes it into the PC register, passing control to it.
So, in order to reduce the time the OS loader needs for completing this procedure, it is good idea to write
the address of each symbol only once, to a dedicated place.
Besides, as we have already figured out, it is impossible to load a 32-bit value into a register while using
only one instruction without a memory access.
Therefore, the optimal solution is to allocate a separate function working in ARM mode with the sole goal of
passing control to the dynamic library and then to jump to this short one-instruction function (the so-called
thunk function) from the Thumb-code.
By the way, in the previous example (compiled for ARM mode) the control is passed by the BL to the same
thunk function. The processor mode, however, is not being switched (hence the absence of an “X” in the
instruction mnemonic).
More about thunk-functions
Thunk-functions are hard to understand, apparently, because of a misnomer. The simplest way to under-
stand it as adaptors or convertors of one type of jack to another. For example, an adaptor allowing the
insertion of a British power plug into an American wall socket, or vice-versa. Thunk functions are also
sometimes called wrappers.
Here are a couple more descriptions of these functions:
“A piece of coding which provides an address:”, according to P. Z. Ingerman, who invented

thunks in 1961 as a way of binding actual parameters to their formal definitions in Algol-60
procedure calls. If a procedure is called with an expression in the place of a formal parameter,
the compiler generates a thunk which computes the expression and leaves the address of
the result in some standard location.
…
Microsoft and IBM have both defined, in their Intel-based systems, a “16-bit environment”
(with bletcherous segment registers and 64K address limits) and a “32-bit environment”
(with flat addressing and semi-real memory management). The two environments can both
be running on the same computer and OS (thanks to what is called, in the Microsoft world,
WOW which stands for Windows On Windows). MS and IBM have both decided that the
process of getting from 16- to 32-bit and vice versa is called a “thunk”; for Windows 95,
there is even a tool, THUNK.EXE, called a “thunk compiler”.
( The Jargon File )

Another example we can find in LAPACK library—a “Linear Algebra PACKage” written in FORTRAN. C/C++
developers also want to use LAPACK, but it’s insane to rewrite it to C/C++ and then maintain several
versions. So there are short C functions callable from C/C++ environment, which are, in turn, call FORTRAN
functions, and do almost anything else:
double Blas_Dot_Prod(const LaVectorDouble &dx, const LaVectorDouble &dy)
{
assert(dx.size()==dy.size());
integer n = dx.size();
integer incx = dx.inc(), incy = dy.inc();
return F77NAME(ddot)(&n, &dx(0), &incx, &dy(0), &incy);

}
Also, functions like that are called “wrappers”.
ARM64
GCC
Let’s compile the example using GCC 4.8.1 in ARM64:
22
Listing 1.29: Non-optimizing GCC 4.8.1 + objdump
1 0000000000400590 <main>:
2 400590: a9bf7bfd stp x29, x30, [sp,#-16]!
3 400594: 910003fd mov x29, sp
4 400598: 90000000 adrp x0, 400000 <_init-0x3b8>
5 40059c: 91192000 add x0, x0, #0x648
6 4005a0: 97ffffa0 bl 400420 <puts@plt>
7 4005a4: 52800000 mov w0, #0x0 // #0
8 4005a8: a8c17bfd ldp x29, x30, [sp],#16
9 4005ac: d65f03c0 ret
10
11 ...
12
13 Contents of section .rodata:
14 400640 01000200 00000000 48656c6c 6f210a00 ........Hello!..
There are no Thumb and Thumb-2 modes in ARM64, only ARM, so there are 32-bit instructions only. The
Register count is doubled: .2.4 on page 1027. 64-bit registers have X- prefixes, while its 32-bit parts—W-.
The STP instruction (Store Pair) saves two registers in the stack simultaneously: X29 and X30.
Of course, this instruction is able to save this pair at an arbitrary place in memory, but the SP register is
specified here, so the pair is saved in the stack.
ARM64 registers are 64-bit ones, each has a size of 8 bytes, so one needs 16 bytes for saving two registers.
The exclamation mark (“!”) after the operand means that 16 is to be subtracted from SP first, and only
then are values from register pair to be written into the stack. This is also called pre-index. About the
difference between post-index and pre-index read here: 1.39.2 on page 437.
Hence, in terms of the more familiar x86, the first instruction is just an analogue to a pair of PUSH X29
and PUSH X30. X29 is used as FP42 in ARM64, and X30 as LR, so that’s why they are saved in the function
prologue and restored in the function epilogue.
The second instruction copies SP in X29 (or FP). This is made so to set up the function stack frame.
ADRP and ADD instructions are used to fill the address of the string “Hello!” into the X0 register, because
the first function argument is passed in this register. There are no instructions, whatsoever, in ARM that
can store a large number into a register (because the instruction length is limited to 4 bytes, read more
about it here: 1.39.3 on page 438). So several instructions must be utilized. The first instruction (ADRP)
writes the address of the 4KiB page, where the string is located, into X0, and the second one (ADD) just
adds the remainder to the address. More about that in: 1.39.4 on page 439.
0x400000 + 0x648 = 0x400648, and we see our “Hello!” C-string in the .rodata data segment at this
address.
puts() is called afterwards using the BL instruction. This was already discussed: 1.5.3 on page 20.
MOV writes 0 into W0. W0 is the lower 32 bits of the 64-bit X0 register:
High 32-bit part low 32-bit part
X0
W0
The function result is returned via X0 and main() returns 0, so that’s how the return result is prepared.
But why use the 32-bit part?
Because the int data type in ARM64, just like in x86-64, is still 32-bit, for better compatibility.
So if a function returns a 32-bit int, only the lower 32 bits of X0 register have to be filled.
In order to verify this, let’s change this example slightly and recompile it. Now main() returns a 64-bit
value:
Listing 1.30: main() returning a value of uint64_t type

#include <stdio.h>
#include <stdint.h>
uint64_t main()
{
printf ("Hello!\n");
42 Frame Pointer
23
return 0;
}
The result is the same, but that’s how MOV at that line looks like now:
Listing 1.31: Non-optimizing GCC 4.8.1 + objdump

4005a4: d2800000 mov x0, #0x0 // #0
LDP (Load Pair) then restores the X29 and X30 registers.
There is no exclamation mark after the instruction: this implies that the values are first loaded from the
stack, and only then is SP increased by 16. This is called post-index.
A new instruction appeared in ARM64: RET. It works just as BX LR, only a special hint bit is added, informing
the CPU that this is a return from a function, not just another jump instruction, so it can execute it more
optimally.
Due to the simplicity of the function, optimizing GCC generates the very same code.
1.5.4 MIPS
A word about the “global pointer”
One important MIPS concept is the “global pointer”. As we may already know, each MIPS instruction has
a size of 32 bits, so it’s impossible to embed a 32-bit address into one instruction: a pair has to be used
for this (like GCC did in our example for the text string address loading). It’s possible, however, to load
data from the address in the range of register − 32768...register + 32767 using one single instruction (because
16 bits of signed offset could be encoded in a single instruction). So we can allocate some register for
this purpose and also allocate a 64KiB area of most used data. This allocated register is called a “global
pointer” and it points to the middle of the 64KiB area. This area usually contains global variables and
addresses of imported functions like printf(), because the GCC developers decided that getting the
address of some function must be as fast as a single instruction execution instead of two. In an ELF file
this 64KiB area is located partly in sections .sbss (“small BSS43 ”) for uninitialized data and .sdata (“small
data”) for initialized data. This implies that the programmer may choose what data he/she wants to
be accessed fast and place it into .sdata/.sbss. Some old-school programmers may recall the MS-DOS
memory model 11.6 on page 986 or the MS-DOS memory managers like XMS/EMS where all memory was
divided in 64KiB blocks.
This concept is not unique to MIPS. At least PowerPC uses this technique as well.
Optimizing GCC
Let’s consider the following example, which illustrates the “global pointer” concept.

1 $LC0:
2 ; \000 is zero byte in octal base:
3 .ascii "Hello, world!\012\000"
4 main:
5 ; function prologue.
6 ; set the GP:
7 lui $28,%hi(__gnu_local_gp)
8 addiu $sp,$sp,-32
9 addiu $28,$28,%lo(__gnu_local_gp)
10 ; save the RA to the local stack:
11 sw $31,28($sp)
12 ; load the address of the puts() function from the GP to $25:
13 lw $25,%call16(puts)($28)
14 ; load the address of the text string to $4 ($a0):
15 lui $4,%hi($LC0)
16 ; jump to puts(), saving the return address in the link register:
17 jalr $25
18 addiu $4,$4,%lo($LC0) ; branch delay slot
19 ; restore the RA:
20 lw $31,28($sp)
21 ; copy 0 from $zero to $v0:
43 Block Started by Symbol
24
22 move $2,$0
23 ; return by jumping to the RA:
24 j $31
25 ; function epilogue:
26 addiu $sp,$sp,32 ; branch delay slot + free local stack
As we see, the $GP register is set in the function prologue to point to the middle of this area. The RA
register is also saved in the local stack. puts() is also used here instead of printf(). The address of the
puts() function is loaded into $25 using LW the instruction (“Load Word”). Then the address of the text
string is loaded to $4 using LUI (“Load Upper Immediate”) and ADDIU (“Add Immediate Unsigned Word”)
instruction pair. LUI sets the high 16 bits of the register (hence “upper” word in instruction name) and
ADDIU adds the lower 16 bits of the address.
ADDIU follows JALR (haven’t you forgot branch delay slots yet?). The register $4 is also called $A0, which
is used for passing the first function argument 44 .
JALR (“Jump and Link Register”) jumps to the address stored in the $25 register (address of puts()) while
saving the address of the next instruction (LW) in RA. This is very similar to ARM. Oh, and one important
thing is that the address saved in RA is not the address of the next instruction (because it’s in a delay slot
and is executed before the jump instruction), but the address of the instruction after the next one (after
the delay slot). Hence, P C + 8 is written to RA during the execution of JALR, in our case, this is the address
of the LW instruction next to ADDIU.
LW (“Load Word”) at line 20 restores RA from the local stack (this instruction is actually part of the function
epilogue).
MOVE at line 22 copies the value from the $0 ($ZERO) register to $2 ($V0).
MIPS has a constant register, which always holds zero. Apparently, the MIPS developers came up with
the idea that zero is in fact the busiest constant in the computer programming, so let’s just use the $0
register every time zero is needed.
Another interesting fact is that MIPS lacks an instruction that transfers data between registers. In fact,
MOVE DST, SRC is ADD DST, SRC, $ZERO (DST = SRC + 0), which does the same. Apparently, the MIPS
developers wanted to have a compact opcode table. This does not mean an actual addition happens at
each MOVE instruction. Most likely, the CPU optimizes these pseudo instructions and the ALU45 is never
used.
J at line 24 jumps to the address in RA, which is effectively performing a return from the function. ADDIU
after J is in fact executed before J (remember branch delay slots?) and is part of the function epilogue.
Here is also a listing generated by IDA. Each register here has its own pseudo name:

1 .text:00000000 main:
2 .text:00000000
3 .text:00000000 var_10 = -0x10
4 .text:00000000 var_4 = -4
5 .text:00000000
7 ; set the GP:
8 .text:00000000 lui $gp, (__gnu_local_gp >> 16)
9 .text:00000004 addiu $sp, -0x20
10 .text:00000008 la $gp, (__gnu_local_gp & 0xFFFF)
11 ; save the RA to the local stack:
12 .text:0000000C sw $ra, 0x20+var_4($sp)
13 ; save the GP to the local stack:
14 ; for some reason, this instruction is missing in the GCC assembly output:
15 .text:00000010 sw $gp, 0x20+var_10($sp)
16 ; load the address of the puts() function from the GP to $t9:
17 .text:00000014 lw $t9, (puts & 0xFFFF)($gp)
18 ; form the address of the text string in $a0:
19 .text:00000018 lui $a0, ($LC0 >> 16) # "Hello, world!"
20 ; jump to puts(), saving the return address in the link register:
21 .text:0000001C jalr $t9
22 .text:00000020 la $a0, ($LC0 & 0xFFFF) # "Hello, world!"
24 .text:00000024 lw $ra, 0x20+var_4($sp)
44 The MIPS registers table is available in appendix .3.1 on page 1028
45 Arithmetic Logic Unit
25
25 ; copy 0 from $zero to $v0:
26 .text:00000028 move $v0, $zero
27 ; return by jumping to the RA:
28 .text:0000002C jr $ra
30 .text:00000030 addiu $sp, 0x20
The instruction at line 15 saves the GP value into the local stack, and this instruction is missing mysteri-
ously from the GCC output listing, maybe by a GCC error 46 . The GP value has to be saved indeed, because
each function can use its own 64KiB data window. The register containing the puts() address is called
$T9, because registers prefixed with T- are called “temporaries” and their contents may not be preserved.
Non-optimizing GCC
Non-optimizing GCC is more verbose.
Listing 1.34: Non-optimizing GCC 4.4.5 (assembly output)

1 $LC0:
2 .ascii "Hello, world!\012\000"
3 main:
5 ; save the RA ($31) and FP in the stack:
6 addiu $sp,$sp,-32
7 sw $31,28($sp)
8 sw $fp,24($sp)
9 ; set the FP (stack frame pointer):
10 move $fp,$sp
11 ; set the GP:
12 lui $28,%hi(__gnu_local_gp)
13 addiu $28,$28,%lo(__gnu_local_gp)
14 ; load the address of the text string:
15 lui $2,%hi($LC0)
16 addiu $4,$2,%lo($LC0)
17 ; load the address of puts() using the GP:
18 lw $2,%call16(puts)($28)
19 nop
20 ; call puts():
21 move $25,$2
22 jalr $25
23 nop ; branch delay slot
24
25 ; restore the GP from the local stack:
26 lw $28,16($fp)
27 ; set register $2 ($V0) to zero:
28 move $2,$0
29 ; function epilogue.
30 ; restore the SP:
31 move $sp,$fp
33 lw $31,28($sp)
34 ; restore the FP:
35 lw $fp,24($sp)
36 addiu $sp,$sp,32
37 ; jump to the RA:
38 j $31
39 nop ; branch delay slot
We see here that register FP is used as a pointer to the stack frame. We also see 3 NOPs. The second
and third of which follow the branch instructions. Perhaps the GCC compiler always adds NOPs (because
of branch delay slots) after branch instructions and then, if optimization is turned on, maybe eliminates
them. So in this case they are left here.
Here is also IDA listing:
Listing 1.35: Non-optimizing GCC 4.4.5 (IDA)

1 .text:00000000 main:
46 Apparently, functions generating listings are not so critical to GCC users, so some unfixed cosmetic bugs may still exist.
26
2 .text:00000000
3 .text:00000000 var_10 = -0x10
4 .text:00000000 var_8 = -8
5 .text:00000000 var_4 = -4
6 .text:00000000
8 ; save the RA and FP in the stack:
9 .text:00000000 addiu $sp, -0x20
10 .text:00000004 sw $ra, 0x20+var_4($sp)
11 .text:00000008 sw $fp, 0x20+var_8($sp)
12 ; set the FP (stack frame pointer):
13 .text:0000000C move $fp, $sp
14 ; set the GP:
15 .text:00000010 la $gp, __gnu_local_gp
16 .text:00000018 sw $gp, 0x20+var_10($sp)
17 ; load the address of the text string:
18 .text:0000001C lui $v0, (aHelloWorld >> 16) # "Hello, world!"
19 .text:00000020 addiu $a0, $v0, (aHelloWorld & 0xFFFF) # "Hello, world!"
20 ; load the address of puts() using the GP:
21 .text:00000024 lw $v0, (puts & 0xFFFF)($gp)
22 .text:00000028 or $at, $zero ; NOP
23 ; call puts():
24 .text:0000002C move $t9, $v0
25 .text:00000030 jalr $t9
27 ; restore the GP from local stack:
28 .text:00000038 lw $gp, 0x20+var_10($fp)
29 ; set register $2 ($V0) to zero:
30 .text:0000003C move $v0, $zero
31 ; function epilogue.
32 ; restore the SP:
33 .text:00000040 move $sp, $fp
35 .text:00000044 lw $ra, 0x20+var_4($sp)
36 ; restore the FP:
37 .text:00000048 lw $fp, 0x20+var_8($sp)
38 .text:0000004C addiu $sp, 0x20
39 ; jump to the RA:
40 .text:00000050 jr $ra
Interestingly, IDA recognized the LUI/ADDIU instructions pair and coalesced them into one LA (“Load Ad-
dress”) pseudo instruction at line 15. We may also see that this pseudo instruction has a size of 8 bytes!
This is a pseudo instruction (or macro) because it’s not a real MIPS instruction, but rather a handy name
for an instruction pair.
Another thing is that IDA doesn’t recognize NOP instructions, so here they are at lines 22, 26 and 41. It is
OR $AT, $ZERO. Essentially, this instruction applies the OR operation to the contents of the $AT register
with zero, which is, of course, an idle instruction. MIPS, like many other ISAs, doesn’t have a separate
NOP instruction.
Role of the stack frame in this example
The address of the text string is passed in the register. Why setup a local stack anyway? The reason
for this lies in the fact that the values of registers RA and GP have to be saved somewhere (because
printf() is called), and the local stack is used for this purpose. If this was a leaf function, it would have
been possible to get rid of the function prologue and epilogue, for example: 1.4.3 on page 8.
Optimizing GCC: load it into GDB
Listing 1.36: sample GDB session

root@debian-mips:~# gcc hw.c -O3 -o hw
root@debian-mips:~# gdb hw
GNU gdb (GDB) 7.0.1-debian
...
Reading symbols from /root/hw...(no debugging symbols found)...done.
27
(gdb) b main
Breakpoint 1 at 0x400654
(gdb) run
Starting program: /root/hw
Breakpoint 1, 0x00400654 in main ()

(gdb) set step-mode on
(gdb) disas
Dump of assembler code for function main:
0x00400640 <main+0>: lui gp,0x42
0x00400644 <main+4>: addiu sp,sp,-32
0x00400648 <main+8>: addiu gp,gp,-30624
0x0040064c <main+12>: sw ra,28(sp)
0x00400650 <main+16>: sw gp,16(sp)
0x00400654 <main+20>: lw t9,-32716(gp)
0x00400658 <main+24>: lui a0,0x40
0x0040065c <main+28>: jalr t9
0x00400660 <main+32>: addiu a0,a0,2080
0x00400664 <main+36>: lw ra,28(sp)
0x00400668 <main+40>: move v0,zero
0x0040066c <main+44>: jr ra
0x00400670 <main+48>: addiu sp,sp,32
End of assembler dump.
(gdb) s
0x00400658 in main ()
(gdb) s
0x0040065c in main ()
(gdb) s
0x2ab2de60 in printf () from /lib/libc.so.6
(gdb) x/s $a0
0x400820: "hello, world"
(gdb)
1.5.5 Conclusion
The main difference between x86/ARM and x64/ARM64 code is that the pointer to the string is now 64-bits
in length. Indeed, modern CPUs are now 64-bit due to both the reduced cost of memory and the greater
demand for it by modern applications. We can add much more memory to our computers than 32-bit
pointers are able to address. As such, all pointers are now 64-bit.
1.5.6 Exercises
• http://challenges.re/48
1.6 Function prologue and epilogue

A function prologue is a sequence of instructions at the start of a function. It often looks something like
the following code fragment:
push ebp
mov ebp, esp
sub esp, X
What these instruction do: save the value of the EBP register on the stack, set the value of the EBP register
to the value of the ESP and then allocate space on the stack for local variables.
The value in the EBP stays the same over the period of the function execution and is to be used for local
variables and arguments access. For the same purpose one can use ESP, but since it changes over time
this approach is not too convenient.
The function epilogue frees the allocated space in the stack, returns the value in the EBP register back to
its initial state and returns the control flow to the caller:
mov esp, ebp
pop ebp
28
ret 0
Function prologues and epilogues are usually detected in disassemblers for function delimitation.
1.6.1 Recursion
Epilogues and prologues can negatively affect the recursion performance.
More about recursion in this book: 3.7.3 on page 481.
1.7 An Empty Function: redux

Let’s back to the empty function example 1.3 on page 5. Now that we know about function prologue and
epilogue, this is an empty function 1.1 on page 5 compiled by non-optimizing GCC:
Listing 1.37: Non-optimizing GCC 8.2 x64 (assembly output)

f:
push rbp
mov rbp, rsp
nop
pop rbp
ret
It’s RET, but function prologue and epilogue, probably, wasn’t optimized and left as is. NOP is seems
another compiler artefact. Anyway, the only effective instruction here is RET. All other instructions can be
removed (or optimized).
1.8 Returning Values: redux

Again, when we know about function prologue and epilogue, let’s recompile an example returning a value
( 1.4 on page 7, 1.8 on page 7) using non-optimizing GCC:

f:
push rbp
mov rbp, rsp
mov eax, 123
pop rbp
ret
Effective instructions here are MOV and RET, others are – prologue and epilogue.
1.9 Stack
47
The stack is one of the most fundamental data structures in computer science . AKA48 LIFO49 .
Technically, it is just a block of memory in process memory along with the ESP or RSP register in x86 or
x64, or the SP register in ARM, as a pointer within that block.
The most frequently used stack access instructions are PUSH and POP (in both x86 and ARM Thumb-mode).
PUSH subtracts from ESP/RSP/SP 4 in 32-bit mode (or 8 in 64-bit mode) and then writes the contents of its
sole operand to the memory address pointed by ESP/RSP/SP.
POP is the reverse operation: retrieve the data from the memory location that SP points to, load it into the
instruction operand (often a register) and then add 4 (or 8) to the stack pointer.
After stack allocation, the stack pointer points at the bottom of the stack. PUSH decreases the stack pointer
and POP increases it. The bottom of the stack is actually at the beginning of the memory allocated for the
stack block. It seems strange, but that’s the way it is.
47 wikipedia.org/wiki/Call_stack
48 Also Known As
49 Last In First Out
29
ARM supports both descending and ascending stacks.
For example the STMFD/LDMFD, STMED50 /LDMED51 instructions are intended to deal with a descending
stack (grows downwards, starting with a high address and progressing to a lower one). The STMFA52 /LDMFA53 ,
STMEA54 /LDMEA55 instructions are intended to deal with an ascending stack (grows upwards, starting from
a low address and progressing to a higher one).
1.9.1 Why does the stack grow backwards?

Intuitively, we might think that the stack grows upwards, i.e. towards higher addresses, like any other
data structure.
The reason that the stack grows backward is probably historical. When the computers were big and
occupied a whole room, it was easy to divide memory into two parts, one for the heap and one for the
stack. Of course, it was unknown how big the heap and the stack would be during program execution, so
this solution was the simplest possible.
Start of heap Start of stack
Heap Stack
In [D. M. Ritchie and K. Thompson, The UNIX Time Sharing System, (1974)]56 we can read:
The user-core part of an image is divided into three logical segments. The program text
segment begins at location 0 in the virtual address space. During execution, this segment
is write-protected and a single copy of it is shared among all processes executing the same
program. At the first 8K byte boundary above the program text segment in the virtual ad-
dress space begins a nonshared, writable data segment, the size of which may be extended
by a system call. Starting at the highest address in the virtual address space is a stack
segment, which automatically grows downward as the hardware’s stack pointer fluctuates.
This reminds us how some students write two lecture notes using only one notebook: notes for the first
lecture are written as usual, and notes for the second one are written from the end of notebook, by flipping
it. Notes may meet each other somewhere in between, in case of lack of free space.
1.9.2 What is the stack used for?

Save the function’s return address
x86
When calling another function with a CALL instruction, the address of the point exactly after the CALL
instruction is saved to the stack and then an unconditional jump to the address in the CALL operand is
executed.
The CALL instruction is equivalent to a
PUSH address_after_call / JMP operand instruction pair.
RET fetches a value from the stack and jumps to it —that is equivalent to a POP tmp / JMP tmp instruction
pair.
Overflowing the stack is straightforward. Just run eternal recursion:
void f()
{
f();
50 Store Multiple Empty Descending (ARM instruction)
51 Load Multiple Empty Descending (ARM instruction)
52 Store Multiple Full Ascending (ARM instruction)
53 Load Multiple Full Ascending (ARM instruction)
54 Store Multiple Empty Ascending (ARM instruction)
55 Load Multiple Empty Ascending (ARM instruction)
30
};
MSVC 2008 reports the problem:

c:\tmp6>cl ss.cpp /Fass.asm
Microsoft (R) 32-bit C/C++ Optimizing Compiler Version 15.00.21022.08 for 80x86
Copyright (C) Microsoft Corporation. All rights reserved.
ss.cpp
c:\tmp6\ss.cpp(4) : warning C4717: 'f' : recursive on all control paths, function will cause ⤦
Ç runtime stack overflow
…but generates the right code anyway:

?f@@YAXXZ PROC ; f
; Line 2
push ebp
mov ebp, esp
; Line 3
call ?f@@YAXXZ ; f
; Line 4
pop ebp
ret 0
?f@@YAXXZ ENDP ; f
…Also if we turn on the compiler optimization (/Ox option) the optimized code will not overflow the stack
and will work correctly57 instead:
?f@@YAXXZ PROC ; f
; Line 2
$LL3@f:
; Line 3
jmp SHORT $LL3@f
?f@@YAXXZ ENDP ; f
GCC 4.4.1 generates similar code in both cases without, however, issuing any warning about the problem.
ARM
ARM programs also use the stack for saving return addresses, but differently. As mentioned in “Hello,
world!” ( 1.5.3 on page 18), the RA is saved to the LR (link register). If one needs, however, to call another
function and use the LR register one more time, its value has to be saved. Usually it is saved in the
function prologue.
Often, we see instructions like PUSH R4-R7,LR along with this instruction in epilogue POP R4-R7,PC—thus
register values to be used in the function are saved in the stack, including LR.
Nevertheless, if a function never calls any other function, in RISC terminology it is called a leaf function58 .
As a consequence, leaf functions do not save the LR register (because they don’t modify it). If such
function is small and uses a small number of registers, it may not use the stack at all. Thus, it is possible
to call leaf functions without using the stack, which can be faster than on older x86 machines because
external RAM is not used for the stack 59 . This can be also useful for situations when memory for the stack
is not yet allocated or not available.
Some examples of leaf functions: 1.14.3 on page 103, 1.14.3 on page 103, 1.282 on page 313, 1.298 on
page 330, 1.28.5 on page 331, 1.192 on page 209, 1.190 on page 207, 1.209 on page 225.
Passing function arguments
The most popular way to pass parameters in x86 is called “cdecl”:

57 irony here
58 infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.faqs/ka13785.html
59 Some time ago, on PDP-11 and VAX, the CALL instruction (calling other functions) was expensive; up to 50% of execution time
might be spent on it, so it was considered that having a big number of small functions is an anti-pattern [Eric S. Raymond, The Art
of UNIX Programming, (2003)Chapter 4, Part II].
31
push arg3
push arg2
push arg1
call f
add esp, 12 ; 4*3=12
Callee functions get their arguments via the stack pointer.

Therefore, this is how the argument values are located in the stack before the execution of the f() func-
tion’s very first instruction:
ESP return address
ESP+4 argument#1, marked in IDA as arg_0
ESP+0xC argument#3, marked in IDA as arg_8
… …
For more information on other calling conventions see also section ( 6.1 on page 725).
By the way, the callee function does not have any information about how many arguments were passed.
C functions with a variable number of arguments (like printf()) determine their number using format
string specifiers (which begin with the % symbol).
If we write something like:
printf("%d %d %d", 1234);
printf() will print 1234, and then two random numbers60 , which were lying next to it in the stack.
That’s why it is not very important how we declare the main() function: as main(),
main(int argc, char *argv[]) or main(int argc, char *argv[], char *envp[]).
In fact, the CRT-code is calling main() roughly as:
push envp
push argv
push argc
call main
...
If you declare main() as main() without arguments, they are, nevertheless, still present in the stack, but
are not used. If you declare main() as main(int argc, char *argv[]), you will be able to use first two
arguments, and the third will remain “invisible” for your function. Even more, it is possible to declare
main(int argc), and it will work.
Another related example: 6.1.10.
Alternative ways of passing arguments
It is worth noting that nothing obliges programmers to pass arguments through the stack. It is not a
requirement. One could implement any other method without using the stack at all.
A somewhat popular way among assembly language newbies is to pass arguments via global variables,
like:
Listing 1.39: Assembly code

...
mov X, 123
mov Y, 456
call do_something
...
X dd ?
Y dd ?
60 Not random in strict sense, but rather unpredictable: 1.9.4 on page 36
32
do_something proc near
; take X
; take Y
; do something
retn
do_something endp
But this method has obvious drawback: do_something() function cannot call itself recursively (or via
another function), because it has to zap its own arguments. The same story with local variables: if you
hold them in global variables, the function couldn’t call itself. And this is also not thread-safe 61 . A
method to store such information in stack makes this easier—it can hold as many function arguments
and/or values, as much space it has.
[Donald E. Knuth, The Art of Computer Programming, Volume 1, 3rd ed., (1997), 189] mentions even
weirder schemes particularly convenient on IBM System/360.
MS-DOS had a way of passing all function arguments via registers, for example, this is piece of code for
ancient 16-bit MS-DOS prints “Hello, world!”:
mov dx, msg ; address of message
mov ah, 9 ; 9 means "print string" function
int 21h ; DOS "syscall"
mov ah, 4ch ; "terminate program" function

int 21h ; DOS "syscall"
msg db 'Hello, World!\$'
This is quite similar to 6.1.3 on page 726 method. And also it’s very similar to calling syscalls in Linux
( 6.3.1 on page 739) and Windows.
If a MS-DOS function is going to return a boolean value (i.e., single bit, usually indicating error state), CF
flag was often used.
For example:
mov ah, 3ch ; create file
lea dx, filename
mov cl, 1
int 21h
jc error
mov file_handle, ax
...
error:
...
In case of error, CF flag is raised. Otherwise, handle of newly created file is returned via AX.
This method is still used by assembly language programmers. In Windows Research Kernel source code
(which is quite similar to Windows 2003) we can find something like this (file base/ntos/ke/i386/cpu.asm):
public Get386Stepping
Get386Stepping proc
call MultiplyTest ; Perform multiplication test

jnc short G3s00 ; if nc, muttest is ok
mov ax, 0
ret
G3s00:
call Check386B0 ; Check for B0 stepping
jnc short G3s05 ; if nc, it's B1/later
mov ax, 100h ; It is B0/earlier stepping
ret
G3s05:
call Check386D1 ; Check for D1 stepping
jc short G3s10 ; if c, it is NOT D1
mov ax, 301h ; It is D1/later stepping
ret
61 Correctly implemented, each thread would have its own stack with its own arguments/variables.
33
G3s10:
mov ax, 101h ; assume it is B1 stepping
ret
...
MultiplyTest proc
xor cx,cx ; 64K times is a nice round number

mlt00: push cx
call Multiply ; does this chip's multiply work?
pop cx
jc short mltx ; if c, No, exit
loop mlt00 ; if nc, YEs, loop to try again
clc
mltx:
ret
MultiplyTest endp
Local variable storage
A function could allocate space in the stack for its local variables just by decreasing the stack pointer
towards the stack bottom.
Hence, it’s very fast, no matter how many local variables are defined. It is also not a requirement to store
local variables in the stack. You could store local variables wherever you like, but traditionally this is how
it’s done.
x86: alloca() function
It is worth noting the alloca() function 62 . This function works like malloc(), but allocates memory
directly on the stack. The allocated memory chunk does not have to be freed via a free() function call,
since the function epilogue ( 1.6 on page 28) returns ESP back to its initial state and the allocated memory
is just dropped. It is worth noting how alloca() is implemented. In simple terms, this function just shifts
ESP downwards toward the stack bottom by the number of bytes you need and sets ESP as a pointer to
the allocated block.
Let’s try:
#ifdef __GNUC__
#include <alloca.h> // GCC
#else
#include <malloc.h> // MSVC
#endif
#include <stdio.h>
void f()
{
char *buf=(char*)alloca (600);
#ifdef __GNUC__
snprintf (buf, 600, "hi! %d, %d, %d\n", 1, 2, 3); // GCC
#else
_snprintf (buf, 600, "hi! %d, %d, %d\n", 1, 2, 3); // MSVC
#endif
puts (buf);
};
_snprintf() function works just like printf(), but instead of dumping the result into stdout (e.g., to
terminal or console), it writes it to the buf buffer. Function puts() copies the contents of buf to stdout.
Of course, these two function calls might be replaced by one printf() call, but we have to illustrate small
buffer usage.
62 In MSVC, the function implementation can be found in alloca16.asm and chkstk.asm in
C:\Program Files (x86)\Microsoft Visual Studio 10.0\VC\crt\src\intel
34
MSVC
Let’s compile (MSVC 2010):

...
mov eax, 600 ; 00000258H

call __alloca_probe_16
mov esi, esp
push 3
push 2
push 1
push OFFSET $SG2672
push 600 ; 00000258H
push esi
call __snprintf
push esi
call _puts
add esp, 28
...
63
The sole alloca() argument is passed via EAX (instead of pushing it into the stack) .
GCC + Intel syntax
GCC 4.4.1 does the same without calling external functions:

.LC0:
.string "hi! %d, %d, %d\n"
f:
push ebp
mov ebp, esp
push ebx
sub esp, 660
lea ebx, [esp+39]
and ebx, -16 ; align pointer by 16-bit border
mov DWORD PTR [esp], ebx ; s
mov DWORD PTR [esp+20], 3
mov DWORD PTR [esp+8], OFFSET FLAT:.LC0 ; "hi! %d, %d, %d\n"
mov DWORD PTR [esp+4], 600 ; maxlen
call _snprintf
mov DWORD PTR [esp], ebx ; s
call puts
mov ebx, DWORD PTR [ebp-4]
leave
ret
GCC + AT&T syntax
Let’s see the same code, but in AT&T syntax:

.LC0:
63 It is because alloca() is rather a compiler intrinsic ( 11.3 on page 981) than a normal function. One of the reasons we need a
separate function instead of just a couple of instructions in the code, is because the MSVC64 alloca() implementation also has code
which reads from the memory just allocated, in order to let the OS map physical memory to this VM65 region. After the alloca()
call, ESP points to the block of 600 bytes and we can use it as memory for the buf array.
35
f:
pushl %ebp
movl %esp, %ebp
pushl %ebx
subl $660, %esp
leal 39(%esp), %ebx
andl $-16, %ebx
movl %ebx, (%esp)
movl $3, 20(%esp)
movl $2, 16(%esp)
movl $1, 12(%esp)
movl $.LC0, 8(%esp)
movl $600, 4(%esp)
call _snprintf
movl %ebx, (%esp)
call puts
movl -4(%ebp), %ebx
leave
ret
The code is the same as in the previous listing.

By the way, movl $3, 20(%esp) corresponds to mov DWORD PTR [esp+20], 3 in Intel-syntax. In the
AT&T syntax, the register+offset format of addressing memory looks like offset(%register).
(Windows) SEH
SEH66 records are also stored on the stack (if they are present). Read more about it: ( 6.5.3 on page 755).
Buffer overflow protection
More about it here ( 1.26.2 on page 271).
Automatic deallocation of data in stack
Perhaps the reason for storing local variables and SEH records in the stack is that they are freed automat-
ically upon function exit, using just one instruction to correct the stack pointer (it is often ADD). Function
arguments, as we could say, are also deallocated automatically at the end of function. In contrast, every-
thing stored in the heap must be deallocated explicitly.
1.9.3 A typical stack layout

A typical stack layout in a 32-bit environment at the start of a function, before the first instruction execution
looks like this:
… …
ESP-0xC local variable#2, marked in IDA as var_8
ESP-8 local variable#1, marked in IDA as var_4
ESP-4 saved value ofEBP
ESP Return Address
ESP+0xC argument#3, marked in IDA as arg_8
… …
1.9.4 Noise in stack

When one says that something seems
random, what one usually means in practice
is that one cannot see any regularities in it.
Stephen Wolfram, A New Kind of Science.

66 Structured Exception Handling
36
Often in this book “noise” or “garbage” values in the stack or memory are mentioned. Where do they
come from? These are what has been left there after other functions’ executions. Short example:
#include <stdio.h>
void f1()
{
int a=1, b=2, c=3;
};
void f2()
{
int a, b, c;
printf ("%d, %d, %d\n", a, b, c);
};
int main()
{
f1();
f2();
};
Compiling …
Listing 1.43: Non-optimizing MSVC 2010

$SG2752 DB '%d, %d, %d', 0aH, 00H
_c$ = -12 ; size = 4

_b$ = -8 ; size = 4
_a$ = -4 ; size = 4
_f1 PROC
push ebp
mov ebp, esp
sub esp, 12
mov DWORD PTR _a$[ebp], 1
mov DWORD PTR _b$[ebp], 2
mov DWORD PTR _c$[ebp], 3
mov esp, ebp
pop ebp
ret 0
_f1 ENDP
_c$ = -12 ; size = 4

_b$ = -8 ; size = 4
_a$ = -4 ; size = 4
_f2 PROC
push ebp
mov ebp, esp
sub esp, 12
mov eax, DWORD PTR _c$[ebp]
push eax
mov ecx, DWORD PTR _b$[ebp]
push ecx
mov edx, DWORD PTR _a$[ebp]
push edx
push OFFSET $SG2752 ; '%d, %d, %d'
call DWORD PTR __imp__printf
add esp, 16
mov esp, ebp
pop ebp
ret 0
_f2 ENDP
_main PROC
push ebp
mov ebp, esp
call _f1
call _f2
xor eax, eax
pop ebp
37
ret 0
_main ENDP
The compiler will grumble a little bit…

c:\Polygon\c>cl st.c /Fast.asm /MD
Microsoft (R) 32-bit C/C++ Optimizing Compiler Version 16.00.40219.01 for 80x86
st.c
c:\polygon\c\st.c(11) : warning C4700: uninitialized local variable 'c' used
c:\polygon\c\st.c(11) : warning C4700: uninitialized local variable 'b' used
c:\polygon\c\st.c(11) : warning C4700: uninitialized local variable 'a' used
Microsoft (R) Incremental Linker Version 10.00.40219.01
/out:st.exe
st.obj
But when we run the compiled program …

c:\Polygon\c>st
1, 2, 3
Oh, what a weird thing! We did not set any variables in f2(). These are “ghosts” values, which are still
in the stack.
38
Let’s load the example into OllyDbg:
Figure 1.6: OllyDbg: f1()
When f1() assigns the variables a, b and c, their values are stored at the address 0x1FF860 and so on.
39
And when f2() executes:
Figure 1.7: OllyDbg: f2()
... a, b and c of f2() are located at the same addresses! No one has overwritten the values yet, so at that
point they are still untouched. So, for this weird situation to occur, several functions have to be called
one after another and SP has to be the same at each function entry (i.e., they have the same number of
arguments). Then the local variables will be located at the same positions in the stack. Summarizing, all
values in the stack (and memory cells in general) have values left there from previous function executions.
They are not random in the strict sense, but rather have unpredictable values. Is there another option?
It would probably be possible to clear portions of the stack before each function execution, but that’s too
much extra (and unnecessary) work.
MSVC 2013
The example was compiled by MSVC 2010. But the reader of this book made attempt to compile this
example in MSVC 2013, ran it, and got all 3 numbers reversed:
c:\Polygon\c>st
3, 2, 1
Why? I also compiled this example in MSVC 2013 and saw this:

_a$ = -12 ; size = 4
_b$ = -8 ; size = 4
_c$ = -4 ; size = 4
_f2 PROC
...
_f2 ENDP
_c$ = -12 ; size = 4

_b$ = -8 ; size = 4
_a$ = -4 ; size = 4
_f1 PROC
40
...
_f1 ENDP
Unlike MSVC 2010, MSVC 2013 allocated a/b/c variables in function f2() in reverse order.And this is
completely correct, because C/C++ standards has no rule, in which order local variables must be allocated
in the local stack, if at all. The reason of difference is because MSVC 2010 has one way to do it, and MSVC
2013 has supposedly something changed inside of compiler guts, so it behaves slightly different.
1.9.5 Exercises
1.10 Almost empty function

This is a real piece of code I found in Boolector67 :
// forward declaration. the function is residing in some other module:
int boolector_main (int argc, char **argv);
// executable
int main (int argc, char **argv)
{
return boolector_main (argc, argv);
}
Why would anyone do so? I don’t know, but my best guess is that boolector_main() may be compiled
in some kind of DLL or dynamic library, and be called from a test suite. Surely, a test suite can prepare
argc/argv variables as CRT would do it.
Interestingly enough, how this compiles:

main:
push rbp
mov rbp, rsp
sub rsp, 16
mov DWORD PTR -4[rbp], edi
mov QWORD PTR -16[rbp], rsi
mov rdx, QWORD PTR -16[rbp]
mov eax, DWORD PTR -4[rbp]
mov rsi, rdx
mov edi, eax
call boolector_main
leave
ret
This is OK, prologue, unnecessary (not optimized) shuffling of two arguments, CALL, epilogue, RET. But
let’s see optimizing version:
Listing 1.46: Optimizing GCC 8.2 x64 (assembly output)

main:
jmp boolector_main
As simple as that: stack/registers are untouched and boolector_main() has the same arguments set. So
all we need to do is pass execution to another address.
This is close to thunk function.
We will see something more advanced later: 1.11.2 on page 53, 1.21.1 on page 154.
67 https://boolector.github.io/
41
1.11 printf() with several arguments
Now let’s extend the Hello, world! ( 1.5 on page 8) example, replacing printf() in the main() function
body with this:
#include <stdio.h>
int main()
{
printf("a=%d; b=%d; c=%d", 1, 2, 3);
return 0;
};
1.11.1 x86
x86: 3 arguments
MSVC
When we compile it with MSVC 2010 Express we get:

$SG3830 DB 'a=%d; b=%d; c=%d', 00H
...
push 3
push 2
push 1
push OFFSET $SG3830
call _printf
add esp, 16 ; 00000010H
Almost the same, but now we can see the printf() arguments are pushed onto the stack in reverse order.
The first argument is pushed last.
By the way, variables of int type in 32-bit environment have 32-bit width, that is 4 bytes.
So, we have 4 arguments here. 4 ∗ 4 = 16 —they occupy exactly 16 bytes in the stack: a 32-bit pointer to
a string and 3 numbers of type int.
When the stack pointer (ESP register) has changed back by the
ADD ESP, X instruction after a function call, often, the number of function arguments could be deduced
by simply dividing X by 4.
Of course, this is specific to the cdecl calling convention, and only for 32-bit environment.
See also the calling conventions section ( 6.1 on page 725).
In certain cases where several functions return right after one another, the compiler could merge multiple
“ADD ESP, X” instructions into one, after the last call:
push a1
push a2
call ...
...
push a1
call ...
...
push a1
push a2
push a3
call ...
add esp, 24
Here is a real-world example:
Listing 1.47: x86

.text:100113E7 push 3
.text:100113E9 call sub_100018B0 ; takes one argument (3)
42
.text:100113EE call sub_100019D0 ; takes no arguments at all
.text:100113F3 call sub_10006A90 ; takes no arguments at all
.text:100113F8 push 1
.text:100113FA call sub_100018B0 ; takes one argument (1)
.text:100113FF add esp, 8 ; drops two arguments from stack at once
43
MSVC and OllyDbg
Now let’s try to load this example in OllyDbg. It is one of the most popular user-land win32 debuggers.
We can compile our example in MSVC 2012 with /MD option, which means to link with MSVCR*.DLL, so we
can see the imported functions clearly in the debugger.
Then load the executable in OllyDbg. The very first breakpoint is in ntdll.dll, press F9 (run). The second
breakpoint is in CRT-code. Now we have to find the main() function.
Find this code by scrolling the code to the very top (MSVC allocates the main() function at the very
beginning of the code section):
Figure 1.8: OllyDbg: the very start of the main() function
Click on the PUSH EBP instruction, press F2 (set breakpoint) and press F9 (run). We have to perform these
actions in order to skip CRT-code, because we aren’t really interested in it yet.
44
Press F8 (step over) 6 times, i.e. skip 6 instructions:
Figure 1.9: OllyDbg: before printf() execution
Now the PC points to the CALL printf instruction. OllyDbg, like other debuggers, highlights the value of
the registers which were changed. So each time you press F8, EIP changes and its value is displayed in
red. ESP changes as well, because the arguments values are pushed into the stack.
Where are the values in the stack? Take a look at the right bottom debugger window:
Figure 1.10: OllyDbg: stack after the argument values have been pushed (The red rectangular border was
added by the author in a graphics editor)
We can see 3 columns there: address in the stack, value in the stack and some additional OllyDbg com-
ments. OllyDbg understands printf()-like strings, so it reports the string here and the 3 values attached
to it.
It is possible to right-click on the format string, click on “Follow in dump”, and the format string will appear
in the debugger left-bottom window, which always displays some part of the memory. These memory
values can be edited. It is possible to change the format string, in which case the result of our example
would be different. It is not very useful in this particular case, but it could be good as an exercise so you
start building a feel of how everything works here.
45
Press F8 (step over).
We see the following output in the console:
a=1; b=2; c=3
Let’s see how the registers and stack state have changed:
Figure 1.11: OllyDbg after printf() execution
Register EAX now contains 0xD (13). That is correct, since printf() returns the number of characters
printed. The value of EIP has changed: indeed, now it contains the address of the instruction coming
after CALL printf. ECX and EDX values have changed as well. Apparently, the printf() function’s hidden
machinery used them for its own needs.
A very important fact is that neither the ESP value, nor the stack state have been changed! We clearly
see that the format string and corresponding 3 values are still there. This is indeed the cdecl calling
convention behavior: callee does not return ESP back to its previous value. The caller is responsible to do
so.
46
Press F8 again to execute ADD ESP, 10 instruction:
Figure 1.12: OllyDbg: after ADD ESP, 10 instruction execution
ESP has changed, but the values are still in the stack! Yes, of course; no one needs to set these values
to zeros or something like that. Everything above the stack pointer (SP) is noise or garbage and has no
meaning at all. It would be time consuming to clear the unused stack entries anyway, and no one really
needs to.
GCC
Now let’s compile the same program in Linux using GCC 4.4.1 and take a look at what we have got in IDA:
main proc near

var_C = dword ptr -0Ch
var_8 = dword ptr -8
push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
sub esp, 10h
mov eax, offset aADBDCD ; "a=%d; b=%d; c=%d"
mov [esp+10h+var_4], 3
mov [esp+10h+var_C], 1
call _printf
mov eax, 0
leave
retn
main endp
Its noticeable that the difference between the MSVC code and the GCC code is only in the way the ar-
guments are stored on the stack. Here the GCC is working directly with the stack without the use of
PUSH/POP.
GCC and GDB
47
Let’s try this example also in GDB68 in Linux.
-g option instructs the compiler to include debug information in the executable file.
$ gcc 1.c -g -o 1
$ gdb 1
GNU gdb (GDB) 7.6.1-ubuntu
...
Reading symbols from /home/dennis/polygon/1...done.
Listing 1.48: let’s set breakpoint on printf()

(gdb) b printf
Breakpoint 1 at 0x80482f0
Run. We don’t have the printf() function source code here, so GDB can’t show it, but may do so.
(gdb) run
Starting program: /home/dennis/polygon/1
Breakpoint 1, __printf (format=0x80484f0 "a=%d; b=%d; c=%d") at printf.c:29

29 printf.c: No such file or directory.
Print 10 stack elements. The most left column contains addresses on the stack.
(gdb) x/10w $esp
0xbffff11c: 0x0804844a 0x080484f0 0x00000001 0x00000002
0xbffff12c: 0x00000003 0x08048460 0x00000000 0x00000000
0xbffff13c: 0xb7e29905 0x00000001
The very first element is the RA (0x0804844a). We can verify this by disassembling the memory at this
address:
(gdb) x/5i 0x0804844a
0x804844a <main+45>: mov $0x0,%eax
0x804844f <main+50>: leave
0x8048450 <main+51>: ret
0x8048451: xchg %ax,%ax
0x8048453: xchg %ax,%ax
The two XCHG instructions are idle instructions, analogous to NOPs.

The second element (0x080484f0) is the format string address:
(gdb) x/s 0x080484f0
0x80484f0: "a=%d; b=%d; c=%d"
Next 3 elements (1, 2, 3) are the printf() arguments. The rest of the elements could be just “garbage”
on the stack, but could also be values from other functions, their local variables, etc. We can ignore them
for now.
Run “finish”. The command instructs GDB to “execute all instructions until the end of the function”. In
this case: execute till the end of printf().
(gdb) finish
Run till exit from #0 __printf (format=0x80484f0 "a=%d; b=%d; c=%d") at printf.c:29
main () at 1.c:6
6 return 0;
Value returned is $2 = 13
GDB shows what printf() returned in EAX (13). This is the number of characters printed out, just like in
the OllyDbg example.
We also see “return 0;” and the information that this expression is in the 1.c file at the line 6. Indeed,
the 1.c file is located in the current directory, and GDB finds the string there. How does GDB know
which C-code line is being currently executed? This is due to the fact that the compiler, while generating
68 GNU Debugger
48
debugging information, also saves a table of relations between source code line numbers and instruction
addresses. GDB is a source-level debugger, after all.
Let’s examine the registers. 13 in EAX:
(gdb) info registers
eax 0xd 13
ecx 0x0 0
edx 0x0 0
ebx 0xb7fc0000 -1208221696
esp 0xbffff120 0xbffff120
ebp 0xbffff138 0xbffff138
esi 0x0 0
edi 0x0 0
eip 0x804844a 0x804844a <main+45>
...
Let’s disassemble the current instructions. The arrow points to the instruction to be executed next.
(gdb) disas
0x0804841d <+0>: push %ebp
0x0804841e <+1>: mov %esp,%ebp
0x08048420 <+3>: and $0xfffffff0,%esp
0x08048423 <+6>: sub $0x10,%esp
0x08048426 <+9>: movl $0x3,0xc(%esp)
0x0804842e <+17>: movl $0x2,0x8(%esp)
0x08048436 <+25>: movl $0x1,0x4(%esp)
0x0804843e <+33>: movl $0x80484f0,(%esp)
0x08048445 <+40>: call 0x80482f0 <printf@plt>
=> 0x0804844a <+45>: mov $0x0,%eax
0x0804844f <+50>: leave
0x08048450 <+51>: ret
GDB uses AT&T syntax by default. But it is possible to switch to Intel syntax:
(gdb) set disassembly-flavor intel
(gdb) disas
0x0804841d <+0>: push ebp
0x0804841e <+1>: mov ebp,esp
0x08048420 <+3>: and esp,0xfffffff0
0x08048423 <+6>: sub esp,0x10
0x08048426 <+9>: mov DWORD PTR [esp+0xc],0x3
0x0804842e <+17>: mov DWORD PTR [esp+0x8],0x2
0x08048436 <+25>: mov DWORD PTR [esp+0x4],0x1
0x0804843e <+33>: mov DWORD PTR [esp],0x80484f0
0x08048445 <+40>: call 0x80482f0 <printf@plt>
=> 0x0804844a <+45>: mov eax,0x0
0x0804844f <+50>: leave
0x08048450 <+51>: ret
Execute next instruction. GDB shows ending bracket, meaning, it ends the block.
(gdb) step
7 };
Let’s examine the registers after the MOV EAX, 0 instruction execution. Indeed EAX is zero at that point.
eax 0x0 0
ecx 0x0 0
edx 0x0 0
ebx 0xb7fc0000 -1208221696
esp 0xbffff120 0xbffff120
ebp 0xbffff138 0xbffff138
esi 0x0 0
edi 0x0 0
49
eip 0x804844f 0x804844f <main+50>
...
x64: 8 arguments
To see how other arguments are passed via the stack, let’s change our example again by increasing the
number of arguments to 9 (printf() format string + 8 int variables):
#include <stdio.h>
int main()
{
printf("a=%d; b=%d; c=%d; d=%d; e=%d; f=%d; g=%d; h=%d\n", 1, 2, 3, 4, 5, 6, 7, 8);
return 0;
};
MSVC
As it was mentioned earlier, the first 4 arguments has to be passed through the RCX, RDX, R8, R9 registers
in Win64, while all the rest—via the stack. That is exactly what we see here. However, the MOV instruction,
instead of PUSH, is used for preparing the stack, so the values are stored to the stack in a straightforward
manner.

$SG2923 DB 'a=%d; b=%d; c=%d; d=%d; e=%d; f=%d; g=%d; h=%d', 0aH, 00H
main PROC
sub rsp, 88
mov DWORD PTR [rsp+64], 8

mov r9d, 3
mov r8d, 2
mov edx, 1
call printf
; return 0
xor eax, eax
add rsp, 88
ret 0
main ENDP
_TEXT ENDS
END
The observant reader may ask why are 8 bytes allocated for int values, when 4 is enough? Yes, one
has to recall: 8 bytes are allocated for any data type shorter than 64 bits. This is established for the
convenience’s sake: it makes it easy to calculate the address of arbitrary argument. Besides, they are all
located at aligned memory addresses. It is the same in the 32-bit environments: 4 bytes are reserved for
all data types.
GCC
The picture is similar for x86-64 *NIX OS-es, except that the first 6 arguments are passed through the
RDI, RSI, RDX, RCX, R8, R9 registers. All the rest—via the stack. GCC generates the code storing the string
pointer into EDI instead of RDI—we noted that previously: 1.5.2 on page 15.
We also noted earlier that the EAX register has been cleared before a printf() call: 1.5.2 on page 15.
50
Listing 1.50: Optimizing GCC 4.4.6 x64
.LC0:
.string "a=%d; b=%d; c=%d; d=%d; e=%d; f=%d; g=%d; h=%d\n"
main:
sub rsp, 40
mov r9d, 5
mov r8d, 4
mov ecx, 3
mov edx, 2
mov esi, 1
mov edi, OFFSET FLAT:.LC0
mov DWORD PTR [rsp], 6
call printf
; return 0
xor eax, eax

add rsp, 40
ret
GCC + GDB
Let’s try this example in GDB.

$ gcc -g 2.c -o 2
$ gdb 2
...
Reading symbols from /home/dennis/polygon/2...done.
Listing 1.51: let’s set the breakpoint to printf(), and run

(gdb) b printf
(gdb) run
Starting program: /home/dennis/polygon/2
Breakpoint 1, __printf (format=0x400628 "a=%d; b=%d; c=%d; d=%d; e=%d; f=%d; g=%d; h=%d\n") at ⤦
Ç printf.c:29
Registers RSI/RDX/RCX/R8/R9 have the expected values. RIP has the address of the very first instruction
of the printf() function.
rax 0x0 0
rbx 0x0 0
rcx 0x3 3
rdx 0x2 2
rsi 0x1 1
rdi 0x400628 4195880
rbp 0x7fffffffdf60 0x7fffffffdf60
rsp 0x7fffffffdf38 0x7fffffffdf38
r8 0x4 4
r9 0x5 5
r10 0x7fffffffdce0 140737488346336
r11 0x7ffff7a65f60 140737348263776
r12 0x400440 4195392
r13 0x7fffffffe040 140737488347200
r14 0x0 0
r15 0x0 0
51
rip 0x7ffff7a65f60 0x7ffff7a65f60 <__printf>
...
Listing 1.52: let’s inspect the format string

(gdb) x/s $rdi
0x400628: "a=%d; b=%d; c=%d; d=%d; e=%d; f=%d; g=%d; h=%d\n"
Let’s dump the stack with the x/g command this time—g stands for giant words, i.e., 64-bit words.
(gdb) x/10g $rsp
0x7fffffffdf38: 0x0000000000400576 0x0000000000000006
0x7fffffffdf48: 0x0000000000000007 0x00007fff00000008
0x7fffffffdf58: 0x0000000000000000 0x0000000000000000
0x7fffffffdf68: 0x00007ffff7a33de5 0x0000000000000000
0x7fffffffdf78: 0x00007fffffffe048 0x0000000100000000
The very first stack element, just like in the previous case, is the RA. 3 values are also passed through
the stack: 6, 7, 8. We also see that 8 is passed with the high 32-bits not cleared: 0x00007fff00000008.
That’s OK, because the values are of int type, which is 32-bit. So, the high register or stack element part
may contain “random garbage”.
If you take a look at where the control will return after the printf() execution, GDB will show the entire
main() function:
(gdb) disas 0x0000000000400576
0x000000000040052d <+0>: push rbp
0x000000000040052e <+1>: mov rbp,rsp
0x0000000000400531 <+4>: sub rsp,0x20
0x0000000000400535 <+8>: mov DWORD PTR [rsp+0x10],0x8
0x000000000040053d <+16>: mov DWORD PTR [rsp+0x8],0x7
0x0000000000400545 <+24>: mov DWORD PTR [rsp],0x6
0x000000000040054c <+31>: mov r9d,0x5
0x0000000000400552 <+37>: mov r8d,0x4
0x0000000000400558 <+43>: mov ecx,0x3
0x000000000040055d <+48>: mov edx,0x2
0x0000000000400562 <+53>: mov esi,0x1
0x0000000000400567 <+58>: mov edi,0x400628
0x000000000040056c <+63>: mov eax,0x0
0x0000000000400571 <+68>: call 0x400410 <printf@plt>
0x0000000000400576 <+73>: mov eax,0x0
0x000000000040057b <+78>: leave
0x000000000040057c <+79>: ret
Let’s finish executing printf(), execute the instruction zeroing EAX, and note that the EAX register has
a value of exactly zero. RIP now points to the LEAVE instruction, i.e., the penultimate one in the main()
function.
(gdb) finish
Run till exit from #0 __printf (format=0x400628 "a=%d; b=%d; c=%d; d=%d; e=%d; f=%d; g=%d; h=%⤦
Ç d\n") at printf.c:29
a=1; b=2; c=3; d=4; e=5; f=6; g=7; h=8
main () at 2.c:6
6 return 0;
Value returned is $1 = 39
(gdb) next
7 };
rax 0x0 0
rbx 0x0 0
rcx 0x26 38
rdx 0x7ffff7dd59f0 140737351866864
rsi 0x7fffffd9 2147483609
rdi 0x0 0
rbp 0x7fffffffdf60 0x7fffffffdf60
rsp 0x7fffffffdf40 0x7fffffffdf40
r8 0x7ffff7dd26a0 140737351853728
52
r9 0x7ffff7a60134 140737348239668
r10 0x7fffffffd5b0 140737488344496
r11 0x7ffff7a95900 140737348458752
r12 0x400440 4195392
r13 0x7fffffffe040 140737488347200
r14 0x0 0
r15 0x0 0
rip 0x40057b 0x40057b <main+78>
...
1.11.2 ARM
ARM: 3 arguments
ARM’s traditional scheme for passing arguments (calling convention) behaves as follows: the first 4 ar-
guments are passed through the R0-R3 registers; the remaining arguments via the stack. This resembles
the arguments passing scheme in fastcall ( 6.1.3 on page 726) or win64 ( 6.1.5 on page 728).
32-bit ARM
Listing 1.53: Non-optimizing Keil 6/2013 (ARM mode)

.text:00000000 main
.text:00000004 03 30 A0 E3 MOV R3, #3
.text:00000008 02 20 A0 E3 MOV R2, #2
.text:0000000C 01 10 A0 E3 MOV R1, #1
.text:00000010 08 00 8F E2 ADR R0, aADBDCD ; "a=%d; b=%d; c=%d"
.text:00000014 06 00 00 EB BL __2printf
.text:00000018 00 00 A0 E3 MOV R0, #0 ; return 0
.text:0000001C 10 80 BD E8 LDMFD SP!, {R4,PC}
So, the first 4 arguments are passed via the R0-R3 registers in this order: a pointer to the printf() format
string in R0, then 1 in R1, 2 in R2 and 3 in R3. The instruction at 0x18 writes 0 to R0—this is return 0
C-statement. There is nothing unusual so far.
Optimizing Keil 6/2013 generates the same code.
Optimizing Keil 6/2013 (Thumb mode)
Listing 1.54: Optimizing Keil 6/2013 (Thumb mode)

.text:00000000 main
.text:00000000 10 B5 PUSH {R4,LR}
.text:00000002 03 23 MOVS R3, #3
.text:00000004 02 22 MOVS R2, #2
.text:00000006 01 21 MOVS R1, #1
.text:00000008 02 A0 ADR R0, aADBDCD ; "a=%d; b=%d; c=%d"
.text:0000000A 00 F0 0D F8 BL __2printf
.text:0000000E 00 20 MOVS R0, #0
.text:00000010 10 BD POP {R4,PC}
There is no significant difference from the non-optimized code for ARM mode.
Optimizing Keil 6/2013 (ARM mode) + let’s remove return
Let’s rework example slightly by removing return 0:
53
#include <stdio.h>
void main()
{
printf("a=%d; b=%d; c=%d", 1, 2, 3);
};
The result is somewhat unusual:
Listing 1.55: Optimizing Keil 6/2013 (ARM mode)

.text:00000014 main
.text:00000014 03 30 A0 E3 MOV R3, #3
.text:00000018 02 20 A0 E3 MOV R2, #2
.text:0000001C 01 10 A0 E3 MOV R1, #1
.text:00000020 1E 0E 8F E2 ADR R0, aADBDCD ; "a=%d; b=%d; c=%d\n"
.text:00000024 CB 18 00 EA B __2printf
This is the optimized (-O3) version for ARM mode and this time we see B as the last instruction instead
of the familiar BL. Another difference between this optimized version and the previous one (compiled
without optimization) is the lack of function prologue and epilogue (instructions preserving the R0 and LR
registers values). The B instruction just jumps to another address, without any manipulation of the LR
register, similar to JMP in x86. Why does it work? Because this code is, in fact, effectively equivalent to
the previous. There are two main reasons: 1) neither the stack nor SP (the stack pointer) is modified; 2)
the call to printf() is the last instruction, so there is nothing going on afterwards. On completion, the
printf() function simply returns the control to the address stored in LR. Since the LR currently stores
the address of the point from where our function has been called then the control from printf() will be
returned to that point. Therefore we do not have to save LR because we do not have necessity to modify
LR. And we do not have necessity to modify LR because there are no other function calls except printf().
Furthermore, after this call we do not to do anything else! That is the reason such optimization is possible.
This optimization is often used in functions where the last statement is a call to another function. A similar
example is presented here: 1.21.1 on page 155.
A somewhat simpler case was described above: 1.10 on page 41.
ARM64
Non-optimizing GCC (Linaro) 4.9
Listing 1.56: Non-optimizing GCC (Linaro) 4.9

.LC1:
.string "a=%d; b=%d; c=%d"
f2:
; save FP and LR in stack frame:
stp x29, x30, [sp, -16]!
; set stack frame (FP=SP):
add x29, sp, 0
adrp x0, .LC1
add x0, x0, :lo12:.LC1
mov w1, 1
mov w2, 2
mov w3, 3
bl printf
mov w0, 0
; restore FP and LR
ldp x29, x30, [sp], 16
ret
The first instruction STP (Store Pair) saves FP (X29) and LR (X30) in the stack. The second ADD X29, SP,
0 instruction forms the stack frame. It is just writing the value of SP into X29.
Next, we see the familiar ADRP/ADD instruction pair, which forms a pointer to the string. lo12 meaning low
12 bits, i.e., linker will write low 12 bits of LC1 address into the opcode of ADD instruction. %d in printf()
string format is a 32-bit int, so the 1, 2 and 3 are loaded into 32-bit register parts.
54
Optimizing GCC (Linaro) 4.9 generates the same code.
ARM: 8 arguments
Let’s use again the example with 9 arguments from the previous section: 1.11.1 on page 50.
#include <stdio.h>
int main()
{
return 0;
};
Optimizing Keil 6/2013: ARM mode
.text:00000028 main
.text:00000028
.text:00000028 var_18 = -0x18
.text:00000028 var_14 = -0x14
.text:00000028 var_4 = -4
.text:00000028
.text:00000028 04 E0 2D E5 STR LR, [SP,#var_4]!
.text:0000002C 14 D0 4D E2 SUB SP, SP, #0x14
.text:00000030 08 30 A0 E3 MOV R3, #8
.text:00000034 07 20 A0 E3 MOV R2, #7
.text:00000038 06 10 A0 E3 MOV R1, #6
.text:0000003C 05 00 A0 E3 MOV R0, #5
.text:00000040 04 C0 8D E2 ADD R12, SP, #0x18+var_14
.text:00000044 0F 00 8C E8 STMIA R12, {R0-R3}
.text:00000048 04 00 A0 E3 MOV R0, #4
.text:0000004C 00 00 8D E5 STR R0, [SP,#0x18+var_18]
.text:00000050 03 30 A0 E3 MOV R3, #3
.text:00000054 02 20 A0 E3 MOV R2, #2
.text:00000058 01 10 A0 E3 MOV R1, #1
.text:0000005C 6E 0F 8F E2 ADR R0, aADBDCDDDEDFDGD ; "a=%d; b=%d; c=%d; d=%d; e=%d; f=%d;
g=%"...
.text:00000060 BC 18 00 EB BL __2printf
.text:00000064 14 D0 8D E2 ADD SP, SP, #0x14
.text:00000068 04 F0 9D E4 LDR PC, [SP+4+var_4],#4
This code can be divided into several parts:

• Function prologue:
The very first STR LR, [SP,#var_4]! instruction saves LR on the stack, because we are going to
use this register for the printf() call. Exclamation mark at the end indicates pre-index.
This implies that SP is to be decreased by 4 first, and then LR will be saved at the address stored in
SP. This is similar to PUSH in x86. Read more about it at: 1.39.2 on page 437.
The second SUB SP, SP, #0x14 instruction decreases SP (the stack pointer) in order to allocate
0x14 (20) bytes on the stack. Indeed, we have to pass 5 32-bit values via the stack to the printf()
function, and each one occupies 4 bytes, which is exactly 5 ∗ 4 = 20. The other 4 32-bit values are to
be passed through registers.
• Passing 5, 6, 7 and 8 via the stack: they are stored in the R0, R1, R2 and R3 registers respectively.
Then, the ADD R12, SP, #0x18+var_14 instruction writes the stack address where these 4 variables
are to be stored, into the R12 register. var_14 is an assembly macro, equal to -0x14, created by IDA
to conveniently display the code accessing the stack. The var_? macros generated by IDA reflect
local variables in the stack.
So, SP+4 is to be stored into the R12 register.
The next STMIA R12, R0-R3 instruction writes registers R0-R3 contents to the memory pointed by
R12. STMIA abbreviates Store Multiple Increment After. Increment After implies that R12 is to be
increased by 4 after each register value is written.
55
• Passing 4 via the stack: 4 is stored in R0 and then this value, with the help of the
STR R0, [SP,#0x18+var_18] instruction is saved on the stack. var_18 is -0x18, so the offset is to
be 0, thus the value from the R0 register (4) is to be written to the address written in SP.
• Passing 1, 2 and 3 via registers: The values of the first 3 numbers (a, b, c) (1, 2, 3 respectively) are
passed through the R1, R2 and R3 registers right before the printf() call, and the other 5 values are
passed via the stack:
• printf() call.
• Function epilogue:
The ADD SP, SP, #0x14 instruction restores the SP pointer back to its former value, thus annulling
everything what has been stored on the stack. Of course, what has been stored on the stack will
stay there, but it will all be rewritten during the execution of subsequent functions.
The LDR PC, [SP+4+var_4],#4 instruction loads the saved LR value from the stack into the PC reg-
ister, thus causing the function to exit. There is no exclamation mark—indeed, PC is loaded first from
the address stored in SP (4 + var_4 = 4 + (−4) = 0, so this instruction is analogous to LDR PC, [SP],#4),
and then SP is increased by 4. This is referred as post-index69 . Why does IDA display the instruction
like that? Because it wants to illustrate the stack layout and the fact that var_4 is allocated for saving
the LR value in the local stack. This instruction is somewhat similar to POP PC in x8670 .
Optimizing Keil 6/2013: Thumb mode
.text:0000001C printf_main2
.text:0000001C
.text:0000001C var_18 = -0x18
.text:0000001C var_14 = -0x14
.text:0000001C var_8 = -8
.text:0000001C
.text:0000001C 00 B5 PUSH {LR}
.text:0000001E 08 23 MOVS R3, #8
.text:00000020 85 B0 SUB SP, SP, #0x14
.text:00000022 04 93 STR R3, [SP,#0x18+var_8]
.text:00000024 07 22 MOVS R2, #7
.text:00000026 06 21 MOVS R1, #6
.text:00000028 05 20 MOVS R0, #5
.text:0000002A 01 AB ADD R3, SP, #0x18+var_14
.text:0000002C 07 C3 STMIA R3!, {R0-R2}
.text:0000002E 04 20 MOVS R0, #4
.text:00000030 00 90 STR R0, [SP,#0x18+var_18]
.text:00000032 03 23 MOVS R3, #3
.text:00000034 02 22 MOVS R2, #2
.text:00000036 01 21 MOVS R1, #1
.text:00000038 A0 A0 ADR R0, aADBDCDDDEDFDGD ; "a=%d; b=%d; c=%d; d=%d; e=%d; f=%d;
g=%"...
.text:0000003A 06 F0 D9 F8 BL __2printf
.text:0000003E
.text:0000003E loc_3E ; CODE XREF: example13_f+16
.text:0000003E 05 B0 ADD SP, SP, #0x14
.text:00000040 00 BD POP {PC}
The output is almost like in the previous example. However, this is Thumb code and the values are packed
into stack differently: 8 goes first, then 5, 6, 7, and 4 goes third.
Optimizing Xcode 4.6.3 (LLVM): ARM mode
__text:0000290C _printf_main2
__text:0000290C
__text:0000290C var_1C = -0x1C
__text:0000290C var_C = -0xC
__text:0000290C
__text:0000290C 80 40 2D E9 STMFD SP!, {R7,LR}
__text:00002910 0D 70 A0 E1 MOV R7, SP
69 Read more about it: 1.39.2 on page 437.
70 It is impossible to set IP/EIP/RIP value using POP in x86, but anyway, you got the analogy right.
56
__text:00002914 14 D0 4D E2 SUB SP, SP, #0x14
__text:00002918 70 05 01 E3 MOV R0, #0x1570
__text:0000291C 07 C0 A0 E3 MOV R12, #7
__text:00002920 00 00 40 E3 MOVT R0, #0
__text:00002924 04 20 A0 E3 MOV R2, #4
__text:00002928 00 00 8F E0 ADD R0, PC, R0
__text:0000292C 06 30 A0 E3 MOV R3, #6
__text:00002930 05 10 A0 E3 MOV R1, #5
__text:00002934 00 20 8D E5 STR R2, [SP,#0x1C+var_1C]
__text:00002938 0A 10 8D E9 STMFA SP, {R1,R3,R12}
__text:0000293C 08 90 A0 E3 MOV R9, #8
__text:00002940 01 10 A0 E3 MOV R1, #1
__text:00002944 02 20 A0 E3 MOV R2, #2
__text:00002948 03 30 A0 E3 MOV R3, #3
__text:0000294C 10 90 8D E5 STR R9, [SP,#0x1C+var_C]
__text:00002950 A4 05 00 EB BL _printf
__text:00002954 07 D0 A0 E1 MOV SP, R7
__text:00002958 80 80 BD E8 LDMFD SP!, {R7,PC}
Almost the same as what we have already seen, with the exception of STMFA (Store Multiple Full Ascending)
instruction, which is a synonym of STMIB (Store Multiple Increment Before) instruction. This instruction
increases the value in the SP register and only then writes the next register value into the memory, rather
than performing those two actions in the opposite order.
Another thing that catches the eye is that the instructions are arranged seemingly random. For example,
the value in the R0 register is manipulated in three places, at addresses 0x2918, 0x2920 and 0x2928, when
it would be possible to do it in one point.
However, the optimizing compiler may have its own reasons on how to order the instructions so to achieve
higher efficiency during the execution.
Usually, the processor attempts to simultaneously execute instructions located side-by-side.
For example, instructions like MOVT R0, #0 and ADD R0, PC, R0 cannot be executed simultaneously
since they both modify the R0 register. On the other hand, MOVT R0, #0 and MOV R2, #4 instructions
can be executed simultaneously since the effects of their execution are not conflicting with each other.
Presumably, the compiler tries to generate code in such a manner (wherever it is possible).
Optimizing Xcode 4.6.3 (LLVM): Thumb-2 mode
__text:00002BA0 _printf_main2
__text:00002BA0
__text:00002BA0 var_1C = -0x1C
__text:00002BA0 var_18 = -0x18
__text:00002BA0 var_C = -0xC
__text:00002BA0
__text:00002BA0 80 B5 PUSH {R7,LR}
__text:00002BA2 6F 46 MOV R7, SP
__text:00002BA4 85 B0 SUB SP, SP, #0x14
__text:00002BA6 41 F2 D8 20 MOVW R0, #0x12D8
__text:00002BAA 4F F0 07 0C MOV.W R12, #7
__text:00002BAE C0 F2 00 00 MOVT.W R0, #0
__text:00002BB2 04 22 MOVS R2, #4
__text:00002BB4 78 44 ADD R0, PC ; char *
__text:00002BB6 06 23 MOVS R3, #6
__text:00002BB8 05 21 MOVS R1, #5
__text:00002BBA 0D F1 04 0E ADD.W LR, SP, #0x1C+var_18
__text:00002BBE 00 92 STR R2, [SP,#0x1C+var_1C]
__text:00002BC0 4F F0 08 09 MOV.W R9, #8
__text:00002BC4 8E E8 0A 10 STMIA.W LR, {R1,R3,R12}
__text:00002BC8 01 21 MOVS R1, #1
__text:00002BCA 02 22 MOVS R2, #2
__text:00002BCC 03 23 MOVS R3, #3
__text:00002BCE CD F8 10 90 STR.W R9, [SP,#0x1C+var_C]
__text:00002BD2 01 F0 0A EA BLX _printf
__text:00002BD6 05 B0 ADD SP, SP, #0x14
__text:00002BD8 80 BD POP {R7,PC}
57
The output is almost the same as in the previous example, with the exception that Thumb-instructions
are used instead.
ARM64

.LC2:
.string "a=%d; b=%d; c=%d; d=%d; e=%d; f=%d; g=%d; h=%d\n"
f3:
; grab more space in stack:
sub sp, sp, #32
stp x29, x30, [sp,16]
; set stack frame (FP=SP):
add x29, sp, 16
adrp x0, .LC2 ; "a=%d; b=%d; c=%d; d=%d; e=%d; f=%d; g=%d; h=%d\n"
mov w1, 8 ; 9th argument
str w1, [sp] ; store 9th argument in the stack
mov w1, 1
mov w2, 2
mov w3, 3
mov w4, 4
mov w5, 5
mov w6, 6
mov w7, 7
bl printf
sub sp, x29, #16
; restore FP and LR
ldp x29, x30, [sp,16]
add sp, sp, 32
ret
The first 8 arguments are passed in X- or W-registers: [Procedure Call Standard for the ARM 64-bit Archi-
tecture (AArch64), (2013)]71 . A string pointer requires a 64-bit register, so it’s passed in X0. All other
values have a int 32-bit type, so they are stored in the 32-bit part of the registers (W-). The 9th argument
(8) is passed via the stack. Indeed: it’s not possible to pass large number of arguments through registers,
because the number of registers is limited.
Optimizing GCC (Linaro) 4.9 generates the same code.
1.11.3 MIPS
3 arguments
Optimizing GCC 4.4.5
The main difference with the “Hello, world!” example is that in this case printf() is called instead of
puts() and 3 more arguments are passed through the registers $5…$7 (or $A0…$A2). That is why these
registers are prefixed with A-, which implies they are used for function arguments passing.

$LC0:
.ascii "a=%d; b=%d; c=%d\000"
main:
; function prologue:
lui $28,%hi(__gnu_local_gp)
addiu $sp,$sp,-32
addiu $28,$28,%lo(__gnu_local_gp)
sw $31,28($sp)
58
; load address of printf():
lw $25,%call16(printf)($28)
; load address of the text string and set 1st argument of printf():
lui $4,%hi($LC0)
addiu $4,$4,%lo($LC0)
; set 2nd argument of printf():
li $5,1 # 0x1
; set 3rd argument of printf():
li $6,2 # 0x2
; call printf():
jalr $25
; set 4th argument of printf() (branch delay slot):
li $7,3 # 0x3
; function epilogue:
lw $31,28($sp)
; set return value to 0:
move $2,$0
; return
j $31
addiu $sp,$sp,32 ; branch delay slot

.text:00000000 main:
.text:00000000
.text:00000000 var_10 = -0x10
.text:00000000 var_4 = -4
.text:00000000
.text:00000000 lui $gp, (__gnu_local_gp >> 16)
.text:00000004 addiu $sp, -0x20
.text:00000008 la $gp, (__gnu_local_gp & 0xFFFF)
.text:0000000C sw $ra, 0x20+var_4($sp)
.text:00000010 sw $gp, 0x20+var_10($sp)
; load address of printf():
.text:00000014 lw $t9, (printf & 0xFFFF)($gp)
; load address of the text string and set 1st argument of printf():
.text:00000018 la $a0, $LC0 # "a=%d; b=%d; c=%d"
.text:00000020 li $a1, 1
.text:00000024 li $a2, 2
; call printf():
.text:00000028 jalr $t9
; set 4th argument of printf() (branch delay slot):
.text:0000002C li $a3, 3
.text:00000030 lw $ra, 0x20+var_4($sp)
.text:00000034 move $v0, $zero
; return
.text:00000038 jr $ra
.text:0000003C addiu $sp, 0x20 ; branch delay slot
IDA has coalesced pair of LUI and ADDIU instructions into one LA pseudo instruction. That’s why there are
no instruction at address 0x1C: because LA occupies 8 bytes.
Non-optimizing GCC 4.4.5
Non-optimizing GCC is more verbose:

$LC0:
.ascii "a=%d; b=%d; c=%d\000"
main:
addiu $sp,$sp,-32
59
sw $31,28($sp)
sw $fp,24($sp)
move $fp,$sp
; load address of the text string:
lui $2,%hi($LC0)
; set 1st argument of printf():
move $4,$2
li $5,1 # 0x1
li $6,2 # 0x2
; set 4th argument of printf():
li $7,3 # 0x3
; get address of printf():
nop
; call printf():
move $25,$2
jalr $25
nop
lw $28,16($fp)
move $2,$0
move $sp,$fp
lw $31,28($sp)
lw $fp,24($sp)
addiu $sp,$sp,32
; return
j $31
nop

.text:00000000
.text:00000000 var_10 = -0x10
.text:00000000 var_8 = -8
.text:00000000 var_4 = -4
.text:00000000
.text:00000000 addiu $sp, -0x20
.text:00000004 sw $ra, 0x20+var_4($sp)
.text:00000008 sw $fp, 0x20+var_8($sp)
.text:0000000C move $fp, $sp
.text:00000010 la $gp, __gnu_local_gp
.text:00000018 sw $gp, 0x20+var_10($sp)
; load address of the text string:
.text:0000001C la $v0, aADBDCD # "a=%d; b=%d; c=%d"
; set 1st argument of printf():
.text:00000024 move $a0, $v0
.text:00000028 li $a1, 1
.text:0000002C li $a2, 2
; set 4th argument of printf():
.text:00000030 li $a3, 3
; get address of printf():
.text:00000034 lw $v0, (printf & 0xFFFF)($gp)
.text:00000038 or $at, $zero
; call printf():
.text:0000003C move $t9, $v0
.text:00000040 jalr $t9
.text:00000044 or $at, $zero ; NOP
.text:00000048 lw $gp, 0x20+var_10($fp)
60
.text:0000004C move $v0, $zero
.text:00000050 move $sp, $fp
.text:00000054 lw $ra, 0x20+var_4($sp)
.text:00000058 lw $fp, 0x20+var_8($sp)
.text:0000005C addiu $sp, 0x20
; return
.text:00000060 jr $ra
8 arguments
Let’s use again the example with 9 arguments from the previous section: 1.11.1 on page 50.
#include <stdio.h>
int main()
{
return 0;
};
Only the first 4 arguments are passed in the $A0 …$A3 registers, the rest are passed via the stack.
This is the O32 calling convention (which is the most common one in the MIPS world). Other calling
conventions (like N32) may use the registers for different purposes.
SW abbreviates “Store Word” (from register to memory). MIPS lacks instructions for storing a value into
memory, so an instruction pair has to be used instead (LI/SW).

$LC0:
.ascii "a=%d; b=%d; c=%d; d=%d; e=%d; f=%d; g=%d; h=%d\012\000"
main:
addiu $sp,$sp,-56
sw $31,52($sp)
; pass 5th argument in stack:
li $2,4 # 0x4
sw $2,16($sp)
li $2,5 # 0x5
sw $2,20($sp)
li $2,6 # 0x6
sw $2,24($sp)
li $2,7 # 0x7
sw $2,28($sp)
; pass 1st argument in $a0:
lui $4,%hi($LC0)
li $2,8 # 0x8
sw $2,32($sp)
; pass 2nd argument in $a1:
li $5,1 # 0x1
; pass 3rd argument in $a2:
li $6,2 # 0x2
; call printf():
jalr $25
; pass 4th argument in $a3 (branch delay slot):
61
li $7,3 # 0x3
lw $31,52($sp)
move $2,$0
; return
j $31

.text:00000000
.text:00000000 var_28 = -0x28
.text:00000000 var_24 = -0x24
.text:00000000 var_20 = -0x20
.text:00000000 var_1C = -0x1C
.text:00000000 var_18 = -0x18
.text:00000000 var_10 = -0x10
.text:00000000 var_4 = -4
.text:00000000
.text:00000004 addiu $sp, -0x38
.text:00000010 sw $gp, 0x38+var_10($sp)
.text:00000014 li $v0, 4
.text:00000018 sw $v0, 0x38+var_28($sp)
.text:0000001C li $v0, 5
.text:00000020 sw $v0, 0x38+var_24($sp)
.text:00000024 li $v0, 6
.text:00000028 sw $v0, 0x38+var_20($sp)
.text:0000002C li $v0, 7
.text:00000034 sw $v0, 0x38+var_1C($sp)
; prepare 1st argument in $a0:
.text:00000038 lui $a0, ($LC0 >> 16) # "a=%d; b=%d; c=%d; d=%d; e=%d; f=%d;
g=%"...
.text:0000003C li $v0, 8
.text:00000040 sw $v0, 0x38+var_18($sp)
.text:00000044 la $a0, ($LC0 & 0xFFFF) # "a=%d; b=%d; c=%d; d=%d; e=%d;
f=%d; g=%"...
.text:00000048 li $a1, 1
.text:0000004C li $a2, 2
; call printf():
.text:00000050 jalr $t9
; pass 4th argument in $a3 (branch delay slot):
.text:00000054 li $a3, 3
.text:00000058 lw $ra, 0x38+var_4($sp)
; return
.text:00000060 jr $ra
.text:00000064 addiu $sp, 0x38 ; branch delay slot
Non-optimizing GCC is more verbose:
62
$LC0:
.ascii "a=%d; b=%d; c=%d; d=%d; e=%d; f=%d; g=%d; h=%d\012\000"
main:
addiu $sp,$sp,-56
sw $31,52($sp)
sw $fp,48($sp)
move $fp,$sp
lui $2,%hi($LC0)
li $3,4 # 0x4
sw $3,16($sp)
li $3,5 # 0x5
sw $3,20($sp)
li $3,6 # 0x6
sw $3,24($sp)
li $3,7 # 0x7
sw $3,28($sp)
li $3,8 # 0x8
sw $3,32($sp)
move $4,$2
li $5,1 # 0x1
li $6,2 # 0x2
; pass 4th argument in $a3:
li $7,3 # 0x3
; call printf():
nop
move $25,$2
jalr $25
nop
lw $28,40($fp)
move $2,$0
move $sp,$fp
lw $31,52($sp)
lw $fp,48($sp)
addiu $sp,$sp,56
; return
j $31
nop

.text:00000000
.text:00000000 var_28 = -0x28
.text:00000000 var_24 = -0x24
.text:00000000 var_20 = -0x20
.text:00000000 var_1C = -0x1C
.text:00000000 var_18 = -0x18
.text:00000000 var_10 = -0x10
.text:00000000 var_8 = -8
.text:00000000 var_4 = -4
.text:00000000
.text:00000000 addiu $sp, -0x38
.text:00000004 sw $ra, 0x38+var_4($sp)
63
.text:00000008 sw $fp, 0x38+var_8($sp)
.text:00000018 sw $gp, 0x38+var_10($sp)
.text:0000001C la $v0, aADBDCDDDEDFDGD # "a=%d; b=%d; c=%d; d=%d; e=%d;
f=%d; g=%"...
.text:00000024 li $v1, 4
.text:00000028 sw $v1, 0x38+var_28($sp)
.text:0000002C li $v1, 5
.text:00000030 sw $v1, 0x38+var_24($sp)
.text:00000034 li $v1, 6
.text:00000038 sw $v1, 0x38+var_20($sp)
.text:0000003C li $v1, 7
.text:00000040 sw $v1, 0x38+var_1C($sp)
.text:00000044 li $v1, 8
.text:00000048 sw $v1, 0x38+var_18($sp)
.text:0000004C move $a0, $v0
.text:00000050 li $a1, 1
.text:00000054 li $a2, 2
; pass 4th argument in $a3:
.text:00000058 li $a3, 3
; call printf():
.text:0000005C lw $v0, (printf & 0xFFFF)($gp)
.text:00000064 move $t9, $v0
.text:00000068 jalr $t9
.text:0000006C or $at, $zero ; NOP
.text:00000070 lw $gp, 0x38+var_10($fp)
.text:00000078 move $sp, $fp
.text:0000007C lw $ra, 0x38+var_4($sp)
.text:00000080 lw $fp, 0x38+var_8($sp)
.text:00000084 addiu $sp, 0x38
; return
.text:00000088 jr $ra
1.11.4 Conclusion
Here is a rough skeleton of the function call:
Listing 1.66: x86
...
PUSH 3rd argument
PUSH 2nd argument
PUSH 1st argument
CALL function
; modify stack pointer (if needed)
Listing 1.67: x64 (MSVC)

MOV RCX, 1st argument
MOV RDX, 2nd argument
MOV R8, 3rd argument
MOV R9, 4th argument
...
PUSH 5th, 6th argument, etc. (if needed)
CALL function
64
Listing 1.68: x64 (GCC)
MOV RDI, 1st argument
MOV RSI, 2nd argument
MOV RDX, 3rd argument
MOV RCX, 4th argument
...
PUSH 7th, 8th argument, etc. (if needed)
CALL function
Listing 1.69: ARM

MOV R0, 1st argument
MOV R1, 2nd argument
MOV R2, 3rd argument
; pass 5th, 6th argument, etc., in stack (if needed)
BL function
Listing 1.70: ARM64

MOV X0, 1st argument
MOV X1, 2nd argument
MOV X2, 3rd argument
MOV X3, 4th argument
BL function
Listing 1.71: MIPS (O32 calling convention)

LI $4, 1st argument ; AKA $A0
LI $5, 2nd argument ; AKA $A1
LI $6, 3rd argument ; AKA $A2
LI $7, 4th argument ; AKA $A3
LW temp_reg, address of function
JALR temp_reg
1.11.5 By the way

By the way, this difference between the arguments passing in x86, x64, fastcall, ARM and MIPS is a good
illustration of the fact that the CPU is oblivious to how the arguments are passed to functions. It is also
possible to create a hypothetical compiler able to pass arguments via a special structure without using
stack at all.
MIPS $A0 …$A3 registers are labeled this way only for convenience (that is in the O32 calling convention).
Programmers may use any other register (well, maybe except $ZERO) to pass data or use any other calling
convention.
The CPU is not aware of calling conventions whatsoever.
We may also recall how new coming assembly language programmers passing arguments into other func-
tions: usually via registers, without any explicit order, or even via global variables. Of course, it works
fine.
1.12 scanf()
Now let’s use scanf().
65
1.12.1 Simple example
#include <stdio.h>
int main()
{
int x;
printf ("Enter X:\n");
scanf ("%d", &x);
printf ("You entered %d...\n", x);
return 0;
};
It’s not clever to use scanf() for user interactions nowadays. But we can, however, illustrate passing a
pointer to a variable of type int.
About pointers
Pointers are one of the fundamental concepts in computer science. Often, passing a large array, structure
or object as an argument to another function is too expensive, while passing their address is much cheaper.
For example, if you going to print a text string to console, it’s much easier to pass its address into OS kernel.
In addition if the callee function needs to modify something in the large array or structure received as a
parameter and return back the entire structure then the situation is close to absurd. So the simplest thing
to do is to pass the address of the array or structure to the callee function, and let it change what needs
to be changed.
A pointer in C/C++—is simply an address of some memory location.
In x86, the address is represented as a 32-bit number (i.e., it occupies 4 bytes), while in x86-64 it is a 64-
bit number (occupying 8 bytes). By the way, that is the reason behind some people’s indignation related
to switching to x86-64—all pointers in the x64-architecture require twice as much space, including cache
memory, which is “expensive” memory.
It is possible to work with untyped pointers only, given some effort; e.g. the standard C function memcpy(),
that copies a block from one memory location to another, takes 2 pointers of type void* as arguments,
since it is impossible to predict the type of the data you would like to copy. Data types are not important,
only the block size matters.
Pointers are also widely used when a function needs to return more than one value (we are going to get
back to this later ( 3.23 on page 596) ).
scanf() function—is such a case.
Besides the fact that the function needs to indicate how many values were successfully read, it also needs
to return all these values.
In C/C++ the pointer type is only needed for compile-time type checking.
Internally, in the compiled code there is no information about pointer types at all.
x86
MSVC
Here is what we get after compiling with MSVC 2010:

CONST SEGMENT
$SG3831 DB 'Enter X:', 0aH, 00H
$SG3832 DB '%d', 00H
$SG3833 DB 'You entered %d...', 0aH, 00H
CONST ENDS
PUBLIC _main
EXTRN _scanf:PROC
EXTRN _printf:PROC
_TEXT SEGMENT
66
_x$ = -4 ; size = 4
_main PROC
push ebp
mov ebp, esp
push ecx
push OFFSET $SG3831 ; 'Enter X:'
call _printf
add esp, 4
lea eax, DWORD PTR _x$[ebp]
push eax
push OFFSET $SG3832 ; '%d'
call _scanf
add esp, 8
mov ecx, DWORD PTR _x$[ebp]
push ecx
push OFFSET $SG3833 ; 'You entered %d...'
call _printf
add esp, 8
; return 0
xor eax, eax
mov esp, ebp
pop ebp
ret 0
_main ENDP
_TEXT ENDS
x is a local variable.
According to the C/C++ standard it must be visible only in this function and not from any other external
scope. Traditionally, local variables are stored on the stack. There are probably other ways to allocate
them, but in x86 that is the way it is.
The goal of the instruction following the function prologue, PUSH ECX, is not to save the ECX state (notice
the absence of corresponding POP ECX at the function’s end).
In fact it allocates 4 bytes on the stack for storing the x variable.
x is to be accessed with the assistance of the _x$ macro (it equals to -4) and the EBP register pointing to
the current frame.
Over the span of the function’s execution, EBP is pointing to the current stack frame making it possible to
access local variables and function arguments via EBP+offset.
It is also possible to use ESP for the same purpose, although that is not very convenient since it changes
frequently. The value of the EBP could be perceived as a frozen state of the value in ESP at the start of
the function’s execution.
Here is a typical stack frame layout in 32-bit environment:
… …
EBP-8 local variable #2, marked in IDA as var_8
EBP-4 local variable #1, marked in IDA as var_4
EBP saved value of EBP
EBP+4 return address
EBP+8 argument#1, marked in IDA as arg_0
EBP+0xC argument#2, marked in IDA as arg_4
EBP+0x10 argument#3, marked in IDA as arg_8
… …
The scanf() function in our example has two arguments.
The first one is a pointer to the string containing %d and the second is the address of the x variable.
First, the x variable’s address is loaded into the EAX register by the
lea eax, DWORD PTR _x$[ebp] instruction.
LEA stands for load effective address, and is often used for forming an address ( .1.6 on page 1014).
We could say that in this case LEA simply stores the sum of the EBP register value and the _x$ macro in
the EAX register.
This is the same as lea eax, [ebp-4].
67
So, 4 is being subtracted from the EBP register value and the result is loaded in the EAX register. Next the
EAX register value is pushed into the stack and scanf() is being called.
printf() is being called after that with its first argument — a pointer to the string: You entered %d...\n.
The second argument is prepared with: mov ecx, [ebp-4]. The instruction stores the x variable value
and not its address, in the ECX register.
Next the value in the ECX is stored on the stack and the last printf() is being called.
68
MSVC + OllyDbg
Let’s try this example in OllyDbg. Let’s load it and keep pressing F8 (step over) until we reach our exe-
cutable file instead of ntdll.dll. Scroll up until main() appears.
Click on the first instruction (PUSH EBP), press F2 (set a breakpoint), then F9 (Run). The breakpoint will
be triggered when main() begins.
Let’s trace to the point where the address of the variable x is calculated:
Figure 1.13: OllyDbg: The address of the local variable is calculated
Right-click the EAX in the registers window and then select “Follow in stack”.
This address will appear in the stack window. The red arrow has been added, pointing to the variable in
the local stack. At that moment this location contains some garbage (0x6E494714). Now with the help
of PUSH instruction the address of this stack element is going to be stored to the same stack on the next
position. Let’s trace with F8 until the scanf() execution completes. During the scanf() execution, we
input, for example, 123, in the console window:
Enter X:
123
69
scanf() completed its execution already:
Figure 1.14: OllyDbg: scanf() executed
scanf() returns 1 in EAX, which implies that it has read successfully one value. If we look again at the
stack element corresponding to the local variable it now contains 0x7B (123).
70
Later this value is copied from the stack to the ECX register and passed to printf():
Figure 1.15: OllyDbg: preparing the value for passing to printf()
GCC
Let’s try to compile this code in GCC 4.4.1 under Linux:

main proc near

var_1C = dword ptr -1Ch
push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
sub esp, 20h
mov [esp+20h+var_20], offset aEnterX ; "Enter X:"
call _puts
mov eax, offset aD ; "%d"
lea edx, [esp+20h+var_4]
mov [esp+20h+var_1C], edx
call ___isoc99_scanf
mov edx, [esp+20h+var_4]
mov eax, offset aYouEnteredD___ ; "You entered %d...\n"
call _printf
mov eax, 0
leave
retn
main endp
GCC has replaced the printf() call with call to puts(). The reason for this was explained in ( 1.5.3 on
page 20).
As in the MSVC example—the arguments are placed on the stack using the MOV instruction.
By the way
71
This simple example is a demonstration of the fact that the compiler translates a list of expressions in
C/C++-block into a sequential list of instructions. There is nothing in between expressions in C/C++, and
so in the resulting machine code, there are nothing between, the control flow slips from one expression
to the next one.
x64
The picture here is similar with the difference that the registers, rather than the stack, are used for argu-
ments passing.
MSVC

_DATA SEGMENT
$SG1291 DB '%d', 00H
_DATA ENDS
_TEXT SEGMENT
x$ = 32
main PROC
$LN3:
sub rsp, 56
lea rcx, OFFSET FLAT:$SG1289 ; 'Enter X:'
call printf
lea rdx, QWORD PTR x$[rsp]
lea rcx, OFFSET FLAT:$SG1291 ; '%d'
call scanf
mov edx, DWORD PTR x$[rsp]
lea rcx, OFFSET FLAT:$SG1292 ; 'You entered %d...'
call printf
; return 0
xor eax, eax
add rsp, 56
ret 0
main ENDP
_TEXT ENDS
GCC

.LC0:
.string "Enter X:"
.LC1:
.string "%d"
.LC2:
.string "You entered %d...\n"
main:
sub rsp, 24
mov edi, OFFSET FLAT:.LC0 ; "Enter X:"
call puts
lea rsi, [rsp+12]
mov edi, OFFSET FLAT:.LC1 ; "%d"
xor eax, eax
call __isoc99_scanf
mov esi, DWORD PTR [rsp+12]
mov edi, OFFSET FLAT:.LC2 ; "You entered %d...\n"
xor eax, eax
call printf
; return 0
72
xor eax, eax
add rsp, 24
ret
ARM
.text:00000042 scanf_main
.text:00000042
.text:00000042 var_8 = -8
.text:00000042
.text:00000042 08 B5 PUSH {R3,LR}
.text:00000044 A9 A0 ADR R0, aEnterX ; "Enter X:\n"
.text:00000046 06 F0 D3 F8 BL __2printf
.text:0000004A 69 46 MOV R1, SP
.text:0000004C AA A0 ADR R0, aD ; "%d"
.text:0000004E 06 F0 CD F8 BL __0scanf
.text:00000052 00 99 LDR R1, [SP,#8+var_8]
.text:00000054 A9 A0 ADR R0, aYouEnteredD___ ; "You entered %d...\n"
.text:00000056 06 F0 CB F8 BL __2printf
.text:0000005A 00 20 MOVS R0, #0
.text:0000005C 08 BD POP {R3,PC}
In order for scanf() to be able to read item it needs a parameter—pointer to an int. int is 32-bit, so we
need 4 bytes to store it somewhere in memory, and it fits exactly in a 32-bit register. A place for the local
variable x is allocated in the stack and IDA has named it var_8. It is not necessary, however, to allocate
a such since SP (stack pointer) is already pointing to that space and it can be used directly.
So, SP’s value is copied to the R1 register and, together with the format-string, passed to scanf().
PUSH/POP instructions behaves differently in ARM than in x86 (it’s the other way around). They are syn-
onyms to STM/STMDB/LDM/LDMIA instructions. And PUSH instruction first writes a value into the stack, and
then subtracts SP by 4. POP first adds 4 to SP, and then reads a value from the stack. Hence, after PUSH,
SP points to an unused space in stack. It is used by scanf(), and by printf() after.
LDMIA means Load Multiple Registers Increment address After each transfer. STMDB means Store Multiple
Registers Decrement address Before each transfer.
Later, with the help of the LDR instruction, this value is moved from the stack to the R1 register in order
to be passed to printf().
ARM64
Listing 1.74: Non-optimizing GCC 4.9.1 ARM64

1 .LC0:
2 .string "Enter X:"
3 .LC1:
4 .string "%d"
5 .LC2:
6 .string "You entered %d...\n"
7 scanf_main:
8 ; subtract 32 from SP, then save FP and LR in stack frame:
9 stp x29, x30, [sp, -32]!
10 ; set stack frame (FP=SP)
11 add x29, sp, 0
12 ; load pointer to the "Enter X:" string:
13 adrp x0, .LC0
14 add x0, x0, :lo12:.LC0
15 ; X0=pointer to the "Enter X:" string
16 ; print it:
17 bl puts
18 ; load pointer to the "%d" string:
19 adrp x0, .LC1
20 add x0, x0, :lo12:.LC1
21 ; find a space in stack frame for "x" variable (X1=FP+28):
73
22 add x1, x29, 28
23 ; X1=address of "x" variable
24 ; pass the address to scanf() and call it:
25 bl __isoc99_scanf
26 ; load 32-bit value from the variable in stack frame:
27 ldr w1, [x29,28]
28 ; W1=x
29 ; load pointer to the "You entered %d...\n" string
30 ; printf() will take text string from X0 and "x" variable from X1 (or W1)
31 adrp x0, .LC2
32 add x0, x0, :lo12:.LC2
33 bl printf
34 ; return 0
35 mov w0, 0
36 ; restore FP and LR, then add 32 to SP:
37 ldp x29, x30, [sp], 32
38 ret
There is 32 bytes are allocated for stack frame, which is bigger than it needed. Perhaps some memory
aligning issue? The most interesting part is finding space for the x variable in the stack frame (line 22).
Why 28? Somehow, compiler decided to place this variable at the end of stack frame instead of beginning.
The address is passed to scanf(), which just stores the user input value in the memory at that address.
This is 32-bit value of type int. The value is fetched at line 27 and then passed to printf().
MIPS
A place in the local stack is allocated for the x variable, and it is to be referred as $sp + 24.
Its address is passed to scanf(), and the user input values is loaded using the LW (“Load Word”) instruction
and then passed to printf().

$LC0:
.ascii "Enter X:\000"
$LC1:
.ascii "%d\000"
$LC2:
.ascii "You entered %d...\012\000"
main:
addiu $sp,$sp,-40
sw $31,36($sp)
; call puts():
lw $25,%call16(puts)($28)
lui $4,%hi($LC0)
jalr $25
addiu $4,$4,%lo($LC0) ; branch delay slot
; call scanf():
lw $28,16($sp)
lui $4,%hi($LC1)
lw $25,%call16(__isoc99_scanf)($28)
; set 2nd argument of scanf(), $a1=$sp+24:
addiu $5,$sp,24
jalr $25
; call printf():
lw $28,16($sp)
; set 2nd argument of printf(),
; load word at address $sp+24:
lw $5,24($sp)
lui $4,%hi($LC2)
jalr $25
74
lw $31,36($sp)
move $2,$0
; return:
j $31
IDA displays the stack layout as follows:

.text:00000000
.text:00000000 var_18 = -0x18
.text:00000000 var_10 = -0x10
.text:00000000 var_4 = -4
.text:00000000
.text:00000004 addiu $sp, -0x28
.text:00000010 sw $gp, 0x28+var_18($sp)
; call puts():
.text:00000014 lw $t9, (puts & 0xFFFF)($gp)
.text:00000018 lui $a0, ($LC0 >> 16) # "Enter X:"
.text:0000001C jalr $t9
.text:00000020 la $a0, ($LC0 & 0xFFFF) # "Enter X:" ; branch delay slot
; call scanf():
.text:00000024 lw $gp, 0x28+var_18($sp)
.text:00000028 lui $a0, ($LC1 >> 16) # "%d"
.text:0000002C lw $t9, (__isoc99_scanf & 0xFFFF)($gp)
; set 2nd argument of scanf(), $a1=$sp+24:
.text:00000030 addiu $a1, $sp, 0x28+var_10
.text:00000034 jalr $t9 ; branch delay slot
.text:00000038 la $a0, ($LC1 & 0xFFFF) # "%d"
; call printf():
.text:0000003C lw $gp, 0x28+var_18($sp)
; set 2nd argument of printf(),
; load word at address $sp+24:
.text:00000040 lw $a1, 0x28+var_10($sp)
.text:00000048 lui $a0, ($LC2 >> 16) # "You entered %d...\n"
.text:00000050 la $a0, ($LC2 & 0xFFFF) # "You entered %d...\n" ; branch delay
slot
.text:00000054 lw $ra, 0x28+var_4($sp)
; return:
.text:0000005C jr $ra
1.12.2 The classic mistake

It’s a very popular mistake (and/or typo) to pass a value of x instead of pointer to x:
#include <stdio.h>
int main()
{
int x;
scanf ("%d", x); // BUG
return 0;
75
};
So what happens here? x is not initialized and contains some random noise from local stack. When
scanf() called, it takes string from user, parses it into number and tries to write it into x, treating it as
an address in memory. But there is a random noise, so scanf() will try to write at random address. Most
likely, the process will crash.
Interestingly enough, some CRT libraries in debug build, put visually distinctive patterns into memory just
allocated, like 0xCCCCCCCC or 0x0BADF00D and so on. In this case, x may contain 0xCCCCCCCC, and
scanf() would try to write at address 0xCCCCCCCC. And if you’ll notice that something in your process
tries to write at address 0xCCCCCCCC, you’ll know that uninitialized variable (or pointer) gets used without
prior initialization. This is better than as if newly allocated memory is just cleared by zero bytes.
1.12.3 Global variables

What if the x variable from the previous example isn’t local but a global one? Then it would have been
accessible from any point, not only from the function body. Global variables are considered anti-pattern,
but for the sake of the experiment, we could do this.
#include <stdio.h>
// now x is global variable

int x;
int main()
{
scanf ("%d", &x);
return 0;
};
MSVC: x86
_DATA SEGMENT
COMM _x:DWORD
$SG2457 DB '%d', 00H
_DATA ENDS
PUBLIC _main
EXTRN _scanf:PROC
EXTRN _printf:PROC
_TEXT SEGMENT
_main PROC
push ebp
mov ebp, esp
push OFFSET $SG2456
call _printf
add esp, 4
push OFFSET _x
push OFFSET $SG2457
call _scanf
add esp, 8
mov eax, DWORD PTR _x
push eax
push OFFSET $SG2458
call _printf
add esp, 8
xor eax, eax
pop ebp
ret 0
_main ENDP
_TEXT ENDS
76
In this case the x variable is defined in the _DATA segment and no memory is allocated in the local stack. It
is accessed directly, not through the stack. Uninitialized global variables take no space in the executable
file (indeed, why one needs to allocate space for variables initially set to zero?), but when someone
accesses their address, the OS will allocate a block of zeros there72 .
Now let’s explicitly assign a value to the variable:
int x=10; // default value
We got:
_DATA SEGMENT
_x DD 0aH
...
Here we see a value 0xA of DWORD type (DD stands for DWORD = 32 bit) for this variable.
If you open the compiled .exe in IDA, you can see the x variable placed at the beginning of the _DATA
segment, and after it you can see text strings.
If you open the compiled .exe from the previous example in IDA, where the value of x hasn’t been set, you
would see something like this:
Listing 1.77: IDA

.data:0040FA80 _x dd ? ; DATA XREF: _main+10
.data:0040FA80 ; _main+22
.data:0040FA84 dword_40FA84 dd ? ; DATA XREF: _memset+1E
.data:0040FA84 ; unknown_libname_1+28
.data:0040FA88 dword_40FA88 dd ? ; DATA XREF: ___sbh_find_block+5
.data:0040FA88 ; ___sbh_free_block+2BC
.data:0040FA8C ; LPVOID lpMem
.data:0040FA8C lpMem dd ? ; DATA XREF: ___sbh_find_block+B
.data:0040FA8C ; ___sbh_free_block+2CA
.data:0040FA90 dword_40FA90 dd ? ; DATA XREF: _V6_HeapAlloc+13
.data:0040FA90 ; __calloc_impl+72
.data:0040FA94 dword_40FA94 dd ? ; DATA XREF: ___sbh_free_block+2FE
_x is marked with ? with the rest of the variables that do not need to be initialized. This implies that after
loading the .exe to the memory, a space for all these variables is to be allocated and filled with zeros
[ISO/IEC 9899:TC3 (C C99 standard), (2007)6.7.8p10]. But in the .exe file these uninitialized variables do
not occupy anything. This is convenient for large arrays, for example.
72 That is how a VM behaves
77
MSVC: x86 + OllyDbg
Things are even simpler here:
Figure 1.16: OllyDbg: after scanf() execution
The variable is located in the data segment. After the PUSH instruction (pushing the address of x) gets
executed, the address appears in the stack window. Right-click on that row and select “Follow in dump”.
The variable will appear in the memory window on the left. After we have entered 123 in the console,
0x7B appears in the memory window (see the highlighted screenshot regions).
But why is the first byte 7B? Thinking logically, 00 00 00 7B must be there. The cause for this is referred
as endianness, and x86 uses little-endian. This implies that the lowest byte is written first, and the highest
written last. Read more about it at: 2.8 on page 460. Back to the example, the 32-bit value is loaded from
this memory address into EAX and passed to printf().
The memory address of x is 0x00C53394.
78
In OllyDbg we can review the process memory map (Alt-M) and we can see that this address is inside the
.data PE-segment of our program:
Figure 1.17: OllyDbg: process memory map
GCC: x86
The picture in Linux is near the same, with the difference that the uninitialized variables are located in
the _bss segment. In ELF73 file this segment has the following attributes:
; Segment type: Uninitialized
; Segment permissions: Read/Write
If you, however, initialize the variable with some value e.g. 10, it is to be placed in the _data segment,
which has the following attributes:
; Segment type: Pure data
; Segment permissions: Read/Write
MSVC: x64

_DATA SEGMENT
COMM x:DWORD
$SG2925 DB '%d', 00H
_DATA ENDS
_TEXT SEGMENT
main PROC
73 Executable File format widely used in *NIX systems including Linux
79
$LN3:
sub rsp, 40

call printf
lea rdx, OFFSET FLAT:x
call scanf
mov edx, DWORD PTR x
call printf
; return 0
xor eax, eax
add rsp, 40
ret 0
main ENDP
_TEXT ENDS
The code is almost the same as in x86. Please note that the address of the x variable is passed to
scanf() using a LEA instruction, while the variable’s value is passed to the second printf() using a MOV
instruction. DWORD PTR—is a part of the assembly language (no relation to the machine code), indicating
that the variable data size is 32-bit and the MOV instruction has to be encoded accordingly.
ARM: Optimizing Keil 6/2013 (Thumb mode)
Listing 1.79: IDA

.text:00000000 ; Segment type: Pure code
.text:00000000 AREA .text, CODE
...
.text:00000000 main
.text:00000000 PUSH {R4,LR}
.text:00000002 ADR R0, aEnterX ; "Enter X:\n"
.text:00000004 BL __2printf
.text:00000008 LDR R1, =x
.text:0000000A ADR R0, aD ; "%d"
.text:0000000C BL __0scanf
.text:00000010 LDR R0, =x
.text:00000012 LDR R1, [R0]
.text:00000014 ADR R0, aYouEnteredD___ ; "You entered %d...\n"
.text:00000016 BL __2printf
.text:0000001A MOVS R0, #0
.text:0000001C POP {R4,PC}
...
.text:00000020 aEnterX DCB "Enter X:",0xA,0 ; DATA XREF: main+2
.text:0000002A DCB 0
.text:0000002B DCB 0
.text:0000002C off_2C DCD x ; DATA XREF: main+8
.text:0000002C ; main+10
.text:00000030 aD DCB "%d",0 ; DATA XREF: main+A
.text:00000033 DCB 0
.text:00000034 aYouEnteredD___ DCB "You entered %d...",0xA,0 ; DATA XREF: main+14
.text:00000047 DCB 0
.text:00000047 ; .text ends
.text:00000047
...
.data:00000048 ; Segment type: Pure data
.data:00000048 AREA .data, DATA
.data:00000048 ; ORG 0x48
.data:00000048 EXPORT x
.data:00000048 x DCD 0xA ; DATA XREF: main+8
.data:00000048 ; main+10
.data:00000048 ; .data ends
So, the x variable is now global and for this reason located in another segment, namely the data segment
(.data). One could ask, why are the text strings located in the code segment (.text) and x is located right
here? Because it is a variable and by definition its value could change. Moreover it could possibly change
80
often. While text strings has constant type, they will not be changed, so they are located in the .text
segment.
The code segment might sometimes be located in a ROM74 chip (keep in mind, we now deal with embedded
microelectronics, and memory scarcity is common here), and changeable variables —in RAM.
It is not very economical to store constant variables in RAM when you have ROM.
Furthermore, constant variables in RAM must be initialized, because after powering on, the RAM, obviously,
contains random information.
Moving forward, we see a pointer to the x (off_2C) variable in the code segment, and that all operations
with the variable occur via this pointer.
That is because the x variable could be located somewhere far from this particular code fragment, so its
address must be saved somewhere in close proximity to the code.
The LDR instruction in Thumb mode can only address variables in a range of 1020 bytes from its location,
and in ARM-mode —variables in range of ±4095 bytes.
And so the address of the x variable must be located somewhere in close proximity, because there is no
guarantee that the linker would be able to accommodate the variable somewhere nearby the code, it may
well be even in an external memory chip!
One more thing: if a variable is declared as const, the Keil compiler allocates it in the .constdata segment.
Perhaps thereafter, the linker could place this segment in ROM too, along with the code segment.
ARM64

1 .comm x,4,4
2 .LC0:
4 .LC1:
5 .string "%d"
6 .LC2:
8 f5:
9 ; save FP and LR in stack frame:
10 stp x29, x30, [sp, -16]!
12 add x29, sp, 0
14 adrp x0, .LC0
15 add x0, x0, :lo12:.LC0
16 bl puts
18 adrp x0, .LC1
19 add x0, x0, :lo12:.LC1
20 ; form address of x global variable:
21 adrp x1, x
22 add x1, x1, :lo12:x
24 ; form address of x global variable again:
25 adrp x0, x
26 add x0, x0, :lo12:x
27 ; load value from memory at this address:
28 ldr w1, [x0]
29 ; load pointer to the "You entered %d...\n" string:
30 adrp x0, .LC2
31 add x0, x0, :lo12:.LC2
32 bl printf
33 ; return 0
34 mov w0, 0
35 ; restore FP and LR:
36 ldp x29, x30, [sp], 16
37 ret
74 Read-Only Memory
81
In this case the x variable is declared as global and its address is calculated using the ADRP/ADD instruction
pair (lines 21 and 25).
MIPS
Uninitialized global variable
So now the x variable is global. Let’s compile to executable file rather than object file and load it into IDA.
IDA displays the x variable in the .sbss ELF section (remember the “Global Pointer”? 1.5.4 on page 24),
since the variable is not initialized at the start.

.text:004006C0 main:
.text:004006C0
.text:004006C0 var_10 = -0x10
.text:004006C0 var_4 = -4
.text:004006C0
.text:004006C0 lui $gp, 0x42
.text:004006C4 addiu $sp, -0x20
.text:004006C8 li $gp, 0x418940
.text:004006CC sw $ra, 0x20+var_4($sp)
.text:004006D0 sw $gp, 0x20+var_10($sp)
; call puts():
.text:004006D4 la $t9, puts
.text:004006D8 lui $a0, 0x40
.text:004006DC jalr $t9 ; puts
.text:004006E0 la $a0, aEnterX # "Enter X:" ; branch delay slot
; call scanf():
.text:004006E4 lw $gp, 0x20+var_10($sp)
.text:004006E8 lui $a0, 0x40
.text:004006EC la $t9, __isoc99_scanf
; prepare address of x:
.text:004006F0 la $a1, x
.text:004006F4 jalr $t9 ; __isoc99_scanf
.text:004006F8 la $a0, aD # "%d" ; branch delay slot
; call printf():
.text:004006FC lw $gp, 0x20+var_10($sp)
.text:00400700 lui $a0, 0x40
; get address of x:
.text:00400704 la $v0, x
.text:00400708 la $t9, printf
; load value from "x" variable and pass it to printf() in $a1:
.text:0040070C lw $a1, (x - 0x41099C)($v0)
.text:00400710 jalr $t9 ; printf
.text:00400714 la $a0, aYouEnteredD___ # "You entered %d...\n" ; branch
delay slot
.text:00400718 lw $ra, 0x20+var_4($sp)
.text:00400720 jr $ra
...
.sbss:0041099C # Segment type: Uninitialized

.sbss:0041099C .sbss
.sbss:0041099C .globl x
.sbss:0041099C x: .space 4
.sbss:0041099C
IDA reduces the amount of information, so we’ll also do a listing using objdump and comment it:
Listing 1.82: Optimizing GCC 4.4.5 (objdump)

1 004006c0 <main>:
2 ; function prologue:
3 4006c0: 3c1c0042 lui gp,0x42
4 4006c4: 27bdffe0 addiu sp,sp,-32
82
5 4006c8: 279c8940 addiu gp,gp,-30400
6 4006cc: afbf001c sw ra,28(sp)
7 4006d0: afbc0010 sw gp,16(sp)
8 ; call puts():
9 4006d4: 8f998034 lw t9,-32716(gp)
10 4006d8: 3c040040 lui a0,0x40
11 4006dc: 0320f809 jalr t9
12 4006e0: 248408f0 addiu a0,a0,2288 ; branch delay slot
13 ; call scanf():
14 4006e4: 8fbc0010 lw gp,16(sp)
15 4006e8: 3c040040 lui a0,0x40
16 4006ec: 8f998038 lw t9,-32712(gp)
17 ; prepare address of x:
18 4006f0: 8f858044 lw a1,-32700(gp)
19 4006f4: 0320f809 jalr t9
20 4006f8: 248408fc addiu a0,a0,2300 ; branch delay slot
21 ; call printf():
22 4006fc: 8fbc0010 lw gp,16(sp)
23 400700: 3c040040 lui a0,0x40
24 ; get address of x:
25 400704: 8f828044 lw v0,-32700(gp)
26 400708: 8f99803c lw t9,-32708(gp)
27 ; load value from "x" variable and pass it to printf() in $a1:
28 40070c: 8c450000 lw a1,0(v0)
29 400710: 0320f809 jalr t9
30 400714: 24840900 addiu a0,a0,2304 ; branch delay slot
32 400718: 8fbf001c lw ra,28(sp)
33 40071c: 00001021 move v0,zero
34 400720: 03e00008 jr ra
35 400724: 27bd0020 addiu sp,sp,32 ; branch delay slot
36 ; pack of NOPs used for aligning next function start on 16-byte boundary:
37 400728: 00200825 move at,at
38 40072c: 00200825 move at,at
Now we see the x variable address is read from a 64KiB data buffer using GP and adding negative offset to
it (line 18). More than that, the addresses of the three external functions which are used in our example
(puts(), scanf(), printf()), are also read from the 64KiB global data buffer using GP (lines 9, 16 and
26). GP points to the middle of the buffer, and such offset suggests that all three function’s addresses,
and also the address of the x variable, are all stored somewhere at the beginning of that buffer. That
make sense, because our example is tiny.
Another thing worth mentioning is that the function ends with two NOPs (MOVE $AT,$AT — an idle instruc-
tion), in order to align next function’s start on 16-byte boundary.
Initialized global variable
Let’s alter our example by giving the x variable a default value:

int x=10; // default value
Now IDA shows that the x variable is residing in the .data section:

.text:004006A0 main:
.text:004006A0
.text:004006A0 var_10 = -0x10
.text:004006A0 var_8 = -8
.text:004006A0 var_4 = -4
.text:004006A0
.text:004006A0 lui $gp, 0x42
.text:004006A4 addiu $sp, -0x20
.text:004006A8 li $gp, 0x418930
.text:004006AC sw $ra, 0x20+var_4($sp)
.text:004006B0 sw $s0, 0x20+var_8($sp)
.text:004006B4 sw $gp, 0x20+var_10($sp)
.text:004006B8 la $t9, puts
.text:004006BC lui $a0, 0x40
83
.text:004006C0 jalr $t9 ; puts
.text:004006C4 la $a0, aEnterX # "Enter X:"
.text:004006C8 lw $gp, 0x20+var_10($sp)
; prepare high part of x address:
.text:004006CC lui $s0, 0x41
.text:004006D0 la $t9, __isoc99_scanf
.text:004006D4 lui $a0, 0x40
; add low part of x address:
.text:004006D8 addiu $a1, $s0, (x - 0x410000)
; now address of x is in $a1.
.text:004006DC jalr $t9 ; __isoc99_scanf
.text:004006E0 la $a0, aD # "%d"
; get a word from memory:
.text:004006E8 lw $a1, x
; value of x is now in $a1.
.text:004006EC la $t9, printf
.text:004006F0 lui $a0, 0x40
.text:004006F4 jalr $t9 ; printf
.text:004006F8 la $a0, aYouEnteredD___ # "You entered %d...\n"
.text:004006FC lw $ra, 0x20+var_4($sp)
.text:00400704 lw $s0, 0x20+var_8($sp)
.text:00400708 jr $ra
...
.data:00410920 .globl x
.data:00410920 x: .word 0xA
Why not .sdata? Perhaps that this depends on some GCC option?
Nevertheless, now x is in .data, which is a general memory area, and we can take a look how to work with
variables there.
The variable’s address must be formed using a pair of instructions.
In our case those are LUI (“Load Upper Immediate”) and ADDIU (“Add Immediate Unsigned Word”).
Here is also the objdump listing for close inspection:

004006a0 <main>:
4006a0: 3c1c0042 lui gp,0x42
4006a4: 27bdffe0 addiu sp,sp,-32
4006a8: 279c8930 addiu gp,gp,-30416
4006ac: afbf001c sw ra,28(sp)
4006b0: afb00018 sw s0,24(sp)
4006b4: afbc0010 sw gp,16(sp)
4006b8: 8f998034 lw t9,-32716(gp)
4006bc: 3c040040 lui a0,0x40
4006c0: 0320f809 jalr t9
4006c4: 248408d0 addiu a0,a0,2256
4006c8: 8fbc0010 lw gp,16(sp)
; prepare high part of x address:
4006cc: 3c100041 lui s0,0x41
4006d0: 8f998038 lw t9,-32712(gp)
4006d4: 3c040040 lui a0,0x40
; add low part of x address:
4006d8: 26050920 addiu a1,s0,2336
; now address of x is in $a1.
4006dc: 0320f809 jalr t9
4006e0: 248408dc addiu a0,a0,2268
4006e4: 8fbc0010 lw gp,16(sp)
; high part of x address is still in $s0.
; add low part to it and load a word from memory:
4006e8: 8e050920 lw a1,2336(s0)
; value of x is now in $a1.
4006ec: 8f99803c lw t9,-32708(gp)
4006f0: 3c040040 lui a0,0x40
84
4006f4: 0320f809 jalr t9
4006f8: 248408e0 addiu a0,a0,2272
4006fc: 8fbf001c lw ra,28(sp)
400700: 00001021 move v0,zero
400704: 8fb00018 lw s0,24(sp)
400708: 03e00008 jr ra
40070c: 27bd0020 addiu sp,sp,32
We see that the address is formed using LUI and ADDIU, but the high part of address is still in the $S0
register, and it is possible to encode the offset in a LW (“Load Word”) instruction, so one single LW is enough
to load a value from the variable and pass it to printf().
Registers holding temporary data are prefixed with T-, but here we also see some prefixed with S-, the
contents of which must be preserved before use in other functions (i.e., saved somewhere).
That is why the value of $S0 has been set at address 0x4006cc and has been used again at address
0x4006e8, after the scanf() call. The scanf() function does not change its value.
1.12.4 scanf()
As was noted before, it is slightly old-fashioned to use scanf() today. But if we have to, we have to check
if scanf() finishes correctly without an error.
#include <stdio.h>
int main()
{
int x;
if (scanf ("%d", &x)==1)

else
printf ("What you entered? Huh?\n");
return 0;
};
By standard, the scanf()75 function returns the number of fields it has successfully read.
In our case, if everything goes fine and the user enters a number scanf() returns 1, or in case of error
(or EOF76 ) — 0.
Let’s add some C code to check the scanf() return value and print error message in case of an error.
This works as expected:
C:\...>ex3.exe
Enter X:
123
You entered 123...
C:\...>ex3.exe
Enter X:
ouch
What you entered? Huh?
MSVC: x86
Here is what we get in the assembly output (MSVC 2010):

lea eax, DWORD PTR _x$[ebp]
push eax
push OFFSET $SG3833 ; '%d', 00H
call _scanf
add esp, 8
cmp eax, 1
75 scanf, wscanf: MSDN
76 End of File
85
jne SHORT $LN2@main
mov ecx, DWORD PTR _x$[ebp]
push ecx
push OFFSET $SG3834 ; 'You entered %d...', 0aH, 00H
call _printf
add esp, 8
jmp SHORT $LN1@main
$LN2@main:
push OFFSET $SG3836 ; 'What you entered? Huh?', 0aH, 00H
call _printf
add esp, 4
$LN1@main:
xor eax, eax
The caller function (main()) needs the callee function (scanf()) result, so the callee returns it in the EAX
register.
We check it with the help of the instruction CMP EAX, 1 (CoMPare). In other words, we compare the value
in the EAX register with 1.
A JNE conditional jump follows the CMP instruction. JNE stands for Jump if Not Equal.
So, if the value in the EAX register is not equal to 1, the CPU will pass the execution to the address
mentioned in the JNE operand, in our case $LN2@main. Passing the control to this address results in
the CPU executing printf() with the argument What you entered? Huh?. But if everything is fine, the
conditional jump is not be taken, and another printf() call is to be executed, with two arguments:
'You entered %d...' and the value of x.
Since in this case the second printf() has not to be executed, there is a JMP preceding it (unconditional
jump). It passes the control to the point after the second printf() and just before the XOR EAX, EAX
instruction, which implements return 0.
So, it could be said that comparing a value with another is usually implemented by CMP/Jcc instruction
pair, where cc is condition code. CMP compares two values and sets processor flags77 . Jcc checks those
flags and decides to either pass the control to the specified address or not.
This could sound paradoxical, but the CMP instruction is in fact SUB (subtract). All arithmetic instructions
set processor flags, not just CMP. If we compare 1 and 1, 1 − 1 is 0 so the ZF flag would be set (meaning that
the last result is 0). In no other circumstances ZF can be set, except when the operands are equal. JNE
checks only the ZF flag and jumps only if it is not set. JNE is in fact a synonym for JNZ (Jump if Not Zero).
Assembler translates both JNE and JNZ instructions into the same opcode. So, the CMP instruction can be
replaced with a SUB instruction and almost everything will be fine, with the difference that SUB alters the
value of the first operand. CMP is SUB without saving the result, but affecting flags.
MSVC: x86: IDA
It is time to run IDA and try to do something in it. By the way, for beginners it is good idea to use /MD
option in MSVC, which means that all these standard functions are not be linked with the executable file,
but are to be imported from the MSVCR*.DLL file instead. Thus it will be easier to see which standard
function are used and where.
While analyzing code in IDA, it is very helpful to leave notes for oneself (and others). In instance, analyzing
this example, we see that JNZ is to be triggered in case of an error. So it is possible to move the cursor to
the label, press “n” and rename it to “error”. Create another label—into “exit”. Here is my result:
.text:00401000 _main proc near
.text:00401000
.text:00401000 var_4 = dword ptr -4
.text:00401000 argc = dword ptr 8
.text:00401000 argv = dword ptr 0Ch
.text:00401000 envp = dword ptr 10h
.text:00401000
.text:00401000 push ebp
.text:00401001 mov ebp, esp
.text:00401003 push ecx
.text:00401004 push offset Format ; "Enter X:\n"
.text:00401009 call ds:printf
.text:0040100F add esp, 4
77 x86 flags, see also: wikipedia.
86
.text:00401012 lea eax, [ebp+var_4]
.text:00401015 push eax
.text:00401016 push offset aD ; "%d"
.text:0040101B call ds:scanf
.text:00401021 add esp, 8
.text:00401024 cmp eax, 1
.text:00401027 jnz short error
.text:00401029 mov ecx, [ebp+var_4]
.text:0040102C push ecx
.text:0040102D push offset aYou ; "You entered %d...\n"
.text:0040103B jmp short exit
.text:0040103D
.text:0040103D error: ; CODE XREF: _main+27
.text:0040103D push offset aWhat ; "What you entered? Huh?\n"
.text:0040104B
.text:0040104B exit: ; CODE XREF: _main+3B
.text:0040104B xor eax, eax
.text:0040104D mov esp, ebp
.text:0040104F pop ebp
.text:00401050 retn
.text:00401050 _main endp
Now it is slightly easier to understand the code. However, it is not a good idea to comment on every
instruction.
You could also hide(collapse) parts of a function in IDA. To do that mark the block, then press “–” on the
numerical pad and enter the text to be displayed instead.
Let’s hide two blocks and give them names:
.text:00401000 _text segment para public 'CODE' use32
.text:00401000 assume cs:_text
.text:00401000 ;org 401000h
.text:00401000 ; ask for X
.text:00401012 ; get X
.text:00401024 cmp eax, 1
.text:00401027 jnz short error
.text:00401029 ; print result
.text:0040103B jmp short exit
.text:0040103D
.text:0040103D error: ; CODE XREF: _main+27
.text:0040103D push offset aWhat ; "What you entered? Huh?\n"
.text:0040104B
.text:0040104B exit: ; CODE XREF: _main+3B
.text:0040104B xor eax, eax
.text:0040104D mov esp, ebp
.text:0040104F pop ebp
.text:00401050 retn
.text:00401050 _main endp
To expand previously collapsed parts of the code, use “+” on the numerical pad.
87
By pressing “space”, we can see how IDA represents a function as a graph:
Figure 1.18: Graph mode in IDA
There are two arrows after each conditional jump: green and red. The green arrow points to the block
which executes if the jump is triggered, and red if otherwise.
88
It is possible to fold nodes in this mode and give them names as well (“group nodes”). Let’s do it for 3
blocks:
Figure 1.19: Graph mode in IDA with 3 nodes folded
That is very useful. It could be said that a very important part of the reverse engineers’ job (and any other
researcher as well) is to reduce the amount of information they deal with.
89
MSVC: x86 + OllyDbg
Let’s try to hack our program in OllyDbg, forcing it to think scanf() always works without error. When an
address of a local variable is passed into scanf(), the variable initially contains some random garbage,
in this case 0x6E494714:
Figure 1.20: OllyDbg: passing variable address into scanf()
90
While scanf() executes, in the console we enter something that is definitely not a number, like “asdasd”.
scanf() finishes with 0 in EAX, which indicates that an error has occurred:
Figure 1.21: OllyDbg: scanf() returning error
We can also check the local variable in the stack and note that it has not changed. Indeed, what would
scanf() write there? It simply did nothing except returning zero.
Let’s try to “hack” our program. Right-click on EAX, Among the options there is “Set to 1”. This is what
we need.
We now have 1 in EAX, so the following check is to be executed as intended, and printf() will print the
value of the variable in the stack.
When we run the program (F9) we can see the following in the console window:
Listing 1.85: console window

Enter X:
asdasd
You entered 1850296084...
Indeed, 1850296084 is a decimal representation of the number in the stack (0x6E494714)!
91
MSVC: x86 + Hiew
This can also be used as a simple example of executable file patching. We may try to patch the executable
so the program would always print the input, no matter what we enter.
Assuming that the executable is compiled against external MSVCR*.DLL (i.e., with /MD option) 78 , we see
the main() function at the beginning of the .text section. Let’s open the executable in Hiew and find the
beginning of the .text section (Enter, F8, F6, Enter, Enter).
We can see this:
Figure 1.22: Hiew: main() function
Hiew finds ASCIIZ79 strings and displays them, as it does with the imported functions’ names.
78 that’s what also called “dynamic linking”

79 ASCII Zero (null-terminated ASCII string )
92
Move the cursor to address .00401027 (where the JNZ instruction, we have to bypass, is located), press
F3, and then type “9090” (meaning two NOPs):
Figure 1.23: Hiew: replacing JNZ with two NOPs
Then press F9 (update). Now the executable is saved to the disk. It will behave as we wanted.
Two NOPs are probably not the most æsthetic approach. Another way to patch this instruction is to write
just 0 to the second opcode byte (jump offset), so that JNZ will always jump to the next instruction.
We could also do the opposite: replace first byte with EB while not touching the second byte (jump offset).
We would get an unconditional jump that is always triggered. In this case the error message would be
printed every time, no matter the input.
MSVC: x64
Since we work here with int-typed variables, which are still 32-bit in x86-64, we see how the 32-bit part
of the registers (prefixed with E-) are used here as well. While working with pointers, however, 64-bit
register parts are used, prefixed with R-.

_DATA SEGMENT
$SG2926 DB '%d', 00H
93
$SG2929 DB 'What you entered? Huh?', 0aH, 00H
_DATA ENDS
_TEXT SEGMENT
x$ = 32
main PROC
$LN5:
sub rsp, 56
call printf
lea rdx, QWORD PTR x$[rsp]
call scanf
cmp eax, 1
jne SHORT $LN2@main
mov edx, DWORD PTR x$[rsp]
call printf
jmp SHORT $LN1@main
$LN2@main:
lea rcx, OFFSET FLAT:$SG2929 ; 'What you entered? Huh?'
call printf
$LN1@main:
; return 0
xor eax, eax
add rsp, 56
ret 0
main ENDP
_TEXT ENDS
END
ARM

var_8 = -8
PUSH {R3,LR}
ADR R0, aEnterX ; "Enter X:\n"
BL __2printf
MOV R1, SP
ADR R0, aD ; "%d"
BL __0scanf
CMP R0, #1
BEQ loc_1E
ADR R0, aWhatYouEntered ; "What you entered? Huh?\n"
BL __2printf
loc_1A ; CODE XREF: main+26

MOVS R0, #0
POP {R3,PC}
loc_1E ; CODE XREF: main+12

LDR R1, [SP,#8+var_8]
ADR R0, aYouEnteredD___ ; "You entered %d...\n"
BL __2printf
B loc_1A
The new instructions here are CMP and BEQ80 .

CMP is analogous to the x86 instruction with the same name, it subtracts one of the arguments from the
other and updates the conditional flags if needed.
80 (PowerPC, ARM) Branch if Equal
94
BEQ jumps to another address if the operands were equal to each other, or, if the result of the last
computation has been 0, or if the Z flag is 1. It behaves as JZ in x86.
Everything else is simple: the execution flow forks in two branches, then the branches converge at the
point where 0 is written into the R0 as a function return value, and then the function ends.
ARM64

1 .LC0:
3 .LC1:
4 .string "%d"
5 .LC2:
7 .LC3:
8 .string "What you entered? Huh?"
9 f6:
10 ; save FP and LR in stack frame:
11 stp x29, x30, [sp, -32]!
13 add x29, sp, 0
15 adrp x0, .LC0
16 add x0, x0, :lo12:.LC0
17 bl puts
19 adrp x0, .LC1
20 add x0, x0, :lo12:.LC1
21 ; calculate address of x variable in the local stack
22 add x1, x29, 28
24 ; scanf() returned result in W0.
25 ; check it:
26 cmp w0, 1
27 ; BNE is Branch if Not Equal
28 ; so if W0<>0, jump to L2 will be occurred
29 bne .L2
30 ; at this moment W0=1, meaning no error
31 ; load x value from the local stack
32 ldr w1, [x29,28]
33 ; load pointer to the "You entered %d...\n" string:
34 adrp x0, .LC2
35 add x0, x0, :lo12:.LC2
36 bl printf
37 ; skip the code, which print the "What you entered? Huh?" string:
38 b .L3
39 .L2:
40 ; load pointer to the "What you entered? Huh?" string:
41 adrp x0, .LC3
42 add x0, x0, :lo12:.LC3
43 bl puts
44 .L3:
45 ; return 0
46 mov w0, 0
47 ; restore FP and LR:
48 ldp x29, x30, [sp], 32
49 ret
Code flow in this case forks with the use of CMP/BNE (Branch if Not Equal) instructions pair.
MIPS

.text:004006A0 main:
.text:004006A0
.text:004006A0 var_18 = -0x18
95
.text:004006A0 var_10 = -0x10
.text:004006A0 var_4 = -4
.text:004006A0
.text:004006A0 lui $gp, 0x42
.text:004006A4 addiu $sp, -0x28
.text:004006A8 li $gp, 0x418960
.text:004006AC sw $ra, 0x28+var_4($sp)
.text:004006B0 sw $gp, 0x28+var_18($sp)
.text:004006B4 la $t9, puts
.text:004006B8 lui $a0, 0x40
.text:004006BC jalr $t9 ; puts
.text:004006C0 la $a0, aEnterX # "Enter X:"
.text:004006C4 lw $gp, 0x28+var_18($sp)
.text:004006C8 lui $a0, 0x40
.text:004006CC la $t9, __isoc99_scanf
.text:004006D0 la $a0, aD # "%d"
.text:004006D4 jalr $t9 ; __isoc99_scanf
.text:004006D8 addiu $a1, $sp, 0x28+var_10 # branch delay slot
.text:004006DC li $v1, 1
.text:004006E4 beq $v0, $v1, loc_40070C
.text:004006E8 or $at, $zero # branch delay slot, NOP
.text:004006EC la $t9, puts
.text:004006F0 lui $a0, 0x40
.text:004006F4 jalr $t9 ; puts
.text:004006F8 la $a0, aWhatYouEntered # "What you entered? Huh?"
.text:004006FC lw $ra, 0x28+var_4($sp)
.text:00400704 jr $ra
.text:00400708 addiu $sp, 0x28
.text:0040070C loc_40070C:
.text:0040070C la $t9, printf
.text:00400710 lw $a1, 0x28+var_10($sp)
.text:00400714 lui $a0, 0x40
.text:00400718 jalr $t9 ; printf
.text:0040071C la $a0, aYouEnteredD___ # "You entered %d...\n"
.text:00400720 lw $ra, 0x28+var_4($sp)
.text:00400728 jr $ra
scanf() returns the result of its work in register $V0. It is checked at address 0x004006E4 by comparing
the values in $V0 with $V1 (1 has been stored in $V1 earlier, at 0x004006DC). BEQ stands for “Branch
Equal”. If the two values are equal (i.e., success), the execution jumps to address 0x0040070C.
Exercise
As we can see, the JNE/JNZ instruction can be easily replaced by the JE/JZ and vice versa (or BNE by BEQ
and vice versa). But then the basic blocks must also be swapped. Try to do this in some of the examples.
1.12.5 Exercise
1.13 Worth noting: global vs. local variables

Now that you know that global variables are filling with zeroes by OS at start ( 1.12.3 on page 77, [ISO/IEC
9899:TC3 (C C99 standard), (2007)6.7.8p10]), but local variables are not ( 1.9.4 on page 36).
Sometimes, you have a global variable that you forgot to initialize and your program relies on the fact
that it has zero at start. Then you edit a program and move the global variable into a function to make it
local. It wouldn’t be zeroed at initialization anymore and this can result in nasty bugs.
96
1.14 Accessing passed arguments
Now we figured out that the caller function is passing arguments to the callee via the stack. But how does
the callee access them?
Listing 1.90: simple example

#include <stdio.h>
int f (int a, int b, int c)

{
return a*b+c;
};
int main()
{
printf ("%d\n", f(1, 2, 3));
return 0;
};
1.14.1 x86
MSVC
Here is what we get after compilation (MSVC 2010 Express):
Listing 1.91: MSVC 2010 Express

_TEXT SEGMENT
_a$ = 8 ; size = 4
_b$ = 12 ; size = 4
_c$ = 16 ; size = 4
_f PROC
push ebp
mov ebp, esp
mov eax, DWORD PTR _a$[ebp]
imul eax, DWORD PTR _b$[ebp]
add eax, DWORD PTR _c$[ebp]
pop ebp
ret 0
_f ENDP
_main PROC
push ebp
mov ebp, esp
push 3 ; 3rd argument
push 2 ; 2nd argument
push 1 ; 1st argument
call _f
add esp, 12
push eax
push OFFSET $SG2463 ; '%d', 0aH, 00H
call _printf
add esp, 8
; return 0
xor eax, eax
pop ebp
ret 0
_main ENDP
What we see is that the main() function pushes 3 numbers onto the stack and calls f(int,int,int).
Argument access inside f() is organized with the help of macros like:
_a$ = 8, in the same way as local variables, but with positive offsets (addressed with plus). So, we are
addressing the outer side of the stack frame by adding the _a$ macro to the value in the EBP register.
Then the value of a is stored into EAX. After IMUL instruction execution, the value in EAX is a product of the
value in EAX and the content of _b.
After that, ADD adds the value in _c to EAX.
97
The value in EAX does not need to be moved: it is already where it must be. On returning to caller, it takes
the EAX value and uses it as an argument to printf().
MSVC + OllyDbg
Let’s illustrate this in OllyDbg. When we trace to the first instruction in f() that uses one of the arguments
(first one), we see that EBP is pointing to the stack frame, which is marked with a red rectangle.
The first element of the stack frame is the saved value of EBP, the second one is RA, the third is the first
function argument, then the second and third ones.
To access the first function argument, one needs to add exactly 8 (2 32-bit words) to EBP.
OllyDbg is aware about this, so it has added comments to the stack elements like
“RETURN from” and “Arg1 = …”, etc.
N.B.: Function arguments are not members of the function’s stack frame, they are rather members of the
stack frame of the caller function.
Hence, OllyDbg marked “Arg” elements as members of another stack frame.
Figure 1.24: OllyDbg: inside of f() function
GCC
Let’s compile the same in GCC 4.4.1 and see the results in IDA:

public f
f proc near
arg_0 = dword ptr 8

arg_4 = dword ptr 0Ch
arg_8 = dword ptr 10h
push ebp
mov ebp, esp
mov eax, [ebp+arg_0] ; 1st argument
imul eax, [ebp+arg_4] ; 2nd argument
add eax, [ebp+arg_8] ; 3rd argument
pop ebp
retn
f endp
public main
98
main proc near

push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
sub esp, 10h
mov [esp+10h+var_8], 3 ; 3rd argument
mov [esp+10h+var_C], 2 ; 2nd argument
mov [esp+10h+var_10], 1 ; 1st argument
call f
mov edx, offset aD ; "%d\n"
mov [esp+10h+var_C], eax
mov [esp+10h+var_10], edx
call _printf
mov eax, 0
leave
retn
main endp
The result is almost the same with some minor differences discussed earlier.
The stack pointer is not set back after the two function calls(f and printf), because the penultimate LEAVE
( .1.6 on page 1014) instruction takes care of this at the end.
1.14.2 x64
The story is a bit different in x86-64. Function arguments (first 4 or first 6 of them) are passed in registers
i.e. the callee reads them from registers instead of reading them from the stack.
MSVC
Optimizing MSVC:
Listing 1.93: Optimizing MSVC 2012 x64

$SG2997 DB '%d', 0aH, 00H
main PROC
sub rsp, 40
mov edx, 2
lea r8d, QWORD PTR [rdx+1] ; R8D=3
lea ecx, QWORD PTR [rdx-1] ; ECX=1
call f
mov edx, eax
call printf
xor eax, eax
add rsp, 40
ret 0
main ENDP
f PROC
; ECX - 1st argument
; EDX - 2nd argument
; R8D - 3rd argument
imul ecx, edx
lea eax, DWORD PTR [r8+rcx]
ret 0
f ENDP
As we can see, the compact function f() takes all its arguments from the registers.
The LEA instruction here is used for addition, apparently the compiler considered it faster than ADD.
99
LEA is also used in the main() function to prepare the first and third f() arguments. The compiler must
have decided that this would work faster than the usual way of loading values into a register using MOV
instruction.
Let’s take a look at the non-optimizing MSVC output:

f proc near
; shadow space:
arg_0 = dword ptr 8
; ECX - 1st argument

; EDX - 2nd argument
; R8D - 3rd argument
mov [rsp+arg_10], r8d
mov [rsp+arg_8], edx
mov [rsp+arg_0], ecx
mov eax, [rsp+arg_0]
imul eax, [rsp+arg_8]
add eax, [rsp+arg_10]
retn
f endp
main proc near

sub rsp, 28h
mov r8d, 3 ; 3rd argument
mov edx, 2 ; 2nd argument
mov ecx, 1 ; 1st argument
call f
mov edx, eax
lea rcx, $SG2931 ; "%d\n"
call printf
; return 0
xor eax, eax
add rsp, 28h
retn
main endp
It looks somewhat puzzling because all 3 arguments from the registers are saved to the stack for some
reason. This is called “shadow space” 81 : every Win64 may (but is not required to) save all 4 register
values there. This is done for two reasons: 1) it is too lavish to allocate a whole register (or even 4
registers) for an input argument, so it will be accessed via stack; 2) the debugger is always aware where
to find the function arguments at a break 82 .
So, some large functions can save their input arguments in the “shadows space” if they want to use them
during execution, but some small functions (like ours) may not do this.
It is a caller responsibility to allocate “shadow space” in the stack.
GCC
Optimizing GCC generates more or less understandable code:

f:
; EDI - 1st argument
; ESI - 2nd argument
; EDX - 3rd argument
imul esi, edi
lea eax, [rdx+rsi]
ret
81 MSDN
82 MSDN
100
main:
sub rsp, 8
mov edx, 3
mov esi, 2
mov edi, 1
call f
mov edi, OFFSET FLAT:.LC0 ; "%d\n"
mov esi, eax
call printf
xor eax, eax
add rsp, 8
ret
Non-optimizing GCC:

f:
; EDI - 1st argument
; ESI - 2nd argument
; EDX - 3rd argument
push rbp
mov rbp, rsp
mov DWORD PTR [rbp-4], edi
mov DWORD PTR [rbp-8], esi
mov DWORD PTR [rbp-12], edx
mov eax, DWORD PTR [rbp-4]
imul eax, DWORD PTR [rbp-8]
add eax, DWORD PTR [rbp-12]
leave
ret
main:
push rbp
mov rbp, rsp
mov edx, 3
mov esi, 2
mov edi, 1
call f
mov edx, eax
mov eax, OFFSET FLAT:.LC0 ; "%d\n"
mov esi, edx
mov rdi, rax
mov eax, 0 ; number of vector registers passed
call printf
mov eax, 0
leave
ret
There are no “shadow space” requirements in System V *NIX ([Michael Matz, Jan Hubicka, Andreas Jaeger,
Mark Mitchell, System V Application Binary Interface. AMD64 Architecture Processor Supplement, (2013)]
83
), but the callee may want to save its arguments somewhere in case of registers shortage.
GCC: uint64_t instead of int
Our example works with 32-bit int, that is why 32-bit register parts are used (prefixed by E-).
It can be altered slightly in order to use 64-bit values:
#include <stdio.h>
#include <stdint.h>
uint64_t f (uint64_t a, uint64_t b, uint64_t c)

{
return a*b+c;
};
int main()
101
{
printf ("%lld\n", f(0x1122334455667788,
0x1111111122222222,
0x3333333344444444));
return 0;
};

f proc near
imul rsi, rdi
lea rax, [rdx+rsi]
retn
f endp
main proc near

sub rsp, 8
mov rdx, 3333333344444444h ; 3rd argument
mov rsi, 1111111122222222h ; 2nd argument
mov rdi, 1122334455667788h ; 1st argument
call f
mov edi, offset format ; "%lld\n"
mov rsi, rax
call _printf
xor eax, eax
add rsp, 8
retn
main endp
The code is the same, but this time the full size registers (prefixed by R-) are used.
1.14.3 ARM
.text:000000A4 00 30 A0 E1 MOV R3, R0

.text:000000A8 93 21 20 E0 MLA R0, R3, R1, R2
.text:000000AC 1E FF 2F E1 BX LR
...
.text:000000B0 main
.text:000000B0 10 40 2D E9 STMFD SP!, {R4,LR}
.text:000000B4 03 20 A0 E3 MOV R2, #3
.text:000000B8 02 10 A0 E3 MOV R1, #2
.text:000000BC 01 00 A0 E3 MOV R0, #1
.text:000000C0 F7 FF FF EB BL f
.text:000000C4 00 40 A0 E1 MOV R4, R0
.text:000000C8 04 10 A0 E1 MOV R1, R4
.text:000000CC 5A 0F 8F E2 ADR R0, aD_0 ; "%d\n"
.text:000000D0 E3 18 00 EB BL __2printf
.text:000000D4 00 00 A0 E3 MOV R0, #0
.text:000000D8 10 80 BD E8 LDMFD SP!, {R4,PC}
The main() function simply calls two other functions, with three values passed to the first one —(f()).
As was noted before, in ARM the first 4 values are usually passed in the first 4 registers (R0-R3).
The f() function, as it seems, uses the first 3 registers (R0-R2) as arguments.
The MLA (Multiply Accumulate) instruction multiplies its first two operands (R3 and R1), adds the third
operand (R2) to the product and stores the result into the zeroth register (R0), via which, by standard,
functions return values.
Multiplication and addition at once (Fused multiply–add) is a very useful operation. By the way, there was
no such instruction in x86 before FMA-instructions appeared in SIMD 84 .
The very first MOV R3, R0, instruction is, apparently, redundant (a single MLA instruction could be used
here instead). The compiler has not optimized it, since this is non-optimizing compilation.
84 wikipedia
102
The BX instruction returns the control to the address stored in the LR register and, if necessary, switches
the processor mode from Thumb to ARM or vice versa. This can be necessary since, as we can see,
function f() is not aware from what kind of code it may be called, ARM or Thumb. Thus, if it gets called
from Thumb code, BX is not only returns control to the calling function, but also switches the processor
mode to Thumb. Or not switch, if the function has been called from ARM code [ARM(R) Architecture
Reference Manual, ARMv7-A and ARMv7-R edition, (2012)A2.3.2].
Optimizing Keil 6/2013 (ARM mode)
.text:00000098 f
.text:00000098 91 20 20 E0 MLA R0, R1, R0, R2
.text:0000009C 1E FF 2F E1 BX LR
And here is the f() function compiled by the Keil compiler in full optimization mode (-O3).
The MOV instruction was optimized out (or reduced) and now MLA uses all input registers and also places
the result right into R0, exactly where the calling function will read and use it.
.text:0000005E 48 43 MULS R0, R1

.text:00000060 80 18 ADDS R0, R0, R2
.text:00000062 70 47 BX LR
The MLA instruction is not available in Thumb mode, so the compiler generates the code doing these two
operations (multiplication and addition) separately.
First the MULS instruction multiplies R0 by R1, leaving the result in register R0. The second instruction
(ADDS) adds the result and R2 leaving the result in register R0.
ARM64
Optimizing GCC (Linaro) 4.9
Everything here is simple. MADD is just an instruction doing fused multiply/add (similar to the MLA we
already saw). All 3 arguments are passed in the 32-bit parts of X-registers. Indeed, the argument types
are 32-bit int’s. The result is returned in W0.
Listing 1.98: Optimizing GCC (Linaro) 4.9

f:
madd w0, w0, w1, w2
ret
main:
; save FP and LR to stack frame:
stp x29, x30, [sp, -16]!
mov w2, 3
mov w1, 2
add x29, sp, 0
mov w0, 1
bl f
mov w1, w0
adrp x0, .LC7
bl printf
; return 0
mov w0, 0
; restore FP and LR
ldp x29, x30, [sp], 16
ret
.LC7:
.string "%d\n"
Let’s also extend all data types to 64-bit uint64_t and test:
103
#include <stdio.h>
#include <stdint.h>
uint64_t f (uint64_t a, uint64_t b, uint64_t c)

{
return a*b+c;
};
int main()
{
printf ("%lld\n", f(0x1122334455667788,
0x1111111122222222,
0x3333333344444444));
return 0;
};
f:
madd x0, x0, x1, x2
ret
main:
mov x1, 13396
adrp x0, .LC8
stp x29, x30, [sp, -16]!
movk x1, 0x27d0, lsl 16
movk x1, 0x122, lsl 32
add x29, sp, 0
movk x1, 0x58be, lsl 48
bl printf
mov w0, 0
ldp x29, x30, [sp], 16
ret
.LC8:
.string "%lld\n"
The f() function is the same, only the whole 64-bit X-registers are now used. Long 64-bit values are
loaded into the registers by parts, this is also described here: 1.39.3 on page 438.
The non-optimizing compiler is more redundant:

f:
sub sp, sp, #16
str w0, [sp,12]
str w1, [sp,8]
str w2, [sp,4]
ldr w1, [sp,12]
ldr w0, [sp,8]
mul w1, w1, w0
ldr w0, [sp,4]
add w0, w1, w0
add sp, sp, 16
ret
The code saves its input arguments in the local stack, in case someone (or something) in this function
needs using the W0...W2 registers. This prevents overwriting the original function arguments, which may
be needed again in the future.
This is called Register Save Area. ([Procedure Call Standard for the ARM 64-bit Architecture (AArch64),
(2013)]85 ). The callee, however, is not obliged to save them. This is somewhat similar to “Shadow
Space”: 1.14.2 on page 100.
104
Why did the optimizing GCC 4.9 drop this argument saving code? Because it did some additional optimizing
work and concluded that the function arguments will not be needed in the future and also that the registers
W0...W2 will not be used.
We also see a MUL/ADD instruction pair instead of single a MADD.
1.14.4 MIPS
Listing 1.99: Optimizing GCC 4.4.5
.text:00000000 f:
; $a0=a
; $a1=b
; $a2=c
.text:00000000 mult $a1, $a0
.text:00000004 mflo $v0
.text:00000008 jr $ra
.text:0000000C addu $v0, $a2, $v0 ; branch delay slot
; result is in $v0 upon return
.text:00000010
.text:00000010 var_10 = -0x10
.text:00000010 var_4 = -4
.text:00000010
.text:00000014 addiu $sp, -0x20
.text:00000020 sw $gp, 0x20+var_10($sp)
; set c:
.text:00000024 li $a2, 3
; set a:
.text:00000028 li $a0, 1
.text:0000002C jal f
; set b:
.text:00000030 li $a1, 2 ; branch delay slot
; result in $v0 now
.text:00000034 lw $gp, 0x20+var_10($sp)
.text:00000038 lui $a0, ($LC0 >> 16)
.text:0000003C lw $t9, (printf & 0xFFFF)($gp)
.text:00000040 la $a0, ($LC0 & 0xFFFF)
.text:00000044 jalr $t9
; take result of f() function and pass it as a second argument to printf():
.text:00000048 move $a1, $v0 ; branch delay slot
.text:0000004C lw $ra, 0x20+var_4($sp)
.text:00000054 jr $ra
The first four function arguments are passed in four registers prefixed by A-.
There are two special registers in MIPS: HI and LO which are filled with the 64-bit result of the multiplication
during the execution of the MULT instruction.
These registers are accessible only by using the MFLO and MFHI instructions. MFLO here takes the low-part
of the multiplication result and stores it into $V0. So the high 32-bit part of the multiplication result is
dropped (the HI register content is not used). Indeed: we work with 32-bit int data types here.
Finally, ADDU (“Add Unsigned”) adds the value of the third argument to the result.
There are two different addition instructions in MIPS: ADD and ADDU. The difference between them is not
related to signedness, but to exceptions. ADD can raise an exception on overflow, which is sometimes
useful86 and supported in Ada PL, for instance. ADDU does not raise exceptions on overflow.
Since C/C++ does not support this, in our example we see ADDU instead of ADD.
The 32-bit result is left in $V0.
There is a new instruction for us in main(): JAL (“Jump and Link”).
86 http://go.yurichev.com/17326
105
The difference between JAL and JALR is that a relative offset is encoded in the first instruction, while JALR
jumps to the absolute address stored in a register (“Jump and Link Register”).
Both f() and main() functions are located in the same object file, so the relative address of f() is known
and fixed.
1.15 More about results returning

In x86, the result of function execution is usually returned 87 in the EAX register. If it is byte type or a
character (char), then the lowest part of register EAX (AL) is used. If a function returns a float number, the
FPU register ST(0) is used instead. In ARM, the result is usually returned in the R0 register.
1.15.1 Attempt to use the result of a function returning void

So, what if the main() function return value was declared of type void and not int? The so-called startup-
code is calling main() roughly as follows:
push envp
push argv
push argc
call main
push eax
call exit
In other words:
exit(main(argc,argv,envp));
If you declare main() as void, nothing is to be returned explicitly (using the return statement), then
something random, that has been stored in the EAX register at the end of main() becomes the sole
argument of the exit() function. Most likely, there will be a random value, left from your function execution,
so the exit code of program is pseudorandom.
We can illustrate this fact. Please note that here the main() function has a void return type:
#include <stdio.h>
void main()
{
printf ("Hello, world!\n");
};
Let’s compile it in Linux.

GCC 4.8.1 replaced printf() with puts() (we have seen this before: 1.5.3 on page 20), but that’s OK,
since puts() returns the number of characters printed out, just like printf(). Please notice that EAX is
not zeroed before main()’s end.
This implies that the value of EAX at the end of main() contains what puts() has left there.

.LC0:
.string "Hello, world!"
main:
push ebp
mov ebp, esp
and esp, -16
sub esp, 16
mov DWORD PTR [esp], OFFSET FLAT:.LC0
call puts
leave
ret
Let’ s write a bash script that shows the exit status:

87 See also: MSDN: Return Values (C++): MSDN
106
Listing 1.101: tst.sh
#!/bin/sh
./hello_world
echo $?
And run it:

$ tst.sh
Hello, world!
14
14 is the number of characters printed. The number of characters printed slips from printf() through
EAX/RAX into “exit code”.
Another example in the book: 3.32 on page 638.
By the way, when we decompile C++ in Hex-Rays, we can often encounter a function which terminated
with destructor of some class:
...
call ??1CString@@QAE@XZ ; CString:: CString(void)

mov ecx, [esp+30h+var_C]
pop edi
pop ebx
mov large fs:0, ecx
add esp, 28h
retn
By C++ standard, destructor doesn’t return anything, but when Hex-Rays don’t know about it, and thinks
that both destructor and this function returns int, we can see something like that in output:
...
return CString::~CString(&Str);
}
1.15.2 What if we do not use the function result?

printf() returns the count of characters successfully output, but the result of this function is rarely used
in practice.
It is also possible to call a function whose essence is in returning a value, and not use it:
int f()
{
// skip first 3 random values:
rand();
rand();
rand();
// and use 4th:
return rand();
};
The result of the rand() function is left in EAX, in all four cases.
But in the first 3 cases, the value in EAX is just not used.
1.15.3 Returning a structure

Let’s go back to the fact that the return value is left in the EAX register.
That is why old C compilers cannot create functions capable of returning something that does not fit in one
register (usually int), but if one needs it, one have to return information via pointers passed as function’s
arguments.
So, usually, if a function needs to return several values, it returns only one, and all the rest—via pointers.
107
Now it has become possible to return, let’s say, an entire structure, but that is still not very popular. If
a function has to return a large structure, the caller must allocate it and pass a pointer to it via the first
argument, transparently for the programmer. That is almost the same as to pass a pointer in the first
argument manually, but the compiler hides it.
Small example:
struct s
{
int a;
int b;
int c;
};
struct s get_some_values (int a)

{
struct s rt;
rt.a=a+1;
rt.b=a+2;
rt.c=a+3;
return rt;
};
…what we got (MSVC 2010 /Ox):

$T3853 = 8 ; size = 4
_a$ = 12 ; size = 4
?get_some_values@@YA?AUs@@H@Z PROC ; get_some_values
mov ecx, DWORD PTR _a$[esp-4]
mov eax, DWORD PTR $T3853[esp-4]
lea edx, DWORD PTR [ecx+1]
mov DWORD PTR [eax], edx
add ecx, 3
mov DWORD PTR [eax+4], edx
mov DWORD PTR [eax+8], ecx
ret 0
?get_some_values@@YA?AUs@@H@Z ENDP ; get_some_values
The macro name for internal passing of pointer to a structure here is $T3853.
This example can be rewritten using the C99 language extensions:
struct s
{
int a;
int b;
int c;
};
struct s get_some_values (int a)

{
return (struct s){.a=a+1, .b=a+2, .c=a+3};
};

_get_some_values proc near
ptr_to_struct = dword ptr 4

a = dword ptr 8
mov edx, [esp+a]

mov eax, [esp+ptr_to_struct]
lea ecx, [edx+1]
mov [eax], ecx
lea ecx, [edx+2]
add edx, 3
mov [eax+4], ecx
108
mov [eax+8], edx
retn
_get_some_values endp
As we see, the function is just filling the structure’s fields allocated by the caller function, as if a pointer
to the structure has been passed. So there are no performance drawbacks.
1.16 Pointers
1.16.1 Returning values
Pointers are often used to return values from functions (recall scanf() case ( 1.12 on page 65)).
For example, when a function needs to return two values.
Global variables example
#include <stdio.h>
void f1 (int x, int y, int *sum, int *product)

{
*sum=x+y;
*product=x*y;
};
int sum, product;
void main()
{
f1(123, 456, &sum, &product);
printf ("sum=%d, product=%d\n", sum, product);
};
This compiles to:
Listing 1.103: Optimizing MSVC 2010 (/Ob0)

COMM _product:DWORD
COMM _sum:DWORD
$SG2803 DB 'sum=%d, product=%d', 0aH, 00H
_x$ = 8 ; size = 4
_y$ = 12 ; size = 4
_sum$ = 16 ; size = 4
_product$ = 20 ; size = 4
_f1 PROC
mov ecx, DWORD PTR _y$[esp-4]
mov eax, DWORD PTR _x$[esp-4]
lea edx, DWORD PTR [eax+ecx]
imul eax, ecx
mov ecx, DWORD PTR _product$[esp-4]
push esi
mov esi, DWORD PTR _sum$[esp]
mov DWORD PTR [esi], edx
mov DWORD PTR [ecx], eax
pop esi
ret 0
_f1 ENDP
_main PROC
push OFFSET _product
push OFFSET _sum
push 456 ; 000001c8H
push 123 ; 0000007bH
call _f1
mov eax, DWORD PTR _product
mov ecx, DWORD PTR _sum
push eax
109
push ecx
push OFFSET $SG2803
add esp, 28
xor eax, eax
ret 0
_main ENDP
110
Let’s see this in OllyDbg:
Figure 1.25: OllyDbg: global variables addresses are passed to f1()
First, global variables’ addresses are passed to f1(). We can click “Follow in dump” on the stack element,
and we can see the place in the data segment allocated for the two variables.
111
These variables are zeroed, because non-initialized data (from BSS) is cleared before
the execution begins, [see ISO/IEC 9899:TC3 (C C99 standard), (2007) 6.7.8p10].
They reside in the data segment, we can verify this by pressing Alt-M and reviewing the memory map:
Figure 1.26: OllyDbg: memory map
112
Let’s trace (F7) to the start of f1():
Figure 1.27: OllyDbg: f1() starts
Two values are visible in the stack: 456 (0x1C8) and 123 (0x7B), and also the addresses of the two global
variables.
113
Let’s trace until the end of f1(). In the left bottom window we see how the results of the calculation
appear in the global variables:
Figure 1.28: OllyDbg: f1() execution completed
114
Now the global variables’ values are loaded into registers ready for passing to printf() (via the stack):
Figure 1.29: OllyDbg: global variables’ values are passed into printf()
Local variables example
Let’s rework our example slightly:

Listing 1.104: now the sum and product variables are local
void main()
{
int sum, product; // now variables are local in this function
f1(123, 456, &sum, &product);

printf ("sum=%d, product=%d\n", sum, product);
};
f1() code will not change. Only the code of main() will do:
Listing 1.105: Optimizing MSVC 2010 (/Ob0)
_product$ = -8 ; size = 4
_sum$ = -4 ; size = 4
_main PROC
; Line 10
sub esp, 8
; Line 13
lea eax, DWORD PTR _product$[esp+8]
push eax
lea ecx, DWORD PTR _sum$[esp+12]
push ecx
push 456 ; 000001c8H
push 123 ; 0000007bH
call _f1
; Line 14
mov edx, DWORD PTR _product$[esp+24]
mov eax, DWORD PTR _sum$[esp+24]
push edx
push eax
push OFFSET $SG2803
; Line 15
xor eax, eax
add esp, 36
ret 0
115
Let’s look again with OllyDbg. The addresses of the local variables in the stack are 0x2EF854 and 0x2EF858.
We see how these are pushed into the stack:
Figure 1.30: OllyDbg: local variables’ addresses are pushed into the stack
116
f1() starts. So far there is only random garbage in the stack at 0x2EF854 and 0x2EF858:
Figure 1.31: OllyDbg: f1() starting
117
f1() completes:
Figure 1.32: OllyDbg: f1() completes execution
We now find 0xDB18 and 0x243 at addresses 0x2EF854 and 0x2EF858. These values are the f1() results.
Conclusion
f1() could return pointers to any place in memory, located anywhere.

This is in essence the usefulness of the pointers.
By the way, C++ references work exactly the same way. Read more about them: ( 3.21.3 on page 558).
1.16.2 Swap input values

This will do the job:
#include <memory.h>
#include <stdio.h>
void swap_bytes (unsigned char* first, unsigned char* second)

{
unsigned char tmp1;
unsigned char tmp2;
tmp1=*first;
tmp2=*second;
*first=tmp2;
*second=tmp1;
};
int main()
{
// copy string into heap, so we will be able to modify it
char *s=strdup("string");
// swap 2nd and 3rd characters

swap_bytes (s+1, s+2);
printf ("%s\n", s);

};
118
As we can see, bytes are loaded into lower 8-bit parts of ECX and EBX using MOVZX (so higher parts of these
registers will be cleared) and then bytes are written back swapped.
Listing 1.106: Optimizing GCC 5.4

swap_bytes:
push ebx
mov edx, DWORD PTR [esp+8]
mov eax, DWORD PTR [esp+12]
movzx ecx, BYTE PTR [edx]
movzx ebx, BYTE PTR [eax]
mov BYTE PTR [edx], bl
mov BYTE PTR [eax], cl
pop ebx
ret
Addresses of both bytes are taken from arguments and through execution of the function are located in
EDX and EAX.
So we use pointers: probably, there is no better way to solve this task without them.
1.17 GOTO operator

The GOTO operator is generally considered as anti-pattern, see [Edgar Dijkstra, Go To Statement Con-
sidered Harmful (1968)88 ]. Nevertheless, it can be used reasonably, see [Donald E. Knuth, Structured
Programming with go to Statements (1974)89 ] 90 .
Here is a very simple example:
#include <stdio.h>
int main()
{
printf ("begin\n");
goto exit;
printf ("skip me!\n");
exit:
printf ("end\n");
};
Here is what we have got in MSVC 2012:

$SG2934 DB 'begin', 0aH, 00H
$SG2936 DB 'skip me!', 0aH, 00H
$SG2937 DB 'end', 0aH, 00H
_main PROC
push ebp
mov ebp, esp
push OFFSET $SG2934 ; 'begin'
call _printf
add esp, 4
jmp SHORT $exit$3
push OFFSET $SG2936 ; 'skip me!'
call _printf
add esp, 4
$exit$3:
push OFFSET $SG2937 ; 'end'
call _printf
add esp, 4
xor eax, eax
pop ebp
ret 0
_main ENDP
88 http://yurichev.com/mirrors/Dijkstra68.pdf
89 http://yurichev.com/mirrors/KnuthStructuredProgrammingGoTo.pdf
90 [Dennis Yurichev, C/C++ programming language notes] also has some examples.
119
The goto statement has been simply replaced by a JMP instruction, which has the same effect: uncondi-
tional jump to another place. The second printf() could be executed only with human intervention, by
using a debugger or by patching the code.
120
This could also be useful as a simple patching exercise. Let’s open the resulting executable in Hiew:
Figure 1.33: Hiew
121
Place the cursor to address JMP (0x410), press F3 (edit), press zero twice, so the opcode becomes EB 00:
Figure 1.34: Hiew
The second byte of the JMP opcode denotes the relative offset for the jump, 0 means the point right after
the current instruction.
So now JMP not skipping the second printf() call.
Press F9 (save) and exit. Now if we run the executable we will see this:
Listing 1.108: Patched executable output

C:\...>goto.exe
begin
skip me!
end
The same result could be achieved by replacing the JMP instruction with 2 NOP instructions.
NOP has an opcode of 0x90 and length of 1 byte, so we need 2 instructions as JMP replacement (which is
2 bytes in size).
1.17.1 Dead code

The second printf() call is also called “dead code” in compiler terms.
This means that the code will never be executed. So when you compile this example with optimizations,
the compiler removes “dead code”, leaving no trace of it:
Listing 1.109: Optimizing MSVC 2012

$SG2981 DB 'begin', 0aH, 00H
$SG2983 DB 'skip me!', 0aH, 00H
$SG2984 DB 'end', 0aH, 00H
_main PROC
push OFFSET $SG2981 ; 'begin'
call _printf
push OFFSET $SG2984 ; 'end'
$exit$4:
call _printf
add esp, 8
xor eax, eax
ret 0
122
_main ENDP
However, the compiler forgot to remove the “skip me!” string.
1.17.2 Exercise
Try to achieve the same result using your favorite compiler and debugger.
1.18 Conditional jumps

#include <stdio.h>
void f_signed (int a, int b)

{
if (a>b)
printf ("a>b\n");
if (a==b)
printf ("a==b\n");
if (a<b)
printf ("a<b\n");
};
void f_unsigned (unsigned int a, unsigned int b)

{
if (a>b)
printf ("a>b\n");
if (a==b)
printf ("a==b\n");
if (a<b)
printf ("a<b\n");
};
int main()
{
f_signed(1, 2);
f_unsigned(1, 2);
return 0;
};
x86
x86 + MSVC
Here is how the f_signed() function looks like:

_a$ = 8
_b$ = 12
_f_signed PROC
push ebp
mov ebp, esp
cmp eax, DWORD PTR _b$[ebp]
jle SHORT $LN3@f_signed
push OFFSET $SG737 ; 'a>b'
call _printf
add esp, 4
$LN3@f_signed:
mov ecx, DWORD PTR _a$[ebp]
cmp ecx, DWORD PTR _b$[ebp]
jne SHORT $LN2@f_signed
push OFFSET $SG739 ; 'a==b'
call _printf
123
add esp, 4
$LN2@f_signed:
cmp edx, DWORD PTR _b$[ebp]
jge SHORT $LN4@f_signed
push OFFSET $SG741 ; 'a<b'
call _printf
add esp, 4
$LN4@f_signed:
pop ebp
ret 0
_f_signed ENDP
The first instruction, JLE, stands for Jump if Less or Equal. In other words, if the second operand is larger
or equal to the first one, the control flow will be passed to the address or label specified in the instruction.
If this condition does not trigger because the second operand is smaller than the first one, the control
flow would not be altered and the first printf() would be executed. The second check is JNE: Jump if Not
Equal. The control flow will not change if the operands are equal.
The third check is JGE: Jump if Greater or Equal—jump if the first operand is larger than the second or if they
are equal. So, if all three conditional jumps are triggered, none of the printf() calls would be executed
whatsoever. This is impossible without special intervention. Now let’s take a look at the f_unsigned()
function. The f_unsigned() function is the same as f_signed(), with the exception that the JBE and JAE
instructions are used instead of JLE and JGE, as follows:
Listing 1.111: GCC

_a$ = 8 ; size = 4
_b$ = 12 ; size = 4
_f_unsigned PROC
push ebp
mov ebp, esp
jbe SHORT $LN3@f_unsigned
push OFFSET $SG2761 ; 'a>b'
call _printf
add esp, 4
$LN3@f_unsigned:
cmp ecx, DWORD PTR _b$[ebp]
jne SHORT $LN2@f_unsigned
push OFFSET $SG2763 ; 'a==b'
call _printf
add esp, 4
$LN2@f_unsigned:
cmp edx, DWORD PTR _b$[ebp]
jae SHORT $LN4@f_unsigned
push OFFSET $SG2765 ; 'a<b'
call _printf
add esp, 4
$LN4@f_unsigned:
pop ebp
ret 0
_f_unsigned ENDP
As already mentioned, the branch instructions are different: JBE—Jump if Below or Equal and JAE—Jump
if Above or Equal. These instructions (JA/JAE/JB/JBE) differ from JG/JGE/JL/JLE in the fact that they work
with unsigned numbers.
See also the section about signed number representations ( 2.2 on page 448). That is why if we see JG/JL
in use instead of JA/JB or vice-versa, we can be almost sure that the variables are signed or unsigned,
respectively. Here is also the main() function, where there is nothing much new to us:
Listing 1.112: main()

_main PROC
push ebp
mov ebp, esp
124
push 2
push 1
call _f_signed
add esp, 8
push 2
push 1
call _f_unsigned
add esp, 8
xor eax, eax
pop ebp
ret 0
_main ENDP
125
x86 + MSVC + OllyDbg
We can see how flags are set by running this example in OllyDbg. Let’s begin with f_unsigned(), which
works with unsigned numbers.
CMP is executed thrice here, but for the same arguments, so the flags are the same each time.
Result of the first comparison:
Figure 1.35: OllyDbg: f_unsigned(): first conditional jump
So, the flags are: C=1, P=1, A=1, Z=0, S=1, T=0, D=0, O=0.
They are named with one character for brevity in OllyDbg.
OllyDbg gives a hint that the (JBE) jump is to be triggered now. Indeed, if we take a look into Intel manuals
( 12.1.4 on page 999), we can read there that JBE is triggering if CF=1 or ZF=1. The condition is true here,
so the jump is triggered.
126
The next conditional jump:
Figure 1.36: OllyDbg: f_unsigned(): second conditional jump
OllyDbg gives a hint that JNZ is to be triggered now. Indeed, JNZ triggering if ZF=0 (zero flag).
127
The third conditional jump, JNB:
Figure 1.37: OllyDbg: f_unsigned(): third conditional jump
In Intel manuals ( 12.1.4 on page 999) we can see that JNB triggers if CF=0 (carry flag). That is not true
in our case, so the third printf() will execute.
128
Now let’s review the f_signed() function, which works with signed values, in OllyDbg. Flags are set in the
same way: C=1, P=1, A=1, Z=0, S=1, T=0, D=0, O=0. The first conditional jump JLE is to be triggered:
Figure 1.38: OllyDbg: f_signed(): first conditional jump
In Intel manuals ( 12.1.4 on page 999) we find that this instruction is triggered if ZF=1 or SF≠OF. SF≠OF in
our case, so the jump triggers.
129
The second JNZ conditional jump triggering: if ZF=0 (zero flag):
Figure 1.39: OllyDbg: f_signed(): second conditional jump
130
The third conditional jump JGE will not trigger because it would only do so if SF=OF, and that is not true
in our case:
Figure 1.40: OllyDbg: f_signed(): third conditional jump
131
x86 + MSVC + Hiew
We can try to patch the executable file in a way that the f_unsigned() function would always print “a==b”,
no matter the input values. Here is how it looks in Hiew:
Figure 1.41: Hiew: f_unsigned() function
Essentially, we have to accomplish three tasks:

• force the first jump to always trigger;
• force the second jump to never trigger;
• force the third jump to always trigger.
Thus we can direct the code flow to always pass through the second printf(), and output “a==b”.
Three instructions (or bytes) has to be patched:
• The first jump becomes JMP, but the jump offset would remain the same.
• The second jump might be triggered sometimes, but in any case it will jump to the next instruction,
because, we set the jump offset to 0.
In these instructions the jump offset is added to the address for the next instruction. So if the offset
is 0, the jump will transfer the control to the next instruction.
• The third jump we replace with JMP just as we do with the first one, so it will always trigger.
132
Here is the modified code:
Figure 1.42: Hiew: let’s modify the f_unsigned() function
If we miss to change any of these jumps, then several printf() calls may execute, while we want to
execute only one.
Non-optimizing GCC
Non-optimizing GCC 4.4.1 produces almost the same code, but with puts() ( 1.5.3 on page 20) instead
of printf().
Optimizing GCC
An observant reader may ask, why execute CMP several times, if the flags has the same values after each
execution?
Perhaps optimizing MSVC cannot do this, but optimizing GCC 4.8.1 can go deeper:
Listing 1.113: GCC 4.8.1 f_signed()

f_signed:
cmp DWORD PTR [esp+4], eax
jg .L6
je .L7
jge .L1
mov DWORD PTR [esp+4], OFFSET FLAT:.LC2 ; "a<b"
jmp puts
.L6:
mov DWORD PTR [esp+4], OFFSET FLAT:.LC0 ; "a>b"
jmp puts
133
.L1:
rep ret
.L7:
mov DWORD PTR [esp+4], OFFSET FLAT:.LC1 ; "a==b"
jmp puts
We also see JMP puts here instead of CALL puts / RETN.

This kind of trick will have explained later: 1.21.1 on page 154.
This type of x86 code is somewhat rare. MSVC 2012 as it seems, can’t generate such code. On the other
hand, assembly language programmers are fully aware of the fact that Jcc instructions can be stacked.
So if you see such stacking somewhere, it is highly probable that the code was hand-written.
The f_unsigned() function is not that æsthetically short:
Listing 1.114: GCC 4.8.1 f_unsigned()

f_unsigned:
push esi
push ebx
sub esp, 20
mov esi, DWORD PTR [esp+32]
mov ebx, DWORD PTR [esp+36]
cmp esi, ebx
ja .L13
cmp esi, ebx ; this instruction could be removed
je .L14
.L10:
jb .L15
add esp, 20
pop ebx
pop esi
ret
.L15:
mov DWORD PTR [esp+32], OFFSET FLAT:.LC2 ; "a<b"
add esp, 20
pop ebx
pop esi
jmp puts
.L13:
mov DWORD PTR [esp], OFFSET FLAT:.LC0 ; "a>b"
call puts
cmp esi, ebx
jne .L10
.L14:
mov DWORD PTR [esp+32], OFFSET FLAT:.LC1 ; "a==b"
add esp, 20
pop ebx
pop esi
jmp puts
Nevertheless, there are two CMP instructions instead of three.

So optimization algorithms of GCC 4.8.1 are probably not perfect yet.
ARM
32-bit ARM

.text:000000B8 EXPORT f_signed
.text:000000B8 f_signed ; CODE XREF: main+C
.text:000000B8 70 40 2D E9 STMFD SP!, {R4-R6,LR}
.text:000000BC 01 40 A0 E1 MOV R4, R1
134
.text:000000C0 04 00 50 E1 CMP R0, R4
.text:000000C4 00 50 A0 E1 MOV R5, R0
.text:000000C8 1A 0E 8F C2 ADRGT R0, aAB ; "a>b\n"
.text:000000CC A1 18 00 CB BLGT __2printf
.text:000000D0 04 00 55 E1 CMP R5, R4
.text:000000D4 67 0F 8F 02 ADREQ R0, aAB_0 ; "a==b\n"
.text:000000D8 9E 18 00 0B BLEQ __2printf
.text:000000DC 04 00 55 E1 CMP R5, R4
.text:000000E0 70 80 BD A8 LDMGEFD SP!, {R4-R6,PC}
.text:000000E4 70 40 BD E8 LDMFD SP!, {R4-R6,LR}
.text:000000E8 19 0E 8F E2 ADR R0, aAB_1 ; "a<b\n"
.text:000000EC 99 18 00 EA B __2printf
.text:000000EC ; End of function f_signed
Many instructions in ARM mode could be executed only when specific flags are set. E.g. this is often used
when comparing numbers.
For instance, the ADD instruction is in fact named ADDAL internally, where AL stands for Always, i.e., execute
always. The predicates are encoded in 4 high bits of the 32-bit ARM instructions (condition field). The
B instruction for unconditional jumping is in fact conditional and encoded just like any other conditional
jump, but has AL in the condition field, and it implies execute ALways, ignoring flags.
The ADRGT instruction works just like ADR but executes only in case the previous CMP instruction founds
one of the numbers greater than the another, while comparing the two (Greater Than).
The next BLGT instruction behaves exactly as BL and is triggered only if the result of the comparison
has been (Greater Than). ADRGT writes a pointer to the string a>b\n into R0 and BLGT calls printf().
Therefore, instructions suffixed with -GT are to execute only in case the value in R0 (which is a) is bigger
than the value in R4 (which is b).
Moving forward we see the ADREQ and BLEQ instructions. They behave just like ADR and BL, but are to be
executed only if operands were equal to each other during the last comparison. Another CMP is located
before them (because the printf() execution may have tampered the flags).
Then we see LDMGEFD, this instruction works just like LDMFD91 , but is triggered only when one of the values
is greater or equal than the other (Greater or Equal). The LDMGEFD SP!, {R4-R6,PC} instruction acts like
a function epilogue, but it will be triggered only if a >= b, and only then the function execution will finish.
But if that condition is not satisfied, i.e., a < b, then the control flow will continue to the next
“LDMFD SP!, {R4-R6,LR}” instruction, which is one more function epilogue. This instruction restores not
only the R4-R6 registers state, but also LR instead of PC, thus, it does not return from the function. The
last two instructions call printf() with the string «a<b\n» as a sole argument. We already examined an
unconditional jump to the printf() function instead of function return in «printf() with several arguments»
section ( 1.11.2 on page 53).
f_unsigned is similar, only the ADRHI, BLHI, and LDMCSFD instructions are used there, these predicates
(HI = Unsigned higher, CS = Carry Set (greater than or equal)) are analogous to those examined before,
but for unsigned values.
There is not much new in the main() function for us:
Listing 1.116: main()

.text:00000128 EXPORT main
.text:00000128 main
.text:0000012C 02 10 A0 E3 MOV R1, #2
.text:00000130 01 00 A0 E3 MOV R0, #1
.text:00000134 DF FF FF EB BL f_signed
.text:00000138 02 10 A0 E3 MOV R1, #2
.text:0000013C 01 00 A0 E3 MOV R0, #1
.text:00000140 EA FF FF EB BL f_unsigned
.text:00000144 00 00 A0 E3 MOV R0, #0
.text:00000148 10 80 BD E8 LDMFD SP!, {R4,PC}
.text:00000148 ; End of function main
That is how you can get rid of conditional jumps in ARM mode.
Why is this so good? Read here: 2.10.1 on page 462.
91 LDMFD
135
There is no such feature in x86, except the CMOVcc instruction, it is the same as MOV, but triggered only
when specific flags are set, usually set by CMP.

.text:00000072 f_signed ; CODE XREF: main+6
.text:00000072 70 B5 PUSH {R4-R6,LR}
.text:00000074 0C 00 MOVS R4, R1
.text:00000076 05 00 MOVS R5, R0
.text:00000078 A0 42 CMP R0, R4
.text:0000007A 02 DD BLE loc_82
.text:0000007C A4 A0 ADR R0, aAB ; "a>b\n"
.text:0000007E 06 F0 B7 F8 BL __2printf
.text:00000082
.text:00000082 loc_82 ; CODE XREF: f_signed+8
.text:00000082 A5 42 CMP R5, R4
.text:00000084 02 D1 BNE loc_8C
.text:00000086 A4 A0 ADR R0, aAB_0 ; "a==b\n"
.text:00000088 06 F0 B2 F8 BL __2printf
.text:0000008C
.text:0000008C loc_8C ; CODE XREF: f_signed+12
.text:0000008C A5 42 CMP R5, R4
.text:0000008E 02 DA BGE locret_96
.text:00000090 A3 A0 ADR R0, aAB_1 ; "a<b\n"
.text:00000092 06 F0 AD F8 BL __2printf
.text:00000096
.text:00000096 locret_96 ; CODE XREF: f_signed+1C
.text:00000096 70 BD POP {R4-R6,PC}
.text:00000096 ; End of function f_signed
Only B instructions in Thumb mode may be supplemented by condition codes, so the Thumb code looks
more ordinary.
BLE is a normal conditional jump Less than or Equal, BNE—Not Equal, BGE—Greater than or Equal.
f_unsigned is similar, only other instructions are used while dealing with unsigned values: BLS (Unsigned
lower or same) and BCS (Carry Set (Greater than or equal)).
ARM64: Optimizing GCC (Linaro) 4.9
Listing 1.118: f_signed()

f_signed:
; W0=a, W1=b
cmp w0, w1
bgt .L19 ; Branch if Greater Than (a>b)
beq .L20 ; Branch if Equal (a==b)
bge .L15 ; Branch if Greater than or Equal (a>=b) (impossible here)
; a<b
adrp x0, .LC11 ; "a<b"
add x0, x0, :lo12:.LC11
b puts
.L19:
adrp x0, .LC9 ; "a>b"
b puts
.L15: ; impossible to get here
ret
.L20:
adrp x0, .LC10 ; "a==b"
add x0, x0, :lo12:.LC10
b puts
Listing 1.119: f_unsigned()

f_unsigned:
136
stp x29, x30, [sp, -48]!
; W0=a, W1=b
cmp w0, w1
add x29, sp, 0
str x19, [sp,16]
mov w19, w0
bhi .L25 ; Branch if HIgher (a>b)
cmp w19, w1
beq .L26 ; Branch if Equal (a==b)
.L23:
bcc .L27 ; Branch if Carry Clear (if less than) (a<b)
; function epilogue, impossible to be here
ldr x19, [sp,16]
ldp x29, x30, [sp], 48
ret
.L27:
ldr x19, [sp,16]
adrp x0, .LC11 ; "a<b"
ldp x29, x30, [sp], 48
add x0, x0, :lo12:.LC11
b puts
.L25:
adrp x0, .LC9 ; "a>b"
str x1, [x29,40]
bl puts
ldr x1, [x29,40]
cmp w19, w1
bne .L23 ; Branch if Not Equal
.L26:
ldr x19, [sp,16]
adrp x0, .LC10 ; "a==b"
ldp x29, x30, [sp], 48
add x0, x0, :lo12:.LC10
b puts
The comments were added by the author of this book. What is striking is that the compiler is not aware
that some conditions are not possible at all, so there is dead code at some places, which can never be
executed.
Exercise
Try to optimize these functions manually for size, removing redundant instructions, without adding new
ones.
MIPS
One distinctive MIPS feature is the absence of flags. Apparently, it was done to simplify the analysis of
data dependencies.
There are instructions similar to SETcc in x86: SLT (“Set on Less Than”: signed version) and SLTU (unsigned
version). These instructions sets destination register value to 1 if the condition is true or to 0 if otherwise.
The destination register is then checked using BEQ (“Branch on Equal”) or BNE (“Branch on Not Equal”)
and a jump may occur. So, this instruction pair has to be used in MIPS for comparison and branch. Let’s
first start with the signed version of our function:

.text:00000000 f_signed: # CODE XREF: main+18
.text:00000000
.text:00000000 var_10 = -0x10
.text:00000000 var_8 = -8
.text:00000000 var_4 = -4
.text:00000000 arg_0 = 0
.text:00000000 arg_4 = 4
.text:00000000
.text:00000000 addiu $sp, -0x20
137
.text:00000004 sw $ra, 0x20+var_4($sp)
.text:00000008 sw $fp, 0x20+var_8($sp)
.text:00000018 sw $gp, 0x20+var_10($sp)
; store input values into local stack:
.text:0000001C sw $a0, 0x20+arg_0($fp)
.text:00000020 sw $a1, 0x20+arg_4($fp)
; reload them.
.text:00000024 lw $v1, 0x20+arg_0($fp)
.text:00000028 lw $v0, 0x20+arg_4($fp)
; $v0=b
; $v1=a
; this is pseudoinstruction. in fact, "slt $v0,$v0,$v1" is there.
; so $v0 will be set to 1 if $v0<$v1 (b<a) or to 0 if otherwise:
.text:00000030 slt $v0, $v1
; jump to loc_5c, if condition is not true.
; this is pseudoinstruction. in fact, "beq $v0,$zero,loc_5c" is there:
.text:00000034 beqz $v0, loc_5C
; print "a>b" and finish
.text:00000038 or $at, $zero ; branch delay slot, NOP
.text:0000003C lui $v0, (unk_230 >> 16) # "a>b"
.text:00000040 addiu $a0, $v0, (unk_230 & 0xFFFF) # "a>b"
.text:00000044 lw $v0, (puts & 0xFFFF)($gp)
.text:0000004C move $t9, $v0
.text:00000050 jalr $t9
.text:00000058 lw $gp, 0x20+var_10($fp)
.text:0000005C
.text:0000005C loc_5C: # CODE XREF: f_signed+34
.text:0000005C lw $v1, 0x20+arg_0($fp)
.text:00000060 lw $v0, 0x20+arg_4($fp)
; check if a==b, jump to loc_90 if its not true:
.text:00000068 bne $v1, $v0, loc_90
.text:0000006C or $at, $zero ; branch delay slot, NOP
; condition is true, so print "a==b" and finish:
.text:00000070 lui $v0, (aAB >> 16) # "a==b"
.text:00000074 addiu $a0, $v0, (aAB & 0xFFFF) # "a==b"
.text:00000080 move $t9, $v0
.text:00000084 jalr $t9
.text:0000008C lw $gp, 0x20+var_10($fp)
.text:00000090
.text:00000090 loc_90: # CODE XREF: f_signed+68
.text:00000090 lw $v1, 0x20+arg_0($fp)
.text:00000094 lw $v0, 0x20+arg_4($fp)
; check if $v1<$v0 (a<b), set $v0 to 1 if condition is true:
.text:0000009C slt $v0, $v1, $v0
; if condition is not true (i.e., $v0==0), jump to loc_c8:
.text:000000A0 beqz $v0, loc_C8
.text:000000A4 or $at, $zero ; branch delay slot, NOP
; condition is true, print "a<b" and finish
.text:000000A8 lui $v0, (aAB_0 >> 16) # "a<b"
.text:000000AC addiu $a0, $v0, (aAB_0 & 0xFFFF) # "a<b"
.text:000000B0 lw $v0, (puts & 0xFFFF)($gp)
.text:000000B4 or $at, $zero ; NOP
.text:000000B8 move $t9, $v0
.text:000000BC jalr $t9
.text:000000C0 or $at, $zero ; branch delay slot, NOP
.text:000000C4 lw $gp, 0x20+var_10($fp)
.text:000000C8
; all 3 conditions were false, so just finish:
.text:000000C8 loc_C8: # CODE XREF: f_signed+A0
.text:000000C8 move $sp, $fp
138
.text:000000CC lw $ra, 0x20+var_4($sp)
.text:000000D0 lw $fp, 0x20+var_8($sp)
.text:000000D4 addiu $sp, 0x20
.text:000000D8 jr $ra
.text:000000DC or $at, $zero ; branch delay slot, NOP
.text:000000DC # End of function f_signed
SLT REG0, REG0, REG1 is reduced by IDA to its shorter form:

SLT REG0, REG1.
We also see there BEQZ pseudo instruction (“Branch if Equal to Zero”),
which are in fact BEQ REG, $ZERO, LABEL.
The unsigned version is just the same, but SLTU (unsigned version, hence “U” in name) is used instead of
SLT:

.text:000000E0 f_unsigned: # CODE XREF: main+28
.text:000000E0
.text:000000E0 var_10 = -0x10
.text:000000E0 var_8 = -8
.text:000000E0 var_4 = -4
.text:000000E0 arg_0 = 0
.text:000000E0 arg_4 = 4
.text:000000E0
.text:000000E0 addiu $sp, -0x20
.text:000000E4 sw $ra, 0x20+var_4($sp)
.text:000000E8 sw $fp, 0x20+var_8($sp)
.text:000000EC move $fp, $sp
.text:000000F0 la $gp, __gnu_local_gp
.text:000000F8 sw $gp, 0x20+var_10($sp)
.text:000000FC sw $a0, 0x20+arg_0($fp)
.text:00000100 sw $a1, 0x20+arg_4($fp)
.text:00000104 lw $v1, 0x20+arg_0($fp)
.text:00000108 lw $v0, 0x20+arg_4($fp)
.text:0000010C or $at, $zero
.text:00000110 sltu $v0, $v1
.text:00000114 beqz $v0, loc_13C
.text:0000011C lui $v0, (unk_230 >> 16)
.text:00000120 addiu $a0, $v0, (unk_230 & 0xFFFF)
.text:0000012C move $t9, $v0
.text:00000130 jalr $t9
.text:00000138 lw $gp, 0x20+var_10($fp)
.text:0000013C
.text:0000013C loc_13C: # CODE XREF: f_unsigned+34
.text:0000013C lw $v1, 0x20+arg_0($fp)
.text:00000140 lw $v0, 0x20+arg_4($fp)
.text:00000148 bne $v1, $v0, loc_170
.text:00000150 lui $v0, (aAB >> 16) # "a==b"
.text:00000154 addiu $a0, $v0, (aAB & 0xFFFF) # "a==b"
.text:00000160 move $t9, $v0
.text:00000164 jalr $t9
.text:0000016C lw $gp, 0x20+var_10($fp)
.text:00000170
.text:00000170 loc_170: # CODE XREF: f_unsigned+68
.text:00000170 lw $v1, 0x20+arg_0($fp)
.text:00000174 lw $v0, 0x20+arg_4($fp)
.text:0000017C sltu $v0, $v1, $v0
.text:00000180 beqz $v0, loc_1A8
139
.text:00000188 lui $v0, (aAB_0 >> 16) # "a<b"
.text:0000018C addiu $a0, $v0, (aAB_0 & 0xFFFF) # "a<b"
.text:00000198 move $t9, $v0
.text:000001A0 or $at, $zero
.text:000001A4 lw $gp, 0x20+var_10($fp)
.text:000001A8
.text:000001A8 loc_1A8: # CODE XREF: f_unsigned+A0
.text:000001A8 move $sp, $fp
.text:000001AC lw $ra, 0x20+var_4($sp)
.text:000001B0 lw $fp, 0x20+var_8($sp)
.text:000001B4 addiu $sp, 0x20
.text:000001B8 jr $ra
.text:000001BC or $at, $zero
.text:000001BC # End of function f_unsigned
1.18.2 Calculating absolute value

A simple function:
int my_abs (int i)
{
if (i<0)
return -i;
else
return i;
};
Optimizing MSVC
This is how the code is usually generated:

i$ = 8
my_abs PROC
; ECX = input
test ecx, ecx
; check for sign of input value
; skip NEG instruction if sign is positive
jns SHORT $LN2@my_abs
; negate value
neg ecx
$LN2@my_abs:
; prepare result in EAX:
mov eax, ecx
ret 0
my_abs ENDP
GCC 4.9 does mostly the same.
Optimizing Keil 6/2013: Thumb mode
Listing 1.123: Optimizing Keil 6/2013: Thumb mode

my_abs PROC
CMP r0,#0
; is input value equal to zero or greater than zero?
; skip RSBS instruction then
BGE |L0.6|
; subtract input value from 0:
RSBS r0,r0,#0
|L0.6|
BX lr
ENDP
140
ARM lacks a negate instruction, so the Keil compiler uses the “Reverse Subtract” instruction, which just
subtracts with reversed operands.
Optimizing Keil 6/2013: ARM mode
It is possible to add condition codes to some instructions in ARM mode, so that is what the Keil compiler
does:
Listing 1.124: Optimizing Keil 6/2013: ARM mode

my_abs PROC
CMP r0,#0
; execute "Reverse Subtract" instruction only if input value is less than 0:
RSBLT r0,r0,#0
BX lr
ENDP
Now there are no conditional jumps and this is good: 2.10.1 on page 462.
Non-optimizing GCC 4.9 (ARM64)
ARM64 has instruction NEG for negating:
Listing 1.125: Optimizing GCC 4.9 (ARM64)

my_abs:
sub sp, sp, #16
str w0, [sp,12]
ldr w0, [sp,12]
; compare input value with contents of WZR register
; (which always holds zero)
cmp w0, wzr
bge .L2
ldr w0, [sp,12]
neg w0, w0
b .L3
.L2:
ldr w0, [sp,12]
.L3:
add sp, sp, 16
ret
MIPS

my_abs:
; jump if $a0<0:
bltz $a0, locret_10
; just return input value ($a0) in $v0:
move $v0, $a0
jr $ra
or $at, $zero ; branch delay slot, NOP
locret_10:
; negate input value and store it in $v0:
jr $ra
; this is pseudoinstruction. in fact, this is "subu $v0,$zero,$a0" ($v0=0-$a0)
negu $v0, $a0
Here we see a new instruction: BLTZ (“Branch if Less Than Zero”).

There is also the NEGU pseudo instruction, which just does subtraction from zero. The “U” suffix in both
SUBU and NEGU implies that no exception to be raised in case of integer overflow.
Branchless version?
You could have also a branchless version of this code. This we will review later: 3.16 on page 517.
141
1.18.3 Ternary conditional operator
The ternary conditional operator in C/C++ has the following syntax:
expression ? expression : expression
Here is an example:
const char* f (int a)
{
return a==10 ? "it is ten" : "it is not ten";
};
x86
Old and non-optimizing compilers generate assembly code just as if an if/else statement was used:

$SG746 DB 'it is ten', 00H
$SG747 DB 'it is not ten', 00H
tv65 = -4 ; this will be used as a temporary variable

_a$ = 8
_f PROC
push ebp
mov ebp, esp
push ecx
; compare input value with 10
cmp DWORD PTR _a$[ebp], 10
; jump to $LN3@f if not equal
jne SHORT $LN3@f
; store pointer to the string into temporary variable:
mov DWORD PTR tv65[ebp], OFFSET $SG746 ; 'it is ten'
; jump to exit
jmp SHORT $LN4@f
$LN3@f:
; store pointer to the string into temporary variable:
mov DWORD PTR tv65[ebp], OFFSET $SG747 ; 'it is not ten'
$LN4@f:
; this is exit. copy pointer to the string from temporary variable to EAX.
mov eax, DWORD PTR tv65[ebp]
mov esp, ebp
pop ebp
ret 0
_f ENDP

_a$ = 8 ; size = 4
_f PROC
cmp DWORD PTR _a$[esp-4], 10
mov eax, OFFSET $SG792 ; 'it is ten'
; jump to $LN4@f if equal
je SHORT $LN4@f
mov eax, OFFSET $SG793 ; 'it is not ten'
$LN4@f:
ret 0
_f ENDP
Newer compilers are more concise:

142
a$ = 8
f PROC
; load pointers to the both strings
lea rdx, OFFSET FLAT:$SG1355 ; 'it is ten'
lea rax, OFFSET FLAT:$SG1356 ; 'it is not ten'
cmp ecx, 10
; if equal, copy value from RDX ("it is ten")
; if not, do nothing. pointer to the string "it is not ten" is still in RAX as for now.
cmove rax, rdx
ret 0
f ENDP
Optimizing GCC 4.8 for x86 also uses the CMOVcc instruction, while the non-optimizing GCC 4.8 uses
conditional jumps.
ARM
Optimizing Keil for ARM mode also uses the conditional instructions ADRcc:

f PROC
CMP r0,#0xa
; if comparison result is EQual, copy pointer to the "it is ten" string into R0
ADREQ r0,|L0.16| ; "it is ten"
; if comparison result is Not Equal, copy pointer to the "it is not ten" string into R0
ADRNE r0,|L0.28| ; "it is not ten"
BX lr
ENDP
|L0.16|
DCB "it is ten",0
|L0.28|
DCB "it is not ten",0
Without manual intervention, the two instructions ADREQ and ADRNE cannot be executed in the same run.
Optimizing Keil for Thumb mode needs to use conditional jump instructions, since there are no load in-
structions that support conditional flags:

f PROC
CMP r0,#0xa
; jump to |L0.8| if EQual
BEQ |L0.8|
ADR r0,|L0.12| ; "it is not ten"
BX lr
|L0.8|
ADR r0,|L0.28| ; "it is ten"
BX lr
ENDP
|L0.12|
DCB "it is not ten",0
|L0.28|
DCB "it is ten",0
ARM64
Optimizing GCC (Linaro) 4.9 for ARM64 also uses conditional jumps:

f:
cmp x0, 10
143
beq .L3 ; branch if equal
adrp x0, .LC1 ; "it is ten"
ret
.L3:
adrp x0, .LC0 ; "it is not ten"
ret
.LC0:
.string "it is ten"
.LC1:
.string "it is not ten"
That is because ARM64 does not have a simple load instruction with conditional flags, like ADRcc in 32-bit
ARM mode or CMOVcc in x86.
It has, however, “Conditional SELect” instruction (CSEL)[ARM Architecture Reference Manual, ARMv8, for
ARMv8-A architecture profile, (2013)p390, C5.5], but GCC 4.9 does not seem to be smart enough to use
it in such piece of code.
MIPS
Unfortunately, GCC 4.4.5 for MIPS is not very smart, either:

$LC0:
.ascii "it is not ten\000"
$LC1:
.ascii "it is ten\000"
f:
li $2,10 # 0xa
; compare $a0 and 10, jump if equal:
beq $4,$2,$L2
nop ; branch delay slot
; leave address of "it is not ten" string in $v0 and return:

lui $2,%hi($LC0)
j $31
$L2:
; leave address of "it is ten" string in $v0 and return:
lui $2,%hi($LC1)
j $31
Let’s rewrite it in an if/else way
const char* f (int a)

{
if (a==10)
return "it is ten";
else
return "it is not ten";
};
Interestingly, optimizing GCC 4.8 for x86 was also able to use CMOVcc in this case:

.LC0:
.string "it is ten"
.LC1:
.string "it is not ten"
f:
.LFB0:
cmp DWORD PTR [esp+4], 10
144
mov edx, OFFSET FLAT:.LC1 ; "it is not ten"
mov eax, OFFSET FLAT:.LC0 ; "it is ten"
; if comparison result is Not Equal, copy EDX value to EAX
; if not, do nothing
cmovne eax, edx
ret
Optimizing Keil in ARM mode generates code identical to listing.1.130.

But the optimizing MSVC 2012 is not that good (yet).
Conclusion
Why optimizing compilers try to get rid of conditional jumps? Read here about it: 2.10.1 on page 462.
1.18.4 Getting minimal and maximal values

32-bit
int my_max(int a, int b)

{
if (a>b)
return a;
else
return b;
};
int my_min(int a, int b)

{
if (a<b)
return a;
else
return b;
};

_a$ = 8
_b$ = 12
_my_min PROC
push ebp
mov ebp, esp
; compare A and B:
; jump, if A is greater or equal to B:
jge SHORT $LN2@my_min
; reload A to EAX if otherwise and jump to exit
jmp SHORT $LN3@my_min
jmp SHORT $LN3@my_min ; this is redundant JMP
$LN2@my_min:
; return B
mov eax, DWORD PTR _b$[ebp]
$LN3@my_min:
pop ebp
ret 0
_my_min ENDP
_a$ = 8
_b$ = 12
_my_max PROC
push ebp
mov ebp, esp
; compare A and B:
; jump if A is less or equal to B:
jle SHORT $LN2@my_max
145
; reload A to EAX if otherwise and jump to exit
jmp SHORT $LN3@my_max
jmp SHORT $LN3@my_max ; this is redundant JMP
$LN2@my_max:
; return B
$LN3@my_max:
pop ebp
ret 0
_my_max ENDP
These two functions differ only in the conditional jump instruction: JGE (“Jump if Greater or Equal”) is used
in the first one and JLE (“Jump if Less or Equal”) in the second.
There is one unneeded JMP instruction in each function, which MSVC presumably left by mistake.
Branchless
ARM for Thumb mode reminds us of x86 code:

my_max PROC
; R0=A
; R1=B
; compare A and B:
CMP r0,r1
; branch if A is greater then B:
BGT |L0.6|
; otherwise (A<=B) return R1 (B):
MOVS r0,r1
|L0.6|
; return
BX lr
ENDP
my_min PROC
; R0=A
; R1=B
; compare A and B:
CMP r0,r1
; branch if A is less then B:
BLT |L0.14|
; otherwise (A>=B) return R1 (B):
MOVS r0,r1
|L0.14|
; return
BX lr
ENDP
The functions differ in the branching instruction: BGT and BLT. It’s possible to use conditional suffixes in
ARM mode, so the code is shorter.
MOVcc is to be executed only if the condition is met:

my_max PROC
; R0=A
; R1=B
; compare A and B:
CMP r0,r1
; return B instead of A by placing B in R0
; this instruction will trigger only if A<=B (hence, LE - Less or Equal)
; if instruction is not triggered (in case of A>B), A is still in R0 register
MOVLE r0,r1
BX lr
ENDP
146
my_min PROC
; R0=A
; R1=B
; compare A and B:
CMP r0,r1
; return B instead of A by placing B in R0
; this instruction will trigger only if A>=B (hence, GE - Greater or Equal)
; if instruction is not triggered (in case of A<B), A value is still in R0 register
MOVGE r0,r1
BX lr
ENDP
Optimizing GCC 4.8.1 and optimizing MSVC 2013 can use CMOVcc instruction, which is analogous to MOVcc
in ARM:

my_max:
; EDX=A
; EAX=B
; compare A and B:
cmp edx, eax
; if A>=B, load A value into EAX
; the instruction idle if otherwise (if A<B)
cmovge eax, edx
ret
my_min:
; EDX=A
; EAX=B
; compare A and B:
cmp edx, eax
; if A<=B, load A value into EAX
; the instruction idle if otherwise (if A>B)
cmovle eax, edx
ret
64-bit
#include <stdint.h>
int64_t my_max(int64_t a, int64_t b)

{
if (a>b)
return a;
else
return b;
};
int64_t my_min(int64_t a, int64_t b)

{
if (a<b)
return a;
else
return b;
};
There is some unneeded value shuffling, but the code is comprehensible:

my_max:
sub sp, sp, #16
str x0, [sp,8]
str x1, [sp]
147
ldr x1, [sp,8]
ldr x0, [sp]
cmp x1, x0
ble .L2
ldr x0, [sp,8]
b .L3
.L2:
ldr x0, [sp]
.L3:
add sp, sp, 16
ret
my_min:
sub sp, sp, #16
str x0, [sp,8]
str x1, [sp]
ldr x1, [sp,8]
ldr x0, [sp]
cmp x1, x0
bge .L5
ldr x0, [sp,8]
b .L6
.L5:
ldr x0, [sp]
.L6:
add sp, sp, 16
ret
Branchless
No need to load function arguments from the stack, as they are already in the registers:

my_max:
; RDI=A
; RSI=B
; compare A and B:
cmp rdi, rsi
; prepare B in RAX for return:
mov rax, rsi
; if A>=B, put A (RDI) in RAX for return.
; this instruction is idle if otherwise (if A<B)
cmovge rax, rdi
ret
my_min:
; RDI=A
; RSI=B
; compare A and B:
cmp rdi, rsi
; prepare B in RAX for return:
mov rax, rsi
; if A<=B, put A (RDI) in RAX for return.
; this instruction is idle if otherwise (if A>B)
cmovle rax, rdi
ret
MSVC 2013 does almost the same.

ARM64 has the CSEL instruction, which works just as MOVcc in ARM or CMOVcc in x86, just the name is
different: “Conditional SELect”.
Listing 1.141: Optimizing GCC 4.9.1 ARM64

my_max:
; X0=A
; X1=B
; compare A and B:
148
cmp x0, x1
; select X0 (A) to X0 if X0>=X1 or A>=B (Greater or Equal)
; select X1 (B) to X0 if A<B
csel x0, x0, x1, ge
ret
my_min:
; X0=A
; X1=B
; compare A and B:
cmp x0, x1
; select X0 (A) to X0 if X0<=X1 or A<=B (Less or Equal)
; select X1 (B) to X0 if A>B
csel x0, x0, x1, le
ret
MIPS
Unfortunately, GCC 4.4.5 for MIPS is not that good:

my_max:
; set $v1 to 1 if $a1<$a0, or clear otherwise (if $a1>$a0):
slt $v1, $a1, $a0
; jump, if $v1 is 0 (or $a1>$a0):
beqz $v1, locret_10
; this is branch delay slot
; prepare $a1 in $v0 in case of branch triggered:
move $v0, $a1
; no branch triggered, prepare $a0 in $v0:
move $v0, $a0
locret_10:
jr $ra
; the min() function is same, but input operands in SLT instruction are swapped:
my_min:
slt $v1, $a0, $a1
beqz $v1, locret_28
move $v0, $a1
move $v0, $a0
locret_28:
jr $ra
Do not forget about the branch delay slots: the first MOVE is executed before BEQZ, the second MOVE is
executed only if the branch hasn’t been taken.
1.18.5 Conclusion
x86
Here’s the rough skeleton of a conditional jump:
Listing 1.143: x86

CMP register, register/value
Jcc true ; cc=condition code
false:
... some code to be executed if comparison result is false ...
JMP exit
true:
... some code to be executed if comparison result is true ...
exit:
149
ARM
Listing 1.144: ARM

Bcc true ; cc=condition code
false:
... some code to be executed if comparison result is false ...
JMP exit
true:
... some code to be executed if comparison result is true ...
exit:
MIPS
Listing 1.145: Check for zero

BEQZ REG, label
...
Listing 1.146: Check for less than zero using pseudoinstruction

BLTZ REG, label
...
Listing 1.147: Check for equal values

BEQ REG1, REG2, label
...
Listing 1.148: Check for non-equal values

BNE REG1, REG2, label
...
Listing 1.149: Check for less than (signed)

SLT REG1, REG2, REG3
BEQ REG1, label
...
Listing 1.150: Check for less than (unsigned)

SLTU REG1, REG2, REG3
BEQ REG1, label
...
Branchless
If the body of a condition statement is very short, the conditional move instruction can be used: MOVcc in
ARM (in ARM mode), CSEL in ARM64, CMOVcc in x86.
ARM
It’s possible to use conditional suffixes in ARM mode for some instructions:
Listing 1.151: ARM (ARM mode)

instr1_cc ; some instruction will be executed if condition code is true
instr2_cc ; some other instruction will be executed if other condition code is true
... etc...
Of course, there is no limit for the number of instructions with conditional code suffixes, as long as the
CPU flags are not modified by any of them.
Thumb mode has the IT instruction, allowing to add conditional suffixes to the next four instructions. Read
more about it: 1.25.7 on page 259.
150
Listing 1.152: ARM (Thumb mode)
ITEEE EQ ; set these suffixes: if-then-else-else-else
instr1 ; instruction will be executed if condition is true
instr2 ; instruction will be executed if condition is false
1.18.6 Exercise
(ARM64) Try rewriting the code in listing.1.132 by removing all conditional jump instructions and using the
CSEL instruction.
1.19 Software cracking

The vast majority of software can be cracked like that — by searching the very place where protection is
checked, a dongle ( 8.8 on page 818), license key, serial number, etc.
Often, it looks like:
...
call check_protection
jz all_OK
call message_box_protection_missing
call exit
all_OK:
; proceed
...
So if you see a patch (or “crack”), that cracks a software, and that patch replaces 0x74/0x75 (JZ/JNZ)
byte(s) by 0xEB (JMP), this is it.
The process of software cracking comes down to a search of that JMP.
There are also a cases, when a software checks protection from time to time, this can be a dongle, or
a license server can be queried through the Internet. Then you have to look for a function that checks
protection. Then to patch it, to put there xor eax, eax / retn, or mov eax, 1 / retn.
It’s important to understand that after patching of function beginning, usually, a garbage follows these
two instructions. The garbage consists of part of one instruction and the several next instructions.
This is a real case. The beginning of a function which we want to replace by return 1;
Listing 1.153: Before

8BFF mov edi,edi
55 push ebp
8BEC mov ebp,esp
81EC68080000 sub esp,000000868
A110C00001 mov eax,[00100C010]
33C5 xor eax,ebp
8945FC mov [ebp][-4],eax
53 push ebx
8B5D08 mov ebx,[ebp][8]
...
Listing 1.154: After

B801000000 mov eax,1
C3 retn
EC in al,dx
68080000A1 push 0A1000008
10C0 adc al,al
0001 add [ecx],al
33C5 xor eax,ebp
8945FC mov [ebp][-4],eax
53 push ebx
151
8B5D08 mov ebx,[ebp][8]
...
Several incorrect instructions appears — IN, PUSH, ADC, ADD, after which, Hiew disassembler (which I just
used) synchronized and continued to disassemble all the rest.
This is not important — all these instructions followed RETN will never be executed, unless a direct jump
would occur from some place, and that wouldn’t be possible in general case.
Also, a global boolean variable can be present, having a flag, was the software registered or not.
init_etc proc
...
call check_protection_or_license_file
mov is_demo, eax
...
retn
init_etc endp
...
save_file proc
...
mov eax, is_demo
cmp eax, 1
jz all_OK1
call message_box_it_is_a_demo_no_saving_allowed
retn
:all_OK1
; continue saving file
...
save_proc endp
somewhere_else proc
mov eax, is_demo

cmp eax, 1
jz all_OK
; check if we run for 15 minutes

; exit if it is so
; or show nagging screen
:all_OK2
; continue
somewhere_else endp
A beginning of the check_protection_or_license_file() function could be patched, so that it will al-

ways return 1, or, if this is better by some reason, all JZ/JNZ instructions can be patched as well.
More about patching: 11.1.
1.20 Impossible shutdown practical joke (Windows 7)

I don’t quite remember how I found the ExitWindowsEx() function in Windows 98’s (it was late 1990s)
user32.dll file. Probably, I just spotted its self-describing name. And then I tried to block it by patching its
beginning by 0xC3 byte (RETN).
The result was funny: Windows 98 cannot be shutted down anymore. Had to press reset button.
These days I tried to do the same in Windows 7, that was created almost 10 years later and based on
completely different Windows NT base. Still, ExitWindowsEx() function present in user32.dll file and
serves the same purpose.
152
First, I turned off Windows File Protection by adding this to registry (Windows would silently restore modi-
fied system files otherwise):
Windows Registry Editor Version 5.00
[HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Windows NT\CurrentVersion\Winlogon]
"SFCDisable"=dword:ffffff9d
Then I renamed c:\windows\system32\user32.dll to user32.dll.bak. I found ExitWindowsEx() ex-

port entry using Hiew (IDA can help as well) and put 0xC3 byte here. I restarted by Windows 7 and now
it can’t be shutted down. ”Restart” and ”Logoff” buttons don’t work anymore.
I don’t know if it’s funny today or not, but back then, in late 1990s, my friend took patched user32.dll file
on a floppy diskette and copied it to all the computers (within his reach, that worked under Windows 98
(almost all)) at his university. No Windows can be shutted down after and his computer science teacher
was lurid. (Hopefully he can forgive us if he is reading this right now.)
If you do this, backup everything. The best idea is to run Windows under a virtual machine.
1.21 switch()/case/default
1.21.1 Small number of cases
#include <stdio.h>
void f (int a)
{
switch (a)
{
case 0: printf ("zero\n"); break;
case 1: printf ("one\n"); break;
case 2: printf ("two\n"); break;
default: printf ("something unknown\n"); break;
};
};
int main()
{
f (2); // test
};
x86
Non-optimizing MSVC
Result (MSVC 2010):

tv64 = -4 ; size = 4
_a$ = 8 ; size = 4
_f PROC
push ebp
mov ebp, esp
push ecx
mov DWORD PTR tv64[ebp], eax
cmp DWORD PTR tv64[ebp], 0
je SHORT $LN4@f
je SHORT $LN3@f
je SHORT $LN2@f
jmp SHORT $LN1@f
$LN4@f:
push OFFSET $SG739 ; 'zero', 0aH, 00H
call _printf
153
add esp, 4
jmp SHORT $LN7@f
$LN3@f:
push OFFSET $SG741 ; 'one', 0aH, 00H
call _printf
add esp, 4
jmp SHORT $LN7@f
$LN2@f:
push OFFSET $SG743 ; 'two', 0aH, 00H
call _printf
add esp, 4
jmp SHORT $LN7@f
$LN1@f:
push OFFSET $SG745 ; 'something unknown', 0aH, 00H
call _printf
add esp, 4
$LN7@f:
mov esp, ebp
pop ebp
ret 0
_f ENDP
Our function with a few cases in switch() is in fact analogous to this construction:
void f (int a)
{
if (a==0)
printf ("zero\n");
else if (a==1)
printf ("one\n");
else if (a==2)
printf ("two\n");
else
printf ("something unknown\n");
};
If we work with switch() with a few cases it is impossible to be sure if it was a real switch() in the source
code, or just a pack of if() statements.
This implies that switch() is like syntactic sugar for a large number of nested if()s.
There is nothing especially new to us in the generated code, with the exception of the compiler moving
input variable a to a temporary local variable tv64 92 .
If we compile this in GCC 4.4.1, we’ll get almost the same result, even with maximal optimization turned
on (-O3 option).
Optimizing MSVC
Now let’s turn on optimization in MSVC (/Ox): cl 1.c /Fa1.asm /Ox
Listing 1.156: MSVC

_a$ = 8 ; size = 4
_f PROC
mov eax, DWORD PTR _a$[esp-4]
sub eax, 0
je SHORT $LN4@f
sub eax, 1
je SHORT $LN3@f
sub eax, 1
je SHORT $LN2@f
mov DWORD PTR _a$[esp-4], OFFSET $SG791 ; 'something unknown', 0aH, 00H
jmp _printf
$LN2@f:
mov DWORD PTR _a$[esp-4], OFFSET $SG789 ; 'two', 0aH, 00H
jmp _printf
$LN3@f:
92 Local variables in stack are prefixed with tv—that’s how MSVC names internal variables for its needs
154
mov DWORD PTR _a$[esp-4], OFFSET $SG787 ; 'one', 0aH, 00H
jmp _printf
$LN4@f:
mov DWORD PTR _a$[esp-4], OFFSET $SG785 ; 'zero', 0aH, 00H
jmp _printf
_f ENDP
Here we can see some dirty hacks.

First: the value of a is placed in EAX and 0 is subtracted from it. Sounds absurd, but it is done to check if
the value in EAX is 0. If yes, the ZF flag is to be set (e.g. subtracting from 0 is 0) and the first conditional
jump JE (Jump if Equal or synonym JZ —Jump if Zero) is to be triggered and control flow is to be passed
to the $LN4@f label, where the 'zero' message is being printed. If the first jump doesn’t get triggered,
1 is subtracted from the input value and if at some stage the result is 0, the corresponding jump is to be
triggered.
And if no jump gets triggered at all, the control flow passes to printf() with string argument
'something unknown'.
Second: we see something unusual for us: a string pointer is placed into the a variable, and then printf()
is called not via CALL, but via JMP. There is a simple explanation for that: the caller pushes a value to the
stack and calls our function via CALL. CALL itself pushes the return address (RA) to the stack and does an
unconditional jump to our function address. Our function at any point of execution (since it do not contain
any instruction that moves the stack pointer) has the following stack layout:
• ESP—points to RA
• ESP+4—points to the a variable
On the other side, when we have to call printf() here we need exactly the same stack layout, except
for the first printf() argument, which needs to point to the string. And that is what our code does.
It replaces the function’s first argument with the address of the string and jumps to printf(), as if we
didn’t call our function f(), but directly printf(). printf() prints a string to stdout and then executes
the RET instruction, which POPs RA from the stack and control flow is returned not to f() but rather to
f()’s caller, bypassing the end of the f() function.
All this is possible because printf() is called right at the end of the f() function in all cases. In some
way, it is similar to the longjmp()93 function. And of course, it is all done for the sake of speed.
A similar case with the ARM compiler is described in “printf() with several arguments” section, here ( 1.11.2
on page 53).
93 wikipedia
155
OllyDbg
Since this example is tricky, let’s trace it in OllyDbg.

OllyDbg can detect such switch() constructs, and it can add some useful comments. EAX is 2 at the
beginning, that’s the function’s input value:
Figure 1.43: OllyDbg: EAX now contain the first (and only) function argument
156
0 is subtracted from 2 in EAX. Of course, EAX still contains 2. But the ZF flag is now 0, indicating that the
resulting value is non-zero:
Figure 1.44: OllyDbg: SUB executed
157
DEC is executed and EAX now contains 1. But 1 is non-zero, so the ZF flag is still 0:
Figure 1.45: OllyDbg: first DEC executed
158
Next DEC is executed. EAX is finally 0 and the ZF flag gets set, because the result is zero:
Figure 1.46: OllyDbg: second DEC executed
OllyDbg shows that this jump is to be taken now.
159
A pointer to the string “two” is to be written into the stack now:
Figure 1.47: OllyDbg: pointer to the string is to be written at the place of the first argument
Please note: the current argument of the function is 2 and 2 is now in the stack at the address 0x001EF850.
160
MOV writes the pointer to the string at address 0x001EF850 (see the stack window). Then, jump happens.
This is the first instruction of the printf() function in MSVCR100.DLL (This example was compiled with
/MD switch):
Figure 1.48: OllyDbg: first instruction of printf() in MSVCR100.DLL
Now printf() treats the string at 0x00FF3010 as its only argument and prints the string.
161
This is the last instruction of printf():
Figure 1.49: OllyDbg: last instruction of printf() in MSVCR100.DLL
The string “two” has just been printed to the console window.
162
Now let’s press F7 or F8 (step over) and return…not to f(), but rather to main():
Figure 1.50: OllyDbg: return to main()
Yes, the jump has been direct, from the guts of printf() to main(). Because RA in the stack points not
to some place in f(), but rather to main(). And CALL 0x00FF1000 has been the actual instruction which
called f().
ARM: Optimizing Keil 6/2013 (ARM mode)
.text:0000014C f1:
.text:0000014C 00 00 50 E3 CMP R0, #0
.text:00000150 13 0E 8F 02 ADREQ R0, aZero ; "zero\n"
.text:00000154 05 00 00 0A BEQ loc_170
.text:00000158 01 00 50 E3 CMP R0, #1
.text:0000015C 4B 0F 8F 02 ADREQ R0, aOne ; "one\n"
.text:00000160 02 00 00 0A BEQ loc_170
.text:00000164 02 00 50 E3 CMP R0, #2
.text:00000168 4A 0F 8F 12 ADRNE R0, aSomethingUnkno ; "something unknown\n"
.text:0000016C 4E 0F 8F 02 ADREQ R0, aTwo ; "two\n"
.text:00000170
.text:00000170 loc_170: ; CODE XREF: f1+8
.text:00000170 ; f1+14
.text:00000170 78 18 00 EA B __2printf
Again, by investigating this code we cannot say if it was a switch() in the original source code, or just a
pack of if() statements.
Anyway, we see here predicated instructions again (like ADREQ (Equal)) which is triggered only in case
R0 = 0, and then loads the address of the string «zero\n» into R0. The next instruction BEQ redirects
control flow to loc_170, if R0 = 0.
An astute reader may ask, will BEQ trigger correctly since ADREQ it has already filled the R0 register before
with another value?
Yes, it will since BEQ checks the flags set by the CMP instruction, and ADREQ does not modify any flags at
all.
The rest of the instructions are already familiar to us. There is only one call to printf(), at the end,
and we have already examined this trick here ( 1.11.2 on page 53). At the end, there are three paths to
printf().
163
The last instruction, CMP R0, #2, is needed to check if a = 2.
If it is not true, then ADRNE loads a pointer to the string «something unknown \n» into R0, since a has
already been checked to be equal to 0 or 1, and we can sure that the a variable is not equal to these
numbers at this point. And if R0 = 2, a pointer to the string «two\n» will be loaded by ADREQ into R0.

.text:000000D4 f1:
.text:000000D4 10 B5 PUSH {R4,LR}
.text:000000D6 00 28 CMP R0, #0
.text:000000D8 05 D0 BEQ zero_case
.text:000000DA 01 28 CMP R0, #1
.text:000000DC 05 D0 BEQ one_case
.text:000000DE 02 28 CMP R0, #2
.text:000000E0 05 D0 BEQ two_case
.text:000000E2 91 A0 ADR R0, aSomethingUnkno ; "something unknown\n"
.text:000000E4 04 E0 B default_case
.text:000000E6 zero_case: ; CODE XREF: f1+4

.text:000000E6 95 A0 ADR R0, aZero ; "zero\n"
.text:000000E8 02 E0 B default_case
.text:000000EA one_case: ; CODE XREF: f1+8

.text:000000EA 96 A0 ADR R0, aOne ; "one\n"
.text:000000EC 00 E0 B default_case
.text:000000EE two_case: ; CODE XREF: f1+C

.text:000000EE 97 A0 ADR R0, aTwo ; "two\n"
.text:000000F0 default_case ; CODE XREF: f1+10
.text:000000F0 ; f1+14
.text:000000F0 06 F0 7E F8 BL __2printf
.text:000000F4 10 BD POP {R4,PC}
As was already mentioned, it is not possible to add conditional predicates to most instructions in Thumb
mode, so the Thumb-code here is somewhat similar to the easily understandable x86 CISC-style code.
ARM64: Non-optimizing GCC (Linaro) 4.9

.LC12:
.string "zero"
.LC13:
.string "one"
.LC14:
.string "two"
.LC15:
.string "something unknown"
f12:
stp x29, x30, [sp, -32]!
add x29, sp, 0
str w0, [x29,28]
ldr w0, [x29,28]
cmp w0, 1
beq .L34
cmp w0, 2
beq .L35
cmp w0, wzr
bne .L38 ; jump to default label
adrp x0, .LC12 ; "zero"
add x0, x0, :lo12:.LC12
bl puts
b .L32
.L34:
adrp x0, .LC13 ; "one"
add x0, x0, :lo12:.LC13
bl puts
b .L32
.L35:
adrp x0, .LC14 ; "two"
164
add x0, x0, :lo12:.LC14
bl puts
b .L32
.L38:
adrp x0, .LC15 ; "something unknown"
add x0, x0, :lo12:.LC15
bl puts
nop
.L32:
ldp x29, x30, [sp], 32
ret
The type of the input value is int, hence register W0 is used to hold it instead of the whole X0 register.
The string pointers are passed to puts() using an ADRP/ADD instructions pair just like it was demonstrated
in the “Hello, world!” example: 1.5.3 on page 23.

f12:
cmp w0, 1
beq .L31
cmp w0, 2
beq .L32
cbz w0, .L35
; default case
adrp x0, .LC15 ; "something unknown"
add x0, x0, :lo12:.LC15
b puts
.L35:
adrp x0, .LC12 ; "zero"
add x0, x0, :lo12:.LC12
b puts
.L32:
adrp x0, .LC14 ; "two"
add x0, x0, :lo12:.LC14
b puts
.L31:
adrp x0, .LC13 ; "one"
add x0, x0, :lo12:.LC13
b puts
Better optimized piece of code. CBZ (Compare and Branch on Zero) instruction does jump if W0 is zero.
There is also a direct jump to puts() instead of calling it, like it was explained before: 1.21.1 on page 154.
MIPS

f:
lui $gp, (__gnu_local_gp >> 16)
; is it 1?
li $v0, 1
beq $a0, $v0, loc_60
la $gp, (__gnu_local_gp & 0xFFFF) ; branch delay slot
; is it 2?
li $v0, 2
beq $a0, $v0, loc_4C
; jump, if not equal to 0:
bnez $a0, loc_38
; zero case:
lui $a0, ($LC0 >> 16) # "zero"
lw $t9, (puts & 0xFFFF)($gp)
or $at, $zero ; load delay slot, NOP
jr $t9 ; branch delay slot, NOP
la $a0, ($LC0 & 0xFFFF) # "zero" ; branch delay slot
165
loc_38: # CODE XREF: f+1C
lui $a0, ($LC3 >> 16) # "something unknown"
jr $t9
la $a0, ($LC3 & 0xFFFF) # "something unknown" ; branch delay slot
loc_4C: # CODE XREF: f+14

lui $a0, ($LC2 >> 16) # "two"
jr $t9
la $a0, ($LC2 & 0xFFFF) # "two" ; branch delay slot
loc_60: # CODE XREF: f+8

lui $a0, ($LC1 >> 16) # "one"
jr $t9
la $a0, ($LC1 & 0xFFFF) # "one" ; branch delay slot
The function always ends with calling puts(), so here we see a jump to puts() (JR: “Jump Register”)
instead of “jump and link”. We talked about this earlier: 1.21.1 on page 154.
We also often see NOP instructions after LW ones. This is “load delay slot”: another delay slot in MIPS.
An instruction next to LW may execute at the moment while LW loads value from memory.
However, the next instruction must not use the result of LW.
Modern MIPS CPUs have a feature to wait if the next instruction uses result of LW, so this is somewhat
outdated, but GCC still adds NOPs for older MIPS CPUs. In general, it can be ignored.
Conclusion
A switch() with few cases is indistinguishable from an if/else construction, for example: listing.1.21.1.
1.21.2 A lot of cases

If a switch() statement contains a lot of cases, it is not very convenient for the compiler to emit too large
code with a lot JE/JNE instructions.
#include <stdio.h>
void f (int a)
{
switch (a)
{
case 0: printf ("zero\n"); break;
case 1: printf ("one\n"); break;
case 2: printf ("two\n"); break;
case 3: printf ("three\n"); break;
case 4: printf ("four\n"); break;
default: printf ("something unknown\n"); break;
};
};
int main()
{
f (2); // test
};
x86
Non-optimizing MSVC
We get (MSVC 2010):
166
tv64 = -4 ; size = 4
_a$ = 8 ; size = 4
_f PROC
push ebp
mov ebp, esp
push ecx
mov DWORD PTR tv64[ebp], eax
ja SHORT $LN1@f
mov ecx, DWORD PTR tv64[ebp]
jmp DWORD PTR $LN11@f[ecx*4]
$LN6@f:
push OFFSET $SG739 ; 'zero', 0aH, 00H
call _printf
add esp, 4
jmp SHORT $LN9@f
$LN5@f:
push OFFSET $SG741 ; 'one', 0aH, 00H
call _printf
add esp, 4
jmp SHORT $LN9@f
$LN4@f:
push OFFSET $SG743 ; 'two', 0aH, 00H
call _printf
add esp, 4
jmp SHORT $LN9@f
$LN3@f:
push OFFSET $SG745 ; 'three', 0aH, 00H
call _printf
add esp, 4
jmp SHORT $LN9@f
$LN2@f:
push OFFSET $SG747 ; 'four', 0aH, 00H
call _printf
add esp, 4
jmp SHORT $LN9@f
$LN1@f:
push OFFSET $SG749 ; 'something unknown', 0aH, 00H
call _printf
add esp, 4
$LN9@f:
mov esp, ebp
pop ebp
ret 0
npad 2 ; align next label
$LN11@f:
DD $LN6@f ; 0
DD $LN5@f ; 1
DD $LN4@f ; 2
DD $LN3@f ; 3
DD $LN2@f ; 4
_f ENDP
What we see here is a set of printf() calls with various arguments. All they have not only addresses in
the memory of the process, but also internal symbolic labels assigned by the compiler. All these labels
are also mentioned in the $LN11@f internal table.
At the function start, if a is greater than 4, control flow is passed to label $LN1@f, where printf() with
argument 'something unknown' is called.
But if the value of a is less or equals to 4, then it gets multiplied by 4 and added with the $LN11@f table
address. That is how an address inside the table is constructed, pointing exactly to the element we need.
For example, let’s say a is equal to 2. 2 ∗ 4 = 8 (all table elements are addresses in a 32-bit process and
that is why all elements are 4 bytes wide). The address of the $LN11@f table + 8 is the table element
where the $LN4@f label is stored. JMP fetches the $LN4@f address from the table and jumps to it.
167
This table is sometimes called jumptable or branch table94 .
Then the corresponding printf() is called with argument 'two'.
Literally, the jmp DWORD PTR $LN11@f[ecx*4] instruction implies jump to the DWORD that is stored at
address $LN11@f + ecx * 4.
npad ( .1.7 on page 1024) is an assembly language macro that align the next label so that it will be stored
at an address aligned on a 4 bytes (or 16 bytes) boundary. This is very suitable for the processor since
it is able to fetch 32-bit values from memory through the memory bus, cache memory, etc., in a more
effective way if it is aligned.
94 The whole method was once called computed GOTO in early versions of Fortran: wikipedia. Not quite relevant these days, but
what a term!
168
OllyDbg
Let’s try this example in OllyDbg. The input value of the function (2) is loaded into EAX:
Figure 1.51: OllyDbg: function’s input value is loaded in EAX
169
The input value is checked, is it bigger than 4? If not, the “default” jump is not taken:
Figure 1.52: OllyDbg: 2 is no bigger than 4: no jump is taken
170
Here we see a jumptable:
Figure 1.53: OllyDbg: calculating destination address using jumptable
Here we’ve clicked “Follow in Dump” → “Address constant”, so now we see the jumptable in the data
window. These are 5 32-bit values95 . ECX is now 2, so the third element (can be indexed as 296 ) of the
table is to be used. It’s also possible to click “Follow in Dump” → “Memory address” and OllyDbg will show
the element addressed by the JMP instruction. That’s 0x010B103A.
95 They are underlined by OllyDbg because these are also FIXUPs: 6.5.2 on page 751, we are going to come back to them later
96 About indexing, see also: 3.22.3 on page 593
171
After the jump we are at 0x010B103A: the code printing “two” will now be executed:
Figure 1.54: OllyDbg: now we at the case: label
Non-optimizing GCC
Let’s see what GCC 4.4.1 generates:

public f
f proc near ; CODE XREF: main+10

arg_0 = dword ptr 8
push ebp
mov ebp, esp
sub esp, 18h
cmp [ebp+arg_0], 4
ja short loc_8048444
mov eax, [ebp+arg_0]
shl eax, 2
mov eax, ds:off_804855C[eax]
jmp eax
loc_80483FE: ; DATA XREF: .rodata:off_804855C

mov [esp+18h+var_18], offset aZero ; "zero"
call _puts
jmp short locret_8048450
loc_804840C: ; DATA XREF: .rodata:08048560

mov [esp+18h+var_18], offset aOne ; "one"
call _puts
loc_804841A: ; DATA XREF: .rodata:08048564

mov [esp+18h+var_18], offset aTwo ; "two"
call _puts
loc_8048428: ; DATA XREF: .rodata:08048568
172
mov [esp+18h+var_18], offset aThree ; "three"
call _puts
loc_8048436: ; DATA XREF: .rodata:0804856C

mov [esp+18h+var_18], offset aFour ; "four"
call _puts
loc_8048444: ; CODE XREF: f+A

mov [esp+18h+var_18], offset aSomethingUnkno ; "something unknown"
call _puts
locret_8048450: ; CODE XREF: f+26

; f+34...
leave
retn
f endp
off_804855C dd offset loc_80483FE ; DATA XREF: f+12

dd offset loc_804840C
dd offset loc_804841A
dd offset loc_8048428
dd offset loc_8048436
It is almost the same, with a little nuance: argument arg_0 is multiplied by 4 by shifting it to left by 2 bits
(it is almost the same as multiplication by 4) ( 1.24.2 on page 216). Then the address of the label is taken
from the off_804855C array, stored in EAX, and then JMP EAX does the actual jump.
ARM: Optimizing Keil 6/2013 (ARM mode)

00000174 f2
00000174 05 00 50 E3 CMP R0, #5 ; switch 5 cases
00000178 00 F1 8F 30 ADDCC PC, PC, R0,LSL#2 ; switch jump
0000017C 0E 00 00 EA B default_case ; jumptable 00000178 default case
00000180
00000180 loc_180 ; CODE XREF: f2+4
00000180 03 00 00 EA B zero_case ; jumptable 00000178 case 0
00000184
00000184 loc_184 ; CODE XREF: f2+4
00000184 04 00 00 EA B one_case ; jumptable 00000178 case 1
00000188
00000188 loc_188 ; CODE XREF: f2+4
00000188 05 00 00 EA B two_case ; jumptable 00000178 case 2
0000018C
0000018C loc_18C ; CODE XREF: f2+4
0000018C 06 00 00 EA B three_case ; jumptable 00000178 case 3
00000190
00000190 loc_190 ; CODE XREF: f2+4
00000190 07 00 00 EA B four_case ; jumptable 00000178 case 4
00000194
00000194 zero_case ; CODE XREF: f2+4
00000194 ; f2:loc_180
00000194 EC 00 8F E2 ADR R0, aZero ; jumptable 00000178 case 0
00000198 06 00 00 EA B loc_1B8
0000019C
0000019C one_case ; CODE XREF: f2+4
0000019C ; f2:loc_184
0000019C EC 00 8F E2 ADR R0, aOne ; jumptable 00000178 case 1
000001A0 04 00 00 EA B loc_1B8
173
000001A4
000001A4 two_case ; CODE XREF: f2+4
000001A4 ; f2:loc_188
000001A4 01 0C 8F E2 ADR R0, aTwo ; jumptable 00000178 case 2
000001A8 02 00 00 EA B loc_1B8
000001AC
000001AC three_case ; CODE XREF: f2+4
000001AC ; f2:loc_18C
000001AC 01 0C 8F E2 ADR R0, aThree ; jumptable 00000178 case 3
000001B0 00 00 00 EA B loc_1B8
000001B4
000001B4 four_case ; CODE XREF: f2+4
000001B4 ; f2:loc_190
000001B4 01 0C 8F E2 ADR R0, aFour ; jumptable 00000178 case 4
000001B8
000001B8 loc_1B8 ; CODE XREF: f2+24
000001B8 ; f2+2C
000001B8 66 18 00 EA B __2printf
000001BC
000001BC default_case ; CODE XREF: f2+4
000001BC ; f2+8
000001BC D4 00 8F E2 ADR R0, aSomethingUnkno ; jumptable 00000178 default case
000001C0 FC FF FF EA B loc_1B8
This code makes use of the ARM mode feature in which all instructions have a fixed size of 4 bytes.
Let’s keep in mind that the maximum value for a is 4 and any greater value will cause «something un-
known\n» string to be printed.
The first CMP R0, #5 instruction compares the input value of a with 5.
97
The next ADDCC PC, PC, R0,LSL#2 instruction is being executed only if R0 < 5 (CC=Carry clear / Less
than). Consequently, if ADDCC does not trigger (it is a R0 ≥ 5 case), a jump to default_case label will occur.
But if R0 < 5 and ADDCC triggers, the following is to be happen:
The value in R0 is multiplied by 4. In fact, LSL#2 at the instruction’s suffix stands for “shift left by 2
bits”. But as we will see later ( 1.24.2 on page 216) in section “Shifts”, shift left by 2 bits is equivalent to
multiplying by 4.
Then we add R0 ∗ 4 to the current value in PC, thus jumping to one of the B (Branch) instructions located
below.
At the moment of the execution of ADDCC, the value in PC is 8 bytes ahead (0x180) than the address at
which the ADDCC instruction is located (0x178), or, in other words, 2 instructions ahead.
This is how the pipeline in ARM processors works: when ADDCC is executed, the processor at the moment
is beginning to process the instruction after the next one, so that is why PC points there. This has to be
memorized.
If a = 0, then is to be added to the value in PC, and the actual value of the PC will be written into PC (which
is 8 bytes ahead) and a jump to the label loc_180 will happen, which is 8 bytes ahead of the point where
the ADDCC instruction is.
If a = 1, then P C + 8 + a ∗ 4 = P C + 8 + 1 ∗ 4 = P C + 12 = 0x184 will be written to PC, which is the address of
the loc_184 label.
With every 1 added to a, the resulting PC is increased by 4.
4 is the instruction length in ARM mode and also, the length of each B instruction, of which there are 5 in
row.
Each of these five B instructions passes control further, to what was programmed in the switch().
Pointer loading of the corresponding string occurs there, etc.
97 ADD—addition
174

000000F6 EXPORT f2
000000F6 f2
000000F6 10 B5 PUSH {R4,LR}
000000F8 03 00 MOVS R3, R0
000000FA 06 F0 69 F8 BL __ARM_common_switch8_thumb ; switch 6 cases
000000FE 05 DCB 5
000000FF 04 06 08 0A 0C 10 DCB 4, 6, 8, 0xA, 0xC, 0x10 ; jump table for switch statement
00000105 00 ALIGN 2
00000106
00000106 zero_case ; CODE XREF: f2+4
00000106 8D A0 ADR R0, aZero ; jumptable 000000FA case 0
00000108 06 E0 B loc_118
0000010A
0000010A one_case ; CODE XREF: f2+4
0000010A 8E A0 ADR R0, aOne ; jumptable 000000FA case 1
0000010C 04 E0 B loc_118
0000010E
0000010E two_case ; CODE XREF: f2+4
0000010E 8F A0 ADR R0, aTwo ; jumptable 000000FA case 2
00000110 02 E0 B loc_118
00000112
00000112 three_case ; CODE XREF: f2+4
00000112 90 A0 ADR R0, aThree ; jumptable 000000FA case 3
00000114 00 E0 B loc_118
00000116
00000116 four_case ; CODE XREF: f2+4
00000116 91 A0 ADR R0, aFour ; jumptable 000000FA case 4
00000118
00000118 loc_118 ; CODE XREF: f2+12
00000118 ; f2+16
00000118 06 F0 6A F8 BL __2printf
0000011C 10 BD POP {R4,PC}
0000011E
0000011E default_case ; CODE XREF: f2+4
0000011E 82 A0 ADR R0, aSomethingUnkno ; jumptable 000000FA default case
00000120 FA E7 B loc_118
000061D0 EXPORT __ARM_common_switch8_thumb

000061D0 __ARM_common_switch8_thumb ; CODE XREF: example6_f2+4
000061D0 78 47 BX PC
000061D2 00 00 ALIGN 4
000061D2 ; End of function __ARM_common_switch8_thumb
000061D2
000061D4 __32__ARM_common_switch8_thumb ; CODE XREF:
__ARM_common_switch8_thumb
000061D4 01 C0 5E E5 LDRB R12, [LR,#-1]
000061D8 0C 00 53 E1 CMP R3, R12
000061DC 0C 30 DE 27 LDRCSB R3, [LR,R12]
000061E0 03 30 DE 37 LDRCCB R3, [LR,R3]
000061E4 83 C0 8E E0 ADD R12, LR, R3,LSL#1
000061E8 1C FF 2F E1 BX R12
000061E8 ; End of function __32__ARM_common_switch8_thumb
One cannot be sure that all instructions in Thumb and Thumb-2 modes has the same size. It can even be
said that in these modes the instructions have variable lengths, just like in x86.
So there is a special table added that contains information about how much cases are there (not including
default-case), and an offset for each with a label to which control must be passed in the corresponding
case.
175
A special function is present here in order to deal with the table and pass control,
named __ARM_common_switch8_thumb. It starts with BX PC, whose function is to switch the processor to
ARM-mode. Then you see the function for table processing.
It is too advanced to describe it here now, so let’s omit it.
It is interesting to note that the function uses the LR register as a pointer to the table.
Indeed, after calling of this function, LR contains the address after
BL __ARM_common_switch8_thumb instruction, where the table starts.
It is also worth noting that the code is generated as a separate function in order to reuse it, so the compiler
doesn’t generate the same code for every switch() statement.
IDA successfully perceived it as a service function and a table, and added comments to the labels like
jumptable 000000FA case 0.
MIPS

f:
; jump to loc_24 if input value is lesser than 5:
sltiu $v0, $a0, 5
bnez $v0, loc_24
la $gp, (__gnu_local_gp & 0xFFFF) ; branch delay slot
; input value is greater or equal to 5.
; print "something unknown" and finish:
lui $a0, ($LC5 >> 16) # "something unknown"
or $at, $zero ; NOP
jr $t9
la $a0, ($LC5 & 0xFFFF) # "something unknown" ; branch delay slot
loc_24: # CODE XREF: f+8

; load address of jumptable
; LA is pseudoinstruction, LUI and ADDIU pair are there in fact:
la $v0, off_120
; multiply input value by 4:
sll $a0, 2
; sum up multiplied value and jumptable address:
addu $a0, $v0, $a0
; load element from jumptable:
lw $v0, 0($a0)
or $at, $zero ; NOP
; jump to the address we got in jumptable:
jr $v0
sub_44: # DATA XREF: .rodata:0000012C

; print "three" and finish
lui $a0, ($LC3 >> 16) # "three"
or $at, $zero ; NOP
jr $t9
la $a0, ($LC3 & 0xFFFF) # "three" ; branch delay slot
sub_58: # DATA XREF: .rodata:00000130

; print "four" and finish
lui $a0, ($LC4 >> 16) # "four"
or $at, $zero ; NOP
jr $t9
la $a0, ($LC4 & 0xFFFF) # "four" ; branch delay slot
sub_6C: # DATA XREF: .rodata:off_120

; print "zero" and finish
lui $a0, ($LC0 >> 16) # "zero"
or $at, $zero ; NOP
176
jr $t9
la $a0, ($LC0 & 0xFFFF) # "zero" ; branch delay slot

; print "one" and finish
lui $a0, ($LC1 >> 16) # "one"
or $at, $zero ; NOP
jr $t9
la $a0, ($LC1 & 0xFFFF) # "one" ; branch delay slot

; print "two" and finish
lui $a0, ($LC2 >> 16) # "two"
or $at, $zero ; NOP
jr $t9
la $a0, ($LC2 & 0xFFFF) # "two" ; branch delay slot
; may be placed in .rodata section:

off_120: .word sub_6C
.word sub_80
.word sub_94
.word sub_44
.word sub_58
The new instruction for us is SLTIU (“Set on Less Than Immediate Unsigned”).
This is the same as SLTU (“Set on Less Than Unsigned”), but “I” stands for “immediate”, i.e., a number
has to be specified in the instruction itself.
BNEZ is “Branch if Not Equal to Zero”.
Code is very close to the other ISAs. SLL (“Shift Word Left Logical”) does multiplication by 4.
MIPS is a 32-bit CPU after all, so all addresses in the jumptable are 32-bit ones.
Conclusion
Rough skeleton of switch():
Listing 1.163: x86

MOV REG, input
CMP REG, 4 ; maximal number of cases
JA default
SHL REG, 2 ; find element in table. shift for 3 bits in x64.
MOV REG, jump_table[REG]
JMP REG
case1:
; do something
JMP exit
case2:
; do something
JMP exit
case3:
; do something
JMP exit
case4:
; do something
JMP exit
case5:
; do something
JMP exit
default:
...
exit:
177
....
jump_table dd case1
dd case2
dd case3
dd case4
dd case5
The jump to the address in the jump table may also be implemented using this instruction:
JMP jump_table[REG*4]. Or JMP jump_table[REG*8] in x64.
A jumptable is just array of pointers, like the one described later: 1.26.5 on page 283.
1.21.3 When there are several case statements in one block

Here is a very widespread construction: several case statements for a single block:
#include <stdio.h>
void f(int a)
{
switch (a)
{
case 1:
case 2:
case 7:
case 10:
printf ("1, 2, 7, 10\n");
break;
case 3:
case 4:
case 5:
case 6:
printf ("3, 4, 5\n");
break;
case 8:
case 9:
case 20:
case 21:
printf ("8, 9, 21\n");
break;
case 22:
printf ("22\n");
break;
default:
printf ("default\n");
break;
};
};
int main()
{
f(4);
};
It’s too wasteful to generate a block for each possible case, so what is usually done is to generate each
block plus some kind of dispatcher.
MSVC

1 $SG2798 DB '1, 2, 7, 10', 0aH, 00H
2 $SG2800 DB '3, 4, 5', 0aH, 00H
3 $SG2802 DB '8, 9, 21', 0aH, 00H
4 $SG2804 DB '22', 0aH, 00H
5 $SG2806 DB 'default', 0aH, 00H
6
178
7 _a$ = 8
8 _f PROC
9 mov eax, DWORD PTR _a$[esp-4]
10 dec eax
11 cmp eax, 21
12 ja SHORT $LN1@f
13 movzx eax, BYTE PTR $LN10@f[eax]
14 jmp DWORD PTR $LN11@f[eax*4]
15 $LN5@f:
16 mov DWORD PTR _a$[esp-4], OFFSET $SG2798 ; '1, 2, 7, 10'
17 jmp DWORD PTR __imp__printf
18 $LN4@f:
19 mov DWORD PTR _a$[esp-4], OFFSET $SG2800 ; '3, 4, 5'
21 $LN3@f:
22 mov DWORD PTR _a$[esp-4], OFFSET $SG2802 ; '8, 9, 21'
24 $LN2@f:
25 mov DWORD PTR _a$[esp-4], OFFSET $SG2804 ; '22'
27 $LN1@f:
28 mov DWORD PTR _a$[esp-4], OFFSET $SG2806 ; 'default'
30 npad 2 ; align $LN11@f table on 16-byte boundary
31 $LN11@f:
32 DD $LN5@f ; print '1, 2, 7, 10'
33 DD $LN4@f ; print '3, 4, 5'
34 DD $LN3@f ; print '8, 9, 21'
35 DD $LN2@f ; print '22'
36 DD $LN1@f ; print 'default'
37 $LN10@f:
38 DB 0 ; a=1
39 DB 0 ; a=2
40 DB 1 ; a=3
41 DB 1 ; a=4
42 DB 1 ; a=5
43 DB 1 ; a=6
44 DB 0 ; a=7
45 DB 2 ; a=8
46 DB 2 ; a=9
47 DB 0 ; a=10
48 DB 4 ; a=11
49 DB 4 ; a=12
50 DB 4 ; a=13
51 DB 4 ; a=14
52 DB 4 ; a=15
53 DB 4 ; a=16
54 DB 4 ; a=17
55 DB 4 ; a=18
56 DB 4 ; a=19
57 DB 2 ; a=20
58 DB 2 ; a=21
59 DB 3 ; a=22
60 _f ENDP
We see two tables here: the first table ($LN10@f) is an index table, and the second one ($LN11@f) is an
array of pointers to blocks.
First, the input value is used as an index in the index table (line 13).
Here is a short legend for the values in the table: 0 is the first case block (for values 1, 2, 7, 10), 1 is the
second one (for values 3, 4, 5), 2 is the third one (for values 8, 9, 21), 3 is the fourth one (for value 22), 4
is for the default block.
There we get an index for the second table of code pointers and we jump to it (line 14).
What is also worth noting is that there is no case for input value 0.
That’s why we see the DEC instruction at line 10, and the table starts at a = 1, because there is no need to
allocate a table element for a = 0.
179
This is a very widespread pattern.
So why is this economical? Why isn’t it possible to make it as before ( 1.21.2 on page 172), just with one
table consisting of block pointers? The reason is that the elements in index table are 8-bit, hence it’s all
more compact.
GCC
GCC does the job in the way we already discussed ( 1.21.2 on page 172), using just one table of pointers.
ARM64: Optimizing GCC 4.9.1
There is no code to be triggered if the input value is 0, so GCC tries to make the jump table more compact
and so it starts at 1 as an input value.
GCC 4.9.1 for ARM64 uses an even cleverer trick. It’s able to encode all offsets as 8-bit bytes.
Let’s recall that all ARM64 instructions have a size of 4 bytes.
GCC is uses the fact that all offsets in my tiny example are in close proximity to each other. So the jump
table consisting of single bytes.

f14:
; input value in W0
sub w0, w0, #1
cmp w0, 21
; branch if less or equal (unsigned):
bls .L9
.L2:
; print "default":
adrp x0, .LC4
b puts
.L9:
; load jumptable address to X1:
adrp x1, .L4
add x1, x1, :lo12:.L4
; W0=input_value-1
; load byte from the table:
ldrb w0, [x1,w0,uxtw]
; load address of the Lrtx label:
adr x1, .Lrtx4
; multiply table element by 4 (by shifting 2 bits left) and add (or subtract) to the address of
Lrtx:
add x0, x1, w0, sxtb #2
; jump to the calculated address:
br x0
; this label is pointing in code (text) segment:
.Lrtx4:
.section .rodata
; everything after ".section" statement is allocated in the read-only data (rodata) segment:
.L4:
.byte (.L3 - .Lrtx4) / 4 ; case 1
.byte (.L3 - .Lrtx4) / 4 ; case 2
.byte (.L5 - .Lrtx4) / 4 ; case 3
.byte (.L5 - .Lrtx4) / 4 ; case 4
.byte (.L5 - .Lrtx4) / 4 ; case 5
.byte (.L5 - .Lrtx4) / 4 ; case 6
.byte (.L3 - .Lrtx4) / 4 ; case 7
.byte (.L6 - .Lrtx4) / 4 ; case 8
.byte (.L6 - .Lrtx4) / 4 ; case 9
.byte (.L3 - .Lrtx4) / 4 ; case 10
.byte (.L2 - .Lrtx4) / 4 ; case 11
.byte (.L2 - .Lrtx4) / 4 ; case 12
.byte (.L2 - .Lrtx4) / 4 ; case 13
.byte (.L2 - .Lrtx4) / 4 ; case 14
.byte (.L2 - .Lrtx4) / 4 ; case 15
.byte (.L2 - .Lrtx4) / 4 ; case 16
.byte (.L2 - .Lrtx4) / 4 ; case 17
180
.byte (.L2 - .Lrtx4) / 4 ; case 18
.byte (.L2 - .Lrtx4) / 4 ; case 19
.byte (.L6 - .Lrtx4) / 4 ; case 20
.byte (.L6 - .Lrtx4) / 4 ; case 21
.byte (.L7 - .Lrtx4) / 4 ; case 22
.text
; everything after ".text" statement is allocated in the code (text) segment:
.L7:
; print "22"
adrp x0, .LC3
b puts
.L6:
; print "8, 9, 21"
adrp x0, .LC2
b puts
.L5:
; print "3, 4, 5"
adrp x0, .LC1
b puts
.L3:
; print "1, 2, 7, 10"
adrp x0, .LC0
b puts
.LC0:
.string "1, 2, 7, 10"
.LC1:
.string "3, 4, 5"
.LC2:
.string "8, 9, 21"
.LC3:
.string "22"
.LC4:
.string "default"
Let’s compile this example to object file and open it in IDA. Here is the jump table:
Listing 1.166: jumptable in IDA
.rodata:0000000000000064 AREA .rodata, DATA, READONLY
.rodata:0000000000000064 ; ORG 0x64
.rodata:0000000000000064 $d DCB 9 ; case 1
.rodata:0000000000000065 DCB 9 ; case 2
.rodata:0000000000000066 DCB 6 ; case 3
.rodata:0000000000000067 DCB 6 ; case 4
.rodata:0000000000000068 DCB 6 ; case 5
.rodata:0000000000000069 DCB 6 ; case 6
.rodata:000000000000006A DCB 9 ; case 7
.rodata:000000000000006B DCB 3 ; case 8
.rodata:000000000000006C DCB 3 ; case 9
.rodata:000000000000006D DCB 9 ; case 10
.rodata:000000000000006E DCB 0xF7 ; case 11
.rodata:000000000000006F DCB 0xF7 ; case 12
.rodata:0000000000000070 DCB 0xF7 ; case 13
.rodata:0000000000000077 DCB 3 ; case 20
.rodata:0000000000000078 DCB 3 ; case 21
.rodata:0000000000000079 DCB 0 ; case 22
.rodata:000000000000007B ; .rodata ends
So in case of 1, 9 is to be multiplied by 4 and added to the address of Lrtx4 label.

In case of 22, 0 is to be multiplied by 4, resulting in 0.
181
Right after the Lrtx4 label is the L7 label, where you can find the code that prints “22”.
There is no jump table in the code segment, it’s allocated in a separate .rodata section (there is no special
necessity to place it in the code section).
There are also negative bytes (0xF7), they are used for jumping back to the code that prints the “default”
string (at .L2).
1.21.4 Fall-through
Another popular usage of switch() operator is so-called “fallthrough”. Here is simple example98 :
1 bool is_whitespace(char c) {
2 switch (c) {
3 case ' ': // fallthrough
4 case '\t': // fallthrough
5 case '\r': // fallthrough
6 case '\n':
7 return true;
8 default: // not whitespace
9 return false;
10 }
11 }
Slightly harder, from Linux kernel99 :

1 char nco1, nco2;
2
3 void f(int if_freq_khz)
4 {
5
6 switch (if_freq_khz) {
7 default:
8 printf("IF=%d KHz is not supportted, 3250 assumed\n", if_freq_khz);
9 /* fallthrough */
10 case 3250: /* 3.25Mhz */
11 nco1 = 0x34;
12 nco2 = 0x00;
13 break;
14 case 3500: /* 3.50Mhz */
15 nco1 = 0x38;
16 nco2 = 0x00;
17 break;
18 case 4000: /* 4.00Mhz */
19 nco1 = 0x40;
20 nco2 = 0x00;
21 break;
22 case 5000: /* 5.00Mhz */
23 nco1 = 0x50;
24 nco2 = 0x00;
25 break;
26 case 5380: /* 5.38Mhz */
27 nco1 = 0x56;
28 nco2 = 0x14;
29 break;
30 }
31 };

1 .LC0:
2 .string "IF=%d KHz is not supportted, 3250 assumed\n"
3 f:
4 sub esp, 12
5 mov eax, DWORD PTR [esp+16]
6 cmp eax, 4000
7 je .L3
98 Copypasted from https://github.com/azonalon/prgraas/blob/master/prog1lib/lecture_examples/is_whitespace.c
99 Copypasted from https://github.com/torvalds/linux/blob/master/drivers/media/dvb-frontends/lgdt3306a.c
182
8 jg .L4
9 cmp eax, 3250
10 je .L5
11 cmp eax, 3500
12 jne .L2
13 mov BYTE PTR nco1, 56
15 add esp, 12
16 ret
17 .L4:
18 cmp eax, 5000
19 je .L7
20 cmp eax, 5380
21 jne .L2
24 add esp, 12
25 ret
26 .L2:
27 sub esp, 8
28 push eax
29 push OFFSET FLAT:.LC0
30 call printf
31 add esp, 16
32 .L5:
35 add esp, 12
36 ret
37 .L3:
40 add esp, 12
41 ret
42 .L7:
45 add esp, 12
46 ret
We can get to .L5 label if there is number 3250 at function’s input. But we can get to this label from the
other side: we see that there are no jumps between printf() call and .L5 label.
Now we can understand why switch() statement is sometimes a source of bugs: one forgotten break will
transform your switch() statement into fallthrough one, and several blocks will be executed instead of
single one.
1.21.5 Exercises
Exercise#1
It’s possible to rework the C example in 1.21.2 on page 166 in such way that the compiler can produce
even smaller code, but will work just the same. Try to achieve it.
1.22 Loops
x86
There is a special LOOP instruction in x86 instruction set for checking the value in register ECX and if it is
not 0, to decrement ECX and pass control flow to the label in the LOOP operand. Probably this instruction
is not very convenient, and there are no any modern compilers which emit it automatically. So, if you see
this instruction somewhere in code, it is most likely that this is a manually written piece of assembly code.
In C/C++ loops are usually constructed using for(), while() or do/while() statements.
Let’s start with for().
183
This statement defines loop initialization (set loop counter to initial value), loop condition (is the counter
bigger than a limit?), what is performed at each iteration (increment/decrement) and of course loop body.
for (initialization; condition; at each iteration)
{
loop_body;
}
The generated code is consisting of four parts as well.

Let’s start with a simple example:
#include <stdio.h>
void printing_function(int i)
{
printf ("f(%d)\n", i);
};
int main()
{
int i;
for (i=2; i<10; i++)

printing_function(i);
return 0;
};
The result (MSVC 2010):

_i$ = -4
_main PROC
push ebp
mov ebp, esp
push ecx
mov DWORD PTR _i$[ebp], 2 ; loop initialization
jmp SHORT $LN3@main
$LN2@main:
mov eax, DWORD PTR _i$[ebp] ; here is what we do after each iteration:
add eax, 1 ; add 1 to (i) value
mov DWORD PTR _i$[ebp], eax
$LN3@main:
cmp DWORD PTR _i$[ebp], 10 ; this condition is checked before each iteration
jge SHORT $LN1@main ; if (i) is biggest or equals to 10, lets finish loop
mov ecx, DWORD PTR _i$[ebp] ; loop body: call printing_function(i)
push ecx
call _printing_function
add esp, 4
jmp SHORT $LN2@main ; jump to loop begin
$LN1@main: ; loop end
xor eax, eax
mov esp, ebp
pop ebp
ret 0
_main ENDP
As we see, nothing special.

GCC 4.4.1 emits almost the same code, with one subtle difference:

main proc near

push ebp
mov ebp, esp
184
and esp, 0FFFFFFF0h
sub esp, 20h
mov [esp+20h+var_4], 2 ; (i) initializing
jmp short loc_8048476
loc_8048465:
mov eax, [esp+20h+var_4]
call printing_function
add [esp+20h+var_4], 1 ; (i) increment
loc_8048476:
cmp [esp+20h+var_4], 9
jle short loc_8048465 ; if i<=9, continue loop
mov eax, 0
leave
retn
main endp
Now let’s see what we get with optimization turned on (/Ox):
Listing 1.170: Optimizing MSVC

_main PROC
push esi
mov esi, 2
$LL3@main:
push esi
call _printing_function
inc esi
add esp, 4
cmp esi, 10 ; 0000000aH
jl SHORT $LL3@main
xor eax, eax
pop esi
ret 0
_main ENDP
What happens here is that space for the i variable is not allocated in the local stack anymore, but uses
an individual register for it, ESI. This is possible in such small functions where there aren’t many local
variables.
One very important thing is that the f() function must not change the value in ESI. Our compiler is sure
here. And if the compiler decides to use the ESI register in f() too, its value would have to be saved
at the function’s prologue and restored at the function’s epilogue, almost like in our listing: please note
PUSH ESI/POP ESI at the function start and end.
Let’s try GCC 4.4.1 with maximal optimization turned on (-O3 option):

main proc near
push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
sub esp, 10h
185
xor eax, eax
leave
retn
main endp
Huh, GCC just unwound our loop.

Loop unwinding has an advantage in the cases when there aren’t much iterations and we could cut some
execution time by removing all loop support instructions. On the other side, the resulting code is obviously
larger.
Big unrolled loops are not recommended in modern times, because bigger functions may require bigger
cache footprint100 .
OK, let’s increase the maximum value of the i variable to 100 and try again. GCC does:
Listing 1.172: GCC

public main
main proc near
push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
push ebx
mov ebx, 2 ; i=2
sub esp, 1Ch
; aligning label loc_80484D0 (loop body begin) by 16-byte border:

nop
loc_80484D0:
; pass (i) as first argument to printing_function():
mov [esp+20h+var_20], ebx
add ebx, 1 ; i++
cmp ebx, 64h ; i==100?
jnz short loc_80484D0 ; if not, continue
add esp, 1Ch
xor eax, eax ; return 0
pop ebx
mov esp, ebp
pop ebp
retn
main endp
It is quite similar to what MSVC 2010 with optimization (/Ox) produce, with the exception that the EBX
register is allocated for the i variable.
GCC is sure this register will not be modified inside of the f() function, and if it will, it will be saved at the
function prologue and restored at epilogue, just like here in the main() function.
100 Avery good article about it: [Ulrich Drepper, What Every Programmer Should Know About Memory, (2007)]101 . Another
recommendations about loop unrolling from Intel are here: [Intel® 64 and IA-32 Architectures Optimization Reference Manual,
(2014)3.4.1.7].
186
x86: OllyDbg
Let’s compile our example in MSVC 2010 with /Ox and /Ob0 options and load it into OllyDbg.
It seems that OllyDbg is able to detect simple loops and show them in square brackets, for convenience:
Figure 1.55: OllyDbg: main() begin
By tracing (F8 — step over) we see ESI incrementing. Here, for instance, ESI = i = 6:
Figure 1.56: OllyDbg: loop body just executed with i = 6
9 is the last loop value. That’s why JL is not triggering after the increment, and the function will finish:
187
Figure 1.57: OllyDbg: ESI = 10, loop end
x86: tracer
As we might see, it is not very convenient to trace manually in the debugger. That’s a reason we will try
tracer.
We open compiled example in IDA, find the address of the instruction PUSH ESI (passing the sole argument
to f()), which is 0x401026 for this case and we run the tracer:
tracer.exe -l:loops_2.exe bpx=loops_2.exe!0x00401026
BPX just sets a breakpoint at the address and tracer will then print the state of the registers.
In the tracer.log, this is what we see:
PID=12884|New process loops_2.exe
(0) loops_2.exe!0x401026
EAX=0x00a328c8 EBX=0x00000000 ECX=0x6f0f4714 EDX=0x00000000
ESI=0x00000002 EDI=0x00333378 EBP=0x0024fbfc ESP=0x0024fbb8
EIP=0x00331026
FLAGS=PF ZF IF
(0) loops_2.exe!0x401026
EAX=0x00000005 EBX=0x00000000 ECX=0x6f0a5617 EDX=0x000ee188
EIP=0x00331026
FLAGS=CF PF AF SF IF
(0) loops_2.exe!0x401026
EIP=0x00331026
(0) loops_2.exe!0x401026
EIP=0x00331026
FLAGS=CF AF SF IF
(0) loops_2.exe!0x401026
EIP=0x00331026
(0) loops_2.exe!0x401026
EIP=0x00331026
FLAGS=CF AF SF IF
(0) loops_2.exe!0x401026
EIP=0x00331026
188
FLAGS=CF AF SF IF
(0) loops_2.exe!0x401026
EIP=0x00331026
PID=12884|Process loops_2.exe exited. ExitCode=0 (0x0)
We see how the value of ESI register changes from 2 to 9.

Even more than that, the tracer can collect register values for all addresses within the function. This is
called trace there. Every instruction gets traced, all interesting register values are recorded.
Then, an IDA.idc-script is generated, that adds comments. So, in the IDA we’ve learned that the main()
function address is 0x00401020 and we run:
tracer.exe -l:loops_2.exe bpf=loops_2.exe!0x00401020,trace:cc
BPF stands for set breakpoint on function.

As a result, we get the loops_2.exe.idc and loops_2.exe_clear.idc scripts.
189
We load loops_2.exe.idc into IDA and see:
Figure 1.58: IDA with .idc-script loaded
We see that ESI can be from 2 to 9 at the start of the loop body, but from 3 to 0xA (10) after the increment.
We can also see that main() is finishing with 0 in EAX.
tracer also generates loops_2.exe.txt, that contains information about how many times each instruction
has been executed and register values:
Listing 1.173: loops_2.exe.txt

0x401020 (.text+0x20), e= 1 [PUSH ESI] ESI=1
0x401021 (.text+0x21), e= 1 [MOV ESI, 2]
0x401026 (.text+0x26), e= 8 [PUSH ESI] ESI=2..9
0x401027 (.text+0x27), e= 8 [CALL 8D1000h] tracing nested maximum level (1) reached, ⤦
Ç skipping this CALL 8D1000h=0x8d1000
0x40102c (.text+0x2c), e= 8 [INC ESI] ESI=2..9
0x40102d (.text+0x2d), e= 8 [ADD ESP, 4] ESP=0x38fcbc
0x401030 (.text+0x30), e= 8 [CMP ESI, 0Ah] ESI=3..0xa
0x401033 (.text+0x33), e= 8 [JL 8D1026h] SF=false,true OF=false
0x401035 (.text+0x35), e= 1 [XOR EAX, EAX]
0x401037 (.text+0x37), e= 1 [POP ESI]
0x401038 (.text+0x38), e= 1 [RETN] EAX=0
We can use grep here.
ARM
main
STMFD SP!, {R4,LR}
MOV R4, #2
B loc_368
loc_35C ; CODE XREF: main+1C
MOV R0, R4
BL printing_function
ADD R4, R4, #1
loc_368 ; CODE XREF: main+8

CMP R4, #0xA
190
BLT loc_35C
MOV R0, #0
LDMFD SP!, {R4,PC}
Iteration counter i is to be stored in the R4 register. The MOV R4, #2 instruction just initializes i. The
MOV R0, R4 and BL printing_function instructions compose the body of the loop, the first instruction
preparing the argument for f() function and the second calling the function. The ADD R4, R4, #1 in-
struction just adds 1 to the i variable at each iteration. CMP R4, #0xA compares i with 0xA (10). The
next instruction BLT (Branch Less Than) jumps if i is less than 10. Otherwise, 0 is to be written into R0
(since our function returns 0) and function execution finishes.
_main
PUSH {R4,LR}
MOVS R4, #2
loc_132 ; CODE XREF: _main+E

MOVS R0, R4
BL printing_function
ADDS R4, R4, #1
CMP R4, #0xA
BLT loc_132
MOVS R0, #0
POP {R4,PC}
Practically the same.
_main
PUSH {R4,R7,LR}
MOVW R4, #0x1124 ; "%d\n"
MOVS R1, #2
MOVT.W R4, #0
ADD R7, SP, #4
ADD R4, PC
MOV R0, R4
BLX _printf
MOV R0, R4
MOVS R1, #3
BLX _printf
MOV R0, R4
MOVS R1, #4
BLX _printf
MOV R0, R4
MOVS R1, #5
BLX _printf
MOV R0, R4
MOVS R1, #6
BLX _printf
MOV R0, R4
MOVS R1, #7
BLX _printf
MOV R0, R4
MOVS R1, #8
BLX _printf
MOV R0, R4
MOVS R1, #9
BLX _printf
MOVS R0, #0
POP {R4,R7,PC}
In fact, this was in my f() function:
191
void printing_function(int i)
{
printf ("%d\n", i);
};
So, LLVM not just unrolled the loop, but also inlined my very simple function f(), and inserted its body 8
times instead of calling it.
This is possible when the function is so simple (like mine) and when it is not called too much (like here).
ARM64: Optimizing GCC 4.9.1

printing_function:
; prepare second argument of printf():
mov w1, w0
; load address of the "f(%d)\n" string
adrp x0, .LC0
; just branch here instead of branch with link and return:
b printf
main:
; save FP and LR in the local stack:
stp x29, x30, [sp, -32]!
; set up stack frame:
add x29, sp, 0
; save contents of X19 register in the local stack:
str x19, [sp,16]
; we will use W19 register as counter.
; set initial value of 2 to it:
mov w19, 2
.L3:
; prepare first argument of printing_function():
mov w0, w19
; increment counter register.
add w19, w19, 1
; W0 here still holds value of counter value before increment.
bl printing_function
; is it end?
cmp w19, 10
; no, jump to the loop body begin:
bne .L3
; return 0
mov w0, 0
; restore contents of X19 register:
ldr x19, [sp,16]
; restore FP and LR values:
ldp x29, x30, [sp], 32
ret
.LC0:
.string "f(%d)\n"
ARM64: Non-optimizing GCC 4.9.1
Listing 1.175: Non-optimizing GCC 4.9.1 -fno-inline

.LC0:
.string "f(%d)\n"
printing_function:
stp x29, x30, [sp, -32]!
add x29, sp, 0
; save contents of W0 register:
str w0, [x29,28]
192
; load address of the "f(%d)\n" string
adrp x0, .LC0
; reload input value from the local stack to W1 register:
ldr w1, [x29,28]
; call printf()
bl printf
ldp x29, x30, [sp], 32
ret
main:
stp x29, x30, [sp, -32]!
add x29, sp, 0
; initialize counter
mov w0, 2
; store it to the place allocated for it in the local stack:
str w0, [x29,28]
; skip loop body and jump to the loop condition check instructions:
b .L3
.L4:
; load counter value to W0.
; it will be the first argument of printing_function():
ldr w0, [x29,28]
; call printing_function():
bl printing_function
; increment counter value:
ldr w0, [x29,28]
add w0, w0, 1
str w0, [x29,28]
.L3:
; loop condition check.
; load counter value:
ldr w0, [x29,28]
; is it 9?
cmp w0, 9
; less or equal? then jump to loop body begin:
; do nothing otherwise.
ble .L4
; return 0
mov w0, 0
ldp x29, x30, [sp], 32
ret
MIPS

main:
; IDA is not aware of variable names in local stack

; We gave them names manually:
i = -0x10
saved_FP = -8
saved_RA = -4
addiu $sp, -0x28
sw $ra, 0x28+saved_RA($sp)
sw $fp, 0x28+saved_FP($sp)
move $fp, $sp
; initialize counter at 2 and store this value in local stack
li $v0, 2
sw $v0, 0x28+i($fp)
; pseudoinstruction. "BEQ $ZERO, $ZERO, loc_9C" there in fact:
b loc_9C
193
loc_80: # CODE XREF: main+48
; load counter value from local stack and call printing_function():
lw $a0, 0x28+i($fp)
jal printing_function
; load counter, increment it, store it back:
lw $v0, 0x28+i($fp)
or $at, $zero ; NOP
addiu $v0, 1
sw $v0, 0x28+i($fp)
loc_9C: # CODE XREF: main+18

; check counter, is it 10?
lw $v0, 0x28+i($fp)
or $at, $zero ; NOP
slti $v0, 0xA
; if it is less than 10, jump to loc_80 (loop body begin):
bnez $v0, loc_80
; finishing, return 0:
move $v0, $zero
move $sp, $fp
lw $ra, 0x28+saved_RA($sp)
lw $fp, 0x28+saved_FP($sp)
addiu $sp, 0x28
jr $ra
The instruction that’s new to us is B. It is actually the pseudo instruction (BEQ).
One more thing
In the generated code we can see: after initializing i, the body of the loop is not to be executed, as the
condition for i is checked first, and only after that loop body can be executed. And that is correct.
Because, if the loop condition is not met at the beginning, the body of the loop must not be executed.
This is possible in the following case:
for (i=0; i<total_entries_to_process; i++)
loop_body;
If total_entries_to_process is 0, the body of the loop must not be executed at all.

This is why the condition checked before the execution.
However, an optimizing compiler may swap the condition check and loop body, if it sure that the situation
described here is not possible (like in the case of our very simple example and using compilers like Keil,
Xcode (LLVM), MSVC in optimization mode).
1.22.2 Memory blocks copying routine

Real-world memory copy routines may copy 4 or 8 bytes at each iteration, use SIMD102 , vectorization, etc.
But for the sake of simplicity, this example is the simplest possible.
#include <stdio.h>
void my_memcpy (unsigned char* dst, unsigned char* src, size_t cnt)
{
size_t i;
for (i=0; i<cnt; i++)
dst[i]=src[i];
};
102 Single Instruction, Multiple Data
194
Straight-forward implementation
Listing 1.177: GCC 4.9 x64 optimized for size (-Os)

my_memcpy:
; RDI = destination address
; RSI = source address
; RDX = size of block
; initialize counter (i) at 0

xor eax, eax
.L2:
; all bytes copied? exit then:
cmp rax, rdx
je .L5
; load byte at RSI+i:
mov cl, BYTE PTR [rsi+rax]
; store byte at RDI+i:
mov BYTE PTR [rdi+rax], cl
inc rax ; i++
jmp .L2
.L5:
ret
Listing 1.178: GCC 4.9 ARM64 optimized for size (-Os)

my_memcpy:
; X0 = destination address
; X1 = source address
; X2 = size of block

mov x3, 0
.L2:
; all bytes copied? exit then:
cmp x3, x2
beq .L5
; load byte at X1+i:
ldrb w4, [x1,x3]
; store byte at X0+i:
strb w4, [x0,x3]
add x3, x3, 1 ; i++
b .L2
.L5:
ret

my_memcpy PROC
; R0 = destination address
; R1 = source address
; R2 = size of block
PUSH {r4,lr}
MOVS r3,#0
; condition checked at the end of function, so jump there:
B |L0.12|
|L0.6|
; load byte at R1+i:
LDRB r4,[r1,r3]
; store byte at R0+i:
STRB r4,[r0,r3]
; i++
ADDS r3,r3,#1
|L0.12|
; i<size?
CMP r3,r2
; jump to the loop begin if its so:
BCC |L0.6|
POP {r4,pc}
195
ENDP
ARM in ARM mode
Keil in ARM mode takes full advantage of conditional suffixes:

my_memcpy PROC
; R0 = destination address
; R1 = source address
; R2 = size of block

MOV r3,#0
|L0.4|
; all bytes copied?
CMP r3,r2
; the following block is executed only if less than condition,
; i.e., if R2<R3 or i<size.
; load byte at R1+i:
LDRBCC r12,[r1,r3]
; store byte at R0+i:
STRBCC r12,[r0,r3]
; i++
ADDCC r3,r3,#1
; the last instruction of the conditional block.
; jump to loop begin if i<size
; do nothing otherwise (i.e., if i>=size)
BCC |L0.4|
; return
BX lr
ENDP
That’s why there is only one branch instruction instead of 2.
MIPS
Listing 1.181: GCC 4.4.5 optimized for size (-Os) (IDA)

my_memcpy:
; jump to loop check part:
b loc_14
; it will always reside in $v0:
move $v0, $zero ; branch delay slot
loc_8: # CODE XREF: my_memcpy+1C

; load byte as unsigned at address in $t0 to $v1:
lbu $v1, 0($t0)
; increment counter (i):
addiu $v0, 1
; store byte at $a3
sb $v1, 0($a3)
loc_14: # CODE XREF: my_memcpy

; check if counter (i) in $v0 is still less then 3rd function argument ("cnt" in $a2):
sltu $v1, $v0, $a2
; form address of byte in source block:
addu $t0, $a1, $v0
; $t0 = $a1+$v0 = src+i
; jump to loop body if counter sill less then "cnt":
bnez $v1, loc_8
; form address of byte in destination block ($a3 = $a0+$v0 = dst+i):
addu $a3, $a0, $v0 ; branch delay slot
; finish if BNEZ wasnt triggered:
jr $ra
196
Here we have two new instructions: LBU (“Load Byte Unsigned”) and SB (“Store Byte”).
Just like in ARM, all MIPS registers are 32-bit wide, there are no byte-wide parts like in x86.
So when dealing with single bytes, we have to allocate whole 32-bit registers for them.
LBU loads a byte and clears all other bits (“Unsigned”).
On the other hand, LB (“Load Byte”) instruction sign-extends the loaded byte to a 32-bit value.
SB just writes a byte from lowest 8 bits of register to memory.
Vectorization
Optimizing GCC can do much more on this example: 1.36.1 on page 411.
1.22.3 Condition check

It’s important to keep in mind that in for() construct, condition is checked not at the end, but at the
beginning, before execution of loop body. But often, it’s more convenient for compiler to check it at the
end, after body. Sometimes, additional check can be appended at the beginning.
For example:
#include <stdio.h>
void f(int start, int finish)

{
for (; start<finish; start++)
printf ("%d\n", start);
};
Optimizing GCC 5.4.0 x64:

f:
; check condition (1):
cmp edi, esi
jge .L9
push rbp
push rbx
mov ebp, esi
mov ebx, edi
sub rsp, 8
.L5:
mov edx, ebx
xor eax, eax
mov esi, OFFSET FLAT:.LC0 ; "%d\n"
mov edi, 1
add ebx, 1
call __printf_chk
; check condition (2):
cmp ebp, ebx
jne .L5
add rsp, 8
pop rbx
pop rbp
.L9:
rep ret
We see two checks.

Hex-Rays (at least version 2.2.0) decompiles this as:
void __cdecl f(unsigned int start, unsigned int finish)
{
unsigned int v2; // ebx@2
__int64 v3; // rdx@3
if ( (signed int)start < (signed int)finish )

{
v2 = start;
197
do
{
v3 = v2++;
_printf_chk(1LL, "%d\n", v3);
}
while ( finish != v2 );
}
}
In this case, do/while() can be replaced by for() without any doubt, and the first check can be removed.
1.22.4 Conclusion
Rough skeleton of loop from 2 to 9 inclusive:
Listing 1.182: x86

mov [counter], 2 ; initialization
jmp check
body:
; loop body
; do something here
; use counter variable in local stack
add [counter], 1 ; increment
check:
cmp [counter], 9
jle body
The increment operation may be represented as 3 instructions in non-optimized code:
Listing 1.183: x86

MOV [counter], 2 ; initialization
JMP check
body:
; loop body
; do something here
MOV REG, [counter] ; increment
INC REG
MOV [counter], REG
check:
CMP [counter], 9
JLE body
If the body of the loop is short, a whole register can be dedicated to the counter variable:
Listing 1.184: x86

MOV EBX, 2 ; initialization
JMP check
body:
; loop body
; do something here
; use counter in EBX, but do not modify it!
INC EBX ; increment
check:
CMP EBX, 9
JLE body
Some parts of the loop may be generated by compiler in different order:
Listing 1.185: x86

MOV [counter], 2 ; initialization
JMP label_check
label_increment:
ADD [counter], 1 ; increment
label_check:
CMP [counter], 10
198
JGE exit
; loop body
; do something here
JMP label_increment
exit:
Usually the condition is checked before loop body, but the compiler may rearrange it in a way that the
condition is checked after loop body.
This is done when the compiler is sure that the condition is always true on the first iteration, so the body
of the loop is to be executed at least once:
Listing 1.186: x86

MOV REG, 2 ; initialization
body:
; loop body
; do something here
; use counter in REG, but do not modify it!
INC REG ; increment
CMP REG, 10
JL body
Using the LOOP instruction. This is rare, compilers are not using it. When you see it, it’s a sign that this
piece of code is hand-written:
Listing 1.187: x86

; count from 10 to 1
MOV ECX, 10
body:
; loop body
; do something here
; use counter in ECX, but do not modify it!
LOOP body
ARM.
The R4 register is dedicated to counter variable in this example:
Listing 1.188: ARM

MOV R4, 2 ; initialization
B check
body:
; loop body
; do something here
; use counter in R4, but do not modify it!
ADD R4,R4, #1 ; increment
check:
CMP R4, #10
BLT body
1.22.5 Exercises
199
1.23 More about strings
1.23.1 strlen()
103
Let’s talk about loops one more time. Often, the strlen() function is implemented using a while()
statement. Here is how it is done in the MSVC standard libraries:
int my_strlen (const char * str)
{
const char *eos = str;
while( *eos++ ) ;
return( eos - str - 1 );

}
int main()
{
// test
return my_strlen("hello!");
};
x86
Non-optimizing MSVC
Let’s compile:
_eos$ = -4 ; size = 4
_str$ = 8 ; size = 4
_strlen PROC
push ebp
mov ebp, esp
push ecx
mov eax, DWORD PTR _str$[ebp] ; place pointer to string from "str"
mov DWORD PTR _eos$[ebp], eax ; place it to local variable "eos"
$LN2@strlen_:
mov ecx, DWORD PTR _eos$[ebp] ; ECX=eos
; take 8-bit byte from address in ECX and place it as 32-bit value to EDX with sign
extension
movsx edx, BYTE PTR [ecx]

mov eax, DWORD PTR _eos$[ebp] ; EAX=eos
add eax, 1 ; increment EAX
mov DWORD PTR _eos$[ebp], eax ; place EAX back to "eos"
test edx, edx ; EDX is zero?
je SHORT $LN1@strlen_ ; yes, then finish loop
jmp SHORT $LN2@strlen_ ; continue loop
$LN1@strlen_:
; here we calculate the difference between two pointers
mov eax, DWORD PTR _eos$[ebp]

sub eax, DWORD PTR _str$[ebp]
sub eax, 1 ; subtract 1 and return result
mov esp, ebp
pop ebp
ret 0
_strlen_ ENDP
We get two new instructions here: MOVSX and TEST.

The first one—MOVSX—takes a byte from an address in memory and stores the value in a 32-bit register.
MOVSX stands for MOV with Sign-Extend. MOVSX sets the rest of the bits, from the 8th to the 31th, to 1 if
the source byte is negative or to 0 if is positive.
103 counting the characters in a string in the C language
200
And here is why.
By default, the char type is signed in MSVC and GCC. If we have two values of which one is char and the
other is int, (int is signed too), and if the first value contain -2 (coded as 0xFE) and we just copy this byte
into the int container, it makes 0x000000FE, and this from the point of signed int view is 254, but not -2.
In signed int, -2 is coded as 0xFFFFFFFE. So if we have to transfer 0xFE from a variable of char type to int,
we have to identify its sign and extend it. That is what MOVSX does.
You can also read about it in “Signed number representations” section ( 2.2 on page 448).
It’s hard to say if the compiler needs to store a char variable in EDX, it could just take a 8-bit register part
(for example DL). Apparently, the compiler’s register allocator works like that.
Then we see TEST EDX, EDX. You can read more about the TEST instruction in the section about bit
fields ( 1.28 on page 303). Here this instruction just checks if the value in EDX equals to 0.
Non-optimizing GCC
Let’s try GCC 4.4.1:

public strlen
strlen proc near
eos = dword ptr -4

arg_0 = dword ptr 8
push ebp
mov ebp, esp
sub esp, 10h
mov [ebp+eos], eax
loc_80483F0:
mov eax, [ebp+eos]
movzx eax, byte ptr [eax]
test al, al
setnz al
add [ebp+eos], 1
test al, al
jnz short loc_80483F0
mov edx, [ebp+eos]
mov ecx, edx
sub ecx, eax
mov eax, ecx
sub eax, 1
leave
retn
strlen endp
The result is almost the same as in MSVC, but here we see MOVZX instead of MOVSX. MOVZX stands for MOV
with Zero-Extend. This instruction copies a 8-bit or 16-bit value into a 32-bit register and sets the rest of
the bits to 0. In fact, this instruction is convenient only because it enable us to replace this instruction
pair:
xor eax, eax / mov al, [...].
On the other hand, it is obvious that the compiler could produce this code:
mov al, byte ptr [eax] / test al, al—it is almost the same, however, the highest bits of the EAX
register will contain random noise. But let’s think it is compiler’s drawback—it cannot produce more
understandable code. Strictly speaking, the compiler is not obliged to emit understandable (to humans)
code at all.
The next new instruction for us is SETNZ. Here, if AL doesn’t contain zero, test al, al sets the ZF flag
to 0, but SETNZ, if ZF==0 (NZ stands for not zero) sets AL to 1. Speaking in natural language, if AL is not
zero, let’s jump to loc_80483F0. The compiler emits some redundant code, but let’s not forget that the
optimizations are turned off.
201
Optimizing MSVC
Now let’s compile all this in MSVC 2012, with optimizations turned on (/Ox):
Listing 1.189: Optimizing MSVC 2012 /Ob0

_str$ = 8 ; size = 4
_strlen PROC
mov edx, DWORD PTR _str$[esp-4] ; EDX -> pointer to the string
mov eax, edx ; move to EAX
$LL2@strlen:
mov cl, BYTE PTR [eax] ; CL = *EAX
inc eax ; EAX++
test cl, cl ; CL==0?
jne SHORT $LL2@strlen ; no, continue loop
sub eax, edx ; calculate pointers difference
dec eax ; decrement EAX
ret 0
_strlen ENDP
Now it is all simpler. Needless to say, the compiler could use registers with such efficiency only in small
functions with a few local variables.
INC/DEC—are increment/decrement instructions, in other words: add or subtract 1 to/from a variable.
202
Optimizing MSVC + OllyDbg
We can try this (optimized) example in OllyDbg. Here is the first iteration:
Figure 1.59: OllyDbg: first iteration start
We see that OllyDbg found a loop and, for convenience, wrapped its instructions in brackets. By clicking
the right button on EAX, we can choose “Follow in Dump” and the memory window scrolls to the right
place. Here we can see the string “hello!” in memory. There is at least one zero byte after it and then
random garbage.
If OllyDbg sees a register with a valid address in it, that points to some string, it is shown as a string.
203
Let’s press F8 (step over) a few times, to get to the start of the body of the loop:
Figure 1.60: OllyDbg: second iteration start
We see that EAX contains the address of the second character in the string.
204
We have to press F8 enough number of times in order to escape from the loop:
Figure 1.61: OllyDbg: pointers difference to be calculated now
We see that EAX now contains the address of zero byte that’s right after the string plus 1 (because INC
EAX was executed regardless of whether we exit from the loop or not). Meanwhile, EDX hasn’t changed,
so it still pointing to the start of the string.
The difference between these two addresses is being calculated now.
205
The SUB instruction just got executed:
Figure 1.62: OllyDbg: EAX to be decremented now
The difference of pointers is in the EAX register now—7. Indeed, the length of the “hello!” string is 6, but
with the zero byte included—7. But strlen() must return the number of non-zero characters in the string.
So the decrement executes and then the function returns.
Optimizing GCC
Let’s check GCC 4.4.1 with optimizations turned on (-O3 key):

public strlen
strlen proc near
arg_0 = dword ptr 8
push ebp
mov ebp, esp
mov ecx, [ebp+arg_0]
mov eax, ecx
loc_8048418:
movzx edx, byte ptr [eax]
add eax, 1
test dl, dl
jnz short loc_8048418
not ecx
add eax, ecx
pop ebp
retn
strlen endp
Here GCC is almost the same as MSVC, except for the presence of MOVZX. However, here MOVZX could be
replaced with
mov dl, byte ptr [eax].
Perhaps it is simpler for GCC’s code generator to remember the whole 32-bit EDX register is allocated for
a char variable and it then can be sure that the highest bits has no any noise at any point.
206
After that we also see a new instruction—NOT. This instruction inverts all bits in the operand.
You can say that it is a synonym to the XOR ECX, 0ffffffffh instruction. NOT and the following ADD
calculate the pointer difference and subtract 1, just in a different way. At the start ECX, where the pointer
to str is stored, gets inverted and 1 is subtracted from it.
See also: “Signed number representations” ( 2.2 on page 448).
In other words, at the end of the function just after loop body, these operations are executed:
ecx=str;
eax=eos;
ecx=(-ecx)-1;
eax=eax+ecx
return eax
… and this is effectively equivalent to:

ecx=str;
eax=eos;
eax=eax-ecx;
eax=eax-1;
return eax
Why did GCC decide it would be better? Hard to guess. But perhaps the both variants are equivalent in
efficiency.
ARM
32-bit ARM
Non-optimizing Xcode 4.6.3 (LLVM) (ARM mode)
Listing 1.190: Non-optimizing Xcode 4.6.3 (LLVM) (ARM mode)

_strlen
eos = -8
str = -4
SUB SP, SP, #8 ; allocate 8 bytes for local variables

STR R0, [SP,#8+str]
LDR R0, [SP,#8+str]
STR R0, [SP,#8+eos]
loc_2CB8 ; CODE XREF: _strlen+28

LDR R0, [SP,#8+eos]
ADD R1, R0, #1
STR R1, [SP,#8+eos]
LDRSB R0, [R0]
CMP R0, #0
BEQ loc_2CD4
B loc_2CB8
loc_2CD4 ; CODE XREF: _strlen+24
LDR R0, [SP,#8+eos]
LDR R1, [SP,#8+str]
SUB R0, R0, R1 ; R0=eos-str
SUB R0, R0, #1 ; R0=R0-1
ADD SP, SP, #8 ; free allocated 8 bytes
BX LR
Non-optimizing LLVM generates too much code, however, here we can see how the function works with
local variables in the stack. There are only two local variables in our function: eos and str. In this listing,
generated by IDA, we have manually renamed var_8 and var_4 to eos and str.
The first instructions just saves the input values into both str and eos.
The body of the loop starts at label loc_2CB8.
207
The first three instruction in the loop body (LDR, ADD, STR) load the value of eos into R0. Then the value is
incremented and saved back into eos, which is located in the stack.
The next instruction, LDRSB R0, [R0] (“Load Register Signed Byte”), loads a byte from memory at the
address stored in R0 and sign-extends it to 32-bit 104 . This is similar to the MOVSX instruction in x86.
The compiler treats this byte as signed since the char type is signed according to the C standard. It was
already written about it ( 1.23.1 on page 200) in this section, in relation to x86.
It has to be noted that it is impossible to use 8- or 16-bit part of a 32-bit register in ARM separately of the
whole register, as it is in x86.
Apparently, it is because x86 has a huge history of backwards compatibility with its ancestors up to the
16-bit 8086 and even 8-bit 8080, but ARM was developed from scratch as a 32-bit RISC-processor.
Consequently, in order to process separate bytes in ARM, one has to use 32-bit registers anyway.
So, LDRSB loads bytes from the string into R0, one by one. The following CMP and BEQ instructions check
if the loaded byte is 0. If it’s not 0, control passes to the start of the body of the loop. And if it’s 0, the
loop ends.
At the end of the function, the difference between eos and str is calculated, 1 is subtracted from it, and
resulting value is returned via R0.
N.B. Registers were not saved in this function.
That’s because in the ARM calling convention registers R0-R3 are “scratch registers”, intended for argu-
ments passing, and we’re not required to restore their value when the function exits, since the calling
function will not use them anymore. Consequently, they may be used for anything we want.
No other registers are used here, so that is why we have nothing to save on the stack.
Thus, control may be returned back to calling function by a simple jump (BX), to the address in the LR
register.
Optimizing Xcode 4.6.3 (LLVM) (Thumb mode)
Listing 1.191: Optimizing Xcode 4.6.3 (LLVM) (Thumb mode)

_strlen
MOV R1, R0
loc_2DF6
LDRB.W R2, [R1],#1
CMP R2, #0
BNE loc_2DF6
MVNS R0, R0
ADD R0, R1
BX LR
As optimizing LLVM concludes, eos and str do not need space on the stack, and can always be stored in
registers.
Before the start of the loop body, str is always in R0, and eos—in R1.
The LDRB.W R2, [R1],#1 instruction loads a byte from the memory at the address stored in R1, to R2,
sign-extending it to a 32-bit value, but not just that. #1 at the instruction’s end is implies “Post-indexed
addressing”, which means that 1 is to be added to R1 after the byte is loaded. Read more about it: 1.39.2
on page 437.
Then you can see CMP and BNE105 in the body of the loop, these instructions continue looping until 0 is
found in the string.
MVNS106 (inverts all bits, like NOT in x86) and ADD instructions compute eos − str − 1. In fact, these two
instructions compute R0 = str + eos, which is effectively equivalent to what was in the source code, and
why it is so, was already explained here ( 1.23.1 on page 206).
Apparently, LLVM, just like GCC, concludes that this code can be shorter (or faster).
104 TheKeil compiler treats the char type as signed, just like MSVC and GCC.
105 (PowerPC, ARM) Branch if Not Equal
106 MoVe Not
208

_strlen
MOV R1, R0
loc_2C8
LDRB R2, [R1],#1
CMP R2, #0
SUBEQ R0, R1, R0
SUBEQ R0, R0, #1
BNE loc_2C8
BX LR
Almost the same as what we saw before, with the exception that the str−eos−1 expression can be computed
not at the function’s end, but right in the body of the loop. The -EQ suffix, as we may recall, implies that
the instruction executes only if the operands in the CMP that has been executed before were equal to each
other. Thus, if R0 contains 0, both SUBEQ instructions executes and result is left in the R0 register.
ARM64
my_strlen:
mov x1, x0
; X1 is now temporary pointer (eos), acting like cursor
.L58:
; load byte from X1 to W2, increment X1 (post-index)
ldrb w2, [x1],1
; Compare and Branch if NonZero: compare W2 with 0, jump to .L58 if it is not
cbnz w2, .L58
; calculate difference between initial pointer in X0 and current address in X1
sub x0, x1, x0
; decrement lowest 32-bit of result
sub w0, w0, #1
ret
The algorithm is the same as in 1.23.1 on page 202: find a zero byte, calculate the difference between
the pointers and decrement the result by 1. Some comments were added by the author of this book.
The only thing worth noting is that our example is somewhat wrong:
my_strlen() returns 32-bit int, while it has to return size_t or another 64-bit type.
The reason is that, theoretically, strlen() can be called for a huge blocks in memory that exceeds 4GB,
so it must able to return a 64-bit value on 64-bit platforms.
Because of my mistake, the last SUB instruction operates on a 32-bit part of register, while the penultimate
SUB instruction works on full the 64-bit register (it calculates the difference between the pointers).
It’s my mistake, it is better to leave it as is, as an example of how the code could look like in such case.
my_strlen:
; function prologue
sub sp, sp, #32
; first argument (str) will be stored in [sp,8]
str x0, [sp,8]
ldr x0, [sp,8]
; copy "str" to "eos" variable
str x0, [sp,24]
nop
.L62:
; eos++
209
ldr x0, [sp,24] ; load "eos" to X0
add x1, x0, 1 ; increment X0
str x1, [sp,24] ; save X0 to "eos"
; load byte from memory at address in X0 to W0
ldrb w0, [x0]
; is it zero? (WZR is the 32-bit register always contain zero)
cmp w0, wzr
; jump if not zero (Branch Not Equal)
bne .L62
; zero byte found. now calculate difference.
; load "eos" to X1
ldr x1, [sp,24]
; load "str" to X0
ldr x0, [sp,8]
; calculate difference
sub x0, x1, x0
; decrement result
sub w0, w0, #1
; function epilogue
add sp, sp, 32
ret
It’s more verbose. The variables are often tossed here to and from memory (local stack). The same
mistake here: the decrement operation happens on a 32-bit register part.
MIPS

my_strlen:
; "eos" variable will always reside in $v1:
move $v1, $a0
loc_4:
; load byte at address in "eos" into $a1:
lb $a1, 0($v1)
; if loaded byte is not zero, jump to loc_4:
bnez $a1, loc_4
; increment "eos" anyway:
addiu $v1, 1 ; branch delay slot
; loop finished. invert "str" variable:
nor $v0, $zero, $a0
; $v0=-str-1
jr $ra
; return value = $v1 + $v0 = eos + ( -str-1 ) = eos - str - 1
addu $v0, $v1, $v0 ; branch delay slot
MIPS lacks a NOT instruction, but has NOR which is OR + NOT operation.
This operation is widely used in digital electronics107 . For example, the Apollo Guidance Computer used
in the Apollo program, was built by only using 5600 NOR gates: [Jens Eickhoff, Onboard Computers,
Onboard Software and Satellite Operations: An Introduction, (2011)]. But NOR element isn’t very popular
in computer programming.
So, the NOT operation is implemented here as NOR DST, $ZERO, SRC.
From fundamentals 2.2 on page 448 we know that bitwise inverting a signed number is the same as
changing its sign and subtracting 1 from the result.
So what NOT does here is to take the value of str and transform it into −str − 1. The addition operation that
follows prepares result.
1.23.2 Boundaries of strings

It’s interesting to note, how parameters are passed into win32 GetOpenFileName() function. In order to
call it, one must set list of allowed file extensions:
107 NOR is called “universal gate”
210
OPENFILENAME *LPOPENFILENAME;
...
char * filter = "Text files (*.txt)\0*.txt\0MS Word files (*.doc)\0*.doc\0\0";
...
LPOPENFILENAME = (OPENFILENAME *)malloc(sizeof(OPENFILENAME));
...
LPOPENFILENAME->lpstrFilter = filter;
...
if(GetOpenFileName(LPOPENFILENAME))
{
...
What happens here is that list of strings are passed into GetOpenFileName(). It is not a problem to parse
it: whenever you encounter single zero byte, this is an item. Whenever you encounter two zero bytes,
this is end of the list. If you will pass this string into printf(), it will treat first item as a single string.
So this is string, or...? It’s better say this is buffer containing several zero-terminated C-strings, which can
be stored and processed as a whole.
Another example is strtok() function. It takes a string and write zero bytes in the middle of it. It thus
transforms input string into some kind of buffer, which has several zero-terminated C-strings.
1.24 Replacing arithmetic instructions to other ones

In the pursuit of optimization, one instruction may be replaced by another, or even with a group of instruc-
tions. For example, ADD and SUB can replace each other: line 18 in listing.3.122.
For example, the LEA instruction is often used for simple arithmetic calculations: .1.6 on page 1014.
1.24.1 Multiplication
Multiplication using addition
Here is a simple example:

unsigned int f(unsigned int a)
{
return a*8;
};
Multiplication by 8 is replaced by 3 addition instructions, which do the same. Apparently, MSVC’s optimizer
decided that this code can be faster.

_TEXT SEGMENT
_a$ = 8 ; size = 4
_f PROC
add eax, eax
add eax, eax
add eax, eax
ret 0
_f ENDP
_TEXT ENDS
END
Multiplication using shifting
Multiplication and division instructions by a numbers that’s a power of 2 are often replaced by shift in-
structions.
{
return a*4;
};
211
_a$ = 8 ; size = 4
_f PROC
push ebp
mov ebp, esp
shl eax, 2
pop ebp
ret 0
_f ENDP
Multiplication by 4 is just shifting the number to the left by 2 bits and inserting 2 zero bits at the right (as
the last two bits). It is just like multiplying 3 by 100 —we just have to add two zeros at the right.
That’s how the shift left instruction works:
7 6 5 4 3 2 1 0
CF 7 6 5 4 3 2 1 0 0
The added bits at right are always zeros.

Multiplication by 4 in ARM:

f PROC
LSL r0,r0,#2
BX lr
ENDP
Multiplication by 4 in MIPS:

jr $ra
sll $v0, $a0, 2 ; branch delay slot
SLL is “Shift Left Logical”.
Multiplication using shifting, subtracting, and adding
It’s still possible to get rid of the multiplication operation when you multiply by numbers like 7 or 17 again
by using shifting. The mathematics used here is relatively easy.
32-bit
#include <stdint.h>
int f1(int a)
{
return a*7;
};
int f2(int a)
{
return a*28;
};
int f3(int a)
{
return a*17;
};
212
x86

; a*7
_a$ = 8
_f1 PROC
; ECX=a
lea eax, DWORD PTR [ecx*8]
; EAX=ECX*8
sub eax, ecx
; EAX=EAX-ECX=ECX*8-ECX=ECX*7=a*7
ret 0
_f1 ENDP
; a*28
_a$ = 8
_f2 PROC
; ECX=a
lea eax, DWORD PTR [ecx*8]
; EAX=ECX*8
sub eax, ecx
; EAX=EAX-ECX=ECX*8-ECX=ECX*7=a*7
shl eax, 2
; EAX=EAX<<2=(a*7)*4=a*28
ret 0
_f2 ENDP
; a*17
_a$ = 8
_f3 PROC
; EAX=a
shl eax, 4
; EAX=EAX<<4=EAX*16=a*16
add eax, DWORD PTR _a$[esp-4]
; EAX=EAX+a=a*16+a=a*17
ret 0
_f3 ENDP
ARM
Keil for ARM mode takes advantage of the second operand’s shift modifiers:

; a*7
||f1|| PROC
RSB r0,r0,r0,LSL #3
; R0=R0<<3-R0=R0*8-R0=a*8-a=a*7
BX lr
ENDP
; a*28
||f2|| PROC
RSB r0,r0,r0,LSL #3
; R0=R0<<3-R0=R0*8-R0=a*8-a=a*7
LSL r0,r0,#2
; R0=R0<<2=R0*4=a*7*4=a*28
BX lr
ENDP
; a*17
||f3|| PROC
ADD r0,r0,r0,LSL #4
; R0=R0+R0<<4=R0+R0*16=R0*17=a*17
213
BX lr
ENDP
But there are no such modifiers in Thumb mode. It also can’t optimize f2():

; a*7
||f1|| PROC
LSLS r1,r0,#3
; R1=R0<<3=a<<3=a*8
SUBS r0,r1,r0
; R0=R1-R0=a*8-a=a*7
BX lr
ENDP
; a*28
||f2|| PROC
MOVS r1,#0x1c ; 28
; R1=28
MULS r0,r1,r0
; R0=R1*R0=28*a
BX lr
ENDP
; a*17
||f3|| PROC
LSLS r1,r0,#4
; R1=R0<<4=R0*16=a*16
ADDS r0,r0,r1
; R0=R0+R1=a+a*16=a*17
BX lr
ENDP
MIPS

_f1:
sll $v0, $a0, 3
; $v0 = $a0<<3 = $a0*8
jr $ra
subu $v0, $a0 ; branch delay slot
; $v0 = $v0-$a0 = $a0*8-$a0 = $a0*7
_f2:
sll $v0, $a0, 5
; $v0 = $a0<<5 = $a0*32
sll $a0, 2
; $a0 = $a0<<2 = $a0*4
jr $ra
; $v0 = $a0*32-$a0*4 = $a0*28
_f3:
sll $v0, $a0, 4
; $v0 = $a0<<4 = $a0*16
jr $ra
addu $v0, $a0 ; branch delay slot
; $v0 = $a0*16+$a0 = $a0*17
64-bit
#include <stdint.h>
int64_t f1(int64_t a)
214
{
return a*7;
};
{
return a*28;
};
{
return a*17;
};
x64

; a*7
f1:
lea rax, [0+rdi*8]
; RAX=RDI*8=a*8
sub rax, rdi
; RAX=RAX-RDI=a*8-a=a*7
ret
; a*28
f2:
lea rax, [0+rdi*4]
; RAX=RDI*4=a*4
sal rdi, 5
; RDI=RDI<<5=RDI*32=a*32
sub rdi, rax
; RDI=RDI-RAX=a*32-a*4=a*28
mov rax, rdi
ret
; a*17
f3:
mov rax, rdi
sal rax, 4
; RAX=RAX<<4=a*16
add rax, rdi
; RAX=a*16+a=a*17
ret
ARM64
GCC 4.9 for ARM64 is also terse, thanks to the shift modifiers:
Listing 1.203: Optimizing GCC (Linaro) 4.9 ARM64

; a*7
f1:
lsl x1, x0, 3
; X1=X0<<3=X0*8=a*8
sub x0, x1, x0
; X0=X1-X0=a*8-a=a*7
ret
; a*28
f2:
lsl x1, x0, 5
; X1=X0<<5=a*32
sub x0, x1, x0, lsl 2
; X0=X1-X0<<2=a*32-a<<2=a*32-a*4=a*28
ret
215
; a*17
f3:
add x0, x0, x0, lsl 4
; X0=X0+X0<<4=a+a*16=a*17
ret
Booth’s multiplication algorithm
There was a time when computers were big and that expensive, that some of them lacked hardware
support of multiplication operation in CPU, like Data General Nova. And when one need multiplication
operation, it can be provided at software level, for example, using Booth’s multiplication algorithm. This
is a multiplication algorithm which uses only addition operation and shifts.
What modern optimizing compilers do, isn’t the same, but the goal (multiplication) and resources (faster
operations) are the same.
1.24.2 Division
Division using shifts
Example of division by 4:
{
return a/4;
};
We get (MSVC 2010):

_a$ = 8 ; size = 4
_f PROC
shr eax, 2
ret 0
_f ENDP
The SHR (SHift Right) instruction in this example is shifting a number by 2 bits to the right. The two freed
bits at left (e.g., two most significant bits) are set to zero. The two least significant bits are dropped. In
fact, these two dropped bits are the division operation remainder.
The SHR instruction works just like SHL, but in the other direction.
7 6 5 4 3 2 1 0
0 7 6 5 4 3 2 1 0 CF
It is easy to understand if you imagine the number 23 in the decimal numeral system. 23 can be easily
divided by 10 just by dropping last digit (3—division remainder). 2 is left after the operation as a quotient.
So the remainder is dropped, but that’s OK, we work on integer values anyway, these are not a real
numbers!
Division by 4 in ARM:

f PROC
LSR r0,r0,#2
BX lr
ENDP
Division by 4 in MIPS:
216
jr $ra
srl $v0, $a0, 2 ; branch delay slot
The SRL instruction is “Shift Right Logical”.
1.24.3 Exercise
1.25 Floating-point unit

The FPU is a device within the main CPU, specially designed to deal with floating point numbers.
It was called “coprocessor” in the past and it stays somewhat aside of the main CPU.
1.25.1 IEEE 754

A number in the IEEE 754 format consists of a sign, a significand (also called fraction) and an exponent.
1.25.2 x86
It is worth looking into stack machines or learning the basics of the Forth language, before studying the
FPU in x86.
It is interesting to know that in the past (before the 80486 CPU) the coprocessor was a separate chip and
it was not always pre-installed on the motherboard. It was possible to buy it separately and install it 108 .
Starting with the 80486 DX CPU, the FPU is integrated in the CPU.
The FWAIT instruction reminds us of that fact—it switches the CPU to a waiting state, so it can wait until
the FPU has finished with its work.
Another rudiment is the fact that the FPU instruction opcodes start with the so called “escape”-opcodes
(D8..DF), i.e., opcodes passed to a separate coprocessor.
The FPU has a stack capable to holding 8 80-bit registers, and each register can hold a number in the IEEE
754 format.
They are ST(0)..ST(7). For brevity, IDA and OllyDbg show ST(0) as ST, which is represented in some
textbooks and manuals as “Stack Top”.
1.25.3 ARM, MIPS, x86/x64 SIMD

In ARM and MIPS the FPU is not a stack, but a set of registers, which can be accessed randomly, like GPR.
The same ideology is used in the SIMD extensions of x86/x64 CPUs.
1.25.4 C/C++
The standard C/C++ languages offer at least two floating number types, float (single-precision, 32 bits)
109
and double (double-precision, 64 bits).
In [Donald E. Knuth, The Art of Computer Programming, Volume 2, 3rd ed., (1997)246] we can find the
single-precision means that the floating point value can be placed into a single [32-bit] machine word,
double-precision means it can be stored in two words (64 bits).
GCC also supports the long double type (extended precision, 80 bit), which MSVC doesn’t.
The float type requires the same number of bits as the int type in 32-bit environments, but the number
representation is completely different.
108 For example, John Carmack used fixed-point arithmetic values in his Doom video game, stored in 32-bit GPR registers (16 bit for
integral part and another 16 bit for fractional part), so Doom could work on 32-bit computers without FPU, i.e., 80386 and 80486
SX.
109 the single precision floating point number format is also addressed in the Handling float data type as a structure ( 1.30.6 on
page 370) section
217
Let’s consider this simple example:
#include <stdio.h>
double f (double a, double b)

{
return a/3.14 + b*4.1;
};
int main()
{
printf ("%f\n", f(1.2, 3.4));
};
x86
MSVC
Compile it in MSVC 2010:
Listing 1.207: MSVC 2010: f()

CONST SEGMENT
__real@4010666666666666 DQ 04010666666666666r ; 4.1
CONST ENDS
CONST SEGMENT
__real@40091eb851eb851f DQ 040091eb851eb851fr ; 3.14
CONST ENDS
_TEXT SEGMENT
_a$ = 8 ; size = 8
_b$ = 16 ; size = 8
_f PROC
push ebp
mov ebp, esp
fld QWORD PTR _a$[ebp]
; current stack state: ST(0) = _a
fdiv QWORD PTR __real@40091eb851eb851f
; current stack state: ST(0) = result of _a divided by 3.14
fld QWORD PTR _b$[ebp]
; current stack state: ST(0) = _b;

; ST(1) = result of _a divided by 3.14
fmul QWORD PTR __real@4010666666666666
; current stack state:

; ST(0) = result of _b * 4.1;
; ST(1) = result of _a divided by 3.14
faddp ST(1), ST(0)
; current stack state: ST(0) = result of addition
pop ebp
ret 0
_f ENDP
FLD takes 8 bytes from stack and loads the number into the ST(0) register, automatically converting it
into the internal 80-bit format (extended precision).
FDIV divides the value in ST(0) by the number stored at address
__real@40091eb851eb851f —the value 3.14 is encoded there. The assembly syntax doesn’t support
218
floating point numbers, so what we see here is the hexadecimal representation of 3.14 in 64-bit IEEE 754
format.
After the execution of FDIV ST(0) holds the quotient.
By the way, there is also the FDIVP instruction, which divides ST(1) by ST(0), popping both these values
from stack and then pushing the result. If you know the Forth language, you can quickly understand that
this is a stack machine.
The subsequent FLD instruction pushes the value of b into the stack.
After that, the quotient is placed in ST(1), and ST(0) has the value of b.
The next FMUL instruction does multiplication: b from ST(0) is multiplied by value at
__real@4010666666666666 (the number 4.1 is there) and leaves the result in the ST(0) register.
The last FADDP instruction adds the two values at top of stack, storing the result in ST(1) and then popping
the value of ST(0), thereby leaving the result at the top of the stack, in ST(0).
The function must return its result in the ST(0) register, so there are no any other instructions except the
function epilogue after FADDP.
219
MSVC + OllyDbg
2 pairs of 32-bit words are marked by red in the stack. Each pair is a double-number in IEEE 754 format
and is passed from main().
We see how the first FLD loads a value (1.2) from the stack and puts it into ST(0):
Figure 1.63: OllyDbg: the first FLD has been executed
Because of unavoidable conversion errors from 64-bit IEEE 754 floating point to 80-bit (used internally in
the FPU), here we see 1.1999…, which is close to 1.2.
EIP now points to the next instruction (FDIV), which loads a double-number (a constant) from memory.
For convenience, OllyDbg shows its value: 3.14
220
Let’s trace further. FDIV has been executed, now ST(0) contains 0.382…(quotient):
Figure 1.64: OllyDbg: FDIV has been executed
221
Third step: the next FLD has been executed, loading 3.4 into ST(0) (here we see the approximate value
3.39999…):
Figure 1.65: OllyDbg: the second FLD has been executed
At the same time, quotient is pushed into ST(1). Right now, EIP points to the next instruction: FMUL. It
loads the constant 4.1 from memory, which OllyDbg shows.
222
Next: FMUL has been executed, so now the product is in ST(0):
Figure 1.66: OllyDbg: the FMUL has been executed
223
Next: the FADDP has been executed, now the result of the addition is in ST(0), and ST(1) is cleared:
Figure 1.67: OllyDbg: FADDP has been executed
The result is left in ST(0), because the function returns its value in ST(0).
main() takes this value from the register later.
We also see something unusual: the 13.93…value is now located in ST(7). Why?
As we have read some time before in this book, the FPU registers are a stack: 1.25.2 on page 217. But
this is a simplification.
Imagine if it was implemented in hardware as it’s described, then all 7 register’s contents must be moved
(or copied) to adjacent registers during pushing and popping, and that’s a lot of work.
In reality, the FPU has just 8 registers and a pointer (called TOP) which contains a register number, which
is the current “top of stack”.
When a value is pushed to the stack, TOP is pointed to the next available register, and then a value is
written to that register.
The procedure is reversed if a value is popped, however, the register which has been freed is not cleared
(it could possibly be cleared, but this is more work which can degrade performance). So that’s what we
see here.
It can be said that FADDP saved the sum in the stack, and then popped one element.
But in fact, this instruction saved the sum and then shifted TOP.
More precisely, the registers of the FPU are a circular buffer.
GCC
GCC 4.4.1 (with -O3 option) emits the same code, just slightly different:
224
public f
f proc near
arg_0 = qword ptr 8

arg_8 = qword ptr 10h
push ebp
fld ds:dbl_8048608 ; 3.14
; stack state now: ST(0) = 3.14
mov ebp, esp

fdivr [ebp+arg_0]
; stack state now: ST(0) = result of division
fld ds:dbl_8048610 ; 4.1
; stack state now: ST(0) = 4.1, ST(1) = result of division
fmul [ebp+arg_8]
; stack state now: ST(0) = result of multiplication, ST(1) = result of division
pop ebp
faddp st(1), st
; stack state now: ST(0) = result of addition
retn
f endp
The difference is that, first of all, 3.14 is pushed to the stack (into ST(0)), and then the value in arg_0 is
divided by the value in ST(0).
FDIVR stands for Reverse Divide —to divide with divisor and dividend swapped with each other. There is
no likewise instruction for multiplication since it is a commutative operation, so we just have FMUL without
its -R counterpart.
FADDP adds the two values but also pops one value from the stack. After that operation, ST(0) holds the
sum.
ARM: Optimizing Xcode 4.6.3 (LLVM) (ARM mode)
Until ARM got standardized floating point support, several processor manufacturers added their own in-
structions extensions. Then, VFP (Vector Floating Point) was standardized.
One important difference from x86 is that in ARM, there is no stack, you work just with registers.

f
VLDR D16, =3.14
VMOV D17, R0, R1 ; load "a"
VMOV D18, R2, R3 ; load "b"
VDIV.F64 D16, D17, D16 ; a/3.14
VLDR D17, =4.1
VMUL.F64 D17, D18, D17 ; b*4.1
VADD.F64 D16, D17, D16 ; +
VMOV R0, R1, D16
BX LR
dbl_2C98 DCFD 3.14 ; DATA XREF: f

dbl_2CA0 DCFD 4.1 ; DATA XREF: f+10
So, we see here new some registers used, with D prefix.

These are 64-bit registers, there are 32 of them, and they can be used both for floating-point numbers
(double) but also for SIMD (it is called NEON here in ARM).
225
There are also 32 32-bit S-registers, intended to be used for single precision floating pointer numbers
(float).
It is easy to memorize: D-registers are for double precision numbers, while S-registers—for single precision
numbers. More about it: .2.3 on page 1026.
Both constants (3.14 and 4.1) are stored in memory in IEEE 754 format.
VLDR and VMOV, as it can be easily deduced, are analogous to the LDR and MOV instructions, but they work
with D-registers.
It has to be noted that these instructions, just like the D-registers, are intended not only for floating point
numbers, but can be also used for SIMD (NEON) operations and this will also be shown soon.
The arguments are passed to the function in a common way, via the R-registers, however each number
that has double precision has a size of 64 bits, so two R-registers are needed to pass each one.
VMOV D17, R0, R1 at the start, composes two 32-bit values from R0 and R1 into one 64-bit value and
saves it to D17.
VMOV R0, R1, D16 is the inverse operation: what has been in D16 is split in two registers, R0 and R1,
because a double-precision number that needs 64 bits for storage, is returned in R0 and R1.
VDIV, VMUL and VADD, are instruction for processing floating point numbers that compute quotient, product
and sum, respectively.
The code for Thumb-2 is same.
f
PUSH {R3-R7,LR}
MOVS R7, R2
MOVS R4, R3
MOVS R5, R0
MOVS R6, R1
LDR R2, =0x66666666 ; 4.1
LDR R3, =0x40106666
MOVS R0, R7
MOVS R1, R4
BL __aeabi_dmul
MOVS R7, R0
MOVS R4, R1
LDR R2, =0x51EB851F ; 3.14
LDR R3, =0x40091EB8
MOVS R0, R5
MOVS R1, R6
BL __aeabi_ddiv
MOVS R2, R7
MOVS R3, R4
BL __aeabi_dadd
POP {R3-R7,PC}
; 4.1 in IEEE 754 form:

dword_364 DCD 0x66666666 ; DATA XREF: f+A
dword_368 DCD 0x40106666 ; DATA XREF: f+C
; 3.14 in IEEE 754 form:
dword_36C DCD 0x51EB851F ; DATA XREF: f+1A
dword_370 DCD 0x40091EB8 ; DATA XREF: f+1C
Keil generated code for a processor without FPU or NEON support.

The double-precision floating-point numbers are passed via generic R-registers, and instead of FPU-instructions,
service library functions are called
(like __aeabi_dmul, __aeabi_ddiv, __aeabi_dadd) which emulate multiplication, division and addition
for floating-point numbers.
Of course, that is slower than FPU-coprocessor, but it’s still better than nothing.
By the way, similar FPU-emulating libraries were very popular in the x86 world when coprocessors were
rare and expensive, and were installed only on expensive computers.
226
The FPU-coprocessor emulation is called soft float or armel (emulation) in the ARM world, while using the
coprocessor’s FPU-instructions is called hard float or armhf.
Very compact code:

f:
; D0 = a, D1 = b
ldr d2, .LC25 ; 3.14
; D2 = 3.14
fdiv d0, d0, d2
; D0 = D0/D2 = a/3.14
ldr d2, .LC26 ; 4.1
; D2 = 4.1
fmadd d0, d1, d2, d0
; D0 = D1*D2+D0 = b*4.1+a/3.14
ret
; constants in IEEE 754 format:

.LC25:
.word 1374389535 ; 3.14
.word 1074339512
.LC26:
.word 1717986918 ; 4.1
.word 1074816614

f:
sub sp, sp, #16
str d0, [sp,8] ; save "a" in Register Save Area
str d1, [sp] ; save "b" in Register Save Area
ldr x1, [sp,8]
; X1 = a
ldr x0, .LC25
; X0 = 3.14
fmov d0, x1
fmov d1, x0
; D0 = a, D1 = 3.14
fdiv d0, d0, d1
; D0 = D0/D1 = a/3.14
fmov x1, d0
; X1 = a/3.14
ldr x2, [sp]
; X2 = b
ldr x0, .LC26
; X0 = 4.1
fmov d0, x2
; D0 = b
fmov d1, x0
; D1 = 4.1
fmul d0, d0, d1
; D0 = D0*D1 = b*4.1
fmov x0, d0
; X0 = D0 = b*4.1
fmov d0, x1
; D0 = a/3.14
fmov d1, x0
; D1 = X0 = b*4.1
fadd d0, d0, d1
; D0 = D0+D1 = a/3.14 + b*4.1
227
fmov x0, d0 ; \ redundant code
fmov d0, x0 ; /
add sp, sp, 16
ret
.LC25:
.word 1374389535 ; 3.14
.word 1074339512
.LC26:
.word 1717986918 ; 4.1
.word 1074816614

There is a lot of unnecessary value shuffling, including some clearly redundant code (the last two FMOV
instructions). Probably, GCC 4.9 is not yet good in generating ARM64 code.
What is worth noting is that ARM64 has 64-bit registers, and the D-registers are 64-bit ones as well.
So the compiler is free to save values of type double in GPRs instead of the local stack. This isn’t possible
on 32-bit CPUs.
And again, as an exercise, you can try to optimize this function manually, without introducing new instruc-
tions like FMADD.
1.25.6 Passing floating point numbers via arguments

#include <math.h>
#include <stdio.h>
int main ()
{
printf ("32.01 ^ 1.54 = %lf\n", pow (32.01,1.54));
return 0;
}
x86
Let’s see what we get in (MSVC 2010):

CONST SEGMENT
__real@40400147ae147ae1 DQ 040400147ae147ae1r ; 32.01
__real@3ff8a3d70a3d70a4 DQ 03ff8a3d70a3d70a4r ; 1.54
CONST ENDS
_main PROC
push ebp
mov ebp, esp
sub esp, 8 ; allocate space for the first variable
fld QWORD PTR __real@3ff8a3d70a3d70a4
fstp QWORD PTR [esp]
sub esp, 8 ; allocate space for the second variable
fld QWORD PTR __real@40400147ae147ae1
call _pow
add esp, 8 ; return back place of one variable.
; in local stack here 8 bytes still reserved for us.

; result now in ST(0)
fstp QWORD PTR [esp] ; move result from ST(0) to local stack for printf()
push OFFSET $SG2651
call _printf
add esp, 12
xor eax, eax
pop ebp
ret 0
_main ENDP
228
FLD and FSTP move variables between the data segment and the FPU stack. pow()110 takes both values
from the stack and returns its result in the ST(0) register. printf() takes 8 bytes from the local stack
and interprets them as double type variable.
By the way, a pair of MOV instructions could be used here for moving values from the memory into the
stack, because the values in memory are stored in IEEE 754 format, and pow() also takes them in this
format, so no conversion is necessary. That’s how it’s done in the next example, for ARM: 1.25.6.
ARM + Non-optimizing Xcode 4.6.3 (LLVM) (Thumb-2 mode)
_main
var_C = -0xC
PUSH {R7,LR}
MOV R7, SP
SUB SP, SP, #4
VLDR D16, =32.01
VMOV R0, R1, D16
VLDR D16, =1.54
VMOV R2, R3, D16
BLX _pow
VMOV D16, R0, R1
MOV R0, 0xFC1 ; "32.01 ^ 1.54 = %lf\n"
ADD R0, PC
VMOV R1, R2, D16
BLX _printf
MOVS R1, 0
STR R0, [SP,#0xC+var_C]
MOV R0, R1
ADD SP, SP, #4
POP {R7,PC}
dbl_2F90 DCFD 32.01 ; DATA XREF: _main+6

dbl_2F98 DCFD 1.54 ; DATA XREF: _main+E
As it was mentioned before, 64-bit floating pointer numbers are passed in R-registers pairs.
This code is a bit redundant (certainly because optimization is turned off), since it is possible to load values
into the R-registers directly without touching the D-registers.
So, as we see, the _pow function receives its first argument in R0 and R1, and its second one in R2 and R3.
The function leaves its result in R0 and R1. The result of _pow is moved into D16, then in the R1 and R2
pair, from where printf() takes the resulting number.
ARM + Non-optimizing Keil 6/2013 (ARM mode)
_main
STMFD SP!, {R4-R6,LR}
LDR R2, =0xA3D70A4 ; y
LDR R3, =0x3FF8A3D7
LDR R0, =0xAE147AE1 ; x
LDR R1, =0x40400147
BL pow
MOV R4, R0
MOV R2, R4
MOV R3, R1
ADR R0, a32_011_54Lf ; "32.01 ^ 1.54 = %lf\n"
BL __2printf
MOV R0, #0
LDMFD SP!, {R4-R6,PC}
y DCD 0xA3D70A4 ; DATA XREF: _main+4

dword_520 DCD 0x3FF8A3D7 ; DATA XREF: _main+8
x DCD 0xAE147AE1 ; DATA XREF: _main+C
dword_528 DCD 0x40400147 ; DATA XREF: _main+10
a32_011_54Lf DCB "32.01 ^ 1.54 = %lf",0xA,0
110 a standard C function, raises a number to the given power (exponentiation)
229
; DATA XREF: _main+24
D-registers are not used here, just R-register pairs.
ARM64 + Optimizing GCC (Linaro) 4.9

f:
stp x29, x30, [sp, -16]!
add x29, sp, 0
ldr d1, .LC1 ; load 1.54 into D1
ldr d0, .LC0 ; load 32.01 into D0
bl pow
; result of pow() in D0
adrp x0, .LC2
bl printf
mov w0, 0
ldp x29, x30, [sp], 16
ret
.LC0:
; 32.01 in IEEE 754 format
.word -1374389535
.word 1077936455
.LC1:
; 1.54 in IEEE 754 format
.word 171798692
.word 1073259479
.LC2:
.string "32.01 ^ 1.54 = %lf\n"
The constants are loaded into D0 and D1: pow() takes them from there. The result will be in D0 after
the execution of pow(). It is to be passed to printf() without any modification and moving, because
printf() takes arguments of integral types and pointers from X-registers, and floating point arguments
from D-registers.
1.25.7 Comparison example

Let’s try this:
#include <stdio.h>
double d_max (double a, double b)

{
if (a>b)
return a;
return b;
};
int main()
{
printf ("%f\n", d_max (1.2, 3.4));
printf ("%f\n", d_max (5.6, -4));
};
Despite the simplicity of the function, it will be harder to understand how it works.
x86
Non-optimizing MSVC
MSVC 2010 generates the following:
230
PUBLIC _d_max
_TEXT SEGMENT
_a$ = 8 ; size = 8
_b$ = 16 ; size = 8
_d_max PROC
push ebp
mov ebp, esp
; current stack state: ST(0) = _b

; compare _b (ST(0)) and _a, and pop register
fcomp QWORD PTR _a$[ebp]
; stack is empty here
fnstsw ax
test ah, 5
jp SHORT $LN1@d_max
; we are here only if a>b
fld QWORD PTR _a$[ebp]

jmp SHORT $LN2@d_max
$LN1@d_max:
$LN2@d_max:
pop ebp
ret 0
_d_max ENDP
So, FLD loads _b into ST(0).

FCOMP compares the value in ST(0) with what is in _a and sets C3/C2/C0 bits in FPU status word register,
accordingly. This is a 16-bit register that reflects the current state of the FPU.
After the bits are set, the FCOMP instruction also pops one variable from the stack. This is what distinguishes
it from FCOM, which is just compares values, leaving the stack in the same state.
Unfortunately, CPUs before Intel P6 111 don’t have any conditional jumps instructions which check the
C3/C2/C0 bits. Perhaps, it is a matter of history (recall: FPU was a separate chip in past).
Modern CPU starting at Intel P6 have FCOMI/FCOMIP/FUCOMI/FUCOMIP instructions —which do the same,
but modify the ZF/PF/CF CPU flags.
The FNSTSW instruction copies FPU the status word register to AX. C3/C2/C0 bits are placed at positions
14/10/8, they are at the same positions in the AX register and all they are placed in the high part of
AX —AH.
• If b > a in our example, then C3/C2/C0 bits are to be set as following: 0, 0, 0.
• If a > b, then the bits are: 0, 0, 1.
• If a = b, then the bits are: 1, 0, 0.
• If the result is unordered (in case of error), then the set bits are: 1, 1, 1.
This is how C3/C2/C0 bits are located in the AX register:
14 10 9 8
C3 C2 C1 C0
This is how C3/C2/C0 bits are located in the AH register:

6 2 1 0
C3 C2 C1 C0
After the execution of test ah, 5112 , only C0 and C2 bits (on 0 and 2 position) are considered, all other
bits are just ignored.
Now let’s talk about the parity flag, another notable historical rudiment.
111 Intel P6 is Pentium Pro, Pentium II, etc.
112 5=101b
231
This flag is set to 1 if the number of ones in the result of the last calculation is even, and to 0 if it is odd.
Let’s look into Wikipedia113 :
One common reason to test the parity flag actually has nothing to do with parity. The
FPU has four condition flags (C0 to C3), but they cannot be tested directly, and must instead
be first copied to the flags register. When this happens, C0 is placed in the carry flag, C2 in
the parity flag and C3 in the zero flag. The C2 flag is set when e.g. incomparable floating
point values (NaN or unsupported format) are compared with the FUCOM instructions.
As noted in Wikipedia, the parity flag used sometimes in FPU code, let’s see how.
The PF flag is to be set to 1 if both C0 and C2 are set to 0 or both are 1, in which case the subsequent JP
(jump if PF==1) is triggering. If we recall the values of C3/C2/C0 for various cases, we can see that the
conditional jump JP is triggering in two cases: if b > a or a = b (C3 bit is not considered here, since it has
been cleared by the test ah, 5 instruction).
It is all simple after that. If the conditional jump has been triggered, FLD loads the value of _b in ST(0),
and if it hasn’t been triggered, the value of _a is loaded there.
And what about checking C2?
The C2 flag is set in case of error (NaN, etc.), but our code doesn’t check it.
If the programmer cares about FPU errors, he/she must add additional checks.
113 https://en.wikipedia.org/wiki/Parity_flag
232
First OllyDbg example: a=1.2 and b=3.4
Let’s load the example into OllyDbg:
Figure 1.68: OllyDbg: first FLD has been executed
Current arguments of the function: a = 1.2 and b = 3.4 (We can see them in the stack: two pairs of 32-bit
values). b (3.4) is already loaded in ST(0). Now FCOMP is being executed. OllyDbg shows the second
FCOMP argument, which is in stack right now.
233
FCOMP has been executed:
Figure 1.69: OllyDbg: FCOMP has been executed
We see the state of the FPU’s condition flags: all zeros. The popped value is reflected as ST(7), it was
written earlier about reason for this: 1.25.5 on page 224.
234
FNSTSW has been executed:
Figure 1.70: OllyDbg: FNSTSW has been executed
We see that the AX register contain zeros: indeed, all condition flags are zero. (OllyDbg disassembles the
FNSTSW instruction as FSTSW—they are synonyms).
235
TEST has been executed:
Figure 1.71: OllyDbg: TEST has been executed
The PF flag is set to 1.

Indeed: the number of bits set in 0 is 0 and 0 is an even number. OllyDbg disassembles JP as JPE114 —they
are synonyms. And it is about to trigger now.
114 Jump Parity Even (x86 instruction)
236
JPE triggered, FLD loads the value of b (3.4) in ST(0):
Figure 1.72: OllyDbg: the second FLD has been executed
The function finishes its work.
237
Second OllyDbg example: a=5.6 and b=-4
Let’s load example into OllyDbg:
Figure 1.73: OllyDbg: first FLD executed
Current function arguments: a = 5.6 and b = −4. b (-4) is already loaded in ST(0). FCOMP about to execute
now. OllyDbg shows the second FCOMP argument, which is in stack right now.
238
FCOMP executed:
Figure 1.74: OllyDbg: FCOMP executed
We see the state of the FPU’s condition flags: all zeros except C0.
239
FNSTSW executed:
Figure 1.75: OllyDbg: FNSTSW executed
We see that the AX register contains 0x100: the C0 flag is at the 8th bit.
240
TEST executed:
Figure 1.76: OllyDbg: TEST executed
The PF flag is cleared. Indeed:

the count of bits set in 0x100 is 1 and 1 is an odd number. JPE is being skipped now.
241
JPE hasn’t been triggered, so FLD loads the value of a (5.6) in ST(0):
Figure 1.77: OllyDbg: second FLD executed
The function finishes its work.
Optimizing MSVC 2010

_a$ = 8 ; size = 8
_b$ = 16 ; size = 8
_d_max PROC
fld QWORD PTR _b$[esp-4]
fld QWORD PTR _a$[esp-4]
; current stack state: ST(0) = _a, ST(1) = _b
fcom ST(1) ; compare _a and ST(1) = (_b)

fnstsw ax
test ah, 65 ; 00000041H
jne SHORT $LN5@d_max
; copy ST(0) to ST(1) and pop register,
; leave (_a) on top
fstp ST(1)
; current stack state: ST(0) = _a
ret 0
$LN5@d_max:
; copy ST(0) to ST(0) and pop register,
; leave (_b) on top
242
fstp ST(0)
; current stack state: ST(0) = _b
ret 0
_d_max ENDP
FCOM differs from FCOMP in the sense that it just compares the values and doesn’t change the FPU stack.
Unlike the previous example, here the operands are in reverse order, which is why the result of the com-
parison in C3/C2/C0 is different:
• If a > b in our example, then C3/C2/C0 bits are to be set as: 0, 0, 0.
• If b > a, then the bits are: 0, 0, 1.
• If a = b, then the bits are: 1, 0, 0.
The test ah, 65 instruction leaves just two bits —C3 and C0. Both will be zero if a > b: in that case the
JNE jump will not be triggered. Then FSTP ST(1) follows —this instruction copies the value from ST(0) to
the operand and pops one value from the FPU stack. In other words, the instruction copies ST(0) (where
the value of _a is now) into ST(1). After that, two copies of _a are at the top of the stack. Then, one value
is popped. After that, ST(0) contains _a and the function is finishes.
The conditional jump JNE is triggering in two cases: if b > a or a = b. ST(0) is copied into ST(0), it is just
like an idle (NOP) operation, then one value is popped from the stack and the top of the stack (ST(0)) is
contain what has been in ST(1) before (that is _b). Then the function finishes. The reason this instruction
is used here probably is because the FPU has no other instruction to pop a value from the stack and discard
it.
243
First OllyDbg example: a=1.2 and b=3.4
Both FLD are executed:
Figure 1.78: OllyDbg: both FLD are executed
FCOM being executed: OllyDbg shows the contents of ST(0) and ST(1) for convenience.
244
FCOM has been executed:
Figure 1.79: OllyDbg: FCOM has been executed
C0 is set, all other condition flags are cleared.
245
FNSTSW has been executed, AX=0x3100:
Figure 1.80: OllyDbg: FNSTSW is executed
246
TEST is executed:
Figure 1.81: OllyDbg: TEST is executed
ZF=0, conditional jump is about to trigger now.
247
FSTP ST (or FSTP ST(0)) has been executed —1.2 has been popped from the stack, and 3.4 was left on
top:
Figure 1.82: OllyDbg: FSTP is executed
We see that the FSTP ST

instruction works just like popping one value from the FPU stack.
248
Second OllyDbg example: a=5.6 and b=-4
Both FLD are executed:
Figure 1.83: OllyDbg: both FLD are executed
FCOM is about to execute.
249
FCOM has been executed:
Figure 1.84: OllyDbg: FCOM is finished
All conditional flags are cleared.
250
FNSTSW done, AX=0x3000:
Figure 1.85: OllyDbg: FNSTSW has been executed
251
TEST has been executed:
Figure 1.86: OllyDbg: TEST has been executed
ZF=1, jump will not happen now.
252
FSTP ST(1) has been executed: a value of 5.6 is now at the top of the FPU stack.
Figure 1.87: OllyDbg: FSTP has been executed
We now see that the FSTP ST(1) instruction works as follows: it leaves what has been at the top of the
stack, but clears ST(1).
GCC 4.4.1

d_max proc near
b = qword ptr -10h

a = qword ptr -8
a_first_half = dword ptr 8
a_second_half = dword ptr 0Ch
b_first_half = dword ptr 10h
b_second_half = dword ptr 14h
push ebp
mov ebp, esp
sub esp, 10h
; put a and b to local stack:
mov eax, [ebp+a_first_half]

mov dword ptr [ebp+a], eax
mov eax, [ebp+a_second_half]
mov dword ptr [ebp+a+4], eax
mov eax, [ebp+b_first_half]
mov dword ptr [ebp+b], eax
253
mov eax, [ebp+b_second_half]
mov dword ptr [ebp+b+4], eax
; load a and b to FPU stack:
fld [ebp+a]
fld [ebp+b]
; current stack state: ST(0) - b; ST(1) - a
fxch st(1) ; this instruction swaps ST(1) and ST(0)
; current stack state: ST(0) - a; ST(1) - b
fucompp ; compare a and b and pop two values from stack, i.e., a and b
fnstsw ax ; store FPU status to AX
sahf ; load SF, ZF, AF, PF, and CF flags state from AH
setnbe al ; store 1 to AL, if CF=0 and ZF=0
test al, al ; AL==0 ?
jz short loc_8048453 ; yes
fld [ebp+a]
loc_8048453:
fld [ebp+b]
locret_8048456:
leave
retn
d_max endp
FUCOMPP is almost like FCOM, but pops both values from the stack and handles “not-a-numbers” differently.
A bit about not-a-numbers.
The FPU is able to deal with special values which are not-a-numbers or NaNs. These are infinity, result of
division by 0, etc. Not-a-numbers can be “quiet” and “signaling”. It is possible to continue to work with
“quiet” NaNs, but if one tries to do any operation with “signaling” NaNs, an exception is to be raised.
FCOM raises an exception if any operand is NaN. FUCOM raises an exception only if any operand is a signaling
NaN (SNaN).
The next instruction is SAHF (Store AH into Flags) —this is a rare instruction in code not related to the FPU.
8 bits from AH are moved into the lower 8 bits of the CPU flags in the following order:
7 6 4 2 0
SF ZF AF PF CF
Let’s recall that FNSTSW moves the bits that interest us (C3/C2/C0) into AH and they are in positions 6, 2, 0
of the AH register:
6 2 1 0
C3 C2 C1 C0
In other words, the fnstsw ax / sahf instruction pair moves C3/C2/C0 into ZF, PF and CF.
Now let’s also recall the values of C3/C2/C0 in different conditions:
• If a is greater than b in our example, then C3/C2/C0 are to be set to: 0, 0, 0.
• if a is less than b, then the bits are to be set to: 0, 0, 1.
• If a = b, then: 1, 0, 0.
In other words, these states of the CPU flags are possible after three
FUCOMPP/FNSTSW/SAHF instructions:
• If a > b, the CPU flags are to be set as: ZF=0, PF=0, CF=0.
• If a < b, then the flags are to be set as: ZF=0, PF=0, CF=1.
• And if a = b, then: ZF=1, PF=0, CF=0.
254
Depending on the CPU flags and conditions, SETNBE stores 1 or 0 to AL. It is almost the counterpart of
JNBE, with the exception that SETcc115 stores 1 or 0 in AL, but Jcc does actually jump or not. SETNBE
stores 1 only if CF=0 and ZF=0. If it is not true, 0 is to be stored into AL.
Only in one case both CF and ZF are 0: if a > b.
Then 1 is to be stored to AL, the subsequent JZ is not to be triggered and the function will return _a. In all
other cases, _b is to be returned.

public d_max
d_max proc near
arg_0 = qword ptr 8

arg_8 = qword ptr 10h
push ebp
mov ebp, esp
fld [ebp+arg_0] ; _a
fld [ebp+arg_8] ; _b
; stack state now: ST(0) = _b, ST(1) = _a

fxch st(1)
; stack state now: ST(0) = _a, ST(1) = _b

fucom st(1) ; compare _a and _b
fnstsw ax
sahf
; store ST(0) to ST(0) (idle operation),

; pop value at top of stack,
; leave _b at top
fstp st
jmp short loc_804844A
loc_8048448:
; store _a to ST(1), pop value at top of stack, leave _a at top
fstp st(1)
loc_804844A:
pop ebp
retn
d_max endp
It is almost the same except that s used after SAHF. Actually, conditional jump instructions that check
“larger”, “lesser” or “equal” for unsigned number comparison (these are JA, JAE, JB, JBE, JE/JZ, JNA,
JNAE, JNB, JNBE, JNE/JNZ) check only flags CF and ZF.
Let’s recall where bits C3/C2/C0 are located in the AH register after the execution of FSTSW/FNSTSW:
6 2 1 0
C3 C2 C1 C0
Let’s also recall, how the bits from AH are stored into the CPU flags after the execution of SAHF:
7 6 4 2 0
SF ZF AF PF CF
After the comparison, the C3 and C0 bits are moved into ZF and CF, so the conditional jumps are able work
after. s triggering if both CF are ZF zero.
Thereby, the conditional jumps instructions listed here can be used after a FNSTSW/SAHF instruction pair.
115 cc is condition code
255
Apparently, the FPU C3/C2/C0 status bits were placed there intentionally, to easily map them to base CPU
flags without additional permutations?
GCC 4.8.1 with -O3 optimization turned on
Some new FPU instructions were added in the P6 Intel family116 . These are FUCOMI (compare operands
and set flags of the main CPU) and FCMOVcc (works like CMOVcc, but on FPU registers).
Apparently, the maintainers of GCC decided to drop support of pre-P6 Intel CPUs (early Pentiums, 80486,
etc.).
And also, the FPU is no longer separate unit in P6 Intel family, so now it is possible to modify/check flags
of the main CPU from the FPU.
So what we get is:

fld QWORD PTR [esp+4] ; load "a"
fld QWORD PTR [esp+12] ; load "b"
; ST0=b, ST1=a
fxch st(1)
; ST0=a, ST1=b
; compare "a" and "b"
fucomi st, st(1)
; copy ST1 ("b" here) to ST0 if a<=b
; leave "a" in ST0 otherwise
fcmovbe st, st(1)
; discard value in ST1
fstp st(1)
ret
Hard to guess why FXCH (swap operands) is here.

It’s possible to get rid of it easily by swapping the first two FLD instructions or by replacing FCMOVBE (below
or equal) by FCMOVA (above). Probably it’s a compiler inaccuracy.
So FUCOMI compares ST(0) (a) and ST(1) (b) and then sets some flags in the main CPU. FCMOVBE checks
the flags and copies ST(1) (b here at the moment) to ST(0) (a here) if ST 0(a) <= ST 1(b). Otherwise (a > b),
it leaves a in ST(0).
The last FSTP leaves ST(0) on top of the stack, discarding the contents of ST(1).
Let’s trace this function in GDB:
Listing 1.219: Optimizing GCC 4.8.1 and GDB

1 dennis@ubuntuvm:~/polygon$ gcc -O3 d_max.c -o d_max -fno-inline
2 dennis@ubuntuvm:~/polygon$ gdb d_max
3 GNU gdb (GDB) 7.6.1-ubuntu
4 ...
5 Reading symbols from /home/dennis/polygon/d_max...(no debugging symbols found)...done.
6 (gdb) b d_max
7 Breakpoint 1 at 0x80484a0
8 (gdb) run
9 Starting program: /home/dennis/polygon/d_max
10
11 Breakpoint 1, 0x080484a0 in d_max ()
12 (gdb) ni
13 0x080484a4 in d_max ()
14 (gdb) disas $eip
15 Dump of assembler code for function d_max:
16 0x080484a0 <+0>: fldl 0x4(%esp)
17 => 0x080484a4 <+4>: fldl 0xc(%esp)
18 0x080484a8 <+8>: fxch %st(1)
19 0x080484aa <+10>: fucomi %st(1),%st
20 0x080484ac <+12>: fcmovbe %st(1),%st
21 0x080484ae <+14>: fstp %st(1)
22 0x080484b0 <+16>: ret
23 End of assembler dump.
116 Starting at Pentium Pro, Pentium-II, etc.
256
24 (gdb) ni
25 0x080484a8 in d_max ()
26 (gdb) info float
27 R7: Valid 0x3fff9999999999999800 +1.199999999999999956
28 =>R6: Valid 0x4000d999999999999800 +3.399999999999999911
29 R5: Empty 0x00000000000000000000
30 R4: Empty 0x00000000000000000000
31 R3: Empty 0x00000000000000000000
32 R2: Empty 0x00000000000000000000
33 R1: Empty 0x00000000000000000000
34 R0: Empty 0x00000000000000000000
35
36 Status Word: 0x3000
37 TOP: 6
38 Control Word: 0x037f IM DM ZM OM UM PM
39 PC: Extended Precision (64-bits)
40 RC: Round to nearest
41 Tag Word: 0x0fff
42 Instruction Pointer: 0x73:0x080484a4
43 Operand Pointer: 0x7b:0xbffff118
44 Opcode: 0x0000
45 (gdb) ni
46 0x080484aa in d_max ()
47 (gdb) info float
48 R7: Valid 0x4000d999999999999800 +3.399999999999999911
49 =>R6: Valid 0x3fff9999999999999800 +1.199999999999999956
50 R5: Empty 0x00000000000000000000
51 R4: Empty 0x00000000000000000000
52 R3: Empty 0x00000000000000000000
53 R2: Empty 0x00000000000000000000
54 R1: Empty 0x00000000000000000000
55 R0: Empty 0x00000000000000000000
56
58 TOP: 6
62 Tag Word: 0x0fff
63 Instruction Pointer: 0x73:0x080484a8
65 Opcode: 0x0000
66 (gdb) disas $eip
68 0x080484a0 <+0>: fldl 0x4(%esp)
69 0x080484a4 <+4>: fldl 0xc(%esp)
70 0x080484a8 <+8>: fxch %st(1)
71 => 0x080484aa <+10>: fucomi %st(1),%st
72 0x080484ac <+12>: fcmovbe %st(1),%st
73 0x080484ae <+14>: fstp %st(1)
74 0x080484b0 <+16>: ret
76 (gdb) ni
77 0x080484ac in d_max ()
78 (gdb) info registers
79 eax 0x1 1
80 ecx 0xbffff1c4 -1073745468
81 edx 0x8048340 134513472
82 ebx 0xb7fbf000 -1208225792
83 esp 0xbffff10c 0xbffff10c
84 ebp 0xbffff128 0xbffff128
85 esi 0x0 0
86 edi 0x0 0
87 eip 0x80484ac 0x80484ac <d_max+12>
88 eflags 0x203 [ CF IF ]
89 cs 0x73 115
90 ss 0x7b 123
91 ds 0x7b 123
92 es 0x7b 123
93 fs 0x0 0
257
94 gs 0x33 51
95 (gdb) ni
96 0x080484ae in d_max ()
97 (gdb) info float
98 R7: Valid 0x4000d999999999999800 +3.399999999999999911
99 =>R6: Valid 0x4000d999999999999800 +3.399999999999999911
100 R5: Empty 0x00000000000000000000
101 R4: Empty 0x00000000000000000000
102 R3: Empty 0x00000000000000000000
103 R2: Empty 0x00000000000000000000
104 R1: Empty 0x00000000000000000000
105 R0: Empty 0x00000000000000000000
106
108 TOP: 6
112 Tag Word: 0x0fff
113 Instruction Pointer: 0x73:0x080484ac
115 Opcode: 0x0000
116 (gdb) disas $eip
118 0x080484a0 <+0>: fldl 0x4(%esp)
119 0x080484a4 <+4>: fldl 0xc(%esp)
120 0x080484a8 <+8>: fxch %st(1)
121 0x080484aa <+10>: fucomi %st(1),%st
122 0x080484ac <+12>: fcmovbe %st(1),%st
123 => 0x080484ae <+14>: fstp %st(1)
124 0x080484b0 <+16>: ret
126 (gdb) ni
127 0x080484b0 in d_max ()
128 (gdb) info float
129 =>R7: Valid 0x4000d999999999999800 +3.399999999999999911
130 R6: Empty 0x4000d999999999999800
131 R5: Empty 0x00000000000000000000
132 R4: Empty 0x00000000000000000000
133 R3: Empty 0x00000000000000000000
134 R2: Empty 0x00000000000000000000
135 R1: Empty 0x00000000000000000000
136 R0: Empty 0x00000000000000000000
137
139 TOP: 7
143 Tag Word: 0x3fff
144 Instruction Pointer: 0x73:0x080484ae
146 Opcode: 0x0000
147 (gdb) quit
148 A debugging session is active.
149
150 Inferior 1 [process 30194] will be killed.
151
152 Quit anyway? (y or n) y
153 dennis@ubuntuvm:~/polygon$
Using “ni”, let’s execute the first two FLD instructions.

Let’s examine the FPU registers (line 33).
As it was mentioned before, the FPU registers set is a circular buffer rather than a stack ( 1.25.5 on
page 224). And GDB doesn’t show STx registers, but internal the FPU registers (Rx). The arrow (at line
35) points to the current top of the stack.
You can also see the TOP register contents in Status Word (line 36-37)—it is 6 now, so the stack top is now
258
pointing to internal register 6.
The values of a and b are swapped after FXCH is executed (line 54).
FUCOMI is executed (line 83). Let’s see the flags: CF is set (line 95).
FCMOVBE has copied the value of b (see line 104).
FSTP leaves one value at the top of stack (line 139). The value of TOP is now 7, so the FPU stack top is
pointing to internal register 7.
ARM

VMOV D16, R2, R3 ; b
VMOV D17, R0, R1 ; a
VCMPE.F64 D17, D16
VMRS APSR_nzcv, FPSCR
VMOVGT.F64 D16, D17 ; copy "a" to D16
VMOV R0, R1, D16
BX LR
A very simple case. The input values are placed into the D17 and D16 registers and then compared using
the VCMPE instruction.
Just like in the x86 coprocessor, the ARM coprocessor has its own status and flags register (FPSCR117 ),
since there is a necessity to store coprocessor-specific flags. And just like in x86, there are no conditional
jump instruction in ARM, that can check bits in the status register of the coprocessor. So there is VMRS,
which copies 4 bits (N, Z, C, V) from the coprocessor status word into bits of the general status register
(APSR118 ).
VMOVGT is the analog of the MOVGT, instruction for D-registers, it executes if one operand is greater than
the other while comparing (GT—Greater Than).
If it gets executed, the value of a is to be written into D16 (that is currently stored in D17). Otherwise the
value of b stays in the D16 register.
The penultimate instruction VMOV prepares the value in the D16 register for returning it via the R0 and R1
register pair.

VMOV D16, R2, R3 ; b
VMOV D17, R0, R1 ; a
VCMPE.F64 D17, D16
IT GT
VMOVGT.F64 D16, D17
VMOV R0, R1, D16
BX LR
Almost the same as in the previous example, however slightly different. As we already know, many
instructions in ARM mode can be supplemented by condition predicate. But there is no such thing in
Thumb mode. There is no space in the 16-bit instructions for 4 more bits in which conditions can be
encoded.
However, Thumb-2 was extended to make it possible to specify predicates to old Thumb instructions. Here,
in the IDA-generated listing, we see the VMOVGT instruction, as in previous example.
In fact, the usual VMOV is encoded there, but IDA adds the -GT suffix to it, since there is a IT GT instruction
placed right before it.
117 (ARM) Floating-Point Status and Control Register
118 (ARM) Application Program Status Register
259
The IT instruction defines a so-called if-then block.
After the instruction it is possible to place up to 4 instructions, each of them has a predicate suffix. In our
example, IT GT implies that the next instruction is to be executed, if the GT (Greater Than) condition is
true.
Here is a more complex code fragment, by the way, from Angry Birds (for iOS):
Listing 1.222: Angry Birds Classic
...
ITE NE
VMOVNE R2, R3, D16
VMOVEQ R2, R3, D17
BLX _objc_msgSend ; not suffixed
...
ITE stands for if-then-else

and it encodes suffixes for the next two instructions.
The first instruction executes if the condition encoded in ITE (NE, not equal) is true at, and the second—if
the condition is not true. (The inverse condition of NE is EQ (equal)).
The instruction followed after the second VMOV (or VMOVEQ) is a normal one, not suffixed (BLX).
One more that’s slightly harder, which is also from Angry Birds:
...
ITTTT EQ
MOVEQ R0, R4
ADDEQ SP, SP, #0x20
POPEQ.W {R8,R10}
POPEQ {R4-R7,PC}
BLX ___stack_chk_fail ; not suffixed
...
Four “T” symbols in the instruction mnemonic mean that the four subsequent instructions are to be exe-
cuted if the condition is true.
That’s why IDA adds the -EQ suffix to each one of them.
And if there was, for example, ITEEE EQ (if-then-else-else-else), then the suffixes would have been set as
follows:
-EQ
-NE
-NE
-NE
Another fragment from Angry Birds:

...
CMP.W R0, #0xFFFFFFFF
ITTE LE
SUBLE.W R10, R0, #1
NEGLE R0, R0
MOVGT R10, R0
MOVS R6, #0 ; not suffixed
CBZ R0, loc_1E7E32 ; not suffixed
...
ITTE (if-then-then-else)
implies that the 1st and 2nd instructions are to be executed if the LE (Less or Equal) condition is true, and
the 3rd—if the inverse condition (GT—Greater Than) is true.
Compilers usually don’t generate all possible combinations.
For example, in the mentioned Angry Birds game (classic version for iOS) only these variants of the IT
instruction are used: IT, ITE, ITT, ITTE, ITTT, ITTTT. How to learn this? In IDA, it is possible to produce
260
listing files, so it was created with an option to show 4 bytes for each opcode. Then, knowing the high
part of the 16-bit opcode (IT is 0xBF), we do the following using grep:
cat AngryBirdsClassic.lst | grep " BF" | grep "IT" > results.lst
By the way, if you program in ARM assembly language manually for Thumb-2 mode, and you add condi-
tional suffixes, the assembler will add the IT instructions automatically with the required flags where it is
necessary.
Non-optimizing Xcode 4.6.3 (LLVM) (ARM mode)
Listing 1.225: Non-optimizing Xcode 4.6.3 (LLVM) (ARM mode)

b = -0x20
a = -0x18
val_to_return = -0x10
saved_R7 = -4
STR R7, [SP,#saved_R7]!

MOV R7, SP
SUB SP, SP, #0x1C
BIC SP, SP, #7
VMOV D16, R2, R3
VMOV D17, R0, R1
VSTR D17, [SP,#0x20+a]
VSTR D16, [SP,#0x20+b]
VLDR D16, [SP,#0x20+a]
VLDR D17, [SP,#0x20+b]
VCMPE.F64 D16, D17
BLE loc_2E08
VLDR D16, [SP,#0x20+a]
VSTR D16, [SP,#0x20+val_to_return]
B loc_2E10
loc_2E08
VLDR D16, [SP,#0x20+b]
VSTR D16, [SP,#0x20+val_to_return]
loc_2E10
VLDR D16, [SP,#0x20+val_to_return]
VMOV R0, R1, D16
MOV SP, R7
LDR R7, [SP+0x20+b],#4
BX LR
Almost the same as we already saw, but there is too much redundant code because the a and b variables
are stored in the local stack, as well as the return value.

PUSH {R3-R7,LR}
MOVS R4, R2
MOVS R5, R3
MOVS R6, R0
MOVS R7, R1
BL __aeabi_cdrcmple
BCS loc_1C0
MOVS R0, R6
MOVS R1, R7
POP {R3-R7,PC}
loc_1C0
MOVS R0, R4
MOVS R1, R5
POP {R3-R7,PC}
261
Keil doesn’t generate FPU-instructions since it cannot rely on them being supported on the target CPU,
and it cannot be done by straightforward bitwise comparing. So it calls an external library function to do
the comparison: __aeabi_cdrcmple.
N.B. The result of the comparison is to be left in the flags by this function, so the following BCS (Carry
set—Greater than or equal) instruction can work without any additional code.
ARM64
d_max:
; D0 - a, D1 - b
fcmpe d0, d1
fcsel d0, d0, d1, gt
; now result in D0
ret
The ARM64 ISA has FPU-instructions which set APSR the CPU flags instead of FPSCR for convenience.
TheFPU is not a separate device here anymore (at least, logically). Here we see FCMPE. It compares the
two values passed in D0 and D1 (which are the first and second arguments of the function) and sets APSR
flags (N, Z, C, V).
FCSEL (Floating Conditional Select) copies the value of D0 or D1 into D0 depending on the condition (GT—
Greater Than), and again, it uses flags in APSR register instead of FPSCR.
This is much more convenient, compared to the instruction set in older CPUs.
If the condition is true (GT), then the value of D0 is copied into D0 (i.e., nothing happens). If the condition
is not true, the value of D1 is copied into D0.
d_max:
; save input arguments in "Register Save Area"
sub sp, sp, #16
str d0, [sp,8]
str d1, [sp]
; reload values
ldr x1, [sp,8]
ldr x0, [sp]
fmov d0, x1
fmov d1, x0
; D0 - a, D1 - b
fcmpe d0, d1
ble .L76
; a>b; load D0 (a) into X0
ldr x0, [sp,8]
b .L74
.L76:
; a<=b; load D1 (b) into X0
ldr x0, [sp]
.L74:
; result in X0
fmov d0, x0
; result in D0
add sp, sp, 16
ret

First, the function saves its input argument values in the local stack (Register Save Area). Then the code
reloads these values into registers X0/X1 and finally copies them to D0/D1 to be compared using FCMPE. A
lot of redundant code, but that is how non-optimizing compilers work. FCMPE compares the values and
sets the APSR flags. At this moment, the compiler is not thinking yet about the more convenient FCSEL
instruction, so it proceed using old methods: using the BLE instruction (Branch if Less than or Equal). In
the first case (a > b), the value of a gets loaded into X0. In the other case (a <= b), the value of b gets
262
loaded into X0. Finally, the value from X0 gets copied into D0, because the return value needs to be in this
register.
Exercise
As an exercise, you can try optimizing this piece of code manually by removing redundant instructions
and not introducing new ones (including FCSEL).
Optimizing GCC (Linaro) 4.9—float
Let’s also rewrite this example to use float instead of double.

float f_max (float a, float b)
{
if (a>b)
return a;
return b;
};
f_max:
; S0 - a, S1 - b
fcmpe s0, s1
fcsel s0, s0, s1, gt
; now result in S0
ret
It is the same code, but the S-registers are used instead of D- ones. It’s because numbers of type float
are passed in 32-bit S-registers (which are in fact the lower parts of the 64-bit D-registers).
MIPS
The co-processor of the MIPS processor has a condition bit which can be set in the FPU and checked in the
CPU.
Earlier MIPS-es have only one condition bit (called FCC0), later ones have 8 (called FCC7-FCC0).
This bit (or bits) are located in the register called FCCR.

d_max:
; set FPU condition bit if $f14<$f12 (b<a):
c.lt.d $f14, $f12
or $at, $zero ; NOP
; jump to locret_14 if condition bit is set
bc1t locret_14
; this instruction is always executed (set return value to "a"):
mov.d $f0, $f12 ; branch delay slot
; this instruction is executed only if branch was not taken (i.e., if b>=a)
; set return value to "b":
mov.d $f0, $f14
locret_14:
jr $ra
C.LT.D compares two values. LT is the condition “Less Than”. D implies values of type double. Depending
on the result of the comparison, the FCC0 condition bit is either set or cleared.
BC1T checks the FCC0 bit and jumps if the bit is set. T means that the jump is to be taken if the bit is set
(“True”). There is also the instruction BC1F which jumps if the bit is cleared (“False”).
Depending on the jump, one of function arguments is placed into $F0.
263
1.25.8 Some constants
It’s easy to find representations of some constants in Wikipedia for IEEE 754 encoded numbers. It’s
interesting to know that 0.0 in IEEE 754 is represented as 32 zero bits (for single precision) or 64 zero bits
(for double). So in order to set a floating point variable to 0.0 in register or memory, one can use MOV or
XOR reg, reg instruction. This is suitable for structures where many variables present of various data
types. With usual memset() function one can set all integer variables to 0, all boolean variables to false,
all pointers to NULL, and all floating point variables (of any precision) to 0.0.
1.25.9 Copying
One may think inertially that FLD/FST instructions must be used to load and store (and hence, copy) IEEE
754 values. Nevertheless, same can be achieved easier by usual MOV instruction, which, of course, copies
values bitwisely.
1.25.10 Stack, calculators and reverse Polish notation

Now we understand why some old programmable calculators use reverse Polish notation.
For example, for addition of 12 and 34 one has to enter 12, then 34, then press “plus” sign.
It’s because old calculators were just stack machine implementations, and this was much simpler than to
handle complex parenthesized expressions.
Such a calculator still present in many Unix distributions: dc.
1.25.11 80 bits?
Internal numbers representation in FPU — 80-bit. Strange number, because the number not in 2n form.
There is a hypothesis that this is probably due to historical reasons—the standard IBM puched card can
encode 12 rows of 80 bits. 80 ⋅ 25 text mode resolution was also popular in past.
Wikipedia has another explanation: https://en.wikipedia.org/wiki/Extended_precision.
If you know better, please a drop email to the author: <first_name @ last_name . com> or <first_name .
last_name @ gmail . com>.
1.25.12 x64
On how floating point numbers are processed in x86-64, read more here: 1.38 on page 425.
1.25.13 Exercises
1.26 Arrays
An array is just a set of variables in memory that lie next to each other and that have the same type119 .

#include <stdio.h>
int main()
{
int a[20];
int i;
for (i=0; i<20; i++)

a[i]=i*2;
for (i=0; i<20; i++)

119 AKA “homogeneous container”
264
printf ("a[%d]=%d\n", i, a[i]);
return 0;
};
x86
MSVC
Let’s compile:

_TEXT SEGMENT
_i$ = -84 ; size = 4
_a$ = -80 ; size = 80
_main PROC
push ebp
mov ebp, esp
sub esp, 84 ; 00000054H
mov DWORD PTR _i$[ebp], 0
jmp SHORT $LN6@main
$LN5@main:
mov eax, DWORD PTR _i$[ebp]
add eax, 1
$LN6@main:
cmp DWORD PTR _i$[ebp], 20 ; 00000014H
jge SHORT $LN4@main
mov ecx, DWORD PTR _i$[ebp]
shl ecx, 1
mov edx, DWORD PTR _i$[ebp]
mov DWORD PTR _a$[ebp+edx*4], ecx
jmp SHORT $LN5@main
$LN4@main:
jmp SHORT $LN3@main
$LN2@main:
add eax, 1
$LN3@main:
jge SHORT $LN1@main
mov edx, DWORD PTR _a$[ebp+ecx*4]
push edx
push eax
push OFFSET $SG2463
call _printf
add esp, 12 ; 0000000cH
jmp SHORT $LN2@main
$LN1@main:
xor eax, eax
mov esp, ebp
pop ebp
ret 0
_main ENDP
Nothing very special, just two loops: the first is a filling loop and second is a printing loop. The shl ecx,
1 instruction is used for value multiplication by 2 in ECX, more about it: 1.24.2 on page 216.
80 bytes are allocated on the stack for the array, 20 elements of 4 bytes.
265
Let’s try this example in OllyDbg.
We see how the array gets filled:
each element is 32-bit word of int type and its value is the index multiplied by 2:
Figure 1.88: OllyDbg: after array filling
Since this array is located in the stack, we can see all its 20 elements there.
GCC
Here is what GCC 4.4.1 does:

public main
main proc near ; DATA XREF: _start+17

i_2 = dword ptr -54h
i = dword ptr -4
push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
266
sub esp, 70h
mov [esp+70h+i], 0 ; i=0
jmp short loc_804840A
loc_80483F7:
mov eax, [esp+70h+i]
mov edx, [esp+70h+i]
add edx, edx ; edx=i*2
mov [esp+eax*4+70h+i_2], edx
add [esp+70h+i], 1 ; i++
loc_804840A:
cmp [esp+70h+i], 13h
jle short loc_80483F7
mov [esp+70h+i], 0
jmp short loc_8048441
loc_804841B:
mov eax, [esp+70h+i]
mov edx, [esp+eax*4+70h+i_2]
mov eax, offset aADD ; "a[%d]=%d\n"
mov edx, [esp+70h+i]
call _printf
add [esp+70h+i], 1
loc_8048441:
cmp [esp+70h+i], 13h
jle short loc_804841B
mov eax, 0
leave
retn
main endp
By the way, variable a is of type int* (the pointer to int)—you can pass a pointer to an array to another
function, but it’s more correct to say that a pointer to the first element of the array is passed (the addresses
of rest of the elements are calculated in an obvious way).
If you index this pointer as a[idx], idx is just to be added to the pointer and the element placed there (to
which calculated pointer is pointing) is to be returned.
An interesting example: a string of characters like string is an array of characters and it has a type of
const char[].
An index can also be applied to this pointer.
And that is why it is possible to write things like “string”[i]—this is a correct C/C++ expression!
ARM
EXPORT _main
_main
STMFD SP!, {R4,LR}
SUB SP, SP, #0x50 ; allocate place for 20 int variables
; first loop
MOV R4, #0 ; i
B loc_4A0
loc_494
MOV R0, R4,LSL#1 ; R0=R4*2
STR R0, [SP,R4,LSL#2] ; store R0 to SP+R4<<2 (same as SP+R4*4)
ADD R4, R4, #1 ; i=i+1
loc_4A0
267
CMP R4, #20 ; i<20?
BLT loc_494 ; yes, run loop body again
; second loop
MOV R4, #0 ; i
B loc_4C4
loc_4B0
LDR R2, [SP,R4,LSL#2] ; (second printf argument) R2=*(SP+R4<<4) (same as
*(SP+R4*4))
MOV R1, R4 ; (first printf argument) R1=i
ADR R0, aADD ; "a[%d]=%d\n"
BL __2printf
ADD R4, R4, #1 ; i=i+1
loc_4C4
CMP R4, #20 ; i<20?
BLT loc_4B0 ; yes, run loop body again
MOV R0, #0 ; value to return
ADD SP, SP, #0x50 ; deallocate chunk, allocated for 20 int variables
LDMFD SP!, {R4,PC}
int type requires 32 bits for storage (or 4 bytes),

so to store 20 int variables 80 (0x50) bytes are needed. So that is why the SUB SP, SP, #0x50
instruction in the function’s prologue allocates exactly this amount of space in the stack.
In both the first and second loops, the loop iterator i is placed in the R4 register.
The number that is to be written into the array is calculated as i ∗ 2, which is effectively equivalent to
shifting it left by one bit,
so MOV R0, R4,LSL#1 instruction does this.
STR R0, [SP,R4,LSL#2] writes the contents of R0 into the array.
Here is how a pointer to array element is calculated: SP points to the start of the array, R4 is i.
So shifting i left by 2 bits is effectively equivalent to multiplication by 4 (since each array element has a
size of 4 bytes) and then it’s added to the address of the start of the array.
The second loop has an inverse LDR R2, [SP,R4,LSL#2] instruction. It loads the value we need from the
array, and the pointer to it is calculated likewise.
_main
PUSH {R4,R5,LR}
; allocate place for 20 int variables + one more variable
SUB SP, SP, #0x54
; first loop
MOVS R0, #0 ; i
MOV R5, SP ; pointer to first array element
loc_1CE
LSLS R1, R0, #1 ; R1=i<<1 (same as i*2)
ADDS R0, R0, #1 ; i=i+1
CMP R0, #20 ; i<20?
STR R1, [R5,R2] ; store R1 to *(R5+R2) (same R5+i*4)
BLT loc_1CE ; yes, i<20, run loop body again
; second loop
MOVS R4, #0 ; i=0

loc_1DC
LDR R2, [R5,R0] ; load from *(R5+R0) (same as R5+i*4)
MOVS R1, R4
268
ADR R0, aADD ; "a[%d]=%d\n"
BL __2printf
ADDS R4, R4, #1 ; i=i+1
CMP R4, #20 ; i<20?
BLT loc_1DC ; yes, i<20, run loop body again
MOVS R0, #0 ; value to return
; deallocate chunk, allocated for 20 int variables + one more variable
ADD SP, SP, #0x54
POP {R4,R5,PC}
Thumb code is very similar.

Thumb mode has special instructions for bit shifting (like LSLS), which calculates the value to be written
into the array and the address of each element in the array as well.
The compiler allocates slightly more space in the local stack, however, the last 4 bytes are not used.
Non-optimizing GCC 4.9.1 (ARM64)
Listing 1.230: Non-optimizing GCC 4.9.1 (ARM64)

.LC0:
.string "a[%d]=%d\n"
main:
stp x29, x30, [sp, -112]!
; set stack frame (FP=SP)
add x29, sp, 0
; setting initial counter variable at 0 (WZR is the register always holding zero):
str wzr, [x29,108]
; jump to loop condition checking code:
b .L2
.L3:
; load value of "i" variable:
ldr w0, [x29,108]
; multiplicate it by 2:
lsl w2, w0, 1
; find a place of an array in local stack:
add x0, x29, 24
; load 32-bit integer from local stack and sign extend it to 64-bit one:
ldrsw x1, [x29,108]
; calculate address of element (X0+X1<<2=array address+i*4) and store W2 (i*2) there:
str w2, [x0,x1,lsl 2]
ldr w0, [x29,108]
add w0, w0, 1
str w0, [x29,108]
.L2:
; check if we finished:
ldr w0, [x29,108]
cmp w0, 19
; jump to L3 (loop body begin) if not:
ble .L3
; second part of the function begins here.
; setting initial counter variable at 0.
; by the way, the same place in the local stack was used for counter,
; because the same local variable (i) is being used as counter.
str wzr, [x29,108]
b .L4
.L5:
; calculate array address:
add x0, x29, 24
; load "i" value:
ldrsw x1, [x29,108]
; load value from the array at the address (X0+X1<<2 = address of array + i*4)
ldr w2, [x0,x1,lsl 2]
; load address of the "a[%d]=%d\n" string:
adrp x0, .LC0
269
; load "i" variable to W1 and pass it to printf() as second argument:
ldr w1, [x29,108]
; W2 still contains the value of array element which was just loaded.
; call printf():
bl printf
; increment "i" variable:
ldr w0, [x29,108]
add w0, w0, 1
str w0, [x29,108]
.L4:
; are we finished?
ldr w0, [x29,108]
cmp w0, 19
; jump to the loop body begin if not:
ble .L5
; return 0
mov w0, 0
; restore FP and LR:
ldp x29, x30, [sp], 112
ret
MIPS
The function uses a lot of S- registers which must be preserved, so that’s why its values are saved in the
function prologue and restored in the epilogue.

main:
var_70 = -0x70
var_68 = -0x68
var_14 = -0x14
var_10 = -0x10
var_C = -0xC
var_8 = -8
var_4 = -4
addiu $sp, -0x80
la $gp, (__gnu_local_gp & 0xFFFF)
sw $ra, 0x80+var_4($sp)
sw $s3, 0x80+var_8($sp)
sw $s2, 0x80+var_C($sp)
sw $s1, 0x80+var_10($sp)
sw $s0, 0x80+var_14($sp)
sw $gp, 0x80+var_70($sp)
addiu $s1, $sp, 0x80+var_68
move $v1, $s1
move $v0, $zero
; that value will be used as a loop terminator.
; it was precalculated by GCC compiler at compile stage:
li $a0, 0x28 # '('
loc_34: # CODE XREF: main+3C

; store value into memory:
sw $v0, 0($v1)
; increase value to be stored by 2 at each iteration:
addiu $v0, 2
; loop terminator reached?
bne $v0, $a0, loc_34
; add 4 to address anyway:
addiu $v1, 4
; array filling loop is ended
; second loop begin
la $s3, $LC0 # "a[%d]=%d\n"
; "i" variable will reside in $s0:
move $s0, $zero
li $s2, 0x14
270
loc_54: # CODE XREF: main+70
; call printf():
lw $t9, (printf & 0xFFFF)($gp)
lw $a2, 0($s1)
move $a1, $s0
move $a0, $s3
jalr $t9
; increment "i":
addiu $s0, 1
lw $gp, 0x80+var_70($sp)
; jump to loop body if end is not reached:
bne $s0, $s2, loc_54
; move memory pointer to the next 32-bit word:
addiu $s1, 4
; function epilogue
lw $ra, 0x80+var_4($sp)
move $v0, $zero
lw $s3, 0x80+var_8($sp)
lw $s2, 0x80+var_C($sp)
lw $s1, 0x80+var_10($sp)
lw $s0, 0x80+var_14($sp)
jr $ra
addiu $sp, 0x80
$LC0: .ascii "a[%d]=%d\n"<0> # DATA XREF: main+44
Something interesting: there are two loops and the first one doesn’t need i, it needs only i ∗ 2 (increased
by 2 at each iteration) and also the address in memory (increased by 4 at each iteration).
So here we see two variables, one (in $V0) increasing by 2 each time, and another (in $V1) — by 4.
The second loop is where printf() is called and it reports the value of i to the user, so there is a variable
which is increased by 1 each time (in $S0) and also a memory address (in $S1) increased by 4 each time.
That reminds us of loop optimizations: 3.10 on page 490.
Their goal is to get rid of multiplications.
1.26.2 Buffer overflow

Reading outside array bounds
So, array indexing is just array[index]. If you study the generated code closely, you’ll probably note the
missing index bounds checking, which could check if it is less than 20. What if the index is 20 or greater?
That’s the one C/C++ feature it is often blamed for.
Here is a code that successfully compiles and works:
#include <stdio.h>
int main()
{
int a[20];
int i;
for (i=0; i<20; i++)

a[i]=i*2;
printf ("a[20]=%d\n", a[20]);
return 0;
};
Compilation results (MSVC 2008):

$SG2474 DB 'a[20]=%d', 0aH, 00H
_i$ = -84 ; size = 4
271
_a$ = -80 ; size = 80
_main PROC
push ebp
mov ebp, esp
sub esp, 84
jmp SHORT $LN3@main
$LN2@main:
add eax, 1
$LN3@main:
cmp DWORD PTR _i$[ebp], 20
jge SHORT $LN1@main
shl ecx, 1
mov edx, DWORD PTR _i$[ebp]
mov DWORD PTR _a$[ebp+edx*4], ecx
jmp SHORT $LN2@main
$LN1@main:
mov eax, DWORD PTR _a$[ebp+80]
push eax
push OFFSET $SG2474 ; 'a[20]=%d'
add esp, 8
xor eax, eax
mov esp, ebp
pop ebp
ret 0
_main ENDP
_TEXT ENDS
END
The code produced this result:
Listing 1.233: OllyDbg: console output

a[20]=1638280
It is just something that has been lying in the stack near to the array, 80 bytes away from its first element.
272
Let’s try to find out where did this value come from, using OllyDbg.
Let’s load and find the value located right after the last array element:
Figure 1.89: OllyDbg: reading of the 20th element and execution of printf()
What is this? Judging by the stack layout, this is the saved value of the EBP register.
273
Let’s trace further and see how it gets restored:
Figure 1.90: OllyDbg: restoring value of EBP
Indeed, how it could be different? The compiler may generate some additional code to check the index
value to be always in the array’s bounds (like in higher-level programming languages120 ) but this makes
the code slower.
Writing beyond array bounds
OK, we read some values from the stack illegally, but what if we could write something to it?
Here is what we have got:
#include <stdio.h>
int main()
{
int a[20];
int i;
for (i=0; i<30; i++)

a[i]=i;
return 0;
};
120 Java, Python, etc.
274
MSVC
And what we get:

_TEXT SEGMENT
_i$ = -84 ; size = 4
_a$ = -80 ; size = 80
_main PROC
push ebp
mov ebp, esp
sub esp, 84
jmp SHORT $LN3@main
$LN2@main:
add eax, 1
$LN3@main:
cmp DWORD PTR _i$[ebp], 30 ; 0000001eH
jge SHORT $LN1@main
mov edx, DWORD PTR _i$[ebp] ; that instruction is obviously redundant
mov DWORD PTR _a$[ebp+ecx*4], edx ; ECX could be used as second operand here instead
jmp SHORT $LN2@main
$LN1@main:
xor eax, eax
mov esp, ebp
pop ebp
ret 0
_main ENDP
The compiled program crashes after running. No wonder. Let’s see where exactly does it crash.
275
Let’s load it into OllyDbg, and trace until all 30 elements are written:
Figure 1.91: OllyDbg: after restoring the value of EBP
276
Trace until the function end:
Figure 1.92: OllyDbg: EIP has been restored, but OllyDbg can’t disassemble at 0x15
Now please keep your eyes on the registers.

EIP is 0x15 now. It is not a legal address for code—at least for win32 code! We got there somehow against
our will. It is also interesting that the EBP register contain 0x14, ECX and EDX contain 0x1D.
Let’s study stack layout a bit more.
After the control flow has been passed to main(), the value in the EBP register was saved on the stack.
Then, 84 bytes were allocated for the array and the i variable. That’s (20+1)*sizeof(int). ESP now
points to the _i variable in the local stack and after the execution of the next PUSH something, something
is appearing next to _i.
That’s the stack layout while the control is in main():
ESP 4 bytes allocated for i variable
ESP+4 80 bytes allocated for a[20] array
ESP+84 saved EBP value
ESP+88 return address
a[19]=something statement writes the last int in the bounds of the array (in bounds so far!).
a[20]=something statement writes something to the place where the value of EBP is saved.
Please take a look at the register state at the moment of the crash. In our case, 20 has been written in the
20th element. At the function end, the function epilogue restores the original EBP value. (20 in decimal
277
is 0x14 in hexadecimal). Then RET gets executed, which is effectively equivalent to POP EIP instruction.
The RET instruction takes the return address from the stack (that is the address in CRT, which has called
main()), and 21 is stored there (0x15 in hexadecimal). The CPU traps at address 0x15, but there is no
executable code there, so exception gets raised.
Welcome! It is called a buffer overflow121 .
Replace the int array with a string (char array), create a long string deliberately and pass it to the program,
to the function, which doesn’t check the length of the string and copies it in a short buffer, and you’ll able
to point the program to an address to which it must jump. It’s not that simple in reality, but that is how it
emerged. Classic article about it: [Aleph One, Smashing The Stack For Fun And Profit, (1996)]122 .
GCC
Let’s try the same code in GCC 4.4.1. We get:

public main
main proc near
a = dword ptr -54h

i = dword ptr -4
push ebp
mov ebp, esp
sub esp, 60h ; 96
mov [ebp+i], 0
jmp short loc_80483D1
loc_80483C3:
mov eax, [ebp+i]
mov edx, [ebp+i]
mov [ebp+eax*4+a], edx
add [ebp+i], 1
loc_80483D1:
cmp [ebp+i], 1Dh
jle short loc_80483C3
mov eax, 0
leave
retn
main endp
Running this in Linux will produce: Segmentation fault.

If we run this in the GDB debugger, we get this:
(gdb) r
Starting program: /home/dennis/RE/1
Program received signal SIGSEGV, Segmentation fault.

0x00000016 in ?? ()
eax 0x0 0
ecx 0xd2f96388 -755407992
edx 0x1d 29
ebx 0x26eff4 2551796
esp 0xbffff4b0 0xbffff4b0
ebp 0x15 0x15
esi 0x0 0
edi 0x0 0
eip 0x16 0x16
eflags 0x10202 [ IF RF ]
cs 0x73 115
ss 0x7b 123
ds 0x7b 123
es 0x7b 123
fs 0x0 0
gs 0x33 51
121 wikipedia
278
(gdb)
The register values are slightly different than in win32 example, since the stack layout is slightly different
too.
1.26.3 Buffer overflow protection methods

There are several methods to protect against this scourge, regardless of the C/C++ programmers’ negli-
gence. MSVC has options like123 :
/RTCs Stack Frame runtime checking
/GZ Enable stack checks (/RTCs)
One of the methods is to write a random value between the local variables in stack at function prologue
and to check it in function epilogue before the function exits. If value is not the same, do not execute the
last instruction RET, but stop (or hang). The process will halt, but that is much better than a remote attack
to your host.
This random value is called a “canary” sometimes, it is related to the miners’ canary124 , they were used
by miners in the past days in order to detect poisonous gases quickly.
Canaries are very sensitive to mine gases, they become very agitated in case of danger, or even die.
If we compile our very simple array example ( 1.26.1 on page 264) in MSVC with RTC1 and RTCs option,
you can see a call to @_RTC_CheckStackVars@8 a function at the end of the function that checks if the
“canary” is correct.
Let’s see how GCC handles this. Let’s take an alloca() ( 1.9.2 on page 34) example:
#ifdef __GNUC__
#include <alloca.h> // GCC
#else
#include <malloc.h> // MSVC
#endif
#include <stdio.h>
void f()
{
char *buf=(char*)alloca (600);
#ifdef __GNUC__
snprintf (buf, 600, "hi! %d, %d, %d\n", 1, 2, 3); // GCC
#else
_snprintf (buf, 600, "hi! %d, %d, %d\n", 1, 2, 3); // MSVC
#endif
puts (buf);
};
By default, without any additional options, GCC 4.7.3 inserts a “canary” check into the code:

.LC0:
f:
push ebp
mov ebp, esp
push ebx
sub esp, 676
lea ebx, [esp+39]
and ebx, -16
mov DWORD PTR [esp+8], OFFSET FLAT:.LC0 ; "hi! %d, %d, %d\n"
mov DWORD PTR [esp], ebx
mov eax, DWORD PTR gs:20 ; canary
123 compiler-side buffer overflow protection methods: wikipedia.org/wiki/Buffer_overflow_protection
124 wikipedia.org/wiki/Domestic_canary#Miner.27s_canary
279
mov DWORD PTR [ebp-12], eax
xor eax, eax
call _snprintf
call puts
mov eax, DWORD PTR [ebp-12]
xor eax, DWORD PTR gs:20 ; check canary
jne .L5
mov ebx, DWORD PTR [ebp-4]
leave
ret
.L5:
call __stack_chk_fail
The random value is located in gs:20. It gets written on the stack and then at the end of the function
the value in the stack is compared with the correct “canary” in gs:20. If the values are not equal, the
__stack_chk_fail function is called and we can see in the console something like that (Ubuntu 13.04
x86):
*** buffer overflow detected ***: ./2_1 terminated
======= Backtrace: =========
/lib/i386-linux-gnu/libc.so.6(__fortify_fail+0x63)[0xb7699bc3]
/lib/i386-linux-gnu/libc.so.6(+0x10593a)[0xb769893a]
/lib/i386-linux-gnu/libc.so.6(+0x105008)[0xb7698008]
/lib/i386-linux-gnu/libc.so.6(_IO_default_xsputn+0x8c)[0xb7606e5c]
/lib/i386-linux-gnu/libc.so.6(_IO_vfprintf+0x165)[0xb75d7a45]
/lib/i386-linux-gnu/libc.so.6(__vsprintf_chk+0xc9)[0xb76980d9]
/lib/i386-linux-gnu/libc.so.6(__sprintf_chk+0x2f)[0xb7697fef]
./2_1[0x8048404]
/lib/i386-linux-gnu/libc.so.6(__libc_start_main+0xf5)[0xb75ac935]
======= Memory map: ========
08048000-08049000 r-xp 00000000 08:01 2097586 /home/dennis/2_1
08049000-0804a000 r--p 00000000 08:01 2097586 /home/dennis/2_1
0804a000-0804b000 rw-p 00001000 08:01 2097586 /home/dennis/2_1
094d1000-094f2000 rw-p 00000000 00:00 0 [heap]
b7560000-b757b000 r-xp 00000000 08:01 1048602 /lib/i386-linux-gnu/libgcc_s.so.1
b757b000-b757c000 r--p 0001a000 08:01 1048602 /lib/i386-linux-gnu/libgcc_s.so.1
b757c000-b757d000 rw-p 0001b000 08:01 1048602 /lib/i386-linux-gnu/libgcc_s.so.1
b7592000-b7593000 rw-p 00000000 00:00 0
b7593000-b7740000 r-xp 00000000 08:01 1050781 /lib/i386-linux-gnu/libc-2.17.so
b7740000-b7742000 r--p 001ad000 08:01 1050781 /lib/i386-linux-gnu/libc-2.17.so
b7742000-b7743000 rw-p 001af000 08:01 1050781 /lib/i386-linux-gnu/libc-2.17.so
b7743000-b7746000 rw-p 00000000 00:00 0
b775a000-b775d000 rw-p 00000000 00:00 0
b775d000-b775e000 r-xp 00000000 00:00 0 [vdso]
b775e000-b777e000 r-xp 00000000 08:01 1050794 /lib/i386-linux-gnu/ld-2.17.so
b777e000-b777f000 r--p 0001f000 08:01 1050794 /lib/i386-linux-gnu/ld-2.17.so
b777f000-b7780000 rw-p 00020000 08:01 1050794 /lib/i386-linux-gnu/ld-2.17.so
bff35000-bff56000 rw-p 00000000 00:00 0 [stack]
Aborted (core dumped)
gs is the so-called segment register. These registers were used widely in MS-DOS and DOS-extenders
times. Today, its function is different.
To say it briefly, the gs register in Linux always points to the TLS125 ( 6.2 on page 734)—some information
specific to thread is stored there. By the way, in win32 the fs register plays the same role, pointing to
TIB126 127 .
More information can be found in the Linux kernel source code (at least in 3.11 version),
in arch/x86/include/asm/stackprotector.h this variable is described in the comments.
Let’s get back to our simple array example ( 1.26.1 on page 264),
again, now we can see how LLVM checks the correctness of the “canary”:
125 Thread Local Storage
126 Thread Information Block
127 wikipedia.org/wiki/Win32_Thread_Information_Block
280
_main
var_64 = -0x64
var_60 = -0x60
var_5C = -0x5C
var_58 = -0x58
var_54 = -0x54
var_50 = -0x50
var_4C = -0x4C
var_48 = -0x48
var_44 = -0x44
var_40 = -0x40
var_3C = -0x3C
var_38 = -0x38
var_34 = -0x34
var_30 = -0x30
var_2C = -0x2C
var_28 = -0x28
var_24 = -0x24
var_20 = -0x20
var_1C = -0x1C
var_18 = -0x18
canary = -0x14
var_10 = -0x10
PUSH {R4-R7,LR}
ADD R7, SP, #0xC
STR.W R8, [SP,#0xC+var_10]!
SUB SP, SP, #0x54
MOVW R0, #aObjc_methtype ; "objc_methtype"
MOVS R2, #0
MOVT.W R0, #0
MOVS R5, #0
ADD R0, PC
LDR.W R8, [R0]
LDR.W R0, [R8]
STR R0, [SP,#0x64+canary]
MOVS R0, #2
STR R2, [SP,#0x64+var_64]
MOVS R0, #4
STR R0, [SP,#0x64+var_5C]
MOVS R0, #6
MOVS R0, #8
MOVS R0, #0xA
MOVS R0, #0xC
MOVS R0, #0xE
MOVS R0, #0x10
MOVS R0, #0x12
MOVS R0, #0x14
MOVS R0, #0x16
MOVS R0, #0x18
MOVS R0, #0x1A
MOVS R0, #0x1C
MOVS R0, #0x1E
MOVS R0, #0x20
281
MOVS R0, #0x22
MOVS R0, #0x24
MOVS R0, #0x26
MOV R4, 0xFDA ; "a[%d]=%d\n"
MOV R0, SP
ADDS R6, R0, #4
ADD R4, PC
B loc_2F1C
; second loop begin
loc_2F14
ADDS R0, R5, #1
LDR.W R2, [R6,R5,LSL#2]
MOV R5, R0
loc_2F1C
MOV R0, R4
MOV R1, R5
BLX _printf
CMP R5, #0x13
BNE loc_2F14
LDR.W R0, [R8]
LDR R1, [SP,#0x64+canary]
CMP R0, R1
ITTTT EQ ; is canary still correct?
MOVEQ R0, #0
ADDEQ SP, SP, #0x54
LDREQ.W R8, [SP+0x64+var_64],#4
POPEQ {R4-R7,PC}
BLX ___stack_chk_fail
First of all, as we see, LLVM “unrolled” the loop and all values were written into an array one-by-one, pre-
calculated, as LLVM concluded it can work faster. By the way, instructions in ARM mode may help to do
this even faster, and finding this could be your homework.
At the function end we see the comparison of the “canaries”—the one in the local stack and the correct
one, to which R8 points.
If they are equal to each other, a 4-instruction block is triggered by ITTTT EQ, which contains writing 0 in
R0, the function epilogue and exit. If the “canaries” are not equal, the block being skipped,
and the jump to ___stack_chk_fail function will occur, which, perhaps will halt execution.
1.26.4 One more word about arrays

Now we understand why it is impossible to write something like this in C/C++ code:
void f(int size)
{
int a[size];
...
};
That’s just because the compiler must know the exact array size to allocate space for it in the local stack
layout on at the compiling stage.
If you need an array of arbitrary size, allocate it by using malloc(), then access the allocated memory
block as an array of variables of the type you need.
Or use the C99 standard feature [ISO/IEC 9899:TC3 (C C99 standard), (2007)6.7.5/2], and it works like
alloca() ( 1.9.2 on page 34) internally.
It’s also possible to use garbage collecting libraries for C.
And there are also libraries supporting smart pointers for C++.
282
1.26.5 Array of pointers to strings
Here is an example for an array of pointers.
Listing 1.236: Get month name

#include <stdio.h>
const char* month1[]=

{
"January", "February", "March", "April",
"May", "June", "July", "August",
"September", "October", "November", "December"
};
// in 0..11 range
const char* get_month1 (int month)
{
return month1[month];
};
x64

_DATA SEGMENT
month1 DQ FLAT:$SG3122
DQ FLAT:$SG3123
DQ FLAT:$SG3124
DQ FLAT:$SG3125
DQ FLAT:$SG3126
DQ FLAT:$SG3127
DQ FLAT:$SG3128
DQ FLAT:$SG3129
DQ FLAT:$SG3130
DQ FLAT:$SG3131
DQ FLAT:$SG3132
DQ FLAT:$SG3133
$SG3122 DB 'January', 00H
$SG3123 DB 'February', 00H
$SG3124 DB 'March', 00H
$SG3125 DB 'April', 00H
$SG3126 DB 'May', 00H
$SG3127 DB 'June', 00H
$SG3128 DB 'July', 00H
$SG3129 DB 'August', 00H
$SG3130 DB 'September', 00H
$SG3156 DB '%s', 0aH, 00H
$SG3131 DB 'October', 00H
$SG3132 DB 'November', 00H
$SG3133 DB 'December', 00H
_DATA ENDS
month$ = 8
get_month1 PROC
movsxd rax, ecx
lea rcx, OFFSET FLAT:month1
mov rax, QWORD PTR [rcx+rax*8]
ret 0
get_month1 ENDP
The code is very simple:

• The first MOVSXD instruction copies a 32-bit value from ECX (where month argument is passed) to RAX
with sign-extension (because the month argument is of type int).
The reason for the sign extension is that this 32-bit value is to be used in calculations with other
64-bit values.
283
Hence, it has to be promoted to 64-bit128 .
• Then the address of the pointer table is loaded into RCX.
• Finally, the input value (month) is multiplied by 8 and added to the address. Indeed: we are in a 64-
bit environment and all address (or pointers) require exactly 64 bits (or 8 bytes) for storage. Hence,
each table element is 8 bytes wide. And that’s why to pick a specific element, month ∗ 8 bytes has to
be skipped from the start. That’s what MOV does. In addition, this instruction also loads the element
at this address. For 1, an element would be a pointer to a string that contains “February”, etc.
129
Optimizing GCC 4.9 can do the job even better :
Listing 1.238: Optimizing GCC 4.9 x64

movsx rdi, edi
mov rax, QWORD PTR month1[0+rdi*8]
ret
32-bit MSVC
Let’s also compile it in the 32-bit MSVC compiler:

_month$ = 8
_get_month1 PROC
mov eax, DWORD PTR _month$[esp-4]
mov eax, DWORD PTR _month1[eax*4]
ret 0
_get_month1 ENDP
The input value does not need to be extended to 64-bit value, so it is used as is.
And it’s multiplied by 4, because the table elements are 32-bit (or 4 bytes) wide.
32-bit ARM
ARM in ARM mode

get_month1 PROC
LDR r1,|L0.100|
LDR r0,[r1,r0,LSL #2]
BX lr
ENDP
|L0.100|
DCD ||.data||
DCB "January",0
DCB "February",0
DCB "March",0
DCB "April",0
DCB "May",0
DCB "June",0
DCB "July",0
DCB "August",0
DCB "September",0
DCB "October",0
DCB "November",0
DCB "December",0
AREA ||.data||, DATA, ALIGN=2

128 It is somewhat weird, but negative array index could be passed here as month (negative array indices will have been explained
later: 3.22 on page 593). And if this happens, the negative input value of int type is sign-extended correctly and the corresponding
element before table is picked. It is not going to work correctly without sign-extension.
129 “0+” was left in the listing because GCC assembler output is not tidy enough to eliminate it. It’s displacement, and it’s zero
here.
284
month1
DCD ||.conststring||
DCD ||.conststring||+0x8
DCD ||.conststring||+0x1d
DCD ||.conststring||+0x2b
DCD ||.conststring||+0x3c
DCD ||.conststring||+0x4d
The address of the table is loaded in R1.

All the rest is done using just one LDR instruction.
Then input value month is shifted left by 2 (which is the same as multiplying by 4), then added to R1 (where
the address of the table is) and then a table element is loaded from this address.
The 32-bit table element is loaded into R0 from the table.
ARM in Thumb mode
The code is mostly the same, but less dense, because the LSL suffix cannot be specified in the LDR
instruction here:
get_month1 PROC
LSLS r0,r0,#2
LDR r1,|L0.64|
LDR r0,[r1,r0]
BX lr
ENDP
ARM64
Listing 1.241: Optimizing GCC 4.9 ARM64

get_month1:
adrp x1, .LANCHOR0
add x1, x1, :lo12:.LANCHOR0
ldr x0, [x1,w0,sxtw 3]
ret
.LANCHOR0 = . + 0
.type month1, %object
.size month1, 96
month1:
.xword .LC2
.xword .LC3
.xword .LC4
.xword .LC5
.xword .LC6
.xword .LC7
.xword .LC8
.xword .LC9
.xword .LC10
.xword .LC11
.xword .LC12
.xword .LC13
.LC2:
.string "January"
.LC3:
.string "February"
.LC4:
.string "March"
.LC5:
285
.string "April"
.LC6:
.string "May"
.LC7:
.string "June"
.LC8:
.string "July"
.LC9:
.string "August"
.LC10:
.string "September"
.LC11:
.string "October"
.LC12:
.string "November"
.LC13:
.string "December"
The address of the table is loaded in X1 using ADRP/ADD pair.

Then corresponding element is picked using just one LDR, which takes W0 (the register where input argu-
ment month is), shifts it 3 bits to the left (which is the same as multiplying by 8), sign-extends it (this is
what “sxtw” suffix implies) and adds to X0. Then the 64-bit value is loaded from the table into X0.
MIPS

get_month1:
; load address of table into $v0:
la $v0, month1
; take input value and multiply it by 4:
sll $a0, 2
; sum up address of table and multiplied value:
addu $a0, $v0
; load table element at this address into $v0:
lw $v0, 0($a0)
; return
jr $ra
.data # .data.rel.local
.globl month1
month1: .word aJanuary # "January"
.word aFebruary # "February"
.word aMarch # "March"
.word aApril # "April"
.word aMay # "May"
.word aJune # "June"
.word aJuly # "July"
.word aAugust # "August"
.word aSeptember # "September"
.word aOctober # "October"
.word aNovember # "November"
.word aDecember # "December"
.data # .rodata.str1.4
aJanuary: .ascii "January"<0>
aFebruary: .ascii "February"<0>
aMarch: .ascii "March"<0>
aApril: .ascii "April"<0>
aMay: .ascii "May"<0>
aJune: .ascii "June"<0>
aJuly: .ascii "July"<0>
aAugust: .ascii "August"<0>
aSeptember: .ascii "September"<0>
aOctober: .ascii "October"<0>
aNovember: .ascii "November"<0>
aDecember: .ascii "December"<0>
286
Array overflow
Our function accepts values in the range of 0..11, but what if 12 is passed? There is no element in table
at this place.
So the function will load some value which happens to be there, and return it.
Soon after, some other function can try to get a text string from this address and may crash.
Let’s compile the example in MSVC for win64 and open it in IDA to see what the linker has placed after
the table:
Listing 1.243: Executable file in IDA

off_140011000 dq offset aJanuary_1 ; DATA XREF: .text:0000000140001003
; "January"
dq offset aFebruary_1 ; "February"
dq offset aMarch_1 ; "March"
dq offset aApril_1 ; "April"
dq offset aMay_1 ; "May"
dq offset aJune_1 ; "June"
dq offset aJuly_1 ; "July"
dq offset aAugust_1 ; "August"
dq offset aSeptember_1 ; "September"
dq offset aOctober_1 ; "October"
dq offset aNovember_1 ; "November"
dq offset aDecember_1 ; "December"
aJanuary_1 db 'January',0 ; DATA XREF: sub_140001020+4
; .data:off_140011000
aFebruary_1 db 'February',0 ; DATA XREF: .data:0000000140011008
align 4
aMarch_1 db 'March',0 ; DATA XREF: .data:0000000140011010
align 4
aApril_1 db 'April',0 ; DATA XREF: .data:0000000140011018
Month names are came right after.

Our program is tiny, so there isn’t much data to pack in the data segment, so it just the month names.
But it has to be noted that there might be really anything that linker has decided to put by chance.
So what if 12 is passed to the function? The 13th element will be returned.
Let’s see how the CPU treats the bytes there as a 64-bit value:
Listing 1.244: Executable file in IDA

off_140011000 dq offset qword_140011060
; DATA XREF: .text:0000000140001003
dq offset aFebruary_1 ; "February"
dq offset aMarch_1 ; "March"
dq offset aApril_1 ; "April"
dq offset aMay_1 ; "May"
dq offset aJune_1 ; "June"
dq offset aJuly_1 ; "July"
dq offset aAugust_1 ; "August"
dq offset aSeptember_1 ; "September"
dq offset aOctober_1 ; "October"
dq offset aNovember_1 ; "November"
dq offset aDecember_1 ; "December"
qword_140011060 dq 797261756E614Ah ; DATA XREF: sub_140001020+4
; .data:off_140011000
aFebruary_1 db 'February',0 ; DATA XREF: .data:0000000140011008
align 4
aMarch_1 db 'March',0 ; DATA XREF: .data:0000000140011010
And this is 0x797261756E614A.

Soon after, some other function (presumably, one that processes strings) may try to read bytes at this
address, expecting a C-string there.
Most likely it is about to crash, because this value doesn’t look like a valid address.
287
Array overflow protection
If something can go wrong, it will
Murphy’s Law
It’s a bit naïve to expect that every programmer who use your function or library will never pass an
argument larger than 11.
There exists the philosophy that says “fail early and fail loudly” or “fail-fast”, which teaches to report
problems as early as possible and halt.
One such method in C/C++ is assertions.
We can modify our program to fail if an incorrect value is passed:
Listing 1.245: assert() added

const char* get_month1_checked (int month)
{
assert (month<12);
return month1[month];
};
The assertion macro checks for valid values at every function start and fails if the expression is false.

$SG3143 DB 'm', 00H, 'o', 00H, 'n', 00H, 't', 00H, 'h', 00H, '.', 00H
DB 'c', 00H, 00H, 00H
$SG3144 DB 'm', 00H, 'o', 00H, 'n', 00H, 't', 00H, 'h', 00H, '<', 00H
DB '1', 00H, '2', 00H, 00H, 00H
month$ = 48
get_month1_checked PROC
$LN5:
push rbx
sub rsp, 32
movsxd rbx, ecx
cmp ebx, 12
jl SHORT $LN3@get_month1
lea rdx, OFFSET FLAT:$SG3143
mov r8d, 29
call _wassert
$LN3@get_month1:
mov rax, QWORD PTR [rcx+rbx*8]
add rsp, 32
pop rbx
ret 0
get_month1_checked ENDP
In fact, assert() is not a function, but macro. It checks for a condition, then passes also the line number
and file name to another function which reports this information to the user.
Here we see that both file name and condition are encoded in UTF-16. The line number is also passed
(it’s 29).
This mechanism is probably the same in all compilers. Here is what GCC does:

.LC1:
.string "month.c"
.LC2:
.string "month<12"
get_month1_checked:
cmp edi, 11
jg .L6
movsx rdi, edi
288
mov rax, QWORD PTR month1[0+rdi*8]
ret
.L6:
push rax
mov ecx, OFFSET FLAT:__PRETTY_FUNCTION__.2423
mov edx, 29
mov esi, OFFSET FLAT:.LC1
call __assert_fail
__PRETTY_FUNCTION__.2423:
.string "get_month1_checked"
So the macro in GCC also passes the function name for convenience.
Nothing is really free, and this is true for the sanitizing checks as well.
They make your program slower, especially if the assert() macros used in small time-critical functions.
So MSVC, for example, leaves the checks in debug builds, but in release builds they all disappear.
130
Microsoft Windows NT kernels come in “checked” and “free” builds .
The first has validation checks (hence, “checked”), the second one doesn’t (hence, “free” of checks).
Of course, “checked” kernel works slower because of all these checks, so it is usually used only in debug
sessions.
Accessing specific character
An array of pointers to strings can be accessed like this:

#include <stdio.h>
const char* month[]=

{
"January", "February", "March", "April",
"May", "June", "July", "August",
"September", "October", "November", "December"
};
int main()
{
// 4th month, 5th character:
printf ("%c\n", month[3][4]);
};
…since month[3] expression has a const char* type. And then, 5th character is taken from that expression
by adding 4 bytes to its address.
By the way, arguments list passed to main() function has the same data type:
#include <stdio.h>
int main(int argc, char *argv[])

{
printf ("3rd argument, 2nd character: %c\n", argv[3][1]);
};
It’s very important to understand, that, despite similar syntax, this is different from two-dimensional
arrays, which we will consider later.
Another important thing to notice: strings to be addressed must be encoded in a system, where each
character occupies single byte, like ASCII131 and extended ASCII. UTF-8 wouldn’t work here.
130 msdn.microsoft.com/en-us/library/windows/hardware/ff543450(v=vs.85).aspx
131 American Standard Code for Information Interchange
289
1.26.6 Multidimensional arrays
Internally, a multidimensional array is essentially the same thing as a linear array.
Since the computer memory is linear, it is an one-dimensional array. For convenience, this multi-dimensional
array can be easily represented as one-dimensional.
For example, this is how the elements of the 3x4 array are placed in one-dimensional array of 12 cells:
Offset in memory array element

0 [0][0]
1 [0][1]
2 [0][2]
3 [0][3]
4 [1][0]
5 [1][1]
6 [1][2]
7 [1][3]
8 [2][0]
9 [2][1]
10 [2][2]
11 [2][3]
Table 1.3: Two-dimensional array represented in memory as one-dimensional
Here is how each cell of 3*4 array are placed in memory:
0 1 2 3
4 5 6 7
8 9 10 11
Table 1.4: Memory addresses of each cell of two-dimensional array
So, in order to calculate the address of the element we need, we first multiply the first index by 4 (array
width) and then add the second index. That’s called row-major order, and this method of array and matrix
representation is used in at least C/C++ and Python. The term row-major order in plain English language
means: “first, write the elements of the first row, then the second row …and finally the elements of the
last row”.
Another method for representation is called column-major order (the array indices are used in reverse
order) and it is used at least in Fortran, MATLAB and R. column-major order term in plain English language
means: “first, write the elements of the first column, then the second column …and finally the elements
of the last column”.
Which method is better?
In general, in terms of performance and cache memory, the best scheme for data organization is the one,
in which the elements are accessed sequentially.
So if your function accesses data per row, row-major order is better, and vice versa.
Two-dimensional array example
We are going to work with an array of type char, which implies that each element requires only one byte
in memory.
Row filling example
Let’s fill the second row with these values 0..3:
Listing 1.248: Row filling example

#include <stdio.h>
char a[3][4];
290
int main()
{
int x, y;
// clear array
for (x=0; x<3; x++)
for (y=0; y<4; y++)
a[x][y]=0;
// fill second row by 0..3:

for (y=0; y<4; y++)
a[1][y]=y;
};
All three rows are marked with red. We see that second row now has values 0, 1, 2 and 3:
Figure 1.93: OllyDbg: array is filled
Column filling example
Let’s fill the third column with values: 0..2:
Listing 1.249: Column filling example

#include <stdio.h>
char a[3][4];
int main()
{
int x, y;
// clear array
for (x=0; x<3; x++)
for (y=0; y<4; y++)
a[x][y]=0;
// fill third column by 0..2:

for (x=0; x<3; x++)
a[x][2]=x;
};
The three rows are also marked in red here.

We see that in each row, at third position these values are written: 0, 1 and 2.
Figure 1.94: OllyDbg: array is filled
Access two-dimensional array as one-dimensional
We can be easily assured that it’s possible to access a two-dimensional array as one-dimensional array in
at least two ways:
#include <stdio.h>
char a[3][4];
291
char get_by_coordinates1 (char array[3][4], int a, int b)
{
return array[a][b];
};
char get_by_coordinates2 (char *array, int a, int b)

{
// treat input array as one-dimensional
// 4 is array width here
return array[a*4+b];
};
char get_by_coordinates3 (char *array, int a, int b)

{
// treat input array as pointer,
// calculate address, get value at it
// 4 is array width here
return *(array+a*4+b);
};
int main()
{
a[2][3]=123;
printf ("%d\n", get_by_coordinates1(a, 2, 3));
};
Compile132 and run it: it shows correct values.

What MSVC 2013 did is fascinating, all three routines are just the same!

array$ = 8
a$ = 16
b$ = 24
get_by_coordinates3 PROC
; RCX=address of array
; RDX=a
; R8=b
movsxd rax, r8d
; EAX=b
movsxd r9, edx
; R9=a
add rax, rcx
; RAX=b+address of array
movzx eax, BYTE PTR [rax+r9*4]
; AL=load byte at address RAX+R9*4=b+address of array+a*4=address of array+a*4+b
ret 0
get_by_coordinates3 ENDP
array$ = 8
a$ = 16
b$ = 24
movsxd rax, r8d
movsxd r9, edx
add rax, rcx
ret 0
array$ = 8
a$ = 16
b$ = 24
movsxd rax, r8d
132 This program is to be compiled as a C program, not C++, save it to a file with .c extension to compile it using MSVC
292
movsxd r9, edx
add rax, rcx
ret 0
GCC also generates equivalent routines, but slightly different:

; RDI=address of array
; RSI=a
; RDX=b
get_by_coordinates1:
; sign-extend input 32-bit int values "a" and "b" to 64-bit ones
movsx rsi, esi
movsx rdx, edx
lea rax, [rdi+rsi*4]
; RAX=RDI+RSI*4=address of array+a*4
movzx eax, BYTE PTR [rax+rdx]
; AL=load byte at address RAX+RDX=address of array+a*4+b
ret
lea eax, [rdx+rsi*4]
; RAX=RDX+RSI*4=b+a*4
cdqe
movzx eax, BYTE PTR [rdi+rax]
; AL=load byte at address RDI+RAX=address of array+b+a*4
ret
sal esi, 2
; ESI=a<<2=a*4
; sign-extend input 32-bit int values "a*4" and "b" to 64-bit ones
movsx rdx, edx
movsx rsi, esi
add rdi, rsi
; RDI=RDI+RSI=address of array+a*4
movzx eax, BYTE PTR [rdi+rdx]
; AL=load byte at address RDI+RDX=address of array+a*4+b
ret
Three-dimensional array example
It’s the same for multidimensional arrays.

Now we are going to work with an array of type int: each element requires 4 bytes in memory.
Let’s see:
Listing 1.252: simple example

#include <stdio.h>
int a[10][20][30];
void insert(int x, int y, int z, int value)

{
a[x][y][z]=value;
};
x86
We get (MSVC 2010):
293
_DATA SEGMENT
COMM _a:DWORD:01770H
_DATA ENDS
PUBLIC _insert
_TEXT SEGMENT
_x$ = 8 ; size = 4
_y$ = 12 ; size = 4
_z$ = 16 ; size = 4
_value$ = 20 ; size = 4
_insert PROC
push ebp
mov ebp, esp
mov eax, DWORD PTR _x$[ebp]
imul eax, 2400 ; eax=600*4*x
mov ecx, DWORD PTR _y$[ebp]
imul ecx, 120 ; ecx=30*4*y
lea edx, DWORD PTR _a[eax+ecx] ; edx=a + 600*4*x + 30*4*y
mov eax, DWORD PTR _z$[ebp]
mov ecx, DWORD PTR _value$[ebp]
mov DWORD PTR [edx+eax*4], ecx ; *(edx+z*4)=value
pop ebp
ret 0
_insert ENDP
_TEXT ENDS
Nothing special. For index calculation, three input arguments are used in the formula address = 600 ⋅ 4 ⋅ x +
30 ⋅ 4 ⋅ y + 4z, to represent the array as multidimensional. Do not forget that the int type is 32-bit (4 bytes),
so all coefficients must be multiplied by 4.

public insert
insert proc near
x = dword ptr 8
y = dword ptr 0Ch
z = dword ptr 10h
value = dword ptr 14h
push ebp
mov ebp, esp
push ebx
mov ebx, [ebp+x]
mov eax, [ebp+y]
mov ecx, [ebp+z]
lea edx, [eax+eax] ; edx=y*2
mov eax, edx ; eax=y*2
shl eax, 4 ; eax=(y*2)<<4 = y*2*16 = y*32
sub eax, edx ; eax=y*32 - y*2=y*30
imul edx, ebx, 600 ; edx=x*600
add eax, edx ; eax=eax+edx=y*30 + x*600
lea edx, [eax+ecx] ; edx=y*30 + x*600 + z
mov eax, [ebp+value]
mov dword ptr ds:a[edx*4], eax ; *(a+edx*4)=value
pop ebx
pop ebp
retn
insert endp
The GCC compiler does it differently.

For one of the operations in the calculation (30y), GCC produces code without multiplication instructions.
This is how it done: (y + y) ≪ 4 − (y + y) = (2y) ≪ 4 − 2y = 2 ⋅ 16 ⋅ y − 2y = 32y − 2y = 30y. Thus, for the 30y
calculation, only one addition operation, one bitwise shift operation and one subtraction operation are
used. This works faster.
ARM + Non-optimizing Xcode 4.6.3 (LLVM) (Thumb mode)
294
Listing 1.255: Non-optimizing Xcode 4.6.3 (LLVM) (Thumb mode)
_insert
value = -0x10
z = -0xC
y = -8
x = -4
; allocate place in local stack for 4 values of int type

SUB SP, SP, #0x10
MOV R9, 0xFC2 ; a
ADD R9, PC
LDR.W R9, [R9] ; get pointer to array
STR R0, [SP,#0x10+x]
STR R1, [SP,#0x10+y]
STR R2, [SP,#0x10+z]
STR R3, [SP,#0x10+value]
LDR R0, [SP,#0x10+value]
LDR R1, [SP,#0x10+z]
LDR R2, [SP,#0x10+y]
LDR R3, [SP,#0x10+x]
MOV R12, 2400
MUL.W R3, R3, R12
ADD R3, R9
MOV R9, 120
MUL.W R2, R2, R9
ADD R2, R3
LSLS R1, R1, #2 ; R1=R1<<2
ADD R1, R2
STR R0, [R1] ; R1 - address of array element
; deallocate chunk in local stack, allocated for 4 values of int type
ADD SP, SP, #0x10
BX LR
Non-optimizing LLVM saves all variables in local stack, which is redundant.

The address of the array element is calculated by the formula we already saw.
ARM + Optimizing Xcode 4.6.3 (LLVM) (Thumb mode)
Listing 1.256: Optimizing Xcode 4.6.3 (LLVM) (Thumb mode)

_insert
MOVW R9, #0x10FC
MOV.W R12, #2400
MOVT.W R9, #0
RSB.W R1, R1, R1,LSL#4 ; R1 - y. R1=y<<4 - y = y*16 - y = y*15
ADD R9, PC
LDR.W R9, [R9] ; R9 = pointer to an array
MLA.W R0, R0, R12, R9 ; R0 - x, R12 - 2400, R9 - pointer to a. R0=x*2400 + ptr to a
ADD.W R0, R0, R1,LSL#3 ; R0 = R0+R1<<3 = R0+R1*8 = x*2400 + ptr to a + y*15*8 =
; ptr to a + y*30*4 + x*600*4
STR.W R3, [R0,R2,LSL#2] ; R2 - z, R3 - value. address=R0+z*4 =
; ptr to a + y*30*4 + x*600*4 + z*4
BX LR
The tricks for replacing multiplication by shift, addition and subtraction which we already saw are also
present here.
Here we also see a new instruction for us: RSB (Reverse Subtract).
It works just as SUB, but it swaps its operands with each other before execution. Why? SUB and RSB are
instructions, to the second operand of which shift coefficient may be applied: (LSL#4).
But this coefficient can be applied only to second operand.
That’s fine for commutative operations like addition or multiplication (operands may be swapped there
without changing the result).
But subtraction is a non-commutative operation, so RSB exist for these cases.
295
MIPS
My example is tiny, so the GCC compiler decided to put the a array into the 64KiB area addressable by
the Global Pointer.
insert:
; $a0=x
; $a1=y
; $a2=z
; $a3=value
sll $v0, $a0, 5
; $v0 = $a0<<5 = x*32
sll $a0, 3
; $a0 = $a0<<3 = x*8
addu $a0, $v0
; $a0 = $a0+$v0 = x*8+x*32 = x*40
sll $v1, $a1, 5
; $v1 = $a1<<5 = y*32
sll $v0, $a0, 4
; $v0 = $a0<<4 = x*40*16 = x*640
sll $a1, 1
; $a1 = $a1<<1 = y*2
subu $a1, $v1, $a1
; $a1 = $v1-$a1 = y*32-y*2 = y*30
subu $a0, $v0, $a0
; $a0 = $v0-$a0 = x*640-x*40 = x*600
la $gp, __gnu_local_gp
addu $a0, $a1, $a0
; $a0 = $a1+$a0 = y*30+x*600
addu $a0, $a2
; $a0 = $a0+$a2 = y*30+x*600+z
; load address of table:
lw $v0, (a & 0xFFFF)($gp)
; multiply index by 4 to seek array element:
sll $a0, 2
; sum up multiplied index and table address:
addu $a0, $v0, $a0
; store value into table and return:
jr $ra
sw $a3, 0($a0)
.comm a:0x1770
Getting dimensions of multidimensional array
Any string processing function, if an array of characters passed to it, can’t deduce a size of the input array.
Likewise, if a function processes 2D array, only one dimension can be deduced.
For example:
int get_element(int array[10][20], int x, int y)
{
return array[x][y];
};
int main()
{
int array[10][20];
get_element(array, 4, 5);
};
... if compiled (by any compiler) and then decompiled by Hex-Rays:

int get_element(int *array, int x, int y)
{
return array[20 * x + y];
}
296
There is no way to find a size of the first dimension. If x value passed is too big, buffer overflow would
occur, an element from some random place of memory would be read.
And 3D array:
int get_element(int array[10][20][30], int x, int y, int z)
{
return array[x][y][z];
};
int main()
{
int array[10][20][30];
get_element(array, 4, 5, 6);
};
Hex-Rays:
int get_element(int *array, int x, int y, int z)
{
return array[600 * x + z + 30 * y];
}
Again, sizes of only two of 3 dimensions can be deduced.
More examples
The computer screen is represented as a 2D array, but the video-buffer is a linear 1D array. We talk about
it here: 8.15.2 on page 891.
Another example in this book is Minesweeper game: it’s field is also two-dimensional array: 8.4 on
page 794.
1.26.7 Pack of strings as a two-dimensional array

Let’s revisit the function that returns the name of a month: listing.1.236.
As you see, at least one memory load operation is needed to prepare a pointer to the string that’s the
month’s name.
Is it possible to get rid of this memory load operation?
In fact yes, if you represent the list of strings as a two-dimensional array:
#include <stdio.h>
#include <assert.h>
const char month2[12][10]=

{
{ 'J','a','n','u','a','r','y', 0, 0, 0 },
{ 'F','e','b','r','u','a','r','y', 0, 0 },
{ 'M','a','r','c','h', 0, 0, 0, 0, 0 },
{ 'A','p','r','i','l', 0, 0, 0, 0, 0 },
{ 'M','a','y', 0, 0, 0, 0, 0, 0, 0 },
{ 'J','u','n','e', 0, 0, 0, 0, 0, 0 },
{ 'J','u','l','y', 0, 0, 0, 0, 0, 0 },
{ 'A','u','g','u','s','t', 0, 0, 0, 0 },
{ 'S','e','p','t','e','m','b','e','r', 0 },
{ 'O','c','t','o','b','e','r', 0, 0, 0 },
{ 'N','o','v','e','m','b','e','r', 0, 0 },
{ 'D','e','c','e','m','b','e','r', 0, 0 }
};
// in 0..11 range
const char* get_month2 (int month)
{
return &month2[month][0];
};
297
Here is what we’ve get:

month2 DB 04aH
DB 061H
DB 06eH
DB 075H
DB 061H
DB 072H
DB 079H
DB 00H
DB 00H
DB 00H
...
get_month2 PROC
; sign-extend input argument and promote to 64-bit value
movsxd rax, ecx
lea rcx, QWORD PTR [rax+rax*4]
; RCX=month+month*4=month*5
lea rax, OFFSET FLAT:month2
; RAX=pointer to table
lea rax, QWORD PTR [rax+rcx*2]
; RAX=pointer to table + RCX*2=pointer to table + month*5*2=pointer to table + month*10
ret 0
get_month2 ENDP
There are no memory accesses at all.

All this function does is to calculate a point at which the first character of the name of the month is:
pointer_to_the_table + month ∗ 10.
There are also two LEA instructions, which effectively work as several MUL and MOV instructions.
The width of the array is 10 bytes.
Indeed, the longest string here—“September”—is 9 bytes, and plus the terminating zero is 10 bytes.
The rest of the month names are padded by zero bytes, so they all occupy the same space (10 bytes).
Thus, our function works even faster, because all string start at an address which can be calculated easily.
Optimizing GCC 4.9 can do it even shorter:

movsx rdi, edi
lea rax, [rdi+rdi*4]
lea rax, month2[rax+rax]
ret
LEA is also used here for multiplication by 10.

Non-optimizing compilers do multiplication differently.
Listing 1.260: Non-optimizing GCC 4.9 x64

get_month2:
push rbp
mov rbp, rsp
movsx rdx, eax
; RDX = sign-extended input value
mov rax, rdx
; RAX = month
sal rax, 2
; RAX = month<<2 = month*4
add rax, rdx
; RAX = RAX+RDX = month*4+month = month*5
add rax, rax
; RAX = RAX*2 = month*5*2 = month*10
add rax, OFFSET FLAT:month2
298
; RAX = month*10 + pointer to the table
pop rbp
ret
Non-optimizing MSVC just uses IMUL instruction:
Listing 1.261: Non-optimizing MSVC 2013 x64

month$ = 8
get_month2 PROC
mov DWORD PTR [rsp+8], ecx
movsxd rax, DWORD PTR month$[rsp]
; RAX = sign-extended input value into 64-bit one
imul rax, rax, 10
; RAX = RAX*10
; RCX = pointer to the table
add rcx, rax
; RCX = RCX+RAX = pointer to the table+month*10
mov rax, rcx
; RAX = pointer to the table+month*10
mov ecx, 1
; RCX = 1
imul rcx, rcx, 0
; RCX = 1*0 = 0
add rax, rcx
; RAX = pointer to the table+month*10 + 0 = pointer to the table+month*10
ret 0
get_month2 ENDP
But one thing is weird here: why add multiplication by zero and adding zero to the final result?
This looks like a compiler code generator quirk, which wasn’t caught by the compiler’s tests (the result-
ing code works correctly, after all). We intentionally consider such pieces of code so the reader would
understand, that sometimes one shouldn’t puzzle over such compiler artifacts.
32-bit ARM
Optimizing Keil for Thumb mode uses the multiplication instruction MULS:

; R0 = month
MOVS r1,#0xa
; R1 = 10
MULS r0,r1,r0
; R0 = R1*R0 = 10*month
LDR r1,|L0.68|
; R1 = pointer to the table
ADDS r0,r0,r1
; R0 = R0+R1 = 10*month + pointer to the table
BX lr
Optimizing Keil for ARM mode uses add and shift operations:

; R0 = month
LDR r1,|L0.104|
; R1 = pointer to the table
ADD r0,r0,r0,LSL #2
; R0 = R0+R0<<2 = R0+R0*4 = month*5
ADD r0,r1,r0,LSL #1
; R0 = R1+R0<<2 = pointer to the table + month*5*2 = pointer to the table + month*10
BX lr
ARM64
299
; W0 = month
sxtw x0, w0
; X0 = sign-extended input value
adrp x1, .LANCHOR1
add x1, x1, :lo12:.LANCHOR1
; X1 = pointer to the table
; X0 = X0+X0<<2 = X0+X0*4 = X0*5
; X0 = X1+X0<<1 = X1+X0*2 = pointer to the table + X0*10
ret
SXTW is used for sign-extension and promoting input 32-bit value into a 64-bit one and storing it in X0.
ADRP/ADD pair is used for loading the address of the table.
The ADD instructions also has a LSL suffix, which helps with multiplications.
MIPS

.globl get_month2
get_month2:
; $a0=month
sll $v0, $a0, 3
; $v0 = $a0<<3 = month*8
sll $a0, 1
; $a0 = $a0<<1 = month*2
addu $a0, $v0
; $a0 = month*2+month*8 = month*10
; load address of the table:
la $v0, month2
; sum up table address and index we calculated and return:
jr $ra
addu $v0, $a0
month2: .ascii "January"<0>

.byte 0, 0
aFebruary: .ascii "February"<0>
.byte 0
aMarch: .ascii "March"<0>
.byte 0, 0, 0, 0
aApril: .ascii "April"<0>
.byte 0, 0, 0, 0
aMay: .ascii "May"<0>
.byte 0, 0, 0, 0, 0, 0
aJune: .ascii "June"<0>
.byte 0, 0, 0, 0, 0
aJuly: .ascii "July"<0>
.byte 0, 0, 0, 0, 0
aAugust: .ascii "August"<0>
.byte 0, 0, 0
aSeptember: .ascii "September"<0>
aOctober: .ascii "October"<0>
.byte 0, 0
aNovember: .ascii "November"<0>
.byte 0
aDecember: .ascii "December"<0>
.byte 0, 0, 0, 0, 0, 0, 0, 0, 0
Conclusion
This is a bit old-school technique to store text strings. You may find a lot of it in Oracle RDBMS, for example.
It’s hard to say if it’s worth doing on modern computers. Nevertheless, it is a good example of arrays, so
it was added to this book.
300
1.26.8 Conclusion
An array is a pack of values in memory located adjacently.
It’s true for any element type, including structures.
Access to a specific array element is just a calculation of its address.
So, a pointer to an array and address of a first element—is the same thing. This is why ptr[0] and *ptr
expressions are equivalent in C/C++. It’s interesting to note that Hex-Rays often replaces the first by the
second. It does so when it have no idea that it works with pointer to the whole array, and thinks that this
is a pointer to single variable.
1.26.9 Exercises
1.27 Example: a bug in Angband

An ancient rogue-like game from 1990’s 133 had a nice bug in the spirit of the “Roadside Picnic”134 by
Strugatsky brothers or “The Lost Room” TV series135 :
The frog-knows version was abundant of bugs. The funniest of them led to a cunning
technique of cheating the game, that was called ”mushroom farming”. If there were more
than a certain number (about five hundred) of objects in the labyrinth, the game would break,
and many old things turned into objects thrown to the floor. Accordingly, the player went
into the maze, he made such longitudinal grooves there (with a special spell), and walked
along the grooves, creating mushrooms with another special spell. When there were a lot of
mushrooms, the player put and took, put and took some useful item, and mushrooms one
by one turned into this subject. After that, the player returned with hundreds of copies of
the useful item.
( Misha “tiphareth” Verbitsky, http://imperium.lenin.ru/CEBEP/arc/3/lightmusic/light.htm )

And some information from usenet:
From: be...@uswest.com (George Bell)
Subject: [Angband] Multiple artifact copies found (bug?)
Date: Fri, 23 Jul 1993 15:55:08 GMT
Up to 2000 ft I found only 4 artifacts, now my house is littered with the

suckers (FYI, most I've gotten from killing nasties, like Dracoliches and the
like). Something really weird is happening now, as I found multiple
copies of the same artifact! My half-elf ranger is down at 2400 ft on one
level which is particularly nasty. There is a graveyard plus monsters
surrounded by permanent rock and 2 or 3 other special monster rooms! I did
so much slashing with my favorite weapon, Crisdurian, that I filled several
rooms nearly to the brim with treasure (as usual, mostly junk).
Then, when I found a way into the big vault, I noticed some of the treasure
had already been identified (in fact it looked strangely familiar!). Then I
found *two* Short Swords named Sting (1d6) (+7,+8), and I just ran across a
third copy! I have seen multiple copies of Gurthang on this level as well.
Is there some limit on the number of items per level which I have exceeded?
This sounds reasonable as all multiple copies I have seen come from this level.
I'm playing PC angband. Anybody else had this problem?

133 https://en.wikipedia.org/wiki/Angband_(video_game), http://rephial.org/
134 https://en.wikipedia.org/wiki/Roadside_Picnic
135 https://en.wikipedia.org/wiki/The_Lost_Room
301
-George Bell
Help! I need a Rod of Restore Life Levels, if there is such a thing. These
Graveyards are nasty (Black Reavers and some speed 2 wraith in particular).
( https://groups.google.com/forum/#!original/rec.games.moria/jItmfrdGyL8/8csctQqA7PQJ )
From: Ceri <cm...@andrew.cmu.edu>
Subject: Re: [Angband] Multiple artifact copies found (bug?)
Date: Fri, 23 Jul 1993 23:32:20 -0400
welcome to the mush bug. if there are more than 256 items
on the floor, things start duplicating. learn to harness
this power and you will win shortly :>
--Rick
( https://groups.google.com/forum/#!search/angband$202.4$20bug$20multiplying$20items/rec.
games.moria/jItmfrdGyL8/FoQeiccewHAJ )
From: nwe...@soda.berkeley.edu (Nicholas C. Weaver)
Subject: Re: [Angband] Multiple artifact copies found (bug?)
Date: 24 Jul 1993 18:18:05 GMT
In article <74348474...@unix1.andrew.cmu.edu> Ceri <cm...@andrew.cmu.edu> writes:

>welcome to the mush bug. if there are more than 256 items
>on the floor, things start duplicating. learn to harness
>this power and you will win shortly :>
>
>--Rick
QUestion on this. Is it only the first 256 items which get

duplicated? What about the origional items? Etc ETc ETc...
Oh, for those who like to know about bugs, though, the -n option
(start new character) has the following behavior:
(this is in version 2.4.Frog.knows on unix)
If you hit controll-p, you keep your old stats.
YOu loose all record of artifacts founds and named monsters killed.
YOu loose all items you are carrying (they get turned into error in
objid()s ).
You loose your gold.
You KEEP all the stuff in your house.
If you kill something, and then quaff a potion of restore life

levels, you are back up to where you were before in EXPERIENCE POINTS!!
Gaining spells will not work right after this, unless you have a
gain int item (for spellcasters) or gain wis item (for priests/palidans), in
which case after performing the above, then take the item back on and off,
you will be able to learn spells normally again.
This can be exploited, if you are a REAL H0ZER (like me), into
getting multiple artifacts early on. Just get to a level where you can
pound wormtongue into the ground, kill him, go up, drop your stuff in your
house, buy a few potions of restore exp and high value spellbooks with your
leftover gold, angband -n yourself back to what you were before, and repeat
the process. Yes, you CAN kill wormtongue multiple times. :)
This also allows the creation of a human rogue with dunedain warrior
starting stats.
302
Of course, such practices are evil, vile, and disgusting. I take no
liability for the results of spreading this information. Yeah, it's another
bug to go onto the pile.
--
Nicholas C. Weaver perpetual ensign guppy nwe...@soda.berkeley.edu
It is a tale, told by an idiot, full of sound and fury, .signifying nothing.
Since C evolved out of B, and a C+ is close to a B,
does that mean that C++ is a devolution of the language?
( https://groups.google.com/forum/#!original/rec.games.moria/jItmfrdGyL8/FoQeiccewHAJ )
The whole thread: https://groups.google.com/forum/#!search/angband$202.4$20bug$20multiplying$20i
rec.games.moria/jItmfrdGyL8/FoQeiccewHAJ.
136
The author of these lines has found the version with the bug (2.4 fk) , and we can clearly see how the
global arrays are declared:
/* Number of dungeon objects */
#define MAX_DUNGEON_OBJ 423
...
int16 sorted_objects[MAX_DUNGEON_OBJ];
/* Identified objects flags */

int8u object_ident[OBJECT_IDENT_SIZE];
int16 t_level[MAX_OBJ_LEVEL+1];
inven_type t_list[MAX_TALLOC];
inven_type inventory[INVEN_ARRAY_SIZE];
Perhaps this is a reason. The MAX_DUNGEON_OBJ constant is too small. Perhaps, authors should use linked
lists or other data structures, which are unlimited by size. But arrays are simpler to use.
Another example of buffer overflow over globally defined arrays: 3.31 on page 637.
1.28 Manipulating specific bit(s)

A lot of functions define their input arguments as flags in bit fields.
Of course, they could be substituted by a set of bool-typed variables, but it is not frugally.
1.28.1 Specific bit checking

x86
Win32 API example:

HANDLE fh;
fh=CreateFile ("file", GENERIC_WRITE | GENERIC_READ, FILE_SHARE_READ, NULL, OPEN_ALWAYS⤦

Ç , FILE_ATTRIBUTE_NORMAL, NULL);
We get (MSVC 2010):

push 0
push 128 ; 00000080H
push 4
push 0
push 1
push -1073741824 ; c0000000H
push OFFSET $SG78813
call DWORD PTR __imp__CreateFileA@28
mov DWORD PTR _fh$[ebp], eax
136 http://rephial.org/release/2.4.fk, https://yurichev.com/mirrors/angband-2.4.fk.tar
303
Let’s take a look in WinNT.h:
Listing 1.267: WinNT.h
#define GENERIC_READ (0x80000000L)
#define GENERIC_WRITE (0x40000000L)
#define GENERIC_EXECUTE (0x20000000L)
#define GENERIC_ALL (0x10000000L)
Everything is clear, GENERIC_READ | GENERIC_WRITE = 0x80000000 | 0x40000000 = 0xC0000000, and

that value is used as the second argument for the CreateFile()137 function.
How would CreateFile() check these flags?
If we look in KERNEL32.DLL in Windows XP SP3 x86, we’ll find this fragment of code in CreateFileW:
Listing 1.268: KERNEL32.DLL (Windows XP SP3 x86)
.text:7C83D429 test byte ptr [ebp+dwDesiredAccess+3], 40h
.text:7C83D42D mov [ebp+var_8], 1
.text:7C83D434 jz short loc_7C83D417
.text:7C83D436 jmp loc_7C810817
Here we see the TEST instruction, however it doesn’t take the whole second argument,
but only the most significant byte (ebp+dwDesiredAccess+3) and checks it for flag 0x40 (which implies
the GENERIC_WRITE flag here).
TEST is basically the same instruction as AND, but without saving the result (recall the fact CMP is merely
the same as SUB, but without saving the result ( 1.12.4 on page 86)).
The logic of this code fragment is as follows:
if ((dwDesiredAccess&0x40000000) == 0) goto loc_7C83D417
If AND instruction leaves this bit, the ZF flag is to be cleared and the JZ conditional jump is not to be
triggered. The conditional jump is triggered only if the 0x40000000 bit is absent in dwDesiredAccess
variable —then the result of AND is 0, ZF is to be set and the conditional jump is to be triggered.
Let’s try GCC 4.4.1 and Linux:
#include <stdio.h>
#include <fcntl.h>
void main()
{
int handle;
handle=open ("file", O_RDWR | O_CREAT);

};
We get:
public main
main proc near

push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
sub esp, 20h
mov [esp+20h+var_1C], 42h
mov [esp+20h+var_20], offset aFile ; "file"
call _open
leave
retn
main endp
137 msdn.microsoft.com/en-us/library/aa363858(VS.85).aspx
304
If we take a look in the open() function in the libc.so.6 library, it is only a syscall:
Listing 1.270: open() (libc.so.6)

.text:000BE69B mov edx, [esp+4+mode] ; mode
.text:000BE69F mov ecx, [esp+4+flags] ; flags
.text:000BE6A3 mov ebx, [esp+4+filename] ; filename
.text:000BE6A7 mov eax, 5
.text:000BE6AC int 80h ; LINUX - sys_open
So, the bit fields for open() are apparently checked somewhere in the Linux kernel.
Of course, it is easy to download both Glibc and the Linux kernel source code, but we are interested in
understanding the matter without it.
So, as of Linux 2.6, when the sys_open syscall is called, control eventually passes to do_sys_open, and
from there—to the do_filp_open() function (it’s located in the kernel source tree in fs/namei.c).
N.B. Aside from passing arguments via the stack, there is also a method of passing some of them via
registers. This is also called fastcall ( 6.1.3 on page 726). This works faster since CPU does not need to
access the stack in memory to read argument values. GCC has the option regparm138 , through which it’s
possible to set the number of arguments that can be passed via registers.
139 140
The Linux 2.6 kernel is compiled with -mregparm=3 option .
What this means to us is that the first 3 arguments are to be passed via registers EAX, EDX and ECX, and
the rest via the stack. Of course, if the number of arguments is less than 3, only part of registers set is to
be used.
So, let’s download Linux Kernel 2.6.31, compile it in Ubuntu: make vmlinux, open it in IDA, and find the
do_filp_open() function. At the beginning, we see (the comments are mine):
Listing 1.271: do_filp_open() (linux kernel 2.6.31)

do_filp_open proc near
...
push ebp
mov ebp, esp
push edi
push esi
push ebx
mov ebx, ecx
add ebx, 1
sub esp, 98h
mov esi, [ebp+arg_4] ; acc_mode (5th argument)
test bl, 3
mov [ebp+var_80], eax ; dfd (1th argument)
mov [ebp+var_7C], edx ; pathname (2th argument)
mov [ebp+var_78], ecx ; open_flag (3th argument)
jnz short loc_C01EF684
mov ebx, ecx ; ebx <- open_flag
GCC saves the values of the first 3 arguments in the local stack. If that wasn’t done, the compiler would
not touch these registers, and that would be too tight environment for the compiler’s register allocator.
Let’s find this fragment of code:
Listing 1.272: do_filp_open() (linux kernel 2.6.31)

loc_C01EF684: ; CODE XREF: do_filp_open+4F
test bl, 40h ; O_CREAT
jnz loc_C01EF810
mov edi, ebx
shr edi, 11h
xor edi, 1
and edi, 1
test ebx, 10000h
jz short loc_C01EF6D3
or edi, 2
138 ohse.de/uwe/articles/gcc-attributes.html#func-regparm
139 kernelnewbies.org/Linux_2_6_20#head-042c62f290834eb1fe0a1942bbf5bb9a4accbc8f
140 See also arch/x86/include/asm/calling.h file in kernel tree
305
0x40—is what the O_CREAT macro equals to. open_flag gets checked for the presence of the 0x40 bit,
and if this bit is 1, the next JNZ instruction is triggered.
ARM
The O_CREAT bit is checked differently in Linux kernel 3.8.0.

Listing 1.273: linux kernel 3.8.0
struct file *do_filp_open(int dfd, struct filename *pathname,
const struct open_flags *op)
{
...
filp = path_openat(dfd, pathname, &nd, op, flags | LOOKUP_RCU);
...
}
static struct file *path_openat(int dfd, struct filename *pathname,

struct nameidata *nd, const struct open_flags *op, int flags)
{
...
error = do_last(nd, &path, file, op, &opened, pathname);
...
}
static int do_last(struct nameidata *nd, struct path *path,

struct file *file, const struct open_flags *op,
int *opened, struct filename *name)
{
...
if (!(open_flag & O_CREAT)) {
...
error = lookup_fast(nd, path, &inode);
...
} else {
...
error = complete_walk(nd);
}
...
}
Here is how the kernel compiled for ARM mode looks in IDA:
Listing 1.274: do_last() from vmlinux (IDA)
...
.text:C0169EA8 MOV R9, R3 ; R3 - (4th argument) open_flag
...
.text:C0169ED4 LDR R6, [R9] ; R6 - open_flag
...
.text:C0169F68 TST R6, #0x40 ; jumptable C0169F00 default case
.text:C0169F6C BNE loc_C016A128
.text:C0169F70 LDR R2, [R4,#0x10]
.text:C0169F74 ADD R12, R4, #8
.text:C0169F78 LDR R3, [R4,#0xC]
.text:C0169F7C MOV R0, R4
.text:C0169F80 STR R12, [R11,#var_50]
.text:C0169F84 LDRB R3, [R2,R3]
.text:C0169F88 MOV R2, R8
.text:C0169F8C CMP R3, #0
.text:C0169F90 ORRNE R1, R1, #3
.text:C0169F94 STRNE R1, [R4,#0x24]
.text:C0169F98 ANDS R3, R6, #0x200000
.text:C0169F9C MOV R1, R12
.text:C0169FA0 LDRNE R3, [R4,#0x24]
.text:C0169FA4 ANDNE R3, R3, #1
.text:C0169FA8 EORNE R3, R3, #1
.text:C0169FAC STR R3, [R11,#var_54]
.text:C0169FB0 SUB R3, R11, #-var_38
.text:C0169FB4 BL lookup_fast
306
...
.text:C016A128 loc_C016A128 ; CODE XREF: do_last.isra.14+DC
.text:C016A128 MOV R0, R4
.text:C016A12C BL complete_walk
...
TST is analogous to the TEST instruction in x86. We can “spot” visually this code fragment by the fact the
lookup_fast() is to be executed in one case and complete_walk() in the other. This corresponds to the
source code of the do_last() function. The O_CREAT macro equals to 0x40 here too.
1.28.2 Setting and clearing specific bits

For example:
#include <stdio.h>
#define IS_SET(flag, bit) ((flag) & (bit))

#define SET_BIT(var, bit) ((var) |= (bit))
#define REMOVE_BIT(var, bit) ((var) &= ~(bit))
int f(int a)
{
int rt=a;
SET_BIT (rt, 0x4000);

REMOVE_BIT (rt, 0x200);
return rt;
};
int main()
{
f(0x12340678);
};
x86
Non-optimizing MSVC
We get (MSVC 2010):

_rt$ = -4 ; size = 4
_a$ = 8 ; size = 4
_f PROC
push ebp
mov ebp, esp
push ecx
mov DWORD PTR _rt$[ebp], eax
mov ecx, DWORD PTR _rt$[ebp]
or ecx, 16384 ; 00004000H
mov DWORD PTR _rt$[ebp], ecx
mov edx, DWORD PTR _rt$[ebp]
and edx, -513 ; fffffdffH
mov DWORD PTR _rt$[ebp], edx
mov eax, DWORD PTR _rt$[ebp]
mov esp, ebp
pop ebp
ret 0
_f ENDP
The OR instruction sets one bit into a register while ignoring other 1 bits.
AND resets one bit. It can be said that AND just copies all bits except one. Indeed, in the second AND
operand only the bits that need to be saved are set, just the one do not want to copy is not (which is 0 in
the bitmask). It is the easier way to memorize the logic.
307
OllyDbg
Let’s try this example in OllyDbg.

First, let’s see the binary form of the constants we are going to use:
0x200 (0b00000000000000000001000000000) (i.e., the 10th bit (counting from 1st)).
Inverted 0x200 is 0xFFFFFDFF (0b11111111111111111110111111111).
0x4000 (0b00000000000000100000000000000) (i.e., the 15th bit).
The input value is: 0x12340678 (0b10010001101000000011001111000). We see how it’s loaded:
Figure 1.95: OllyDbg: value is loaded into ECX
308
OR got executed:
Figure 1.96: OllyDbg: OR executed
15th bit is set: 0x12344678 (0b10010001101000100011001111000).
309
The value is reloaded again (because the compiler is not in optimizing mode):
Figure 1.97: OllyDbg: value has been reloaded into EDX
310
AND got executed:
Figure 1.98: OllyDbg: AND executed
The 10th bit has been cleared (or, in other words, all bits were left except the 10th) and the final value
now is
0x12344478 (0b10010001101000100010001111000).
Optimizing MSVC
If we compile it in MSVC with optimization turned on (/Ox), the code is even shorter:

_a$ = 8 ; size = 4
_f PROC
and eax, -513 ; fffffdffH
or eax, 16384 ; 00004000H
ret 0
_f ENDP
Non-optimizing GCC
Let’s try GCC 4.4.1 without optimization:
Listing 1.277: Non-optimizing GCC

public f
f proc near

arg_0 = dword ptr 8
push ebp
mov ebp, esp
sub esp, 10h
mov [ebp+var_4], eax
or [ebp+var_4], 4000h
and [ebp+var_4], 0FFFFFDFFh
mov eax, [ebp+var_4]
leave
311
retn
f endp
There is a redundant code present, however, it is shorter than the MSVC version without optimization.
Now let’s try GCC with optimization turned on -O3:
Optimizing GCC
Listing 1.278: Optimizing GCC

public f
f proc near
arg_0 = dword ptr 8
push ebp
mov ebp, esp
pop ebp
or ah, 40h
and ah, 0FDh
retn
f endp
That’s shorter. It is worth noting the compiler works with the EAX register part via the AH register—that is
the EAX register part from the 8th to the 15th bits included.
Byte number:
RAXx64
EAX
AX
AH AL
N.B. The 16-bit CPU 8086 accumulator was named AX and consisted of two 8-bit halves—AL (lower byte)
and AH (higher byte). In 80386 almost all registers were extended to 32-bit, the accumulator was named
EAX, but for the sake of compatibility, its older parts may be still accessed as AX/AH/AL.
Since all x86 CPUs are successors of the 16-bit 8086 CPU, these older 16-bit opcodes are shorter than the
newer 32-bit ones. That’s why the or ah, 40h instruction occupies only 3 bytes. It would be more logical
way to emit here or eax, 04000h but that is 5 bytes, or even 6 (in case the register in the first operand
is not EAX).
Optimizing GCC and regparm
It would be even shorter if to turn on the -O3 optimization flag and also set regparm=3.
Listing 1.279: Optimizing GCC

public f
f proc near
push ebp
or ah, 40h
mov ebp, esp
and ah, 0FDh
pop ebp
retn
f endp
Indeed, the first argument is already loaded in EAX, so it is possible to work with it in-place. It is worth
noting that both the function prologue (push ebp / mov ebp,esp) and epilogue (pop ebp) can easily be
omitted here, but GCC probably is not good enough to do such code size optimizations. However, such
short functions are better to be inlined functions ( 3.14 on page 506).
312
ARM + Optimizing Keil 6/2013 (ARM mode)

02 0C C0 E3 BIC R0, R0, #0x200
01 09 80 E3 ORR R0, R0, #0x4000
1E FF 2F E1 BX LR
BIC (BItwise bit Clear) is an instruction for clearing specific bits. This is just like the AND instruction, but
with inverted operand. I.e., it’s analogous to a NOT +AND instruction pair.
ORR is “logical or”, analogous to OR in x86.
So far it’s easy.
ARM + Optimizing Keil 6/2013 (Thumb mode)

01 21 89 03 MOVS R1, 0x4000
08 43 ORRS R0, R1
49 11 ASRS R1, R1, #5 ; generate 0x200 and place to R1
88 43 BICS R0, R1
70 47 BX LR
Seems like Keil decided that the code in Thumb mode, making 0x200 from 0x4000, is more compact than
the code for writing 0x200 to an arbitrary register.
So that is why, with the help of ASRS (arithmetic shift right), this value is calculated as 0x4000 ≫ 5.
ARM + Optimizing Xcode 4.6.3 (LLVM) (ARM mode)

42 0C C0 E3 BIC R0, R0, #0x4200
01 09 80 E3 ORR R0, R0, #0x4000
1E FF 2F E1 BX LR
The code that was generated by LLVM, in source code form could be something like this:
SET_BIT (rt, 0x4000);
141
And it does exactly what we need. But why 0x4200? Perhaps that an artifact from LLVM’s optimizer .
Probably a compiler’s optimizer error, but the generated code works correctly anyway.
You can read more about compiler anomalies here ( 11.4 on page 981).
Optimizing Xcode 4.6.3 (LLVM) for Thumb mode generates the same code.
ARM: more about the BIC instruction
Let’s rework the example slightly:

int f(int a)
{
int rt=a;
return rt;
};
Then the optimizing Keil 5.03 in ARM mode does:

141 It was LLVM build 2410.2.00 bundled with Apple Xcode 4.6.3
313
f PROC
BIC r0,r0,#0x1000
BIC r0,r0,#0x234
BX lr
ENDP
There are two BIC instructions, i.e., bits 0x1234 are cleared in two passes.
This is because it’s not possible to encode 0x1234 in a BIC instruction, but it’s possible to encode 0x1000
and 0x234.
Optimizing GCCcompiling for ARM64 can use the AND instruction instead of BIC:

f:
and w0, w0, -513 ; 0xFFFFFFFFFFFFFDFF
orr w0, w0, 16384 ; 0x4000
ret
Non-optimizing GCC generates more redundant code, but works just like optimized:

f:
sub sp, sp, #32
str w0, [sp,12]
ldr w0, [sp,12]
str w0, [sp,28]
ldr w0, [sp,28]
orr w0, w0, 16384 ; 0x4000
str w0, [sp,28]
ldr w0, [sp,28]
and w0, w0, -513 ; 0xFFFFFFFFFFFFFDFF
str w0, [sp,28]
ldr w0, [sp,28]
add sp, sp, 32
ret
MIPS

f:
; $a0=a
ori $a0, 0x4000
; $a0=a|0x4000
li $v0, 0xFFFFFDFF
jr $ra
and $v0, $a0, $v0
; at finish: $v0 = $a0 & $v0 = a|0x4000 & 0xFFFFFDFF
ORI is, of course, the OR operation. “I” in the instruction name means that the value is embedded in the
machine code.
But after that we have AND. There is no way to use ANDI because it’s not possible to embed the 0xFFFFFDFF
number in a single instruction, so the compiler has to load 0xFFFFFDFF into register $V0 first and then
generates AND which takes all its values from registers.
314
1.28.3 Shifts
Bit shifts in C/C++ are implemented using ≪ and ≫ operators. The x86 ISA has the SHL (SHift Left) and
SHR (SHift Right) instructions for this. Shift instructions are often used in division and multiplications by
powers of two: 2n (e.g., 1, 2, 4, 8, etc.): 1.24.1 on page 211, 1.24.2 on page 216.
Shifting operations are also so important because they are often used for specific bit isolation or for
constructing a value of several scattered bits.
1.28.4 Setting and clearing specific bits: FPU example

Here is how bits are located in the float type in IEEE 754 form:
31 30 23 22 0
S exponent mantissa or fraction
( S — sign )
The sign of number is in the MSB142 . Will it be possible to change the sign of a floating point number
without any FPU instructions?
#include <stdio.h>
float my_abs (float i)

{
unsigned int tmp=(*(unsigned int*)&i) & 0x7FFFFFFF;
return *(float*)&tmp;
};
float set_sign (float i)

{
unsigned int tmp=(*(unsigned int*)&i) | 0x80000000;
};
float negate (float i)

{
unsigned int tmp=(*(unsigned int*)&i) ^ 0x80000000;
};
int main()
{
printf ("my_abs():\n");
printf ("%f\n", my_abs (123.456));
printf ("%f\n", my_abs (-456.123));
printf ("set_sign():\n");
printf ("%f\n", set_sign (123.456));
printf ("%f\n", set_sign (-456.123));
printf ("negate():\n");
printf ("%f\n", negate (123.456));
printf ("%f\n", negate (-456.123));
};
We need this trickery in C/C++ to copy to/from float value without actual conversion. So there are three
functions: my_abs() resets MSB; set_sign() sets MSB and negate() flips it.
XOR can be used to flip a bit: 2.6 on page 456.
x86
The code is pretty straightforward:

_tmp$ = 8
_i$ = 8
_my_abs PROC
142 Most Significant Bit
315
and DWORD PTR _i$[esp-4], 2147483647 ; 7fffffffH
fld DWORD PTR _tmp$[esp-4]
ret 0
_my_abs ENDP
_tmp$ = 8
_i$ = 8
_set_sign PROC
or DWORD PTR _i$[esp-4], -2147483648 ; 80000000H
ret 0
_set_sign ENDP
_tmp$ = 8
_i$ = 8
_negate PROC
xor DWORD PTR _i$[esp-4], -2147483648 ; 80000000H
ret 0
_negate ENDP
An input value of type float is taken from the stack, but treated as an integer value.
AND and OR reset and set the desired bit. XOR flips it.
Finally, the modified value is loaded into ST0, because floating-point numbers are returned in this register.
Now let’s try optimizing MSVC 2012 for x64:

tmp$ = 8
i$ = 8
my_abs PROC
movss DWORD PTR [rsp+8], xmm0
mov eax, DWORD PTR i$[rsp]
btr eax, 31
mov DWORD PTR tmp$[rsp], eax
movss xmm0, DWORD PTR tmp$[rsp]
ret 0
my_abs ENDP
_TEXT ENDS
tmp$ = 8
i$ = 8
set_sign PROC
bts eax, 31
ret 0
set_sign ENDP
tmp$ = 8
i$ = 8
negate PROC
btc eax, 31
ret 0
negate ENDP
The input value is passed in XMM0, then it is copied into the local stack and then we see some instructions
that are new to us: BTR, BTS, BTC.
These instructions are used for resetting (BTR), setting (BTS) and inverting (or complementing: BTC) spe-
cific bits. The 31st bit is MSB, counting from 0.
316
Finally, the result is copied into XMM0, because floating point values are returned through XMM0 in Win64
environment.
MIPS
GCC 4.4.5 for MIPS does mostly the same:

my_abs:
; move from coprocessor 1:
mfc1 $v1, $f12
li $v0, 0x7FFFFFFF
; $v0=0x7FFFFFFF
; do AND:
and $v0, $v1
; move to coprocessor 1:
mtc1 $v0, $f0
; return
jr $ra
or $at, $zero ; branch delay slot
set_sign:
mfc1 $v0, $f12
lui $v1, 0x8000
; $v1=0x80000000
; do OR:
or $v0, $v1, $v0
mtc1 $v0, $f0
; return
jr $ra
negate:
mfc1 $v0, $f12
lui $v1, 0x8000
; $v1=0x80000000
; do XOR:
xor $v0, $v1, $v0
mtc1 $v0, $f0
; return
jr $ra
One single LUI instruction is used to load 0x80000000 into a register, because LUI is clearing the low 16
bits and these are zeros in the constant, so one LUI without subsequent ORI is enough.
ARM

my_abs PROC
; clear bit:
BIC r0,r0,#0x80000000
BX lr
ENDP
set_sign PROC
; do OR:
ORR r0,r0,#0x80000000
BX lr
ENDP
317
negate PROC
; do XOR:
EOR r0,r0,#0x80000000
BX lr
ENDP
So far so good.
ARM has the BIC instruction, which explicitly clears specific bit(s). EOR is the ARM instruction name for
XOR (“Exclusive OR”).

my_abs PROC
LSLS r0,r0,#1
; r0=i<<1
LSRS r0,r0,#1
; r0=(i<<1)>>1
BX lr
ENDP
set_sign PROC
MOVS r1,#1
; r1=1
LSLS r1,r1,#31
; r1=1<<31=0x80000000
ORRS r0,r0,r1
; r0=r0 | 0x80000000
BX lr
ENDP
negate PROC
MOVS r1,#1
; r1=1
LSLS r1,r1,#31
; r1=1<<31=0x80000000
EORS r0,r0,r1
; r0=r0 ^ 0x80000000
BX lr
ENDP
Thumb mode in ARM offers 16-bit instructions and not much data can be encoded in them, so here a
MOVS/LSLS instruction pair is used for forming the 0x80000000 constant. It works like this: 1 << 31 =
0x80000000.
The code of my_abs is weird and it effectively works like this expression: (i << 1) >> 1. This statement looks
meaningless. But nevertheless, when input << 1 is executed, the MSB (sign bit) is just dropped. When
the subsequent result >> 1 statement is executed, all bits are now in their own places, but MSB is zero,
because all “new” bits appearing from the shift operations are always zeros. That is how the LSLS/LSRS
instruction pair clears MSB.
Optimizing GCC 4.6.3 (Raspberry Pi, ARM mode)
Listing 1.291: Optimizing GCC 4.6.3 for Raspberry Pi (ARM mode)

my_abs
; copy from S0 to R2:
FMRS R2, S0
; clear bit:
BIC R3, R2, #0x80000000
; copy from R3 to S0:
FMSR S0, R3
BX LR
set_sign
318
FMRS R2, S0
; do OR:
ORR R3, R2, #0x80000000
FMSR S0, R3
BX LR
negate
FMRS R2, S0
; do ADD:
ADD R3, R2, #0x80000000
FMSR S0, R3
BX LR
Let’s run Raspberry Pi Linux in QEMU and it emulates an ARM FPU, so S-registers are used here for floating
point numbers instead of R-registers.
The FMRS instruction copies data from GPR to the FPU and back.
my_abs() and set_sign() looks as expected, but negate()? Why is there ADD instead of XOR?
It’s hard to believe, but the instruction ADD register, 0x80000000 works just like
XOR register, 0x80000000. First of all, what’s our goal? The goal is to flip the MSB, so let’s forget
about the XOR operation. From school-level mathematics we may recall that adding values like 1000 to
other values never affects the last 3 digits. For example: 1234567 + 10000 = 1244567 (last 4 digits are never
affected).
But here we operate in binary base and
0x80000000 is 0b100000000000000000000000000000000, i.e., only the highest bit is set.
Adding 0x80000000 to any value never affects the lowest 31 bits, but affects only the MSB. Adding 1 to
0 is resulting in 1.
Adding 1 to 1 is resulting in 0b10 in binary form, but the 32th bit (counting from zero) gets dropped,
because our registers are 32 bit wide, so the result is 0. That’s why XOR can be replaced by ADD here.
It’s hard to say why GCC decided to do this, but it works correctly.
1.28.5 Counting bits set to 1

Here is a simple example of a function that calculates the number of bits set in the input value.
This operation is also called “population count”143 .
#include <stdio.h>
int f(unsigned int a)

{
int i;
int rt=0;
for (i=0; i<32; i++)

if (IS_SET (a, 1<<i))
rt++;
return rt;
};
int main()
{
f(0x12345678); // test
};
143 modern x86 CPUs (supporting SSE4) even have a POPCNT instruction for it
319
In this loop, the iteration count value i is counting from 0 to 31, so the 1 ≪ i statement is counting from 1
to 0x80000000. Describing this operation in natural language, we would say shift 1 by n bits left. In other
words, 1 ≪ i statement consequently produces all possible bit positions in a 32-bit number. The freed bit
at right is always cleared.
Here is a table of all possible 1 ≪ i for i = 0 . . . 31:
C/C++ expression Power of two Decimal form Hexadecimal form
1≪0 20 1 1
1≪1 21 2 2
1≪2 22 4 4
1≪3 23 8 8
1≪4 24 16 0x10
1≪5 25 32 0x20
1≪6 26 64 0x40
1≪7 27 128 0x80
1≪8 28 256 0x100
1≪9 29 512 0x200
1 ≪ 10 210 1024 0x400
1 ≪ 11 211 2048 0x800
1 ≪ 12 212 4096 0x1000
1 ≪ 13 213 8192 0x2000
1 ≪ 14 214 16384 0x4000
1 ≪ 15 215 32768 0x8000
1 ≪ 16 216 65536 0x10000
1 ≪ 17 217 131072 0x20000
1 ≪ 18 218 262144 0x40000
1 ≪ 19 219 524288 0x80000
1 ≪ 20 220 1048576 0x100000
1 ≪ 21 221 2097152 0x200000
1 ≪ 22 222 4194304 0x400000
1 ≪ 23 223 8388608 0x800000
1 ≪ 24 224 16777216 0x1000000
1 ≪ 25 225 33554432 0x2000000
1 ≪ 26 226 67108864 0x4000000
1 ≪ 27 227 134217728 0x8000000
1 ≪ 28 228 268435456 0x10000000
1 ≪ 29 229 536870912 0x20000000
1 ≪ 30 230 1073741824 0x40000000
1 ≪ 31 231 2147483648 0x80000000
These constant numbers (bit masks) very often appear in code and a practicing reverse engineer must
be able to spot them quickly.
Decimal numbers below 65536 and hexadecimal ones are very easy to memorize. While decimal numbers
above 65536 are, probably, not worth memorizing.
These constants are very often used for mapping flags to specific bits. For example, here is excerpt from
ssl_private.h from Apache 2.4.6 source code:
/**
* Define the SSL options
*/
#define SSL_OPT_NONE (0)
#define SSL_OPT_RELSET (1<<0)
#define SSL_OPT_STDENVVARS (1<<1)
#define SSL_OPT_EXPORTCERTDATA (1<<3)
#define SSL_OPT_FAKEBASICAUTH (1<<4)
#define SSL_OPT_STRICTREQUIRE (1<<5)
#define SSL_OPT_OPTRENEGOTIATE (1<<6)
#define SSL_OPT_LEGACYDNFORMAT (1<<7)
Let’s get back to our example.

The IS_SET macro checks bit presence in a.
The IS_SET macro is in fact the logical AND operation (AND) and it returns 0 if the specific bit is absent
there, or the bit mask, if the bit is present. The if() operator in C/C++ triggers if the expression in it is not
320
zero, it might be even 123456, that is why it always works correctly.
x86
MSVC

_rt$ = -8 ; size = 4
_i$ = -4 ; size = 4
_a$ = 8 ; size = 4
_f PROC
push ebp
mov ebp, esp
sub esp, 8
mov DWORD PTR _rt$[ebp], 0
jmp SHORT $LN4@f
$LN3@f:
mov eax, DWORD PTR _i$[ebp] ; increment of i
add eax, 1
$LN4@f:
jge SHORT $LN2@f ; loop finished?
mov edx, 1
shl edx, cl ; EDX=EDX<<CL
and edx, DWORD PTR _a$[ebp]
je SHORT $LN1@f ; result of AND instruction was 0?
; then skip next instructions
mov eax, DWORD PTR _rt$[ebp] ; no, not zero
add eax, 1 ; increment rt
$LN1@f:
jmp SHORT $LN3@f
$LN2@f:
mov esp, ebp
pop ebp
ret 0
_f ENDP
321
OllyDbg
Let’s load this example into OllyDbg. Let the input value be 0x12345678.
For i = 1, we see how i is loaded into ECX:
Figure 1.99: OllyDbg: i = 1, i is loaded into ECX
EDX is 1. SHL is to be executed now.
322
SHL has been executed:
Figure 1.100: OllyDbg: i = 1, EDX =1 ≪ 1 = 2
EDX contain 1 ≪ 1 (or 2). This is a bit mask.
323
AND sets ZF to 1, which implies that the input value (0x12345678) ANDed with 2 results in 0:
Figure 1.101: OllyDbg: i = 1, is there that bit in the input value? No. (ZF =1)
So, there is no corresponding bit in the input value.

The piece of code, which increments the counter is not to be executed: the JZ instruction bypassing it.
324
Let’s trace a bit further and i is now 4. SHL is to be executed now:
Figure 1.102: OllyDbg: i = 4, i is loaded into ECX
325
EDX =1 ≪ 4 (or 0x10 or 16):
Figure 1.103: OllyDbg: i = 4, EDX =1 ≪ 4 = 0x10
This is another bit mask.
326
AND is executed:
Figure 1.104: OllyDbg: i = 4, is there that bit in the input value? Yes. (ZF =0)
ZF is 0 because this bit is present in the input value.

Indeed, 0x12345678 & 0x10 = 0x10.
This bit counts: the jump is not triggering and the bit counter incrementing.
The function returns 13. This is total number of bits set in 0x12345678.
GCC
Let’s compile it in GCC 4.4.1:

public f
f proc near
rt = dword ptr -0Ch

i = dword ptr -8
arg_0 = dword ptr 8
push ebp
mov ebp, esp
push ebx
sub esp, 10h
mov [ebp+rt], 0
mov [ebp+i], 0
jmp short loc_80483EF
loc_80483D0:
mov eax, [ebp+i]
mov edx, 1
mov ebx, edx
mov ecx, eax
shl ebx, cl
mov eax, ebx
and eax, [ebp+arg_0]
test eax, eax
jz short loc_80483EB
add [ebp+rt], 1
loc_80483EB:
add [ebp+i], 1
loc_80483EF:
327
cmp [ebp+i], 1Fh
jle short loc_80483D0
mov eax, [ebp+rt]
add esp, 10h
pop ebx
pop ebp
retn
f endp
x64
Let’s modify the example slightly to extend it to 64-bit:

#include <stdio.h>
#include <stdint.h>
int f(uint64_t a)
{
uint64_t i;
int rt=0;
for (i=0; i<64; i++)

if (IS_SET (a, 1ULL<<i))
rt++;
return rt;
};
So far so easy.
Listing 1.294: Non-optimizing GCC 4.8.2
f:
push rbp
mov rbp, rsp
mov QWORD PTR [rbp-24], rdi ; a
mov DWORD PTR [rbp-12], 0 ; rt=0
mov QWORD PTR [rbp-8], 0 ; i=0
jmp .L2
.L4:
mov rax, QWORD PTR [rbp-8]
mov rdx, QWORD PTR [rbp-24]
; RAX = i, RDX = a
mov ecx, eax
; ECX = i
shr rdx, cl
; RDX = RDX>>CL = a>>i
mov rax, rdx
; RAX = RDX = a>>i
and eax, 1
; EAX = EAX&1 = (a>>i)&1
test rax, rax
; the last bit is zero?
; skip the next ADD instruction, if it was so.
je .L3
add DWORD PTR [rbp-12], 1 ; rt++
.L3:
add QWORD PTR [rbp-8], 1 ; i++
.L2:
cmp QWORD PTR [rbp-8], 63 ; i<63?
jbe .L4 ; jump to the loop body begin, if so
mov eax, DWORD PTR [rbp-12] ; return rt
pop rbp
ret
328

1 f:
2 xor eax, eax ; rt variable will be in EAX register
3 xor ecx, ecx ; i variable will be in ECX register
4 .L3:
5 mov rsi, rdi ; load input value
6 lea edx, [rax+1] ; EDX=EAX+1
7 ; EDX here is a new version of rt,
8 ; which will be written into rt variable, if the last bit is 1
9 shr rsi, cl ; RSI=RSI>>CL
10 and esi, 1 ; ESI=ESI&1
11 ; the last bit is 1? If so, write new version of rt into EAX
12 cmovne eax, edx
13 add rcx, 1 ; RCX++
14 cmp rcx, 64
15 jne .L3
16 rep ret ; AKA fatret
This code is terser, but has a quirk.

In all examples that we see so far, we were incrementing the “rt” value after comparing a specific bit, but
the code here increments “rt” before (line 6), writing the new value into register EDX. Thus, if the last bit is
1, the CMOVNE144 instruction (which is a synonym for CMOVNZ145 ) commits the new value of “rt” by moving
EDX (“proposed rt value”) into EAX (“current rt” to be returned at the end).
Hence, the incrementing is performed at each step of loop, i.e., 64 times, without any relation to the input
value.
The advantage of this code is that it contain only one conditional jump (at the end of the loop) instead of
two jumps (skipping the “rt” value increment and at the end of loop). And that might work faster on the
modern CPUs with branch predictors: 2.10.1 on page 462.
The last instruction is REP RET (opcode F3 C3) which is also called FATRET by MSVC. This is somewhat
optimized version of RET, which is recommended by AMD to be placed at the end of function, if RET goes
right after conditional jump: [Software Optimization Guide for AMD Family 16h Processors, (2013)p.15]
146
.

a$ = 8
f PROC
; RCX = input value
xor eax, eax
mov edx, 1
lea r8d, QWORD PTR [rax+64]
; R8D=64
npad 5
$LL4@f:
test rdx, rcx
; there are no such bit in input value?
; skip the next INC instruction then.
je SHORT $LN3@f
inc eax ; rt++
$LN3@f:
rol rdx, 1 ; RDX=RDX<<1
dec r8 ; R8--
jne SHORT $LL4@f
fatret 0
f ENDP
144 ConditionalMOVe if Not Equal

145 ConditionalMOVe if Not Zero
146 More information on it: http://go.yurichev.com/17328
329
Here the ROL instruction is used instead of SHL, which is in fact “rotate left” instead of “shift left”, but in
this example it works just as SHL.
You can read more about the rotate instruction here: .1.6 on page 1020.
R8 here is counting from 64 to 0. It’s just like an inverted i.
Here is a table of some registers during the execution:
RDX R8
0x0000000000000001 64
0x0000000000000002 63
0x0000000000000004 62
0x0000000000000008 61
... ...
0x4000000000000000 2
0x8000000000000000 1
At the end we see the FATRET instruction, which was explained here: 1.28.5 on the previous page.

a$ = 8
f PROC
; RCX = input value
xor eax, eax
mov edx, 1
lea r8d, QWORD PTR [rax+32]
; EDX = 1, R8D = 32
npad 5
$LL4@f:
; pass 1 ------------------------------
test rdx, rcx
je SHORT $LN3@f
inc eax ; rt++
$LN3@f:
; -------------------------------------
; pass 2 ------------------------------
test rdx, rcx
je SHORT $LN11@f
inc eax ; rt++
$LN11@f:
; -------------------------------------
dec r8 ; R8--
jne SHORT $LL4@f
fatret 0
f ENDP
Optimizing MSVC 2012 does almost the same job as optimizing MSVC 2010, but somehow, it generates
two identical loop bodies and the loop count is now 32 instead of 64.
To be honest, it’s not possible to say why. Some optimization trick? Maybe it’s better for the loop body to
be slightly longer?
Anyway, such code is relevant here to show that sometimes the compiler output may be really weird and
illogical, but perfectly working.
ARM + Optimizing Xcode 4.6.3 (LLVM) (ARM mode)

MOV R1, R0
MOV R0, #0
MOV R2, #1
MOV R3, R0
loc_2E54
330
TST R1, R2,LSL R3 ; set flags according to R1 & (R2<<R3)
ADD R3, R3, #1 ; R3++
ADDNE R0, R0, #1 ; if ZF flag is cleared by TST, then R0++
CMP R3, #32
BNE loc_2E54
BX LR
TST is the same thing as TEST in x86.

As was noted before ( 3.12.3 on page 499), there are no separate shifting instructions in ARM mode.
However, there are modifiers LSL (Logical Shift Left), LSR (Logical Shift Right), ASR (Arithmetic Shift Right),
ROR (Rotate Right) and RRX (Rotate Right with Extend), which may be added to such instructions as MOV,
TST, CMP, ADD, SUB, RSB147 .
These modificators define how to shift the second operand and by how many bits.
Thus the “TST R1, R2,LSL R3” instruction works here as R1 ∧ (R2 ≪ R3).
ARM + Optimizing Xcode 4.6.3 (LLVM) (Thumb-2 mode)
Almost the same, but here are two LSL.W/TST instructions are used instead of a single TST, because in
Thumb mode it is not possible to define LSL modifier directly in TST.
MOV R1, R0
MOVS R0, #0
MOV.W R9, #1
MOVS R3, #0
loc_2F7A
LSL.W R2, R9, R3
TST R2, R1
ADD.W R3, R3, #1
IT NE
ADDNE R0, #1
CMP R3, #32
BNE loc_2F7A
BX LR
ARM64 + Optimizing GCC 4.9
Let’s take the 64-bit example which has been already used: 1.28.5 on page 328.

f:
mov w2, 0 ; rt=0
mov x5, 1
mov w1, w2
.L2:
lsl x4, x5, x1 ; w4 = w5<<w1 = 1<<i
add w3, w2, 1 ; new_rt=rt+1
tst x4, x0 ; (1<<i) & a
add w1, w1, 1 ; i++
; result of TST was non-zero?
; then w2=w3 or rt=new_rt.
; otherwise: w2=w2 or rt=rt (idle operation)
csel w2, w3, w2, ne
cmp w1, 64 ; i<64?
bne .L2 ; yes
mov w0, w2 ; return rt
ret
The result is very similar to what GCC generates for x64: 1.295 on page 329.
The CSEL instruction is “Conditional SELect”. It just chooses one variable of two depending on the flags
set by TST and copies the value into W2, which holds the “rt” variable.
147 These instructions are also called “data processing instructions”
331
ARM64 + Non-optimizing GCC 4.9
And again, we’ll work on the 64-bit example which was already used: 1.28.5 on page 328. The code is
more verbose, as usual.

f:
sub sp, sp, #32
str x0, [sp,8] ; store "a" value to Register Save Area
str wzr, [sp,24] ; rt=0
str wzr, [sp,28] ; i=0
b .L2
.L4:
ldr w0, [sp,28]
mov x1, 1
lsl x0, x1, x0 ; X0 = X1<<X0 = 1<<i
mov x1, x0
; X1 = 1<<i
ldr x0, [sp,8]
; X0 = a
and x0, x1, x0
; X0 = X1&X0 = (1<<i) & a
; X0 contain zero? then jump to .L3, skipping "rt" increment
cmp x0, xzr
beq .L3
; rt++
ldr w0, [sp,24]
add w0, w0, 1
str w0, [sp,24]
.L3:
; i++
ldr w0, [sp,28]
add w0, w0, 1
str w0, [sp,28]
.L2:
; i<=63? then jump to .L4
ldr w0, [sp,28]
cmp w0, 63
ble .L4
; return rt
ldr w0, [sp,24]
add sp, sp, 32
ret
MIPS
Non-optimizing GCC

f:
; IDA is not aware of variable names, we gave them manually:
rt = -0x10
i = -0xC
var_4 = -4
a = 0
addiu $sp, -0x18

sw $fp, 0x18+var_4($sp)
move $fp, $sp
sw $a0, 0x18+a($fp)
; initialize rt and i variables to zero:
sw $zero, 0x18+rt($fp)
sw $zero, 0x18+i($fp)
; jump to loop check instructions:
b loc_68
332
loc_20:
li $v1, 1
lw $v0, 0x18+i($fp)
sllv $v0, $v1, $v0
; $v0 = 1<<i
move $v1, $v0
lw $v0, 0x18+a($fp)
and $v0, $v1, $v0
; $v0 = a & (1<<i)
; is a & (1<<i) equals to zero? jump to loc_58 then:
beqz $v0, loc_58
or $at, $zero
; no jump occurred, that means a & (1<<i)!=0, so increment "rt" then:
lw $v0, 0x18+rt($fp)
addiu $v0, 1
sw $v0, 0x18+rt($fp)
loc_58:
; increment i:
lw $v0, 0x18+i($fp)
addiu $v0, 1
sw $v0, 0x18+i($fp)
loc_68:
; load i and compare it with 0x20 (32).
; jump to loc_20 if it is less then 0x20 (32):
lw $v0, 0x18+i($fp)
slti $v0, 0x20 # ' '
bnez $v0, loc_20
; function epilogue. return rt:
lw $v0, 0x18+rt($fp)
move $sp, $fp ; load delay slot
lw $fp, 0x18+var_4($sp)
addiu $sp, 0x18 ; load delay slot
jr $ra
That is verbose: all local variables are located in the local stack and reloaded each time they’re needed.
The SLLV instruction is “Shift Word Left Logical Variable”, it differs from SLL only in that the shift amount
is encoded in the SLL instruction (and is fixed, as a consequence), but SLLV takes shift amount from a
register.
Optimizing GCC
That is terser. There are two shift instructions instead of one. Why?
It’s possible to replace the first SLLV instruction with an unconditional branch instruction that jumps right
to the second SLLV. But this is another branching instruction in the function, and it’s always favorable to
get rid of them: 2.10.1 on page 462.

f:
; $a0=a
; rt variable will reside in $v0:
move $v0, $zero
; i variable will reside in $v1:
move $v1, $zero
li $t0, 1
li $a3, 32
sllv $a1, $t0, $v1
; $a1 = $t0<<$v1 = 1<<i
333
loc_14:
and $a1, $a0
; $a1 = a&(1<<i)
; increment i:
addiu $v1, 1
; jump to loc_28 if a&(1<<i)==0 and increment rt:
beqz $a1, loc_28
addiu $a2, $v0, 1
; if BEQZ was not triggered, save updated rt into $v0:
move $v0, $a2
loc_28:
; if i!=32, jump to loc_14 and also prepare next shifted value:
bne $v1, $a3, loc_14
sllv $a1, $t0, $v1
; return
jr $ra
1.28.6 Conclusion
Analogous to the C/C++ shifting operators ≪ and ≫, the shift instructions in x86 are SHR/SHL (for unsigned
values) and SAR/SHL (for signed values).
The shift instructions in ARM are LSR/LSL (for unsigned values) and ASR/LSL (for signed values).
It’s also possible to add shift suffix to some instructions (which are called “data processing instructions”).
Check for specific bit (known at compile stage)
Test if the 0b1000000 bit (0x40) is present in the register’s value:
Listing 1.303: C/C++

if (input&0x40)
...
Listing 1.304: x86

TEST REG, 40h
JNZ is_set
; bit is not set
Listing 1.305: x86

TEST REG, 40h
JZ is_cleared
; bit is set

TST REG, #0x40
BNE is_set
; bit is not set
Sometimes, AND is used instead of TEST, but the flags that are set are the same.
Check for specific bit (specified at runtime)
This is usually done by this C/C++ code snippet (shift value by n bits right, then cut off lowest bit):

if ((value>>n)&1)
....
This is usually implemented in x86 code as:
334
Listing 1.308: x86
; REG=input_value
; CL=n
SHR REG, CL
AND REG, 1
Or (shift 1 bit n times left, isolate this bit in input value and check if it’s not zero):

if (value & (1<<n))
....
Listing 1.310: x86

; CL=n
MOV REG, 1
SHL REG, CL
AND input_value, REG
Set specific bit (known at compile stage)

value=value|0x40;
Listing 1.312: x86

OR REG, 40h
Listing 1.313: ARM (ARM mode) and ARM64

ORR R0, R0, #0x40
Set specific bit (specified at runtime)

value=value|(1<<n);
Listing 1.315: x86

; CL=n
MOV REG, 1
SHL REG, CL
OR input_value, REG
Clear specific bit (known at compile stage)
Just apply AND operation with the inverted value:

value=value&(~0x40);
Listing 1.317: x86

AND REG, 0FFFFFFBFh
Listing 1.318: x64

AND REG, 0FFFFFFFFFFFFFFBFh
335
This is actually leaving all bits set except one.
ARM in ARM mode has BIC instruction, which works like the NOT +AND instruction pair:

BIC R0, R0, #0x40
Clear specific bit (specified at runtime)

value=value&(~(1<<n));
Listing 1.321: x86

; CL=n
MOV REG, 1
SHL REG, CL
NOT REG
AND input_value, REG
1.28.7 Exercises
1.29 Linear congruential generator as pseudorandom number

generator
Perhaps, the linear congruential generator is the simplest possible way to generate random numbers.
It’s not in favour nowadays148 , but it’s so simple (just one multiplication, one addition and AND operation),
that we can use it as an example.
#include <stdint.h>
// constants from the Numerical Recipes book

#define RNG_a 1664525
#define RNG_c 1013904223
static uint32_t rand_state;
void my_srand (uint32_t init)

{
rand_state=init;
}
int my_rand ()
{
rand_state=rand_state*RNG_a;
rand_state=rand_state+RNG_c;
return rand_state & 0x7fff;
}
There are two functions: the first one is used to initialize the internal state, and the second one is called
to generate pseudorandom numbers.
We see that two constants are used in the algorithm. They are taken from [William H. Press and Saul A.
Teukolsky and William T. Vetterling and Brian P. Flannery, Numerical Recipes, (2007)].
Let’s define them using a #define C/C++ statement. It’s a macro.
148 Mersenne twister is better
336
The difference between a C/C++ macro and a constant is that all macros are replaced with their value by
C/C++ preprocessor, and they don’t take any memory, unlike variables.
In contrast, a constant is a read-only variable.
It’s possible to take a pointer (or address) of a constant variable, but impossible to do so with a macro.
The last AND operation is needed because by C-standard my_rand() has to return a value in the 0..32767
range.
If you want to get 32-bit pseudorandom values, just omit the last AND operation.
1.29.1 x86
_BSS SEGMENT
_rand_state DD 01H DUP (?)
_BSS ENDS
_init$ = 8
_srand PROC
mov eax, DWORD PTR _init$[esp-4]
mov DWORD PTR _rand_state, eax
ret 0
_srand ENDP
_TEXT SEGMENT
_rand PROC
imul eax, DWORD PTR _rand_state, 1664525
add eax, 1013904223 ; 3c6ef35fH
and eax, 32767 ; 00007fffH
ret 0
_rand ENDP
_TEXT ENDS
Here we see it: both constants are embedded into the code. There is no memory allocated for them.
The my_srand() function just copies its input value into the internal
rand_state variable.
my_rand() takes it, calculates the next rand_state, cuts it and leaves it in the EAX register.
The non-optimized version is more verbose:

_BSS SEGMENT
_BSS ENDS
_init$ = 8
_srand PROC
push ebp
mov ebp, esp
mov eax, DWORD PTR _init$[ebp]
pop ebp
ret 0
_srand ENDP
_TEXT SEGMENT
_rand PROC
push ebp
mov ebp, esp
imul eax, DWORD PTR _rand_state, 1664525
mov ecx, DWORD PTR _rand_state
add ecx, 1013904223 ; 3c6ef35fH
mov DWORD PTR _rand_state, ecx
337
mov eax, DWORD PTR _rand_state
and eax, 32767 ; 00007fffH
pop ebp
ret 0
_rand ENDP
_TEXT ENDS
1.29.2 x64
The x64 version is mostly the same and uses 32-bit registers instead of 64-bit ones (because we are
working with int values here).
But my_srand() takes its input argument from the ECX register rather than from stack:

_BSS SEGMENT
rand_state DD 01H DUP (?)
_BSS ENDS
init$ = 8
my_srand PROC
; ECX = input argument
mov DWORD PTR rand_state, ecx
ret 0
my_srand ENDP
_TEXT SEGMENT
my_rand PROC
imul eax, DWORD PTR rand_state, 1664525 ; 0019660dH
add eax, 1013904223 ; 3c6ef35fH
mov DWORD PTR rand_state, eax
and eax, 32767 ; 00007fffH
ret 0
my_rand ENDP
_TEXT ENDS
GCC compiler generates mostly the same code.
1.29.3 32-bit ARM

my_srand PROC
LDR r1,|L0.52| ; load pointer to rand_state
STR r0,[r1,#0] ; save rand_state
BX lr
ENDP
my_rand PROC
LDR r0,|L0.52| ; load pointer to rand_state
LDR r2,|L0.56| ; load RNG_a
LDR r1,[r0,#0] ; load rand_state
MUL r1,r2,r1
LDR r2,|L0.60| ; load RNG_c
ADD r1,r1,r2
STR r1,[r0,#0] ; save rand_state
; AND with 0x7FFF:
LSL r0,r1,#17
LSR r0,r0,#17
BX lr
ENDP
|L0.52|
DCD ||.data||
|L0.56|
DCD 0x0019660d
338
|L0.60|
DCD 0x3c6ef35f
AREA ||.data||, DATA, ALIGN=2
rand_state
DCD 0x00000000
It’s not possible to embed 32-bit constants into ARM instructions, so Keil has to place them externally and
load them additionally. One interesting thing is that it’s not possible to embed the 0x7FFF constant as
well. So what Keil does is shifting rand_state left by 17 bits and then shifting it right by 17 bits. This is
analogous to the (rand_state ≪ 17) ≫ 17 statement in C/C++. It seems to be useless operation, but what it
does is clearing the high 17 bits, leaving the low 15 bits intact, and that’s our goal after all.
Optimizing Keil for Thumb mode generates mostly the same code.
1.29.4 MIPS
my_srand:
; store $a0 to rand_state:
lui $v0, (rand_state >> 16)
jr $ra
sw $a0, rand_state
my_rand:
; load rand_state to $v0:
lui $v1, (rand_state >> 16)
lw $v0, rand_state
or $at, $zero ; load delay slot
; multiplicate rand_state in $v0 by 1664525 (RNG_a):
sll $a1, $v0, 2
sll $a0, $v0, 4
addu $a0, $a1, $a0
sll $a1, $a0, 6
subu $a0, $a1, $a0
addu $a0, $v0
sll $a1, $a0, 5
addu $a0, $a1
sll $a0, 3
addu $v0, $a0, $v0
sll $a0, $v0, 2
addu $v0, $a0
; add 1013904223 (RNG_c)
; the LI instruction is coalesced by IDA from LUI and ORI
li $a0, 0x3C6EF35F
addu $v0, $a0
; store to rand_state:
sw $v0, (rand_state & 0xFFFF)($v1)
jr $ra
andi $v0, 0x7FFF ; branch delay slot
Wow, here we see only one constant (0x3C6EF35F or 1013904223). Where is the other one (1664525)?
It seems that multiplication by 1664525 is performed by just using shifts and additions! Let’s check this
assumption:
int f (int a)
{
return a*RNG_a;
}

f:
sll $v1, $a0, 2
sll $v0, $a0, 4
339
addu $v0, $v1, $v0
sll $v1, $v0, 6
subu $v0, $v1, $v0
addu $v0, $a0
sll $v1, $v0, 5
addu $v0, $v1
sll $v0, 3
addu $a0, $v0, $a0
sll $v0, $a0, 2
jr $ra
addu $v0, $a0, $v0 ; branch delay slot
Indeed!
MIPS relocations
We will also focus on how such operations as load from memory and store to memory actually work.
The listings here are produced by IDA, which hides some details.
We’ll run objdump twice: to get a disassembled listing and also relocations list:
# objdump -D rand_O3.o
...
00000000 <my_srand>:
0: 3c020000 lui v0,0x0
4: 03e00008 jr ra
8: ac440000 sw a0,0(v0)
0000000c <my_rand>:
c: 3c030000 lui v1,0x0
10: 8c620000 lw v0,0(v1)
14: 00200825 move at,at
18: 00022880 sll a1,v0,0x2
1c: 00022100 sll a0,v0,0x4
20: 00a42021 addu a0,a1,a0
24: 00042980 sll a1,a0,0x6
28: 00a42023 subu a0,a1,a0
2c: 00822021 addu a0,a0,v0
30: 00042940 sll a1,a0,0x5
34: 00852021 addu a0,a0,a1
38: 000420c0 sll a0,a0,0x3
3c: 00821021 addu v0,a0,v0
40: 00022080 sll a0,v0,0x2
44: 00441021 addu v0,v0,a0
48: 3c043c6e lui a0,0x3c6e
4c: 3484f35f ori a0,a0,0xf35f
50: 00441021 addu v0,v0,a0
54: ac620000 sw v0,0(v1)
58: 03e00008 jr ra
5c: 30427fff andi v0,v0,0x7fff
...
# objdump -r rand_O3.o
...
RELOCATION RECORDS FOR [.text]:

OFFSET TYPE VALUE
00000000 R_MIPS_HI16 .bss
00000008 R_MIPS_LO16 .bss
0000000c R_MIPS_HI16 .bss
...
340
Let’s consider the two relocations for the my_srand() function.
The first one, for address 0 has a type of R_MIPS_HI16 and the second one for address 8 has a type of
R_MIPS_LO16.
That implies that address of the beginning of the .bss segment is to be written into the instructions at
address of 0 (high part of address) and 8 (low part of address).
The rand_state variable is at the very start of the .bss segment.
So we see zeros in the operands of instructions LUI and SW, because nothing is there yet— the compiler
don’t know what to write there.
The linker will fix this, and the high part of the address will be written into the operand of LUI and the low
part of the address—to the operand of SW.
SW will sum up the low part of the address and what is in register $V0 (the high part is there).
It’s the same story with the my_rand() function: R_MIPS_HI16 relocation instructs the linker to write the
high part of the .bss segment address into instruction LUI.
So the high part of the rand_state variable address is residing in register $V1.
The LW instruction at address 0x10 sums up the high and low parts and loads the value of the rand_state
variable into $V0.
The SW instruction at address 0x54 do the summing again and then stores the new value to the rand_state
global variable.
IDA processes relocations while loading, thus hiding these details, but we should keep them in mind.
1.29.5 Thread-safe version of the example

The thread-safe version of the example is to be demonstrated later: 6.2.1 on page 734.
1.30 Structures
A C/C++ structure, with some assumptions, is just a set of variables, always stored in memory together,
not necessary of the same type 149 .
1.30.1 MSVC: SYSTEMTIME example

Let’s take the SYSTEMTIME150 win32 structure that describes time.
This is how it’s defined:
Listing 1.329: WinBase.h

typedef struct _SYSTEMTIME {
WORD wYear;
WORD wMonth;
WORD wDayOfWeek;
WORD wDay;
WORD wHour;
WORD wMinute;
WORD wSecond;
WORD wMilliseconds;
} SYSTEMTIME, *PSYSTEMTIME;
Let’s write a C function to get the current time:

#include <windows.h>
#include <stdio.h>
void main()
{
149 AKA “heterogeneous container”
150 MSDN: SYSTEMTIME structure
341
SYSTEMTIME t;
GetSystemTime (&t);
printf ("%04d-%02d-%02d %02d:%02d:%02d\n",

t.wYear, t.wMonth, t.wDay,
t.wHour, t.wMinute, t.wSecond);
return;
};
We get (MSVC 2010):
Listing 1.330: MSVC 2010 /GS-

_t$ = -16 ; size = 16
_main PROC
push ebp
mov ebp, esp
sub esp, 16
lea eax, DWORD PTR _t$[ebp]
push eax
call DWORD PTR __imp__GetSystemTime@4
movzx ecx, WORD PTR _t$[ebp+12] ; wSecond
push ecx
movzx edx, WORD PTR _t$[ebp+10] ; wMinute
push edx
movzx eax, WORD PTR _t$[ebp+8] ; wHour
push eax
movzx ecx, WORD PTR _t$[ebp+6] ; wDay
push ecx
movzx edx, WORD PTR _t$[ebp+2] ; wMonth
push edx
movzx eax, WORD PTR _t$[ebp] ; wYear
push eax
push OFFSET $SG78811 ; '%04d-%02d-%02d %02d:%02d:%02d', 0aH, 00H
call _printf
add esp, 28
xor eax, eax
mov esp, ebp
pop ebp
ret 0
_main ENDP
16 bytes are allocated for this structure in the local stack —that is exactly sizeof(WORD)*8 (there are 8
WORD variables in the structure).
Pay attention to the fact that the structure begins with the wYear field. It can be said that a pointer to the
SYSTEMTIME structure is passed to the GetSystemTime()151 , but it is also can be said that a pointer to
the wYear field is passed, and that is the same! GetSystemTime() writes the current year to the WORD
pointer pointing to, then shifts 2 bytes ahead, writes current month, etc., etc.
151 MSDN: SYSTEMTIME structure
342
OllyDbg
Let’s compile this example in MSVC 2010 with /GS- /MD keys and run it in OllyDbg.
Let’s open windows for data and stack at the address which is passed as the first argument of the
GetSystemTime() function, and let’s wait until it’s executed. We see this:
Figure 1.105: OllyDbg: GetSystemTime() just executed
The system time of the function execution on my computer is 9 December 2014, 22:29:52:
Listing 1.331: printf() output

2014-12-09 22:29:52
So we see these 16 bytes in the data window:

DE 07 0C 00 02 00 09 00 16 00 1D 00 34 00 D4 03
Each two bytes represent one field of the structure. Since the endianness is little endian, we see the low
byte first and then the high one.
Hence, these are the values currently stored in memory:
Hexadecimal number decimal number field name
0x07DE 2014 wYear
0x000C 12 wMonth
0x0002 2 wDayOfWeek
0x0009 9 wDay
0x0016 22 wHour
0x001D 29 wMinute
0x0034 52 wSecond
0x03D4 980 wMilliseconds
The same values are seen in the stack window, but they are grouped as 32-bit values.
And then printf() just takes the values it needs and outputs them to the console.
Some values aren’t output by printf() (wDayOfWeek and wMilliseconds), but they are in memory right
now, available for use.
Replacing the structure with array
The fact that the structure fields are just variables located side-by-side, can be easily demonstrated by
doing the following. Keeping in mind the SYSTEMTIME structure description, it’s possible to rewrite this
simple example like this:
343
#include <stdio.h>
void main()
{
WORD array[8];
GetSystemTime (array);
printf ("%04d-%02d-%02d %02d:%02d:%02d\n",

array[0] /* wYear */, array[1] /* wMonth */, array[3] /* wDay */,
array[4] /* wHour */, array[5] /* wMinute */, array[6] /* wSecond */);
return;
};
The compiler grumbles a bit:

systemtime2.c(7) : warning C4133: 'function' : incompatible types - from 'WORD [8]' to '⤦
Ç LPSYSTEMTIME'
But nevertheless, it produces this code:

$SG78573 DB '%04d-%02d-%02d %02d:%02d:%02d', 0aH, 00H
_array$ = -16 ; size = 16

_main PROC
push ebp
mov ebp, esp
sub esp, 16
lea eax, DWORD PTR _array$[ebp]
push eax
movzx ecx, WORD PTR _array$[ebp+12] ; wSecond
push ecx
movzx edx, WORD PTR _array$[ebp+10] ; wMinute
push edx
movzx eax, WORD PTR _array$[ebp+8] ; wHoure
push eax
movzx ecx, WORD PTR _array$[ebp+6] ; wDay
push ecx
movzx edx, WORD PTR _array$[ebp+2] ; wMonth
push edx
movzx eax, WORD PTR _array$[ebp] ; wYear
push eax
push OFFSET $SG78573 ; '%04d-%02d-%02d %02d:%02d:%02d', 0aH, 00H
call _printf
add esp, 28
xor eax, eax
mov esp, ebp
pop ebp
ret 0
_main ENDP
And it works just as the same!

It is very interesting that the result in assembly form cannot be distinguished from the result of the previous
compilation.
So by looking at this code, one cannot say for sure if there was a structure declared, or an array.
Nevertheless, no sane person would do it, as it is not convenient.
Also the structure fields may be changed by developers, swapped, etc.
We will not study this example in OllyDbg, because it will be just the same as in the case with the structure.
1.30.2 Let’s allocate space for a structure using malloc()

Sometimes it is simpler to place structures not the in local stack, but in the heap:
344
#include <stdio.h>
void main()
{
SYSTEMTIME *t;
t=(SYSTEMTIME *)malloc (sizeof (SYSTEMTIME));
GetSystemTime (t);
printf ("%04d-%02d-%02d %02d:%02d:%02d\n",

t->wYear, t->wMonth, t->wDay,
t->wHour, t->wMinute, t->wSecond);
free (t);
return;
};
Let’s compile it now with optimization (/Ox) so it would be easy to see what we need.

_main PROC
push esi
push 16
call _malloc
add esp, 4
mov esi, eax
push esi
movzx eax, WORD PTR [esi+12] ; wSecond
movzx ecx, WORD PTR [esi+10] ; wMinute
movzx edx, WORD PTR [esi+8] ; wHour
push eax
movzx eax, WORD PTR [esi+6] ; wDay
push ecx
movzx ecx, WORD PTR [esi+2] ; wMonth
push edx
movzx edx, WORD PTR [esi] ; wYear
push eax
push ecx
push edx
call _printf
push esi
call _free
add esp, 32
xor eax, eax
pop esi
ret 0
_main ENDP
So, sizeof(SYSTEMTIME) = 16 and that is exact number of bytes to be allocated by malloc(). It returns
a pointer to a freshly allocated memory block in the EAX register, which is then moved into the ESI register.
GetSystemTime() win32 function takes care of saving value in ESI, and that is why it is not saved here
and continues to be used after the GetSystemTime() call.
New instruction —MOVZX (Move with Zero eXtend). It may be used in most cases as MOVSX, but it sets the
remaining bits to 0. That’s because printf() requires a 32-bit int, but we got a WORD in the structure —
that is 16-bit unsigned type. That’s why by copying the value from a WORD into int, bits from 16 to 31
must be cleared, because a random noise may be there, which is left from the previous operations on the
register(s).
In this example, it’s possible to represent the structure as an array of 8 WORDs:
#include <stdio.h>
345
void main()
{
WORD *t;
t=(WORD *)malloc (16);
GetSystemTime (t);
printf ("%04d-%02d-%02d %02d:%02d:%02d\n",

t[0] /* wYear */, t[1] /* wMonth */, t[3] /* wDay */,
t[4] /* wHour */, t[5] /* wMinute */, t[6] /* wSecond */);
free (t);
return;
};
We get:

$SG78594 DB '%04d-%02d-%02d %02d:%02d:%02d', 0aH, 00H
_main PROC
push esi
push 16
call _malloc
add esp, 4
mov esi, eax
push esi
movzx eax, WORD PTR [esi+12]
movzx ecx, WORD PTR [esi+10]
movzx edx, WORD PTR [esi+8]
push eax
movzx eax, WORD PTR [esi+6]
push ecx
movzx ecx, WORD PTR [esi+2]
push edx
movzx edx, WORD PTR [esi]
push eax
push ecx
push edx
call _printf
push esi
call _free
add esp, 32
xor eax, eax
pop esi
ret 0
_main ENDP
Again, we got the code that cannot be distinguished from the previous one.
And again it has to be noted, you haven’t to do this in practice, unless you really know what you are doing.
1.30.3 UNIX: struct tm

Linux
Let’s take the tm structure from time.h in Linux for example:

#include <stdio.h>
#include <time.h>
void main()
{
struct tm t;
time_t unix_time;
346
unix_time=time(NULL);
localtime_r (&unix_time, &t);
printf ("Year: %d\n", t.tm_year+1900);

printf ("Month: %d\n", t.tm_mon);
printf ("Day: %d\n", t.tm_mday);
printf ("Hour: %d\n", t.tm_hour);
printf ("Minutes: %d\n", t.tm_min);
printf ("Seconds: %d\n", t.tm_sec);
};
Let’s compile it in GCC 4.4.1:

main proc near
push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
sub esp, 40h
mov dword ptr [esp], 0 ; first argument for time()
call time
mov [esp+3Ch], eax
lea eax, [esp+3Ch] ; take pointer to what time() returned
lea edx, [esp+10h] ; at ESP+10h struct tm will begin
mov [esp+4], edx ; pass pointer to the structure begin
mov [esp], eax ; pass pointer to result of time()
call localtime_r
mov eax, [esp+24h] ; tm_year
lea edx, [eax+76Ch] ; edx=eax+1900
mov eax, offset format ; "Year: %d\n"
mov [esp+4], edx
mov [esp], eax
call printf
mov edx, [esp+20h] ; tm_mon
mov eax, offset aMonthD ; "Month: %d\n"
mov [esp+4], edx
mov [esp], eax
call printf
mov edx, [esp+1Ch] ; tm_mday
mov eax, offset aDayD ; "Day: %d\n"
mov [esp+4], edx
mov [esp], eax
call printf
mov edx, [esp+18h] ; tm_hour
mov eax, offset aHourD ; "Hour: %d\n"
mov [esp+4], edx
mov [esp], eax
call printf
mov edx, [esp+14h] ; tm_min
mov eax, offset aMinutesD ; "Minutes: %d\n"
mov [esp+4], edx
mov [esp], eax
call printf
mov edx, [esp+10h]
mov eax, offset aSecondsD ; "Seconds: %d\n"
mov [esp+4], edx ; tm_sec
mov [esp], eax
call printf
leave
retn
main endp
Somehow, IDA did not write the local variables’ names in the local stack. But since we already are expe-
rienced reverse engineers :-) we may do it without this information in this simple example.
Please also pay attention to the lea edx, [eax+76Ch] —this instruction just adds 0x76C (1900) to value
in EAX, but doesn’t modify any flags. See also the relevant section about LEA ( .1.6 on page 1014).
347
GDB
152
Let’s try to load the example into GDB :
Listing 1.336: GDB

dennis@ubuntuvm:~/polygon$ date
Mon Jun 2 18:10:37 EEST 2014
dennis@ubuntuvm:~/polygon$ gcc GCC_tm.c -o GCC_tm
dennis@ubuntuvm:~/polygon$ gdb GCC_tm
...
Reading symbols from /home/dennis/polygon/GCC_tm...(no debugging symbols found)...done.
(gdb) b printf
(gdb) run
Starting program: /home/dennis/polygon/GCC_tm
Breakpoint 1, __printf (format=0x80485c0 "Year: %d\n") at printf.c:29

(gdb) x/20x $esp
0xbffff0dc: 0x080484c3 0x080485c0 0x000007de 0x00000000
0xbffff0ec: 0x08048301 0x538c93ed 0x00000025 0x0000000a
0xbffff0fc: 0x00000012 0x00000002 0x00000005 0x00000072
0xbffff10c: 0x00000001 0x00000098 0x00000001 0x00002a30
0xbffff11c: 0x0804b090 0x08048530 0x00000000 0x00000000
(gdb)
We can easily find our structure in the stack. First, let’s see how it’s defined in time.h:
Listing 1.337: time.h

struct tm
{
int tm_sec;
int tm_min;
int tm_hour;
int tm_mday;
int tm_mon;
int tm_year;
int tm_wday;
int tm_yday;
int tm_isdst;
};
Pay attention that 32-bit int is used here instead of WORD in SYSTEMTIME. So, each field occupies 32-bit.
Here are the fields of our structure in the stack:
0xbffff0dc: 0x080484c3 0x080485c0 0x000007de 0x00000000
0xbffff0ec: 0x08048301 0x538c93ed 0x00000025 sec 0x0000000a min
0xbffff0fc: 0x00000012 hour 0x00000002 mday 0x00000005 mon 0x00000072 year
0xbffff10c: 0x00000001 wday 0x00000098 yday 0x00000001 isdst0x00002a30
0xbffff11c: 0x0804b090 0x08048530 0x00000000 0x00000000
Or as a table:
Hexadecimal number decimal number field name
0x00000025 37 tm_sec
0x0000000a 10 tm_min
0x00000012 18 tm_hour
0x00000002 2 tm_mday
0x00000005 5 tm_mon
0x00000072 114 tm_year
0x00000001 1 tm_wday
0x00000098 152 tm_yday
0x00000001 1 tm_isdst
152 The date result is slightly corrected for demonstration purposes. Of course, it’s not possible to run GDB that quickly, in the same
second.
348
Just like SYSTEMTIME ( 1.30.1 on page 341),
there are also other fields available that are not used, like tm_wday, tm_yday, tm_isdst.
ARM
Same example:

var_38 = -0x38
var_34 = -0x34
var_30 = -0x30
var_2C = -0x2C
var_28 = -0x28
var_24 = -0x24
timer = -0xC
PUSH {LR}
MOVS R0, #0 ; timer
SUB SP, SP, #0x34
BL time
STR R0, [SP,#0x38+timer]
MOV R1, SP ; tp
ADD R0, SP, #0x38+timer ; timer
BL localtime_r
LDR R1, =0x76C
LDR R0, [SP,#0x38+var_24]
ADDS R1, R0, R1
ADR R0, aYearD ; "Year: %d\n"
BL __2printf
ADR R0, aMonthD ; "Month: %d\n"
BL __2printf
LDR R1, [SP,#0x38+var_2C]
ADR R0, aDayD ; "Day: %d\n"
BL __2printf
ADR R0, aHourD ; "Hour: %d\n"
BL __2printf
ADR R0, aMinutesD ; "Minutes: %d\n"
BL __2printf
ADR R0, aSecondsD ; "Seconds: %d\n"
BL __2printf
ADD SP, SP, #0x34
POP {PC}
IDA “knows” the tm structure (because IDA “knows” the types of the arguments of library functions like
localtime_r()),
so it shows here structure elements accesses and their names.

var_38 = -0x38
var_34 = -0x34
PUSH {R7,LR}
MOV R7, SP
SUB SP, SP, #0x30
MOVS R0, #0 ; time_t *
BLX _time
ADD R1, SP, #0x38+var_34 ; struct tm *
349
MOV R0, SP ; time_t *
BLX _localtime_r
LDR R1, [SP,#0x38+var_34.tm_year]
MOV R0, 0xF44 ; "Year: %d\n"
ADD R0, PC ; char *
ADDW R1, R1, #0x76C
BLX _printf
LDR R1, [SP,#0x38+var_34.tm_mon]
MOV R0, 0xF3A ; "Month: %d\n"
ADD R0, PC ; char *
BLX _printf
LDR R1, [SP,#0x38+var_34.tm_mday]
MOV R0, 0xF35 ; "Day: %d\n"
ADD R0, PC ; char *
BLX _printf
LDR R1, [SP,#0x38+var_34.tm_hour]
MOV R0, 0xF2E ; "Hour: %d\n"
ADD R0, PC ; char *
BLX _printf
LDR R1, [SP,#0x38+var_34.tm_min]
MOV R0, 0xF28 ; "Minutes: %d\n"
ADD R0, PC ; char *
BLX _printf
MOV R0, 0xF25 ; "Seconds: %d\n"
ADD R0, PC ; char *
BLX _printf
ADD SP, SP, #0x30
POP {R7,PC}
...
00000000 tm struc ; (sizeof=0x2C, standard type)

00000000 tm_sec DCD ?
00000004 tm_min DCD ?
00000008 tm_hour DCD ?
0000000C tm_mday DCD ?
00000010 tm_mon DCD ?
00000014 tm_year DCD ?
00000018 tm_wday DCD ?
0000001C tm_yday DCD ?
00000020 tm_isdst DCD ?
00000024 tm_gmtoff DCD ?
00000028 tm_zone DCD ? ; offset
0000002C tm ends
MIPS

1 main:
2
3 ; IDA is not aware of structure field names, we named them manually:
4
5 var_40 = -0x40
6 var_38 = -0x38
7 seconds = -0x34
8 minutes = -0x30
9 hour = -0x2C
10 day = -0x28
11 month = -0x24
12 year = -0x20
13 var_4 = -4
14
15 lui $gp, (__gnu_local_gp >> 16)
16 addiu $sp, -0x50
17 la $gp, (__gnu_local_gp & 0xFFFF)
18 sw $ra, 0x50+var_4($sp)
350
19 sw $gp, 0x50+var_40($sp)
20 lw $t9, (time & 0xFFFF)($gp)
21 or $at, $zero ; load delay slot, NOP
22 jalr $t9
23 move $a0, $zero ; branch delay slot, NOP
24 lw $gp, 0x50+var_40($sp)
25 addiu $a0, $sp, 0x50+var_38
26 lw $t9, (localtime_r & 0xFFFF)($gp)
27 addiu $a1, $sp, 0x50+seconds
28 jalr $t9
29 sw $v0, 0x50+var_38($sp) ; branch delay slot
30 lw $gp, 0x50+var_40($sp)
31 lw $a1, 0x50+year($sp)
32 lw $t9, (printf & 0xFFFF)($gp)
33 la $a0, $LC0 # "Year: %d\n"
34 jalr $t9
35 addiu $a1, 1900 ; branch delay slot
36 lw $gp, 0x50+var_40($sp)
37 lw $a1, 0x50+month($sp)
39 lui $a0, ($LC1 >> 16) # "Month: %d\n"
40 jalr $t9
41 la $a0, ($LC1 & 0xFFFF) # "Month: %d\n" ; branch delay slot
42 lw $gp, 0x50+var_40($sp)
43 lw $a1, 0x50+day($sp)
45 lui $a0, ($LC2 >> 16) # "Day: %d\n"
46 jalr $t9
47 la $a0, ($LC2 & 0xFFFF) # "Day: %d\n" ; branch delay slot
48 lw $gp, 0x50+var_40($sp)
49 lw $a1, 0x50+hour($sp)
51 lui $a0, ($LC3 >> 16) # "Hour: %d\n"
52 jalr $t9
53 la $a0, ($LC3 & 0xFFFF) # "Hour: %d\n" ; branch delay slot
54 lw $gp, 0x50+var_40($sp)
55 lw $a1, 0x50+minutes($sp)
57 lui $a0, ($LC4 >> 16) # "Minutes: %d\n"
58 jalr $t9
59 la $a0, ($LC4 & 0xFFFF) # "Minutes: %d\n" ; branch delay slot
60 lw $gp, 0x50+var_40($sp)
61 lw $a1, 0x50+seconds($sp)
63 lui $a0, ($LC5 >> 16) # "Seconds: %d\n"
64 jalr $t9
65 la $a0, ($LC5 & 0xFFFF) # "Seconds: %d\n" ; branch delay slot
66 lw $ra, 0x50+var_4($sp)
68 jr $ra
69 addiu $sp, 0x50
70
71 $LC0: .ascii "Year: %d\n"<0>
72 $LC1: .ascii "Month: %d\n"<0>
73 $LC2: .ascii "Day: %d\n"<0>
74 $LC3: .ascii "Hour: %d\n"<0>
75 $LC4: .ascii "Minutes: %d\n"<0>
76 $LC5: .ascii "Seconds: %d\n"<0>
This is an example where the branch delay slots can confuse us.
For example, there is the instruction addiu $a1, 1900 at line 35 which adds 1900 to the year number.
It’s executed before the corresponding JALR at line 34, do not forget about it.
Structure as a set of values
In order to illustrate that the structure is just variables laying side-by-side in one place, let’s rework our
example while looking at the tm structure definition again: listing.1.337.
351
#include <stdio.h>
#include <time.h>
void main()
{
int tm_sec, tm_min, tm_hour, tm_mday, tm_mon, tm_year, tm_wday, tm_yday, tm_isdst;
time_t unix_time;
localtime_r (&unix_time, &tm_sec);
printf ("Year: %d\n", tm_year+1900);

printf ("Month: %d\n", tm_mon);
printf ("Day: %d\n", tm_mday);
printf ("Hour: %d\n", tm_hour);
printf ("Minutes: %d\n", tm_min);
printf ("Seconds: %d\n", tm_sec);
};
N.B. The pointer to the tm_sec field is passed into localtime_r, i.e., to the first element of the “structure”.
The compiler warns us:

GCC_tm2.c: In function 'main':
GCC_tm2.c:11:5: warning: passing argument 2 of 'localtime_r' from incompatible pointer type [⤦
Ç enabled by default]
In file included from GCC_tm2.c:2:0:
/usr/include/time.h:59:12: note: expected 'struct tm *' but argument is of type 'int *'
But nevertheless, it generates this:

main proc near

unix_time = dword ptr -1Ch
tm_sec = dword ptr -18h
tm_min = dword ptr -14h
tm_hour = dword ptr -10h
tm_mday = dword ptr -0Ch
tm_mon = dword ptr -8
tm_year = dword ptr -4
push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
sub esp, 30h
call __main
mov [esp+30h+var_30], 0 ; arg 0
call time
mov [esp+30h+unix_time], eax
lea eax, [esp+30h+tm_sec]
mov [esp+30h+var_2C], eax
lea eax, [esp+30h+unix_time]
call localtime_r
mov eax, [esp+30h+tm_year]
add eax, 1900
mov [esp+30h+var_30], offset aYearD ; "Year: %d\n"
call printf
mov eax, [esp+30h+tm_mon]
mov [esp+30h+var_30], offset aMonthD ; "Month: %d\n"
call printf
352
mov eax, [esp+30h+tm_mday]
mov [esp+30h+var_30], offset aDayD ; "Day: %d\n"
call printf
mov eax, [esp+30h+tm_hour]
mov [esp+30h+var_30], offset aHourD ; "Hour: %d\n"
call printf
mov eax, [esp+30h+tm_min]
mov [esp+30h+var_30], offset aMinutesD ; "Minutes: %d\n"
call printf
mov eax, [esp+30h+tm_sec]
mov [esp+30h+var_30], offset aSecondsD ; "Seconds: %d\n"
call printf
leave
retn
main endp
This code is identical to what we saw previously and it is not possible to say, was it a structure in original
source code or just a pack of variables.
And this works. However, it is not recommended to do this in practice.
Usually, non-optimizing compilers allocates variables in the local stack in the same order as they were
declared in the function.
Nevertheless, there is no guarantee.
By the way, some other compiler may warn about the tm_year, tm_mon, tm_mday, tm_hour, tm_min vari-
ables, but not tm_sec are used without being initialized.
Indeed, the compiler is not aware that these are to be filled by
localtime_r() function.
We chose this example, since all structure fields are of type int.
This would not work if structure fields are 16-bit (WORD), like in the case of the SYSTEMTIME structure—
GetSystemTime() will fill them incorrectly (because the local variables are aligned on a 32-bit boundary).
Read more about it in next section: “Fields packing in structure” ( 1.30.4 on page 356).
So, a structure is just a pack of variables laying in one place, side-by-side. We could say that the structure
is the instruction to the compiler, directing it to hold variables in one place. By the way, in some very
early C versions (before 1972), there were no structures at all [Dennis M. Ritchie, The development of the
C language, (1993)]153 .
There is no debugger example here: it is just the same as you already saw.
Structure as an array of 32-bit words

#include <stdio.h>
#include <time.h>
void main()
{
struct tm t;
time_t unix_time;
int i;
for (i=0; i<9; i++)

{
int tmp=((int*)&t)[i];
printf ("0x%08X (%d)\n", tmp, tmp);
};
};
353
We just cast a pointer to structure to an array of int’s. And that works! We run the example at 23:51:45
26-July-2014.
0x0000002D (45)
0x00000033 (51)
0x00000017 (23)
0x0000001A (26)
0x00000006 (6)
0x00000072 (114)
0x00000006 (6)
0x000000CE (206)
0x00000001 (1)
The variables here are in the same order as they are enumerated in the definition of the structure: 1.337
on page 348.
Here is how it gets compiled:

main proc near
push ebp
mov ebp, esp
push esi
push ebx
and esp, 0FFFFFFF0h
sub esp, 40h
mov dword ptr [esp], 0 ; timer
lea ebx, [esp+14h]
call _time
lea esi, [esp+38h]
mov [esp+4], ebx ; tp
mov [esp+10h], eax
lea eax, [esp+10h]
mov [esp], eax ; timer
call _localtime_r
nop
lea esi, [esi+0] ; NOP
loc_80483D8:
; EBX here is pointer to structure, ESI is the pointer to the end of it.
mov eax, [ebx] ; get 32-bit word from array
add ebx, 4 ; next field in structure
mov dword ptr [esp+4], offset a0x08xD ; "0x%08X (%d)\n"
mov dword ptr [esp], 1
mov [esp+0Ch], eax ; pass value to printf()
mov [esp+8], eax ; pass value to printf()
call ___printf_chk
cmp ebx, esi ; meet structure end?
jnz short loc_80483D8 ; no - load next value then
lea esp, [ebp-8]
pop ebx
pop esi
pop ebp
retn
main endp
Indeed: the space in the local stack is first treated as a structure, and then it’s treated as an array.
It’s even possible to modify the fields of the structure through this pointer.
And again, it’s dubiously hackish way to do things, not recommended for use in production code.
Exercise
As an exercise, try to modify (increase by 1) the current month number, treating the structure as an array.
Structure as an array of bytes
We can go even further. Let’s cast the pointer to an array of bytes and dump it:
354
#include <stdio.h>
#include <time.h>
void main()
{
struct tm t;
time_t unix_time;
int i, j;
for (i=0; i<9; i++)

{
for (j=0; j<4; j++)
printf ("0x%02X ", ((unsigned char*)&t)[i*4+j]);
printf ("\n");
};
};
0x2D 0x00 0x00 0x00

0x33 0x00 0x00 0x00
0x17 0x00 0x00 0x00
0x1A 0x00 0x00 0x00
0x06 0x00 0x00 0x00
0x72 0x00 0x00 0x00
0x06 0x00 0x00 0x00
0xCE 0x00 0x00 0x00
0x01 0x00 0x00 0x00
We also run this example at 23:51:45 26-July-2014 154 . The values are just the same as in the previous
dump ( 1.30.3 on the preceding page), and of course, the lowest byte goes first, because this is a little-
endian architecture ( 2.8 on page 460).

main proc near
push ebp
mov ebp, esp
push edi
push esi
push ebx
and esp, 0FFFFFFF0h
sub esp, 40h
mov dword ptr [esp], 0 ; timer
lea esi, [esp+14h]
call _time
lea edi, [esp+38h] ; struct end
mov [esp+4], esi ; tp
mov [esp+10h], eax
lea eax, [esp+10h]
mov [esp], eax ; timer
call _localtime_r
lea esi, [esi+0] ; NOP
; ESI here is the pointer to structure in local stack. EDI is the pointer to structure end.
loc_8048408:
xor ebx, ebx ; j=0
loc_804840A:
movzx eax, byte ptr [esi+ebx] ; load byte
add ebx, 1 ; j=j+1
mov dword ptr [esp+4], offset a0x02x ; "0x%02X "
mov [esp+8], eax ; pass loaded byte to printf()
call ___printf_chk
cmp ebx, 4
154 The time and date are the same for demonstration purposes. Byte values are fixed up.
355
jnz short loc_804840A
; print carriage return character (CR)
mov dword ptr [esp], 0Ah ; c
add esi, 4
call _putchar
cmp esi, edi ; meet struct end?
jnz short loc_8048408 ; j=0
lea esp, [ebp-0Ch]
pop ebx
pop esi
pop edi
pop ebp
retn
main endp
GNU Scientific Library: Representation of complex numbers
This is a relatively rare case when an array is used instead of a structure, on purpose:
Representation of complex numbers
=================================
Complex numbers are represented using the type :code:`gsl_complex`. The

internal representation of this type may vary across platforms and
should not be accessed directly. The functions and macros described
below allow complex numbers to be manipulated in a portable way.
For reference, the default form of the :code:`gsl_complex` type is

given by the following struct::
typedef struct
{
double dat[2];
} gsl_complex;
The real and imaginary part are stored in contiguous elements of a two
element array. This eliminates any padding between the real and
imaginary parts, :code:`dat[0]` and :code:`dat[1]`, allowing the struct to
be mapped correctly onto packed complex arrays.
( https://www.gnu.org/software/gsl/doc/html/complex.html#representation-of-complex-numbers
)
1.30.4 Fields packing in structure

One important thing is fields packing in structures.
Let’s take a simple example:
#include <stdio.h>
struct s
{
char a;
int b;
char c;
int d;
};
void f(struct s s)
{
printf ("a=%d; b=%d; c=%d; d=%d\n", s.a, s.b, s.c, s.d);
};
int main()
{
struct s tmp;
tmp.a=1;
tmp.b=2;
356
tmp.c=3;
tmp.d=4;
f(tmp);
};
As we see, we have two char fields (each is exactly one byte) and two more —int (each — 4 bytes).
x86
This compiles to:
Listing 1.345: MSVC 2012 /GS- /Ob0

1 _tmp$ = -16
2 _main PROC
3 push ebp
4 mov ebp, esp
5 sub esp, 16
6 mov BYTE PTR _tmp$[ebp], 1 ; set field a
7 mov DWORD PTR _tmp$[ebp+4], 2 ; set field b
8 mov BYTE PTR _tmp$[ebp+8], 3 ; set field c
9 mov DWORD PTR _tmp$[ebp+12], 4 ; set field d
10 sub esp, 16 ; allocate place for temporary structure
11 mov eax, esp
12 mov ecx, DWORD PTR _tmp$[ebp] ; copy our structure to the temporary one
13 mov DWORD PTR [eax], ecx
14 mov edx, DWORD PTR _tmp$[ebp+4]
15 mov DWORD PTR [eax+4], edx
16 mov ecx, DWORD PTR _tmp$[ebp+8]
17 mov DWORD PTR [eax+8], ecx
20 call _f
21 add esp, 16
22 xor eax, eax
23 mov esp, ebp
24 pop ebp
25 ret 0
26 _main ENDP
27
28 _s$ = 8 ; size = 16
29 ?f@@YAXUs@@@Z PROC ; f
30 push ebp
31 mov ebp, esp
32 mov eax, DWORD PTR _s$[ebp+12]
33 push eax
34 movsx ecx, BYTE PTR _s$[ebp+8]
35 push ecx
36 mov edx, DWORD PTR _s$[ebp+4]
37 push edx
38 movsx eax, BYTE PTR _s$[ebp]
39 push eax
40 push OFFSET $SG3842
41 call _printf
42 add esp, 20
43 pop ebp
44 ret 0
45 ?f@@YAXUs@@@Z ENDP ; f
46 _TEXT ENDS
We pass the structure as a whole, but in fact, as we can see, the structure is being copied to a temporary
one (a place in stack is allocated in line 10 for it, and then all 4 fields, one by one, are copied in lines 12
… 19), and then its pointer (address) is to be passed.
The structure is copied because it’s not known whether the f() function going to modify the structure or
not. If it gets changed, then the structure in main() has to remain as it has been.
We could use C/C++ pointers, and the resulting code will be almost the same, but without the copying.
357
As we can see, each field’s address is aligned on a 4-byte boundary. That’s why each char occupies 4
bytes here (like int). Why? Because it is easier for the CPU to access memory at aligned addresses and
to cache data from it.
However, it is not very economical.
Let’s try to compile it with option (/Zp1) (/Zp[n] pack structures on n-byte boundary).
Listing 1.346: MSVC 2012 /GS- /Zp1

1 _main PROC
2 push ebp
3 mov ebp, esp
4 sub esp, 12
5 mov BYTE PTR _tmp$[ebp], 1 ; set field a
6 mov DWORD PTR _tmp$[ebp+1], 2 ; set field b
7 mov BYTE PTR _tmp$[ebp+5], 3 ; set field c
8 mov DWORD PTR _tmp$[ebp+6], 4 ; set field d
9 sub esp, 12 ; allocate place for temporary structure
10 mov eax, esp
11 mov ecx, DWORD PTR _tmp$[ebp] ; copy 10 bytes
12 mov DWORD PTR [eax], ecx
15 mov cx, WORD PTR _tmp$[ebp+8]
16 mov WORD PTR [eax+8], cx
17 call _f
18 add esp, 12
19 xor eax, eax
20 mov esp, ebp
21 pop ebp
22 ret 0
23 _main ENDP
24
25 _TEXT SEGMENT
26 _s$ = 8 ; size = 10
27 ?f@@YAXUs@@@Z PROC ; f
28 push ebp
29 mov ebp, esp
30 mov eax, DWORD PTR _s$[ebp+6]
31 push eax
32 movsx ecx, BYTE PTR _s$[ebp+5]
33 push ecx
34 mov edx, DWORD PTR _s$[ebp+1]
35 push edx
36 movsx eax, BYTE PTR _s$[ebp]
37 push eax
38 push OFFSET $SG3842
39 call _printf
40 add esp, 20
41 pop ebp
42 ret 0
43 ?f@@YAXUs@@@Z ENDP ; f
Now the structure takes only 10 bytes and each char value takes 1 byte. What does it give to us? Size
economy. And as drawback —the CPU accessing these fields slower than it could.
The structure is also copied in main(). Not field-by-field, but directly 10 bytes, using three pairs of MOV.
Why not 4?
The compiler decided that it’s better to copy 10 bytes using 3 MOV pairs than to copy two 32-bit words and
two bytes using 4 MOV pairs.
By the way, such copy implementation using MOV instead of calling the memcpy() function is widely used,
because it’s faster than a call to memcpy()—for short blocks, of course: 3.14.1 on page 511.
As it can be easily guessed, if the structure is used in many source and object files, all these must be
compiled with the same convention about structures packing.
Aside from MSVC /Zp option which sets how to align each structure field, there is also the #pragma pack
358
compiler option, which can be defined right in the source code. It is available in both MSVC155 and GCC156 .
Let’s get back to the SYSTEMTIME structure that consists of 16-bit fields. How does our compiler know to
pack them on 1-byte alignment boundary?
WinNT.h file has this:

#include "pshpack1.h"
And this:

#include "pshpack4.h" // 4 byte packing is the default
The file PshPack1.h looks like:
Listing 1.349: PshPack1.h

#if ! (defined(lint) || defined(RC_INVOKED))
#if ( _MSC_VER >= 800 && !defined(_M_I86)) || defined(_PUSHPOP_SUPPORTED)
#pragma warning(disable:4103)
#if !(defined( MIDL_PASS )) || defined( __midl )
#pragma pack(push,1)
#else
#pragma pack(1)
#endif
#else
#pragma pack(1)
#endif
#endif /* ! (defined(lint) || defined(RC_INVOKED)) */
This tell the compiler how to pack the structures defined after #pragma pack.
155 MSDN: Working with Packing Structures

156 Structure-Packing Pragmas
359
OllyDbg + fields are packed by default
Let’s try our example (where the fields are aligned by default (4 bytes)) in OllyDbg:
Figure 1.106: OllyDbg: Before printf() execution
We see our 4 fields in the data window.

But where do the random bytes (0x30, 0x37, 0x01) come from, that are next to the first (a) and third (c)
fields?
By looking at our listing 1.345 on page 357, we can see that the first and third fields are char, therefore
only one byte is written, 1 and 3 respectively (lines 6 and 8).
The remaining 3 bytes of the 32-bit words are not being modified in memory! Hence, random garbage is
left there.
This garbage doesn’t influence the printf() output in any way, because the values for it are prepared
using the MOVSX instruction, which takes bytes, not words: listing.1.345 (lines 34 and 38).
By the way, the MOVSX (sign-extending) instruction is used here, because char is signed by default in MSVC
and GCC. If the unsigned char data type or uint8_t was used here, MOVZX instruction would have been
used instead.
360
OllyDbg + fields aligning on 1 byte boundary
Things are much clearer here: 4 fields occupy 10 bytes and the values are stored side-by-side
ARM

.text:0000003E exit ; CODE XREF: f+16
.text:0000003E 05 B0 ADD SP, SP, #0x14
.text:00000040 00 BD POP {PC}
.text:00000280 f
.text:00000280
.text:00000280 var_18 = -0x18
.text:00000280 a = -0x14
.text:00000280 b = -0x10
.text:00000280 c = -0xC
.text:00000280 d = -8
.text:00000280
.text:00000280 0F B5 PUSH {R0-R3,LR}
.text:00000282 81 B0 SUB SP, SP, #4
.text:00000284 04 98 LDR R0, [SP,#16] ; d
.text:00000286 02 9A LDR R2, [SP,#8] ; b
.text:00000288 00 90 STR R0, [SP]
.text:0000028A 68 46 MOV R0, SP
.text:0000028C 03 7B LDRB R3, [R0,#12] ; c
.text:0000028E 01 79 LDRB R1, [R0,#4] ; a
.text:00000290 59 A0 ADR R0, aADBDCDDD ; "a=%d; b=%d; c=%d; d=%d\n"
.text:00000292 05 F0 AD FF BL __2printf
.text:00000296 D2 E6 B exit
As we may recall, here a structure is passed instead of pointer to one, and since the first 4 function
arguments in ARM are passed via registers, the structure’s fields are passed via R0-R3.
LDRB loads one byte from memory and extends it to 32-bit, taking its sign into account. This is similar to
MOVSX in x86. Here it is used to load fields a and c from the structure.
361
One more thing we spot easily is that instead of function epilogue, there is jump to another function’s
epilogue! Indeed, that was quite different function, not related in any way to ours, however, it has exactly
the same epilogue (probably because, it hold 5 local variables too (5 ∗ 4 = 0x14)).
Also it is located nearby (take a look at the addresses).
Indeed, it doesn’t matter which epilogue gets executed, if it works just as we need.
Apparently, Keil decides to reuse a part of another function to economize.
The epilogue takes 4 bytes while jump—only 2.
ARM + Optimizing Xcode 4.6.3 (LLVM) (Thumb-2 mode)

var_C = -0xC
PUSH {R7,LR}
MOV R7, SP
SUB SP, SP, #4
MOV R9, R1 ; b
MOV R1, R0 ; a
MOVW R0, #0xF10 ; "a=%d; b=%d; c=%d; d=%d\n"
SXTB R1, R1 ; prepare a
MOVT.W R0, #0
STR R3, [SP,#0xC+var_C] ; place d to stack for printf()
ADD R0, PC ; format-string
SXTB R3, R2 ; prepare c
MOV R2, R9 ; b
BLX _printf
ADD SP, SP, #4
POP {R7,PC}
SXTB (Signed Extend Byte) is analogous to MOVSX in x86. All the rest—just the same.
MIPS

1 f:
2
3 var_18 = -0x18
4 var_10 = -0x10
5 var_4 = -4
6 arg_0 = 0
7 arg_4 = 4
8 arg_8 = 8
9 arg_C = 0xC
10
11 ; $a0=s.a
12 ; $a1=s.b
13 ; $a2=s.c
14 ; $a3=s.d
15 lui $gp, (__gnu_local_gp >> 16)
16 addiu $sp, -0x28
17 la $gp, (__gnu_local_gp & 0xFFFF)
18 sw $ra, 0x28+var_4($sp)
19 sw $gp, 0x28+var_10($sp)
20 ; prepare a byte from 32-bit big-endian integer:
21 sra $t0, $a0, 24
22 move $v1, $a1
23 ; prepare a byte from 32-bit big-endian integer:
24 sra $v0, $a2, 24
26 sw $a0, 0x28+arg_0($sp)
27 lui $a0, ($LC0 >> 16) # "a=%d; b=%d; c=%d; d=%d\n"
28 sw $a3, 0x28+var_18($sp)
29 sw $a1, 0x28+arg_4($sp)
30 sw $a2, 0x28+arg_8($sp)
362
31 sw $a3, 0x28+arg_C($sp)
32 la $a0, ($LC0 & 0xFFFF) # "a=%d; b=%d; c=%d; d=%d\n"
33 move $a1, $t0
34 move $a2, $v1
35 jalr $t9
36 move $a3, $v0 ; branch delay slot
37 lw $ra, 0x28+var_4($sp)
39 jr $ra
40 addiu $sp, 0x28 ; branch delay slot
41
42 $LC0: .ascii "a=%d; b=%d; c=%d; d=%d\n"<0>
Structure fields come in registers $A0..$A3 and then get reshuffled into $A1..$A3 for printf(), while 4th
field (from $A3) is passed via local stack using SW.
But there are two SRA (“Shift Word Right Arithmetic”) instructions, which prepare char fields. Why?
MIPS is a big-endian architecture by default 2.8 on page 460, and the Debian Linux we work in is big-endian
as well.
So when byte variables are stored in 32-bit structure slots, they occupy the high 31..24 bits.
And when a char variable needs to be extended into a 32-bit value, it must be shifted right by 24 bits.
char is a signed type, so an arithmetical shift is used here instead of logical.
One more word
Passing a structure as a function argument (instead of a passing pointer to structure) is the same as
passing all structure fields one by one.
If the structure fields are packed by default, the f() function can be rewritten as:
void f(char a, int b, char c, int d)
{
printf ("a=%d; b=%d; c=%d; d=%d\n", a, b, c, d);
};
And that leads to the same code.
1.30.5 Nested structures

Now what about situations when one structure is defined inside of another?
#include <stdio.h>
struct inner_struct
{
int a;
int b;
};
struct outer_struct
{
char a;
int b;
struct inner_struct c;
char d;
int e;
};
void f(struct outer_struct s)

{
printf ("a=%d; b=%d; c.a=%d; c.b=%d; d=%d; e=%d\n",
s.a, s.b, s.c.a, s.c.b, s.d, s.e);
};
int main()
{
363
struct outer_struct s;
s.a=1;
s.b=2;
s.c.a=100;
s.c.b=101;
s.d=3;
s.e=4;
f(s);
};
…in this case, both inner_struct fields are to be placed between the a,b and d,e fields of the outer_struct.
$SG2802 DB 'a=%d; b=%d; c.a=%d; c.b=%d; d=%d; e=%d', 0aH, 00H
_TEXT SEGMENT
_s$ = 8
_f PROC
mov eax, DWORD PTR _s$[esp+16]
movsx ecx, BYTE PTR _s$[esp+12]
mov edx, DWORD PTR _s$[esp+8]
push eax
push ecx
mov ecx, DWORD PTR _s$[esp+8]
push edx
movsx edx, BYTE PTR _s$[esp+8]
push eax
push ecx
push edx
push OFFSET $SG2802 ; 'a=%d; b=%d; c.a=%d; c.b=%d; d=%d; e=%d'
call _printf
add esp, 28
ret 0
_f ENDP
_s$ = -24
_main PROC
sub esp, 24
push ebx
push esi
push edi
mov ecx, 2
sub esp, 24
mov eax, esp
; from this moment, EAX is synonymous to ESP:
mov BYTE PTR _s$[esp+60], 1
mov ebx, DWORD PTR _s$[esp+60]
mov DWORD PTR [eax], ebx
lea esi, DWORD PTR [ecx+99]
lea edi, DWORD PTR [ecx+2]
mov BYTE PTR _s$[esp+76], 3
mov ecx, DWORD PTR _s$[esp+76]
mov DWORD PTR [eax+12], esi
mov DWORD PTR [eax+20], edi
call _f
add esp, 24
pop edi
pop esi
xor eax, eax
pop ebx
add esp, 24
ret 0
_main ENDP
364
One curious thing here is that by looking onto this assembly code, we do not even see that another
structure was used inside of it! Thus, we would say, nested structures are unfolded into linear or one-
dimensional structure.
Of course, if we replace the struct inner_struct c; declaration with struct inner_struct *c; (thus
making a pointer here) the situation will be quite different.
365
OllyDbg
Let’s load the example into OllyDbg and take a look at outer_struct in memory:
That’s how the values are located in memory:

• (outer_struct.a) (byte) 1 + 3 bytes of random garbage;
• (outer_struct.b) (32-bit word) 2;
• (inner_struct.a) (32-bit word) 0x64 (100);
• (inner_struct.b) (32-bit word) 0x65 (101);
• (outer_struct.d) (byte) 3 + 3 bytes of random garbage;
• (outer_struct.e) (32-bit word) 4.
1.30.6 Bit fields in a structure

CPUID example
The C/C++ language allows to define the exact number of bits for each structure field. It is very useful if
one needs to save memory space. For example, one bit is enough for a bool variable. But of course, it is
not rational if speed is important.
Let’s consider the CPUID157 instruction example. This instruction returns information about the current
CPU and its features.
If the EAX is set to 1 before the instruction’s execution, CPUID returning this information packed into the
EAX register:
3:0 (4 bits) Stepping
7:4 (4 bits) Model
11:8 (4 bits) Family
13:12 (2 bits) Processor Type
19:16 (4 bits) Extended Model
27:20 (8 bits) Extended Family
MSVC 2010 has CPUID macro, but GCC 4.4.1 does not. So let’s make this function by ourselves for GCC
with the help of its built-in assembler158 .
157 wikipedia
158 More about internal GCC assembler
366
#include <stdio.h>
#ifdef __GNUC__
static inline void cpuid(int code, int *a, int *b, int *c, int *d) {
asm volatile("cpuid":"=a"(*a),"=b"(*b),"=c"(*c),"=d"(*d):"a"(code));
}
#endif
#ifdef _MSC_VER
#include <intrin.h>
#endif
struct CPUID_1_EAX
{
unsigned int stepping:4;
unsigned int model:4;
unsigned int family_id:4;
unsigned int processor_type:2;
unsigned int reserved1:2;
unsigned int extended_model_id:4;
unsigned int extended_family_id:8;
unsigned int reserved2:4;
};
int main()
{
struct CPUID_1_EAX *tmp;
int b[4];
#ifdef _MSC_VER
__cpuid(b,1);
#endif
#ifdef __GNUC__
cpuid (1, &b[0], &b[1], &b[2], &b[3]);
#endif
tmp=(struct CPUID_1_EAX *)&b[0];
printf ("stepping=%d\n", tmp->stepping);

printf ("model=%d\n", tmp->model);
printf ("family_id=%d\n", tmp->family_id);
printf ("processor_type=%d\n", tmp->processor_type);
printf ("extended_model_id=%d\n", tmp->extended_model_id);
printf ("extended_family_id=%d\n", tmp->extended_family_id);
return 0;
};
After CPUID fills EAX/EBX/ECX/EDX, these registers are to be written in the b[] array. Then, we have a
pointer to the CPUID_1_EAX structure and we point it to the value in EAX from the b[] array.
In other words, we treat a 32-bit int value as a structure. Then we read specific bits from the structure.
MSVC
Let’s compile it in MSVC 2008 with /Ox option:

_b$ = -16 ; size = 16
_main PROC
sub esp, 16
push ebx
xor ecx, ecx

mov eax, 1
cpuid
push esi
367
lea esi, DWORD PTR _b$[esp+24]
mov DWORD PTR [esi], eax
mov DWORD PTR [esi+4], ebx
mov DWORD PTR [esi+8], ecx
mov DWORD PTR [esi+12], edx
mov esi, DWORD PTR _b$[esp+24]

mov eax, esi
and eax, 15
push eax
push OFFSET $SG15435 ; 'stepping=%d', 0aH, 00H
call _printf
mov ecx, esi

shr ecx, 4
and ecx, 15
push ecx
push OFFSET $SG15436 ; 'model=%d', 0aH, 00H
call _printf
mov edx, esi

shr edx, 8
and edx, 15
push edx
push OFFSET $SG15437 ; 'family_id=%d', 0aH, 00H
call _printf
mov eax, esi

shr eax, 12
and eax, 3
push eax
push OFFSET $SG15438 ; 'processor_type=%d', 0aH, 00H
call _printf
mov ecx, esi

shr ecx, 16
and ecx, 15
push ecx
push OFFSET $SG15439 ; 'extended_model_id=%d', 0aH, 00H
call _printf
shr esi, 20
and esi, 255
push esi
push OFFSET $SG15440 ; 'extended_family_id=%d', 0aH, 00H
call _printf
add esp, 48
pop esi
xor eax, eax

pop ebx
add esp, 16
ret 0
_main ENDP
The SHR instruction shifting the value in EAX by the number of bits that must be skipped, e.g., we ignore
some bits at the right side.
The AND instruction clears the unneeded bits on the left, or, in other words, leaves only those bits in the
EAX register we need.
368
MSVC + OllyDbg
Let’s load our example into OllyDbg and see, what values are set in EAX/EBX/ECX/EDX after the execution
of CPUID:
Figure 1.109: OllyDbg: After CPUID execution
EAX has 0x000206A7 (my CPU is Intel Xeon E3-1220).

This is 0b00000000000000100000011010100111 in binary form.
Here is how the bits are distributed by fields:
field in binary form in decimal form
reserved2 0000 0
extended_family_id 00000000 0
extended_model_id 0010 2
reserved1 00 0
processor_id 00 0
family_id 0110 6
model 1010 10
stepping 0111 7
Listing 1.355: Console output

stepping=7
model=10
family_id=6
processor_type=0
extended_model_id=2
extended_family_id=0
GCC
Let’s try GCC 4.4.1 with -O3 option.

main proc near ; DATA XREF: _start+17
push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
push esi
mov esi, 1
369
push ebx
mov eax, esi
sub esp, 18h
cpuid
mov esi, eax
and eax, 0Fh
mov [esp+8], eax
mov dword ptr [esp+4], offset aSteppingD ; "stepping=%d\n"
call ___printf_chk
mov eax, esi
shr eax, 4
and eax, 0Fh
mov [esp+8], eax
mov dword ptr [esp+4], offset aModelD ; "model=%d\n"
call ___printf_chk
mov eax, esi
shr eax, 8
and eax, 0Fh
mov [esp+8], eax
mov dword ptr [esp+4], offset aFamily_idD ; "family_id=%d\n"
call ___printf_chk
mov eax, esi
shr eax, 0Ch
and eax, 3
mov [esp+8], eax
mov dword ptr [esp+4], offset aProcessor_type ; "processor_type=%d\n"
call ___printf_chk
mov eax, esi
shr eax, 10h
shr esi, 14h
and eax, 0Fh
and esi, 0FFh
mov [esp+8], eax
mov dword ptr [esp+4], offset aExtended_model ; "extended_model_id=%d\n"
call ___printf_chk
mov [esp+8], esi
mov dword ptr [esp+4], offset unk_80486D0
call ___printf_chk
add esp, 18h
xor eax, eax
pop ebx
pop esi
mov esp, ebp
pop ebp
retn
main endp
Almost the same. The only thing worth noting is that GCC somehow combines the calculation of
extended_model_id and extended_family_id into one block, instead of calculating them separately be-
fore each printf() call.
Handling float data type as a structure
As we already noted in the section about FPU ( 1.25 on page 217), both float and double types consist of
a sign, a significand (or fraction) and an exponent. But will we be able to work with these fields directly?
Let’s try this with float.
31 30 23 22 0
S exponent mantissa or fraction
( S — sign )
370
#include <stdio.h>
#include <assert.h>
#include <stdlib.h>
#include <memory.h>
struct float_as_struct
{
unsigned int fraction : 23; // fractional part
unsigned int exponent : 8; // exponent + 0x3FF
unsigned int sign : 1; // sign bit
};
float f(float _in)

{
float f=_in;
struct float_as_struct t;
assert (sizeof (struct float_as_struct) == sizeof (float));
memcpy (&t, &f, sizeof (float));
t.sign=1; // set negative sign

t.exponent=t.exponent+2; // multiply d by 2n (n here is 2)
memcpy (&f, &t, sizeof (float));
return f;
};
int main()
{
printf ("%f\n", f(1.234));
};
The float_as_struct structure occupies the same amount of memory as float, i.e., 4 bytes or 32 bits.
Now we are setting the negative sign in the input value and also, by adding 2 to the exponent, we thereby
multiply the whole number by 22 , i.e., by 4.
Let’s compile in MSVC 2008 without optimization turned on:

_t$ = -8 ; size = 4
_f$ = -4 ; size = 4
__in$ = 8 ; size = 4
?f@@YAMM@Z PROC ; f
push ebp
mov ebp, esp
sub esp, 8
fld DWORD PTR __in$[ebp]

fstp DWORD PTR _f$[ebp]
push 4
lea eax, DWORD PTR _f$[ebp]
push eax
lea ecx, DWORD PTR _t$[ebp]
push ecx
call _memcpy
add esp, 12
mov edx, DWORD PTR _t$[ebp]

or edx, -2147483648 ; 80000000H - set minus sign
mov DWORD PTR _t$[ebp], edx
mov eax, DWORD PTR _t$[ebp]

shr eax, 23 ; 00000017H - drop significand
and eax, 255 ; 000000ffH - leave here only exponent
add eax, 2 ; add 2 to it
371
and eax, 255 ; 000000ffH
shl eax, 23 ; 00000017H - shift result to place of bits 30:23
mov ecx, DWORD PTR _t$[ebp]
and ecx, -2139095041 ; 807fffffH - drop exponent
; add original value without exponent with new calculated exponent:

or ecx, eax
mov DWORD PTR _t$[ebp], ecx
push 4
lea edx, DWORD PTR _t$[ebp]
push edx
lea eax, DWORD PTR _f$[ebp]
push eax
call _memcpy
add esp, 12
fld DWORD PTR _f$[ebp]
mov esp, ebp

pop ebp
ret 0
?f@@YAMM@Z ENDP ; f
A bit redundant. If it was compiled with /Ox flag there would be no memcpy() call, the f variable is used
directly. But it is easier to understand by looking at the unoptimized version.
What would GCC 4.4.1 with -O3 do?

; f(float)
public _Z1ff
_Z1ff proc near

arg_0 = dword ptr 8
push ebp
mov ebp, esp
sub esp, 4
or eax, 80000000h ; set minus sign
mov edx, eax
and eax, 807FFFFFh ; leave only sign and significand in EAX
shr edx, 23 ; prepare exponent
add edx, 2 ; add 2
movzx edx, dl ; clear all bits except 7:0 in EDX
shl edx, 23 ; shift new calculated exponent to its place
or eax, edx ; join new exponent and original value without exponent
fld [ebp+var_4]
leave
retn
_Z1ff endp
public main
main proc near
push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
sub esp, 10h
fld ds:dword_8048614 ; -4.936
fstp qword ptr [esp+8]
mov dword ptr [esp+4], offset asc_8048610 ; "%f\n"
call ___printf_chk
xor eax, eax
leave
retn
main endp
372
The f() function is almost understandable. However, what is interesting is that GCC was able to calcu-
late the result of f(1.234) during compilation despite all this hodge-podge with the structure fields and
prepared this argument to printf() as precalculated at compile time!
1.30.7 Exercises
1.31 The classic struct bug

This is a classic struct bug.
Here is a sample definition:
struct test
{
int field1;
int field2;
};
And then C files:

void setter(struct test *t, int a, int b)
{
t->field1=a;
t->field2=b;
};
#include <stdio.h>
void printer(struct test *t)

{
printf ("%d\n", t->field1);
printf ("%d\n", t->field2);
};
So far so good.
Now you add a third field into the structure, some place between two fields:
struct test
{
int field1;
int inserted;
int field2;
};
And you probably modify setter() function, but forget about printer():
void setter(struct test *t, int a, int b, int c)
{
t->field1=a;
t->inserted=b;
t->field2=c;
};
You compile your project, but the C file where printer() is residing, isn’t recompiling, because your
IDE159 or build system has no idea that module depends on a test struct definition. Maybe because
#include <new.h> is omitted. Or maybe, new.h header file is included in printer.c via some other
header file. The object file remains untouched (IDE thinks it doesn’t need to be recompiled), while
setter() function is already a new version. These two object files (old and new) eventually linked into
an executable file.
159 Integrated development environment
373
Then you run it, and the setter() sets 3 fields at +0, +4 and +8 offsets. However, the printer() only
knows about 2 fields, and gets them from +0 and +4 offsets during printing.
This leads to very obscure and nasty bugs. The reason is that IDE or build system or Makefile doesn’t
know the fact that both C files (or modules) depends on the header file with test definition. A popular
remedy is to clean everything and recompile.
This is true for C++ classes as well, since they works just like structures: 3.21.1 on page 541.
This is a C/C++’s malady, and a source of criticism, yes. Many newer PLs has better support of modules
and interfaces. But keep in mind, when C compiler was created: 1970s, on old PDP computers. So
everything was simplified down to this by C creators.
1.32 Unions
C/C++ union is mostly used for interpreting a variable (or memory block) of one data type as a variable
of another data type.
1.32.1 Pseudo-random number generator example

If we need float random numbers between 0 and 1, the simplest thing is to use a PRNG like the Mersenne
twister. It produces random unsigned 32-bit values (in other words, it produces random 32 bits). Then we
can transform this value to float and then divide it by RAND_MAX (0xFFFFFFFF in our case)—we getting a
value in the 0..1 interval.
But as we know, division is slow. Also, we would like to issue as few FPU operations as possible. Can we
get rid of the division?
Let’s recall what a floating point number consists of: sign bit, significand bits and exponent bits. We just
have to store random bits in all significand bits to get a random float number!
The exponent cannot be zero (the floating number is denormalized in this case), so we are storing 0b01111111
to exponent—this means that the exponent is 1. Then we filling the significand with random bits, set the
sign bit to 0 (which means a positive number) and voilà. The generated numbers is to be between 1 and
2, so we must also subtract 1.
A very simple linear congruential random numbers generator is used in my example160 , it produces 32-bit
numbers. The PRNG is initialized with the current time in UNIX timestamp format.
Here we represent the float type as an union—it is the C/C++ construction that enables us to interpret a
piece of memory as different types. In our case, we are able to create a variable of type union and then
access to it as it is float or as it is uint32_t. It can be said, it is just a hack. A dirty one.
The integer PRNG code is the same as we already considered: 1.29 on page 336. So this code in compiled
form is omitted.
#include <stdio.h>
#include <stdint.h>
#include <time.h>
// integer PRNG definitions, data and routines:
// constants from the Numerical Recipes book

const uint32_t RNG_a=1664525;
const uint32_t RNG_c=1013904223;
uint32_t RNG_state; // global variable
void my_srand(uint32_t i)
{
RNG_state=i;
};
uint32_t my_rand()
{
RNG_state=RNG_state*RNG_a+RNG_c;
return RNG_state;
};
160 the idea was taken from: http://go.yurichev.com/17308
374
// FPU PRNG definitions and routines:
union uint32_t_float
{
uint32_t i;
float f;
};
float float_rand()
{
union uint32_t_float tmp;
tmp.i=my_rand() & 0x007fffff | 0x3F800000;
return tmp.f-1;
};
// test
int main()
{
my_srand(time(NULL)); // PRNG initialization
for (int i=0; i<100; i++)

printf ("%f\n", float_rand());
return 0;
};
x86

$SG4238 DB '%f', 0aH, 00H
__real@3ff0000000000000 DQ 03ff0000000000000r ; 1
tv130 = -4
_tmp$ = -4
?float_rand@@YAMXZ PROC
push ecx
call ?my_rand@@YAIXZ
; EAX=pseudorandom value
and eax, 8388607 ; 007fffffH
or eax, 1065353216 ; 3f800000H
; EAX=pseudorandom value & 0x007fffff | 0x3f800000
; store it into local stack:
mov DWORD PTR _tmp$[esp+4], eax
; reload it as float point number:
fld DWORD PTR _tmp$[esp+4]
; subtract 1.0:
fsub QWORD PTR __real@3ff0000000000000
; store value we got into local stack and reload it:
fstp DWORD PTR tv130[esp+4] ; \ these instructions are redundant
fld DWORD PTR tv130[esp+4] ; /
pop ecx
ret 0
?float_rand@@YAMXZ ENDP
_main PROC
push esi
xor eax, eax
call _time
push eax
call ?my_srand@@YAXI@Z
add esp, 4
mov esi, 100
$LL3@main:
call ?float_rand@@YAMXZ
sub esp, 8
375
push OFFSET $SG4238
call _printf
add esp, 12
dec esi
jne SHORT $LL3@main
xor eax, eax
pop esi
ret 0
_main ENDP
Function names are so strange here because this example was compiled as C++ and this is name mangling
in C++, we will talk about it later: 3.21.1 on page 542. If we compile this in MSVC 2012, it uses the SIMD
instructions for the FPU, read more about it here: 1.38.5 on page 436.
ARM (ARM mode)

float_rand
STMFD SP!, {R3,LR}
BL my_rand
; R0=pseudorandom value
FLDS S0, =1.0
; S0=1.0
BIC R3, R0, #0xFF000000
BIC R3, R3, #0x800000
ORR R3, R3, #0x3F800000
; R3=pseudorandom value & 0x007fffff | 0x3f800000
; copy from R3 to FPU (register S15).
; it behaves like bitwise copy, no conversion done:
FMSR S15, R3
; subtract 1.0 and leave result in S0:
FSUBS S0, S15, S0
LDMFD SP!, {R3,PC}
flt_5C DCFS 1.0
main
STMFD SP!, {R4,LR}
MOV R0, #0
BL time
BL my_srand
MOV R4, #0x64 ; 'd'
loc_78
BL float_rand
; S0=pseudorandom value
LDR R0, =aF ; "%f"
; convert float type value into double type value (printf() will need it):
FCVTDS D7, S0
; bitwise copy from D7 into R2/R3 pair of registers (for printf()):
FMRRD R2, R3, D7
BL printf
SUBS R4, R4, #1
BNE loc_78
MOV R0, R4
LDMFD SP!, {R4,PC}
aF DCB "%f",0xA,0
We’ll also make a dump in objdump and we’ll see that the FPU instructions have different names than in
IDA. Apparently, IDA and binutils developers used different manuals? Perhaps it would be good to know
both instruction name variants.

00000038 <float_rand>:
38: e92d4008 push {r3, lr}
3c: ebfffffe bl 10 <my_rand>
376
40: ed9f0a05 vldr s0, [pc, #20] ; 5c <float_rand+0x24>
44: e3c034ff bic r3, r0, #-16777216 ; 0xff000000
48: e3c33502 bic r3, r3, #8388608 ; 0x800000
4c: e38335fe orr r3, r3, #1065353216 ; 0x3f800000
50: ee073a90 vmov s15, r3
54: ee370ac0 vsub.f32 s0, s15, s0
58: e8bd8008 pop {r3, pc}
5c: 3f800000 svccc 0x00800000
00000000 <main>:
0: e92d4010 push {r4, lr}
4: e3a00000 mov r0, #0
8: ebfffffe bl 0 <time>
c: ebfffffe bl 0 <main>
10: e3a04064 mov r4, #100 ; 0x64
14: ebfffffe bl 38 <main+0x38>
18: e59f0018 ldr r0, [pc, #24] ; 38 <main+0x38>
1c: eeb77ac0 vcvt.f64.f32 d7, s0
20: ec532b17 vmov r2, r3, d7
24: ebfffffe bl 0 <printf>
28: e2544001 subs r4, r4, #1
2c: 1afffff8 bne 14 <main+0x14>
30: e1a00004 mov r0, r4
34: e8bd8010 pop {r4, pc}
38: 00000000 andeq r0, r0, r0
The instructions at 0x5c in float_rand() and at 0x38 in main() are (pseudo-)random noise.
1.32.2 Calculating machine epsilon

The machine epsilon is the smallest possible value the FPU can work with. The more bits allocated for
floating point number, the smaller the machine epsilon. It is 2−23 = 1.19e − 07 for float and 2−52 = 2.22e − 16
for double. See also: Wikipedia article.
It’s interesting, how easy it’s to calculate the machine epsilon:
#include <stdio.h>
#include <stdint.h>
union uint_float
{
uint32_t i;
float f;
};
float calculate_machine_epsilon(float start)

{
union uint_float v;
v.f=start;
v.i++;
return v.f-start;
}
void main()
{
printf ("%g\n", calculate_machine_epsilon(1.0));
};
What we do here is just treat the fraction part of the IEEE 754 number as integer and add 1 to it. The
resulting floating number is equal to starting_value+machine_epsilon, so we just have to subtract the starting
value (using floating point arithmetic) to measure, what difference one bit reflects in the single precision
(float). The union serves here as a way to access IEEE 754 number as a regular integer. Adding 1 to it in
fact adds 1 to the fraction part of the number, however, needless to say, overflow is possible, which will
add another 1 to the exponent part.
x86
377
tv130 = 8
_v$ = 8
_start$ = 8
_calculate_machine_epsilon PROC
fld DWORD PTR _start$[esp-4]
fst DWORD PTR _v$[esp-4] ; this instruction is redundant
inc DWORD PTR _v$[esp-4]
fsubr DWORD PTR _v$[esp-4]
fstp DWORD PTR tv130[esp-4] ; \ this instruction pair is also redundant
fld DWORD PTR tv130[esp-4] ; /
ret 0
_calculate_machine_epsilon ENDP
The second FST instruction is redundant: there is no necessity to store the input value in the same place
(the compiler decided to allocate the v variable at the same point in the local stack as the input argument).
Then it is incremented with INC, as it is a normal integer variable. Then it is loaded into the FPU as a 32-bit
IEEE 754 number, FSUBR does the rest of job and the resulting value is stored in ST0. The last FSTP/FLD
instruction pair is redundant, but the compiler didn’t optimize it out.
ARM64
Let’s extend our example to 64-bit:

#include <stdio.h>
#include <stdint.h>
typedef union
{
uint64_t i;
double d;
} uint_double;
double calculate_machine_epsilon(double start)

{
uint_double v;
v.d=start;
v.i++;
return v.d-start;
}
void main()
{
printf ("%g\n", calculate_machine_epsilon(1.0));
};
ARM64 has no instruction that can add a number to a FPU D-register, so the input value (that came in D0)
is first copied into GPR, incremented, copied to FPU register D1, and then subtraction occurs.

calculate_machine_epsilon:
fmov x0, d0 ; load input value of double type into X0
add x0, x0, 1 ; X0++
fmov d1, x0 ; move it to FPU register
fsub d0, d1, d0 ; subtract
ret
See also this example compiled for x64 with SIMD instructions: 1.38.4 on page 435.
MIPS
The new instruction here is MTC1 (“Move To Coprocessor 1”), it just transfers data from GPR to the FPU’s
registers.

calculate_machine_epsilon:
mfc1 $v0, $f12
378
or $at, $zero ; NOP
addiu $v1, $v0, 1
mtc1 $v1, $f2
jr $ra
sub.s $f0, $f2, $f12 ; branch delay slot
Conclusion
It’s hard to say whether someone may need this trickery in real-world code, but as was mentioned many
times in this book, this example serves well for explaining the IEEE 754 format and unions in C/C++.
1.32.3 FSCALE instruction replacement

Agner Fog in his Optimizing subroutines in assembly language / An optimization guide for x86 platforms
work 161 states that FSCALE FPU instruction (calculating 2n ) may be slow on many CPUs, and he offers
faster replacement.
Here is my translation of his assembly code to C/C++:
#include <stdint.h>
#include <stdio.h>
union uint_float
{
uint32_t i;
float f;
};
float flt_2n(int N)
{
union uint_float tmp;
tmp.i=(N<<23)+0x3f800000;
return tmp.f;
};
struct float_as_struct
{
unsigned int fraction : 23;
unsigned int exponent : 8;
unsigned int sign : 1;
};
float flt_2n_v2(int N)
{
struct float_as_struct tmp;
tmp.fraction=0;
tmp.sign=0;
tmp.exponent=N+0x7f;
return *(float*)(&tmp);
};
union uint64_double
{
uint64_t i;
double d;
};
double dbl_2n(int N)
{
union uint64_double tmp;
tmp.i=((uint64_t)N<<52)+0x3ff0000000000000UL;
return tmp.d;
};
161 http://www.agner.org/optimize/optimizing_assembly.pdf
379
struct double_as_struct
{
uint64_t fraction : 52;
int exponent : 11;
int sign : 1;
};
double dbl_2n_v2(int N)
{
struct double_as_struct tmp;
tmp.fraction=0;
tmp.sign=0;
tmp.exponent=N+0x3ff;
return *(double*)(&tmp);
};
int main()
{
// 211 = 2048
printf ("%f\n", flt_2n(11));
printf ("%f\n", flt_2n_v2(11));
printf ("%lf\n", dbl_2n(11));
printf ("%lf\n", dbl_2n_v2(11));
};
FSCALE instruction may be faster in your environment, but still, it’s a good example of union’s and the
fact that exponent is stored in 2n form, so an input n value is shifted to the exponent in IEEE 754 encoded
number. Then exponent is then corrected with addition of 0x3f800000 or 0x3ff0000000000000.
The same can be done without shift using struct, but internally, shift operations still occurred.
1.32.4 Fast square root calculation

Another well-known algorithm where float is interpreted as integer is fast calculation of square root.
Listing 1.365: The source code is taken from Wikipedia: http://go.yurichev.com/17364

/* Assumes that float is in the IEEE 754 single precision floating point format
* and that int is 32 bits. */
float sqrt_approx(float z)
{
int val_int = *(int*)&z; /* Same bits, but as an int */
/*
* To justify the following code, prove that
*
* ((((val_int / 2^m) - b) / 2) + b) * 2^m = ((val_int - 2^m) / 2) + ((b + 1) / 2) * 2^m)
*
* where
*
* b = exponent bias
* m = number of mantissa bits
*
* .
*/
val_int -= 1 << 23; /* Subtract 2^m. */

val_int >>= 1; /* Divide by 2. */
val_int += 1 << 29; /* Add ((b + 1) / 2) * 2^m. */
return *(float*)&val_int; /* Interpret again as float */

}
As an exercise, you can try to compile this function and to understand, how it works.
There is also well-known algorithm of fast calculation of √1 . Algorithm became popular, supposedly,
x
because it was used in Quake III Arena.
Algorithm description can be found in Wikipedia: http://go.yurichev.com/17360.
380
1.33 Pointers to functions
A pointer to a function, as any other pointer, is just the address of the function’s start in its code segment.
They are often used for calling callback functions.
Well-known examples are:
• qsort(), atexit() from the standard C library;
• *NIX OS signals.
• thread starting: CreateThread() (win32), pthread_create() (POSIX);
• lots of win32 functions, like EnumChildWindows().
• lots of places in the Linux kernel, for example the filesystem driver functions are called via callbacks.
• The GCC plugin functions are also called via callbacks.
So, the qsort() function is an implementation of quicksort in the C/C++ standard library. The function is
able to sort anything, any type of data, as long as you have a function to compare these two elements
and qsort() is able to call it.
The comparison function can be defined as:
int (*compare)(const void *, const void *)
Let’s use the following example:

1 /* ex3 Sorting ints with qsort */
2
3 #include <stdio.h>
4 #include <stdlib.h>
5
6 int comp(const void * _a, const void * _b)
7 {
8 const int *a=(const int *)_a;
9 const int *b=(const int *)_b;
10
11 if (*a==*b)
12 return 0;
13 else
14 if (*a < *b)
15 return -1;
16 else
17 return 1;
18 }
19
20 int main(int argc, char* argv[])
21 {
22 int numbers[10]={1892,45,200,-98,4087,5,-12345,1087,88,-100000};
23 int i;
24
25 /* Sort the array */
26 qsort(numbers,10,sizeof(int),comp) ;
27 for (i=0;i<9;i++)
28 printf("Number = %d\n",numbers[ i ]) ;
29 return 0;
30 }
1.33.1 MSVC
Let’s compile it in MSVC 2010 (some parts were omitted for the sake of brevity) with /Ox option:
Listing 1.366: Optimizing MSVC 2010: /GS- /MD

__a$ = 8 ; size = 4
__b$ = 12 ; size = 4
_comp PROC
mov eax, DWORD PTR __a$[esp-4]
mov ecx, DWORD PTR __b$[esp-4]
mov eax, DWORD PTR [eax]
381
mov ecx, DWORD PTR [ecx]
cmp eax, ecx
jne SHORT $LN4@comp
xor eax, eax
ret 0
$LN4@comp:
xor edx, edx
cmp eax, ecx
setge dl
lea eax, DWORD PTR [edx+edx-1]
ret 0
_comp ENDP
_numbers$ = -40 ; size = 40

_argc$ = 8 ; size = 4
_argv$ = 12 ; size = 4
_main PROC
sub esp, 40 ; 00000028H
push esi
push OFFSET _comp
push 4
lea eax, DWORD PTR _numbers$[esp+52]
push 10 ; 0000000aH
push eax
mov DWORD PTR _numbers$[esp+60], 1892 ; 00000764H
mov DWORD PTR _numbers$[esp+64], 45 ; 0000002dH
mov DWORD PTR _numbers$[esp+68], 200 ; 000000c8H
mov DWORD PTR _numbers$[esp+72], -98 ; ffffff9eH
mov DWORD PTR _numbers$[esp+76], 4087 ; 00000ff7H
mov DWORD PTR _numbers$[esp+80], 5
mov DWORD PTR _numbers$[esp+84], -12345 ; ffffcfc7H
mov DWORD PTR _numbers$[esp+88], 1087 ; 0000043fH
mov DWORD PTR _numbers$[esp+92], 88 ; 00000058H
mov DWORD PTR _numbers$[esp+96], -100000 ; fffe7960H
call _qsort
add esp, 16 ; 00000010H
...
Nothing surprising so far. As a fourth argument, the address of label _comp is passed, which is just a place
where comp() is located, or, in other words, the address of the very first instruction of that function.
How does qsort() call it?
Let’s take a look at this function, located in MSVCR80.DLL (a MSVC DLL module with C standard library
functions):
Listing 1.367: MSVCR80.DLL

.text:7816CBF0 ; void __cdecl qsort(void *, unsigned int, unsigned int, int (__cdecl *)(const
void *, const void *))
.text:7816CBF0 public _qsort
.text:7816CBF0 _qsort proc near
.text:7816CBF0
.text:7816CBF0 lo = dword ptr -104h
.text:7816CBF0 hi = dword ptr -100h
.text:7816CBF0 var_FC = dword ptr -0FCh
.text:7816CBF0 stkptr = dword ptr -0F8h
.text:7816CBF0 lostk = dword ptr -0F4h
.text:7816CBF0 histk = dword ptr -7Ch
.text:7816CBF0 base = dword ptr 4
.text:7816CBF0 num = dword ptr 8
.text:7816CBF0 width = dword ptr 0Ch
.text:7816CBF0 comp = dword ptr 10h
.text:7816CBF0
.text:7816CBF0 sub esp, 100h
....
.text:7816CCE0 loc_7816CCE0: ; CODE XREF: _qsort+B1
382
.text:7816CCE0 shr eax, 1
.text:7816CCE2 imul eax, ebp
.text:7816CCE5 add eax, ebx
.text:7816CCE7 mov edi, eax
.text:7816CCE9 push edi
.text:7816CCEA push ebx
.text:7816CCEB call [esp+118h+comp]
.text:7816CCF2 add esp, 8
.text:7816CCF5 test eax, eax
.text:7816CCF7 jle short loc_7816CD04
comp—is the fourth function argument. Here the control gets passed to the address in the comp argument.
Before it, two arguments are prepared for comp(). Its result is checked after its execution.
That’s why it is dangerous to use pointers to functions. First of all, if you call qsort() with an incorrect
function pointer, qsort() may pass control flow to an incorrect point, the process may crash and this bug
will be hard to find.
The second reason is that the callback function types must comply strictly, calling the wrong function with
wrong arguments of wrong types may lead to serious problems, however, the crashing of the process is
not a problem here —the problem is how to determine the reason for the crash —because the compiler
may be silent about the potential problems while compiling.
383
MSVC + OllyDbg
Let’s load our example into OllyDbg and set a breakpoint on comp(). We can see how the values are
compared at the first comp() call:
Figure 1.110: OllyDbg: first call of comp()
OllyDbg shows the compared values in the window under the code window, for convenience. We can also
see that the SP points to RA, where the qsort() function is (located in MSVCR100.DLL).
384
By tracing (F8) until the RETN instruction and pressing F8 one more time, we return to the qsort() function:
Figure 1.111: OllyDbg: the code in qsort() right after comp() call
That has been a call to the comparison function.
385
Here is also a screenshot of the moment of the second call of comp()—now values that have to be com-
pared are different:
Figure 1.112: OllyDbg: second call of comp()
MSVC + tracer
Let’s also see which pairs are compared. These 10 numbers are being sorted: 1892, 45, 200, -98, 4087,
5, -12345, 1087, 88, -100000.
We got the address of the first CMP instruction in comp(), it is 0x0040100C and we’ve set a breakpoint on
it:
tracer.exe -l:17_1.exe bpx=17_1.exe!0x0040100C
Now we get some information about the registers at the breakpoint:

PID=4336|New process 17_1.exe
(0) 17_1.exe!0x40100c
EAX=0x00000764 EBX=0x0051f7c8 ECX=0x00000005 EDX=0x00000000
ESI=0x0051f7d8 EDI=0x0051f7b4 EBP=0x0051f794 ESP=0x0051f67c
EIP=0x0028100c
FLAGS=IF
(0) 17_1.exe!0x40100c
EAX=0x00000005 EBX=0x0051f7c8 ECX=0xfffe7960 EDX=0x00000000
EIP=0x0028100c
FLAGS=PF ZF IF
(0) 17_1.exe!0x40100c
EAX=0x00000764 EBX=0x0051f7c8 ECX=0x00000005 EDX=0x00000000
EIP=0x0028100c
FLAGS=CF PF ZF IF
...
Let’s filter out EAX and ECX and we got:

EAX=0x00000764 ECX=0x00000005
EAX=0x00000005 ECX=0xfffe7960
EAX=0x00000764 ECX=0x00000005
EAX=0x0000002d ECX=0x00000005
EAX=0x00000058 ECX=0x00000005
EAX=0x0000043f ECX=0x00000005
EAX=0xffffcfc7 ECX=0x00000005
EAX=0x000000c8 ECX=0x00000005
386
EAX=0xffffff9e ECX=0x00000005
EAX=0x00000ff7 ECX=0x00000005
EAX=0x00000ff7 ECX=0x00000005
EAX=0xffffcfc7 ECX=0xfffe7960
EAX=0x00000005 ECX=0xffffcfc7
EAX=0xffffff9e ECX=0xffffcfc7
EAX=0x000000c8 ECX=0x00000ff7
EAX=0x0000002d ECX=0x00000ff7
EAX=0x0000043f ECX=0x00000ff7
EAX=0x00000058 ECX=0x00000ff7
EAX=0x00000764 ECX=0x00000ff7
EAX=0x000000c8 ECX=0x00000764
EAX=0x0000002d ECX=0x00000764
EAX=0x0000043f ECX=0x00000764
EAX=0x00000058 ECX=0x00000764
EAX=0x000000c8 ECX=0x00000058
EAX=0x0000002d ECX=0x000000c8
EAX=0x0000043f ECX=0x000000c8
EAX=0x000000c8 ECX=0x00000058
EAX=0x0000002d ECX=0x000000c8
EAX=0x0000002d ECX=0x00000058
That’s 34 pairs. Therefore, the quick sort algorithm needs 34 comparison operations to sort these 10
numbers.
387
MSVC + tracer (code coverage)
We can also use the tracer’s feature to collect all possible register values and show them in IDA.
Let’s trace all instructions in comp():
tracer.exe -l:17_1.exe bpf=17_1.exe!0x00401000,trace:cc
We get an .idc-script for loading into IDA and load it:
Figure 1.113: tracer and IDA. N.B.: some values are cut at right
IDA gave the function a name (PtFuncCompare)—because IDA sees that the pointer to this function is
passed to qsort().
We see that the a and b pointers are pointing to various places in the array, but the step between them is
4, as 32-bit values are stored in the array.
We see that the instructions at 0x401010 and 0x401012 were never executed (so they left as white):
indeed, comp() has never returned 0, because there no equal elements in the array.
1.33.2 GCC
Not a big difference:
Listing 1.368: GCC

lea eax, [esp+40h+var_28]
mov [esp+40h+var_28], 764h
mov [esp+40h+var_24], 2Dh
mov [esp+40h+var_20], 0C8h
mov [esp+40h+var_1C], 0FFFFFF9Eh
mov [esp+40h+var_18], 0FF7h
mov [esp+40h+var_10], 0FFFFCFC7h
mov [esp+40h+var_C], 43Fh
mov [esp+40h+var_8], 58h
mov [esp+40h+var_4], 0FFFE7960h
mov [esp+40h+var_34], offset comp
mov [esp+40h+var_3C], 0Ah
call _qsort
388
comp() function:
public comp
comp proc near
arg_0 = dword ptr 8

push ebp
mov ebp, esp
mov edx, [eax]
xor eax, eax
cmp [ecx], edx
pop ebp
retn
loc_8048458:
setnl al
movzx eax, al
lea eax, [eax+eax-1]
pop ebp
retn
comp endp
162
The implementation of qsort() is located in libc.so.6 and it is in fact just a wrapper for qsort_r().
In turn, it is calling quicksort(), where our defined function is called via a passed pointer:
Listing 1.369: (file libc.so.6, glibc version—2.10.1)

...
.text:0002DDF6 mov edx, [ebp+arg_10]
.text:0002DDF9 mov [esp+4], esi
.text:0002DDFD mov [esp], edi
.text:0002DE00 mov [esp+8], edx
.text:0002DE04 call [ebp+arg_C]
...
GCC + GDB (with source code)
Obviously, we have the C-source code of our example ( 1.33 on page 381), so we can set a breakpoint (b)
on line number (11—the line where the first comparison occurs). We also have to compile the example
with debugging information included (-g), so the table with addresses and corresponding line numbers is
present.
We can also print values using variable names (p): the debugging information also has tells us which
register and/or local stack element contains which variable.
We can also see the stack (bt) and find out that there is some intermediate function msort_with_tmp()
used in Glibc.
Listing 1.370: GDB session

dennis@ubuntuvm:~/polygon$ gcc 17_1.c -g
dennis@ubuntuvm:~/polygon$ gdb ./a.out
Copyright (C) 2013 Free Software Foundation, Inc.
...
Reading symbols from /home/dennis/polygon/a.out...done.
(gdb) b 17_1.c:11
Breakpoint 1 at 0x804845f: file 17_1.c, line 11.
(gdb) run
Starting program: /home/dennis/polygon/./a.out
Breakpoint 1, comp (_a=0xbffff0f8, _b=_b@entry=0xbffff0fc) at 17_1.c:11

11 if (*a==*b)
(gdb) p *a
162 a concept like thunk function
389
$1 = 1892
(gdb) p *b
$2 = 45
(gdb) c
Continuing.
Breakpoint 1, comp (_a=0xbffff104, _b=_b@entry=0xbffff108) at 17_1.c:11

11 if (*a==*b)
(gdb) p *a
$3 = -98
(gdb) p *b
$4 = 4087
(gdb) bt
#0 comp (_a=0xbffff0f8, _b=_b@entry=0xbffff0fc) at 17_1.c:11
#1 0xb7e42872 in msort_with_tmp (p=p@entry=0xbffff07c, b=b@entry=0xbffff0f8, n=n@entry=2)
at msort.c:65
#2 0xb7e4273e in msort_with_tmp (n=2, b=0xbffff0f8, p=0xbffff07c) at msort.c:45
#3 msort_with_tmp (p=p@entry=0xbffff07c, b=b@entry=0xbffff0f8, n=n@entry=5) at msort.c:53
#6 0xb7e42cef in msort_with_tmp (n=10, b=0xbffff0f8, p=0xbffff07c) at msort.c:45
#7 __GI_qsort_r (b=b@entry=0xbffff0f8, n=n@entry=10, s=s@entry=4, cmp=cmp@entry=0x804844d <⤦
Ç comp>,
arg=arg@entry=0x0) at msort.c:297
#8 0xb7e42dcf in __GI_qsort (b=0xbffff0f8, n=10, s=4, cmp=0x804844d <comp>) at msort.c:307
#9 0x0804850d in main (argc=1, argv=0xbffff1c4) at 17_1.c:26
(gdb)
GCC + GDB (no source code)
But often there is no source code at all, so we can disassemble the comp() function (disas), find the very
first CMP instruction and set a breakpoint (b) at that address.
At each breakpoint, we are going to dump all register contents
(info registers). The stack information is also available (bt),
but partially: there is no line number information for comp().
Listing 1.371: GDB session

dennis@ubuntuvm:~/polygon$ gcc 17_1.c
dennis@ubuntuvm:~/polygon$ gdb ./a.out
Copyright (C) 2013 Free Software Foundation, Inc.
...
Reading symbols from /home/dennis/polygon/a.out...(no debugging symbols found)...done.
(gdb) disas comp
Dump of assembler code for function comp:
0x0804844d <+0>: push ebp
0x0804844e <+1>: mov ebp,esp
0x08048450 <+3>: sub esp,0x10
0x08048453 <+6>: mov eax,DWORD PTR [ebp+0x8]
0x08048456 <+9>: mov DWORD PTR [ebp-0x8],eax
0x08048459 <+12>: mov eax,DWORD PTR [ebp+0xc]
0x0804845c <+15>: mov DWORD PTR [ebp-0x4],eax
0x0804845f <+18>: mov eax,DWORD PTR [ebp-0x8]
0x08048462 <+21>: mov edx,DWORD PTR [eax]
0x08048464 <+23>: mov eax,DWORD PTR [ebp-0x4]
0x08048467 <+26>: mov eax,DWORD PTR [eax]
0x08048469 <+28>: cmp edx,eax
0x0804846b <+30>: jne 0x8048474 <comp+39>
0x0804846d <+32>: mov eax,0x0
0x08048472 <+37>: jmp 0x804848e <comp+65>
0x08048477 <+42>: mov edx,DWORD PTR [eax]
0x0804847c <+47>: mov eax,DWORD PTR [eax]
0x0804847e <+49>: cmp edx,eax
0x08048480 <+51>: jge 0x8048489 <comp+60>
390
0x08048482 <+53>: mov eax,0xffffffff
0x08048487 <+58>: jmp 0x804848e <comp+65>
0x08048489 <+60>: mov eax,0x1
0x0804848e <+65>: leave
0x0804848f <+66>: ret
(gdb) b *0x08048469
(gdb) run
Starting program: /home/dennis/polygon/./a.out
Breakpoint 1, 0x08048469 in comp ()

eax 0x2d 45
ecx 0xbffff0f8 -1073745672
edx 0x764 1892
ebx 0xb7fc0000 -1208221696
esp 0xbfffeeb8 0xbfffeeb8
ebp 0xbfffeec8 0xbfffeec8
esi 0xbffff0fc -1073745668
edi 0xbffff010 -1073745904
eip 0x8048469 0x8048469 <comp+28>
eflags 0x286 [ PF SF IF ]
cs 0x73 115
ss 0x7b 123
ds 0x7b 123
es 0x7b 123
fs 0x0 0
gs 0x33 51
(gdb) c
Continuing.

eax 0xff7 4087
ecx 0xbffff104 -1073745660
edx 0xffffff9e -98
ebx 0xb7fc0000 -1208221696
esp 0xbfffee58 0xbfffee58
ebp 0xbfffee68 0xbfffee68
esi 0xbffff108 -1073745656
edi 0xbffff010 -1073745904
eip 0x8048469 0x8048469 <comp+28>
eflags 0x282 [ SF IF ]
cs 0x73 115
ss 0x7b 123
ds 0x7b 123
es 0x7b 123
fs 0x0 0
gs 0x33 51
(gdb) c
Continuing.

eax 0xffffff9e -98
ecx 0xbffff100 -1073745664
edx 0xc8 200
ebx 0xb7fc0000 -1208221696
esp 0xbfffeeb8 0xbfffeeb8
ebp 0xbfffeec8 0xbfffeec8
esi 0xbffff104 -1073745660
edi 0xbffff010 -1073745904
eip 0x8048469 0x8048469 <comp+28>
eflags 0x286 [ PF SF IF ]
cs 0x73 115
ss 0x7b 123
ds 0x7b 123
es 0x7b 123
fs 0x0 0
391
gs 0x33 51
(gdb) bt
#0 0x08048469 in comp ()
#1 0xb7e42872 in msort_with_tmp (p=p@entry=0xbffff07c, b=b@entry=0xbffff0f8, n=n@entry=2)
at msort.c:65
#6 0xb7e42cef in msort_with_tmp (n=10, b=0xbffff0f8, p=0xbffff07c) at msort.c:45
#7 __GI_qsort_r (b=b@entry=0xbffff0f8, n=n@entry=10, s=s@entry=4, cmp=cmp@entry=0x804844d <⤦
Ç comp>,
arg=arg@entry=0x0) at msort.c:297
#8 0xb7e42dcf in __GI_qsort (b=0xbffff0f8, n=10, s=4, cmp=0x804844d <comp>) at msort.c:307
#9 0x0804850d in main ()
1.33.3 Danger of pointers to functions

As we can see, qsort() function expects a pointer to function which takes two void* arguments and
returning integer. If you have several comparison functions in your code (one compares string, another—
integers, etc), it’s very easy to mix them up with each other. You could try to sort array of string using
function which compares integers, and compiler will not warn you about bug.
1.34 64-bit values in 32-bit environment

In a 32-bit environment, GPR’s are 32-bit, so 64-bit values are stored and passed as 32-bit value pairs 163 .
1.34.1 Returning of 64-bit value

#include <stdint.h>
uint64_t f ()
{
return 0x1234567890ABCDEF;
};
x86
In a 32-bit environment, 64-bit values are returned from functions in the EDX:EAX register pair.
_f PROC
mov eax, -1867788817 ; 90abcdefH
mov edx, 305419896 ; 12345678H
ret 0
_f ENDP
ARM
A 64-bit value is returned in the R0-R1 register pair (R1 is for the high part and R0 for the low part):
||f|| PROC
LDR r0,|L0.12|
LDR r1,|L0.16|
BX lr
ENDP
|L0.12|
DCD 0x90abcdef
|L0.16|
DCD 0x12345678
163 By the way, 32-bit values are passed as pairs in 16-bit environment in the same way: 3.34.4 on page 645
392
MIPS
A 64-bit value is returned in the V0-V1 ($2-$3) register pair (V0 ($2) is for the high part and V1 ($3) for the
low part):
Listing 1.374: Optimizing GCC 4.4.5 (assembly listing)

li $3,-1867841536 # 0xffffffff90ab0000
li $2,305397760 # 0x12340000
ori $3,$3,0xcdef
j $31
ori $2,$2,0x5678

lui $v1, 0x90AB
lui $v0, 0x1234
li $v1, 0x90ABCDEF
jr $ra
li $v0, 0x12345678
1.34.2 Arguments passing, addition, subtraction

#include <stdint.h>
uint64_t f_add (uint64_t a, uint64_t b)

{
return a+b;
};
void f_add_test ()
{
#ifdef __GNUC__
printf ("%lld\n", f_add(12345678901234, 23456789012345));
#else
printf ("%I64d\n", f_add(12345678901234, 23456789012345));
#endif
};
uint64_t f_sub (uint64_t a, uint64_t b)

{
return a-b;
};
x86

_a$ = 8 ; size = 8
_b$ = 16 ; size = 8
_f_add PROC
add eax, DWORD PTR _b$[esp-4]
mov edx, DWORD PTR _a$[esp]
adc edx, DWORD PTR _b$[esp]
ret 0
_f_add ENDP
_f_add_test PROC
push 5461 ; 00001555H
push 1972608889 ; 75939f79H
push 2874 ; 00000b3aH
push 1942892530 ; 73ce2ff2H
call _f_add
push edx
push eax
push OFFSET $SG1436 ; '%I64d', 0aH, 00H
call _printf
393
add esp, 28
ret 0
_f_add_test ENDP
_f_sub PROC
sub eax, DWORD PTR _b$[esp-4]
sbb edx, DWORD PTR _b$[esp]
ret 0
_f_sub ENDP
We can see in the f_add_test() function that each 64-bit value is passed using two 32-bit values, high
part first, then low part.
Addition and subtraction occur in pairs as well.
In addition, the low 32-bit part are added first. If carry has been occurred while adding, the CF flag is set.
The following ADC instruction adds the high parts of the values, and also adds 1 if CF = 1.
Subtraction also occurs in pairs. The first SUB may also turn on the CF flag, which is to be checked in the
subsequent SBB instruction: if the carry flag is on, then 1 is also to be subtracted from the result.
It is easy to see how the f_add() function result is then passed to printf().
Listing 1.377: GCC 4.8.1 -O1 -fno-inline

_f_add:
add eax, DWORD PTR [esp+4]
adc edx, DWORD PTR [esp+8]
ret
_f_add_test:
sub esp, 28
mov DWORD PTR [esp+8], 1972608889 ; 75939f79H
mov DWORD PTR [esp+12], 5461 ; 00001555H
mov DWORD PTR [esp], 1942892530 ; 73ce2ff2H
mov DWORD PTR [esp+4], 2874 ; 00000b3aH
call _f_add
mov DWORD PTR [esp+4], eax
mov DWORD PTR [esp+8], edx
mov DWORD PTR [esp], OFFSET FLAT:LC0 ; "%lld\n"
call _printf
add esp, 28
ret
_f_sub:
sub eax, DWORD PTR [esp+12]
sbb edx, DWORD PTR [esp+16]
ret
GCC code is the same.
ARM

f_add PROC
ADDS r0,r0,r2
ADC r1,r1,r3
BX lr
ENDP
f_sub PROC
SUBS r0,r0,r2
SBC r1,r1,r3
394
BX lr
ENDP
f_add_test PROC
PUSH {r4,lr}
LDR r2,|L0.68| ; 0x75939f79
LDR r3,|L0.72| ; 0x00001555
LDR r0,|L0.76| ; 0x73ce2ff2
LDR r1,|L0.80| ; 0x00000b3a
BL f_add
POP {r4,lr}
MOV r2,r0
MOV r3,r1
ADR r0,|L0.84| ; "%I64d\n"
B __2printf
ENDP
|L0.68|
DCD 0x75939f79
|L0.72|
DCD 0x00001555
|L0.76|
DCD 0x73ce2ff2
|L0.80|
DCD 0x00000b3a
|L0.84|
DCB "%I64d\n",0
The first 64-bit value is passed in R0 and R1 register pair, the second in R2 and R3 register pair. ARM
has the ADC instruction as well (which counts carry flag) and SBC (“subtract with carry”). Important thing:
when the low parts are added/subtracted, ADDS and SUBS instructions with -S suffix are used. The -S suffix
stands for “set flags”, and flags (esp. carry flag) is what consequent ADC/SBC instructions definitely need.
Otherwise, instructions without the -S suffix would do the job (ADD and SUB).
MIPS

f_add:
; $a0 - high part of a
; $a1 - low part of a
; $a2 - high part of b
; $a3 - low part of b
addu $v1, $a3, $a1 ; sum up low parts
addu $a0, $a2, $a0 ; sum up high parts
; will carry generated while summing up low parts?
; if yes, set $v0 to 1
sltu $v0, $v1, $a3
jr $ra
; add 1 to high part of result if carry should be generated:
addu $v0, $a0 ; branch delay slot
; $v0 - high part of result
; $v1 - low part of result
f_sub:
; $a0 - high part of a
; $a1 - low part of a
; $a2 - high part of b
; $a3 - low part of b
subu $v1, $a1, $a3 ; subtract low parts
subu $v0, $a0, $a2 ; subtract high parts
; will carry generated while subtracting low parts?
; if yes, set $a0 to 1
sltu $a1, $v1
jr $ra
; subtract 1 from high part of result if carry should be generated:
; $v0 - high part of result
; $v1 - low part of result
395
f_add_test:
var_10 = -0x10
var_4 = -4

addiu $sp, -0x20
lui $a1, 0x73CE
lui $a3, 0x7593
li $a0, 0xB3A
li $a3, 0x75939F79
li $a2, 0x1555
jal f_add
li $a1, 0x73CE2FF2
lw $gp, 0x20+var_10($sp)
lui $a0, ($LC0 >> 16) # "%lld\n"
lw $t9, (printf & 0xFFFF)($gp)
la $a0, ($LC0 & 0xFFFF) # "%lld\n"
move $a3, $v1
move $a2, $v0
jr $t9
addiu $sp, 0x20
$LC0: .ascii "%lld\n"<0>
MIPS has no flags register, so there is no such information present after the execution of arithmetic op-
erations. So there are no instructions like x86’s ADC and SBB. To know if the carry flag would be set, a
comparison (using SLTU instruction) also occurs, which sets the destination register to 1 or 0. This 1 or 0
is then added or subtracted to/from the final result.
1.34.3 Multiplication, division

#include <stdint.h>
uint64_t f_mul (uint64_t a, uint64_t b)

{
return a*b;
};
uint64_t f_div (uint64_t a, uint64_t b)

{
return a/b;
};
uint64_t f_rem (uint64_t a, uint64_t b)

{
return a % b;
};
x86

_a$ = 8 ; size = 8
_b$ = 16 ; size = 8
_f_mul PROC
push ebp
mov ebp, esp
mov eax, DWORD PTR _b$[ebp+4]
push eax
push ecx
mov edx, DWORD PTR _a$[ebp+4]
396
push edx
push eax
call __allmul ; long long multiplication
pop ebp
ret 0
_f_mul ENDP
_a$ = 8 ; size = 8
_b$ = 16 ; size = 8
_f_div PROC
push ebp
mov ebp, esp
push eax
push ecx
push edx
push eax
call __aulldiv ; unsigned long long division
pop ebp
ret 0
_f_div ENDP
_a$ = 8 ; size = 8
_b$ = 16 ; size = 8
_f_rem PROC
push ebp
mov ebp, esp
push eax
push ecx
push edx
push eax
call __aullrem ; unsigned long long remainder
pop ebp
ret 0
_f_rem ENDP
Multiplication and division are more complex operations, so usually the compiler embeds calls to a library
functions doing that.
These functions are described here: .5 on page 1029.
Listing 1.381: Optimizing GCC 4.8.1 -fno-inline

_f_mul:
push ebx
mov ebx, DWORD PTR [esp+12]
mov ecx, DWORD PTR [esp+20]
imul ebx, eax
imul ecx, edx
mul edx
add ecx, ebx
add edx, ecx
pop ebx
ret
_f_div:
sub esp, 28
397
mov DWORD PTR [esp], eax
call ___udivdi3 ; unsigned division
add esp, 28
ret
_f_rem:
sub esp, 28
call ___umoddi3 ; unsigned modulo
add esp, 28
ret
GCC does the expected, but the multiplication code is inlined right in the function, thinking it could be
more efficient. GCC has different library function names: .4 on page 1029.
ARM
Keil for Thumb mode inserts library subroutine calls:

||f_mul|| PROC
PUSH {r4,lr}
BL __aeabi_lmul
POP {r4,pc}
ENDP
||f_div|| PROC
PUSH {r4,lr}
BL __aeabi_uldivmod
POP {r4,pc}
ENDP
||f_rem|| PROC
PUSH {r4,lr}
BL __aeabi_uldivmod
MOVS r0,r2
MOVS r1,r3
POP {r4,pc}
ENDP
Keil for ARM mode, on the other hand, is able to produce 64-bit multiplication code:

||f_mul|| PROC
PUSH {r4,lr}
UMULL r12,r4,r0,r2
MLA r1,r2,r1,r4
MLA r1,r0,r3,r1
MOV r0,r12
POP {r4,pc}
ENDP
||f_div|| PROC
PUSH {r4,lr}
BL __aeabi_uldivmod
POP {r4,pc}
ENDP
398
||f_rem|| PROC
PUSH {r4,lr}
BL __aeabi_uldivmod
MOV r0,r2
MOV r1,r3
POP {r4,pc}
ENDP
MIPS
Optimizing GCC for MIPS can generate 64-bit multiplication code, but has to call a library routine for 64-bit
division:

f_mul:
mult $a2, $a1
mflo $v0
or $at, $zero ; NOP
or $at, $zero ; NOP
mult $a0, $a3
mflo $a0
addu $v0, $a0
or $at, $zero ; NOP
multu $a3, $a1
mfhi $a2
mflo $v1
jr $ra
addu $v0, $a2
f_div:
var_10 = -0x10
var_4 = -4

addiu $sp, -0x20
lw $t9, (__udivdi3 & 0xFFFF)($gp)
or $at, $zero
jalr $t9
or $at, $zero
or $at, $zero
jr $ra
addiu $sp, 0x20
f_rem:
var_10 = -0x10
var_4 = -4

addiu $sp, -0x20
lw $t9, (__umoddi3 & 0xFFFF)($gp)
or $at, $zero
jalr $t9
or $at, $zero
or $at, $zero
jr $ra
addiu $sp, 0x20
399
There are a lot of NOPs, probably delay slots filled after the multiplication instruction (it’s slower than
other instructions, after all).
1.34.4 Shifting right

#include <stdint.h>
uint64_t f (uint64_t a)
{
return a>>7;
};
x86

_a$ = 8 ; size = 8
_f PROC
shrd eax, edx, 7
shr edx, 7
ret 0
_f ENDP
Listing 1.386: Optimizing GCC 4.8.1 -fno-inline

_f:
shrd eax, edx, 7
shr edx, 7
ret
Shifting also occurs in two passes: first the lower part is shifted, then the higher part. But the lower part
is shifted with the help of the SHRD instruction, it shifts the value of EAX by 7 bits, but pulls new bits from
EDX, i.e., from the higher part. In other words, 64-bit value from EDX:EAX register’s pair, as a whole, is
shifted by 7 bits and lowest 32 bits of result are placed into EAX. The higher part is shifted using the much
more popular SHR instruction: indeed, the freed bits in the higher part must be filled with zeros.
ARM
ARM doesn’t have such instruction as SHRD in x86, so the Keil compiler ought to do this using simple shifts
and OR operations:

||f|| PROC
LSR r0,r0,#7
ORR r0,r0,r1,LSL #25
LSR r1,r1,#7
BX lr
ENDP

||f|| PROC
LSLS r2,r1,#25
LSRS r0,r0,#7
ORRS r0,r0,r2
LSRS r1,r1,#7
BX lr
ENDP
400
MIPS
GCC for MIPS follows the same algorithm as Keil does for Thumb mode:

f:
sll $v0, $a0, 25
srl $v1, $a1, 7
or $v1, $v0, $v1
jr $ra
srl $v0, $a0, 7
1.34.5 Converting 32-bit value into 64-bit one

#include <stdint.h>
int64_t f (int32_t a)
{
return a;
};
x86

_a$ = 8
_f PROC
cdq
ret 0
_f ENDP
Here we also run into necessity to extend a 32-bit signed value into a 64-bit signed one. Unsigned values
are converted straightforwardly: all bits in the higher part must be set to 0. But this is not appropriate for
signed data types: the sign has to be copied into the higher part of the resulting number.
The CDQ instruction does that here, it takes its input value in EAX, extends it to 64-bit and leaves it in the
EDX:EAX register pair. In other words, CDQ gets the number sign from EAX (by getting the most significant
bit in EAX), and depending of it, sets all 32 bits in EDX to 0 or 1. Its operation is somewhat similar to the
MOVSX instruction.
ARM

||f|| PROC
ASR r1,r0,#31
BX lr
ENDP
Keil for ARM is different: it just arithmetically shifts right the input value by 31 bits. As we know, the sign
bit is MSB, and the arithmetical shift copies the sign bit into the “emerged” bits. So after “ASR r1,r0,#31”,
R1 containing 0xFFFFFFFF if the input value has been negative and 0 otherwise. R1 contains the high part
of the resulting 64-bit value. In other words, this code just copies the MSB (sign bit) from the input value
in R0 to all bits of the high 32-bit part of the resulting 64-bit value.
MIPS
GCC for MIPS does the same as Keil did for ARM mode:

f:
sra $v0, $a0, 31
jr $ra
move $v1, $a0
401
1.35 LARGE_INTEGER structure case
Imagine this: late 1980s, you’re Microsoft, and you’re developing a new serious OS (Windows NT), that
will compete with Unices. Target platforms has both 32-bit and 64-bit CPUs. And you need a 64-bit integer
datatype for all sort of purposes, starting at FILETIME164 structure.
The problem: not all target C/C++ compilers support 64-bit integer yet (this is late 1980s). Surely, this
will be changed in (near) future, but not now. What would you do?
While reading this, try to stop (and/or close this book) and think, how can you solve this problem.
164 https://docs.microsoft.com/en-us/windows/desktop/api/minwinbase/ns-minwinbase-filetime
402
This is what Microsoft did, something like this165 :
union ULARGE_INTEGER
{
struct backward_compatibility
{
DWORD LowPart;
DWORD HighPart;
};
#ifdef NEW_FANCY_COMPILER_SUPPORTING_64_BIT
ULONGLONG QuadPart;
#endif
};
This is a chunk of 8 bytes, which can be accessed via 64-bit integer QuadPart (if compiled using newer
compiler), or using two 32-bit integers (if compiled using old one).
QuadPart field is just absent here when compiled using old compiler.
Order is crucial: first field (LowPart) maps to lower 4 bytes of 64-bit value, second field (HighPart) maps
to higher 4 bytes.
Microsoft also added utility functions for all the arithmetical operation, in a same manner as I already
described: 1.34 on page 392.
And this is from the leaked Windows 2000 source code base:
Listing 1.393: i386 arch

;++
;
; LARGE_INTEGER
; RtlLargeIntegerAdd (
; IN LARGE_INTEGER Addend1,
; IN LARGE_INTEGER Addend2
; )
;
; Routine Description:
;
; This function adds a signed large integer to a signed large integer and
; returns the signed large integer result.
;
; Arguments:
;
; (TOS+4) = Addend1 - first addend value
; (TOS+12) = Addend2 - second addend value
;
; Return Value:
;
; The large integer result is stored in (edx:eax)
;
;--
cPublicProc _RtlLargeIntegerAdd ,4
cPublicFpo 4,0
mov eax,[esp]+4 ; (eax)=add1.low

add eax,[esp]+12 ; (eax)=sum.low
mov edx,[esp]+8 ; (edx)=add1.hi
adc edx,[esp]+16 ; (edx)=sum.hi
stdRET _RtlLargeIntegerAdd
stdENDP _RtlLargeIntegerAdd
Listing 1.394: MIPS arch

LEAF_ENTRY(RtlLargeIntegerAdd)
lw t0,4 * 4(sp) // get low part of addend2 value

lw t1,4 * 5(sp) // get high part of addend2 value
165 Not a copypasted source code, I wrote this
403
addu t0,t0,a2 // add low parts of large integer
addu t1,t1,a3 // add high parts of large integer
sltu t2,t0,a2 // generate carry from low part
addu t1,t1,t2 // add carry to high part
sw t0,0(a0) // store low part of result
sw t1,4(a0) // store high part of result
move v0,a0 // set function return register
j ra // return
.end RtlLargeIntegerAdd
Now two 64-bit architectures:
Listing 1.395: Itanium arch

add v0 = a0, a1 // add both quadword arguments

LEAF_RETURN
LEAF_EXIT(RtlLargeIntegerAdd)
Listing 1.396: DEC Alpha arch

addq a0, a1, v0 // add both quadword arguments

ret zero, (ra) // return
.end RtlLargeIntegerAdd
No need using 32-bit instructions on Itanium and DEC Alpha—64-bit ones are here already.
And this is what we can find in Windows Research Kernel:
DECLSPEC_DEPRECATED_DDK // Use native __int64 math
__inline
LARGE_INTEGER
NTAPI
RtlLargeIntegerAdd (
LARGE_INTEGER Addend1,
LARGE_INTEGER Addend2
)
{
LARGE_INTEGER Sum;
Sum.QuadPart = Addend1.QuadPart + Addend2.QuadPart;

return Sum;
}
All these functions can be dropped (in future), but now they just operate on QuadPart field. If this piece
of code is to be compiled using a modern 32-bit compiler (that supports 64-bit integer), it will generate
two 32-bit additions under the hood. From this moment, LowPart/HighPart fields can be dropped from the
LARGE_INTEGER union/structure.
Would you use such a technique today? Probably not, but if someone would need 128-bit integer data
type, you can implement it just like this.
Also, needless to say, this works thanks to little-endian ( 2.8 on page 460) (all architectures Windows NT
was developed for are little-endian). This trick wouldn’t be possible on a big-endian architecture.
1.36 SIMD
SIMD is an acronym: Single Instruction, Multiple Data.
As its name implies, it processes multiple data using only one instruction.
Like the FPU, that CPU subsystem looks like a separate processor inside x86.
SIMD began as MMX in x86. 8 new 64-bit registers appeared: MM0-MM7.
404
Each MMX register can hold 2 32-bit values, 4 16-bit values or 8 bytes. For example, it is possible to add
8 8-bit values (bytes) simultaneously by adding two values in MMX registers.
One simple example is a graphics editor that represents an image as a two dimensional array. When the
user changes the brightness of the image, the editor must add or subtract a coefficient to/from each pixel
value. For the sake of brevity if we say that the image is grayscale and each pixel is defined by one 8-bit
byte, then it is possible to change the brightness of 8 pixels simultaneously.
By the way, this is the reason why the saturation instructions are present in SIMD.
When the user changes the brightness in the graphics editor, overflow and underflow are not desirable,
so there are addition instructions in SIMD which are not adding anything if the maximum value is reached,
etc.
When MMX appeared, these registers were actually located in the FPU’s registers. It was possible to use
either FPU or MMX at the same time. One might think that Intel saved on transistors, but in fact the reason
of such symbiosis was simpler —older OSes that are not aware of the additional CPU registers would not
save them at the context switch, but saving the FPU registers. Thus, MMX-enabled CPU + old OS + process
utilizing MMX features will still work.
SSE—is extension of the SIMD registers to 128 bits, now separate from the FPU.
AVX—another extension, to 256 bits.
Now about practical usage.
Of course, this is memory copy routines (memcpy), memory comparing (memcmp) and so on.
One more example: the DES encryption algorithm takes a 64-bit block and a 56-bit key, encrypt the block
and produces a 64-bit result. The DES algorithm may be considered as a very large electronic circuit, with
wires and AND/OR/NOT gates.
Bitslice DES166 —is the idea of processing groups of blocks and keys simultaneously. Let’s say, variable
of type unsigned int on x86 can hold up to 32 bits, so it is possible to store there intermediate results for
32 block-key pairs simultaneously, using 64+56 variables of type unsigned int.
There is an utility to brute-force Oracle RDBMS passwords/hashes (ones based on DES), using slightly
modified bitslice DES algorithm for SSE2 and AVX—now it is possible to encrypt 128 or 256 block-keys
pairs simultaneously.
http://go.yurichev.com/17313
1.36.1 Vectorization
Vectorization167 is when, for example, you have a loop taking couple of arrays for input and producing
one array. The loop body takes values from the input arrays, does something and puts the result into the
output array. Vectorization is to process several elements simultaneously.
Vectorization is not very fresh technology: the author of this textbook saw it at least on the Cray Y-MP
supercomputer line from 1988 when he played with its “lite” version Cray Y-MP EL 168 .
For example:
for (i = 0; i < 1024; i++)
{
C[i] = A[i]*B[i];
}
This fragment of code takes elements from A and B, multiplies them and saves the result into C.
If each array element we have is 32-bit int, then it is possible to load 4 elements from A into a 128-bit
XMM-register, from B to another XMM-registers, and by executing PMULLD (Multiply Packed Signed Dword
Integers and Store Low Result) and PMULHW (Multiply Packed Signed Integers and Store High Result), it is
possible to get 4 64-bit products at once.
Thus, loop body execution count is 1024/4 instead of 1024, that is 4 times less and, of course, faster.
167 Wikipedia: vectorization
168 Remotely. It is installed in the museum of supercomputers: http://go.yurichev.com/17081
405
Addition example
Some compilers can do vectorization automatically in simple cases, e.g., Intel C++169 .
Here is tiny function:
int f (int sz, int *ar1, int *ar2, int *ar3)
{
for (int i=0; i<sz; i++)
ar3[i]=ar1[i]+ar2[i];
return 0;
};
Intel C++
Let’s compile it with Intel C++ 11.1.051 win32:

icl intel.cpp /QaxSSE2 /Faintel.asm /Ox
We got (in IDA):
; int __cdecl f(int, int *, int *, int *)
public ?f@@YAHHPAH00@Z
?f@@YAHHPAH00@Z proc near

sz = dword ptr 4
ar1 = dword ptr 8
ar2 = dword ptr 0Ch
ar3 = dword ptr 10h
push edi
push esi
push ebx
push esi
mov edx, [esp+10h+sz]
test edx, edx
jle loc_15B
mov eax, [esp+10h+ar3]
cmp edx, 6
jle loc_143
cmp eax, [esp+10h+ar2]
jbe short loc_36
mov esi, [esp+10h+ar2]
sub esi, eax
lea ecx, ds:0[edx*4]
neg esi
cmp ecx, esi
jbe short loc_55
loc_36: ; CODE XREF: f(int,int *,int *,int *)+21

jnb loc_143
sub esi, eax
lea ecx, ds:0[edx*4]
cmp esi, ecx
jb loc_143

jbe short loc_67
sub esi, eax
neg esi
cmp ecx, esi
jbe short loc_7F
169 More about Intel C++ automatic vectorization: Excerpt: Effective Automatic Vectorization
406
jnb loc_143
sub esi, eax
cmp esi, ecx
jb loc_143
loc_7F: ; CODE XREF: f(int,int *,int *,int *)+65

mov edi, eax ; edi = ar3
and edi, 0Fh ; is ar3 16-byte aligned?
jz short loc_9A ; yes
test edi, 3
jnz loc_162
neg edi
add edi, 10h
shr edi, 2
loc_9A: ; CODE XREF: f(int,int *,int *,int *)+84

lea ecx, [edi+4]
cmp edx, ecx
jl loc_162
mov ecx, edx
sub ecx, edi
and ecx, 3
neg ecx
add ecx, edx
test edi, edi
jbe short loc_D6
mov ebx, [esp+10h+ar2]
mov [esp+10h+var_10], ecx
mov ecx, [esp+10h+ar1]
xor esi, esi
loc_C1: ; CODE XREF: f(int,int *,int *,int *)+CD

mov edx, [ecx+esi*4]
add edx, [ebx+esi*4]
mov [eax+esi*4], edx
inc esi
cmp esi, edi
jb short loc_C1
mov ecx, [esp+10h+var_10]
mov edx, [esp+10h+sz]
loc_D6: ; CODE XREF: f(int,int *,int *,int *)+B2

lea esi, [esi+edi*4] ; is ar2+i*4 16-byte aligned?
test esi, 0Fh
jz short loc_109 ; yes!
loc_ED: ; CODE XREF: f(int,int *,int *,int *)+105

movdqu xmm1, xmmword ptr [ebx+edi*4] ; ar1+i*4
movdqu xmm0, xmmword ptr [esi+edi*4] ; ar2+i*4 is not 16-byte aligned, so load it to
XMM0
paddd xmm1, xmm0
movdqa xmmword ptr [eax+edi*4], xmm1 ; ar3+i*4
add edi, 4
cmp edi, ecx
jb short loc_ED
jmp short loc_127
loc_109: ; CODE XREF: f(int,int *,int *,int *)+E3


movdqu xmm0, xmmword ptr [ebx+edi*4]
407
paddd xmm0, xmmword ptr [esi+edi*4]
movdqa xmmword ptr [eax+edi*4], xmm0
add edi, 4
cmp edi, ecx
jb short loc_111

; f(int,int *,int *,int *)+164
cmp ecx, edx
jnb short loc_15B
mov edi, [esp+10h+ar2]
loc_133: ; CODE XREF: f(int,int *,int *,int *)+13F

mov ebx, [esi+ecx*4]
add ebx, [edi+ecx*4]
mov [eax+ecx*4], ebx
inc ecx
cmp ecx, edx
jb short loc_133
jmp short loc_15B

; f(int,int *,int *,int *)+3A ...
mov edi, [esp+10h+ar2]
xor ecx, ecx
loc_14D: ; CODE XREF: f(int,int *,int *,int *)+159

mov ebx, [esi+ecx*4]
add ebx, [edi+ecx*4]
mov [eax+ecx*4], ebx
inc ecx
cmp ecx, edx
jb short loc_14D
loc_15B: ; CODE XREF: f(int,int *,int *,int *)+A

; f(int,int *,int *,int *)+129 ...
xor eax, eax
pop ecx
pop ebx
pop esi
pop edi
retn
loc_162: ; CODE XREF: f(int,int *,int *,int *)+8C

; f(int,int *,int *,int *)+9F
xor ecx, ecx
jmp short loc_127
?f@@YAHHPAH00@Z endp
The SSE2-related instructions are:

• MOVDQU (Move Unaligned Double Quadword)—just loads 16 bytes from memory into a XMM-register.
• PADDD (Add Packed Integers)—adds 4 pairs of 32-bit numbers and leaves the result in the first operand.
By the way, no exception is raised in case of overflow and no flags are to be set, just the low 32 bits of
the result are to be stored. If one of PADDD’s operands is the address of a value in memory, then the
address must be aligned on a 16-byte boundary. If it is not aligned, an exception will be triggered.
• MOVDQA (Move Aligned Double Quadword) is the same as MOVDQU, but requires the address of the
value in memory to be aligned on a 16-bit boundary. If it is not aligned, exception will be raised.
MOVDQA works faster than MOVDQU, but requires aforesaid.
So, these SSE2-instructions are to be executed only in case there are more than 4 pairs to work on and
the pointer ar3 is aligned on a 16-byte boundary.
Also, if ar2 is aligned on a 16-byte boundary as well, this fragment of code is to be executed:
paddd xmm0, xmmword ptr [esi+edi*4] ; ar2+i*4
408
Otherwise, the value from ar2 is to be loaded into XMM0 using MOVDQU, which does not require aligned
pointer, but may work slower:
movdqu xmm0, xmmword ptr [esi+edi*4] ; ar2+i*4 is not 16-byte aligned, so load it to XMM0
paddd xmm1, xmm0
In all other cases, non-SSE2 code is to be executed.
GCC
GCC may also vectorize in simple cases170 , if the -O3 option is used and SSE2 support is turned on: -msse2.
What we get (GCC 4.4.1):
; f(int, int *, int *, int *)
public _Z1fiPiS_S_
_Z1fiPiS_S_ proc near

arg_0 = dword ptr 8
arg_C = dword ptr 14h
push ebp
mov ebp, esp
push edi
push esi
push ebx
sub esp, 0Ch
mov esi, [ebp+arg_4]
mov edi, [ebp+arg_8]
mov ebx, [ebp+arg_C]
test ecx, ecx
jle short loc_80484D8
cmp ecx, 6
lea eax, [ebx+10h]
ja short loc_80484E8
loc_80484C1: ; CODE XREF: f(int,int *,int *,int *)+4B

; f(int,int *,int *,int *)+61 ...
xor eax, eax
nop
lea esi, [esi+0]
loc_80484C8: ; CODE XREF: f(int,int *,int *,int *)+36

mov edx, [edi+eax*4]
add edx, [esi+eax*4]
mov [ebx+eax*4], edx
add eax, 1
cmp eax, ecx
jnz short loc_80484C8
loc_80484D8: ; CODE XREF: f(int,int *,int *,int *)+17

; f(int,int *,int *,int *)+A5
add esp, 0Ch
xor eax, eax
pop ebx
pop esi
pop edi
170 More about GCC vectorization support: http://go.yurichev.com/17083
409
pop ebp
retn
align 8
loc_80484E8: ; CODE XREF: f(int,int *,int *,int *)+1F

test bl, 0Fh
jnz short loc_80484C1
lea edx, [esi+10h]
cmp ebx, edx
jbe loc_8048578
loc_80484F8: ; CODE XREF: f(int,int *,int *,int *)+E0

lea edx, [edi+10h]
cmp ebx, edx
cmp edi, eax
jbe short loc_80484C1
loc_8048503: ; CODE XREF: f(int,int *,int *,int *)+5D

mov eax, ecx
shr eax, 2
shl eax, 2
test eax, eax
jz short loc_8048547
mov [ebp+var_18], ecx
mov ecx, [ebp+var_14]
xor eax, eax
xor edx, edx
nop
loc_8048520: ; CODE XREF: f(int,int *,int *,int *)+9B

movdqu xmm1, xmmword ptr [edi+eax]
movdqu xmm0, xmmword ptr [esi+eax]
add edx, 1
paddd xmm0, xmm1
movdqa xmmword ptr [ebx+eax], xmm0
add eax, 10h
cmp edx, ecx
jb short loc_8048520
mov ecx, [ebp+var_18]
mov eax, [ebp+var_10]
cmp ecx, eax
jz short loc_80484D8

lea edx, ds:0[eax*4]
add esi, edx
add edi, edx
add ebx, edx
lea esi, [esi+0]
loc_8048558: ; CODE XREF: f(int,int *,int *,int *)+CC

mov edx, [edi]
add eax, 1
add edi, 4
add edx, [esi]
add esi, 4
mov [ebx], edx
add ebx, 4
cmp ecx, eax
jg short loc_8048558
add esp, 0Ch
xor eax, eax
pop ebx
pop esi
pop edi
pop ebp
410
retn

cmp eax, esi
jnb loc_80484C1
jmp loc_80484F8
_Z1fiPiS_S_ endp
Almost the same, however, not as meticulously as Intel C++.
Memory copy example
Let’s revisit the simple memcpy() example ( 1.22.2 on page 194):

#include <stdio.h>
void my_memcpy (unsigned char* dst, unsigned char* src, size_t cnt)
{
size_t i;
dst[i]=src[i];
};
And that’s what optimizations GCC 4.9.1 did:

my_memcpy:
; RDI = destination address
; RSI = source address
; RDX = size of block
test rdx, rdx
je .L41
lea rax, [rdi+16]
cmp rsi, rax
lea rax, [rsi+16]
setae cl
cmp rdi, rax
setae al
or cl, al
je .L13
cmp rdx, 22
jbe .L13
mov rcx, rsi
push rbp
push rbx
neg rcx
and ecx, 15
cmp rcx, rdx
cmova rcx, rdx
xor eax, eax
test rcx, rcx
je .L4
movzx eax, BYTE PTR [rsi]
cmp rcx, 1
mov BYTE PTR [rdi], al
je .L15
movzx eax, BYTE PTR [rsi+1]
cmp rcx, 2
mov BYTE PTR [rdi+1], al
je .L16
cmp rcx, 3
je .L17
cmp rcx, 4
je .L18
411
cmp rcx, 5
je .L19
cmp rcx, 6
je .L20
cmp rcx, 7
je .L21
cmp rcx, 8
je .L22
cmp rcx, 9
je .L23
cmp rcx, 10
je .L24
cmp rcx, 11
je .L25
cmp rcx, 12
je .L26
cmp rcx, 13
je .L27
cmp rcx, 15
jne .L28
mov eax, 15
.L4:
mov r10, rdx
lea r9, [rdx-1]
sub r10, rcx
lea r8, [r10-16]
sub r9, rcx
shr r8, 4
add r8, 1
mov r11, r8
sal r11, 4
cmp r9, 14
jbe .L6
lea rbp, [rsi+rcx]
xor r9d, r9d
add rcx, rdi
xor ebx, ebx
.L7:
movdqa xmm0, XMMWORD PTR [rbp+0+r9]
add rbx, 1
movups XMMWORD PTR [rcx+r9], xmm0
add r9, 16
cmp rbx, r8
jb .L7
add rax, r11
cmp r10, r11
je .L1
.L6:
movzx ecx, BYTE PTR [rsi+rax]
412
lea rcx, [rax+1]
cmp rdx, rcx
jbe .L1
movzx ecx, BYTE PTR [rsi+1+rax]
mov BYTE PTR [rdi+1+rax], cl
lea rcx, [rax+2]
cmp rdx, rcx
jbe .L1
lea rcx, [rax+3]
cmp rdx, rcx
jbe .L1
lea rcx, [rax+4]
cmp rdx, rcx
jbe .L1
lea rcx, [rax+5]
cmp rdx, rcx
jbe .L1
lea rcx, [rax+6]
cmp rdx, rcx
jbe .L1
lea rcx, [rax+7]
cmp rdx, rcx
jbe .L1
lea rcx, [rax+8]
cmp rdx, rcx
jbe .L1
lea rcx, [rax+9]
cmp rdx, rcx
jbe .L1
lea rcx, [rax+10]
cmp rdx, rcx
jbe .L1
lea rcx, [rax+11]
cmp rdx, rcx
jbe .L1
lea rcx, [rax+12]
cmp rdx, rcx
jbe .L1
lea rcx, [rax+13]
cmp rdx, rcx
jbe .L1
lea rcx, [rax+14]
cmp rdx, rcx
jbe .L1
movzx edx, BYTE PTR [rsi+14+rax]
413
mov BYTE PTR [rdi+14+rax], dl
.L1:
pop rbx
pop rbp
.L41:
rep ret
.L13:
xor eax, eax
.L3:
movzx ecx, BYTE PTR [rsi+rax]
add rax, 1
cmp rax, rdx
jne .L3
rep ret
.L28:
mov eax, 14
jmp .L4
.L15:
mov eax, 1
jmp .L4
.L16:
mov eax, 2
jmp .L4
.L17:
mov eax, 3
jmp .L4
.L18:
mov eax, 4
jmp .L4
.L19:
mov eax, 5
jmp .L4
.L20:
mov eax, 6
jmp .L4
.L21:
mov eax, 7
jmp .L4
.L22:
mov eax, 8
jmp .L4
.L23:
mov eax, 9
jmp .L4
.L24:
mov eax, 10
jmp .L4
.L25:
mov eax, 11
jmp .L4
.L26:
mov eax, 12
jmp .L4
.L27:
mov eax, 13
jmp .L4
1.36.2 SIMD strlen() implementation

It has to be noted that the SIMD instructions can be inserted in C/C++ code via special macros171 . For
MSVC, some of them are located in the intrin.h file.
It is possible to implement the strlen() function172 using SIMD instructions that works 2-2.5 times faster
than the common implementation. This function loads 16 characters into a XMM-register and check each
171 MSDN: MMX, SSE, and SSE2 Intrinsics
172 strlen() —standard C library function for calculating string length
414
173
against zero .
size_t strlen_sse2(const char *str)
{
register size_t len = 0;
const char *s=str;
bool str_is_aligned=(((unsigned int)str)&0xFFFFFFF0) == (unsigned int)str;
if (str_is_aligned==false)
return strlen (str);
__m128i xmm0 = _mm_setzero_si128();

__m128i xmm1;
int mask = 0;
for (;;)
{
xmm1 = _mm_load_si128((__m128i *)s);
xmm1 = _mm_cmpeq_epi8(xmm1, xmm0);
if ((mask = _mm_movemask_epi8(xmm1)) != 0)
{
unsigned long pos;
_BitScanForward(&pos, mask);
len += (size_t)pos;
break;
}
s += sizeof(__m128i);
len += sizeof(__m128i);
};
return len;
}
Let’s compile it in MSVC 2010 with /Ox option:

_pos$75552 = -4 ; size = 4
_str$ = 8 ; size = 4
?strlen_sse2@@YAIPBD@Z PROC ; strlen_sse2
push ebp
mov ebp, esp
and esp, -16 ; fffffff0H
mov eax, DWORD PTR _str$[ebp]
sub esp, 12 ; 0000000cH
push esi
mov esi, eax
and esi, -16 ; fffffff0H
xor edx, edx
mov ecx, eax
cmp esi, eax
je SHORT $LN4@strlen_sse
lea edx, DWORD PTR [eax+1]
$LL11@strlen_sse:
mov cl, BYTE PTR [eax]
inc eax
test cl, cl
jne SHORT $LL11@strlen_sse
sub eax, edx
pop esi
mov esp, ebp
pop ebp
ret 0
$LN4@strlen_sse:
movdqa xmm1, XMMWORD PTR [eax]
pxor xmm0, xmm0
173 The example is based on source code from: http://go.yurichev.com/17330.
415
pcmpeqb xmm1, xmm0
pmovmskb eax, xmm1
test eax, eax
jne SHORT $LN9@strlen_sse
$LL3@strlen_sse:
movdqa xmm1, XMMWORD PTR [ecx+16]
add ecx, 16 ; 00000010H
pcmpeqb xmm1, xmm0
add edx, 16 ; 00000010H
pmovmskb eax, xmm1
test eax, eax
je SHORT $LL3@strlen_sse
$LN9@strlen_sse:
bsf eax, eax
mov ecx, eax
mov DWORD PTR _pos$75552[esp+16], eax
lea eax, DWORD PTR [ecx+edx]
pop esi
mov esp, ebp
pop ebp
ret 0
?strlen_sse2@@YAIPBD@Z ENDP ; strlen_sse2
How it works? First of all, we must understand goal of the function. It calculates C-string length, but we
can use different terms: it’s task is searching for zero byte, and then calculating its position relatively to
string start.
First, we check if the str pointer is aligned on a 16-byte boundary. If not, we call the generic strlen()
implementation.
Then, we load the next 16 bytes into the XMM1 register using MOVDQA.
An observant reader might ask, why can’t MOVDQU be used here since it can load data from the memory
regardless pointer alignment?
Yes, it might be done in this way: if the pointer is aligned, load data using MOVDQA, if not —use the slower
MOVDQU.
But here we are may hit another caveat:
In the Windows NT line of OS (but not limited to it), memory is allocated by pages of 4 KiB (4096 bytes).
Each win32-process has 4 GiB available, but in fact, only some parts of the address space are connected
to real physical memory. If the process is accessing an absent memory block, an exception is to be raised.
That’s how VM works174 .
So, a function loading 16 bytes at once may step over the border of an allocated memory block. Let’s
say that the OS has allocated 8192 (0x2000) bytes at address 0x008c0000. Thus, the block is the bytes
starting from address 0x008c0000 to 0x008c1fff inclusive.
After the block, that is, starting from address 0x008c2000 there is nothing at all, e.g. the OS not allocated
any memory there. Any attempt to access memory starting from that address will raise an exception.
And let’s consider the example in which the program is holding a string that contains 5 characters almost
at the end of a block, and that is not a crime.
0x008c1ff8 ’h’
0x008c1ff9 ’e’
0x008c1ffa ’l’
0x008c1ffb ’l’
0x008c1ffc ’o’
0x008c1ffd ’\x00’
0x008c1ffe random noise
0x008c1fff random noise
So, in normal conditions the program calls strlen(), passing it a pointer to the string 'hello' placed
in memory at address 0x008c1ff8. strlen() reads one byte at a time until 0x008c1ffd, where there’s a
zero byte, and then it stops.
Now if we implement our own strlen() reading 16 bytes at once, starting at any address, aligned or
not, MOVDQU may attempt to load 16 bytes at once at address 0x008c1ff8 up to 0x008c2008, and then an
174 wikipedia
416
exception will be raised. That situation is to be avoided, of course.
So then we’ll work only with the addresses aligned on a 16 bytes boundary, which in combination with
the knowledge that the OS’ page size is usually aligned on a 16-byte boundary gives us some warranty
that our function will not read from unallocated memory.
Let’s get back to our function.
_mm_setzero_si128()—is a macro generating pxor xmm0, xmm0 —it just clears the XMM0 register.
_mm_load_si128()—is a macro for MOVDQA, it just loads 16 bytes from the address into the XMM1 register.
_mm_cmpeq_epi8()—is a macro for PCMPEQB, an instruction that compares two XMM-registers bytewise.
And if some byte is equals to the one in the other register, there will be 0xff at this point in the result or
0 if otherwise.
For example:
XMM1: 0x11223344556677880000000000000000
XMM0: 0x11ab3444007877881111111111111111
After the execution of pcmpeqb xmm1, xmm0, the XMM1 register contains:
XMM1: 0xff0000ff0000ffff0000000000000000
In our case, this instruction compares each 16-byte block with a block of 16 zero-bytes, which has been
set in the XMM0 register by pxor xmm0, xmm0.
The next macro is _mm_movemask_epi8() —that is the PMOVMSKB instruction.
It is very useful with PCMPEQB.
pmovmskb eax, xmm1
This instruction sets first EAX bit to 1 if the most significant bit of the first byte in XMM1 is 1. In other words,
if the first byte of the XMM1 register is 0xff, then the first bit of EAX is to be 1, too.
If the second byte in the XMM1 register is 0xff, then the second bit in EAX is to be set to 1. In other words,
the instruction is answering the question “which bytes in XMM1 has the most significant bit set, or greater
than 0x7f”, and returns 16 bits in the EAX register. The other bits in the EAX register are to be cleared.
By the way, do not forget about this quirk of our algorithm. There might be 16 bytes in the input like:
15 14 13 12 11 10 9 3 2 1 0
’h’ ’e’ ’l’ ’l’ ’o’ 0 garbage 0 garbage
It is the 'hello' string, terminating zero, and some random noise in memory.
If we load these 16 bytes into XMM1 and compare them with the zeroed XMM0, we are getting something
like 175 :
XMM1: 0x0000ff00000000000000ff0000000000
This means that the instruction found two zero bytes, and it is not surprising.
PMOVMSKB in our case will set EAX to
0b0010000000100000.
Obviously, our function must take only the first zero bit and ignore the rest.
The next instruction is BSF (Bit Scan Forward).
This instruction finds the first bit set to 1 and stores its position into the first operand.
EAX=0b0010000000100000
After the execution of bsf eax, eax, EAX contains 5, meaning 1 has been found at the 5th bit position
(starting from zero).
MSVC has a macro for this instruction: _BitScanForward.
Now it is simple. If a zero byte has been found, its position is added to what we have already counted and
now we have the return result.
Almost all.
175 An order from MSB to LSB176 is used here.
417
By the way, it is also has to be noted that the MSVC compiler emitted two loop bodies side by side, for
optimization.
By the way, SSE 4.2 (that appeared in Intel Core i7) offers more instructions where these string manipula-
tions might be even easier: http://go.yurichev.com/17331
1.37 64 bits
1.37.1 x86-64
It is a 64-bit extension to the x86 architecture.
From the reverse engineer’s perspective, the most important changes are:
• Almost all registers (except FPU and SIMD) were extended to 64 bits and got a R- prefix. 8 additional
registers wer added. Now GPR’s are: RAX, RBX, RCX, RDX, RBP, RSP, RSI, RDI, R8, R9, R10, R11, R12,
R13, R14, R15.
It is still possible to access the older register parts as usual. For example, it is possible to access the
lower 32-bit part of the RAX register using EAX:
Byte number:
RAXx64
EAX
AX
AH AL
The new R8-R15 registers also have their lower parts: R8D-R15D (lower 32-bit parts), R8W-R15W (lower
16-bit parts), R8L-R15L (lower 8-bit parts).
Byte number:
R8
R8D
R8W
R8L
The number of SIMD registers was doubled from 8 to 16: XMM0-XMM15.
• In Win64, the function calling convention is slightly different, somewhat resembling fastcall ( 6.1.3
on page 726). The first 4 arguments are stored in the RCX, RDX, R8, R9 registers, the rest —in the
stack. The caller function must also allocate 32 bytes so the callee may save there 4 first arguments
and use these registers for its own needs. Short functions may use arguments just from registers,
but larger ones may save their values on the stack.
System V AMD64 ABI (Linux, *BSD, Mac OS X)[Michael Matz, Jan Hubicka, Andreas Jaeger, Mark
Mitchell, System V Application Binary Interface. AMD64 Architecture Processor Supplement, (2013)]
177
also somewhat resembles fastcall, it uses 6 registers RDI, RSI, RDX, RCX, R8, R9 for the first 6
arguments. All the rest are passed via the stack.
See also the section on calling conventions ( 6.1 on page 725).
• The C/C++ int type is still 32-bit for compatibility.
• All pointers are 64-bit now.
Since now the number of registers is doubled, the compilers have more space for maneuvering called
register allocation. For us this implies that the emitted code containing less number of local variables.
For example, the function that calculates the first S-box of the DES encryption algorithm processes 32/64/128/256
values at once (depending on DES_type type (uint32, uint64, SSE2 or AVX)) using the bitslice DES method
(read more about this technique here ( 1.36 on page 405)):
/*
* Generated S-box files.
*
* This software may be modified, redistributed, and used for any purpose,
* so long as its origin is acknowledged.
*
418
* Produced by Matthew Kwan - March 1998
*/
#ifdef _WIN64
#define DES_type unsigned __int64
#else
#define DES_type unsigned int
#endif
void
s1 (
DES_type a1,
DES_type a2,
DES_type a3,
DES_type a4,
DES_type a5,
DES_type a6,
DES_type *out1,
DES_type *out2,
DES_type *out3,
DES_type *out4
) {
DES_type x1, x2, x3, x4, x5, x6, x7, x8;
x1 = a3 & ~a5;
x2 = x1 ^ a4;
x3 = a3 & ~a4;
x4 = x3 | a5;
x5 = a6 & x4;
x6 = x2 ^ x5;
x7 = a4 & ~a5;
x8 = a3 ^ a4;
x9 = a6 & ~x8;
x10 = x7 ^ x9;
x11 = a2 | x10;
x12 = x6 ^ x11;
x13 = a5 ^ x5;
x14 = x13 & x8;
x15 = a5 & ~a4;
x16 = x3 ^ x14;
x17 = a6 | x16;
x18 = x15 ^ x17;
x19 = a2 | x18;
x20 = x14 ^ x19;
x21 = a1 & x20;
x22 = x12 ^ ~x21;
*out2 ^= x22;
x23 = x1 | x5;
x24 = x23 ^ x8;
x25 = x18 & ~x2;
x26 = a2 & ~x25;
x27 = x24 ^ x26;
x28 = x6 | x7;
x29 = x28 ^ x25;
x30 = x9 ^ x24;
x31 = x18 & ~x30;
x32 = a2 & x31;
x33 = x29 ^ x32;
x34 = a1 & x33;
x35 = x27 ^ x34;
*out4 ^= x35;
x36 = a3 & x28;
x37 = x18 & ~x36;
x38 = a2 | x3;
419
x39 = x37 ^ x38;
x40 = a3 | x31;
x41 = x24 & ~x37;
x42 = x41 | x3;
x43 = x42 & ~a2;
x44 = x40 ^ x43;
x45 = a1 & ~x44;
x46 = x39 ^ ~x45;
*out1 ^= x46;
x47 = x33 & ~x9;
x48 = x47 ^ x39;
x49 = x4 ^ x36;
x50 = x49 & ~x5;
x51 = x42 | x18;
x52 = x51 ^ a5;
x53 = a2 & ~x52;
x54 = x50 ^ x53;
x55 = a1 | x54;
x56 = x48 ^ ~x55;
*out3 ^= x56;
}
There are a lot of local variables. Of course, not all those going into the local stack. Let’s compile it with
MSVC 2008 with /Ox option:

PUBLIC _s1
; Function compile flags: /Ogtpy
_TEXT SEGMENT
_x6$ = -20 ; size = 4
_x3$ = -16 ; size = 4
_x1$ = -12 ; size = 4
_x8$ = -8 ; size = 4
_x4$ = -4 ; size = 4
_a1$ = 8 ; size = 4
_a2$ = 12 ; size = 4
_a3$ = 16 ; size = 4
_x33$ = 20 ; size = 4
_x7$ = 20 ; size = 4
_a4$ = 20 ; size = 4
_a5$ = 24 ; size = 4
tv326 = 28 ; size = 4
_x36$ = 28 ; size = 4
_x28$ = 28 ; size = 4
_a6$ = 28 ; size = 4
_out1$ = 32 ; size = 4
_x24$ = 36 ; size = 4
_out2$ = 36 ; size = 4
_out3$ = 40 ; size = 4
_out4$ = 44 ; size = 4
_s1 PROC
sub esp, 20 ; 00000014H
mov edx, DWORD PTR _a5$[esp+16]
push ebx
mov ebx, DWORD PTR _a4$[esp+20]
push ebp
push esi
mov esi, DWORD PTR _a3$[esp+28]
push edi
mov edi, ebx
not edi
mov ebp, edi
and edi, DWORD PTR _a5$[esp+32]
mov ecx, edx
not ecx
and ebp, esi
mov eax, ecx
and eax, esi
and ecx, ebx
420
mov DWORD PTR _x1$[esp+36], eax
xor eax, ebx
mov esi, ebp
or esi, edx
mov DWORD PTR _x4$[esp+36], esi
and esi, DWORD PTR _a6$[esp+32]
mov DWORD PTR _x7$[esp+32], ecx
mov edx, esi
xor edx, eax
mov DWORD PTR _x6$[esp+36], edx
mov edx, DWORD PTR _a3$[esp+32]
xor edx, ebx
mov ebx, esi
xor ebx, DWORD PTR _a5$[esp+32]
mov DWORD PTR _x8$[esp+36], edx
and ebx, edx
mov ecx, edx
mov edx, ebx
xor edx, ebp
or edx, DWORD PTR _a6$[esp+32]
not ecx
and ecx, DWORD PTR _a6$[esp+32]
xor edx, edi
mov edi, edx
or edi, DWORD PTR _a2$[esp+32]
mov DWORD PTR _x3$[esp+36], ebp
mov ebp, DWORD PTR _a2$[esp+32]
xor edi, ebx
and edi, DWORD PTR _a1$[esp+32]
mov ebx, ecx
xor ebx, DWORD PTR _x7$[esp+32]
not edi
or ebx, ebp
xor edi, ebx
mov ebx, edi
mov edi, DWORD PTR _out2$[esp+32]
xor ebx, DWORD PTR [edi]
not eax
and eax, edx
mov DWORD PTR [edi], ebx
mov ebx, DWORD PTR _x7$[esp+32]
or ebx, DWORD PTR _x6$[esp+36]
mov edi, esi
or edi, DWORD PTR _x1$[esp+36]
mov DWORD PTR _x28$[esp+32], ebx
xor edi, DWORD PTR _x8$[esp+36]
mov DWORD PTR _x24$[esp+32], edi
xor edi, ecx
not edi
and edi, edx
mov ebx, edi
and ebx, ebp
xor ebx, eax
not eax
mov DWORD PTR _x33$[esp+32], ebx
and ebx, DWORD PTR _a1$[esp+32]
and eax, ebp
xor eax, ebx
mov ebx, DWORD PTR _out4$[esp+32]
xor eax, DWORD PTR [ebx]
xor eax, DWORD PTR _x24$[esp+32]
mov DWORD PTR [ebx], eax
mov eax, DWORD PTR _x28$[esp+32]
and eax, DWORD PTR _a3$[esp+32]
mov ebx, DWORD PTR _x3$[esp+36]
or edi, DWORD PTR _a3$[esp+32]
mov DWORD PTR _x36$[esp+32], eax
not eax
421
and eax, edx
or ebx, ebp
xor ebx, eax
not eax
and eax, DWORD PTR _x24$[esp+32]
not ebp
or eax, DWORD PTR _x3$[esp+36]
not esi
and ebp, eax
or eax, edx
xor eax, DWORD PTR _a5$[esp+32]
mov edx, DWORD PTR _x36$[esp+32]
xor edx, DWORD PTR _x4$[esp+36]
xor ebp, edi
mov edi, DWORD PTR _out1$[esp+32]
not eax
and eax, DWORD PTR _a2$[esp+32]
not ebp
and ebp, DWORD PTR _a1$[esp+32]
and edx, esi
xor eax, edx
or eax, DWORD PTR _a1$[esp+32]
not ebp
xor ebp, DWORD PTR [edi]
not ecx
and ecx, DWORD PTR _x33$[esp+32]
xor ebp, ebx
not eax
mov DWORD PTR [edi], ebp
xor eax, ecx
mov ecx, DWORD PTR _out3$[esp+32]
xor eax, DWORD PTR [ecx]
pop edi
pop esi
xor eax, ebx
pop ebp
pop ebx
add esp, 20
ret 0
_s1 ENDP
5 variables were allocated in the local stack by the compiler.

Now let’s try the same thing in the 64-bit version of MSVC 2008:

a1$ = 56
a2$ = 64
a3$ = 72
a4$ = 80
x36$1$ = 88
a5$ = 88
a6$ = 96
out1$ = 104
out2$ = 112
out3$ = 120
out4$ = 128
s1 PROC
$LN3:
mov QWORD PTR [rsp+24], rbx
mov QWORD PTR [rsp+32], rbp
mov QWORD PTR [rsp+16], rdx
mov QWORD PTR [rsp+8], rcx
push rsi
push rdi
push r12
push r13
push r14
push r15
422
mov r15, QWORD PTR a5$[rsp]
mov rcx, QWORD PTR a6$[rsp]
mov rbp, r8
mov r10, r9
mov rax, r15
mov rdx, rbp
not rax
xor rdx, r9
not r10
mov r11, rax
and rax, r9
mov rsi, r10
mov QWORD PTR x36$1$[rsp], rax
and r11, r8
and rsi, r8
and r10, r15
mov r13, rdx
mov rbx, r11
xor rbx, r9
mov r9, QWORD PTR a2$[rsp]
mov r12, rsi
or r12, r15
not r13
and r13, rcx
mov r14, r12
and r14, rcx
mov rax, r14
mov r8, r14
xor r8, rbx
xor rax, r15
not rbx
and rax, rdx
mov rdi, rax
xor rdi, rsi
or rdi, rcx
xor rdi, r10
and rbx, rdi
mov rcx, rdi
or rcx, r9
xor rcx, rax
mov rax, r13
xor rax, QWORD PTR x36$1$[rsp]
and rcx, QWORD PTR a1$[rsp]
or rax, r9
not rcx
xor rcx, rax
mov rax, QWORD PTR out2$[rsp]
xor rcx, QWORD PTR [rax]
xor rcx, r8
mov QWORD PTR [rax], rcx
mov rax, QWORD PTR x36$1$[rsp]
mov rcx, r14
or rax, r8
or rcx, r11
mov r11, r9
xor rcx, rdx
mov QWORD PTR x36$1$[rsp], rax
mov r8, rsi
mov rdx, rcx
xor rdx, r13
not rdx
and rdx, rdi
mov r10, rdx
and r10, r9
xor r10, rax
xor r10, rbx
not rbx
and rbx, r9
mov rax, r10
and rax, QWORD PTR a1$[rsp]
423
xor rbx, rax
xor rbx, QWORD PTR [rax]
xor rbx, rcx
mov QWORD PTR [rax], rbx
mov rbx, QWORD PTR x36$1$[rsp]
and rbx, rbp
mov r9, rbx
not r9
and r9, rdi
or r8, r11
xor r8, r9
not r9
and r9, rcx
or rdx, rbp
mov rbp, QWORD PTR [rsp+80]
or r9, rsi
xor rbx, r12
mov rcx, r11
not rcx
not r14
not r13
and rcx, r9
or r9, rdi
and rbx, r14
xor r9, r15
xor rcx, rdx
mov rdx, QWORD PTR a1$[rsp]
not r9
not rcx
and r13, r10
and r9, r11
and rcx, rdx
xor r9, rbx
mov rbx, QWORD PTR [rsp+72]
not rcx
xor rcx, QWORD PTR [rax]
or r9, rdx
not r9
xor rcx, r8
mov QWORD PTR [rax], rcx
xor r9, r13
xor r9, QWORD PTR [rax]
xor r9, r8
mov QWORD PTR [rax], r9
pop r15
pop r14
pop r13
pop r12
pop rdi
pop rsi
ret 0
s1 ENDP
Nothing was allocated in the local stack by the compiler, x36 is synonym for a5.
By the way, there are CPUs with much more GPR’s, e.g. Itanium (128 registers).
1.37.2 ARM
64-bit instructions appeared in ARMv8.
1.37.3 Float point numbers

How floating point numbers are processed in x86-64 is explained here: 1.38 on the next page.
424
1.37.4 64-bit architecture criticism
Some people has irritation sometimes: now one needs twice as much memory for storing pointers, includ-
ing cache memory, despite the fact that x64 CPUs can address only 48 bits of external RAM.
Pointers have gone out of favor to the point now where I had to flame about it because
on my 64-bit computer that I have here, if I really care about using the capability of my
machine I find that I’d better not use pointers because I have a machine that has 64-bit
registers but it only has 2 gigabytes of RAM. So a pointer never has more than 32 significant
bits to it. But every time I use a pointer it’s costing me 64 bits and that doubles the size
of my data structure. Worse, it goes into the cache and half of my cache is gone and that
costs cash—cache is expensive.
So if I’m really trying to push the envelope now, I have to use arrays instead of pointers.
I make complicated macros so that it looks like I’m using pointers, but I’m not really.
( Donald Knuth in “Coders at Work: Reflections on the Craft of Programming ”. )

Some people make their own memory allocators. It’s interesting to know about CryptoMiniSat178 case.
This program rarely uses more than 4GiB of RAM, but it uses pointers heavily. So it requires less memory
on 32-bit architecture than on 64-bit one. To mitigate this problem, author made his own allocator (in
clauseallocator.(h|cpp) files), which allows to have access to allocated memory using 32-bit identifiers
instead of 64-bit pointers.
1.38 Working with floating point numbers using SIMD

Of course, the FPU has remained in x86-compatible processors when the SIMD extensions were added.
The SIMD extensions (SSE2) offer an easier way to work with floating-point numbers.
The number format remains the same (IEEE 754).
So, modern compilers (including those generating for x86-64) usually use SIMD instructions instead of
FPU ones.
It can be said that it’s good news, because it’s easier to work with them.
We are going to reuse the examples from the FPU section here: 1.25 on page 217.

#include <stdio.h>
double f (double a, double b)

{
return a/3.14 + b*4.1;
};
int main()
{
printf ("%f\n", f(1.2, 3.4));
};
x64

__real@4010666666666666 DQ 04010666666666666r ; 4.1
a$ = 8
b$ = 16
f PROC
divsd xmm0, QWORD PTR __real@40091eb851eb851f
mulsd xmm1, QWORD PTR __real@4010666666666666
addsd xmm0, xmm1
178 https://github.com/msoos/cryptominisat/
425
ret 0
f ENDP
179
The input floating point values are passed in the XMM0-XMM3 registers, all the rest—via the stack .
a is passed in XMM0, b—via XMM1.
The XMM-registers are 128-bit (as we know from the section about SIMD: 1.36 on page 404), but the
double values are 64 bit, so only lower register half is used.
DIVSD is an SSE-instruction that stands for “Divide Scalar Double-Precision Floating-Point Values”, it just
divides one value of type double by another, stored in the lower halves of operands.
The constants are encoded by compiler in IEEE 754 format.
MULSD and ADDSD work just as the same, but do multiplication and addition.
The result of the function’s execution in type double is left in the in XMM0 register.
That is how non-optimizing MSVC works:

__real@4010666666666666 DQ 04010666666666666r ; 4.1
a$ = 8
b$ = 16
f PROC
movsdx QWORD PTR [rsp+16], xmm1
movsdx xmm0, QWORD PTR a$[rsp]
movsdx xmm1, QWORD PTR b$[rsp]
addsd xmm0, xmm1
ret 0
f ENDP
Slightly redundant. The input arguments are saved in the “shadow space” ( 1.14.2 on page 100), but only
their lower register halves, i.e., only 64-bit values of type double. GCC produces the same code.
x86
Let’s also compile this example for x86. Despite the fact it’s generating for x86, MSVC 2012 uses SSE2
instructions:

tv70 = -8 ; size = 8
_a$ = 8 ; size = 8
_b$ = 16 ; size = 8
_f PROC
push ebp
mov ebp, esp
sub esp, 8
movsd xmm0, QWORD PTR _a$[ebp]
movsd xmm1, QWORD PTR _b$[ebp]
addsd xmm0, xmm1
movsd QWORD PTR tv70[ebp], xmm0
fld QWORD PTR tv70[ebp]
mov esp, ebp
pop ebp
ret 0
_f ENDP
179 MSDN: Parameter Passing
426
tv67 = 8 ; size = 8
_a$ = 8 ; size = 8
_b$ = 16 ; size = 8
_f PROC
movsd xmm1, QWORD PTR _a$[esp-4]
movsd xmm0, QWORD PTR _b$[esp-4]
addsd xmm1, xmm0
movsd QWORD PTR tv67[esp-4], xmm1
fld QWORD PTR tv67[esp-4]
ret 0
_f ENDP
It’s almost the same code, however, there are some differences related to calling conventions: 1) the
arguments are passed not in XMM registers, but in the stack, like in the FPU examples ( 1.25 on page 217);
2) the result of the function is returned in ST(0) — in order to do so, it’s copied (through local variable tv)
from one of the XMM registers to ST(0).
427
Let’s try the optimized example in OllyDbg:
Figure 1.114: OllyDbg: MOVSD loads the value of a into XMM1
428
Figure 1.115: OllyDbg: DIVSD calculated quotient and stored it in XMM1
429
Figure 1.116: OllyDbg: MULSD calculated product and stored it in XMM0
430
Figure 1.117: OllyDbg: ADDSD adds value in XMM0 to XMM1
431
Figure 1.118: OllyDbg: FLD left function result in ST(0)
We see that OllyDbg shows the XMM registers as pairs of double numbers, but only the lower part is used.
Apparently, OllyDbg shows them in that format because the SSE2 instructions (suffixed with -SD) are
executed right now.
But of course, it’s possible to switch the register format and to see their contents as 4 float-numbers or
just as 16 bytes.
432
1.38.2 Passing floating point number via arguments
#include <math.h>
#include <stdio.h>
int main ()
{
printf ("32.01 ^ 1.54 = %lf\n", pow (32.01,1.54));
return 0;
}
They are passed in the lower halves of the XMM0-XMM3 registers.

$SG1354 DB '32.01 ^ 1.54 = %lf', 0aH, 00H
__real@40400147ae147ae1 DQ 040400147ae147ae1r ; 32.01

__real@3ff8a3d70a3d70a4 DQ 03ff8a3d70a3d70a4r ; 1.54
main PROC
sub rsp, 40 ; 00000028H
movsdx xmm1, QWORD PTR __real@3ff8a3d70a3d70a4
movsdx xmm0, QWORD PTR __real@40400147ae147ae1
call pow
movaps xmm1, xmm0
movd rdx, xmm1
call printf
xor eax, eax
add rsp, 40 ; 00000028H
ret 0
main ENDP
There is no MOVSDX instruction in Intel and AMD manuals ( 12.1.4 on page 999), there it is called just MOVSD.
So there are two instructions sharing the same name in x86 (about the other see: .1.6 on page 1015).
Apparently, Microsoft developers wanted to get rid of the mess, so they renamed it to MOVSDX. It just loads
a value into the lower half of a XMM register.
pow() takes arguments from XMM0 and XMM1, and returns result in XMM0. It is then moved to RDX for
printf(). Why? Maybe because printf()—is a variable arguments function?
.LC2:
.string "32.01 ^ 1.54 = %lf\n"
main:
sub rsp, 8
movsd xmm1, QWORD PTR .LC0[rip]
movsd xmm0, QWORD PTR .LC1[rip]
call pow
; result is now in XMM0
mov eax, 1 ; number of vector registers passed
call printf
xor eax, eax
add rsp, 8
ret
.LC0:
.long 171798692
.long 1073259479
.LC1:
.long 2920577761
.long 1077936455
GCC generates clearer output. The value for printf() is passed in XMM0. By the way, here is a case when
1 is written into EAX for printf()—this implies that one argument will be passed in vector registers, just
as the standard requires [Michael Matz, Jan Hubicka, Andreas Jaeger, Mark Mitchell, System V Application
Binary Interface. AMD64 Architecture Processor Supplement, (2013)] 180 .
433
1.38.3 Comparison example
#include <stdio.h>
double d_max (double a, double b)

{
if (a>b)
return a;
return b;
};
int main()
{
printf ("%f\n", d_max (1.2, 3.4));
printf ("%f\n", d_max (5.6, -4));
};
x64

a$ = 8
b$ = 16
d_max PROC
comisd xmm0, xmm1
ja SHORT $LN2@d_max
movaps xmm0, xmm1
$LN2@d_max:
fatret 0
d_max ENDP
Optimizing MSVC generates a code very easy to understand.

COMISD is “Compare Scalar Ordered Double-Precision Floating-Point Values and Set EFLAGS”. Essentially,
that is what it does.
Non-optimizing MSVC generates more redundant code, but it is still not hard to understand:

a$ = 8
b$ = 16
d_max PROC
comisd xmm0, QWORD PTR b$[rsp]
jbe SHORT $LN1@d_max
jmp SHORT $LN2@d_max
$LN1@d_max:
movsdx xmm0, QWORD PTR b$[rsp]
$LN2@d_max:
fatret 0
d_max ENDP
However, GCC 4.4.6 did more optimizations and used the MAXSD (“Return Maximum Scalar Double-Precision
Floating-Point Value”) instruction, which just choose the maximum value!

d_max:
maxsd xmm0, xmm1
ret
434
x86
Let’s compile this example in MSVC 2012 with optimization turned on:

_a$ = 8 ; size = 8
_b$ = 16 ; size = 8
_d_max PROC
movsd xmm0, QWORD PTR _a$[esp-4]
comisd xmm0, QWORD PTR _b$[esp-4]
jbe SHORT $LN1@d_max
fld QWORD PTR _a$[esp-4]
ret 0
$LN1@d_max:
fld QWORD PTR _b$[esp-4]
ret 0
_d_max ENDP
Almost the same, but the values of a and b are taken from the stack and the function result is left in ST(0).
If we load this example in OllyDbg, we can see how the COMISD instruction compares values and sets/clears
the CF and PF flags:
Figure 1.119: OllyDbg: COMISD changed CF and PF flags
1.38.4 Calculating machine epsilon: x64 and SIMD

Let’s revisit the “calculating machine epsilon” example for double listing.1.32.2.
Now we compile it for x64:
435
v$ = 8
calculate_machine_epsilon PROC
movsdx QWORD PTR v$[rsp], xmm0
movaps xmm1, xmm0
inc QWORD PTR v$[rsp]
movsdx xmm0, QWORD PTR v$[rsp]
subsd xmm0, xmm1
ret 0
calculate_machine_epsilon ENDP
There is no way to add 1 to a value in 128-bit XMM register, so it must be placed into memory.
There is, however, the ADDSD instruction (Add Scalar Double-Precision Floating-Point Values) which can
add a value to the lowest 64-bit half of a XMM register while ignoring the higher one, but MSVC 2012
probably is not that good yet 181 .
Nevertheless, the value is then reloaded to a XMM register and subtraction occurs. SUBSD is “Subtract
Scalar Double-Precision Floating-Point Values”, i.e., it operates on the lower 64-bit part of 128-bit XMM
register. The result is returned in the XMM0 register.
1.38.5 Pseudo-random number generator example revisited

Let’s revisit “pseudo-random number generator example” example listing.1.32.1.
If we compile this in MSVC 2012, it will use the SIMD instructions for the FPU.

__real@3f800000 DD 03f800000r ; 1
tv128 = -4
_tmp$ = -4
?float_rand@@YAMXZ PROC
push ecx
call ?my_rand@@YAIXZ
; EAX=pseudorandom value
and eax, 8388607 ; 007fffffH
or eax, 1065353216 ; 3f800000H
; EAX=pseudorandom value & 0x007fffff | 0x3f800000
; store it into local stack:
mov DWORD PTR _tmp$[esp+4], eax
; reload it as float point number:
movss xmm0, DWORD PTR _tmp$[esp+4]
; subtract 1.0:
subss xmm0, DWORD PTR __real@3f800000
; move value to ST0 by placing it in temporary variable...
movss DWORD PTR tv128[esp+4], xmm0
; ... and reloading it into ST0:
fld DWORD PTR tv128[esp+4]
pop ecx
ret 0
?float_rand@@YAMXZ ENDP
All instructions have the -SS suffix, which stands for “Scalar Single”.
“Scalar” implies that only one value is stored in the register.
“Single”182 stands for float data type.
1.38.6 Summary
Only the lower half of XMM registers is used in all examples here, to store number in IEEE 754 format.
Essentially, all instructions prefixed by -SD (“Scalar Double-Precision”)—are instructions working with float-
ing point numbers in IEEE 754 format, stored in the lower 64-bit half of a XMM register.
And it is easier than in the FPU, probably because the SIMD extensions were evolved in a less chaotic way
than the FPU ones in the past. The stack register model is not used.
181 As an exercise, you may try to rework this code to eliminate the usage of the local stack.
182 I.e., single precision.
436
If you would try to replace double with float
in these examples, the same instructions will be used, but prefixed with -SS (“Scalar Single-Precision”),
for example, MOVSS, COMISS, ADDSS, etc.
“Scalar” implies that the SIMD register containing only one value instead of several.
Instructions working with several values in a register simultaneously have “Packed” in their name.
Needless to say, the SSE2 instructions work with 64-bit IEEE 754 numbers (double), while the internal
representation of the floating-point numbers in FPU is 80-bit numbers.
Hence, the FPU may produce less round-off errors and as a consequence, FPU may give more precise
calculation results.
1.39 ARM-specific details

1.39.1 Number sign (#) before number
The Keil compiler, IDA and objdump precede all numbers with the “#” number sign, for example: list-
ing.1.22.1.
But when GCC 4.9 generates assembly language output, it doesn’t, for example: listing.3.18.
The ARM listings in this book are somewhat mixed.
It’s hard to say, which method is right. Supposedly, one has to obey the rules accepted in environment
he/she works in.
1.39.2 Addressing modes

This instruction is possible in ARM64:
ldr x0, [x29,24]
This means add 24 to the value in X29 and load the value from this address.
Please note that 24 is inside the brackets. The meaning is different if the number is outside the brackets:
ldr w4, [x1],28
This means load the value at the address in X1, then add 28 to X1.
ARM allows you to add or subtract a constant to/from the address used for loading.
And it’s possible to do that both before and after loading.
There is no such addressing mode in x86, but it is present in some other processors, even on PDP-11.
There is a legend that the pre-increment, post-increment, pre-decrement and post-decrement modes in
PDP-11,
were “guilty” for the appearance of such C language (which developed on PDP-11) constructs as *ptr++,
*++ptr, *ptr--, *--ptr.
By the way, this is one of the hard to memorize C features. This is how it is:
C term ARM term C statement how it works
Post-increment post-indexed addressing *ptr++ use *ptr value,
then increment
ptr pointer
Post-decrement post-indexed addressing *ptr-- use *ptr value,
then decrement
ptr pointer
Pre-increment pre-indexed addressing *++ptr increment ptr pointer,
then use
*ptr value
Pre-decrement pre-indexed addressing *--ptr decrement ptr pointer,
then use
*ptr value
437
Pre-indexing is marked with an exclamation mark in the ARM assembly language. For example, see line
2 in listing.1.29.
Dennis Ritchie (one of the creators of the C language) mentioned that it presumably was invented by Ken
Thompson (another C creator) because this processor feature was present in PDP-7 183 , [Dennis M. Ritchie,
The development of the C language, (1993)]184 .
Thus, C language compilers may use it, if it is present on the target processor.
That’s very convenient for array processing.
1.39.3 Loading a constant into a register

32-bit ARM
As we already know, all instructions have a length of 4 bytes in ARM mode and 2 bytes in Thumb mode.
Then how can we load a 32-bit value into a register, if it’s not possible to encode it in one instruction?
Let’s try:
unsigned int f()
{
return 0x12345678;
};
Listing 1.413: GCC 4.6.3 -O3 ARM mode

f:
ldr r0, .L2
bx lr
.L2:
.word 305419896 ; 0x12345678
So, the 0x12345678 value is just stored aside in memory and loaded if needed.
But it’s possible to get rid of the additional memory access.
Listing 1.414: GCC 4.6.3 -O3 -march=armv7-a (ARM mode)

movw r0, #22136 ; 0x5678
movt r0, #4660 ; 0x1234
bx lr
We see that the value is loaded into the register by parts, the lower part first (using MOVW), then the higher
(using MOVT).
This implies that 2 instructions are necessary in ARM mode for loading a 32-bit value into a register.
It’s not a real problem, because in fact there are not many constants in real code (except of 0 and 1).
Does it mean that the two-instruction version is slower than one-instruction version?
Doubtfully. Most likely, modern ARM processors are able to detect such sequences and execute them fast.
On the other hand, IDA is able to detect such patterns in the code and disassembles this function as:
MOV R0, 0x12345678
BX LR
ARM64
uint64_t f()
{
return 0x12345678ABCDEF01;
};
183 http://yurichev.com/mirrors/C/c_dmr_postincrement.txt
438
Listing 1.415: GCC 4.9.1 -O3
mov x0, 61185 ; 0xef01
movk x0, 0xabcd, lsl 16
movk x0, 0x5678, lsl 32
movk x0, 0x1234, lsl 48
ret
MOVK stands for “MOV Keep”, i.e., it writes a 16-bit value into the register, not touching the rest of the bits.
The LSL suffix shifts left the value by 16, 32 and 48 bits at each step. The shifting is done before loading.
This implies that 4 instructions are necessary to load a 64-bit value into a register.
Storing floating-point number into register
It’s possible to store a floating-point number into a D-register using only one instruction.
For example:
double a()
{
return 1.5;
};
Listing 1.416: GCC 4.9.1 -O3 + objdump

0000000000000000 <a>:
0: 1e6f1000 fmov d0, #1.500000000000000000e+000
4: d65f03c0 ret
The number 1.5 was indeed encoded in a 32-bit instruction. But how?
In ARM64, there are 8 bits in the FMOV instruction for encoding some floating-point numbers.
The algorithm is called VFPExpandImm() in [ARM Architecture Reference Manual, ARMv8, for ARMv8-A
architecture profile, (2013)]185 . This is also called minifloat 186 .
We can try different values: the compiler is able to encode 30.0 and 31.0, but it couldn’t encode 32.0, as 8
bytes have to be allocated for this number in the IEEE 754 format:
double a()
{
return 32;
};
Listing 1.417: GCC 4.9.1 -O3

a:
ldr d0, .LC0
ret
.LC0:
.word 0
.word 1077936128
1.39.4 Relocs in ARM64

As we know, there are 4-byte instructions in ARM64, so it is impossible to write a large number into a
register using a single instruction.
Nevertheless, an executable image can be loaded at any random address in memory, so that’s why relocs
exists. Read more about them (in relation to Win32 PE): 6.5.2 on page 751.
The address is formed using the ADRP and ADD instruction pair in ARM64.
The first loads a 4KiB-page address and the second one adds the remainder. Let’s compile the example
from “Hello, world!” (listing.1.11) in GCC (Linaro) 4.9 under win32:
185 Also available as http://yurichev.com/mirrors/ARMv8-A_Architecture_Reference_Manual_(Issue_A.a).pdf
186 wikipedia
439
Listing 1.418: GCC (Linaro) 4.9 and objdump of object file
...>aarch64-linux-gnu-gcc.exe hw.c -c
...>aarch64-linux-gnu-objdump.exe -d hw.o
...
0000000000000000 <main>:
0: a9bf7bfd stp x29, x30, [sp,#-16]!
4: 910003fd mov x29, sp
8: 90000000 adrp x0, 0 <main>
c: 91000000 add x0, x0, #0x0
10: 94000000 bl 0 <printf>
14: 52800000 mov w0, #0x0 // #0
18: a8c17bfd ldp x29, x30, [sp],#16
1c: d65f03c0 ret
...>aarch64-linux-gnu-objdump.exe -r hw.o
...

OFFSET TYPE VALUE
0000000000000008 R_AARCH64_ADR_PREL_PG_HI21 .rodata
000000000000000c R_AARCH64_ADD_ABS_LO12_NC .rodata
0000000000000010 R_AARCH64_CALL26 printf
So there are 3 relocs in this object file.

• The first one takes the page address, cuts the lowest 12 bits and writes the remaining high 21 bits
to the ADRP instruction’s bit fields. This is because we don’t need to encode the low 12 bits, and the
ADRP instruction has space only for 21 bits.
• The second one puts the 12 bits of the address relative to the page start into the ADD instruction’s
bit fields.
• The last, 26-bit one, is applied to the instruction at address 0x10 where the jump to the printf()
function is.
All ARM64 (and in ARM in ARM mode) instruction addresses have zeros in the two lowest bits (because
all instructions have a size of 4 bytes), so one have to encode only the highest 26 bits of 28-bit address
space (±128MB).
There are no such relocs in the executable file: because it’s known where the “Hello!” string is located,
in which page, and the address of puts() is also known.
So there are values set already in the ADRP, ADD and BL instructions (the linker has written them while
linking):
Listing 1.419: objdump of executable file

0000000000400590 <main>:
400590: a9bf7bfd stp x29, x30, [sp,#-16]!
400594: 910003fd mov x29, sp
400598: 90000000 adrp x0, 400000 <_init-0x3b8>
40059c: 91192000 add x0, x0, #0x648
4005a0: 97ffffa0 bl 400420 <puts@plt>
4005a4: 52800000 mov w0, #0x0 // #0
4005a8: a8c17bfd ldp x29, x30, [sp],#16
4005ac: d65f03c0 ret
...
Contents of section .rodata:

400640 01000200 00000000 48656c6c 6f210000 ........Hello!..
As an example, let’s try to disassemble the BL instruction manually.

0x97ffffa0 is 0b10010111111111111111111110100000. According to [ARM Architecture Reference Manual, ARMv8,
for ARMv8-A architecture profile, (2013)C5.6.26], imm26 is the last 26 bits:
imm26 = 0b11111111111111111110100000. It is 0x3FFFFA0, but the MSB is 1, so the number is negative, and we
can convert it manually to convenient form for us. By the rules of negation ( 2.2 on page 449), just invert
440
all bits: (it is 0b1011111=0x5F), and add 1 (0x5F+1=0x60). So the number in signed form is -0x60. Let’s
multiply -0x60 by 4 (because address stored in opcode is divided by 4): it is -0x180. Now let’s calculate
destination address: 0x4005a0 + (-0x180) = 0x400420 (please note: we consider the address of the BL
instruction, not the current value of PC, which may be different!). So the destination address is 0x400420.
More about ARM64-related relocs: [ELF for the ARM 64-bit Architecture (AArch64), (2013)]187 .
1.40 MIPS-specific details

1.40.1 Loading a 32-bit constant into register
unsigned int f()
{
return 0x12345678;
};
All instructions in MIPS, just like ARM, have a size of 32-bit, so it’s not possible to embed a 32-bit constant
into one instruction.
So one have to use at least two instructions: the first loads the high part of the 32-bit number and the
second one applies an OR operation, which effectively sets the low 16-bit part of the target register:
Listing 1.420: GCC 4.4.5 -O3 (assembly output)

li $2,305397760 # 0x12340000
j $31
ori $2,$2,0x5678 ; branch delay slot
IDA is fully aware of such frequently encountered code patterns, so, for convenience it shows the last ORI
instruction as the LI pseudo instruction, which allegedly loads a full 32-bit number into the $V0 register.
Listing 1.421: GCC 4.4.5 -O3 (IDA)

lui $v0, 0x1234
jr $ra
li $v0, 0x12345678 ; branch delay slot
The GCC assembly output has the LI pseudo instruction, but in fact, LUI (“Load Upper Immediate”) is
there, which stores a 16-bit value into the high part of the register.
Let’s see in objdump output:
Listing 1.422: objdump

00000000 <f>:
0: 3c021234 lui v0,0x1234
4: 03e00008 jr ra
8: 34425678 ori v0,v0,0x5678
Loading a 32-bit global variable into register
unsigned int global_var=0x12345678;
unsigned int f2()

{
return global_var;
};
This is slightly different: LUI loads upper 16-bit from global_var into $2 (or $V0) and then LW loads lower
16-bits summing it with the contents of $2:
Listing 1.423: GCC 4.4.5 -O3 (assembly output)

f2:
lui $2,%hi(global_var)
lw $2,%lo(global_var)($2)
441
j $31
nop ; branch delay slot
...
global_var:
.word 305419896
IDA is fully aware of often used LUI/LW instruction pair, so it coalesces both into a single LW instruction:
Listing 1.424: GCC 4.4.5 -O3 (IDA)

_f2:
lw $v0, global_var
jr $ra
...
.data
.globl global_var
global_var: .word 0x12345678 # DATA XREF: _f2
objdump’s output is the same as GCC’s assembly output. Let’s also dump relocs of the object file:
Listing 1.425: objdump

objdump -D filename.o
...
0000000c <f2>:
c: 3c020000 lui v0,0x0
10: 8c420000 lw v0,0(v0)
14: 03e00008 jr ra
18: 00200825 move at,at ; branch delay slot
1c: 00200825 move at,at
Disassembly of section .data:
00000000 <global_var>:
0: 12345678 beq s1,s4,159e4 <f2+0x159d8>
...
objdump -r filename.o
...

OFFSET TYPE VALUE
0000000c R_MIPS_HI16 global_var
00000010 R_MIPS_LO16 global_var
...
We can see that address of global_var is to be written right into LUI and LW instructions during executable
file loading: high 16-bit part of global_var goes into the first one (LUI), lower 16-bit part goes into the
second one (LW).
1.40.2 Further reading about MIPS

Dominic Sweetman, See MIPS Run, Second Edition, (2010).
442
Chapter 2
Important fundamentals
2.1 Integral datatypes

Integral datatype is a type for a value which can be converted to number. These are numbers, enumera-
tions, booleans.
2.1.1 Bit
Obvious usage for bits are boolean values: 0 for false and 1 for true.
Set of booleans can be packed into word: there will be 32 booleans in 32-bit word, etc. This way is called
bitmap or bitfield.
But it has obvious overhead: a bit jiggling, isolating, etc. While using word (or int type) for boolean
variable is not economic, but highly efficient.
In C/C++ environment, 0 is for false and any non-zero value is for true. For example:
if (1234)
printf ("this will always be executed\n");
else
printf ("this will never\n");
This is popular way of enumerating characters in a C-string:

char *input=...;
while(*input) // execute body if *input character is non-zero

{
// do something with *input
input++;
};
2.1.2 Nibble AKA nybble

AKA half-byte, tetrade. Equals to 4 bits.
All these terms are still in use today.
Binary-coded decimal (BCD1 )
4-bit nibbles were used in 4-bit CPUs like legendary Intel 4004 (used in calculators).
It’s interesting to know that there was binary-coded decimal (BCD) way of representing decimal digit using
4 bits. Decimal 0 is represented as 0b0000, decimal 9 as 0b1001 and higher values are not used. Decimal
1234 is represented as 0x1234. Of course, this way is not economical.
Nevertheless, it has one advantage: decimal to BCD-packed number conversion and back is extremely
easy. BCD-numbers can be added, subtracted, etc., but an additional correction is needed. x86 CPUs has
1 Binary-Coded Decimal
443
rare instructions for that: AAA/DAA (adjust after addition), AAS/DAS (adjust after subtraction), AAM (after
multiplication), AAD (after division).
The need for CPUs to support BCD numbers is a reason why half-carry flag (on 8080/Z80) and auxiliary
flag (AF on x86) are exist: this is carry-flag generated after proceeding of lower 4 bits. The flag is then
used for adjustment instructions.
The fact of easy conversion had led to popularity of [Peter Abel, IBM PC assembly language and pro-
gramming (1987)] book. But aside of this book, the author of these notes never seen BCD numbers in
practice, except for magic numbers ( 5.6.1 on page 704), like when someone’s birthday is encoded like
0x19791011—this is indeed packed BCD number.
Surprisingly, the author found a use of BCD-encoded numbers in SAP software: https://yurichev.com/
blog/SAP/. Some numbers, including prices, are encoded in BCD form in database. Perhaps, they used
it to make it compatible with some ancient software/hardware?
BCD instructions in x86 were often used for other purposes, especially in undocumented ways, for exam-
ple:
cmp al,10
sbb al,69h
das
This obscure code converts number in 0..15 range into ASCII character ’0’..’9’, ’A’..’F’.
Z80
Z80 was clone of 8-bit Intel 8080 CPU, and because of space constraints, it has 4-bit ALU, i.e., each
operation over two 8-bit numbers had to be proceeded in two steps. One side-effect of this was easy and
natural generation of half-carry flag.
2.1.3 Byte
Byte is primarily used for character storage. 8-bit bytes were not common as today. Punched tapes for
teletypes had 5 and 6 possible holes, this is 5 or 6 bits for byte.
To emphasize the fact the byte has 8 bits, byte is sometimes called octet: at least fetchmail uses this
terminology.
9-bit bytes used to exist in 36-bit architectures: 4 9-bit bytes would fit in a single word. Probably because
of this fact, C/C++ standard tells that char has to have a room for at least 8 bits, but more bits are
allowable.
For example, in the early C language manual2 , we can find this:
char one byte character (PDP-11, IBM360: 8 bits; H6070: 9 bits)
By H6070 they probably meant Honeywell 6070, with 36-bit words.
Standard ASCII table
7-bit ASCII table is standard, which has only 128 possible characters. Early E-Mail transport software were
operating only on 7-bit ASCII codes, so a MIME3 standard needed to encode messages in non-Latin writing
systems. 7-bit ASCII code was augmented by parity bit, resulting in 8 bits.
Data Encryption Standard (DES4 ) has a 56 bits key, this is 8 7-bit bytes, leaving a space to parity bit for
each character.
There is no need to memorize whole ASCII table, but rather ranges. [0..0x1F] are control characters (non-
printable). [0x20..0x7E] are printable ones. Codes starting at 0x80 are usually used for non-Latin writing
systems and/or pseudographics.
Significant codes which will be easily memorized are: 0 (end of C-string, '\0' in C/C++); 0xA or 10 (line
feed, '\n' in C/C++); 0xD or 13 (carriage return, '\r' in C/C++).
0x20 (space) is also often memorized.
2 https://yurichev.com/mirrors/C/bwk-tutor.html
3 Multipurpose Internet Mail Extensions
4 Data Encryption Standard
444
8-bit CPUs
x86 has capability to work with byte(s) on register level (because they are descendants of 8-bit 8080 CPU),
RISC CPUs like ARM and MIPS—not.
2.1.4 Wide char

This is an attempt to support multi-lingual environment by extending byte to 16-bit. Most well-known
example is Windows NT kernel and win32 functions with W suffix. This is why each Latin character in plain
English text string is interleaved with zero byte. This encoding is called UCS-2 or UTF-16
Usually, wchar_t is synonym to 16-bit short data type.
2.1.5 Signed integer vs unsigned

Some may argue, why unsigned data types exist at first place, since any unsigned number can be rep-
resented as signed. Yes, but absence of sign bit in a value extends its range twice. Hence, signed byte
has range of -128..127, and unsigned one: 0..255. Another benefit of using unsigned data types is self-
documenting: you define a variable which can’t be assigned to negative values.
Unsigned data types are absent in Java, for which it’s criticized. It’s hard to implement cryptographical
algorithms using boolean operations over signed data types.
Values like 0xFFFFFFFF (-1) are used often, mostly as error codes.
2.1.6 Word
Word word is somewhat ambiguous term and usually denotes a data type fitting in GPR. Bytes are practical
for characters, but impractical for other arithmetical calculations.
Hence, many CPUs have GPRs with width of 16, 32 or 64 bits. Even 8-bit CPUs like 8080 and Z80 offer to
work with 8-bit register pairs, each pair forming a 16-bit pseudoregister (BC, DE, HL, etc.). Z80 has some
capability to work with register pairs, and this is, in a sense, some kind of 16-bit CPU emulation.
In general, if a CPU marketed as “n-bit CPU”, this usually means it has n-bit GPRs.
There was a time when hard disks and RAM modules were marketed as having n kilo-words instead of b
kilobytes/megabytes.
For example, Apollo Guidance Computer has 2048 words of RAM. This was a 16-bit computer, so there
was 4096 bytes of RAM.
TX-0 had 64K of 18-bit words of magnetic core memory, i.e., 64 kilo-words.
DECSYSTEM-2060 could have up to 4096 kilowords of solid state memory (i.e., hard disks, tapes, etc).
This was 36-bit computer, so this is 18432 kilobytes or 18 megabytes.
Essentially, why do you need bytes if you have words? Mostly for text strings processing. Words can be
used in almost any other situations.
int in C/C++ is almost always mapped to word. (Except of AMD64 architecture where int is still 32-bit one,
perhaps, for the reason of better portability.)
int is 16-bit on PDP-11 and old MS-DOS compilers. int is 32-bit on VAX, on x86 starting at 80386, etc.
Even more than that, if type declaration for a variable is omitted in C/C++ program, int is used silently by
default. Perhaps, this is inheritance of B programming language5 .
GPR is usually fastest container for variable, faster than packed bit, and sometimes even faster than byte
(because there is no need to isolate a single bit/byte from GPR). Even if you use it as a container for loop
counter in 0..99 range.
5 http://yurichev.com/blog/typeless/
445
Word in assembly language is still 16-bit for x86, because it was so for 16-bit 8086. Double word is 32-bit,
quad word is 64-bit. That’s why 16-bit words are declared using DW in x86 assembly, 32-bit ones using DD
and 64-bit ones using DQ.
Word is 32-bit for ARM, MIPS, etc., 16-bit data types are called half-word there. Hence, double word on
32-bit RISC is 64-bit data type.
GDB has the following terminology: halfword for 16-bit, word for 32-bit and giant word for 64-bit.
16-bit C/C++ environment on PDP-11 and MS-DOS has long data type with width of 32 bits, perhaps, they
meant long word or long int?
32-bit C/C++ environment has long long data type with width of 64 bits.
Now you see why the word word is ambiguous.
Should I use int?
Some people argue that int shouldn’t be used at all, because it ambiguity can lead to bugs. For example,
well-known lzhuf library uses int at one point and everything works fine on 16-bit architecture. But if
ported to architecture with 32-bit int, it can crash: http://yurichev.com/blog/lzhuf/.
Less ambiguous types are defined in stdint.h file: uint8_t, uint16_t, uint32_t, uint64_t, etc.
Some people like Donald E. Knuth proposed6 more sonorous words for these types: byte/wyde/tetra-
byte/octabyte. But these names are less popular than clear terms with inclusion of u (unsigned) character
and number right into the type name.
Word-oriented computers
Despite the ambiguity of the word term, modern computers are still word-oriented: RAM and all levels of
cache are still organized by words, not by bytes. However, size in bytes is used in marketing.
Access to RAM/cache by address aligned by word boundary is often cheaper than non-aligned.
During data structures development, which are supposed to be fast and efficient, one should always take
into consideration length of the word on the CPU to be executed on. Sometimes the compiler will do this
for programmer, sometimes not.
2.1.7 Address register

For those who fostered on 32-bit and/or 64-bit x86, and/or RISC of 90s like ARM, MIPS, PowerPC, it’s natural
that address bus has the same width as GPR or word. Nevertheless, width of address bus can be different
on other architectures.
8-bit Z80 can address 216 bytes, using 8-bit registers pairs or dedicated registers (IX, IY). SP and PC
registers are also 16-bit ones.
Cray-1 supercomputer has 64-bit GPRs, but 24-bit address registers, so it can address 224 (16 megawords
or 128 megabytes). RAM was very expensive in 1970s, and a typical Cray had 1048576 (0x100000) words
of RAM or 8MB. So why to allocate 64-bit register for address or pointer?
8086/8088 CPUs had a really weird addressing scheme: values of two 16-bit registers were summed in a
weird manner resulting in a 20-bit address. Perhaps, this was some kind of toy-level virtualization ( 11.6
on page 986)? 8086 could run several programs (not simultaneously, though).
Early ARM1 has an interesting artifact:
Another interesting thing about the register file is the PC register is missing a few bits.
Since the ARM1 uses 26-bit addresses, the top 6 bits are not used. Because all instructions
are aligned on a 32-bit boundary, the bottom two address bits in the PC are always zero.
These 8 bits are not only unused, they are omitted from the chip entirely.
( http://www.righto.com/2015/12/reverse-engineering-arm1-ancestor-of.html )
Hence, it’s physically not possible to push a value with one of two last bits set into PC register. Nor it’s
possible to set any bits in high 6 bits of PC.
6 http://www-cs-faculty.stanford.edu/~uno/news98.html
446
x86-64 architecture has virtual 64-bit pointers/addresses, but internally, width of address bus is 48 bits
(seems enough to address 256TB of RAM).
2.1.8 Numbers
What are numbers used for?
When you see some number(s) altering in a CPU register, you may be interested in what this number
means. It’s an important skill for a reverse engineer to determine possible data type from a set of changing
numbers.
Boolean
If the number is switching from 0 to 1 and back, most chances that this value has boolean data type.
Loop counter, array index
Variable increasing from 0, like: 0, 1, 2, 3…—a good chance this is a loop counter and/or array index.
Signed numbers
If you see a variable which holds very low numbers and sometimes very high numbers, like 0, 1, 2, 3, and
0xFFFFFFFF, 0xFFFFFFFE, 0xFFFFFFFD, there’s a good chance it is a signed variable in two’s complement
form ( 2.2 on the following page), and last 3 numbers are -1, -2, -3.
32-bit numbers
There are numbers so large, that there is even a special notation which exists to represent them (Knuth’s
up-arrow notation). These numbers are so large so these are not practical for engineering, science and
mathematics.
Almost all engineers and scientists are happy with IEEE 754 double precision floating point, which has
maximal value around 1.8 ⋅ 10308 . (As a comparison, the number of atoms in the observable universe, is
estimated to be between 4 ⋅ 1079 and 4 ⋅ 1081 .)
In fact, upper bound in practical computing is much, much lower. In MS-DOS era 16-bit int was used
almost for everything (array indices, loop counters), while 32-bit long was used rarely.
During advent of x86-64, it was decided for int to stay as 32 bit size integer, because, probably, usage of
64-bit int is even rarer.
I would say, 16-bit numbers in range 0..65535 are probably most used numbers in computing.
Given that, if you see unusually large 32-bit value like 0x87654321, this is a good chance this can be:
• this can still be a 16-bit number, but signed, between 0xFFFF8000 (-32768) and 0xFFFFFFFF (-1).
• address of memory cell (can be checked using memory map feature of debugger).
• packed bytes (can be checked visually).
• bit flags.
• something related to (amateur) cryptography.
• magic number ( 5.6.1 on page 704).
• IEEE 754 floating point number (can also be checked).
Almost same story for 64-bit values.
…so 16-bit int is enough for almost everything?
It’s interesting to note: in [Michael Abrash, Graphics Programming Black Book, 1997 chapter 13] we can
find that there are plenty cases in which 16-bit variables are just enough. In a meantime, Michael Abrash
has a pity that 80386 and 80486 CPUs has so little available registers, so he offers to put two 16-bit values
into one 32-bit register and then to rotate it using ROR reg, 16 (on 80386 and later) (ROL reg, 16 will
also work) or BSWAP (on 80486 and later) instruction.
447
That reminds us Z80 with alternate pack of registers (suffixed with apostrophe), to which CPU can switch
(and then switch back) using EXX instruction.
Size of buffer
When a programmer needs to declare the size of some buffer, values in form of 2x are usually used (512
bytes, 1024, etc.). Values in 2x form are easily recognizable ( 1.28.5 on page 320) in decimal, hexadecimal
and binary base.
But needless to say, programmers are still humans with their decimal culture. And somehow, in DBMS
area, size of textual database fields is often chosen as 10x number, like 100, 200. They just think “Okay,
100 is enough, wait, 200 will be better”. And they are right, of course.
Maximum width of VARCHAR2 data type in Oracle RDBMS is 4000 characters, not 4096.
There is nothing wrong with this, this is just a place where numbers like 10x can be encountered.
Address
It’s always a good idea to keep in mind an approximate memory map of the process you currently debug.
For example, many win32 executables started at 0x00401000, so an address like 0x00451230 is probably
located inside executable section. You’ll see addresses like these in the EIP register.
Stack is usually located somewhere below.
Many debuggers are able to show the memory map of the debuggee, for example: 1.12.3 on page 79.
If a value is increasing by step 4 on 32-bit architecture or by step 8 on 64-bit one, this probably sliding
address of some elements of array.
It’s important to know that win32 doesn’t use addresses below 0x10000, so if you see some number be-
low this constant, this cannot be an address (see also: https://msdn.microsoft.com/en-us/library/
ms810627.aspx).
Anyway, many debuggers can show you if the value in a register can be an address to something. OllyDbg
can also show an ASCII string if the value is an address of it.
Bit field
If you see a value where one (or more) bit(s) are flipping from time to time like 0xABCD1234 → 0xABCD1434
and back, this is probably a bit field (or bitmap).
Packed bytes
When strcmp() or memcmp() copies a buffer, it loads/stores 4 (or 8) bytes simultaneously, so if a string
containing “4321”, and it would be copied to another place, at one point you’ll see 0x31323334 value in
some register. This is 4 packed bytes into a 32-bit value.
2.2 Signed number representations

There are several methods for representing signed numbers, but “two’s complement” is the most popular
one in computers.
Here is a table for some byte values:
448
binary hexadecimal unsigned signed
01111111 0x7f 127 127
01111110 0x7e 126 126
...
00000110 0x6 6 6
00000101 0x5 5 5
00000100 0x4 4 4
00000011 0x3 3 3
00000010 0x2 2 2
00000001 0x1 1 1
00000000 0x0 0 0
11111111 0xff 255 -1
11111110 0xfe 254 -2
11111101 0xfd 253 -3
11111100 0xfc 252 -4
11111011 0xfb 251 -5
11111010 0xfa 250 -6
...
10000010 0x82 130 -126
10000001 0x81 129 -127
10000000 0x80 128 -128
The difference between signed and unsigned numbers is that if we represent 0xFFFFFFFE and 0x00000002
as unsigned, then the first number (4294967294) is bigger than the second one (2). If we represent them
both as signed, the first one becomes −2, and it is smaller than the second (2). That is the reason why
conditional jumps ( 1.18 on page 123) are present both for signed (e.g. JG, JL) and unsigned (JA, JB)
operations.
For the sake of simplicity, this is what one needs to know:
• Numbers can be signed or unsigned.
• C/C++ signed types:
– int64_t (-9,223,372,036,854,775,808 .. 9,223,372,036,854,775,807) (- 9.2.. 9.2 quintillions) or
0x8000000000000000..0x7FFFFFFFFFFFFFFF),
– int (-2,147,483,648..2,147,483,647 (- 2.15.. 2.15Gb) or
0x80000000..0x7FFFFFFF),
– char (-128..127 or 0x80..0x7F),
– ssize_t.
Unsigned:
– uint64_t (0..18,446,744,073,709,551,615 ( 18 quintillions) or 0..0xFFFFFFFFFFFFFFFF),
– unsigned int (0..4,294,967,295 ( 4.3Gb) or 0..0xFFFFFFFF),
– unsigned char (0..255 or 0..0xFF),
– size_t.
• Signed types have the sign in the MSB: 1 means “minus”, 0 means “plus”.
• Promoting to a larger data types is simple: 1.34.5 on page 401.
• Negation is simple: just invert all bits and add 1.
We can keep in mind that a number of inverse sign is located on the opposite side at the same
proximity from zero. The addition of one is needed because zero is present in the middle.
• The addition and subtraction operations work well for both signed and unsigned values. But for mul-
tiplication and division operations, x86 has different instructions: IDIV/IMUL for signed and DIV/MUL
for unsigned.
• Here are some more instructions that work with signed numbers:
CBW/CWD/CWDE/CDQ/CDQE ( .1.6 on page 1017), MOVSX ( 1.23.1 on page 200), SAR ( .1.6 on page 1021).
A table of some negative and positive values ( ?? on page ??) looks like thermometer with Celsius scale.
This is why addition and subtraction works equally well for both signed and unsigned numbers: if the first
addend is represented as mark on thermometer, and one need to add a second addend, and it’s positive,
449
we just shift mark up on thermometer by the value of second addend. If the second addend is negative,
then we shift mark down to absolute value of the second addend.
Addition of two negative numbers works as follows. For example, we need to add -2 and -3 using 16-bit
registers. -2 and -3 is 0xfffe and 0xfffd respectively. If we add these numbers as unsigned, we will get
0xfffe+0xfffd=0x1fffb. But we work on 16-bit registers, so the result is cut off, the first 1 is dropped, 0xfffb
is left, and this is -5. This works because -2 (or 0xfffe) can be represented using plain English like this: “2
lacks in this value up to maximal value in 16-bit register + 1”. -3 can be represented as “…3 lacks in this
value up to …”. Maximal value of 16-bit register + 1 is 0x10000. During addition of two numbers and
cutting off by 216 modulo, 2 + 3 = 5 will be lacking.
2.2.1 Using IMUL over MUL

Example like listing.3.23.2 where two unsigned values are multiplied compiles into listing.3.23.2 where
IMUL is used instead of MUL.
This is important property of both MUL and IMUL instructions. First of all, they both produce 64-bit value
if two 32-bit values are multiplied, or 128-bit value if two 64-bit values are multiplied (biggest possible
product in 32-bit environment is
0xffffffff*0xffffffff=0xfffffffe00000001). But C/C++ standards have no way to access higher half
of result, and a product always has the same size as multiplicands. And both MUL and IMUL instructions
works in the same way if higher half is ignored, i.e., they both generate the same lower half. This is
important property of “two’s complement” way of representing signed numbers.
So C/C++ compiler can use any of these instructions.
But IMUL is more versatile than MUL because it can take any register(s) as source, while MUL requires one
of multiplicands stored in AX/EAX/RAX register. Even more than that: MUL stores result in EDX:EAX pair in
32-bit environment, or RDX:RAX in 64-bit one, so it always calculates the whole result. On contrary, it’s
possible to set a single destination register while using IMUL instead of pair, and then CPU will calculate
only lower half, which works faster [see Torborn Granlund, Instruction latencies and throughput for AMD
and Intel x86 processors7 ).
Given than, C/C++ compilers may generate IMUL instruction more often then MUL.
Nevertheless, using compiler intrinsic, it’s still possible to do unsigned multiplication and get full result.
This is sometimes called extended multiplication. MSVC has intrinsic for this called __emul8 and another
one: _umul1289 . GCC offer __int128 data type, and if 64-bit multiplicands are first promoted to 128-bit
ones, then a product is stored into another __int128 value, then result is shifted by 64 bits right, you’ll get
higher half of result10 .
MulDiv() function in Windows
Windows has MulDiv() function 11 , fused multiply/divide function, it multiplies two 32-bit integers into
intermediate 64-bit value and then divides it by a third 32-bit integer. It is easier than to use two compiler
intrinsic, so Microsoft developers made a special function for it. And it seems, this is busy function, judging
by its usage.
2.2.2 Couple of additions about two’s complement form
Exercise 2-1. Write a program to determine

the ranges of char, short, int, and long
variables, both signed and unsigned, by
printing appropriate values from standard
headers and by direct computation.
Brian W. Kernighan, Dennis M. Ritchie, The C

Programming Language, 2ed, (1988)
7 http://yurichev.com/mirrors/x86-timing.pdf]
8 https://msdn.microsoft.com/en-us/library/d2s81xt0(v=vs.80).aspx
9 https://msdn.microsoft.com/library/3dayytw9%28v=vs.100%29.aspx
10 Example: http://stackoverflow.com/a/13187798
11 https://msdn.microsoft.com/en-us/library/windows/desktop/aa383718(v=vs.85).aspx
450
Getting maximum number of some word
Maximum unsigned number is just a number where all bits are set: 0xFF....FF (this is -1 if the word is
treated as signed integer). So you take a word, set all bits and get the value:
#include <stdio.h>
int main()
{
unsigned int val=~0; // change to "unsigned char" to get maximal value for the unsigned
8-bit byte
// 0-1 will also work, or just -1
printf ("%u\n", val); // ;
This is 4294967295 for 32-bit integer.
Getting minimum number for some signed word
Minimum signed number is encoded as 0x80....00, i.e., most significant bit is set, while others are cleared.
Maximum signed number is encoded in the same way, but all bits are inverted: 0x7F....FF.
Let’s shift a lone bit left until it disappears:
#include <stdio.h>
int main()
{
signed int val=1; // change to "signed char" to find values for signed byte
while (val!=0)
{
printf ("%d %d\n", val, ~val);
val=val<<1;
};
};
Output is:
...
536870912 -536870913
1073741824 -1073741825
-2147483648 2147483647
Two last numbers are minimum and maximum signed 32-bit int respectively.
2.2.3 -1
Now you see that −1 is when all bits are set. Often, you can find the −1 constant in all sorts of code, where
a constant with all bits set are needed, for example, a mask.
For example: 3.18.1 on page 526.
2.3 Integer overflow

I intentionally put this section after the section about signed number representation.
First, take a look at this implementation of itoa() function from [Brian W. Kernighan, Dennis M. Ritchie, The
C Programming Language, 2ed, (1988)]:
void itoa(int n, char s[])
{
int i, sign;
if ((sign = n) < 0) /* record sign */
n = -n; /* make n positive */
i = 0;
do { /* generate digits in reverse order */
s[i++] = n % 10 + '0'; /* get next digit */
} while ((n /= 10) > 0); /* delete it */
451
if (sign < 0)
s[i++] = '-';
s[i] = '\0';
strrev(s);
}
( The full source code: https://beginners.re/current-tree/fundamentals/itoa_KR.c )

It has a subtle bug. Try to find it. You can download source code, compile it, etc. The answer on the next
page.
452
From [Brian W. Kernighan, Dennis M. Ritchie, The C Programming Language, 2ed, (1988)]:
Exercise 3-4. In a two’s complement number representation, our version of itoa does not
handle the largest negative number, that is, the value of n equal to −(2wordsize−1 ). Explain
why not. Modify it to print that value correctly, regardless of the machine on which it runs.
The answer is: the function cannot process largest negative number (INT_MIN or 0x80000000 or -2147483648)
correctly.
How to change sign? Invert all bits and add 1. If to invert all bits in INT_MIN value (0x80000000), this
is 0x7fffffff. Add 1 and this is 0x80000000 again. So changing sign has no effect. This is an important
artifact of two’s complement system.
Further reading:
• blexim – Basic Integer Overflows12
• Yannick Moy, Nikolaj Bjørner, and David Sielaff – Modular Bug-finding for Integer Overflows in the
Large: Sound, Efficient, Bit-precise Static Analysis13
2.4 AND
2.4.1 Checking if a value is on 2n boundary

If you need to check if your value is divisible by 2n number (like 1024, 4096, etc.) without remainder, you
can use a % operator in C/C++, but there is a simpler way. 4096 is 0x1000, so it always has 4 ∗ 3 = 12 lower
bits cleared.
What you need is just:
if (value&0xFFF)
{
printf ("value is not divisible by 0x1000 (or 4096)\n");
printf ("by the way, remainder is %d\n", value&0xFFF);
}
else
printf ("value is divisible by 0x1000 (or 4096)\n");
In other words, this code checks if there are any bit set among lower 12 bits. As a side effect, lower 12
bits is always a remainder from division a value by 4096 (because division by 2n is merely a right shift,
and shifted (and dropped) bits are bits of remainder).
Same story if you want to check if the number is odd or even:
if (value&1)
// odd
else
// even
This is merely the same as if to divide by 2 and get 1-bit remainder.
2.4.2 KOI-8R Cyrillic encoding

It was a time when 8-bit ASCII table wasn’t supported by some Internet services, including email. Some
supported, some others—not.
It was also a time, when non-Latin writing systems used second half of 8-bit ASCII table to accommodate
non-Latin characters. There were several popular Cyrillic encodings, but KOI-8R (devised by Andrey “ache”
Chernov) is somewhat unique in comparison with others.
12 http://phrack.org/issues/60/10.html
13 https://yurichev.com/mirrors/SMT/z3prefix.pdf
453
Figure 2.1: KOI8-R table
Someone may notice that Cyrillic characters are allocated almost in the same sequence as Latin ones.
This leads to one important property: if all 8th bits in Cyrillic text encoded in KOI-8R are to be reset, a
text transforms into transliterated text with Latin characters in place of Cyrillic. For example, Russian
sentence:
Мой дядя самых честных правил, Когда не в шутку занемог, Он уважать себя
заставил, И лучше выдумать не мог.
…if encoded in KOI-8R and then 8th bit stripped, transforms into:
mOJ DQDQ SAMYH ÊSTNYH PRAWIL, kOGDA NE W [UTKU ZANEMOG, oN UWAVATX SEBQ
ZASTAWIL, i LU^[E WYDUMATX NE MOG.
…perhaps this is not very appealing æsthetically, but this text is still readable to Russian language natives.
Hence, Cyrillic text encoded in KOI-8R, passed through an old 7-bit service will survive into transliterated,
but still readable text.
Stripping 8th bit is automatically transposes any character from the second half of the (any) 8-bit ASCII
table to the first one, into the same place (take a look at red arrow right of table). If the character has
already been placed in the first half (i.e., it has been in standard 7-bit ASCII table), it’s not transposed.
Perhaps, transliterated text is still recoverable, if you’ll add 8th bit to the characters which were seems
transliterated.
Drawback is obvious: Cyrillic characters allocated in KOI-8R table are not in the same sequence as in
Russian/Bulgarian/Ukrainian/etc. alphabet, and this isn’t suitable for sorting, for example.
2.5 AND and OR as subtraction and addition

2.5.1 ZX Spectrum ROM text strings
Those who once investigated ZX Spectrum ROM internals, probably noticed that the last symbol of each
text string is seemingly absent.
454
Figure 2.2: Part of ZX Spectrum ROM
There are present, in fact.

Here is excerpt of ZX Spectrum 128K ROM disassembled:
L048C: DEFM "MERGE erro" ; Report 'a'.
DEFB 'r'+$80
L0497: DEFM "Wrong file typ" ; Report 'b'.
DEFB 'e'+$80
L04A6: DEFM "CODE erro" ; Report 'c'.
DEFB 'r'+$80
L04B0: DEFM "Too many bracket" ; Report 'd'.
DEFB 's'+$80
L04C1: DEFM "File already exist" ; Report 'e'.
DEFB 's'+$80
( http://www.matthew-wilson.net/spectrum/rom/128_ROM0.html )
Last character has most significant bit set, which marks string end. Presumably, it was done to save some
space? Old 8-bit computers have very tight environment.
Characters of all messages are always in standard 7-bit ASCII table, so it’s guaranteed 7th bit is never
used for characters.
To print such string, we must check MSB of each byte, and if it’s set, we must clear it, then print character,
and then stop. Here is a C example:
unsigned char hw[]=
{
'H',
'e',
'l',
'l',
'o'|0x80
};
void print_string()
{
for (int i=0; ;i++)
{
if (hw[i]&0x80) // check MSB
{
// clear MSB
// (in other words, clear all, but leave 7 lower bits intact)
printf ("%c", hw[i] & 0x7F);
// stop
455
break;
};
printf ("%c", hw[i]);
};
};
Now what is interesting, since 7th bit is the most significant bit (in byte), we can check it, set it and remove
it using arithmetical operations instead of logical.
I can rewrite my C example:
unsigned char hw[]=
{
'H',
'e',
'l',
'l',
'o'+0x80
};
void print()
{
for (int i=0; ;i++)
{
// hw[] must have 'unsigned char' type
if (hw[i] >= 0x80) // check for MSB
{
printf ("%c", hw[i]-0x80); // clear MSB
// stop
break;
};
printf ("%c", hw[i]);
};
};
By default, char is signed type in C/C++, so to compare it with variable like 0x80 (which is negative (−128)
if treated as signed), we must treat each character in text message as unsigned.
Now if 7th bit is set, the number is always larger or equal to 0x80. If 7th bit is clear, the number is always
smaller than 0x80.
Even more than that: if 7th bit is set, it can be cleared by subtracting 0x80, nothing else. If it’s not set
beforehand, however, subtracting will destruct other bits.
Likewise, if 7th bit is clear, it’s possible to set it by adding 0x80. But if it’s set beforehand, addition
operation will destruct some other bits.
In fact, this is valid for any bit. If the 4th bit is clear, you can set it just by adding 0x10: 0x100+0x10 =
0x110. If the 4th bit is set, you can clear it by subtracting 0x10: 0x1234-0x10 = 0x1224.
It works, because carry isn’t happened during addition/subtraction. It will, however, happen, if the bit is
already set there before addition, or absent before subtraction.
Likewise, addition/subtraction can be replaced using OR/AND operation if two conditions are met: 1) you
want to add/subtract by a number in form of 2n ; 2) this bit in source value is clear/set.
For example, addition of 0x20 is the same as ORing value with 0x20 under condition that this bit is clear
before: 0x1204|0x20 = 0x1204+0x20 = 0x1224.
Subtraction of 0x20 is the same as ANDing value with 0x20 (0x....FFDF), but if this bit is set before:
0x1234&(~0x20) = 0x1234&0xFFDF = 0x1234-0x20 = 0x1214.
Again, it works because carry not happened when you add 2n number and this bit isn’t set before.
This property of boolean algebra is important, worth understanding and keeping it in mind.
Another example in this book: 3.19.3 on page 536.
2.6 XOR (exclusive OR)

XOR is widely used when one needs just to flip specific bit(s). Indeed, the XOR operation applied with 1
effectively inverts a bit:
456
input A input B output
0 0 0
0 1 1
1 0 1
1 1 0
And vice-versa, the XOR operation applied with 0 does nothing, i.e., it’s an idle operation. This is a very
important property of the XOR operation and it’s highly recommended to memorize it.
2.6.1 Logical difference

14
In Cray-1 supercomputer (1976-1977) manual , you can find XOR instruction was called logical differ-
ence.
Indeed, XOR(a,b)=1 if a!=b.
2.6.2 Everyday speech

XOR operation present in common everyday speech. When someone asks “please buy apples or bananas”,
this usually means “buy the first object or the second, but not both”—this is exactly exclusive OR, because
logical OR would mean “both objects are also fine”.
Some people suggest “and/or” should be used in everyday speech to make emphasis that logical OR is
used instead of exclusive OR: https://en.wikipedia.org/wiki/And/or.
2.6.3 Encryption
XOR is heavily used in both amateur ( 9.1 on page 908) and real encryption (at least in Feistel network).
XOR is very useful here because: cipher_text = plain_text ⊕ key and then: (plain_text ⊕ key) ⊕ key = plain_text.
2.6.4 RAID4
RAID4 offers a very simple method to protect hard disks. For example, there are several disks (D1 , D2 , D3 ,
etc.) and one parity disk (P ). Each bit/byte written to parity disk is calculated and written on-fly:
P = D1 ⊕ D2 ⊕ D3 (2.1)
If any of disks is failed, for example, D2 , it’s restored using the very same way:
D2 = D1 ⊕ P ⊕ D3 (2.2)
If parity disk failed, it is restored using 2.1 way. If two of any disks are failed, then it wouldn’t be possible
to restore both.
RAID5 is more advanced, but this XOR property is still exploited there.
That’s why RAID controllers has hardware “XOR accelerators” helping to XOR large chunks of written data
on-fly. When computers get faster and faster, it now can be done at software level, using SIMD.
2.6.5 XOR swap algorithm

Hard to believe, but this code swaps values in EAX and EBX without aid of any other additional register or
memory cell:
xor eax, ebx
xor ebx, eax
xor eax, ebx
Let’s find out, how it works. First, we will rewrite it to step aside from x86 assembly language:
14 http://www.bitsavers.org/pdf/cray/CRAY-1/HR-0004-CRAY_1_Hardware_Reference_Manual-PRELIMINARY-1975.OCR.
pdf
457
X = X XOR Y
Y = Y XOR X
X = X XOR Y
What X and Y has at each step? Just keep in mind the simple rule: (X ⊕ Y ) ⊕ Y = X for any values of X and
Y.
Let’s see, X after 1st step has X⊕Y ; Y after 2nd step has Y ⊕(X⊕Y ) = X; X after 3rd step has (X⊕Y )⊕X = Y .
Hard to say if anyone should use this trick, but it servers as a good demonstration example of XOR prop-
erties.
Wikipedia article (https://en.wikipedia.org/wiki/XOR_swap_algorithm) has also yet another expla-
nation: addition and subtraction operations can be used instead of XOR:
X = X + Y
Y = X - Y
X = X - Y
Let’s see: X after 1st step has X +Y ; Y after 2nd step has X +Y −Y = X; X after 3rd step has X +Y −X = Y .
2.6.6 XOR linked list

Doubly linked list is a list in which each element has link to the previous element and to the next one.
Hence, it’s very easy to traverse list backwards or forward. std::list in C++ implements doubly linked
list which also is examined in this book: 3.21.4 on page 566.
So each element has two pointers. Is it possible, perhaps in environment of small RAM footprint, to
preserve all functionality with one pointer instead of two? Yes, if it a value of prev ⊕ next will be stored in
this memory cell, which is usually called “link”.
Maybe, we could say that address to the previous element is “encrypted” using address of next element
and otherwise: next element address is “encrypted” using previous element address.
When we traverse this list forward, we always know address of the previous element, so we can “decrypt”
this field and get address of the next element. Likewise, it’s possible to traverse this list backwards,
“decrypting” this field using next element’s address.
But it’s not possible to find address of previous or next element of some specific element without knowing
address of the first one.
Couple of things to complete this solution: first element will have address of next element without any
XOR-ing, last element will have address of previous element without any XOR-ing.
Now let’s sum it up. This is example of doubly linked list of 5 elements. Ax is address of element.
address link field contents
A0 A1
A1 A0 ⊕ A2
A2 A1 ⊕ A3
A3 A2 ⊕ A4
A4 A3
And again, hard to say if anyone should use this tricky hacks, but this is also a good demonstration of
XOR properties. As with XOR swap algorithm, Wikipedia article about it also offers way to use addition or
subtraction instead of XOR: https://en.wikipedia.org/wiki/XOR_linked_list.
2.6.7 Switching value trick

15
... found in Jorg Arndt — Matters Computational / Ideas, Algorithms, Source Code .
You want a variable to be switching between 123 and 456. You may write something like:
if (a==123)
a=456;
else
a=123;
15 https://www.jjj.de/fxt/fxtbook.pdf
458
But this can be done using a single operation:
#include <stdio.h>
int main()
{
int a=123;
#define C 123^456
a=a^C;
printf ("%d\n", a);
a=a^C;
printf ("%d\n", a);
a=a^C;
printf ("%d\n", a);
};
It works because 123 ⊕ 123 ⊕ 456 = 0 ⊕ 456 = 456 and 456 ⊕ 123 ⊕ 456 = 456 ⊕ 456 ⊕ 123 = 0 ⊕ 123 = 123.
One can argue, worth it using or not, especially keeping in mind code readability. But this is yet another
demonstration of XOR properties.
2.6.8 Zobrist hashing / tabulation hashing

If you work on a chess engine, you traverse a game tree many times per second, and often, you can
encounter the same position, which has already been processed.
So you have to use a method to store already calculated positions somewhere. But chess position can
require a lot of memory, and a hash function would be used instead.
Here is a way to compress a chess position into 64-bit value, called Zobrist hashing:
// we have 8*8 board and 12 pieces (6 for white side and 6 for black)
uint64_t table[12][8][8]; // filled with random values
int position[8][8]; // for each square on board. 0 - no piece. 1..12 - piece
uint64_t hash;
for (int row=0; row<8; row++)

for (int col=0; col<8; col++)
{
int piece=position[row][col];
if (piece!=0)
hash=hash^table[piece][row][col];
};
return hash;
Now the most interesting part: if the next (modified) chess position differs only by one (moved) piece,
you don’t need to recalculate hash for the whole position, all you need is:
hash=...; // (already calculated)
// subtract information about the old piece:

hash=hash^table[old_piece][old_row][old_col];
// add information about the new piece:

hash=hash^table[new_piece][new_row][new_col];
2.6.9 By the way

The usual OR also sometimes called inclusive OR (or even IOR), as opposed to exclusive OR. One place is
operator Python’s library: it’s called operator.ior here.
459
2.6.10 AND/OR/XOR as MOV
OR reg, 0xFFFFFFFF sets all bits to 1, hence, no matter what has been in register before, it will be set
to −1. OR reg, -1 is shorter than MOV reg, -1, so MSVC uses OR instead the latter, for example: 3.18.1
on page 526.
Likewise, AND reg, 0 always resets all bits, hence, it acts like MOV reg, 0.
XOR reg, reg, no matter what has been in register beforehand, resets all bits, and also acts like MOV reg,
0.
2.7 Population count

POPCNT instruction is population count (AKA Hamming weight). It just counts number of bits set in an input
value.
As a side effect, POPCNT instruction (or operation) can be used to determine, if the value has 2n form.
Since, 2n number always has just one single bit, POPCNT’s result will always be just 1.
For example, I once wrote a base64 strings scanner for hunting something interesting in binary files16 .
And there is a lot of garbage and false positives, so I add an option to filter out data blocks which has size
of 2n bytes (i.e., 256 bytes, 512, 1024, etc.). The size of block is checked just like this:
if (popcnt(size)==1)
// OK
...
The instruction is also known as “NSA17 instruction” due to rumors:
This branch of cryptography is fast-paced and very politically charged. Most designs
are secret; a majority of military encryptions systems in use today are based on LFSRs. In
fact, most Cray computers (Cray 1, Cray X-MP, Cray Y-MP) have a rather curious instruction
generally known as “population count.” It counts the 1 bits in a register and can be used both
to efficiently calculate the Hamming distance between two binary words and to implement a
vectorized version of a LFSR. I’ve heard this called the canonical NSA instruction, demanded
by almost all computer contracts.
[Bruce Schneier, Applied Cryptography, (John Wiley & Sons, 1994)]
2.8 Endianness
The endianness is a way of representing values in memory.
2.8.1 Big-endian
The 0x12345678 value is represented in memory as:
address in memory byte value
+0 0x12
+1 0x34
+2 0x56
+3 0x78
Big-endian CPUs include Motorola 68k, IBM POWER.
2.8.2 Little-endian
The 0x12345678 value is represented in memory as:
16 https://github.com/DennisYurichev/base64scanner
17 National Security Agency
460
address in memory byte value
+0 0x78
+1 0x56
+2 0x34
+3 0x12
Little-endian CPUs include Intel x86. One important example of little-endian using in this book is: ?? on
page ??.
2.8.3 Example
18
Let’s take big-endian MIPS Linux installed and ready in QEMU .
And let’s compile this simple example:
#include <stdio.h>
int main()
{
int v;
v=123;
printf ("%02X %02X %02X %02X\n",

*(char*)&v,
*(((char*)&v)+1),
*(((char*)&v)+2),
*(((char*)&v)+3));
};
After running it we get:

root@debian-mips:~# ./a.out
00 00 00 7B
That is it. 0x7B is 123 in decimal. In little-endian architectures, 7B is the first byte (you can check on x86
or x86-64), but here it is the last one, because the highest byte goes first.
That’s why there are separate Linux distributions for MIPS (“mips” (big-endian) and “mipsel” (little-endian)).
It is impossible for a binary compiled for one endianness to work on an OS with different endianness.
There is another example of MIPS big-endiannes in this book: 1.30.4 on page 362.
2.8.4 Bi-endian
CPUs that may switch between endianness are ARM, PowerPC, SPARC, MIPS, IA6419 , etc.
2.8.5 Converting data

The BSWAP instruction can be used for conversion.
TCP/IP network data packets use the big-endian conventions, so that is why a program working on a little-
endian architecture has to convert the values. The htonl() and htons() functions are usually used.
In TCP/IP, big-endian is also called “network byte order”, while byte order on the computer “host byte
order”. “host byte order” is little-endian on Intel x86 and other little-endian architectures, but it is big-
endian on IBM POWER, so htonl() and htons() don’t shuffle any bytes on the latter.
2.9 Memory
There are 3 main types of memory:
18 Available for download here: http://go.yurichev.com/17008
19 Intel Architecture 64 (Itanium)
461
• Global memory AKA “static memory allocation”. No need to allocate explicitly, the allocation is
performed just by declaring variables/arrays globally. These are global variables, residing in the
data or constant segments. They are available globally (hence, considered as an anti-pattern). Not
convenient for buffers/arrays, because they must have a fixed size. Buffer overflows that occur here
usually overwrite variables or buffers residing next to them in memory. There’s an example in this
book: 1.12.3 on page 76.
• Stack AKA “allocate on stack”. The allocation is performed just by declaring variables/arrays locally
in the function. These are usually local variables for the function. Sometimes these local variable
are also available to descending functions (to callee functions, if caller passes a pointer to a variable
to the callee to be executed). Allocation and deallocation are very fast, it just SP needs to be shifted.
But they’re also not convenient for buffers/arrays, because the buffer size has to be fixed, unless
alloca() ( 1.9.2 on page 34) (or a variable-length array) is used. Buffer overflows usually overwrite
important stack structures: 1.26.2 on page 271.
• Heap AKA “dynamic memory allocation”. Allocation/deallocation is performed by calling
malloc()/free() or new/delete in C++. This is the most convenient method: the block size may
be set at runtime.
Resizing is possible (using realloc()), but can be slow. This is the slowest way to allocate memory:
the memory allocator must support and update all control structures while allocating and deallocating.
Buffer overflows usually overwrite these structures. Heap allocations are also source of memory leak
problems: each memory block has to be deallocated explicitly, but one may forget about it, or do it
incorrectly.
Another problem is the “use after free”—using a memory block after free() has been called on it,
which is very dangerous.
Example in this book: 1.30.2 on page 344.
2.10 CPU
2.10.1 Branch predictors
Some latest compilers try to get rid of conditional jump instructions. Examples in this book are: 1.18.1 on
page 134, 1.18.3 on page 142, 1.28.5 on page 328.
This is because the branch predictor is not always perfect, so the compilers try to do without conditional
jumps, if possible.
Conditional instructions in ARM (like ADRcc) are one way, another one is the CMOVcc x86 instruction.
2.10.2 Data dependencies

Modern CPUs are able to execute instructions simultaneously (OOE20 ), but in order to do so, the results of
one instruction in a group must not influence the execution of others. Hence, the compiler endeavors to
use instructions with minimal influence on the CPU state.
That’s why the LEA instruction is so popular, because it does not modify CPU flags, while other arithmetic
instructions does.
2.11 Hash functions

A very simple example is CRC32, an algorithm that provides “stronger” checksum for integrity checking
purposes. It is impossible to restore the original text from the hash value, it has much less information:
But CRC32 is not cryptographically secure: it is known how to alter a text in a way that the resulting CRC32
hash value will be the one we need. Cryptographic hash functions are protected from this.
MD5, SHA1, etc. are such functions and they are widely used to hash user passwords in order to
store them in a database. Indeed: an Internet forum database may not contain user passwords (a stolen
database can compromise all users’ passwords) but only hashes (so a cracker can’t reveal the passwords).
Besides, an Internet forum engine does not need to know your password exactly, it needs only to check
if its hash is the same as the one in the database, and give you access if they match. One of the simplest
20 Out-of-Order Execution
462
password cracking methods is just to try hashing all possible passwords in order to see which matches
the resulting value that we need. Other methods are much more complex.
2.11.1 How do one-way functions work?

A one-way function is a function which is able to transform one value into another, while it is impossible
(or very hard) to reverse it. Some people have difficulties while understanding how this is possible at all.
Here is a simple demonstration.
We have a vector of 10 numbers in range 0..9, each is present only once, for example:
4 6 0 1 3 5 7 8 9 2
The algorithm for the simplest possible one-way function is:

• take the number at zeroth position (4 in our case);
• take the number at first position (6 in our case);
• swap numbers at positions of 4 and 6.
Let’s mark the numbers at positions 4 and 6:
4 6 0 1 3 5 7 8 9 2
^ ^
Let’s swap them and we get this result:

4 6 0 1 7 5 3 8 9 2
While looking at the result, and even if we know the algorithm, we can’t know unambiguously the initial
state, because the first two numbers could be 0 and/or 1, and then they could participate in the swapping
procedure.
This is an utterly simplified example for demonstration. Real one-way functions are much more com-
plex.
463
Chapter 3
Slightly more advanced examples
3.1 Zero register

x86 architecture lacks zero register, unlike MIPS and ARM. However, it’s often a case, when a compiler
assigns a zero to a register, that will hold it upon the function’s exit.
This is a case of Mahjong game from Windows 7 x86. Zeroed EBX (at 0x010281BF) is used to initialize
local variables, to pass zero argument into other functions and to compare values against it.
Listing 3.1: Mahjong.exe from Windows 7 x86

.text:010281AE sub_10281AE proc near ; CODE XREF: sub_1028790+4FFp
.text:010281AE ; sub_102909A+357p ...
.text:010281AE
.text:010281AE var_34 = dword ptr -34h
.text:010281AE var_2C = dword ptr -2Ch
.text:010281AE var_1C = dword ptr -1Ch
.text:010281AE var_4 = dword ptr -4
.text:010281AE arg_0 = dword ptr 8
.text:010281AE arg_4 = byte ptr 0Ch
.text:010281AE
.text:010281AE push 28h
.text:010281B0 mov eax, offset __ehhandler$?⤦
Ç enable_segment@_Helper@_Concurrent_vector_base_v4@details@Concurrency@@SAIAAV234@II@Z
.text:010281B5 call __EH_prolog3
.text:010281BA mov edi, ecx
.text:010281BC mov esi, [ebp+arg_0]
.text:010281BF xor ebx, ebx ; *
.text:010281C1 mov [ebp+var_10], ebx ; *
.text:010281C4 cmp [esi], ebx ; *
.text:010281C6 jbe short loc_10281E8
.text:010281C8
.text:010281C8 loc_10281C8: ; CODE XREF: sub_10281AE+38j
.text:010281C8 mov eax, [esi+0Ch]
.text:010281CB mov ecx, [ebp+var_10]
.text:010281CE push dword ptr [eax+ecx*4]
.text:010281D1 call sub_10506C9
.text:010281D6 mov eax, [ebp+var_10]
.text:010281D9 pop ecx
.text:010281DA mov ecx, [esi+0Ch]
.text:010281DD mov [ecx+eax*4], ebx ; *
.text:010281E0 inc eax
.text:010281E1 mov [ebp+var_10], eax
.text:010281E4 cmp eax, [esi]
.text:010281E6 jb short loc_10281C8
.text:010281E8
464
.text:010281E8 loc_10281E8: ; CODE XREF: sub_10281AE+18j
.text:010281E8 mov [esi], ebx ; *
.text:010281EA mov [edi+14h], ebx ; *
.text:010281ED mov [ebp+var_34], ebx ; *
.text:010281F0 mov [ebp+var_30], ebx ; *
.text:010281F3 mov [ebp+var_2C], 10h
.text:010281FA mov [ebp+var_28], ebx ; *
.text:010281FD mov [ebp+var_4], ebx ; *
.text:01028200 mov [ebp+arg_0], ebx ; *
.text:01028203 cmp [edi+0B0h], ebx ; *
.text:01028209 jbe loc_10282C3
.text:0102820F
.text:0102820F loc_102820F: ; CODE XREF: sub_10281AE+10Fj
.text:0102820F mov eax, [edi+0BCh]
.text:01028215 mov ecx, [ebp+arg_0]
.text:01028218 mov eax, [eax+ecx*4]
.text:0102821B mov [ebp+var_14], eax
.text:0102821E cmp eax, ebx ; *
.text:01028220 jz loc_10282A6
.text:01028226 push ebx ; *
.text:01028228 mov ecx, edi
.text:0102822A call sub_1026B3D
.text:0102822F test al, al
.text:01028231 jz short loc_10282A6
.text:01028233 mov [ebp+var_24], ebx ; *
.text:01028239 mov [ebp+var_1C], 10h
.text:0102824A push eax
.text:0102824B push [ebp+var_14]
.text:0102824E mov ecx, edi
.text:01028250 mov byte ptr [ebp+var_4], 1
.text:01028254 call sub_1026E4F
.text:0102825C cmp [ebp+var_24], ebx ; *
.text:0102825F jbe short loc_102829B
.text:01028261
.text:01028261 loc_1028261: ; CODE XREF: sub_10281AE+EBj
.text:01028261 push 0Ch ; Size
.text:01028263 call sub_102E741
.text:01028268 pop ecx
.text:01028269 cmp eax, ebx ; *
.text:0102826B jz short loc_1028286
.text:0102826D mov edx, [ebp+var_10]
.text:01028270 mov ecx, [ebp+var_18]
.text:01028273 mov ecx, [ecx+edx*4]
.text:01028276 mov edx, [ebp+var_14]
.text:01028279 mov edx, [edx+4]
.text:0102827C mov [eax], edx
.text:0102827E mov [eax+4], ecx
.text:01028281 mov [eax+8], ebx ; *
.text:01028284 jmp short loc_1028288
.text:01028286 ; ---------------------------------------------------------------------------
.text:01028286
.text:01028286 loc_1028286: ; CODE XREF: sub_10281AE+BDj
.text:01028286 xor eax, eax
.text:01028288
.text:01028288 loc_1028288: ; CODE XREF: sub_10281AE+D6j
.text:01028289 mov ecx, esi
.text:0102828B call sub_104922B
.text:01028290 inc [ebp+var_10]
.text:01028293 mov eax, [ebp+var_10]
.text:01028296 cmp eax, [ebp+var_24]
.text:01028299 jb short loc_1028261
.text:0102829B
465
.text:0102829B loc_102829B: ; CODE XREF: sub_10281AE+B1j
.text:0102829B lea ecx, [ebp+var_24]
.text:0102829E mov byte ptr [ebp+var_4], bl
.text:010282A1 call sub_10349DB
.text:010282A6
.text:010282A6 loc_10282A6: ; CODE XREF: sub_10281AE+72j
.text:010282A6 ; sub_10281AE+83j
.text:010282A6 push [ebp+arg_0]
.text:010282A9 lea ecx, [ebp+var_34]
.text:010282AC call sub_104922B
.text:010282B1 inc [ebp+arg_0]
.text:010282B4 mov eax, [ebp+arg_0]
.text:010282B7 cmp eax, [edi+0B0h]
.text:010282BD jb loc_102820F
.text:010282C3
.text:010282C3 loc_10282C3: ; CODE XREF: sub_10281AE+5Bj
.text:010282C3 cmp [ebp+arg_4], bl
.text:010282C6 jz short loc_1028337
.text:010282C8 mov eax, dword_1088AD8
.text:010282CD mov esi, ds:EnableMenuItem
.text:010282D3 mov edi, 40002
.text:010282D8 cmp [eax+8], ebx ; *
.text:010282DB jnz short loc_10282EC
.text:010282DD push 3 ; uEnable
.text:010282DF push edi ; uIDEnableItem
.text:010282E0 push hMenu ; hMenu
.text:010282E6 call esi ; EnableMenuItem
.text:010282E8 push 3
.text:010282EA jmp short loc_10282F7
.text:010282EC ; ---------------------------------------------------------------------------
.text:010282EC
.text:010282EC loc_10282EC: ; CODE XREF: sub_10281AE+12Dj
.text:010282EC push ebx ; *
.text:010282ED push edi ; uIDEnableItem
.text:010282EE push hMenu ; hMenu
.text:010282F4 call esi ; EnableMenuItem
.text:010282F6 push ebx ; *
.text:010282F7
.text:010282F7 loc_10282F7: ; CODE XREF: sub_10281AE+13Cj
.text:010282F7 push edi ; uIDEnableItem
.text:010282F8 push hmenu ; hMenu
.text:010282FE call esi ; EnableMenuItem
.text:01028300 mov ecx, dword_1088AD8
.text:01028306 call sub_1020402
.text:0102830B mov edi, 40001
.text:01028310 test al, al
.text:01028312 jz short loc_1028321
.text:01028314 push ebx ; *
.text:01028315 push edi ; uIDEnableItem
.text:01028316 push hMenu ; hMenu
.text:0102831C call esi ; EnableMenuItem
.text:0102831E push ebx ; *
.text:0102831F jmp short loc_102832E
.text:01028321 ; ---------------------------------------------------------------------------
.text:01028321
.text:01028321 loc_1028321: ; CODE XREF: sub_10281AE+164j
.text:01028321 push 3 ; uEnable
.text:01028323 push edi ; uIDEnableItem
.text:01028324 push hMenu ; hMenu
.text:0102832A call esi ; EnableMenuItem
.text:0102832C push 3 ; uEnable
.text:0102832E
.text:0102832E loc_102832E: ; CODE XREF: sub_10281AE+171j
.text:0102832E push edi ; uIDEnableItem
.text:0102832F push hmenu ; hMenu
.text:01028335 call esi ; EnableMenuItem
.text:01028337
.text:01028337 loc_1028337: ; CODE XREF: sub_10281AE+118j
.text:01028337 lea ecx, [ebp+var_34]
.text:0102833A call sub_10349DB
466
.text:0102833F call __EH_epilog3
.text:01028344 retn 8
.text:01028344 sub_10281AE endp
By the way, IDA can rename a register inside a function (press n):
Listing 3.2: Mahjong.exe from Windows 7 x86

.text:010281AE sub_10281AE proc near
...
.text:010281AE zero = ebx
...
.text:010281BF xor zero, zero ; *

.text:010281C1 mov [ebp+var_10], zero ; *
.text:010281C4 cmp [esi], zero ; *
.text:010281C6 jbe short loc_10281E8
...
3.2 Double negation

A popular way1 to convert non-zero value into 1 (or boolean true) and zero value into 0 (or boolean false)
is !!statement:
int convert_to_bool(int a)
{
return !!a;
};
Optimizing GCC 5.4 x86:

convert_to_bool:
xor eax, eax
test edx, edx
setne al
ret
XOR always clears return value in EAX, even in case if SETNE will not trigger. I.e., XOR sets default return
value to zero.
If the input value is not equal to zero (-NE suffix in SET instruction), 1 is set to AL, otherwise AL isn’t
touched.
Why SETNE operates on low 8-bit part of EAX register? Because the matter is just in the last bit (0 or 1),
while other bits are cleared by XOR.
Therefore, that C/C++ code could be rewritten like this:
{
if (a!=0)
return 1;
else
return 0;
};
…or even:
{
if (a)
return 1;
1 This way is also controversial, because it leads to hard-to-read code
467
else
return 0;
};
Compilers targeting CPUs lacking instruction similar to SET, in this case, generates branching instructions,
etc.
3.3 const correctness

This is undeservedly underused feature of many programming languages. Read here about its importance:
1, 2.
Ideally, everything you don’t modify should have const modifier.
Interestingly, how const correctness is implemented at low level. There are no runtime checks of local
const variables and function arguments (only compile-time checks). But global variables of such a type
are to be allocated in read-only data segments.
This is example is to be crashed, because if compiled by MSVC for win32, the a global variable is allocated
in .rdata read-only segment:
const a=123;
void f(int *i)

{
*i=11; // crash
};
int main()
{
f(&a);
return a;
};
Anonymous (not linked to a variable name) C strings also have const char* type. You can’t modify them:
#include <string.h>
#include <stdio.h>
void alter_string(char *s)

{
strcpy (s, "Goodbye!");
printf ("Result: %s\n", s);
};
int main()
{
alter_string ("Hello, world!\n");
};
This code will crash on Linux (“segmentation fault”) and on Windows if compiled by MinGW.
GCC for Linux places all text strings info .rodata data segment, which is explicitly read-only (“read only
data”):
$ objdump -s 1
...
Contents of section .rodata:

400600 01000200 52657375 6c743a20 25730a00 ....Result: %s..
400610 48656c6c 6f2c2077 6f726c64 210a00 Hello, world!..
When the alter_string() function tries to write there, exception occurred.

Things are different in the code generated by MSVC, strings are located in .data segment, which has no
READONLY flag. MSVC’s developers misstep?
468
C:\...>objdump -s 1.exe
...
Contents of section .data:

40b000 476f6f64 62796521 00000000 52657375 Goodbye!....Resu
40b010 6c743a20 25730a00 48656c6c 6f2c2077 lt: %s..Hello, w
40b020 6f726c64 210a0000 00000000 00000000 orld!...........
40b030 01000000 00000000 c0cb4000 00000000 ..........@.....
...
C:\...>objdump -x 1.exe
...
Sections:
Idx Name Size VMA LMA File off Algn
0 .text 00006d2a 00401000 00401000 00000400 2**2
CONTENTS, ALLOC, LOAD, READONLY, CODE
1 .rdata 00002262 00408000 00408000 00007200 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
2 .data 00000e00 0040b000 0040b000 00009600 2**2
CONTENTS, ALLOC, LOAD, DATA
3 .reloc 00000b98 0040e000 0040e000 0000a400 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
However, MinGW hasn’t this fault and allocates text strings in .rdata segment.
3.3.1 Overlapping const strings

The fact that an anonymous C-string has const type ( 1.5.1 on page 9), and that C-strings allocated in
constants segment are guaranteed to be immutable, has an interesting consequence: the compiler may
use a specific part of the string.
Let’s try this example:
#include <stdio.h>
int f1()
{
printf ("world\n");
}
int f2()
{
printf ("hello world\n");
}
int main()
{
f1();
f2();
}
Common C/C++-compilers (including MSVC) allocate two strings, but let’s see what GCC 4.8.1 does:
Listing 3.3: GCC 4.8.1 + IDA listing

f1 proc near
s = dword ptr -1Ch
sub esp, 1Ch

mov [esp+1Ch+s], offset s ; "world\n"
call _puts
add esp, 1Ch
retn
f1 endp
469
f2 proc near
s = dword ptr -1Ch
sub esp, 1Ch

mov [esp+1Ch+s], offset aHello ; "hello "
call _puts
add esp, 1Ch
retn
f2 endp
aHello db 'hello '

s db 'world',0xa,0
Indeed: when we print the “hello world” string these two words are positioned in memory adjacently and
puts() called from f2() function is not aware that this string is divided. In fact, it’s not divided; it’s
divided only virtually, in this listing.
When puts() is called from f1(), it uses the “world” string plus a zero byte. puts() is not aware that
there is something before this string!
This clever trick is often used by at least GCC and can save some memory. This is close to string interning.
Another related example is here: 3.4.
3.4 strstr() example

Let’s back to the fact that GCC sometimes can use part of string: 3.3.1 on the previous page.
The strstr() C/C++ standard library function is used to find any occurrence in a string. This is what we will
do:
#include <string.h>
#include <stdio.h>
int main()
{
char *s="Hello, world!";
char *w=strstr(s, "world");
printf ("%p, [%s]\n", s, s);

printf ("%p, [%s]\n", w, w);
};
The output is:

0x8048530, [Hello, world!]
0x8048537, [world!]
The difference between the address of the original string and the address of the substring that strstr() has
returned is 7. Indeed, “Hello, ” string has length of 7 characters.
The printf() function during second call has no idea there are some other characters before the passed
string and it prints characters from the middle of original string till the end (marked by zero byte).
3.5 qsort() revisited

(Let’s back to the fact that CMP instruction works like SUB: 1.12.4 on page 86.)
Now that you already familiar with qsort() function ( 1.33 on page 381), here is a nice example where
comparison operation (CMP) can be replaced by subtraction operation (SUB).
/* qsort int comparison function */
int int_cmp(const void *a, const void *b)
{
const int *ia = (const int *)a; // casting pointer types
const int *ib = (const int *)b;
470
return *ia - *ib;
/* integer comparison: returns negative if b > a
and positive if a > b */
}
( http://www.anyexample.com/programming/c/qsort__sorting_array_of_strings__integers_and_
structs.xml http://archive.is/Hh3jz )
Also, a typical implementation of strcmp() (from OpenBSD):
int
strcmp(const char *s1, const char *s2)
{
while (*s1 == *s2++)
if (*s1++ == 0)
return (0);
return (*(unsigned char *)s1 - *(unsigned char *)--s2);
}
3.6 Temperature converting

Another very popular example in programming books for beginners is a small program that converts
Fahrenheit temperature to Celsius or back.
5 ⋅ (F − 32)
C=
9
We can also add simple error handling: 1) we must check if the user has entered a correct number; 2) we
must check if the Celsius temperature is not below −273 (which is below absolute zero, as we may recall
from school physics lessons).
The exit() function terminates the program instantly, without returning to the caller function.
3.6.1 Integer values

#include <stdio.h>
#include <stdlib.h>
int main()
{
int celsius, fahr;
printf ("Enter temperature in Fahrenheit:\n");
if (scanf ("%d", &fahr)!=1)
{
printf ("Error while parsing your input\n");
exit(0);
};
celsius = 5 * (fahr-32) / 9;
if (celsius<-273)
{
printf ("Error: incorrect temperature!\n");
exit(0);
};
printf ("Celsius: %d\n", celsius);
};
Optimizing MSVC 2012 x86

$SG4228 DB 'Enter temperature in Fahrenheit:', 0aH, 00H
$SG4230 DB '%d', 00H
$SG4231 DB 'Error while parsing your input', 0aH, 00H
$SG4233 DB 'Error: incorrect temperature!', 0aH, 00H
471
$SG4234 DB 'Celsius: %d', 0aH, 00H
_fahr$ = -4 ; size = 4
_main PROC
push ecx
push esi
mov esi, DWORD PTR __imp__printf
push OFFSET $SG4228 ; 'Enter temperature in Fahrenheit:'
call esi ; call printf()
lea eax, DWORD PTR _fahr$[esp+12]
push eax
call DWORD PTR __imp__scanf
add esp, 12
cmp eax, 1
je SHORT $LN2@main
push OFFSET $SG4231 ; 'Error while parsing your input'
add esp, 4
push 0
call DWORD PTR __imp__exit
$LN9@main:
$LN2@main:
mov eax, DWORD PTR _fahr$[esp+8]
add eax, -32 ; ffffffe0H
lea ecx, DWORD PTR [eax+eax*4]
mov eax, 954437177 ; 38e38e39H
imul ecx
sar edx, 1
mov eax, edx
shr eax, 31 ; 0000001fH
add eax, edx
cmp eax, -273 ; fffffeefH
jge SHORT $LN1@main
push OFFSET $SG4233 ; 'Error: incorrect temperature!'
add esp, 4
push 0
$LN10@main:
$LN1@main:
push eax
push OFFSET $SG4234 ; 'Celsius: %d'
add esp, 8
; return 0 - by C99 standard
xor eax, eax
pop esi
pop ecx
ret 0
$LN8@main:
_main ENDP
What we can say about it:

• The address of printf() is first loaded in the ESI register, so the subsequent printf() calls are
done just by the CALL ESI instruction. It’s a very popular compiler technique, possible if several
consequent calls to the same function are present in the code, and/or if there is a free register which
can be used for this.
• We see the ADD EAX, -32 instruction at the place where 32 has to be subtracted from the value.
EAX = EAX + (−32) is equivalent to EAX = EAX − 32 and somehow, the compiler decided to use ADD
instead of SUB. Maybe it’s worth it, it’s hard to be sure.
• The LEA instruction is used when the value is to be multiplied by 5: lea ecx, DWORD PTR [eax+eax*4].
Yes, i + i ∗ 4 is equivalent to i ∗ 5 and LEA works faster then IMUL.
By the way, the SHL EAX, 2 / ADD EAX, EAX instruction pair could be also used here instead—
some compilers do it like.
• The division by multiplication trick ( 3.12 on page 496) is also used here.
472
• main() returns 0 if we don’t have return 0 at its end. The C99 standard tells us [ISO/IEC 9899:TC3
(C C99 standard), (2007)5.1.2.2.3] that main() will return 0 in case the return statement is missing.
This rule works only for the main() function.
Though, MSVC doesn’t officially support C99, but maybe it support it partially?
The code is almost the same, but we can find INT 3 instructions after each exit() call.
xor ecx, ecx
call QWORD PTR __imp_exit
int 3
INT 3 is a debugger breakpoint.

It is known that exit() is one of the functions which can never return 2 , so if it does, something really
odd has happened and it’s time to load the debugger.
3.6.2 Floating-point values

#include <stdio.h>
#include <stdlib.h>
int main()
{
double celsius, fahr;
printf ("Enter temperature in Fahrenheit:\n");
if (scanf ("%lf", &fahr)!=1)
{
printf ("Error while parsing your input\n");
exit(0);
};
celsius = 5 * (fahr-32) / 9;
if (celsius<-273)
{
printf ("Error: incorrect temperature!\n");
exit(0);
};
printf ("Celsius: %lf\n", celsius);
};
MSVC 2010 x86 uses FPU instructions…

$SG4040 DB '%lf', 00H
$SG4044 DB 'Celsius: %lf', 0aH, 00H
__real@c071100000000000 DQ 0c071100000000000r ; -273

__real@4022000000000000 DQ 04022000000000000r ; 9
__real@4014000000000000 DQ 04014000000000000r ; 5
__real@4040000000000000 DQ 04040000000000000r ; 32
_fahr$ = -8 ; size = 8
_main PROC
sub esp, 8
push esi
push eax
2 another popular one is longjmp()
473
push OFFSET $SG4040 ; '%lf'
add esp, 12
cmp eax, 1
je SHORT $LN2@main
add esp, 4
push 0
$LN2@main:
fld QWORD PTR _fahr$[esp+12]
fsub QWORD PTR __real@4040000000000000 ; 32
fmul QWORD PTR __real@4014000000000000 ; 5
fdiv QWORD PTR __real@4022000000000000 ; 9
fld QWORD PTR __real@c071100000000000 ; -273
fcomp ST(1)
fnstsw ax
test ah, 65 ; 00000041H
jne SHORT $LN1@main
fstp ST(0)
add esp, 4
push 0
$LN1@main:
sub esp, 8
push OFFSET $SG4044 ; 'Celsius: %lf'
call esi
add esp, 12
xor eax, eax
pop esi
add esp, 8
ret 0
$LN10@main:
_main ENDP
…but MSVC 2012 uses SIMD instructions instead:

$SG4230 DB '%lf', 00H
$SG4234 DB 'Celsius: %lf', 0aH, 00H
__real@c071100000000000 DQ 0c071100000000000r ; -273
__real@4040000000000000 DQ 04040000000000000r ; 32
__real@4022000000000000 DQ 04022000000000000r ; 9
__real@4014000000000000 DQ 04014000000000000r ; 5
_fahr$ = -8 ; size = 8
_main PROC
sub esp, 8
push esi
push eax
push OFFSET $SG4230 ; '%lf'
add esp, 12
cmp eax, 1
je SHORT $LN2@main
add esp, 4
474
push 0
$LN9@main:
$LN2@main:
movsd xmm1, QWORD PTR _fahr$[esp+12]
subsd xmm1, QWORD PTR __real@4040000000000000 ; 32
movsd xmm0, QWORD PTR __real@c071100000000000 ; -273
mulsd xmm1, QWORD PTR __real@4014000000000000 ; 5
divsd xmm1, QWORD PTR __real@4022000000000000 ; 9
comisd xmm0, xmm1
jbe SHORT $LN1@main
add esp, 4
push 0
$LN10@main:
$LN1@main:
sub esp, 8
movsd QWORD PTR [esp], xmm1
push OFFSET $SG4234 ; 'Celsius: %lf'
add esp, 12
xor eax, eax
pop esi
add esp, 8
ret 0
$LN8@main:
_main ENDP
Of course, SIMD instructions are available in x86 mode, including those working with floating point num-
bers.
It’s somewhat easier to use them for calculations, so the new Microsoft compiler uses them.
We can also see that the −273 value is loaded into XMM0 register too early. And that’s OK, because the
compiler may emit instructions not in the order they are in the source code.
3.7 Fibonacci numbers

Another widespread example used in programming textbooks is a recursive function that generates the
Fibonacci numbers3 . The sequence is very simple: each consecutive number is the sum of the previous
two. The first two numbers are 0 and 1, or 1 and 1.
The sequence starts like this:
0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610, 987, 1597, 2584, 4181...
3.7.1 Example #1
The implementation is simple. This program generates the sequence until 21.
#include <stdio.h>
void fib (int a, int b, int limit)

{
printf ("%d\n", a+b);
if (a+b > limit)
return;
fib (b, a+b, limit);
};
int main()
{
printf ("0\n1\n1\n");
fib (1, 1, 20);
475
};

_a$ = 8 ; size = 4
_b$ = 12 ; size = 4
_limit$ = 16 ; size = 4
_fib PROC
push ebp
mov ebp, esp
add eax, DWORD PTR _b$[ebp]
push eax
push OFFSET $SG2643
add esp, 8
add ecx, DWORD PTR _b$[ebp]
cmp ecx, DWORD PTR _limit$[ebp]
jle SHORT $LN1@fib
jmp SHORT $LN2@fib
$LN1@fib:
mov edx, DWORD PTR _limit$[ebp]
push edx
push eax
push ecx
call _fib
add esp, 12
$LN2@fib:
pop ebp
ret 0
_fib ENDP
_main PROC
push ebp
mov ebp, esp
push OFFSET $SG2647 ; "0\n1\n1\n"
add esp, 4
push 20
push 1
push 1
call _fib
add esp, 12
xor eax, eax
pop ebp
ret 0
_main ENDP
We will illustrate the stack frames with this.
476
Let’s load the example in OllyDbg and trace to the last call of f():
Figure 3.1: OllyDbg: last call of f()
477
Let’s investigate the stack more closely. Comments were added by the author of this book 4 :
0035F940 00FD1039 RETURN to fib.00FD1039 from fib.00FD1000
0035F944 00000008 1st argument: a
0035F948 0000000D 2nd argument b
0035F94C 00000014 3rd argument: limit
0035F950 /0035F964 saved EBP register
0035F954 |00FD1039 RETURN to fib.00FD1039 from fib.00FD1000
0035F958 |00000005 1st argument: a
0035F95C |00000008 2nd argument: b
0035F960 |00000014 3rd argument: limit
0035F964 ]0035F978 saved EBP register
0035F96C |00000003 1st argument: a
0035F970 |00000005 2nd argument: b
0035F978 ]0035F98C saved EBP register
0035F97C |00FD1039 RETURN to fib.00FD1039 from fib.00FD1000
0035F980 |00000002 1st argument: a
0035F984 |00000003 2nd argument: b
0035F98C ]0035F9A0 saved EBP register
0035F994 |00000001 1st argument: a
0035F998 |00000002 2nd argument: b
0035F99C |00000014 3rd argument: limit
0035F9A0 ]0035F9B4 saved EBP register
0035F9A4 |00FD105C RETURN to fib.00FD105C from fib.00FD1000
0035F9A8 |00000001 1st argument: a \
0035F9AC |00000001 2nd argument: b | prepared in main() for f1()
0035F9B0 |00000014 3rd argument: limit /
0035F9B4 ]0035F9F8 saved EBP register
0035F9B8 |00FD11D0 RETURN to fib.00FD11D0 from fib.00FD1040
0035F9BC |00000001 main() 1st argument: argc \
0035F9C0 |006812C8 main() 2nd argument: argv | prepared in CRT for main()
0035F9C4 |00682940 main() 3rd argument: envp /
The function is recursive 5 , hence stack looks like a “sandwich”.

We see that the limit argument is always the same (0x14 or 20), but the a and b arguments are different
for each call.
There are also the RA-s and the saved EBP values. OllyDbg is able to determine the EBP-based frames, so
it draws these brackets. The values inside each bracket make the stack frame, in other words, the stack
area which each function incarnation uses as scratch space.
We can also say that each function incarnation must not access stack elements beyond the boundaries of
its frame (excluding function arguments), although it’s technically possible.
It’s usually true, unless the function has bugs.
Each saved EBP value is the address of the previous stack frame: this is the reason why some debuggers
can easily divide the stack in frames and dump each function’s arguments.
As we see here, each function incarnation prepares the arguments for the next function call.
At the end we see the 3 arguments for main(). argc is 1 (yes, indeed, we have ran the program without
command-line arguments).
This easily to lead to a stack overflow: just remove (or comment out) the limit check and it will crash with
exception 0xC00000FD (stack overflow.)
3.7.2 Example #2
My function has some redundancy, so let’s add a new local variable next and replace all “a+b” with it:
#include <stdio.h>
4 By the way, it’s possible to select several entries in OllyDbg and copy them to the clipboard (Ctrl-C). That’s what was done by
author for this example.

5 i.e., it calls itself
478
void fib (int a, int b, int limit)
{
int next=a+b;
printf ("%d\n", next);
if (next > limit)
return;
fib (b, next, limit);
};
int main()
{
printf ("0\n1\n1\n");
fib (1, 1, 20);
};
This is the output of non-optimizing MSVC, so the next variable is actually allocated in the local stack:

_next$ = -4 ; size = 4
_a$ = 8 ; size = 4
_b$ = 12 ; size = 4
_limit$ = 16 ; size = 4
_fib PROC
push ebp
mov ebp, esp
push ecx
mov DWORD PTR _next$[ebp], eax
mov ecx, DWORD PTR _next$[ebp]
push ecx
add esp, 8
mov edx, DWORD PTR _next$[ebp]
cmp edx, DWORD PTR _limit$[ebp]
jle SHORT $LN1@fib
jmp SHORT $LN2@fib
$LN1@fib:
mov eax, DWORD PTR _limit$[ebp]
push eax
mov ecx, DWORD PTR _next$[ebp]
push ecx
mov edx, DWORD PTR _b$[ebp]
push edx
call _fib
add esp, 12
$LN2@fib:
mov esp, ebp
pop ebp
ret 0
_fib ENDP
_main PROC
push ebp
mov ebp, esp
push OFFSET $SG2753 ; "0\n1\n1\n"
add esp, 4
push 20
push 1
push 1
call _fib
add esp, 12
xor eax, eax
pop ebp
ret 0
_main ENDP
479
Let’s load it in OllyDbg once again:
Figure 3.2: OllyDbg: last call of f()
Now the next variable is present in each frame.
480
Let’s investigate the stack more closely. The author has again added his comments:
0029FC14 00E0103A RETURN to fib2.00E0103A from fib2.00E01000
0029FC18 00000008 1st argument: a
0029FC1C 0000000D 2nd argument: b
0029FC20 00000014 3rd argument: limit
0029FC24 0000000D "next" variable
0029FC28 /0029FC40 saved EBP register
0029FC2C |00E0103A RETURN to fib2.00E0103A from fib2.00E01000
0029FC30 |00000005 1st argument: a
0029FC34 |00000008 2nd argument: b
0029FC38 |00000014 3rd argument: limit
0029FC3C |00000008 "next" variable
0029FC40 ]0029FC58 saved EBP register
0029FC44 |00E0103A RETURN to fib2.00E0103A from fib2.00E01000
0029FC48 |00000003 1st argument: a
0029FC4C |00000005 2nd argument: b
0029FC54 |00000005 "next" variable
0029FC5C |00E0103A RETURN to fib2.00E0103A from fib2.00E01000
0029FC60 |00000002 1st argument: a
0029FC64 |00000003 2nd argument: b
0029FC6C |00000003 "next" variable
0029FC74 |00E0103A RETURN to fib2.00E0103A from fib2.00E01000
0029FC78 |00000001 1st argument: a \
0029FC7C |00000002 2nd argument: b | prepared in f1() for next f1() call
0029FC80 |00000014 3rd argument: limit /
0029FC84 |00000002 "next" variable
0029FC88 ]0029FC9C saved EBP register
0029FC8C |00E0106C RETURN to fib2.00E0106C from fib2.00E01000
0029FC90 |00000001 1st argument: a \
0029FC94 |00000001 2nd argument: b | prepared in main() for f1()
0029FC98 |00000014 3rd argument: limit /
0029FC9C ]0029FCE0 saved EBP register
0029FCA0 |00E011E0 RETURN to fib2.00E011E0 from fib2.00E01050
0029FCA4 |00000001 main() 1st argument: argc \
0029FCA8 |000812C8 main() 2nd argument: argv | prepared in CRT for main()
0029FCAC |00082940 main() 3rd argument: envp /
Here we see it: the next value is calculated in each function incarnation, then passed as argument b to
the next incarnation.
3.7.3 Summary
Recursive functions are æsthetically nice, but technically may degrade performance because of their
heavy stack usage. Everyone who writes performance critical code probably should avoid recursion.
For example, the author of this book once wrote a function to seek a particular node in a binary tree. As a
recursive function it looked quite stylish but since additional time was spent at each function call for the
prologue/epilogue, it was working a couple of times slower than an iterative (recursion-free) implementa-
tion.
By the way, that is the reason that some functional PL6 compilers (where recursion is used heavily) use
tail call. We talk about tail call when a function has only one single call to itself located at the end of it,
like:
Listing 3.9: Scheme, example is copypasted from Wikipedia

;; factorial : number -> number
;; to calculate the product of all positive
;; integers less than or equal to n.
(define (factorial n)
(if (= n 1)
1
(* n (factorial (- n 1)))))
6 LISP, Python, Lua, etc.
481
Tail call is important because compiler can rework this code easily into iterative one, to get rid of recursion.
3.8 CRC32 calculation example

This is a very popular table-based CRC32 hash calculation technique7 .
/* By Bob Jenkins, (c) 2006, Public Domain */
#include <stdio.h>
#include <stddef.h>
#include <string.h>
typedef unsigned long ub4;

typedef unsigned char ub1;
static const ub4 crctab[256] = {

0x00000000, 0x77073096, 0xee0e612c, 0x990951ba, 0x076dc419,
0x706af48f, 0xe963a535, 0x9e6495a3, 0x0edb8832, 0x79dcb8a4,
0xe0d5e91e, 0x97d2d988, 0x09b64c2b, 0x7eb17cbd, 0xe7b82d07,
0x90bf1d91, 0x1db71064, 0x6ab020f2, 0xf3b97148, 0x84be41de,
0x1adad47d, 0x6ddde4eb, 0xf4d4b551, 0x83d385c7, 0x136c9856,
0x646ba8c0, 0xfd62f97a, 0x8a65c9ec, 0x14015c4f, 0x63066cd9,
0xfa0f3d63, 0x8d080df5, 0x3b6e20c8, 0x4c69105e, 0xd56041e4,
0xa2677172, 0x3c03e4d1, 0x4b04d447, 0xd20d85fd, 0xa50ab56b,
0x35b5a8fa, 0x42b2986c, 0xdbbbc9d6, 0xacbcf940, 0x32d86ce3,
0x45df5c75, 0xdcd60dcf, 0xabd13d59, 0x26d930ac, 0x51de003a,
0xc8d75180, 0xbfd06116, 0x21b4f4b5, 0x56b3c423, 0xcfba9599,
0xb8bda50f, 0x2802b89e, 0x5f058808, 0xc60cd9b2, 0xb10be924,
0x2f6f7c87, 0x58684c11, 0xc1611dab, 0xb6662d3d, 0x76dc4190,
0x01db7106, 0x98d220bc, 0xefd5102a, 0x71b18589, 0x06b6b51f,
0x9fbfe4a5, 0xe8b8d433, 0x7807c9a2, 0x0f00f934, 0x9609a88e,
0xe10e9818, 0x7f6a0dbb, 0x086d3d2d, 0x91646c97, 0xe6635c01,
0x6b6b51f4, 0x1c6c6162, 0x856530d8, 0xf262004e, 0x6c0695ed,
0x1b01a57b, 0x8208f4c1, 0xf50fc457, 0x65b0d9c6, 0x12b7e950,
0x8bbeb8ea, 0xfcb9887c, 0x62dd1ddf, 0x15da2d49, 0x8cd37cf3,
0xfbd44c65, 0x4db26158, 0x3ab551ce, 0xa3bc0074, 0xd4bb30e2,
0x4adfa541, 0x3dd895d7, 0xa4d1c46d, 0xd3d6f4fb, 0x4369e96a,
0x346ed9fc, 0xad678846, 0xda60b8d0, 0x44042d73, 0x33031de5,
0xaa0a4c5f, 0xdd0d7cc9, 0x5005713c, 0x270241aa, 0xbe0b1010,
0xc90c2086, 0x5768b525, 0x206f85b3, 0xb966d409, 0xce61e49f,
0x5edef90e, 0x29d9c998, 0xb0d09822, 0xc7d7a8b4, 0x59b33d17,
0x2eb40d81, 0xb7bd5c3b, 0xc0ba6cad, 0xedb88320, 0x9abfb3b6,
0x03b6e20c, 0x74b1d29a, 0xead54739, 0x9dd277af, 0x04db2615,
0x73dc1683, 0xe3630b12, 0x94643b84, 0x0d6d6a3e, 0x7a6a5aa8,
0xe40ecf0b, 0x9309ff9d, 0x0a00ae27, 0x7d079eb1, 0xf00f9344,
0x8708a3d2, 0x1e01f268, 0x6906c2fe, 0xf762575d, 0x806567cb,
0x196c3671, 0x6e6b06e7, 0xfed41b76, 0x89d32be0, 0x10da7a5a,
0x67dd4acc, 0xf9b9df6f, 0x8ebeeff9, 0x17b7be43, 0x60b08ed5,
0xd6d6a3e8, 0xa1d1937e, 0x38d8c2c4, 0x4fdff252, 0xd1bb67f1,
0xa6bc5767, 0x3fb506dd, 0x48b2364b, 0xd80d2bda, 0xaf0a1b4c,
0x36034af6, 0x41047a60, 0xdf60efc3, 0xa867df55, 0x316e8eef,
0x4669be79, 0xcb61b38c, 0xbc66831a, 0x256fd2a0, 0x5268e236,
0xcc0c7795, 0xbb0b4703, 0x220216b9, 0x5505262f, 0xc5ba3bbe,
0xb2bd0b28, 0x2bb45a92, 0x5cb36a04, 0xc2d7ffa7, 0xb5d0cf31,
0x2cd99e8b, 0x5bdeae1d, 0x9b64c2b0, 0xec63f226, 0x756aa39c,
0x026d930a, 0x9c0906a9, 0xeb0e363f, 0x72076785, 0x05005713,
0x95bf4a82, 0xe2b87a14, 0x7bb12bae, 0x0cb61b38, 0x92d28e9b,
0xe5d5be0d, 0x7cdcefb7, 0x0bdbdf21, 0x86d3d2d4, 0xf1d4e242,
0x68ddb3f8, 0x1fda836e, 0x81be16cd, 0xf6b9265b, 0x6fb077e1,
0x18b74777, 0x88085ae6, 0xff0f6a70, 0x66063bca, 0x11010b5c,
0x8f659eff, 0xf862ae69, 0x616bffd3, 0x166ccf45, 0xa00ae278,
0xd70dd2ee, 0x4e048354, 0x3903b3c2, 0xa7672661, 0xd06016f7,
0x4969474d, 0x3e6e77db, 0xaed16a4a, 0xd9d65adc, 0x40df0b66,
0x37d83bf0, 0xa9bcae53, 0xdebb9ec5, 0x47b2cf7f, 0x30b5ffe9,
0xbdbdf21c, 0xcabac28a, 0x53b39330, 0x24b4a3a6, 0xbad03605,
0xcdd70693, 0x54de5729, 0x23d967bf, 0xb3667a2e, 0xc4614ab8,
0x5d681b02, 0x2a6f2b94, 0xb40bbe37, 0xc30c8ea1, 0x5a05df1b,
7 The source code has been taken from here: http://go.yurichev.com/17327
482
0x2d02ef8d
};
/* how to derive the values in crctab[] from polynomial 0xedb88320 */

void build_table()
{
ub4 i, j;
for (i=0; i<256; ++i) {
j = i;
j = (j>>1) ^ ((j&1) ? 0xedb88320 : 0);
j = (j>>1) ^ ((j&1) ? 0xedb88320 : 0);
j = (j>>1) ^ ((j&1) ? 0xedb88320 : 0);
j = (j>>1) ^ ((j&1) ? 0xedb88320 : 0);
j = (j>>1) ^ ((j&1) ? 0xedb88320 : 0);
j = (j>>1) ^ ((j&1) ? 0xedb88320 : 0);
j = (j>>1) ^ ((j&1) ? 0xedb88320 : 0);
j = (j>>1) ^ ((j&1) ? 0xedb88320 : 0);
printf("0x%.8lx, ", j);
if (i%6 == 5) printf("\n");
}
}
/* the hash function */

ub4 crc(const void *key, ub4 len, ub4 hash)
{
ub4 i;
const ub1 *k = key;
for (hash=len, i=0; i<len; ++i)
hash = (hash >> 8) ^ crctab[(hash & 0xff) ^ k[i]];
return hash;
}
/* To use, try "gcc -O crc.c -o crc; crc < crc.c" */

int main()
{
char s[1000];
while (gets(s)) printf("%.8lx\n", crc(s, strlen(s), 0));
return 0;
}
We are interested in the crc() function only. By the way, pay attention to the two loop initializers in the
for() statement: hash=len, i=0. The C/C++ standard allows this, of course. The emitted code will
contain two operations in the loop initialization part instead of one.
Let’s compile it in MSVC with optimization (/Ox). For the sake of brevity, only the crc() function is listed
here, with my comments.
_key$ = 8 ; size = 4
_len$ = 12 ; size = 4
_hash$ = 16 ; size = 4
_crc PROC
mov edx, DWORD PTR _len$[esp-4]
xor ecx, ecx ; i will be stored in ECX
mov eax, edx
test edx, edx
jbe SHORT $LN1@crc
push ebx
push esi
mov esi, DWORD PTR _key$[esp+4] ; ESI = key
push edi
$LL3@crc:
; work with bytes using only 32-bit registers. byte from address key+i we store into EDI
movzx edi, BYTE PTR [ecx+esi]

mov ebx, eax ; EBX = (hash = len)
and ebx, 255 ; EBX = hash & 0xff
; XOR EDI, EBX (EDI=EDIÊBX) - this operation uses all 32 bits of each register
483
; but other bits (8-31) are cleared all time, so it is OK
; these are cleared because, as for EDI, it was done by MOVZX instruction above
; high bits of EBX was cleared by AND EBX, 255 instruction above (255 = 0xff)
xor edi, ebx
; EAX=EAX>>8; bits 24-31 taken from nowhere will be cleared

shr eax, 8
; EAX=EAX^crctab[EDI*4] - choose EDI-th element from crctab[] table

xor eax, DWORD PTR _crctab[edi*4]
inc ecx ; i++
cmp ecx, edx ; i<len ?
jb SHORT $LL3@crc ; yes
pop edi
pop esi
pop ebx
$LN1@crc:
ret 0
_crc ENDP
Let’s try the same in GCC 4.4.1 with -O3 option:

public crc
crc proc near
key = dword ptr 8

hash = dword ptr 0Ch
push ebp
xor edx, edx
mov ebp, esp
push esi
mov esi, [ebp+key]
push ebx
mov ebx, [ebp+hash]
test ebx, ebx
mov eax, ebx
nop ; padding
lea esi, [esi+0] ; padding; works as NOP (ESI does not change here)
loc_80484B8:
mov ecx, eax ; save previous state of hash to ECX
xor al, [esi+edx] ; AL=*(key+i)
add edx, 1 ; i++
shr ecx, 8 ; ECX=hash>>8
movzx eax, al ; EAX=*(key+i)
mov eax, dword ptr ds:crctab[eax*4] ; EAX=crctab[EAX]
xor eax, ecx ; hash=EAXÊCX
cmp ebx, edx
ja short loc_80484B8
loc_80484D3:
pop ebx
pop esi
pop ebp
retn
crc endp
GCC has aligned the loop start on a 8-byte boundary by adding NOP and lea esi, [esi+0] (that is an
idle operation too). Read more about it in npad section ( .1.7 on page 1024).
3.9 Network address calculation example

As we know, a TCP/IP address (IPv4) consists of four numbers in the 0 . . . 255 range, i.e., four bytes.
484
Four bytes can be fit in a 32-bit variable easily, so an IPv4 host address, network mask or network address
can all be 32-bit integers.
From the user’s point of view, the network mask is defined as four numbers and is formatted like 255.255.255.0
or so, but network engineers (sysadmins) use a more compact notation (CIDR8 ), like “/8”, “/16”, etc.
This notation just defines the number of bits the mask has, starting at the MSB.
Mask Hosts Usable Netmask Hex mask
/30 4 2 255.255.255.252 0xfffffffc
/29 8 6 255.255.255.248 0xfffffff8
/28 16 14 255.255.255.240 0xfffffff0
/27 32 30 255.255.255.224 0xffffffe0
/26 64 62 255.255.255.192 0xffffffc0
/24 256 254 255.255.255.0 0xffffff00 class C network
/23 512 510 255.255.254.0 0xfffffe00
/22 1024 1022 255.255.252.0 0xfffffc00
/21 2048 2046 255.255.248.0 0xfffff800
/20 4096 4094 255.255.240.0 0xfffff000
/19 8192 8190 255.255.224.0 0xffffe000
/18 16384 16382 255.255.192.0 0xffffc000
/17 32768 32766 255.255.128.0 0xffff8000
/16 65536 65534 255.255.0.0 0xffff0000 class B network
/8 16777216 16777214 255.0.0.0 0xff000000 class A network
Here is a small example, which calculates the network address by applying the network mask to the host
address.
#include <stdio.h>
#include <stdint.h>
uint32_t form_IP (uint8_t ip1, uint8_t ip2, uint8_t ip3, uint8_t ip4)
{
return (ip1<<24) | (ip2<<16) | (ip3<<8) | ip4;
};
void print_as_IP (uint32_t a)

{
printf ("%d.%d.%d.%d\n",
(a>>24)&0xFF,
(a>>16)&0xFF,
(a>>8)&0xFF,
(a)&0xFF);
};
// bit=31..0
uint32_t set_bit (uint32_t input, int bit)
{
return input=input|(1<<bit);
};
uint32_t form_netmask (uint8_t netmask_bits)

{
uint32_t netmask=0;
uint8_t i;
for (i=0; i<netmask_bits; i++)

netmask=set_bit(netmask, 31-i);
return netmask;
};
void calc_network_address (uint8_t ip1, uint8_t ip2, uint8_t ip3, uint8_t ip4, uint8_t ⤦
Ç netmask_bits)
{
uint32_t netmask=form_netmask(netmask_bits);
uint32_t ip=form_IP(ip1, ip2, ip3, ip4);
uint32_t netw_adr;
8 Classless Inter-Domain Routing
485
printf ("netmask=");
print_as_IP (netmask);
netw_adr=ip&netmask;
printf ("network address=");

print_as_IP (netw_adr);
};
int main()
{
calc_network_address (10, 1, 2, 4, 24); // 10.1.2.4, /24
};
3.9.1 calc_network_address()
calc_network_address() function is simplest one: it just ANDs the host address with the network mask,
resulting in the network address.

1 _ip1$ = 8 ; size = 1
2 _ip2$ = 12 ; size = 1
3 _ip3$ = 16 ; size = 1
4 _ip4$ = 20 ; size = 1
5 _netmask_bits$ = 24 ; size = 1
6 _calc_network_address PROC
7 push edi
8 push DWORD PTR _netmask_bits$[esp]
9 call _form_netmask
10 push OFFSET $SG3045 ; 'netmask='
11 mov edi, eax
12 call DWORD PTR __imp__printf
13 push edi
14 call _print_as_IP
15 push OFFSET $SG3046 ; 'network address='
16 call DWORD PTR __imp__printf
17 push DWORD PTR _ip4$[esp+16]
21 call _form_IP
22 and eax, edi ; network address = host address & netmask
23 push eax
24 call _print_as_IP
25 add esp, 36
26 pop edi
27 ret 0
28 _calc_network_address ENDP
At line 22 we see the most important AND—here the network address is calculated.
3.9.2 form_IP()
The form_IP() function just puts all 4 bytes into a 32-bit value.
Here is how it is usually done:
• Allocate a variable for the return value. Set it to 0.
• Take the fourth (lowest) byte, apply OR operation to this byte and return the value. The return value
contain the 4th byte now.
• Take the third byte, shift it left by 8 bits. You’ll get a value like 0x0000bb00 where bb is your third byte.
Apply the OR operation to the resulting value and returning value. The return value has contained
0x000000aa so far, so ORing the values will produce a value like 0x0000bbaa.
486
• Take the second byte, shift it left by 16 bits. You’ll get a value like 0x00cc0000, where cc is your
second byte. Apply the OR operation to the resulting value and returning value. The return value
has contained 0x0000bbaa so far, so ORing the values will produce a value like 0x00ccbbaa.
• Take the first byte, shift it left by 24 bits. You’ll get a value like 0xdd000000, where dd is your first byte.
Apply the OR operation to the resulting value and returning value. The return value has contained
0x00ccbbaa so far, so ORing the values will produce a value like 0xddccbbaa.
And this is how it’s done by non-optimizing MSVC 2012:

; denote ip1 as "dd", ip2 as "cc", ip3 as "bb", ip4 as "aa".
_ip1$ = 8 ; size = 1
_ip2$ = 12 ; size = 1
_ip3$ = 16 ; size = 1
_ip4$ = 20 ; size = 1
_form_IP PROC
push ebp
mov ebp, esp
movzx eax, BYTE PTR _ip1$[ebp]
; EAX=000000dd
shl eax, 24
; EAX=dd000000
movzx ecx, BYTE PTR _ip2$[ebp]
; ECX=000000cc
shl ecx, 16
; ECX=00cc0000
or eax, ecx
; EAX=ddcc0000
movzx edx, BYTE PTR _ip3$[ebp]
; EDX=000000bb
shl edx, 8
; EDX=0000bb00
or eax, edx
; EAX=ddccbb00
movzx ecx, BYTE PTR _ip4$[ebp]
; ECX=000000aa
or eax, ecx
; EAX=ddccbbaa
pop ebp
ret 0
_form_IP ENDP
Well, the order is different, but, of course, the order of the operations doesn’t matter.
Optimizing MSVC 2012 does essentially the same, but in a different way:

; denote ip1 as "dd", ip2 as "cc", ip3 as "bb", ip4 as "aa".
_ip1$ = 8 ; size = 1
_ip2$ = 12 ; size = 1
_ip3$ = 16 ; size = 1
_ip4$ = 20 ; size = 1
_form_IP PROC
movzx eax, BYTE PTR _ip1$[esp-4]
; EAX=000000dd
movzx ecx, BYTE PTR _ip2$[esp-4]
; ECX=000000cc
shl eax, 8
; EAX=0000dd00
or eax, ecx
; EAX=0000ddcc
; ECX=000000bb
shl eax, 8
; EAX=00ddcc00
or eax, ecx
; EAX=00ddccbb
; ECX=000000aa
487
shl eax, 8
; EAX=ddccbb00
or eax, ecx
; EAX=ddccbbaa
ret 0
_form_IP ENDP
We could say that each byte is written to the lowest 8 bits of the return value, and then the return value
is shifted left by one byte at each step.
Repeat 4 times for each input byte.
That’s it! Unfortunately, there are probably no other ways to do it.
There are no popular CPUs or ISAs which has instruction for composing a value from bits or bytes.
It’s all usually done by bit shifting and ORing.
3.9.3 print_as_IP()
print_as_IP() does the inverse: splitting a 32-bit value into 4 bytes.
Slicing works somewhat simpler: just shift input value by 24, 16, 8 or 0 bits, take the bits from zeroth to
seventh (lowest byte), and that’s it:

_a$ = 8 ; size = 4
_print_as_IP PROC
push ebp
mov ebp, esp
; EAX=ddccbbaa
and eax, 255
; EAX=000000aa
push eax
; ECX=ddccbbaa
shr ecx, 8
; ECX=00ddccbb
and ecx, 255
; ECX=000000bb
push ecx
; EDX=ddccbbaa
shr edx, 16
; EDX=0000ddcc
and edx, 255
; EDX=000000cc
push edx
; EAX=ddccbbaa
shr eax, 24
; EAX=000000dd
and eax, 255 ; probably redundant instruction
; EAX=000000dd
push eax
push OFFSET $SG2973 ; '%d.%d.%d.%d'
add esp, 20
pop ebp
ret 0
_print_as_IP ENDP
Optimizing MSVC 2012 does almost the same, but without unnecessary reloading of the input value:

_a$ = 8 ; size = 4
_print_as_IP PROC
; ECX=ddccbbaa
488
movzx eax, cl
; EAX=000000aa
push eax
mov eax, ecx
; EAX=ddccbbaa
shr eax, 8
; EAX=00ddccbb
and eax, 255
; EAX=000000bb
push eax
mov eax, ecx
; EAX=ddccbbaa
shr eax, 16
; EAX=0000ddcc
and eax, 255
; EAX=000000cc
push eax
; ECX=ddccbbaa
shr ecx, 24
; ECX=000000dd
push ecx
push OFFSET $SG3020 ; '%d.%d.%d.%d'
add esp, 20
ret 0
_print_as_IP ENDP
3.9.4 form_netmask() and set_bit()

form_netmask() makes a network mask value from CIDR notation. Of course, it would be much effective
to use here some kind of a precalculated table, but we consider it in this way intentionally, to demonstrate
bit shifts.
We will also write a separate function set_bit(). It’s a not very good idea to create a function for such
primitive operation, but it would be easy to understand how it all works.

_input$ = 8 ; size = 4
_bit$ = 12 ; size = 4
_set_bit PROC
mov ecx, DWORD PTR _bit$[esp-4]
mov eax, 1
shl eax, cl
or eax, DWORD PTR _input$[esp-4]
ret 0
_set_bit ENDP
_netmask_bits$ = 8 ; size = 1
_form_netmask PROC
push ebx
push esi
movzx esi, BYTE PTR _netmask_bits$[esp+4]
xor ecx, ecx
xor bl, bl
test esi, esi
jle SHORT $LN9@form_netma
xor edx, edx
$LL3@form_netma:
mov eax, 31
sub eax, edx
push eax
push ecx
call _set_bit
inc bl
movzx edx, bl
add esp, 8
mov ecx, eax
cmp edx, esi
489
jl SHORT $LL3@form_netma
$LN9@form_netma:
pop esi
mov eax, ecx
pop ebx
ret 0
_form_netmask ENDP
set_bit() is primitive: it just shift left 1 to number of bits we need and then ORs it with the “input” value.
form_netmask() has a loop: it will set as many bits (starting from the MSB) as passed in the netmask_bits
argument
3.9.5 Summary
That’s it! We run it and getting:
netmask=255.255.255.0
network address=10.1.2.0
netmask=255.0.0.0
netmask=255.255.255.128
netmask=255.255.255.192
3.10 Loops: several iterators

In most cases loops have only one iterator, but there could be several in the resulting code.
Here is a very simple example:
#include <stdio.h>
void f(int *a1, int *a2, size_t cnt)

{
size_t i;
// copy from one array to another in some weird scheme

a1[i*3]=a2[i*7];
};
There are two multiplications at each iteration and they are costly operations. Can we optimize it some-
how?
Yes, if we notice that both array indices are jumping on values that we can easily calculate without multi-
plication.
3.10.1 Three iterators

f PROC
; RCX=a1
; RDX=a2
; R8=cnt
test r8, r8 ; cnt==0? exit then
je SHORT $LN1@f
npad 11
$LL3@f:
mov eax, DWORD PTR [rdx]
lea rcx, QWORD PTR [rcx+12]
lea rdx, QWORD PTR [rdx+28]
mov DWORD PTR [rcx-12], eax
dec r8
jne SHORT $LL3@f
$LN1@f:
490
ret 0
f ENDP
Now there are 3 iterators: the cnt variable and two indices, which are increased by 12 and 28 at each
iteration. We can rewrite this code in C/C++:
#include <stdio.h>

{
size_t i;
size_t idx1=0; idx2=0;

{
a1[idx1]=a2[idx2];
idx1+=3;
idx2+=7;
};
};
So, at the cost of updating 3 iterators at each iteration instead of one, we can remove two multiplication
operations.
3.10.2 Two iterators

GCC 4.9 does even more, leaving only 2 iterators:

; RDI=a1
; RSI=a2
; RDX=cnt
f:
test rdx, rdx ; cnt==0? exit then
je .L1
; calculate last element address in "a2" and leave it in RDX
lea rax, [0+rdx*4]
; RAX=RDX*4=cnt*4
sal rdx, 5
; RDX=RDX<<5=cnt*32
sub rdx, rax
; RDX=RDX-RAX=cnt*32-cnt*4=cnt*28
add rdx, rsi
; RDX=RDX+RSI=a2+cnt*28
.L3:
mov eax, DWORD PTR [rsi]
add rsi, 28
add rdi, 12
mov DWORD PTR [rdi-12], eax
cmp rsi, rdx
jne .L3
.L1:
rep ret
There is no counter variable any more: GCC concluded that it is not needed.
The last element of the a2 array is calculated before the loop begins (which is easy: cnt ∗ 7) and that’s how
the loop is to be stopped: just iterate until the second index reaches this precalculated value.
You can read more about multiplication using shifts/additions/subtractions here: 1.24.1 on page 212.
This code can be rewritten into C/C++ like that:
#include <stdio.h>

{
size_t idx1=0; idx2=0;
491
size_t last_idx2=cnt*7;

for (;;)
{
a1[idx1]=a2[idx2];
idx1+=3;
idx2+=7;
if (idx2==last_idx2)
break;
};
};
GCC (Linaro) 4.9 for ARM64 does the same, but it precalculates the last index of a1 instead of a2, which,
of course has the same effect:
Listing 3.18: Optimizing GCC (Linaro) 4.9 ARM64
; X0=a1
; X1=a2
; X2=cnt
f:
cbz x2, .L1 ; cnt==0? exit then
; calculate last element of "a1" array
; X2=X2+X2<<1=X2+X2*2=X2*3
mov x3, 0
lsl x2, x2, 2
; X2=X2<<2=X2*4=X2*3*4=X2*12
.L3:
ldr w4, [x1],28 ; load at X1, add 28 to X1 (post-increment)
str w4, [x0,x3] ; store at X0+X3=a1+X3
add x3, x3, 12 ; shift X3
cmp x3, x2 ; end?
bne .L3
.L1:
ret
GCC 4.4.5 for MIPS does the same:

Listing 3.19: Optimizing GCC 4.4.5 for MIPS (IDA)
; $a0=a1
; $a1=a2
; $a2=cnt
f:
; jump to loop check code:
beqz $a2, locret_24
; initialize counter (i) at 0:
move $v0, $zero ; branch delay slot, NOP
loc_8:
; load 32-bit word at $a1
lw $a3, 0($a1)
addiu $v0, 1
; check for finish (compare "i" in $v0 and "cnt" in $a2):
sltu $v1, $v0, $a2
; store 32-bit word at $a0:
sw $a3, 0($a0)
; add 0x1C (28) to $a1 at each iteration:
addiu $a1, 0x1C
; jump to loop body if i<cnt:
bnez $v1, loc_8
; add 0xC (12) to $a0 at each iteration:
addiu $a0, 0xC ; branch delay slot
locret_24:
jr $ra
492
3.10.3 Intel C++ 2011 case
Compiler optimizations can also be weird, but nevertheless, still correct. Here is what the Intel C++
compiler 2011 does:
Listing 3.20: Optimizing Intel C++ 2011 x64

f PROC
; parameter 1: rcx = a1
; parameter 2: rdx = a2
; parameter 3: r8 = cnt
.B1.1::
test r8, r8
jbe exit
.B1.2::
cmp r8, 6
jbe just_copy
.B1.3::
cmp rcx, rdx
jbe .B1.5
.B1.4::
mov r10, r8
mov r9, rcx
shl r10, 5
lea rax, QWORD PTR [r8*4]
sub r9, rdx
sub r10, rax
cmp r9, r10
jge just_copy2
.B1.5::
cmp rdx, rcx
jbe just_copy
.B1.6::
mov r9, rdx
lea rax, QWORD PTR [r8*8]
sub r9, rcx
lea r10, QWORD PTR [rax+r8*4]
cmp r9, r10
jl just_copy
just_copy2::
; R8 = cnt
; RDX = a2
; RCX = a1
xor r10d, r10d
xor r9d, r9d
xor eax, eax
.B1.8::
mov r11d, DWORD PTR [rax+rdx]
inc r10
mov DWORD PTR [r9+rcx], r11d
add r9, 12
add rax, 28
cmp r10, r8
jb .B1.8
jmp exit
just_copy::
; R8 = cnt
; RDX = a2
; RCX = a1
xor r10d, r10d
xor r9d, r9d
xor eax, eax
493
.B1.11::
mov r11d, DWORD PTR [rax+rdx]
inc r10
mov DWORD PTR [r9+rcx], r11d
add r9, 12
add rax, 28
cmp r10, r8
jb .B1.11
exit::
ret
First, there are some decisions taken, then one of the routines is executed.
Looks like it is a check if arrays intersect.
This is very well known way of optimizing memory block copy routines. But copy routines are the same!
This is has to be an error of the Intel C++ optimizer, which still produces workable code, though.
We intentionally considering such example code in this book so the reader would understand that compiler
output is weird at times, but still correct, because when the compiler was tested, it passed the tests.
3.11 Duff’s device

Duff’s device is an unrolled loop with the possibility to jump to the middle of it. The unrolled loop is
implemented using a fallthrough switch() statement. We would use here a slightly simplified version of
Tom Duff’s original code. Let’s say, we have to write a function that clears a region in memory. One
can come with a simple loop, clearing byte by byte. It’s obviously slow, since all modern computers have
much wider memory bus. So the better way is to clear the memory region using 4 or 8 bytes blocks. Since
we are going to work with a 64-bit example here, we are going to clear the memory in 8 bytes blocks. So
far so good. But what about the tail? Memory clearing routine can also be called for regions of size that’s
not a multiple of 8. So here is the algorithm:
• calculate the number of 8-bytes blocks, clear them using 8-bytes (64-bit) memory accesses;
• calculate the size of the tail, clear it using 1-byte memory accesses.
The second step can be implemented using a simple loop. But let’s implement it as an unrolled loop:
#include <stdint.h>
#include <stdio.h>
void bzero(uint8_t* dst, size_t count)

{
int i;
if (count&(~7))
// work out 8-byte blocks
for (i=0; i<count>>3; i++)
{
*(uint64_t*)dst=0;
dst=dst+8;
};
// work out the tail

switch(count & 7)
{
case 7: *dst++ = 0;
case 6: *dst++ = 0;
case 5: *dst++ = 0;
case 4: *dst++ = 0;
case 3: *dst++ = 0;
case 2: *dst++ = 0;
case 1: *dst++ = 0;
case 0: // do nothing
break;
}
}
494
Let’s first understand how the calculation is performed. The memory region size comes as a 64-bit value.
And this value can be divided in two parts:
7 6 5 4 3 2 1 0
… B B B B B S S S
( “B” is number of 8-byte blocks and “S” is length of the tail in bytes ).
When we divide the input memory region size by 8, the value is just shifted right by 3 bits. But to calculate
the remainder, we can just to isolate the lowest 3 bits! So the number of 8-byte blocks is calculated as
count >> 3 and remainder as count&7. We also have to find out if we are going to execute the 8-byte
procedure at all, so we need to check if the value of count is greater than 7. We do this by clearing the
3 lowest bits and comparing the resulting number with zero, because all we need here is to answer the
question, is the high part of count non-zero. Of course, this works because 8 is 23 and division by numbers
that are 2n is easy. It’s not possible for other numbers. It’s actually hard to say if these hacks are worth
using, because they lead to hard-to-read code. However, these tricks are very popular and a practicing
programmer, even if he/she is not using them, nevertheless has to understand them.
So the first part is simple: get the number of 8-byte blocks and write 64-bit zero values to memory. The
second part is an unrolled loop implemented as fallthrough switch() statement.
First, let’s express in plain English what we have to do here.
We have to “write as many zero bytes in memory, as count&7 value tells us”. If it’s 0, jump to the end, there
is no work to do. If it’s 1, jump to the place inside switch() statement where only one storage operation is
to be executed. If it’s 2, jump to another place, where two storage operation are to be executed, etc. 7 as
input value leads to the execution of all 7 operations. There is no 8, because a memory region of 8 bytes
is to be processed by the first part of our function. So we wrote an unrolled loop. It was definitely faster
on older computers than normal loops (and conversely, latest CPUs works better for short loops than for
unrolled ones). Maybe this is still meaningful on modern low-cost embedded MCU9 s.
Let’s see what the optimizing MSVC 2012 does:
dst$ = 8
count$ = 16
bzero PROC
test rdx, -8
je SHORT $LN11@bzero
; work out 8-byte blocks
xor r10d, r10d
mov r9, rdx
shr r9, 3
mov r8d, r10d
test r9, r9
je SHORT $LN11@bzero
npad 5
$LL19@bzero:
inc r8d
mov QWORD PTR [rcx], r10
add rcx, 8
movsxd rax, r8d
cmp rax, r9
jb SHORT $LL19@bzero
$LN11@bzero:
; work out the tail
and edx, 7
dec rdx
cmp rdx, 6
ja SHORT $LN9@bzero
lea r8, OFFSET FLAT:__ImageBase
mov eax, DWORD PTR $LN22@bzero[r8+rdx*4]
add rax, r8
jmp rax
$LN8@bzero:
mov BYTE PTR [rcx], 0
inc rcx
$LN7@bzero:
inc rcx
9 Microcontroller Unit
495
$LN6@bzero:
inc rcx
$LN5@bzero:
inc rcx
$LN4@bzero:
inc rcx
$LN3@bzero:
inc rcx
$LN2@bzero:
$LN9@bzero:
fatret 0
npad 1
$LN22@bzero:
DD $LN2@bzero
DD $LN3@bzero
DD $LN4@bzero
DD $LN5@bzero
DD $LN6@bzero
DD $LN7@bzero
DD $LN8@bzero
bzero ENDP
The first part of the function is predictable. The second part is just an unrolled loop and a jump passing
control flow to the correct instruction inside it. There is no other code between the MOV/INC instruction
pairs, so the execution is to fall until the very end, executing as many pairs as needed. By the way, we
can observe that the MOV/INC pair consumes a fixed number of bytes (3+3). So the pair consumes 6 bytes.
Knowing that, we can get rid of the switch() jumptable, we can just multiple the input value by 6 and jump
to current_RIP + input_value ∗ 6.
This can also be faster because we are not in need to fetch a value from the jumptable.
It’s possible that 6 probably is not a very good constant for fast multiplication and maybe it’s not worth
it, but you get the idea10 .
That is what old-school demomakers did in the past with unrolled loops.
3.11.1 Should one use unrolled loops?

Unrolled loops can have benefits if there is no fast cache memory between RAM and CPU, and the CPU, in
order to get the code of the next instruction, must load it from RAM each time. This is a case of modern
low-cost MCU and old CPUs.
Unrolled loops are slower than short loops if there is a fast cache between RAM and CPU and the body
of loop can fit into cache, and CPU will load the code from it not touching the RAM. Fast loops are the
loops which body’s size can fit into L1 cache, but even faster loops are those small ones which can fit into
micro-operation cache.
3.12 Division using multiplication

A very simple function:
int f(int a)
{
return a/9;
};
3.12.1 x86
…is compiled in a very predictable way:
10 As an exercise, you can try to rework the code to get rid of the jumptable. The instruction pair can be rewritten in a way that it
will consume 4 bytes or maybe 8. 1 byte is also possible (using STOSB instruction).
496
Listing 3.21: MSVC
_a$ = 8 ; size = 4
_f PROC
push ebp
mov ebp, esp
cdq ; sign extend EAX to EDX:EAX
mov ecx, 9
idiv ecx
pop ebp
ret 0
_f ENDP
IDIV divides the 64-bit number stored in the EDX:EAX register pair by the value in the ECX. As a result,
EAX will contain the quotient, and EDX— the remainder. The result is returned from the f() function in the
EAX register, so the value is not moved after the division operation, it is in right place already.
Since IDIV uses the value in the EDX:EAX register pair, the CDQ instruction (before IDIV) extends the value
in EAX to a 64-bit value taking its sign into account, just as MOVSX does.
If we turn optimization on (/Ox), we get:

_a$ = 8 ; size = 4
_f PROC

mov eax, 954437177 ; 38e38e39H
imul ecx
sar edx, 1
mov eax, edx
shr eax, 31 ; 0000001fH
add eax, edx
ret 0
_f ENDP
This is division by multiplication. Multiplication operations work much faster. And it is possible to use this
trick 11 to produce code which is effectively equivalent and faster.
This is also called “strength reduction” in compiler optimizations.
GCC 4.4.1 generates almost the same code even without additional optimization flags, just like MSVC with
optimization turned on:
Listing 3.23: Non-optimizing GCC 4.4.1

public f
f proc near
arg_0 = dword ptr 8
push ebp
mov ebp, esp
mov edx, 954437177 ; 38E38E39h
mov eax, ecx
imul edx
sar edx, 1
mov eax, ecx
sar eax, 1Fh
mov ecx, edx
sub ecx, eax
mov eax, ecx
pop ebp
retn
f endp
11 Read more about division by multiplication in [Henry S. Warren, Hacker’s Delight, (2002)10-3]
497
3.12.2 How it works
From school-level mathematics, we can remember that division by 9 can be replaced by multiplication by
1
9
. In fact, sometimes compilers do so for floating-point arithmetics, for example, FDIV instruction in x86
code can be replaced by FMUL. At least MSVC 6.0 will replace division by 9 by multiplication by 0.111111...
and sometimes it’s hard to be sure, what operation was in the original source code.
But when we operate over integer values and integer CPU registers, we can’t use fractions. However, we
can rework fraction like that:
result = x
9
=x⋅ 1
9
=x⋅ 1⋅M agicN umber
9⋅M agicN umber
Given the fact that division by 2n is very fast (using shifts), we now should find that M agicN umber, for
which the following equation will be true: 2n = 9 ⋅ M agicN umber.
Division by 232 is somewhat hidden: lower 32-bit of product in EAX is not used (dropped), only higher
32-bit of product (in EDX) is used and then shifted by additional 1 bit.
954437177 232+1
In other words, the assembly code we have just seen multiplicates by 232+1 , or divides by 954437177 .
To find a divisor we just have to divide numerator by denominator. Using Wolfram Alpha, we can get
8.99999999.... as result (which is close to 9).
Read more about it in [Henry S. Warren, Hacker’s Delight, (2002)10-3].
Many people miss “hidden” division by 232 or 264 , when lower 32-bit part (or 64-bit part) of product is not
used. This is why division by multiplication is difficult to understand at the beginning.
Mathematics for Programmers12 has yet another explanation.
3.12.3 ARM
The ARM processor, just like in any other “pure” RISC processor lacks an instruction for division. It also
lacks a single instruction for multiplication by a 32-bit constant (recall that a 32-bit constant cannot fit
into a 32-bit opcode).
By taking advantage of this clever trick (or hack), it is possible to do division using only three instructions:
addition, subtraction and bit shifts ( 1.28 on page 303).
Here is an example that divides a 32-bit number by 10, from [Advanced RISC Machines Ltd, The ARM
Cookbook, (1994)3.3 Division by a Constant]. The output consists of the quotient and the remainder.
; takes argument in a1
; returns quotient in a1, remainder in a2
; cycles could be saved if only divide or remainder is required
SUB a2, a1, #10 ; keep (x-10) for later
SUB a1, a1, a1, lsr #2
ADD a1, a1, a1, lsr #4
ADD a1, a1, a1, lsr #8
ADD a1, a1, a1, lsr #16
MOV a1, a1, lsr #3
ADD a3, a1, a1, asl #2
SUBS a2, a2, a3, asl #1 ; calc (x-10) - (x/10)*10
ADDPL a1, a1, #1 ; fix-up quotient
ADDMI a2, a2, #10 ; fix-up remainder
MOV pc, lr
__text:00002C58 39 1E 08 E3 E3 18 43 E3 MOV R1, 0x38E38E39

__text:00002C60 10 F1 50 E7 SMMUL R0, R0, R1
__text:00002C64 C0 10 A0 E1 MOV R1, R0,ASR#1
__text:00002C68 A0 0F 81 E0 ADD R0, R1, R0,LSR#31
__text:00002C6C 1E FF 2F E1 BX LR
This code is almost the same as the one generated by the optimizing MSVC and GCC.
Apparently, LLVM uses the same algorithm for generating constants.
12 https://yurichev.com/writings/Math-for-programmers.pdf
498
The observant reader may ask, how does MOV writes a 32-bit value in a register, when this is not possible
in ARM mode.
it is impossible indeed, but, as we see, there are 8 bytes per instruction instead of the standard 4, in fact,
there are two instructions.
The first instruction loads 0x8E39 into the low 16 bits of register and the second instruction is MOVT, it
loads 0x383E into the high 16 bits of the register. IDA is fully aware of such sequences, and for the sake
of compactness reduces them to one single “pseudo-instruction”.
The SMMUL (Signed Most Significant Word Multiply) instruction two multiplies numbers, treating them as
signed numbers and leaving the high 32-bit part of result in the R0 register, dropping the low 32-bit part
of the result.
The“MOV R1, R0,ASR#1” instruction is an arithmetic shift right by one bit.
“ADD R0, R1, R0,LSR#31” is R0 = R1 + R0 >> 31
There is no separate shifting instruction in ARM mode. Instead, an instructions like (MOV, ADD, SUB, RSB)13
can have a suffix added, that says if the second operand must be shifted, and if yes, by what value and
how. ASR stands for Arithmetic Shift Right, LSR—Logical Shift Right.
MOV R1, 0x38E38E39

SMMUL.W R0, R0, R1
ASRS R1, R0, #1
ADD.W R0, R1, R0,LSR#31
BX LR
There are separate instructions for shifting in Thumb mode, and one of them is used here—ASRS (arithmetic
shift right).
Non-optimizing Xcode 4.6.3 (LLVM) and Keil 6/2013
Non-optimizing LLVM does not generate the code we saw before in this section, but instead inserts a call
to the library function ___divsi3.
What about Keil: it inserts a call to the library function __aeabi_idivmod in all cases.
3.12.4 MIPS
For some reason, optimizing GCC 4.4.5 generate just a division instruction:

f:
li $v0, 9
bnez $v0, loc_10
div $a0, $v0 ; branch delay slot
break 0x1C00 ; "break 7" in assembly output and objdump
loc_10:
mflo $v0
jr $ra
Here we see here a new instruction: BREAK. It just raises an exception.

In this case, an exception is raised if the divisor is zero (it’s not possible to divide by zero in conventional
math).
But GCC probably did not do very well the optimization job and did not see that $V0 is never zero.
So the check is left here. So if $V0 is zero somehow, BREAK is to be executed, signaling to the OS about
the exception.
Otherwise, MFLO executes, which takes the result of the division from the LO register and copies it in $V0.
13 These instructions are also called “data processing instructions”
499
By the way, as we may know, the MUL instruction leaves the high 32 bits of the result in register HI and
the low 32 bits in register LO.
DIV leaves the result in the LO register, and remainder in the HI register.
If we alter the statement to “a % 9”, the MFHI instruction is to be used here instead of MFLO.
3.12.5 Exercise
3.13 String to number conversion (atoi())

Let’s try to reimplement the standard atoi() C function.

Here is the simplest possible way to read a number represented in ASCII encoding.
It’s not error-prone: a character other than a digit leads to incorrect result.
#include <stdio.h>
int my_atoi (char *s)

{
int rt=0;
while (*s)
{
rt=rt*10 + (*s-'0');
s++;
};
return rt;
};
int main()
{
printf ("%d\n", my_atoi ("1234"));
printf ("%d\n", my_atoi ("1234567890"));
};
So what the algorithm does is just reading digits from left to right.
The zero ASCII character is subtracted from each digit.
The digits from “0” to “9” are consecutive in the ASCII table, so we do not even need to know the exact
value of the “0” character.
All we have to know is that “0” minus “0” is 0, “9” minus “0”’is 9 and so on.
Subtracting “0” from each character results in a number from 0 to 9 inclusive.
Any other character leads to an incorrect result, of course!
Each digit has to be added to the final result (in variable “rt”), but the final result is also multiplied by 10
at each digit.
In other words, the result is shifted left by one position in decimal form on each iteration.
The last digit is added, but there is no shift.

s$ = 8
my_atoi PROC
; load first character
movzx r8d, BYTE PTR [rcx]
500
; EAX is allocated for "rt" variable
; its 0 at start
xor eax, eax
; first character is zero-byte, i.e., string terminator?
; exit then.
test r8b, r8b
je SHORT $LN9@my_atoi
$LL2@my_atoi:
lea edx, DWORD PTR [rax+rax*4]
; EDX=RAX+RAX*4=rt+rt*4=rt*5
movsx eax, r8b
; EAX=input character
; load next character to R8D
movzx r8d, BYTE PTR [rcx+1]
; shift pointer in RCX to the next character:
lea rcx, QWORD PTR [rcx+1]
lea eax, DWORD PTR [rax+rdx*2]
; EAX=RAX+RDX*2=input character + rt*5*2=input character + rt*10
; correct digit by subtracting 48 (0x30 or '0')
add eax, -48 ; ffffffffffffffd0H
; was the last character zero?
test r8b, r8b
; jump to loop begin, if not
jne SHORT $LL2@my_atoi
$LN9@my_atoi:
ret 0
my_atoi ENDP
A character can be loaded in two places: the first character and all subsequent characters. This is arranged
so for loop regrouping.
There is no instruction for multiplication by 10, two LEA instruction do this instead.
MSVC sometimes uses the ADD instruction with a negative constant instead of SUB. This is the case.
It’s very hard to say why this is better then SUB. But MSVC does this often.
Optimizing GCC 4.9.1 x64
Optimizing GCC 4.9.1 is more concise, but there is one redundant RET instruction at the end. One would
be enough.

my_atoi:
; load input character into EDX
movsx edx, BYTE PTR [rdi]
; EAX is allocated for "rt" variable
xor eax, eax
; exit, if loaded character is null byte
test dl, dl
je .L4
.L3:
lea eax, [rax+rax*4]
; EAX=RAX*5=rt*5
; shift pointer to the next character:
add rdi, 1
lea eax, [rdx-48+rax*2]
; EAX=input character - 48 + RAX*2 = input character - '0' + rt*10
; load next character:
; goto loop begin, if loaded character is not null byte
test dl, dl
jne .L3
rep ret
.L4:
rep ret
501
my_atoi PROC
; R1 will contain pointer to character
MOV r1,r0
; R0 will contain "rt" variable
MOV r0,#0
B |L0.28|
|L0.12|
ADD r0,r0,r0,LSL #2
; R0=R0+R0<<2=rt*5
ADD r0,r2,r0,LSL #1
; R0=input character + rt*5<<1 = input character + rt*10
; correct whole thing by subtracting '0' from rt:
SUB r0,r0,#0x30
ADD r1,r1,#1
|L0.28|
; load input character to R2
LDRB r2,[r1,#0]
; is it null byte? if no, jump to loop body.
CMP r2,#0
BNE |L0.12|
; exit if null byte.
; "rt" variable is still in R0 register, ready to be used in caller function
BX lr
ENDP

my_atoi PROC
; R1 will be pointer to the input character
MOVS r1,r0
; R0 is allocated to "rt" variable
MOVS r0,#0
B |L0.16|
|L0.6|
MOVS r3,#0xa
; R3=10
MULS r0,r3,r0
; R0=R3*R0=rt*10
ADDS r1,r1,#1
; correct whole thing by subtracting '0' character from it:
SUBS r0,r0,#0x30
ADDS r0,r2,r0
; rt=R2+R0=input character + (rt*10 - '0')
|L0.16|
; load input character to R2
LDRB r2,[r1,#0]
; is it zero?
CMP r2,#0
; jump to loop body if it is not
BNE |L0.6|
; rt variable in R0 now, ready to be used in caller function
BX lr
ENDP
Interestingly, from school mathematics we may recall that the order of addition and subtraction operations
doesn’t matter.
That’s our case: first, the rt ∗ 10 −′ 0′ expression is computed, then the input character value is added to it.
Indeed, the result is the same, but the compiler did some regrouping.
Optimizing GCC 4.9.1 ARM64
The ARM64 compiler can use the pre-increment instruction suffix:
502
my_atoi:
; load input character into W1
ldrb w1, [x0]
mov x2, x0
; X2=address of input string
; is loaded character zero?
; jump to exit if its so
; W1 will contain 0 in this case.
; it will be reloaded into W0 at L4.
cbz w1, .L4
; W0 will contain "rt" variable
; initialize it at zero:
mov w0, 0
.L3:
; subtract 48 or '0' from input variable and put result into W3:
sub w3, w1, #48
; load next character at address X2+1 into W1 with pre-increment:
ldrb w1, [x2,1]!
add w0, w0, w0, lsl 2
; W0=W0+W0<<2=W0+W0*4=rt*5
add w0, w3, w0, lsl 1
; W0=input digit + W0<<1 = input digit + rt*5*2 = input digit + rt*10
; if the character we just loaded is not null byte, jump to the loop begin
cbnz w1, .L3
; variable to be returned (rt) is in W0, ready to be used in caller function
ret
.L4:
mov w0, w1
ret
3.13.2 A slightly advanced example

My new code snippet is more advanced, now it checks for the “minus” sign at the first character and
reports an error if a non-digit has been found in the input string:
#include <stdio.h>
int my_atoi (char *s)

{
int negative=0;
int rt=0;
if (*s=='-')
{
negative=1;
s++;
};
while (*s)
{
if (*s<'0' || *s>'9')
{
printf ("Error! Unexpected char: '%c'\n", *s);
exit(0);
};
rt=rt*10 + (*s-'0');
s++;
};
if (negative)
return -rt;
return rt;
};
int main()
{
printf ("%d\n", my_atoi ("1234"));
printf ("%d\n", my_atoi ("1234567890"));
503
printf ("%d\n", my_atoi ("-1234"));
printf ("%d\n", my_atoi ("-1234567890"));
printf ("%d\n", my_atoi ("-a1234567890")); // error
};
Optimizing GCC 4.9.1 x64

.LC0:
.string "Error! Unexpected char: '%c'\n"
my_atoi:
sub rsp, 8
; check for minus sign
cmp dl, 45 ; '-'
je .L22
xor esi, esi
test dl, dl
je .L20
.L10:
; ESI=0 here if there was no minus sign and 1 if it was
lea eax, [rdx-48]
; any character other than digit will result in unsigned number greater than 9 after subtraction
; so if it is not digit, jump to L4, where error will be reported:
cmp al, 9
ja .L4
xor eax, eax
jmp .L6
.L7:
lea ecx, [rdx-48]
cmp cl, 9
ja .L4
.L6:
lea eax, [rax+rax*4]
add rdi, 1
lea eax, [rdx-48+rax*2]
test dl, dl
jne .L7
; if there was no minus sign, skip NEG instruction
; if it was, execute it.
test esi, esi
je .L18
neg eax
.L18:
add rsp, 8
ret
.L22:
movsx edx, BYTE PTR [rdi+1]
lea rax, [rdi+1]
test dl, dl
je .L20
mov rdi, rax
mov esi, 1
jmp .L10
.L20:
xor eax, eax
jmp .L18
.L4:
; report error. character is in EDX
mov edi, 1
mov esi, OFFSET FLAT:.LC0 ; "Error! Unexpected char: '%c'\n"
xor eax, eax
call __printf_chk
xor edi, edi
call exit
504
If the “minus” sign has been encountered at the string start, the NEG instruction is to be executed at the
end. It just negates the number.
There is one more thing that needs mentioning.
How would a common programmer check if the character is not a digit? Just how we have it in the source
code:
if (*s<'0' || *s>'9')
...
There are two comparison operations.

What is interesting is that we can replace both operations by single one: just subtract “0” from character
value,
treat result as unsigned value (this is important) and check if it’s greater than 9.
For example, let’s say that the user input contains the dot character (“.”) which has ASCII code 46.
46 − 48 = −2 if we treat the result as a signed number.
Indeed, the dot character is located two places earlier than the “0” character in the ASCII table. But it is
0xFFFFFFFE (4294967294) if we treat the result as an unsigned value, and that’s definitely bigger than 9!
The compilers do this often, so it’s important to recognize these tricks.
Another example of it in this book: 3.19.1 on page 534.
Optimizing MSVC 2013 x64 does the same tricks.

1 my_atoi PROC
2 PUSH {r4-r6,lr}
3 MOV r4,r0
4 LDRB r0,[r0,#0]
5 MOV r6,#0
6 MOV r5,r6
7 CMP r0,#0x2d '-'
8 ; R6 will contain 1 if minus was encountered, 0 if otherwise
9 MOVEQ r6,#1
10 ADDEQ r4,r4,#1
11 B |L0.80|
12 |L0.36|
13 SUB r0,r1,#0x30
14 CMP r0,#0xa
15 BCC |L0.64|
16 ADR r0,|L0.220|
17 BL __2printf
18 MOV r0,#0
19 BL exit
20 |L0.64|
21 LDRB r0,[r4],#1
22 ADD r1,r5,r5,LSL #2
23 ADD r0,r0,r1,LSL #1
24 SUB r5,r0,#0x30
25 |L0.80|
26 LDRB r1,[r4,#0]
27 CMP r1,#0
28 BNE |L0.36|
29 CMP r6,#0
30 ; negate result
31 RSBNE r0,r5,#0
32 MOVEQ r0,r5
33 POP {r4-r6,pc}
34 ENDP
35
36 |L0.220|
37 DCB "Error! Unexpected char: '%c'\n",0
505
There is no NEG instruction in 32-bit ARM, so the “Reverse Subtraction” operation (line 31) is used here.
It is triggered if the result of the CMP instruction (at line 29) has been “Not Equal” (hence -NE suffix).
So what RSBNE does is to subtract the resulting value from 0.
It works just like the regular subtraction operation, but swaps operands.
Subtracting any number from 0 results in negation: 0 − x = −x.
Thumb mode code is mostly the same.
GCC 4.9 for ARM64 can use the NEG instruction, which is available in ARM64.
3.13.3 Exercise
Oh, by the way, security researchers deals often with unpredictable behavior of program while handling
of incorrect data.
For example, while fuzzing. As an exercise, you may try to enter non-digit characters and see what
happens.
Try to explain, what happened and why.
3.14 Inline functions

Inlined code is when the compiler, instead of placing a call instruction to a small or tiny function, just
places its body right in-place.
Listing 3.32: A simple example

#include <stdio.h>
int celsius_to_fahrenheit (int celsius)

{
return celsius * 9 / 5 + 32;
};
int main(int argc, char *argv[])

{
int celsius=atol(argv[1]);
printf ("%d\n", celsius_to_fahrenheit (celsius));
};
…is compiled in very predictable way, however, if we turn on GCC optimizations (-O3), we’ll see:

_main:
push ebp
mov ebp, esp
and esp, -16
sub esp, 16
call ___main
mov eax, DWORD PTR [ebp+12]
mov eax, DWORD PTR [eax+4]
call _atol
mov edx, 1717986919
mov DWORD PTR [esp], OFFSET FLAT:LC2 ; "%d\n"
lea ecx, [eax+eax*8]
mov eax, ecx
imul edx
sar ecx, 31
sar edx
sub edx, ecx
add edx, 32
call _printf
leave
ret
506
(Here the division is performed by multiplication( 3.12 on page 496).)
Yes, our small function celsius_to_fahrenheit() has just been placed before the printf() call.
Why? It can be faster than executing this function’s code plus the overhead of calling/returning.
Modern optimizing compilers are choosing small functions for inlining automatically. But it’s possible to
force compiler additionally to inline some function, if to mark it with the “inline” keyword in its declaration.
3.14.1 Strings and memory functions

Another very common automatic optimization tactic is the inlining of string functions like strcpy(), str-
cmp(), strlen(), memset(), memcmp(), memcpy(), etc..
Sometimes it’s faster than to call a separate function.
These are very frequent patterns and it is highly advisable for reverse engineers to learn to detect auto-
matically.
strcmp()
Listing 3.34: strcmp() example

bool is_bool (char *s)
{
if (strcmp (s, "true")==0)
return true;
if (strcmp (s, "false")==0)
return false;
assert(0);
};

.LC0:
.string "true"
.LC1:
.string "false"
is_bool:
.LFB0:
push edi
mov ecx, 5
push esi
sub esp, 20
repz cmpsb
je .L3
mov ecx, 6
repz cmpsb
seta cl
setb dl
xor eax, eax
cmp cl, dl
jne .L8
add esp, 20
pop esi
pop edi
ret
.L8:
mov DWORD PTR [esp], 0
call assert
add esp, 20
pop esi
pop edi
ret
.L3:
add esp, 20
507
mov eax, 1
pop esi
pop edi
ret

$SG3454 DB 'true', 00H
$SG3456 DB 'false', 00H
_s$ = 8 ; size = 4
?is_bool@@YA_NPAD@Z PROC ; is_bool
push esi
mov esi, DWORD PTR _s$[esp]
mov ecx, OFFSET $SG3454 ; 'true'
mov eax, esi
$LL6@is_bool:
mov dl, BYTE PTR [eax]
cmp dl, BYTE PTR [ecx]
jne SHORT $LN7@is_bool
test dl, dl
je SHORT $LN8@is_bool
mov dl, BYTE PTR [eax+1]
cmp dl, BYTE PTR [ecx+1]
add eax, 2
add ecx, 2
test dl, dl
jne SHORT $LL6@is_bool
$LN8@is_bool:
xor eax, eax
jmp SHORT $LN9@is_bool
$LN7@is_bool:
sbb eax, eax
sbb eax, -1
$LN9@is_bool:
test eax, eax
mov al, 1
pop esi
ret 0
$LN2@is_bool:
mov ecx, OFFSET $SG3456 ; 'false'

mov eax, esi
$LL10@is_bool:
mov dl, BYTE PTR [eax]
cmp dl, BYTE PTR [ecx]
test dl, dl
je SHORT $LN12@is_bool
mov dl, BYTE PTR [eax+1]
cmp dl, BYTE PTR [ecx+1]
add eax, 2
add ecx, 2
test dl, dl
jne SHORT $LL10@is_bool
$LN12@is_bool:
xor eax, eax
jmp SHORT $LN13@is_bool
$LN11@is_bool:
sbb eax, eax
sbb eax, -1
$LN13@is_bool:
test eax, eax
508
xor al, al
pop esi
ret 0
$LN1@is_bool:
push 11
push OFFSET $SG3458
push OFFSET $SG3459
call DWORD PTR __imp___wassert
add esp, 12
pop esi
ret 0
?is_bool@@YA_NPAD@Z ENDP ; is_bool
strlen()
Listing 3.37: strlen() example

int strlen_test(char *s1)
{
return strlen(s1);
};

_s1$ = 8 ; size = 4
_strlen_test PROC
mov eax, DWORD PTR _s1$[esp-4]
lea edx, DWORD PTR [eax+1]
$LL3@strlen_tes:
inc eax
test cl, cl
jne SHORT $LL3@strlen_tes
sub eax, edx
ret 0
_strlen_test ENDP
strcpy()
Listing 3.39: strcpy() example

void strcpy_test(char *s1, char *outbuf)
{
strcpy(outbuf, s1);
};

_s1$ = 8 ; size = 4
_outbuf$ = 12 ; size = 4
_strcpy_test PROC
mov eax, DWORD PTR _s1$[esp-4]
mov edx, DWORD PTR _outbuf$[esp-4]
sub edx, eax
$LL3@strcpy_tes:
mov BYTE PTR [edx+eax], cl
inc eax
test cl, cl
jne SHORT $LL3@strcpy_tes
ret 0
_strcpy_test ENDP
509
memset()
Example#1
Listing 3.41: 32 bytes

#include <stdio.h>
void f(char *out)

{
memset(out, 0, 32);
};
Many compilers don’t generate a call to memset() for short blocks, but rather insert a pack of MOVs:

f:
mov QWORD PTR [rdi], 0
mov QWORD PTR [rdi+8], 0
ret
By the way, that remind us of unrolled loops: 1.22.1 on page 191.
Example#2
Listing 3.43: 67 bytes

#include <stdio.h>
void f(char *out)

{
memset(out, 0, 67);
};
When the block size is not a multiple of 4 or 8, the compilers can behave differently.
For instance, MSVC 2012 continues to insert MOVs:

out$ = 8
f PROC
xor eax, eax
mov QWORD PTR [rcx], rax
mov QWORD PTR [rcx+8], rax
mov WORD PTR [rcx+64], ax
mov BYTE PTR [rcx+66], al
ret 0
f ENDP
…while GCC uses REP STOSQ, concluding that this would be shorter than a pack of MOVs:

f:
mov QWORD PTR [rdi], 0
mov rcx, rdi
lea rdi, [rdi+8]
xor eax, eax
and rdi, -8
510
sub rcx, rdi
add ecx, 67
shr ecx, 3
rep stosq
ret
memcpy()
Short blocks
The routine to copy short blocks is often implemented as a sequence of MOV instructions.
Listing 3.46: memcpy() example

void memcpy_7(char *inbuf, char *outbuf)
{
memcpy(outbuf+10, inbuf, 7);
};

_inbuf$ = 8 ; size = 4
_memcpy_7 PROC
mov ecx, DWORD PTR _inbuf$[esp-4]
mov edx, DWORD PTR [ecx]
mov eax, DWORD PTR _outbuf$[esp-4]
mov dx, WORD PTR [ecx+4]
mov WORD PTR [eax+14], dx
mov cl, BYTE PTR [ecx+6]
mov BYTE PTR [eax+16], cl
ret 0
_memcpy_7 ENDP

memcpy_7:
push ebx
mov ecx, DWORD PTR [esp+12]
mov ebx, DWORD PTR [eax]
lea edx, [ecx+10]
mov DWORD PTR [ecx+10], ebx
movzx ecx, WORD PTR [eax+4]
mov WORD PTR [edx+4], cx
movzx eax, BYTE PTR [eax+6]
mov BYTE PTR [edx+6], al
pop ebx
ret
That’s usually done as follows: 4-byte blocks are copied first, then a 16-bit word (if needed), then the last
byte (if needed).
Structures are also copied using MOV: 1.30.4 on page 358.
Long blocks
The compilers behave differently in this case.
Listing 3.49: memcpy() example

{
};
511
{
};
For copying 128 bytes, MSVC uses a single MOVSD instruction (because 128 divides evenly by 4):

_memcpy_128 PROC
push esi
mov esi, DWORD PTR _inbuf$[esp]
push edi
mov edi, DWORD PTR _outbuf$[esp+4]
add edi, 10
mov ecx, 32
rep movsd
pop edi
pop esi
ret 0
_memcpy_128 ENDP
When copying 123 bytes, 30 32-bit words are copied first using MOVSD (that’s 120 bytes), then 2 bytes are
copied using MOVSW, then one more byte using MOVSB.

_memcpy_123 PROC
push esi
mov esi, DWORD PTR _inbuf$[esp]
push edi
mov edi, DWORD PTR _outbuf$[esp+4]
add edi, 10
mov ecx, 30
rep movsd
movsw
movsb
pop edi
pop esi
ret 0
_memcpy_123 ENDP
GCC uses one big universal functions, that works for any block size:

memcpy_123:
.LFB3:
push edi
mov eax, 123
push esi
lea edi, [edx+10]
test edi, 1
jne .L24
test edi, 2
jne .L25
.L7:
mov ecx, eax
xor edx, edx
shr ecx, 2
test al, 2
rep movsd
je .L8
mov WORD PTR [edi], dx
mov edx, 2
512
.L8:
test al, 1
je .L5
movzx eax, BYTE PTR [esi+edx]
mov BYTE PTR [edi+edx], al
.L5:
pop esi
pop edi
ret
.L24:
movzx eax, BYTE PTR [esi]
lea edi, [edx+11]
add esi, 1
test edi, 2
mov BYTE PTR [edx+10], al
mov eax, 122
je .L7
.L25:
add edi, 2
add esi, 2
sub eax, 2
mov WORD PTR [edi-2], dx
jmp .L7
.LFE3:
Universal memory copy functions usually work as follows: calculate how many 32-bit words can be copied,
then copy them using MOVSD, then copy the remaining bytes.
More advanced and complex copy functions use SIMD instructions and also take the memory alignment
in consideration.
As an example of SIMD strlen() function: 1.36.2 on page 414.
memcmp()
Listing 3.53: memcmp() example

int memcmp_1235(char *buf1, char *buf2)
{
return memcmp(buf1, buf2, 1235);
};
For any block size, MSVC 2013 inserts the same universal function:

_buf1$ = 8 ; size = 4
_buf2$ = 12 ; size = 4
_memcmp_1235 PROC
mov ecx, DWORD PTR _buf1$[esp-4]
mov edx, DWORD PTR _buf2$[esp-4]
push esi
mov esi, 1231
npad 2
$LL5@memcmp_123:
mov eax, DWORD PTR [ecx]
cmp eax, DWORD PTR [edx]
jne SHORT $LN4@memcmp_123
add ecx, 4
add edx, 4
sub esi, 4
jae SHORT $LL5@memcmp_123
$LN4@memcmp_123:
mov al, BYTE PTR [ecx]
cmp al, BYTE PTR [edx]
mov al, BYTE PTR [ecx+1]
cmp al, BYTE PTR [edx+1]
513
cmp esi, -1
je SHORT $LN3@memcmp_123
$LN3@memcmp_123:
xor eax, eax
pop esi
ret 0
$LN6@memcmp_123:
sbb eax, eax
or eax, 1
pop esi
ret 0
_memcmp_1235 ENDP
strcat()
This is inlined strcat() as it has been generated by MSVC 6.0. There are 3 parts visible: 1) getting source
string length (first scasb); 2) getting destination string length (second scasb); 3) copying source string
into the end of destination string (movsd/movsb pair).
Listing 3.55: strcat()

lea edi, [src]
or ecx, 0FFFFFFFFh
repne scasb
not ecx
sub edi, ecx
mov esi, edi
mov edi, [dst]
mov edx, ecx
or ecx, 0FFFFFFFFh
repne scasb
mov ecx, edx
dec edi
shr ecx, 2
rep movsd
mov ecx, edx
and ecx, 3
rep movsb
IDA script
There is also a small IDA script for searching and folding such very frequently seen pieces of inline code:
GitHub.
3.15 C99 restrict

Here is a reason why Fortran programs, in some cases, work faster than C/C++ ones.
void f1 (int* x, int* y, int* sum, int* product, int* sum_product, int* update_me, size_t s)
{
for (int i=0; i<s; i++)
{
sum[i]=x[i]+y[i];
product[i]=x[i]*y[i];
update_me[i]=i*123; // some dummy value
sum_product[i]=sum[i]+product[i];
};
};
514
That’s very simple example with one specific thing in it: the pointer to the update_me array could be a
pointer to the sum array, product array, or even the sum_product array—nothing forbids that, right?
The compiler is fully aware of this, so it generates code with four stages in the loop body:
• calculate next sum[i]
• calculate next product[i]
• calculate next update_me[i]
• calculate next sum_product[i]—on this stage, we need to load from memory the already calculated
sum[i] and product[i]
Is it possible to optimize the last stage? Since we have already calculated sum[i] and product[i] it is
not necessary to load them again from memory.
Yes, but compiler is not sure that nothing has been overwritten at the 3rd stage! This is called “pointer
aliasing”, a situation when the compiler cannot be sure that a memory to which a pointer is pointing hasn’t
been changed.
restrict in the C99 standard [ISO/IEC 9899:TC3 (C C99 standard), (2007) 6.7.3/1] is a promise, given by
programmer to the compiler that the function arguments marked by this keyword always points to different
memory locations and never intersects.
To be more precise and describe this formally, restrict shows that only this pointer is to be used to access
an object, and no other pointer will be used for it.
It can be even said the object will be accessed only via one single pointer, if it is marked as restrict.
Let’s add this keyword to each pointer argument:
void f2 (int* restrict x, int* restrict y, int* restrict sum, int* restrict product, int* ⤦
Ç restrict sum_product,
int* restrict update_me, size_t s)
{
for (int i=0; i<s; i++)
{
sum[i]=x[i]+y[i];
product[i]=x[i]*y[i];
update_me[i]=i*123; // some dummy value
sum_product[i]=sum[i]+product[i];
};
};
Let’s see results:
Listing 3.56: GCC x64: f1()

f1:
push r15 r14 r13 r12 rbp rdi rsi rbx
mov r13, QWORD PTR 120[rsp]
mov rbp, QWORD PTR 104[rsp]
test r13, r13
je .L1
add r13, 1
xor ebx, ebx
mov edi, 1
xor r11d, r11d
jmp .L4
.L6:
mov r11, rdi
mov rdi, rax
.L4:
lea rax, 0[0+r11*4]
lea r10, [rcx+rax]
lea r14, [rdx+rax]
lea rsi, [r8+rax]
add rax, r9
mov r15d, DWORD PTR [r10]
add r15d, DWORD PTR [r14]
mov DWORD PTR [rsi], r15d ; store to sum[]
mov r10d, DWORD PTR [r10]
515
imul r10d, DWORD PTR [r14]
mov DWORD PTR [rax], r10d ; store to product[]
mov DWORD PTR [r12+r11*4], ebx ; store to update_me[]
add ebx, 123
mov r10d, DWORD PTR [rsi] ; reload sum[i]
add r10d, DWORD PTR [rax] ; reload product[i]
lea rax, 1[rdi]
cmp rax, r13
mov DWORD PTR 0[rbp+r11*4], r10d ; store to sum_product[]
jne .L6
.L1:
pop rbx rsi rdi rbp r12 r13 r14 r15
ret
Listing 3.57: GCC x64: f2()

f2:
push r13 r12 rbp rdi rsi rbx
mov rbp, QWORD PTR 88[rsp]
test r13, r13
je .L7
add r13, 1
xor r10d, r10d
mov edi, 1
xor eax, eax
jmp .L10
.L11:
mov rax, rdi
mov rdi, r11
.L10:
mov esi, DWORD PTR [rcx+rax*4]
mov r11d, DWORD PTR [rdx+rax*4]
mov DWORD PTR [r12+rax*4], r10d ; store to update_me[]
add r10d, 123
lea ebx, [rsi+r11]
imul r11d, esi
mov DWORD PTR [r8+rax*4], ebx ; store to sum[]
mov DWORD PTR [r9+rax*4], r11d ; store to product[]
add r11d, ebx
mov DWORD PTR 0[rbp+rax*4], r11d ; store to sum_product[]
lea r11, 1[rdi]
cmp r11, r13
jne .L11
.L7:
pop rbx rsi rdi rbp r12 r13
ret
The difference between the compiled f1() and f2() functions is as follows: in f1(), sum[i] and product[i]
are reloaded in the middle of the loop, and in f2() there is no such thing, the already calculated values
are used, since we “promised” the compiler that no one and nothing will change the values in sum[i] and
product[i] during the execution of the loop’s body, so it is “sure” that there is no need to load the value
from memory again.
Obviously, the second example works faster.
But what if the pointers in the function’s arguments intersect somehow?
This is on the programmer’s conscience, and the results will be incorrect.
Let’s go back to Fortran.
Compilers of this programming language treats all pointers as such, so when it was not possible to set
restrict in C, Fortran could generate faster code in these cases.
How practical is it?
In the cases when the function works with several big blocks in memory.
There are a lot of such in linear algebra, for instance.
516
Supercomputers/HPC14 are very busy with linear algebra, so probably that is why, traditionally, Fortran is
still used there [Eugene Loh, The Ideal HPC Programming Language, (2010)].
But when the number of iterations is not very big, certainly, the speed boost may not to be significant.
3.16 Branchless abs() function

Let’s revisit an example we considered earlier 1.18.2 on page 140 and ask ourselves, is it possible to make
a branchless version of the function in x86 code?
int my_abs (int i)
{
if (i<0)
return -i;
else
return i;
};
And the answer is yes.
3.16.1 Optimizing GCC 4.9.1 x64

We could see it if we compile it using optimizing GCC 4.9:

my_abs:
mov edx, edi
mov eax, edi
sar edx, 31
; EDX is 0xFFFFFFFF here if sign of input value is minus
; EDX is 0 if sign of input value is plus (including 0)
; the following two instructions have effect only if EDX is 0xFFFFFFFF
; or idle if EDX is 0
xor eax, edx
sub eax, edx
ret
This is how it works:

Arithmetically shift the input value right by 31.
Arithmetical shift implies sign extension, so if the MSB is 1, all 32 bits are to be filled with 1, or with 0 if
otherwise.
In other words, the SAR REG, 31 instruction makes 0xFFFFFFFF if the sign has been negative or 0 if
positive.
After the execution of SAR, we have this value in EDX.
Then, if the value is 0xFFFFFFFF (i.e., the sign is negative), the input value is inverted
(because XOR REG, 0xFFFFFFFF is effectively an inverse all bits operation).
Then, again, if the value is 0xFFFFFFFF (i.e., the sign is negative), 1 is added to the final result (because
subtracting −1 from some value resulting in incrementing it).
Inversion of all bits and incrementing is exactly how two’s complement value is negated: 2.2 on page 448.
We may observe that the last two instruction do something if the sign of the input value is negative.
Otherwise (if the sign is positive) they do nothing at all, leaving the input value untouched.
The algorithm is explained in [Henry S. Warren, Hacker’s Delight, (2002)2-4].
It’s hard to say, how GCC did it, deduced it by itself or found a suitable pattern among known ones?
14 High-Performance Computing
517
3.16.2 Optimizing GCC 4.9 ARM64
GCC 4.9 for ARM64 generates mostly the same, just decides to use the full 64-bit registers.
There are less instructions, because the input value can be shifted using a suffixed instruction (“asr”)
instead of using a separate instruction.

my_abs:
; sign-extend input 32-bit value to X0 64-bit register:
sxtw x0, w0
eor x1, x0, x0, asr 63
; X1=X0^(X0>>63) (shift is arithmetical)
sub x0, x1, x0, asr 63
; X0=X1-(X0>>63)=X0^(X0>>63)-(X0>>63) (all shifts are arithmetical)
ret
3.17 Variadic functions

Functions like printf() and scanf() can have a variable number of arguments. How are these arguments
accessed?
3.17.1 Computing arithmetic mean

Let’s imagine that we want to calculate arithmetic mean, and for some weird reason we want to specify
all the values as function arguments.
But it’s impossible to get the number of arguments in a variadic function in C/C++, so let’s denote the
value of −1 as a terminator.
Using va_arg macro
There is the standard stdarg.h header file which define macros for dealing with such arguments.
The printf() and scanf() functions use them as well.
#include <stdio.h>
#include <stdarg.h>
int arith_mean(int v, ...)

{
va_list args;
int sum=v, count=1, i;
va_start(args, v);
while(1)
{
i=va_arg(args, int);
if (i==-1) // terminator
break;
sum=sum+i;
count++;
}
va_end(args);
return sum/count;
};
int main()
{
printf ("%d\n", arith_mean (1, 2, 7, 10, 15, -1 /* terminator */));
};
The first argument has to be treated just like a normal argument.

All other arguments are loaded using the va_arg macro and then summed.
So what is inside?
518
cdecl calling conventions
Listing 3.60: Optimizing MSVC 6.0

_v$ = 8
_arith_mean PROC NEAR
mov eax, DWORD PTR _v$[esp-4] ; load 1st argument into sum
push esi
mov esi, 1 ; count=1
lea edx, DWORD PTR _v$[esp] ; address of the 1st argument
$L838:
mov ecx, DWORD PTR [edx+4] ; load next argument
add edx, 4 ; shift pointer to the next argument
cmp ecx, -1 ; is it -1?
je SHORT $L856 ; exit if so
add eax, ecx ; sum = sum + loaded argument
inc esi ; count++
jmp SHORT $L838
$L856:
; calculate quotient
cdq
idiv esi
pop esi
ret 0
_arith_mean ENDP
$SG851 DB '%d', 0aH, 00H
_main PROC NEAR

push -1
push 15
push 10
push 7
push 2
push 1
call _arith_mean
push eax
push OFFSET FLAT:$SG851 ; '%d'
call _printf
add esp, 32
ret 0
_main ENDP
The arguments, as we may see, are passed to main() one-by-one.

The first argument is pushed into the local stack as first.
The terminating value (−1) is pushed last.
The arith_mean() function takes the value of the first argument and stores it in the sum variable.
Then, it sets the EDX register to the address of the second argument, takes the value from it, adds it to
sum, and does this in an infinite loop, until −1 is found.
When it’s found, the sum is divided by the number of all values (excluding −1) and the quotient is returned.
So, in other words, the function treats the stack fragment as an array of integer values of infinite length.
Now we can understand why the cdecl calling convention forces us to push the first argument into the
stack as last.
Because otherwise, it would not be possible to find the first argument, or, for printf-like functions, it would
not be possible to find the address of the format-string.
Register-based calling conventions
The observant reader may ask, what about calling conventions where the first few arguments are passed
in registers? Let’s see:
519
$SG3013 DB '%d', 0aH, 00H
v$ = 8
arith_mean PROC
mov DWORD PTR [rsp+8], ecx ; 1st argument
mov QWORD PTR [rsp+16], rdx ; 2nd argument
mov QWORD PTR [rsp+24], r8 ; 3rd argument
mov eax, ecx ; sum = 1st argument
lea rcx, QWORD PTR v$[rsp+8] ; pointer to the 2nd argument
mov QWORD PTR [rsp+32], r9 ; 4th argument
mov edx, DWORD PTR [rcx] ; load 2nd argument
mov r8d, 1 ; count=1
cmp edx, -1 ; 2nd argument is -1?
je SHORT $LN8@arith_mean ; exit if so
$LL3@arith_mean:
add eax, edx ; sum = sum + loaded argument
mov edx, DWORD PTR [rcx+8] ; load next argument
lea rcx, QWORD PTR [rcx+8] ; shift pointer to point to the argument after next
inc r8d ; count++
cmp edx, -1 ; is loaded argument -1?
jne SHORT $LL3@arith_mean ; go to loop begin if its not
$LN8@arith_mean:
; calculate quotient
cdq
idiv r8d
ret 0
arith_mean ENDP
main PROC
sub rsp, 56
mov edx, 2
mov DWORD PTR [rsp+40], -1
lea r9d, QWORD PTR [rdx+8]
lea ecx, QWORD PTR [rdx-1]
call arith_mean
mov edx, eax
call printf
xor eax, eax
add rsp, 56
ret 0
main ENDP
We see that the first 4 arguments are passed in the registers and two more—in the stack.
The arith_mean() function first places these 4 arguments into the Shadow Space and then treats the
Shadow Space and stack behind it as a single continuous array!
What about GCC? Things are slightly clumsier here, because now the function is divided in two parts: the
first part saves the registers into the “red zone”, processes that space, and the second part of the function
processes the stack:
arith_mean:
lea rax, [rsp+8]
; save 6 input registers in
; red zone in the local stack
mov QWORD PTR [rsp-40], rsi
mov QWORD PTR [rsp-32], rdx
mov QWORD PTR [rsp-16], r8
mov QWORD PTR [rsp-24], rcx
mov esi, 8
mov QWORD PTR [rsp-64], rax
lea rax, [rsp-48]
mov QWORD PTR [rsp-8], r9
mov DWORD PTR [rsp-72], 8
lea rdx, [rsp+8]
520
mov r8d, 1
mov QWORD PTR [rsp-56], rax
jmp .L5
.L7:
; work out saved arguments
lea rax, [rsp-48]
mov ecx, esi
add esi, 8
add rcx, rax
mov ecx, DWORD PTR [rcx]
cmp ecx, -1
je .L4
.L8:
add edi, ecx
add r8d, 1
.L5:
; decide, which part we will work out now.
; is current argument number less or equal 6?
cmp esi, 47
jbe .L7 ; no, process saved arguments then
; work out arguments from stack
mov rcx, rdx
add rdx, 8
mov ecx, DWORD PTR [rcx]
cmp ecx, -1
jne .L8
.L4:
mov eax, edi
cdq
idiv r8d
ret
.LC1:
.string "%d\n"
main:
sub rsp, 8
mov edx, 7
mov esi, 2
mov edi, 1
mov r9d, -1
mov r8d, 15
mov ecx, 10
xor eax, eax
call arith_mean
mov edx, eax
mov edi, 1
xor eax, eax
add rsp, 8
jmp __printf_chk
By the way, a similar usage of the Shadow Space is also considered here: 6.1.8 on page 731.
Using pointer to the first function argument
The example can be rewritten without va_arg macro:

#include <stdio.h>
int arith_mean(int v, ...)

{
int *i=&v;
int sum=*i, count=1;
i++;
while(1)
{
if ((*i)==-1) // terminator
break;
521
sum=sum+(*i);
count++;
i++;
}
return sum/count;
};
int main()
{
printf ("%d\n", arith_mean (1, 2, 7, 10, 15, -1 /* terminator */));
// test: https://www.wolframalpha.com/input/?i=mean(1,2,7,10,15)
};
In other words, if an argument set is array of words (32-bit or 64-bit), we just enumerate array elements
starting at first one.
3.17.2 vprintf() function case

Many programmers define their own logging functions which take a printf-like format string + a variable
number of arguments.
Another popular example is the die() function, which prints some message and exits.
We need some way to pack input arguments of unknown number and pass them to the printf() function.
But how?
That’s why there are functions with “v” in name.
One of them is vprintf(): it takes a format-string and a pointer to a variable of type va_list:
#include <stdlib.h>
#include <stdarg.h>
void die (const char * fmt, ...)

{
va_list va;
va_start (va, fmt);
vprintf (fmt, va);

exit(0);
};
By closer examination, we can see that va_list is a pointer to an array. Let’s compile:

_fmt$ = 8
_die PROC
; load 1st argument (format-string)
mov ecx, DWORD PTR _fmt$[esp-4]
; get pointer to the 2nd argument
lea eax, DWORD PTR _fmt$[esp]
push eax ; pass a pointer
push ecx
call _vprintf
add esp, 8
push 0
call _exit
$LN3@die:
int 3
_die ENDP
We see that all our function does is just taking a pointer to the arguments and passing it to vprintf(), and
that function is treating it like an infinite array of arguments!

fmt$ = 48
die PROC
; save first 4 arguments in Shadow Space
522
mov QWORD PTR [rsp+8], rcx
mov QWORD PTR [rsp+16], rdx
mov QWORD PTR [rsp+24], r8
mov QWORD PTR [rsp+32], r9
sub rsp, 40
lea rdx, QWORD PTR fmt$[rsp+8] ; pass pointer to the 1st argument
; RCX here is still points to the 1st argument (format-string) of die()
; so vprintf() will take it right from RCX
call vprintf
xor ecx, ecx
call exit
int 3
die ENDP
3.17.3 Pin case

It’s interesting to note how some functions from Pin DBI15 framework takes number of arguments:
INS_InsertPredicatedCall(
ins, IPOINT_BEFORE, (AFUNPTR)RecordMemRead,
IARG_INST_PTR,
IARG_MEMORYOP_EA, memOp,
IARG_END);
( pinatrace.cpp )
And this is how INS_InsertPredicatedCall() function is declared:
extern VOID INS_InsertPredicatedCall(INS ins, IPOINT ipoint, AFUNPTR funptr, ...);
( pin_client.PH )
Hence, constants with names starting with IARG_ are some kinds of arguments to the function, which are
handled inside of INS_InsertPredicatedCall(). You can pass as many arguments, as you need. Some
commands has additional argument(s), some are not. Full list of arguments: https://software.intel.
com/sites/landingpage/pintool/docs/58423/Pin/html/group__INST__ARGS.html. And it has to be
a way to detect an end of arguments list, so the list must be terminated with IARG_END constant, without
which, the function will (try to) handle random noise in the local stack, treating it as additional arguments.
Also, in [Brian W. Kernighan, Rob Pike, Practice of Programming, (1999)] we can find a nice example of
C/C++ routines very similar to pack/unpack 16 in Python.
3.17.4 Format string exploit

It’s a popular mistake, to write printf(string) instead of puts(string) or printf("%s", string). If
the attacker can put his/her own text into string, he/she can crash process, or get insight into variables
in the local stack.
Take a look at this:
#include <stdio.h>
int main()
{
char *s1="hello";
char *s2="world";
char buf[128];
// do something mundane here

strcpy (buf, s1);
strcpy (buf, " ");
strcpy (buf, s2);
printf ("%s");
};
15 Dynamic Binary Instrumentation

16 https://docs.python.org/3/library/struct.html
523
Please note, that printf() has no additional arguments besides single format string.
Now let’s imagine, that was the attacker who put %s string into the last printf() first arguments. I
compile this example using GCC 5.4.0 on x86 Ubuntu, and the resulting executable prints “world” string
if it gets executed!
If I turn optimization on, printf() outputs some garbage, though—probably, strcpy() calls has been
optimized and/or local variables as well. Also, result will be different for x64 code, different compiler, OS,
etc.
Now, let’s say, attacker could pass the following string to printf() call: %x %x %x %x %x. In may case,
output is: “80485c6 b7751b48 1 0 80485c0” (these are just values from local stack). You see, there are 1
and 0 values, and some pointers (first is probably pointer to “world” string). So if the attacker passes %s
%s %s %s %s string, the process will crash, because printf() treats 1 and/or 0 as pointer to string, tries
to read characters from there and fails.
Even worse, there could be sprintf (buf, string) in code, where buf is a buffer in the local stack with
size of 1024 bytes or so, attacker can craft string in such a way that buf will be overflown, maybe even
in a way that would lead to code execution.
Many popular and well-known software was (or even still) vulnerable:
QuakeWorld went up, got to around 4000 users, then the master server exploded.
Disrupter and cohorts are working on more robust code now.
If anyone did it on purpose, how about letting us know... (It wasn’t all the people that
tried %s as a name)
( John Carmack’s .plan file, 17-Dec-199617 )

Nowadays, almost all decent compilers warn about this.
Another problem is the lesser known %n printf() argument: whenever printf() reaches it in a for-
mat string, it writes the number of characters printed so far into the corresponding argument: http://
stackoverflow.com/questions/3401156/what-is-the-use-of-the-n-format-specifier-in-c. Thus,
an attacker could zap local variables by passing many %n commands in format string.
3.18 Strings trimming

A very common string processing task is to remove some characters at the start and/or at the end.
In this example, we are going to work with a function which removes all newline characters (CR18 /LF19 )
from the end of the input string:
#include <stdio.h>
#include <string.h>
char* str_trim (char *s)

{
char c;
size_t str_len;
// work as long as \r or \n is at the end of string

// stop if some other character there or its an empty string
// (at start or due to our operation)
for (str_len=strlen(s); str_len>0 && (c=s[str_len-1]); str_len--)
{
if (c=='\r' || c=='\n')
s[str_len-1]=0;
else
break;
};
return s;
};
17 https://github.com/ESWAT/john-carmack-plan-archive/blob/33ae52fdba46aa0d1abfed6fc7598233748541c0/by_day/
johnc_plan_19961217.txt
18 Carriage Return (13 or ’\r’ in C/C++)
19 Line Feed (10 or ’\n’ in C/C++)
524
int main()
{
// test
// strdup() is used to copy text string into data segment,

// because it will crash on Linux otherwise,
// where text strings are allocated in constant data segment,
// and not modifiable.
printf ("[%s]\n", str_trim (strdup("")));

printf ("[%s]\n", str_trim (strdup("\n")));
printf ("[%s]\n", str_trim (strdup("\r")));
printf ("[%s]\n", str_trim (strdup("\n\r")));
printf ("[%s]\n", str_trim (strdup("\r\n")));
printf ("[%s]\n", str_trim (strdup("test1\r\n")));
printf ("[%s]\n", str_trim (strdup("test2\n\r")));
printf ("[%s]\n", str_trim (strdup("test3\n\r\n\r")));
printf ("[%s]\n", str_trim (strdup("test4\n")));
printf ("[%s]\n", str_trim (strdup("test5\r")));
printf ("[%s]\n", str_trim (strdup("test6\r\r\r")));
};
The input argument is always returned on exit, this is convenient when you want to chain string processing
functions, like it has done here in the main() function.
The second part of for() (str_len>0 && (c=s[str_len-1])) is the so called “short-circuit” in C/C++ and
is very convenient [Dennis Yurichev, C/C++ programming language notes1.3.8].
The C/C++ compilers guarantee an evaluation sequence from left to right.
So if the first clause is false after evaluation, the second one is never to be evaluated.
3.18.1 x64: Optimizing MSVC 2013

s$ = 8
str_trim PROC
; RCX is the first function argument and it always holds pointer to the string
mov rdx, rcx
; this is strlen() function inlined right here:
; set RAX to 0xFFFFFFFFFFFFFFFF (-1)
or rax, -1
$LL14@str_trim:
inc rax
cmp BYTE PTR [rcx+rax], 0
jne SHORT $LL14@str_trim
; is the input string length zero? exit then:
test rax, rax
je SHORT $LN15@str_trim
; RAX holds string length
dec rcx
; RCX = s-1
mov r8d, 1
add rcx, rax
; RCX = s-1+strlen(s), i.e., this is the address of the last character in the string
sub r8, rdx
; R8 = 1-s
$LL6@str_trim:
; load the last character of the string:
; jump, if its code is 13 or 10:
movzx eax, BYTE PTR [rcx]
cmp al, 13
je SHORT $LN2@str_trim
cmp al, 10
jne SHORT $LN15@str_trim
$LN2@str_trim:
; the last character has a 13 or 10 code
; write zero at this place:
525
; decrement address of the last character,
; so it will point to the character before the one which has just been erased:
dec rcx
lea rax, QWORD PTR [r8+rcx]
; RAX = 1 - s + address of the current last character
; thus we can determine if we reached the first character and we need to stop, if it is so
test rax, rax
jne SHORT $LL6@str_trim
$LN15@str_trim:
mov rax, rdx
ret 0
str_trim ENDP
First, MSVC inlined the strlen() function code, because it concluded this is to be faster than the usual
strlen() work + the cost of calling it and returning from it. This is called inlining: 3.14 on page 506.
The first instruction of the inlined strlen() is
OR RAX, 0xFFFFFFFFFFFFFFFF.
MSVC often uses OR instead of MOV RAX, 0xFFFFFFFFFFFFFFFF, because resulting opcode is shorter.
And of course, it is equivalent: all bits are set, and a number with all bits set is −1 in two’s complement
arithmetic: 2.2 on page 448.
Why would the −1 number be used in strlen(), one might ask. Due to optimizations, of course. Here is
the code that MSVC generated:
Listing 3.66: Inlined strlen() by MSVC 2013 x64

; RCX = pointer to the input string
; RAX = current string length
or rax, -1
label:
inc rax
cmp BYTE PTR [rcx+rax], 0
jne SHORT label
; RAX = string length
Try to write shorter if you want to initialize the counter at 0! OK, let’ try:
Listing 3.67: Our version of strlen()

; RCX = pointer to the input string
; RAX = current string length
xor rax, rax
label:
cmp byte ptr [rcx+rax], 0
jz exit
inc rax
jmp label
exit:
; RAX = string length
We failed. We have to use additional JMP instruction!

So what the MSVC 2013 compiler did is to move the INC instruction to the place before the actual character
loading.
If the first character is 0, that’s OK, RAX is 0 at this moment, so the resulting string length is 0.
The rest in this function seems easy to understand.
3.18.2 x64: Non-optimizing GCC 4.9.1

str_trim:
push rbp
mov rbp, rsp
sub rsp, 32
mov QWORD PTR [rbp-24], rdi
; for() first part begins here
526
mov rdi, rax
call strlen
mov QWORD PTR [rbp-8], rax ; str_len
; for() first part ends here
jmp .L2
; for() body begins here
.L5:
cmp BYTE PTR [rbp-9], 13 ; c=='\r'?
je .L3
cmp BYTE PTR [rbp-9], 10 ; c=='\n'?
jne .L4
.L3:
mov rax, QWORD PTR [rbp-8] ; str_len
lea rdx, [rax-1] ; EDX=str_len-1
mov rax, QWORD PTR [rbp-24] ; s
add rax, rdx ; RAX=s+str_len-1
mov BYTE PTR [rax], 0 ; s[str_len-1]=0
; for() body ends here
; for() third part begins here
sub QWORD PTR [rbp-8], 1 ; str_len--
; for() third part ends here
.L2:
; for() second part begins here
cmp QWORD PTR [rbp-8], 0 ; str_len==0?
je .L4 ; exit then
; check second clause, and load "c"
mov rax, QWORD PTR [rbp-8] ; RAX=str_len
lea rdx, [rax-1] ; RDX=str_len-1
mov rax, QWORD PTR [rbp-24] ; RAX=s
add rax, rdx ; RAX=s+str_len-1
movzx eax, BYTE PTR [rax] ; AL=s[str_len-1]
mov BYTE PTR [rbp-9], al ; store loaded char into "c"
cmp BYTE PTR [rbp-9], 0 ; is it zero?
jne .L5 ; yes? exit then
; for() second part ends here
.L4:
; return "s"
leave
ret
Comments are added by the author of the book.

After the execution of strlen(), the control is passed to the L2 label, and there two clauses are checked,
one after another.
The second will never be checked, if the first one (str_len==0) is false (this is “short-circuit”).
Now let’s see this function in short form:
• First for() part (call to strlen())
• goto L2
• L5: for() body. goto exit, if needed
• for() third part (decrement of str_len)
• L2: for() second part: check first clause, then second. goto loop body begin or exit.
• L4: // exit
• return s
3.18.3 x64: Optimizing GCC 4.9.1

str_trim:
push rbx
mov rbx, rdi
; RBX will always be "s"
call strlen
; check for str_len==0 and exit if its so
527
test rax, rax
je .L9
lea rdx, [rax-1]
; RDX will always contain str_len-1 value, not str_len
; so RDX is more like buffer index variable
lea rsi, [rbx+rdx] ; RSI=s+str_len-1
movzx ecx, BYTE PTR [rsi] ; load character
test cl, cl
je .L9 ; exit if its zero
cmp cl, 10
je .L4
cmp cl, 13 ; exit if its not '\n' and not '\r'
jne .L9
.L4:
; this is weird instruction. we need RSI=s-1 here.
; its possible to get it by MOV RSI, EBX / DEC RSI
; but this is two instructions instead of one
sub rsi, rax
; RSI = s+str_len-1-str_len = s-1
; main loop begin
.L12:
test rdx, rdx
; store zero at address s-1+str_len-1+1 = s-1+str_len = s+str_len-1
mov BYTE PTR [rsi+1+rdx], 0
; check for str_len-1==0. exit if so.
je .L9
sub rdx, 1 ; equivalent to str_len--
; load next character at address s+str_len-1
movzx ecx, BYTE PTR [rbx+rdx]
test cl, cl ; is it zero? exit then
je .L9
cmp cl, 10 ; is it '\n'?
je .L12
cmp cl, 13 ; is it '\r'?
je .L12
.L9:
; return "s"
mov rax, rbx
pop rbx
ret
Now this is more complex.

The code before the loop’s body start is executed only once, but it has the CR/LF characters check too!
What is this code duplication for?
The common way to implement the main loop is probably this:
• (loop start) check for CR/LF characters, make decisions
• store zero character
But GCC has decided to reverse these two steps.
Of course, store zero character cannot be first step, so another check is needed:
• workout first character. match it to CR/LF, exit if character is not CR/LF
• (loop begin) store zero character
• check for CR/LF characters, make decisions
Now the main loop is very short, which is good for latest CPUs.
The code doesn’t use the str_len variable, but str_len-1. So this is more like an index in a buffer.
Apparently, GCC notices that the str_len-1 statement is used twice.
So it’s better to allocate a variable which always holds a value that’s smaller than the current string length
by one, and decrement it (this is the same effect as decrementing the str_len variable).
528
3.18.4 ARM64: Non-optimizing GCC (Linaro) 4.9
This implementation is straightforward:

str_trim:
stp x29, x30, [sp, -48]!
add x29, sp, 0
str x0, [x29,24] ; copy input argument into local stack
ldr x0, [x29,24] ; s
bl strlen
str x0, [x29,40] ; str_len variable in local stack
b .L2
; main loop begin
.L5:
ldrb w0, [x29,39]
; W0=c
cmp w0, 13 ; is it '\r'?
beq .L3
ldrb w0, [x29,39]
; W0=c
cmp w0, 10 ; is it '\n'?
bne .L4 ; goto exit if it is not
.L3:
ldr x0, [x29,40]
; X0=str_len
sub x0, x0, #1
; X0=str_len-1
ldr x1, [x29,24]
; X1=s
add x0, x1, x0
; X0=s+str_len-1
strb wzr, [x0] ; write byte at s+str_len-1
; decrement str_len:
ldr x0, [x29,40]
; X0=str_len
sub x0, x0, #1
; X0=str_len-1
str x0, [x29,40]
; save X0 (or str_len-1) to local stack
.L2:
ldr x0, [x29,40]
; str_len==0?
cmp x0, xzr
; goto exit then
beq .L4
ldr x0, [x29,40]
; X0=str_len
sub x0, x0, #1
; X0=str_len-1
ldr x1, [x29,24]
; X1=s
add x0, x1, x0
; X0=s+str_len-1
; load byte at address s+str_len-1 to W0
ldrb w0, [x0]
strb w0, [x29,39] ; store loaded byte to "c"
ldrb w0, [x29,39] ; reload it
; is it zero byte?
cmp w0, wzr
; goto exit, if its zero or to L5 if its not
bne .L5
.L4:
; return s
ldr x0, [x29,24]
ldp x29, x30, [sp], 48
ret
529
3.18.5 ARM64: Optimizing GCC (Linaro) 4.9
This is a more advanced optimization.
The first character is loaded at the beginning, and compared against 10 (the LF character).
Characters are also loaded in the main loop, for the characters after first one.
This is somewhat similar to the 3.18.3 on page 527 example.

str_trim:
stp x29, x30, [sp, -32]!
add x29, sp, 0
str x19, [sp,16]
mov x19, x0
; X19 will always hold value of "s"
bl strlen
; X0=str_len
cbz x0, .L9 ; goto L9 (exit) if str_len==0
sub x1, x0, #1
; X1=X0-1=str_len-1
add x3, x19, x1
; X3=X19+X1=s+str_len-1
ldrb w2, [x19,x1] ; load byte at address X19+X1=s+str_len-1
; W2=loaded character
cbz w2, .L9 ; is it zero? jump to exit then
cmp w2, 10 ; is it '\n'?
bne .L15
.L12:
; main loop body. loaded character is always 10 or 13 at this moment!
sub x2, x1, x0
; X2=X1-X0=str_len-1-str_len=-1
add x2, x3, x2
; X2=X3+X2=s+str_len-1+(-1)=s+str_len-2
strb wzr, [x2,1] ; store zero byte at address s+str_len-2+1=s+str_len-1
cbz x1, .L9 ; str_len-1==0? goto exit, if so
sub x1, x1, #1 ; str_len--
ldrb w2, [x19,x1] ; load next character at address X19+X1=s+str_len-1
cmp w2, 10 ; is it '\n'?
cbz w2, .L9 ; jump to exit, if its zero
beq .L12 ; jump to begin loop, if its '\n'
.L15:
cmp w2, 13 ; is it '\r'?
beq .L12 ; yes, jump to the loop body begin
.L9:
; return "s"
mov x0, x19
ldr x19, [sp,16]
ldp x29, x30, [sp], 32
ret
3.18.6 ARM: Optimizing Keil 6/2013 (ARM mode)

And again, the compiler took advantage of ARM mode’s conditional instructions, so the code is much more
compact.

str_trim PROC
PUSH {r4,lr}
; R0=s
MOV r4,r0
; R4=s
BL strlen ; strlen() takes "s" value from R0
; R0=str_len
MOV r3,#0
; R3 will always hold 0
|L0.16|
CMP r0,#0 ; str_len==0?
530
ADDNE r2,r4,r0 ; (if str_len!=0) R2=R4+R0=s+str_len
LDRBNE r1,[r2,#-1] ; (if str_len!=0) R1=load byte at address R2-1=s+str_len-1
CMPNE r1,#0 ; (if str_len!=0) compare loaded byte against 0
BEQ |L0.56| ; jump to exit if str_len==0 or loaded byte is 0
CMP r1,#0xd ; is loaded byte '\r'?
CMPNE r1,#0xa ; (if loaded byte is not '\r') is loaded byte '\r'?
SUBEQ r0,r0,#1 ; (if loaded byte is '\r' or '\n') R0-- or str_len--
STRBEQ r3,[r2,#-1] ; (if loaded byte is '\r' or '\n') store R3 (zero) at address
R2-1=s+str_len-1
BEQ |L0.16| ; jump to loop begin if loaded byte was '\r' or '\n'
|L0.56|
; return "s"
MOV r0,r4
POP {r4,pc}
ENDP
3.18.7 ARM: Optimizing Keil 6/2013 (Thumb mode)

There are less conditional instructions in Thumb mode, so the code is simpler.
But there are is really weird thing with the 0x20 and 0x1F offsets (lines 22 and 23). Why did the Keil
compiler do so? Honestly, it’s hard to say.
It has to be a quirk of Keil’s optimization process. Nevertheless, the code works correctly.

1 str_trim PROC
2 PUSH {r4,lr}
3 MOVS r4,r0
4 ; R4=s
5 BL strlen ; strlen() takes "s" value from R0
6 ; R0=str_len
7 MOVS r3,#0
8 ; R3 will always hold 0
9 B |L0.24|
10 |L0.12|
11 CMP r1,#0xd ; is loaded byte '\r'?
12 BEQ |L0.20|
13 CMP r1,#0xa ; is loaded byte '\n'?
14 BNE |L0.38| ; jump to exit, if no
15 |L0.20|
16 SUBS r0,r0,#1 ; R0-- or str_len--
17 STRB r3,[r2,#0x1f] ; store 0 at address R2+0x1F=s+str_len-0x20+0x1F=s+str_len-1
18 |L0.24|
19 CMP r0,#0 ; str_len==0?
20 BEQ |L0.38| ; yes? jump to exit
21 ADDS r2,r4,r0 ; R2=R4+R0=s+str_len
22 SUBS r2,r2,#0x20 ; R2=R2-0x20=s+str_len-0x20
23 LDRB r1,[r2,#0x1f] ; load byte at address R2+0x1F=s+str_len-0x20+0x1F=s+str_len-1 to
R1
24 CMP r1,#0 ; is loaded byte 0?
25 BNE |L0.12| ; jump to loop begin, if its not 0
26 |L0.38|
27 ; return "s"
28 MOVS r0,r4
29 POP {r4,pc}
30 ENDP
3.18.8 MIPS
str_trim:
; IDA is not aware of local variable names, we gave them manually:
saved_GP = -0x10
saved_S0 = -8
saved_RA = -4
531
addiu $sp, -0x20
sw $ra, 0x20+saved_RA($sp)
sw $s0, 0x20+saved_S0($sp)
sw $gp, 0x20+saved_GP($sp)
; call strlen(). input string address is still in $a0, strlen() will take it from there:
lw $t9, (strlen & 0xFFFF)($gp)
jalr $t9
; input string address is still in $a0, put it to $s0:
move $s0, $a0 ; branch delay slot
; result of strlen() (i.e, length of string) is in $v0 now
; jump to exit if $v0==0 (i.e., if length of string is 0):
beqz $v0, exit
addiu $a1, $v0, -1
; $a1 = $v0-1 = str_len-1
addu $a1, $s0, $a1
; $a1 = input string address + $a1 = s+strlen-1
; load byte at address $a1:
lb $a0, 0($a1)
; loaded byte is zero? jump to exit if its so:
beqz $a0, exit
addiu $v1, $v0, -2
; $v1 = str_len-2
addu $v1, $s0, $v1
; $v1 = $s0+$v1 = s+str_len-2
li $a2, 0xD
; skip loop body:
b loc_6C
li $a3, 0xA ; branch delay slot
loc_5C:
; load next byte from memory to $a0:
lb $a0, 0($v1)
move $a1, $v1
; $a1=s+str_len-2
; jump to exit if loaded byte is zero:
beqz $a0, exit
; decrement str_len:
addiu $v1, -1 ; branch delay slot
loc_6C:
; at this moment, $a0=loaded byte, $a2=0xD (CR symbol) and $a3=0xA (LF symbol)
; loaded byte is CR? jump to loc_7C then:
beq $a0, $a2, loc_7C
addiu $v0, -1 ; branch delay slot
; loaded byte is LF? jump to exit if its not LF:
bne $a0, $a3, exit
loc_7C:
; loaded byte is CR at this moment
; jump to loc_5c (loop body begin) if str_len (in $v0) is not zero:
bnez $v0, loc_5C
; simultaneously, store zero at that place in memory:
sb $zero, 0($a1) ; branch delay slot
; "exit" label was named by me manually:
exit:
lw $ra, 0x20+saved_RA($sp)
move $v0, $s0
lw $s0, 0x20+saved_S0($sp)
jr $ra
addiu $sp, 0x20 ; branch delay slot
Registers prefixed with S- are also called “saved temporaries”, so $S0 value is saved in the local stack
and restored upon finish.
532
3.19 toupper() function
Another very popular function transforms a symbol from lower case to upper case, if needed:
char toupper (char c)
{
if(c>='a' && c<='z')
return c-'a'+'A';
else
return c;
}
The 'a'+'A' expression is left in the source code for better readability, it will be optimized by compiler,
of course 20 .
The ASCII code of “a” is 97 (or 0x61), and 65 (or 0x41) for “A”.
The difference (or distance) between them in the ASCII table is 32 (or 0x20).
For better understanding, the reader may take a look at the 7-bit standard ASCII table:
Figure 3.3: 7-bit ASCII table in Emacs
3.19.1 x64
Two comparison operations
Non-optimizing MSVC is straightforward: the code checks if the input symbol is in [97..122] range (or in
[‘a’..‘z’] range) and subtracts 32 if it’s true.
There are also some minor compiler artifact:
Listing 3.73: Non-optimizing MSVC 2013 (x64)

1 c$ = 8
2 toupper PROC
3 mov BYTE PTR [rsp+8], cl
4 movsx eax, BYTE PTR c$[rsp]
5 cmp eax, 97
6 jl SHORT $LN2@toupper
8 cmp eax, 122
9 jg SHORT $LN2@toupper
11 sub eax, 32
12 jmp SHORT $LN3@toupper
13 jmp SHORT $LN1@toupper ; compiler artefact
14 $LN2@toupper:
15 movzx eax, BYTE PTR c$[rsp] ; unnecessary casting
16 $LN1@toupper:
17 $LN3@toupper: ; compiler artefact
18 ret 0
19 toupper ENDP
It’s important to notice that the input byte is loaded into a 64-bit local stack slot at line 3.
20 However, to be meticulous, there still could be compilers which can’t optimize such expressions and will leave them right in the
code.
533
All the remaining bits ([8..63]) are untouched, i.e., contain some random noise (you’ll see it in debugger).
All instructions operate only on byte-level, so it’s fine.
The last MOVZX instruction at line 15 takes the byte from the local stack slot and zero-extends it to a int
32-bit data type.
Non-optimizing GCC does mostly the same:
Listing 3.74: Non-optimizing GCC 4.9 (x64)

toupper:
push rbp
mov rbp, rsp
mov eax, edi
mov BYTE PTR [rbp-4], al
cmp BYTE PTR [rbp-4], 96
jle .L2
cmp BYTE PTR [rbp-4], 122
jg .L2
movzx eax, BYTE PTR [rbp-4]
sub eax, 32
jmp .L3
.L2:
movzx eax, BYTE PTR [rbp-4]
.L3:
pop rbp
ret
One comparison operation
Optimizing MSVC does a better job, it generates only one comparison operation:
Listing 3.75: Optimizing MSVC 2013 (x64)

toupper PROC
lea eax, DWORD PTR [rcx-97]
cmp al, 25
ja SHORT $LN2@toupper
movsx eax, cl
sub eax, 32
ret 0
$LN2@toupper:
movzx eax, cl
ret 0
toupper ENDP
It was explained earlier how to replace the two comparison operations with a single one: 3.13.2 on
page 505.
We will now rewrite this in C/C++:
int tmp=c-97;
if (tmp>25)
return c;
else
return c-32;
The tmp variable must be signed.

This makes two subtraction operations in case of a transformation plus one comparison.
In contrast the original algorithm uses two comparison operations plus one subtracting.
Optimizing GCC is even better, it gets rid of the jumps (which is good: 2.10.1 on page 462) by using the
CMOVcc instruction:
Listing 3.76: Optimizing GCC 4.9 (x64)

1 toupper:
2 lea edx, [rdi-97] ; 0x61
3 lea eax, [rdi-32] ; 0x20
534
4 cmp dl, 25
5 cmova eax, edi
6 ret
At line 3 the code prepares the subtracted value in advance, as if the conversion will always happen.
At line 5 the subtracted value in EAX is replaced by the untouched input value if a conversion is not needed.
And then this value (of course incorrect) is dropped.
Advance subtracting is a price the compiler pays for the absence of conditional jumps.
3.19.2 ARM
Optimizing Keil for ARM mode also generates only one comparison:

toupper PROC
SUB r1,r0,#0x61
CMP r1,#0x19
SUBLS r0,r0,#0x20
ANDLS r0,r0,#0xff
BX lr
ENDP
The SUBLS and ANDLS instructions are executed only if the value in R1 is less than 0x19 (or equal). They
also do the actual conversion.
Optimizing Keil for Thumb mode generates only one comparison operation as well:

toupper PROC
MOVS r1,r0
SUBS r1,r1,#0x61
CMP r1,#0x19
BHI |L0.14|
SUBS r0,r0,#0x20
LSLS r0,r0,#24
LSRS r0,r0,#24
|L0.14|
BX lr
ENDP
The last two LSLS and LSRS instructions work like AND reg, 0xFF: they are equivalent to the C/C++-
expression (i << 24) >> 24.
Seems like that Keil for Thumb mode deduced that two 2-byte instructions are shorter than the code that
loads the 0xFF constant into a register plus an AND instruction.
GCC for ARM64
Listing 3.79: Non-optimizing GCC 4.9 (ARM64)

toupper:
sub sp, sp, #16
strb w0, [sp,15]
ldrb w0, [sp,15]
cmp w0, 96
bls .L2
ldrb w0, [sp,15]
cmp w0, 122
bhi .L2
ldrb w0, [sp,15]
sub w0, w0, #32
uxtb w0, w0
b .L3
.L2:
ldrb w0, [sp,15]
.L3:
535
add sp, sp, 16
ret
Listing 3.80: Optimizing GCC 4.9 (ARM64)

toupper:
uxtb w0, w0
sub w1, w0, #97
uxtb w1, w1
cmp w1, 25
bhi .L2
sub w0, w0, #32
uxtb w0, w0
.L2:
ret
3.19.3 Using bit operations

Given the fact that 5th bit (counting from 0th) is always present after the check, subtracting is merely
clearing this sole bit, but the very same effect can be achieved with ANDing ( 2.5 on page 454).
Even simpler, with XOR-ing:
char toupper (char c)
{
if(c>='a' && c<='z')
return c^0x20;
else
return c;
}
The code is close to what the optimized GCC has produced for the previous example ( 3.76 on page 534):
Listing 3.81: Optimizing GCC 5.4 (x86)

toupper:
lea ecx, [edx-97]
mov eax, edx
xor eax, 32
cmp cl, 25
cmova eax, edx
ret
…but XOR is used instead of SUB.

Flipping 5th bit is just moving a cursor in ASCII table up and down by two rows.
Some people say that lowercase/uppercase letters has been placed in the ASCII table in such a way
deliberately, because:
Very old keyboards used to do Shift just by toggling the 32 or 16 bit, depending on the
key; this is why the relationship between small and capital letters in ASCII is so regular, and
the relationship between numbers and symbols, and some pairs of symbols, is sort of regular
if you squint at it.
( Eric S. Raymond, http://www.catb.org/esr/faqs/things-every-hacker-once-knew/ )

Therefore, we can write this piece of code, which just flips the case of letters:
#include <stdio.h>
char flip (char c)

{
if((c>='a' && c<='z') || (c>='A' && c<='Z'))
return c^0x20;
else
return c;
}
536
int main()
{
// will produce "hELLO, WORLD!"
for (char *s="Hello, world!"; *s; s++)
printf ("%c", flip(*s));
};
3.19.4 Summary
All these compiler optimizations are very popular nowadays and a practicing reverse engineer usually
sees such code patterns often.
3.20 Obfuscation
The obfuscation is an attempt to hide the code (or its meaning) from reverse engineers.
3.20.1 Text strings

As we know from ( 5.4 on page 697), text strings may be really helpful.
Programmers who are aware of this try to hide them, making it impossible to find the string in IDA or any
hex editor.
Here is the simplest method.
This is how the string can be constructed:
mov byte ptr [ebx], 'h'
mov byte ptr [ebx+1], 'e'
mov byte ptr [ebx+2], 'l'
mov byte ptr [ebx+4], 'o'
mov byte ptr [ebx+5], ' '
mov byte ptr [ebx+6], 'w'
mov byte ptr [ebx+7], 'o'
mov byte ptr [ebx+8], 'r'
mov byte ptr [ebx+10], 'd'
The string can also be compared with another one like this:
mov ebx, offset username
cmp byte ptr [ebx], 'j'
jnz fail
cmp byte ptr [ebx+1], 'o'
jnz fail
cmp byte ptr [ebx+2], 'h'
jnz fail
cmp byte ptr [ebx+3], 'n'
jnz fail
jz it_is_john
In both cases, it is impossible to find these strings straightforwardly in a hex editor.

By the way, this is a way to work with the strings when it is impossible to allocate space for them in the
data segment, for example in a PIC21 or in shellcode.
Another method is to use sprintf() for the construction:
sprintf(buf, "%s%c%s%c%s", "hel",'l',"o w",'o',"rld");
The code looks weird, but as a simple anti-reversing measure, it may be helpful.
Text strings may also be present in encrypted form, then every string usage is to be preceded by a string
decrypting routine. For example: 8.8.2 on page 826.
21 Position Independent Code
537
3.20.2 Executable code
Inserting garbage
Executable code obfuscation implies inserting random garbage code between real one, which executes
but does nothing useful.
A simple example:
Listing 3.82: original code

add eax, ebx
mul ecx
Listing 3.83: obfuscated code

xor esi, 011223344h ; garbage
add esi, eax ; garbage
add eax, ebx
mov edx, eax ; garbage
shl edx, 4 ; garbage
mul ecx
xor esi, ecx ; garbage
Here the garbage code uses registers which are not used in the real code (ESI and EDX). However, the
intermediate results produced by the real code may be used by the garbage instructions for some extra
mess—why not?
Replacing instructions with bloated equivalents
• MOV op1, op2 can be replaced by the PUSH op2 / POP op1 pair.
• JMP label can be replaced by the PUSH label / RET pair. IDA will not show the references to the
label.
• CALL label can be replaced by the following instructions triplet:
PUSH label_after_CALL_instruction / PUSH label / RET.
• PUSH op can also be replaced with the following instructions pair:
SUB ESP, 4 (or 8) / MOV [ESP], op.
Always executed/never executed code
If the developer is sure that ESI at always 0 at that point:

mov esi, 1
... ; some code not touching ESI
dec esi
... ; some code not touching ESI
cmp esi, 0
jz real_code
; fake luggage
real_code:
The reverse engineer needs some time to get into it.

This is also called an opaque predicate.
Another example (and again, the developer is sure that ESI is always zero):
; ESI=0
add eax, ebx ; real code
mul ecx ; real code
add eax, esi ; opaque predicate. XOR, AND or SHL, etc, can be here instead of ADD.
Making a lot of mess
instruction 1
instruction 2
instruction 3
538
Can be replaced with:
begin: jmp ins1_label
ins2_label: instruction 2
jmp ins3_label
jmp exit:
jmp ins2_label
exit:
Using indirect pointers

dummy_data1 db 100h dup (0)
message1 db 'hello world',0
dummy_data2 db 200h dup (0)

message2 db 'another message',0
func proc
...
mov eax, offset dummy_data1 ; PE or ELF reloc here
add eax, 100h
push eax
call dump_string
...
mov eax, offset dummy_data2 ; PE or ELF reloc here
add eax, 200h
push eax
call dump_string
...
func endp
IDA will show references only to dummy_data1 and dummy_data2, but not to the text strings.
Global variables and even functions may be accessed like that.
Now something slightly more advanced.
Honestly, I don’t know its exact name, but I would call it shifted pointer. This technique is quite common,
at least in copy protection schemes.
In short: while writing a value into global memory you use an address, but by reading you use a sum of
(other) addresses, or maybe a difference. The goal is to hide a real address from a reverse engineer who
debugs the code or exploring it in IDA (or another disassembler).
This can be a nuisance.
#include <stdio.h>
// 64KiB, but it's OK

unsigned char secret_array[0x10000];
void check_lic_key()
{
// pretend licence check has been failed
secret_array[0x6123]=1; // 1 mean failed
printf ("check failed\n"); // exit(0); // / a cracker may patch here
// or put there another value if check is succeeded

secret_array[0x6123]=0;
};
unsigned char get_byte_at_0x6000(unsigned char *a)

{
return *(a+0x6000);
};
539
void check_again()
{
if (get_byte_at_0x6000(secret_array+0x123)==1)
{
// do something mean (add watermark maybe) or report error:
printf ("check failed\n");
}
else
{
// proceed further
};
};
int main()
{
// at start:
check_lic_key();
// do something
// ... and while in some very critical part:

check_again();
};
If compiled by non-optimizing MSVC 2015:

_check_lic_key proc near
push ebp
mov ebp, esp
mov eax, 1
imul ecx, eax, 6123h
mov _secret_array[ecx], 1
pop ebp
retn
_check_lic_key endp
_get_byte_at_0x6000 proc near
a = dword ptr 8
push ebp
mov ebp, esp
mov eax, [ebp+a]
mov al, [eax+6000h]
pop ebp
retn
_get_byte_at_0x6000 endp
_check_again proc near

push ebp
mov ebp, esp
push offset point_passed_to_get_byte_at_0x6000
call j__get_byte_at_0x6000
add esp, 4
movzx eax, al
cmp eax, 1
push offset _Format ; "check failed\n"
call j__printf
add esp, 4
loc_406735:
pop ebp
retn
_check_again endp
.data:0045F5C0 ; char secret_array[65536]

.data:0045F5C0 _secret_array db 123h dup(?)
540
.data:0045F6E3 ; char point_passed_to_get_byte_at_0x6000[65245]
.data:0045F6E3 point_passed_to_get_byte_at_0x6000 db 0FEDDh dup(?)
You see, IDA can only get two addresses: secret_array[] (start of the array) and
point_passed_to_get_byte_at_0x6000.
How to deal with it: you can use hardware breakpoints on memory access operations tracer has BPMx
options) or symbolic execution engine or maybe you can write a plugin for IDA...
Surely, one array can be used for many values, not limited to boolean ones...
N.B.: Optimizing MSVC 2015 is smart enough to optimize the get_byte_at_0x6000() function out.
3.20.3 Virtual machine / pseudo-code

A programmer can construct his/her own PL or ISA and interpreter for it.
(Like the pre-5.0 Visual Basic, .NET or Java machines). The reverse engineer will have to spend some time
to understand the meaning and details of all of the ISA’s instructions.
He/she will also have to write a disassembler/decompiler of some sort.
3.20.4 Other things to mention

My own (yet weak) attempt to patch the Tiny C compiler to produce obfuscated code: http://go.yurichev.
com/17220.
Using the MOV instruction for really complicated things: [Stephen Dolan, mov is Turing-complete, (2013)]
22
.
3.20.5 Exercise
3.21 C++
3.21.1 Classes
A simple example
Internally, the representation of C++ classes is almost the same as the structures.
Let’s try an example with two variables, two constructors and one method:
#include <stdio.h>
class c
{
private:
int v1;
int v2;
public:
c() // default ctor
{
v1=667;
v2=999;
};
c(int a, int b) // ctor

{
v1=a;
v2=b;
};
void dump()
{
printf ("%d; %d\n", v1, v2);
22 Also available as http://www.cl.cam.ac.uk/~sd601/papers/mov.pdf
541
};
};
int main()
{
class c c1;
class c c2(5,6);
c1.dump();
c2.dump();
return 0;
};
MSVC: x86
Here is how the main() function looks like, translated into assembly language:
Listing 3.84: MSVC

_c2$ = -16 ; size = 8
_c1$ = -8 ; size = 8
_main PROC
push ebp
mov ebp, esp
sub esp, 16
lea ecx, DWORD PTR _c1$[ebp]
call ??0c@@QAE@XZ ; c::c
push 6
push 5
call ??0c@@QAE@HH@Z ; c::c
call ?dump@c@@QAEXXZ ; c::dump
call ?dump@c@@QAEXXZ ; c::dump
xor eax, eax
mov esp, ebp
pop ebp
ret 0
_main ENDP
Here’s what’s going on. For each object (instance of class c) 8 bytes are allocated, exactly the size needed
to store the 2 variables.
For c1 a default argumentless constructor ??0c@@QAE@XZ is called. For c2 another constructor ??0c@@QAE@HH@Z
is called and two numbers are passed as arguments.
A pointer to the object (this in C++ terminology) is passed in the ECX register. This is called thiscall ( 3.21.1)—
the method for passing a pointer to the object.
MSVC does it using the ECX register. Needless to say, it is not a standardized method, other compilers can
do it differently, e.g., via the first function argument (like GCC).
Why do these functions have such odd names? That’s name mangling.
A C++ class may contain several methods sharing the same name but having different arguments—that
is polymorphism. And of course, different classes may have their own methods with the same name.
Name mangling enable us to encode the class name + method name + all method argument types in one
ASCII string, which is then used as an internal function name. That’s all because neither the linker, nor
the DLL OS loader (mangled names may be among the DLL exports as well) knows anything about C++
or OOP23 .
The dump() function is called two times.
Now let’s see the constructors’ code:
23 Object-Oriented Programming
542
Listing 3.85: MSVC
_this$ = -4 ; size = 4
??0c@@QAE@XZ PROC ; c::c, COMDAT
; _this$ = ecx
push ebp
mov ebp, esp
push ecx
mov DWORD PTR _this$[ebp], ecx
mov eax, DWORD PTR _this$[ebp]
mov DWORD PTR [eax], 667
mov ecx, DWORD PTR _this$[ebp]
mov DWORD PTR [ecx+4], 999
mov esp, ebp
pop ebp
ret 0
??0c@@QAE@XZ ENDP ; c::c
_this$ = -4 ; size = 4
_a$ = 8 ; size = 4
_b$ = 12 ; size = 4
??0c@@QAE@HH@Z PROC ; c::c, COMDAT
; _this$ = ecx
push ebp
mov ebp, esp
push ecx
mov DWORD PTR [eax], ecx
mov edx, DWORD PTR _this$[ebp]
mov DWORD PTR [edx+4], eax
mov esp, ebp
pop ebp
ret 8
??0c@@QAE@HH@Z ENDP ; c::c
The constructors are just functions, they use a pointer to the structure in ECX, copying the pointer into
their own local variable, however, it is not necessary.
From the C++ standard (C++11 12.1) we know that constructors are not required to return any values.
In fact, internally, the constructors return a pointer to the newly created object, i.e., this.
Now the dump() method:
Listing 3.86: MSVC

_this$ = -4 ; size = 4
?dump@c@@QAEXXZ PROC ; c::dump, COMDAT
; _this$ = ecx
push ebp
mov ebp, esp
push ecx
mov ecx, DWORD PTR [eax+4]
push ecx
mov edx, DWORD PTR _this$[ebp]
mov eax, DWORD PTR [edx]
push eax
push OFFSET ??_C@_07NJBDCIEC@?$CFd?$DL?5?$CFd?6?$AA@
call _printf
add esp, 12
mov esp, ebp
pop ebp
ret 0
?dump@c@@QAEXXZ ENDP ; c::dump
543
Simple enough: dump() takes a pointer to the structure that contains the two int’s from ECX, takes both
values from it and passes them to printf().
The code is much shorter if compiled with optimizations (/Ox):
Listing 3.87: MSVC

??0c@@QAE@XZ PROC ; c::c, COMDAT
; _this$ = ecx
mov eax, ecx
mov DWORD PTR [eax+4], 999
ret 0
??0c@@QAE@XZ ENDP ; c::c
_a$ = 8 ; size = 4
_b$ = 12 ; size = 4
??0c@@QAE@HH@Z PROC ; c::c, COMDAT
; _this$ = ecx
mov edx, DWORD PTR _b$[esp-4]
mov eax, ecx
ret 8
??0c@@QAE@HH@Z ENDP ; c::c
?dump@c@@QAEXXZ PROC ; c::dump, COMDAT

; _this$ = ecx
mov eax, DWORD PTR [ecx+4]
push eax
push ecx
push OFFSET ??_C@_07NJBDCIEC@?$CFd?$DL?5?$CFd?6?$AA@
call _printf
add esp, 12
ret 0
?dump@c@@QAEXXZ ENDP ; c::dump
That’s all. The other thing we must note is that the stack pointer hasn’t been corrected with add esp, X
after the constructor has been called. At the same time, the constructor has ret 8 instead of RET at the
end.
This is all because the thiscall ( 3.21.1 on page 542) calling convention is used here, which together with
the stdcall ( 6.1.2 on page 725) method offers the callee to correct the stack instead of the caller. The
ret x instruction adds X to the value in ESP, then passes the control to the caller function.
See also the section about calling conventions ( 6.1 on page 725).
It also has to be noted that the compiler decides when to call the constructor and destructor—but we
already know that from the C++ language basics.
MSVC: x86-64
As we already know, the first 4 function arguments in x86-64 are passed in RCX, RDX, R8 and R9 registers,
all the rest—via the stack.
Nevertheless, the this pointer to the object is passed in RCX, the first argument of the method in RDX, etc.
We can see this in the c(int a, int b) method internals:

; void dump()
?dump@c@@QEAAXXZ PROC ; c::dump

mov r8d, DWORD PTR [rcx+4]
mov edx, DWORD PTR [rcx]
lea rcx, OFFSET FLAT:??_C@_07NJBDCIEC@?$CFd?$DL?5?$CFd?6?$AA@ ; '%d; %d'
jmp printf
?dump@c@@QEAAXXZ ENDP ; c::dump
544
; c(int a, int b)
??0c@@QEAA@HH@Z PROC ; c::c

mov DWORD PTR [rcx], edx ; 1st argument: a
mov DWORD PTR [rcx+4], r8d ; 2nd argument: b
mov rax, rcx
ret 0
??0c@@QEAA@HH@Z ENDP ; c::c
; default ctor
??0c@@QEAA@XZ PROC ; c::c

mov DWORD PTR [rcx], 667
mov DWORD PTR [rcx+4], 999
mov rax, rcx
ret 0
??0c@@QEAA@XZ ENDP ; c::c
24
The int data type is still 32-bit in x64 , so that is why 32-bit register parts are used here.
We also see JMP printf instead of RET in the dump() method, that hack we already saw earlier: 1.21.1
on page 154.
GCC: x86
It is almost the same story in GCC 4.4.1, with a few exceptions.

public main
main proc near

push ebp
mov ebp, esp
and esp, 0FFFFFFF0h
sub esp, 20h
call _ZN1cC1Ev
mov [esp+20h+var_1C], 5
call _ZN1cC1Eii
call _ZN1c4dumpEv
call _ZN1c4dumpEv
mov eax, 0
leave
retn
main endp
Here we see another name mangling style, specific to GNU 25 . It can also be noted that the pointer to the
object is passed as the first function argument—invisible to programmer, of course.
First constructor:
24 Apparently, for easier porting of 32-bit C/C++ code to x64.
25 Thereis a good document about the various name mangling conventions in different compilers:
[Agner Fog, Calling conventions (2015)].
545
public _ZN1cC1Ev ; weak
_ZN1cC1Ev proc near ; CODE XREF: main+10
arg_0 = dword ptr 8
push ebp
mov ebp, esp
mov dword ptr [eax], 667
mov dword ptr [eax+4], 999
pop ebp
retn
_ZN1cC1Ev endp
It just writes two numbers using the pointer passed in the first (and only) argument.
Second constructor:
public _ZN1cC1Eii
_ZN1cC1Eii proc near
arg_0 = dword ptr 8

push ebp
mov ebp, esp
mov edx, [ebp+arg_4]
mov [eax], edx
mov [eax+4], edx
pop ebp
retn
_ZN1cC1Eii endp
This is a function, the analog of which can look like this:

void ZN1cC1Eii (int *obj, int a, int b)
{
*obj=a;
*(obj+1)=b;
};
…and that is completely predictable.

Now the dump() function:
public _ZN1c4dumpEv
_ZN1c4dumpEv proc near

arg_0 = dword ptr 8
push ebp
mov ebp, esp
sub esp, 18h
mov edx, [eax+4]
mov eax, [eax]
mov [esp+18h+var_18], offset aDD ; "%d; %d\n"
call _printf
leave
546
retn
_ZN1c4dumpEv endp
This function in its internal representation has only one argument, used as pointer to the object (this).
This function could be rewritten in C like this:
void ZN1c4dumpEv (int *obj)
{
printf ("%d; %d\n", *obj, *(obj+1));
};
Thus, if we base our judgment on these simple examples, the difference between MSVC and GCC is the
style of the encoding of function names (name mangling) and the method for passing a pointer to the
object (via the ECX register or via the first argument).
GCC: x86-64
The first 6 arguments, as we already know, are passed in the RDI, RSI, RDX, RCX, R8 and R9 ([Michael Matz,
Jan Hubicka, Andreas Jaeger, Mark Mitchell, System V Application Binary Interface. AMD64 Architecture
Processor Supplement, (2013)] 26 ) registers, and the pointer to this via the first one (RDI) and that is what
we see here. The int data type is also 32-bit here.
The JMP instead of RET hack is also used here.

; default ctor
_ZN1cC2Ev:
mov DWORD PTR [rdi], 667
mov DWORD PTR [rdi+4], 999
ret
; c(int a, int b)
_ZN1cC2Eii:
mov DWORD PTR [rdi], esi
mov DWORD PTR [rdi+4], edx
ret
; dump()
_ZN1c4dumpEv:
mov edx, DWORD PTR [rdi+4]
mov esi, DWORD PTR [rdi]
xor eax, eax
mov edi, OFFSET FLAT:.LC0 ; "%d; %d\n"
jmp printf
Class inheritance
Inherited classes are similar to the simple structures we already discussed, but extended in inheritable
classes.
Let’s take this simple example:
#include <stdio.h>
class object
{
public:
int color;
object() { };
object (int color) { this->color=color; };
void print_color() { printf ("color=%d\n", color); };
};
547
class box : public object
{
private:
int width, height, depth;
public:
box(int color, int width, int height, int depth)
{
this->color=color;
this->width=width;
this->height=height;
this->depth=depth;
};
void dump()
{
printf ("this is a box. color=%d, width=%d, height=%d, depth=%d\n", color, width, ⤦
Ç height, depth);
};
};
class sphere : public object

{
private:
int radius;
public:
sphere(int color, int radius)
{
this->color=color;
this->radius=radius;
};
void dump()
{
printf ("this is sphere. color=%d, radius=%d\n", color, radius);
};
};
int main()
{
box b(1, 10, 20, 30);
sphere s(2, 40);
b.print_color();
s.print_color();
b.dump();
s.dump();
return 0;
};
Let’s investigate the generated code of the dump() functions/methods and also object::print_color(),
and see the memory layout for the structures-objects (for 32-bit code).
So, here are the dump() methods for several classes, generated by MSVC 2008 with /Ox and /Ob0 options
27
.

??_C@_09GCEDOLPA@color?$DN?$CFd?6?$AA@ DB 'color=%d', 0aH, 00H ; `string'
?print_color@object@@QAEXXZ PROC ; object::print_color, COMDAT
; _this$ = ecx
push eax
; 'color=%d', 0aH, 00H

push OFFSET ??_C@_09GCEDOLPA@color?$DN?$CFd?6?$AA@
call _printf
add esp, 8
ret 0
27 The /Ob0 option stands for disabling inline expansion since function inlining can make our experiment harder.
548
?print_color@object@@QAEXXZ ENDP ; object::print_color

?dump@box@@QAEXXZ PROC ; box::dump, COMDAT
; _this$ = ecx
mov edx, DWORD PTR [ecx+8]
push eax
push edx
push eax
push ecx
; 'this is a box. color=%d, width=%d, height=%d, depth=%d', 0aH, 00H ; `string'

push OFFSET ??_C@_0DG@NCNGAADL@this?5is?5box?4?5color?$DN?$CFd?0?5width?$DN?$CFd?0@
call _printf
add esp, 20
ret 0
?dump@box@@QAEXXZ ENDP ; box::dump

?dump@sphere@@QAEXXZ PROC ; sphere::dump, COMDAT
; _this$ = ecx
push eax
push ecx
; 'this is sphere. color=%d, radius=%d', 0aH, 00H

push OFFSET ??_C@_0CF@EFEDJLDC@this?5is?5sphere?4?5color?$DN?$CFd?0?5radius@
call _printf
add esp, 12
ret 0
?dump@sphere@@QAEXXZ ENDP ; sphere::dump
So, here is the memory layout:

(base class object)
offset description
+0x0 int color
(inherited classes)
box:
offset description
+0x0 int color
+0x4 int width
+0x8 int height
+0xC int depth
sphere:
offset description
+0x0 int color
+0x4 int radius
Let’s see main() function body:

PUBLIC _main
_TEXT SEGMENT
_s$ = -24 ; size = 8
_b$ = -16 ; size = 16
_main PROC
sub esp, 24
push 30
549
push 20
push 10
push 1
lea ecx, DWORD PTR _b$[esp+40]
call ??0box@@QAE@HHHH@Z ; box::box
push 40
push 2
lea ecx, DWORD PTR _s$[esp+32]
call ??0sphere@@QAE@HH@Z ; sphere::sphere
call ?print_color@object@@QAEXXZ ; object::print_color
call ?print_color@object@@QAEXXZ ; object::print_color
call ?dump@box@@QAEXXZ ; box::dump
call ?dump@sphere@@QAEXXZ ; sphere::dump
xor eax, eax
add esp, 24
ret 0
_main ENDP
The inherited classes must always add their fields after the base classes’ fields, to make it possible for
the base class methods to work with their own fields.
When the object::print_color() method is called, a pointers to both the box and sphere objects are
passed as this, and it can work with these objects easily since the color field in these objects is always
at the pinned address (at offset +0x0).
It can be said that the object::print_color() method is agnostic in relation to the input object type as
long as the fields are pinned at the same addresses, and this condition is always true.
And if you create inherited class of the box class, the compiler will add the new fields after the depth field,
leaving the box class fields at the pinned addresses.
Thus, the box::dump() method will work fine for accessing the color, width, height and depths fields,
which are always pinned at known addresses.
The code generated by GCC is almost the same, with the sole exception of passing the this pointer (as it
has been explained above, it is passed as the first argument instead of using the ECX register).
Encapsulation
Encapsulation is hiding the data in the private sections of the class, e.g. to allow access to them only from
this class methods.
However, are there any marks in code the about the fact that some field is private and some other—not?
No, there are no such marks.
Let’s try this simple example:
#include <stdio.h>
class box
{
private:
int color, width, height, depth;
public:
{
this->color=color;
this->width=width;
this->depth=depth;
};
void dump()
{
Ç height, depth);
};
};
550
Let’s compile it again in MSVC 2008 with /Ox and /Ob0 options and see the box::dump() method code:
; _this$ = ecx
push eax
push edx
push eax
push ecx
; 'this is a box. color=%d, width=%d, height=%d, depth=%d', 0aH, 00H
push OFFSET ??_C@_0DG@NCNGAADL@this?5is?5box?4?5color?$DN?$CFd?0?5width?$DN?$CFd?0@
call _printf
add esp, 20
ret 0
Here is a memory layout of the class:

offset description
+0x0 int color
+0x4 int width
+0x8 int height
+0xC int depth
All fields are private and not allowed to be accessed from any other function, but knowing this layout, can
we create code that modifies these fields?
To do this we’ll add the hack_oop_encapsulation() function, which is not going to compile if it looked
like this:
void hack_oop_encapsulation(class box * o)
{
o->width=1; // that code can't be compiled:
// "error C2248: 'box::width' : cannot access private member declared in class
'box'"
};
Nevertheless, if we cast the box type to a pointer to an int array, and we modify the array of int-s that we
have, we can succeed.
void hack_oop_encapsulation(class box * o)
{
unsigned int *ptr_to_object=reinterpret_cast<unsigned int*>(o);
ptr_to_object[1]=123;
};
This function’s code is very simple—it can be said that the function takes a pointer to an array of int-s for
input and writes 123 to the second int:
?hack_oop_encapsulation@@YAXPAVbox@@@Z PROC ; hack_oop_encapsulation
mov eax, DWORD PTR _o$[esp-4]
ret 0
?hack_oop_encapsulation@@YAXPAVbox@@@Z ENDP ; hack_oop_encapsulation
Let’s check how it works:

int main()
{
box b(1, 10, 20, 30);
b.dump();
hack_oop_encapsulation(&b);
551
b.dump();
return 0;
};
Let’s run:
this is a box. color=1, width=10, height=20, depth=30
this is a box. color=1, width=123, height=20, depth=30
We see that the encapsulation is just protection of class fields only in the compilation stage.
The C++ compiler is not allowing the generation of code that modifies protected fields straightforwardly,
nevertheless, it is possible with the help of dirty hacks.
Multiple inheritance
Multiple inheritance is creating a class which inherits fields and methods from two or more classes.
Let’s write a simple example again:
#include <stdio.h>
class box
{
public:
box() { };
box(int width, int height, int depth)
{
this->width=width;
this->depth=depth;
};
void dump()
{
printf ("this is a box. width=%d, height=%d, depth=%d\n", width, height, depth);
};
int get_volume()
{
return width * height * depth;
};
};
class solid_object
{
public:
int density;
solid_object() { };
solid_object(int density)
{
this->density=density;
};
int get_density()
{
return density;
};
void dump()
{
printf ("this is a solid_object. density=%d\n", density);
};
};
class solid_box: box, solid_object

{
public:
solid_box (int width, int height, int depth, int density)
{
this->width=width;
552
this->depth=depth;
this->density=density;
};
void dump()
{
printf ("this is a solid_box. width=%d, height=%d, depth=%d, density=%d\n", width, ⤦
Ç height, depth, density);
};
int get_weight() { return get_volume() * get_density(); };
};
int main()
{
box b(10, 20, 30);
solid_object so(100);
solid_box sb(10, 20, 30, 3);
b.dump();
so.dump();
sb.dump();
printf ("%d\n", sb.get_weight());
return 0;
};
Let’s compile it in MSVC 2008 with the /Ox and /Ob0 options and see the code of box::dump(),
solid_object::dump() and solid_box::dump():

; _this$ = ecx
push eax
push edx
push eax
; 'this is a box. width=%d, height=%d, depth=%d', 0aH, 00H
push OFFSET ??_C@_0CM@DIKPHDFI@this?5is?5box?4?5width?$DN?$CFd?0?5height?$DN?$CFd@
call _printf
add esp, 16
ret 0

?dump@solid_object@@QAEXXZ PROC ; solid_object::dump, COMDAT
; _this$ = ecx
push eax
; 'this is a solid_object. density=%d', 0aH
push OFFSET ??_C@_0CC@KICFJINL@this?5is?5solid_object?4?5density?$DN?$CFd@
call _printf
add esp, 8
ret 0
?dump@solid_object@@QAEXXZ ENDP ; solid_object::dump

?dump@solid_box@@QAEXXZ PROC ; solid_box::dump, COMDAT
; _this$ = ecx
push eax
push edx
push eax
push ecx
553
; 'this is a solid_box. width=%d, height=%d, depth=%d, density=%d', 0aH
push OFFSET ??_C@_0DO@HNCNIHNN@this?5is?5solid_box?4?5width?$DN?$CFd?0?5hei@
call _printf
add esp, 20
ret 0
?dump@solid_box@@QAEXXZ ENDP ; solid_box::dump
So, the memory layout for all three classes is:

box class:
offset description
+0x0 width
+0x4 height
+0x8 depth
solid_object class:
offset description
+0x0 density
It can be said that the solid_box class memory layout is united:
solid_box class:
offset description
+0x0 width
+0x4 height
+0x8 depth
+0xC density
The code of the box::get_volume() and solid_object::get_density() methods is trivial:

?get_volume@box@@QAEHXZ PROC ; box::get_volume, COMDAT
; _this$ = ecx
imul eax, DWORD PTR [ecx+4]
imul eax, DWORD PTR [ecx]
ret 0
?get_volume@box@@QAEHXZ ENDP ; box::get_volume

?get_density@solid_object@@QAEHXZ PROC ; solid_object::get_density, COMDAT
; _this$ = ecx
ret 0
?get_density@solid_object@@QAEHXZ ENDP ; solid_object::get_density
But the code of the solid_box::get_weight() method is much more interesting:

?get_weight@solid_box@@QAEHXZ PROC ; solid_box::get_weight, COMDAT
; _this$ = ecx
push esi
mov esi, ecx
push edi
lea ecx, DWORD PTR [esi+12]
call ?get_density@solid_object@@QAEHXZ ; solid_object::get_density
mov ecx, esi
mov edi, eax
call ?get_volume@box@@QAEHXZ ; box::get_volume
imul eax, edi
pop edi
pop esi
ret 0
?get_weight@solid_box@@QAEHXZ ENDP ; solid_box::get_weight
554
get_weight() just calls two methods, but for get_volume() it just passes pointer to this, and for get_density()
it passes a pointer to this incremented by 12 (or 0xC) bytes, and there, in the solid_box class memory
layout, the fields of the solid_object class start.
Thus, the solid_object::get_density() method will believe like it is dealing with the usual solid_object
class, and the box::get_volume() method will work with its three fields, believing this is just the usual
object of class box.
Thus, we can say, an object of a class, that inherits from several other classes, is representing in memory
as a united class, that contains all inherited fields. And each inherited method is called with a pointer to
the corresponding structure’s part.
Virtual methods
Yet another simple example:

#include <stdio.h>
class object
{
public:
int color;
object() { };
object (int color) { this->color=color; };
virtual void dump()
{
printf ("color=%d\n", color);
};
};
class box : public object

{
private:
public:
{
this->color=color;
this->width=width;
this->depth=depth;
};
void dump()
{
Ç height, depth);
};
};
class sphere : public object

{
private:
int radius;
public:
sphere(int color, int radius)
{
this->color=color;
this->radius=radius;
};
void dump()
{
printf ("this is sphere. color=%d, radius=%d\n", color, radius);
};
};
int main()
{
box b(1, 10, 20, 30);
sphere s(2, 40);
555
object *o1=&b;
object *o2=&s;
o1->dump();
o2->dump();
return 0;
};
Class object has a virtual method dump() that is being replaced in the inheriting box and sphere classes.
If we are in an environment where it is not known the type of an object, as in the main() function in exam-
ple, where the virtual method dump() is called, the information about its type must be stored somewhere,
to be able to call the relevant virtual method.
Let’s compile it in MSVC 2008 with the /Ox and /Ob0 options and see the code of main():
_s$ = -32 ; size = 12
_b$ = -20 ; size = 20
_main PROC
sub esp, 32
push 30
push 20
push 10
push 1
call ??0box@@QAE@HHHH@Z ; box::box
push 40
push 2
call ??0sphere@@QAE@HH@Z ; sphere::sphere
mov eax, DWORD PTR _b$[esp+32]
mov edx, DWORD PTR [eax]
call edx
call edx
xor eax, eax
add esp, 32
ret 0
_main ENDP
A pointer to the dump() function is taken somewhere from the object. Where could we store the address
of the new method? Only somewhere in the constructors: there is no other place since nothing else is
called in the main() function. 28
Let’s see the code of the constructor of the box class:
??_R0?AVbox@@@8 DD FLAT:??_7type_info@@6B@ ; box `RTTI Type Descriptor'
DD 00H
DB '.?AVbox@@', 00H
??_R1A@?0A@EA@box@@8 DD FLAT:??_R0?AVbox@@@8 ; box::`RTTI Base Class Descriptor at (0,-1,0,64)'

DD 01H
DD 00H
DD 0ffffffffH
DD 00H
DD 040H
DD FLAT:??_R3box@@8
??_R2box@@8 DD FLAT:??_R1A@?0A@EA@box@@8 ; box::`RTTI Base Class Array'

DD FLAT:??_R1A@?0A@EA@object@@8
??_R3box@@8 DD 00H ; box::`RTTI Class Hierarchy Descriptor'

DD 00H
DD 02H
DD FLAT:??_R2box@@8
28 You can read more about pointers to functions in the relevant section:( 1.33 on page 381)
556
??_R4box@@6B@ DD 00H ; box::`RTTI Complete Object Locator'
DD 00H
DD 00H
DD FLAT:??_R0?AVbox@@@8
DD FLAT:??_R3box@@8
??_7box@@6B@ DD FLAT:??_R4box@@6B@ ; box::`vftable'

DD FLAT:?dump@box@@UAEXXZ
_color$ = 8 ; size = 4
_width$ = 12 ; size = 4
_height$ = 16 ; size = 4
_depth$ = 20 ; size = 4
??0box@@QAE@HHHH@Z PROC ; box::box, COMDAT
; _this$ = ecx
push esi
mov esi, ecx
call ??0object@@QAE@XZ ; object::object
mov eax, DWORD PTR _color$[esp]
mov ecx, DWORD PTR _width$[esp]
mov edx, DWORD PTR _height$[esp]
mov DWORD PTR [esi+4], eax
mov eax, DWORD PTR _depth$[esp]
mov DWORD PTR [esi], OFFSET ??_7box@@6B@
mov DWORD PTR [esi+8], ecx
mov DWORD PTR [esi+12], edx
mov eax, esi
pop esi
ret 16
??0box@@QAE@HHHH@Z ENDP ; box::box
Here we see a slightly different memory layout: the first field is a pointer to some table box::`vftable'
(the name has been set by the MSVC compiler).
In this table we see a link to a table named
box::`RTTI Complete Object Locator' and also a link
to the box::dump() method.
These are called virtual methods table and RTTI29 . The table of virtual methods has the addresses of
methods and the RTTI table contains information about types.
By the way, the RTTI tables are used while calling dynamic_cast and typeid in C++. You can also see here
the class name as a plain text string.
Thus, a method of the base object class may call the virtual method object::dump(), which in turn will call
a method of an inherited class, since that information is present right in the object’s structure.
Some additional CPU time is needed for doing look-ups in these tables and finding the right virtual method
address, thus virtual methods are widely considered as slightly slower than common methods.
In GCC-generated code the RTTI tables are constructed slightly differently.
3.21.2 ostream
Let’s start again with a “hello world” example, but now we are going to use ostream:
#include <iostream>
int main()
{
std::cout << "Hello, world!\n";
}
Almost any C++ textbook tells us that the << operation can be defined (overloaded) for other types. That
is what is done in ostream. We see that operator<< is called for ostream:
Listing 3.101: MSVC 2012 (reduced listing)
29 Run-Time Type Information
557
$SG37112 DB 'Hello, world!', 0aH, 00H
_main PROC
push OFFSET ?cout@std@@3V?$basic_ostream@DU?$char_traits@D@std@@@1@A ; std::cout
call ??$?6U?$char_traits@D@std@@@std@@YAAAV?$basic_ostream@DU?⤦
Ç $char_traits@D@std@@@0@AAV10@PBD@Z ; std::operator<<<std::char_traits<char>
>
add esp, 8
xor eax, eax
ret 0
_main ENDP
Let’s modify the example:

#include <iostream>
int main()
{
std::cout << "Hello, " << "world!\n";
}
And again, from many C++ textbooks we know that the result of each operator<< in ostream is forwarded
to the next one. Indeed:

$SG37112 DB 'world!', 0aH, 00H
$SG37113 DB 'Hello, ', 00H
_main PROC
push OFFSET $SG37113 ; 'Hello, '
push OFFSET ?cout@std@@3V?$basic_ostream@DU?$char_traits@D@std@@@1@A ; std::cout
>
add esp, 8
push OFFSET $SG37112 ; 'world!'

push eax ; result of previous function execution
>
add esp, 8
xor eax, eax

ret 0
_main ENDP
If we would rename operator<< method name to f(), that code will looks like:
f(f(std::cout, "Hello, "), "world!");
GCC generates almost the same code as MSVC.
3.21.3 References
In C++, references are pointers ( 3.23 on page 596) as well, but they are called safe, because it is harder
to make a mistake while dealing with them (C++11 8.3.2).
For example, reference must always be pointing to an object of the corresponding type and cannot be
NULL [Marshall Cline, C++ FAQ8.6].
Even more than that, references cannot be changed, it is impossible to point them to another object
(reseat) [Marshall Cline, C++ FAQ8.5].
If we are going to change the example with pointers ( 3.23 on page 596) to use references instead …
void f2 (int x, int y, int & sum, int & product)
{
sum=x+y;
product=x*y;
558
};
…then we can see that the compiled code is just the same as in the pointers example ( 3.23 on page 596):

_x$ = 8 ; size = 4
_y$ = 12 ; size = 4
_sum$ = 16 ; size = 4
_product$ = 20 ; size = 4
?f2@@YAXHHAAH0@Z PROC ; f2
mov ecx, DWORD PTR _y$[esp-4]
mov eax, DWORD PTR _x$[esp-4]
lea edx, DWORD PTR [eax+ecx]
imul eax, ecx
mov ecx, DWORD PTR _product$[esp-4]
push esi
mov esi, DWORD PTR _sum$[esp]
mov DWORD PTR [esi], edx
pop esi
ret 0
?f2@@YAXHHAAH0@Z ENDP ; f2
(The reason why C++ functions has such strange names is explained here: 3.21.1 on page 542.)
Hence, C++ references are as much efficient as usual pointers.
3.21.4 STL
N.B.: all examples here were checked only in 32-bit environment. x64 wasn’t checked.
std::string
Internals
Many string libraries [Dennis Yurichev, C/C++ programming language notes2.2] implement a structure
that contains a pointer to a string buffer, a variable that always contains the current string length (which
is very convenient for many functions: [Dennis Yurichev, C/C++ programming language notes2.2.1]) and
a variable containing the current buffer size.
The string in the buffer is usually terminated with zero, in order to be able to pass a pointer to the buffer
into the functions that take usual C ASCIIZ strings.
It is not specified in the C++ standard how std::string has to be implemented, however, it is usually
implemented as explained above.
The C++ string is not a class (as QString in Qt, for instance) but a template (basic_string), this is made in
order to support various character types: at least char and wchar_t.
So, std::string is a class with char as its base type.
And std::wstring is a class with wchar_t as its base type.
MSVC
The MSVC implementation may store the buffer in place instead of using a pointer to a buffer (if the string
is shorter than 16 symbols).
This implies that a short string is to occupy at least 16 + 4 + 4 = 24 bytes in 32-bit environment or at least
16 + 8 + 8 = 32
bytes in 64-bit one, and if the string is longer than 16 characters, we also have to add the length of the
string itself.
Listing 3.104: example for MSVC

#include <string>
#include <stdio.h>
559
struct std_string
{
union
{
char buf[16];
char* ptr;
} u;
size_t size; // AKA 'Mysize' in MSVC
size_t capacity; // AKA 'Myres' in MSVC
};
void dump_std_string(std::string s)
{
struct std_string *p=(struct std_string*)&s;
printf ("[%s] size:%d capacity:%d\n", p->size>16 ? p->u.ptr : p->u.buf, p->size, p->⤦
Ç capacity);
};
int main()
{
std::string s1="a short string";
std::string s2="a string longer than 16 bytes";
dump_std_string(s1);
// that works without using c_str()

printf ("%s\n", &s1);
printf ("%s\n", s2);
};
Almost everything is clear from the source code.

A couple of notes:
If the string is shorter than 16 symbols, a buffer for the string is not to be allocated in the heap.
This is convenient because in practice, a lot of strings are short indeed.
Looks like that Microsoft’s developers chose 16 characters as a good balance.
One very important thing here can be seen at the end of main(): we’re not using the c_str() method,
nevertheless, if we compile and run this code, both strings will appear in the console!
This is why it works.
In the first case the string is shorter than 16 characters and the buffer with the string is located in the
beginning of the std::string object (it can be treated as a structure). printf() treats the pointer as a pointer
to the null-terminated array of characters, hence it works.
Printing the second string (longer than 16 characters) is even more dangerous: it is a typical programmer’s
mistake (or typo) to forget to write c_str().
This works because at the moment a pointer to buffer is located at the start of structure.
This may stay unnoticed for a long time, until a longer string appears there at some time, then the process
will crash.
GCC
GCC’s implementation of this structure has one more variable—reference count.

One interesting fact is that in GCC, a pointer to an instance of std::string instance points not to the be-
ginning of the structure, but to the buffer pointer. In libstdc++-v3\include\bits\basic_string.h we can read
that it was done for more convenient debugging:
* The reason you want _M_data pointing to the character %array and
* not the _Rep is so that the debugger can see the string
* contents. (Probably we should add a non-inline member to get
* the _Rep for the debugger to use, so users can check the actual
* string length.)
560
basic_string.h source code
We consider this in our example:
Listing 3.105: example for GCC

#include <string>
#include <stdio.h>
struct std_string
{
size_t length;
size_t capacity;
size_t refcount;
};
void dump_std_string(std::string s)
{
char *p1=*(char**)&s; // GCC type checking workaround
struct std_string *p2=(struct std_string*)(p1-sizeof(struct std_string));
printf ("[%s] size:%d capacity:%d\n", p1, p2->length, p2->capacity);
};
int main()
{
std::string s1="a short string";
std::string s2="a string longer than 16 bytes";
// GCC type checking workaround:

printf ("%s\n", *(char**)&s1);
printf ("%s\n", *(char**)&s2);
};
A trickery has to be used to imitate the mistake we already have seen above because GCC has stronger
type checking, nevertheless, printf() works here without c_str() as well.
A more advanced example
#include <string>
#include <stdio.h>
int main()
{
std::string s1="Hello, ";
std::string s2="world!\n";
std::string s3=s1+s2;
printf ("%s\n", s3.c_str());

}

$SG39512 DB 'Hello, ', 00H
$SG39514 DB 'world!', 0aH, 00H
$SG39581 DB '%s', 0aH, 00H
_s2$ = -72 ; size = 24

_s3$ = -48 ; size = 24
_s1$ = -24 ; size = 24
_main PROC
sub esp, 72
push 7
lea ecx, DWORD PTR _s1$[esp+80]
mov DWORD PTR _s1$[esp+100], 15
561
mov BYTE PTR _s1$[esp+80], 0
call ?assign@?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@QAEAAV12@PBDI@Z ;
std::basic_string<char,std::char_traits<char>,std::allocator<char> >::assign
push 7
lea ecx, DWORD PTR _s2$[esp+80]
lea eax, DWORD PTR _s2$[esp+72]

push eax
push eax
push eax
call ??$?HDU?$char_traits@D@std@@V?$allocator@D@1@@std@@YA?AV?$basic_string@DU?⤦
Ç $char_traits@D@std@@V?$allocator@D@2@@0@ABV10@0@Z ;
std::operator+<char,std::char_traits<char>,std::allocator<char> >
; inlined c_str() method:

cmp DWORD PTR _s3$[esp+104], 16
cmovae eax, DWORD PTR _s3$[esp+84]
push eax
call _printf
add esp, 20

jb SHORT $LN119@main
push DWORD PTR _s3$[esp+72]
call ??3@YAXPAX@Z ; operator delete
add esp, 4
$LN119@main:
add esp, 4
$LN151@main:
add esp, 4
$LN195@main:
xor eax, eax
add esp, 72
ret 0
_main ENDP
The compiler does not construct strings statically: it would not be possible anyway if the buffer needs to
be located in the heap.
Instead, the ASCIIZ strings are stored in the data segment, and later, at runtime, with the help of the
“assign” method, the s1 and s2 strings are constructed. And with the help of operator+, the s3 string is
constructed.
Please note that there is no call to the c_str() method, because its code is tiny enough so the compiler
562
inlined it right there: if the string is shorter than 16 characters, a pointer to buffer is left in EAX, otherwise
the address of the string buffer located in the heap is fetched.
Next, we see calls to the 3 destructors, they are called if the string is longer than 16 characters: then the
buffers in the heap have to be freed. Otherwise, since all three std::string objects are stored in the stack,
they are freed automatically, when the function ends.
As a consequence, processing short strings is faster, because of less heap accesses.
GCC code is even simpler (because the GCC way, as we saw above, is to not store shorter strings right in
the structure):
.LC0:
.string "Hello, "
.LC1:
.string "world!\n"
main:
push ebp
mov ebp, esp
push edi
push esi
push ebx
and esp, -16
sub esp, 32
lea ebx, [esp+28]
lea edi, [esp+20]
mov DWORD PTR [esp+8], ebx
lea esi, [esp+24]
mov DWORD PTR [esp+4], OFFSET FLAT:.LC0
mov DWORD PTR [esp], edi
call _ZNSsC1EPKcRKSaIcE

mov DWORD PTR [esp], esi
mov DWORD PTR [esp+4], edi

call _ZNSsC1ERKSs
mov DWORD PTR [esp+4], esi

call _ZNSs6appendERKSs
; inlined c_str():
call puts

lea ebx, [esp+19]
sub eax, 12
call _ZNSs4_Rep10_M_disposeERKSaIcE
sub eax, 12
sub eax, 12
563
lea esp, [ebp-12]
xor eax, eax
pop ebx
pop esi
pop edi
pop ebp
ret
It can be seen that it’s not a pointer to the object that is passed to destructors, but rather an address 12
bytes (or 3 words) before, i.e., a pointer to the real start of the structure.
std::string as a global variable
Experienced C++ programmers knows that global variables of STL30 types can be defined without prob-
lems.
Yes, indeed:
#include <stdio.h>
#include <string>
std::string s="a string";
int main()
{
printf ("%s\n", s.c_str());
};
But how and where std::string constructor will be called?

In fact, this variable is to be initialized even before main() start.
Listing 3.108: MSVC 2012: here is how a global variable is constructed and also its destructor is registered
??__Es@@YAXXZ PROC
push 8
push OFFSET $SG39512 ; 'a string'
mov ecx, OFFSET ?s@@3V?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@A ; s
push OFFSET ??__Fs@@YAXXZ ; `dynamic atexit destructor for 's''
call _atexit
pop ecx
ret 0
??__Es@@YAXXZ ENDP
Listing 3.109: MSVC 2012: here a global variable is used in main()

$SG39512 DB 'a string', 00H
$SG39519 DB '%s', 0aH, 00H
_main PROC
cmp DWORD PTR ?s@@3V?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@A+20, 16
mov eax, OFFSET ?s@@3V?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@A ; s
cmovae eax, DWORD PTR ?s@@3V?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@A
push eax
push OFFSET $SG39519 ; '%s'
call _printf
add esp, 8
xor eax, eax
ret 0
_main ENDP
Listing 3.110: MSVC 2012: this destructor function is called before exit
??__Fs@@YAXXZ PROC
push ecx
30 (C++) Standard Template Library
564
cmp DWORD PTR ?s@@3V?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@A+20, 16
jb SHORT $LN23@dynamic
push esi
mov esi, DWORD PTR ?s@@3V?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@A
lea ecx, DWORD PTR $T2[esp+8]
call ??0?$_Wrap_alloc@V?$allocator@D@std@@@std@@QAE@XZ
push OFFSET ?s@@3V?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@A ; s
call ??$destroy@PAD@?$_Wrap_alloc@V?$allocator@D@std@@@std@@QAEXPAPAD@Z
call ??0?$_Wrap_alloc@V?$allocator@D@std@@@std@@QAE@XZ
push esi
add esp, 4
pop esi
$LN23@dynamic:
mov DWORD PTR ?s@@3V?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@A+20, 15
mov DWORD PTR ?s@@3V?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@A+16, 0
mov BYTE PTR ?s@@3V?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@A, 0
pop ecx
ret 0
??__Fs@@YAXXZ ENDP
In fact, a special function with all constructors of global variables is called from CRT, before main().
More than that: with the help of atexit() another function is registered, which contain calls to all destructors
of such global variables.
GCC works likewise:

main:
push ebp
mov ebp, esp
and esp, -16
sub esp, 16
mov eax, DWORD PTR s
call puts
xor eax, eax
leave
ret
.LC0:
.string "a string"
_GLOBAL__sub_I_s:
sub esp, 44
lea eax, [esp+31]
mov DWORD PTR [esp], OFFSET FLAT:s
mov DWORD PTR [esp+8], OFFSET FLAT:__dso_handle
mov DWORD PTR [esp+4], OFFSET FLAT:s
mov DWORD PTR [esp], OFFSET FLAT:_ZNSsD1Ev
call __cxa_atexit
add esp, 44
ret
.LFE645:
.size _GLOBAL__sub_I_s, .-_GLOBAL__sub_I_s
.section .init_array,"aw"
.align 4
.long _GLOBAL__sub_I_s
.globl s
.bss
.align 4
.type s, @object
.size s, 4
s:
.zero 4
.hidden __dso_handle
565
But it does not create a separate function for this, each destructor is passed to atexit(), one by one.
std::list
This is the well-known doubly-linked list: each element has two pointers, to the previous and next ele-
ments.
This implies that the memory footprint is enlarged by 2 words for each element (8 bytes in 32-bit environ-
ment or 16 bytes in 64-bit).
C++ STL just adds the “next” and “previous” pointers to the existing structure of the type that you want
to unite in a list.
Let’s work out an example with a simple 2-variable structure that we want to store in a list.
Although the C++ standard does not say how to implement it, both MSVC’s and GCC’s implementations
are straightforward and similar, so here is only one source code for both:
#include <stdio.h>
#include <list>
#include <iostream>
struct a
{
int x;
int y;
};
struct List_node
{
struct List_node* _Next;
struct List_node* _Prev;
int x;
int y;
};
void dump_List_node (struct List_node *n)

{
printf ("ptr=0x%p _Next=0x%p _Prev=0x%p x=%d y=%d\n",
n, n->_Next, n->_Prev, n->x, n->y);
};
void dump_List_vals (struct List_node* n)

{
struct List_node* current=n;
for (;;)
{
dump_List_node (current);
current=current->_Next;
if (current==n) // end
break;
};
};
void dump_List_val (unsigned int *a)

{
#ifdef _MSC_VER
// GCC implementation does not have "size" field
printf ("_Myhead=0x%p, _Mysize=%d\n", a[0], a[1]);
#endif
dump_List_vals ((struct List_node*)a[0]);
};
int main()
{
std::list<struct a> l;
printf ("* empty list:\n");

dump_List_val((unsigned int*)(void*)&l);
566
struct a t1;
t1.x=1;
t1.y=2;
l.push_front (t1);
t1.x=3;
t1.y=4;
l.push_front (t1);
t1.x=5;
t1.y=6;
l.push_back (t1);
printf ("* 3-elements list:\n");

std::list<struct a>::iterator tmp;

printf ("node at .begin:\n");
tmp=l.begin();
dump_List_node ((struct List_node *)*(void**)&tmp);
printf ("node at .end:\n");
tmp=l.end();
dump_List_node ((struct List_node *)*(void**)&tmp);
printf ("* let's count from the beginning:\n");

std::list<struct a>::iterator it=l.begin();
printf ("1st element: %d %d\n", (*it).x, (*it).y);
it++;
printf ("2nd element: %d %d\n", (*it).x, (*it).y);
it++;
printf ("3rd element: %d %d\n", (*it).x, (*it).y);
it++;
printf ("element at .end(): %d %d\n", (*it).x, (*it).y);
printf ("* let's count from the end:\n");

std::list<struct a>::iterator it2=l.end();
printf ("element at .end(): %d %d\n", (*it2).x, (*it2).y);
it2--;
printf ("3rd element: %d %d\n", (*it2).x, (*it2).y);
it2--;
printf ("2nd element: %d %d\n", (*it2).x, (*it2).y);
it2--;
printf ("1st element: %d %d\n", (*it2).x, (*it2).y);
printf ("removing last element...\n");

l.pop_back();
};
GCC
Let’s start with GCC.

When we run the example, we’ll see a long dump, let’s work with it in pieces.
* empty list:
ptr=0x0028fe90 _Next=0x0028fe90 _Prev=0x0028fe90 x=3 y=0
Here we see an empty list.

Despite the fact it is empty, it has one element with garbage (AKA dummy node) in x and y. Both the
“next” and “prev” pointers are pointing to the self node:
567
Variable
list.begin() list.end()
std::list
Next
Prev
X=garbage
Y=garbage
At this moment, the .begin and .end iterators are equal to each other.
If we push 3 elements, the list internally will be:
* 3-elements list:
ptr=0x000349a0 _Next=0x00034988 _Prev=0x0028fe90 x=3 y=4
ptr=0x00034988 _Next=0x00034b40 _Prev=0x000349a0 x=1 y=2
ptr=0x00034b40 _Next=0x0028fe90 _Prev=0x00034988 x=5 y=6
ptr=0x0028fe90 _Next=0x000349a0 _Prev=0x00034b40 x=5 y=6
The last element is still at 0x0028fe90, it not to be moved until the list’s disposal.
It still contain random garbage in x and y (5 and 6). By coincidence, these values are the same as in the
last element, but it doesn’t mean that they are meaningful.
Here is how these 3 elements are stored in memory:
Variable
std::list
list.begin() list.end()
Next Next Next Next
Prev Prev Prev Prev
X=1st ele- X=2nd ele- X=3rd ele-

X=garbage
ment ment ment
Y=1st ele- Y=2nd ele- Y=3rd ele-
Y=garbage
ment ment ment
The l variable always points to the first node.

The .begin() and .end() iterators are not variables, but functions, which when called return pointers to the
corresponding nodes.
Having a dummy element (AKA sentinel node) is a very popular practice in implementing doubly-linked
lists.
Without it, a lot of operations may become slightly more complex and, hence, slower.
The iterator is in fact just a pointer to a node. list.begin() and list.end() just return pointers.
node at .begin:
node at .end:
The fact that the last element has a pointer to the first and the first element has a pointer to the last one
remind us of circular lists.
568
This is very helpful here: having a pointer to the first list element, i.e., that is in the l variable, it is easy
to get a pointer to the last one quickly, without the necessity to traverse the whole list.
Inserting an element at the end of the list is also quick, thanks to this feature.
operator-- and operator++ just set the current iterator’s value to the
current_node->prev or current_node->next values.
The reverse iterators (.rbegin, .rend) work just as the same, but in reverse.
operator* just returns a pointer to the point in the node structure, where the user’s structure starts, i.e.,
a pointer to the first element of the structure (x).
The list insertion and deletion are trivial: just allocate a new node (or deallocate) and update all pointers
to be valid.
That’s why an iterator may become invalid after element deletion: it may still point to the node that has
been already deallocated. This is also called a dangling pointer.
And of course, the information from the freed node (to which iterator still points) cannot be used anymore.
The GCC implementation (as of 4.8.1) doesn’t store the current size of the list: this implies a slow .size()
method: it has to traverse the whole list to count the elements, because it doesn’t have any other way to
get the information.
This means that this operation is O(n), i.e., it steadily gets slower as the list grows.
Listing 3.112: Optimizing GCC 4.8.1 -fno-inline-small-functions

main proc near
push ebp
mov ebp, esp
push esi
push ebx
and esp, 0FFFFFFF0h
sub esp, 20h
lea ebx, [esp+10h]
mov dword ptr [esp], offset s ; "* empty list:"
mov [esp+10h], ebx
mov [esp+14h], ebx
call puts
mov [esp], ebx
call _Z13dump_List_valPj ; dump_List_val(uint *)
lea esi, [esp+18h]
mov [esp+4], esi
mov [esp], ebx
mov dword ptr [esp+18h], 1 ; X for new element
mov dword ptr [esp+1Ch], 2 ; Y for new element
call _ZNSt4listI1aSaIS0_EE10push_frontERKS0_ ;
std::list<a,std::allocator<a>>::push_front(a const&)
mov [esp+4], esi
mov [esp], ebx
call _ZNSt4listI1aSaIS0_EE10push_frontERKS0_ ;
std::list<a,std::allocator<a>>::push_front(a const&)
mov dword ptr [esp], 10h
call _Znwj ; operator new(uint)
cmp eax, 0FFFFFFF8h
jz short loc_80002A6
mov ecx, [esp+1Ch]
mov edx, [esp+18h]
mov [eax+0Ch], ecx
mov [eax+8], edx
loc_80002A6: ; CODE XREF: main+86

mov [esp+4], ebx
mov [esp], eax
call _ZNSt8__detail15_List_node_base7_M_hookEPS0_ ;
std::__detail::_List_node_base::_M_hook(std::__detail::_List_node_base*)
mov dword ptr [esp], offset a3ElementsList ; "* 3-elements list:"
call puts
569
mov [esp], ebx
mov dword ptr [esp], offset aNodeAt_begin ; "node at .begin:"
call puts
mov eax, [esp+10h]
mov [esp], eax
call _Z14dump_List_nodeP9List_node ; dump_List_node(List_node *)
mov dword ptr [esp], offset aNodeAt_end ; "node at .end:"
call puts
mov [esp], ebx
call _Z14dump_List_nodeP9List_node ; dump_List_node(List_node *)
mov dword ptr [esp], offset aLetSCountFromT ; "* let's count from the beginning:"
call puts
mov esi, [esp+10h]
mov eax, [esi+0Ch]
mov [esp+0Ch], eax
mov eax, [esi+8]
mov dword ptr [esp+4], offset a1stElementDD ; "1st element: %d %d\n"
mov [esp+8], eax
call __printf_chk
mov esi, [esi] ; operator++: get ->next pointer
mov eax, [esi+0Ch]
mov [esp+0Ch], eax
mov eax, [esi+8]
mov dword ptr [esp+4], offset a2ndElementDD ; "2nd element: %d %d\n"
mov [esp+8], eax
call __printf_chk
mov esi, [esi] ; operator++: get ->next pointer
mov eax, [esi+0Ch]
mov [esp+0Ch], eax
mov eax, [esi+8]
mov dword ptr [esp+4], offset a3rdElementDD ; "3rd element: %d %d\n"
mov [esp+8], eax
call __printf_chk
mov eax, [esi] ; operator++: get ->next pointer
mov edx, [eax+0Ch]
mov [esp+0Ch], edx
mov eax, [eax+8]
mov dword ptr [esp+4], offset aElementAt_endD ; "element at .end(): %d %d\n"
mov [esp+8], eax
call __printf_chk
mov dword ptr [esp], offset aLetSCountFro_0 ; "* let's count from the end:"
call puts
mov eax, [esp+1Ch]
mov dword ptr [esp+4], offset aElementAt_endD ; "element at .end(): %d %d\n"
mov [esp+0Ch], eax
mov eax, [esp+18h]
mov [esp+8], eax
call __printf_chk
mov esi, [esp+14h]
mov eax, [esi+0Ch]
mov [esp+0Ch], eax
mov eax, [esi+8]
mov dword ptr [esp+4], offset a3rdElementDD ; "3rd element: %d %d\n"
mov [esp+8], eax
call __printf_chk
mov esi, [esi+4] ; operator--: get ->prev pointer
mov eax, [esi+0Ch]
mov [esp+0Ch], eax
mov eax, [esi+8]
mov dword ptr [esp+4], offset a2ndElementDD ; "2nd element: %d %d\n"
mov [esp+8], eax
call __printf_chk
570
mov eax, [esi+4] ; operator--: get ->prev pointer
mov edx, [eax+0Ch]
mov [esp+0Ch], edx
mov eax, [eax+8]
mov dword ptr [esp+4], offset a1stElementDD ; "1st element: %d %d\n"
mov [esp+8], eax
call __printf_chk
mov dword ptr [esp], offset aRemovingLastEl ; "removing last element..."
call puts
mov esi, [esp+14h]
mov [esp], esi
call _ZNSt8__detail15_List_node_base9_M_unhookEv ;
std::__detail::_List_node_base::_M_unhook(void)
mov [esp], esi ; void *
call _ZdlPv ; operator delete(void *)
mov [esp], ebx
mov [esp], ebx
call _ZNSt10_List_baseI1aSaIS0_EE8_M_clearEv ;
std::_List_base<a,std::allocator<a>>::_M_clear(void)
lea esp, [ebp-8]
xor eax, eax
pop ebx
pop esi
pop ebp
retn
main endp
Listing 3.113: The whole output

* empty list:
ptr=0x0028fe90 _Next=0x0028fe90 _Prev=0x0028fe90 x=3 y=0
* 3-elements list:
ptr=0x00034988 _Next=0x00034b40 _Prev=0x000349a0 x=1 y=2
ptr=0x00034b40 _Next=0x0028fe90 _Prev=0x00034988 x=5 y=6
node at .begin:
node at .end:
* let's count from the beginning:
1st element: 3 4
2nd element: 1 2
3rd element: 5 6
element at .end(): 5 6
* let's count from the end:
element at .end(): 5 6
3rd element: 5 6
2nd element: 1 2
1st element: 3 4
removing last element...
ptr=0x00034988 _Next=0x0028fe90 _Prev=0x000349a0 x=1 y=2
ptr=0x0028fe90 _Next=0x000349a0 _Prev=0x00034988 x=5 y=6
MSVC
MSVC’s implementation (2012) is just the same, but it also stores the current size of the list.
This implies that the .size() method is very fast (O(1)): it just reads one value from memory.
On the other hand, the size variable must be updated at each insertion/deletion.
MSVC’s implementation is also slightly different in the way it arranges the nodes:
571
Variable
std::list
list.end() list.begin()
Next Next Next Next
Prev Prev Prev Prev
X=1st ele- X=2nd ele- X=3rd ele-

X=garbage
ment ment ment
Y=2nd ele-
Y=1st ele- Y=3rd ele-
Y=garbage ment2nd élé-
ment ment
ment
GCC has its dummy element at the end of the list, while MSVC’s is at the beginning.
Listing 3.114: Optimizing MSVC 2012 /Fa2.asm /GS- /Ob1

_l$ = -16 ; size = 8
_t1$ = -8 ; size = 8
_main PROC
sub esp, 16
push ebx
push esi
push edi
push 0
push 0
lea ecx, DWORD PTR _l$[esp+36]
mov DWORD PTR _l$[esp+40], 0
; allocate first garbage element
call ?_Buynode0@?$_List_alloc@$0A@U?$_List_base_types@Ua@@V?⤦
Ç $allocator@Ua@@@std@@@std@@@std@@QAEPAU?$_List_node@Ua@@PAX@2@PAU32@0@Z ;
std::_List_alloc<0,std::_List_base_types<a,std::allocator<a> > >::_Buynode0
mov edi, DWORD PTR __imp__printf
mov ebx, eax
push OFFSET $SG40685 ; '* empty list:'
mov DWORD PTR _l$[esp+32], ebx
call edi ; printf
lea eax, DWORD PTR _l$[esp+32]
push eax
call ?dump_List_val@@YAXPAI@Z ; dump_List_val
mov esi, DWORD PTR [ebx]
add esp, 8
lea eax, DWORD PTR _t1$[esp+28]
push eax
push DWORD PTR [esi+4]
push esi
mov DWORD PTR _t1$[esp+40], 1 ; data for a new node
; allocate new node
call ??$_Buynode@ABUa@@@?$_List_buy@Ua@@V?$allocator@Ua@@@std@@@std@@QAEPAU?⤦
Ç $_List_node@Ua@@PAX@1@PAU21@0ABUa@@@Z ;
std::_List_buy<a,std::allocator<a> >::_Buynode<a const &>
mov esi, DWORD PTR [ebx]
push eax
push esi
572
; allocate new node
push eax
push DWORD PTR [ebx+4]
push ebx
; allocate new node
mov DWORD PTR [ebx+4], eax
push OFFSET $SG40689 ; '* 3-elements list:'
call edi ; printf
push eax
push OFFSET $SG40831 ; 'node at .begin:'
call edi ; printf
push DWORD PTR [ebx] ; get next field of node "l" variable points to
call ?dump_List_node@@YAXPAUList_node@@@Z ; dump_List_node
push OFFSET $SG40835 ; 'node at .end:'
call edi ; printf
push ebx ; pointer to the node "l" variable points to!
call ?dump_List_node@@YAXPAUList_node@@@Z ; dump_List_node
push OFFSET $SG40839 ; '* let''s count from the begin:'
call edi ; printf
mov esi, DWORD PTR [ebx] ; operator++: get ->next pointer
push OFFSET $SG40846 ; '1st element: %d %d'
call edi ; printf
mov esi, DWORD PTR [esi] ; operator++: get ->next pointer
push OFFSET $SG40848 ; '2nd element: %d %d'
call edi ; printf
mov esi, DWORD PTR [esi] ; operator++: get ->next pointer
push OFFSET $SG40850 ; '3rd element: %d %d'
call edi ; printf
mov eax, DWORD PTR [esi] ; operator++: get ->next pointer
add esp, 64
push DWORD PTR [eax+12]
push OFFSET $SG40852 ; 'element at .end(): %d %d'
call edi ; printf
push OFFSET $SG40853 ; '* let''s count from the end:'
call edi ; printf
push DWORD PTR [ebx+12] ; use x and y fields from the node "l" variable points to
push DWORD PTR [ebx+8]
push OFFSET $SG40860 ; 'element at .end(): %d %d'
call edi ; printf
mov esi, DWORD PTR [ebx+4] ; operator--: get ->prev pointer
push OFFSET $SG40862 ; '3rd element: %d %d'
call edi ; printf
mov esi, DWORD PTR [esi+4] ; operator--: get ->prev pointer
573
push OFFSET $SG40864 ; '2nd element: %d %d'
call edi ; printf
mov eax, DWORD PTR [esi+4] ; operator--: get ->prev pointer
push OFFSET $SG40866 ; '1st element: %d %d'
call edi ; printf
add esp, 64
push OFFSET $SG40867 ; 'removing last element...'
call edi ; printf
mov edx, DWORD PTR [ebx+4]
add esp, 4
; prev=next?
; it is the only element, garbage one?
; if yes, do not delete it!
cmp edx, ebx
je SHORT $LN349@main
mov ecx, DWORD PTR [edx+4]
mov ecx, DWORD PTR [edx]
mov eax, DWORD PTR [edx+4]
push edx
mov DWORD PTR [ecx+4], eax
add esp, 4
$LN349@main:
push eax
mov eax, DWORD PTR [ebx]
add esp, 4
mov DWORD PTR [ebx], ebx
mov DWORD PTR [ebx+4], ebx
cmp eax, ebx
je SHORT $LN412@main
$LL414@main:
mov esi, DWORD PTR [eax]
push eax
add esp, 4
mov eax, esi
cmp esi, ebx
jne SHORT $LL414@main
$LN412@main:
push ebx
add esp, 4
xor eax, eax
pop edi
pop esi
pop ebx
add esp, 16
ret 0
_main ENDP
Unlike GCC, MSVC’s code allocates the dummy element at the start of the function with the help of the
“Buynode” function, it is also used to allocate the rest of the nodes ( GCC’s code allocates the first element
in the local stack).
Listing 3.115: The whole output

* empty list:
_Myhead=0x003CC258, _Mysize=0
ptr=0x003CC258 _Next=0x003CC258 _Prev=0x003CC258 x=6226002 y=4522072
* 3-elements list:
574
ptr=0x003CC258 _Next=0x003CC288 _Prev=0x003CC2A0 x=6226002 y=4522072
ptr=0x003CC270 _Next=0x003CC2A0 _Prev=0x003CC288 x=1 y=2
ptr=0x003CC2A0 _Next=0x003CC258 _Prev=0x003CC270 x=5 y=6
node at .begin:
node at .end:
ptr=0x003CC258 _Next=0x003CC288 _Prev=0x003CC2A0 x=6226002 y=4522072
* let's count from the beginning:
1st element: 3 4
2nd element: 1 2
3rd element: 5 6
element at .end(): 6226002 4522072
* let's count from the end:
element at .end(): 6226002 4522072
3rd element: 5 6
2nd element: 1 2
1st element: 3 4
removing last element...
C++11 std::forward_list
The same thing as std::list, but singly-linked one, i.e., having only the “next” field at each node.
It has a smaller memory footprint, but also don’t offer the ability to traverse list backwards.
std::vector
We would call std::vector a safe wrapper of the PODT31 C array. Internally it is somewhat similar to
std::string ( 3.21.4 on page 559): it has a pointer to the allocated buffer, a pointer to the end of the
array, and a pointer to the end of the allocated buffer.
The array’s elements lie in memory adjacently to each other, just like in a normal array ( 1.26 on page 264).
In C++11 there is a new method called .data() , that returns a pointer to the buffer, like .c_str() in
std::string.
The buffer allocated in the heap can be larger than the array itself.
Both MSVC’s and GCC’s implementations are similar, just the names of the structure’s fields are slightly
different32 , so here is one source code that works for both compilers. Here is again the C-like code for
dumping the structure of std::vector:
#include <stdio.h>
#include <vector>
#include <algorithm>
#include <functional>
struct vector_of_ints
{
// MSVC names:
int *Myfirst;
int *Mylast;
int *Myend;
// GCC structure is the same, but names are: _M_start, _M_finish, _M_end_of_storage
};
void dump(struct vector_of_ints *in)

{
printf ("_Myfirst=%p, _Mylast=%p, _Myend=%p\n", in->Myfirst, in->Mylast, in->Myend);
size_t size=(in->Mylast-in->Myfirst);
31 (C++) Plain Old Data Type
32 GCC internals: http://go.yurichev.com/17086
575
size_t capacity=(in->Myend-in->Myfirst);
printf ("size=%d, capacity=%d\n", size, capacity);
for (size_t i=0; i<size; i++)
printf ("element %d: %d\n", i, in->Myfirst[i]);
};
int main()
{
std::vector<int> c;
dump ((struct vector_of_ints*)(void*)&c);
c.push_back(1);
c.push_back(2);
c.push_back(3);
c.push_back(4);
c.reserve (6);
c.push_back(5);
c.push_back(6);
printf ("%d\n", c.at(5)); // with bounds checking
printf ("%d\n", c[8]); // operator[], without bounds checking
};
Here is the output of this program when compiled in MSVC:

_Myfirst=00000000, _Mylast=00000000, _Myend=00000000
size=0, capacity=0
_Myfirst=0051CF48, _Mylast=0051CF4C, _Myend=0051CF4C
size=1, capacity=1
element 0: 1
_Myfirst=0051CF58, _Mylast=0051CF60, _Myend=0051CF60
size=2, capacity=2
element 0: 1
element 1: 2
_Myfirst=0051C278, _Mylast=0051C284, _Myend=0051C284
size=3, capacity=3
element 0: 1
element 1: 2
element 2: 3
_Myfirst=0051C290, _Mylast=0051C2A0, _Myend=0051C2A0
size=4, capacity=4
element 0: 1
element 1: 2
element 2: 3
element 3: 4
_Myfirst=0051B180, _Mylast=0051B190, _Myend=0051B198
size=4, capacity=6
element 0: 1
element 1: 2
element 2: 3
element 3: 4
size=5, capacity=6
element 0: 1
element 1: 2
element 2: 3
element 3: 4
element 4: 5
size=6, capacity=6
element 0: 1
element 1: 2
element 2: 3
element 3: 4
element 4: 5
576
element 5: 6
6
6619158
As it can be seen, there is no allocated buffer when main() starts. After the first push_back() call, a
buffer is allocated. And then, after each push_back() call, both array size and buffer size (capacity) are
increased. But the buffer address changes as well, because push_back() reallocates the buffer in the
heap each time. It is costly operation, that’s why it is very important to predict the size of the array in the
future and reserve enough space for it with the .reserve() method.
The last number is garbage: there are no array elements at this point, so a random number is printed. This
illustrates the fact that operator[] of std::vector does not check of the index is in the array’s bounds.
The slower .at() method, however, does this checking and throws an std::out_of_range exception in
case of error.
Let’s see the code:
Listing 3.116: MSVC 2012 /GS- /Ob1

$SG52650 DB '%d', 0aH, 00H
$SG52651 DB '%d', 0aH, 00H
_this$ = -4 ; size = 4
__Pos$ = 8 ; size = 4
?at@?$vector@HV?$allocator@H@std@@@std@@QAEAAHI@Z PROC ;
std::vector<int,std::allocator<int> >::at, COMDAT
; _this$ = ecx
push ebp
mov ebp, esp
push ecx
mov ecx, DWORD PTR _this$[ebp]
mov edx, DWORD PTR [eax+4]
sub edx, DWORD PTR [ecx]
sar edx, 2
cmp edx, DWORD PTR __Pos$[ebp]
ja SHORT $LN1@at
push OFFSET ??_C@_0BM@NMJKDPPO@invalid?5vector?$DMT?$DO?5subscript?$AA@
call DWORD PTR __imp_?_Xout_of_range@std@@YAXPBD@Z
$LN1@at:
mov ecx, DWORD PTR [eax]
mov edx, DWORD PTR __Pos$[ebp]
lea eax, DWORD PTR [ecx+edx*4]
$LN3@at:
mov esp, ebp
pop ebp
ret 4
?at@?$vector@HV?$allocator@H@std@@@std@@QAEAAHI@Z ENDP ; std::vector<int,std::allocator<int>
>::at
_c$ = -36 ; size = 12

$T1 = -24 ; size = 4
$T2 = -20 ; size = 4
$T3 = -16 ; size = 4
$T4 = -12 ; size = 4
$T5 = -8 ; size = 4
$T6 = -4 ; size = 4
_main PROC
push ebp
mov ebp, esp
sub esp, 36
mov DWORD PTR _c$[ebp], 0 ; Myfirst
mov DWORD PTR _c$[ebp+4], 0 ; Mylast
mov DWORD PTR _c$[ebp+8], 0 ; Myend
lea eax, DWORD PTR _c$[ebp]
push eax
call ?dump@@YAXPAUvector_of_ints@@@Z ; dump
add esp, 4
mov DWORD PTR $T6[ebp], 1
577
lea ecx, DWORD PTR $T6[ebp]
push ecx
lea ecx, DWORD PTR _c$[ebp]
call ?push_back@?$vector@HV?$allocator@H@std@@@std@@QAEX$$QAH@Z ;
std::vector<int,std::allocator<int> >::push_back
lea edx, DWORD PTR _c$[ebp]
push edx
add esp, 4
lea eax, DWORD PTR $T5[ebp]
push eax
push ecx
add esp, 4
lea edx, DWORD PTR $T4[ebp]
push edx
push eax
add esp, 4
push ecx
push edx
add esp, 4
push 6
call ?reserve@?$vector@HV?$allocator@H@std@@@std@@QAEXI@Z ;
std::vector<int,std::allocator<int> >::reserve
push eax
add esp, 4
push ecx
push edx
add esp, 4
push eax
push ecx
add esp, 4
push 5
call ?at@?$vector@HV?$allocator@H@std@@@std@@QAEAAHI@Z ; std::vector<int,std::allocator<int>
>::at
578
push edx
add esp, 8
mov eax, 8
shl eax, 2
mov ecx, DWORD PTR _c$[ebp]
mov edx, DWORD PTR [ecx+eax]
push edx
add esp, 8
call ?_Tidy@?$vector@HV?$allocator@H@std@@@std@@IAEXXZ ;
std::vector<int,std::allocator<int> >::_Tidy
xor eax, eax
mov esp, ebp
pop ebp
ret 0
_main ENDP
We see how the .at() method checks the bounds and throws an exception in case of error. The number
that the last printf() call prints is just taken from the memory, without any checks.
One may ask, why not use the variables like “size” and “capacity”, like it was done in std::string.
Supposedly, this was done for faster bounds checking.
The code GCC generates is in general almost the same, but the .at() method is inlined:
Listing 3.117: GCC 4.8.1 -fno-inline-small-functions -O1
main proc near
push ebp
mov ebp, esp
push edi
push esi
push ebx
and esp, 0FFFFFFF0h
sub esp, 20h
mov dword ptr [esp+14h], 0
mov dword ptr [esp+1Ch], 0
lea eax, [esp+14h]
mov [esp], eax
call _Z4dumpP14vector_of_ints ; dump(vector_of_ints *)
lea eax, [esp+10h]
mov [esp+4], eax
lea eax, [esp+14h]
mov [esp], eax
call _ZNSt6vectorIiSaIiEE9push_backERKi ;
std::vector<int,std::allocator<int>>::push_back(int const&)
lea eax, [esp+14h]
mov [esp], eax
lea eax, [esp+10h]
mov [esp+4], eax
lea eax, [esp+14h]
mov [esp], eax
lea eax, [esp+14h]
mov [esp], eax
lea eax, [esp+10h]
mov [esp+4], eax
lea eax, [esp+14h]
mov [esp], eax
579
lea eax, [esp+14h]
mov [esp], eax
lea eax, [esp+10h]
mov [esp+4], eax
lea eax, [esp+14h]
mov [esp], eax
lea eax, [esp+14h]
mov [esp], eax
mov ebx, [esp+14h]
mov eax, [esp+1Ch]
sub eax, ebx
cmp eax, 17h
ja short loc_80001CF
mov edi, [esp+18h]
sub edi, ebx
sar edi, 2
mov dword ptr [esp], 18h
call _Znwj ; operator new(uint)
mov esi, eax
test edi, edi
jz short loc_80001AD
lea eax, ds:0[edi*4]
mov [esp+8], eax ; n
mov [esp+4], ebx ; src
mov [esp], esi ; dest
call memmove
loc_80001AD: ; CODE XREF: main+F8

mov eax, [esp+14h]
test eax, eax
jz short loc_80001BD
mov [esp], eax ; void *
loc_80001BD: ; CODE XREF: main+117

mov [esp+14h], esi
lea eax, [esi+edi*4]
mov [esp+18h], eax
add esi, 18h
mov [esp+1Ch], esi
loc_80001CF: ; CODE XREF: main+DD

lea eax, [esp+14h]
mov [esp], eax
lea eax, [esp+10h]
mov [esp+4], eax
lea eax, [esp+14h]
mov [esp], eax
lea eax, [esp+14h]
mov [esp], eax
lea eax, [esp+10h]
mov [esp+4], eax
lea eax, [esp+14h]
mov [esp], eax
lea eax, [esp+14h]
mov [esp], eax
580
mov eax, [esp+14h]
mov edx, [esp+18h]
sub edx, eax
cmp edx, 17h
mov dword ptr [esp], offset aVector_m_range ; "vector::_M_range_check"
call _ZSt20__throw_out_of_rangePKc ; std::__throw_out_of_range(char const*)
loc_8000246: ; CODE XREF: main+19C

mov eax, [eax+14h]
mov [esp+8], eax
mov dword ptr [esp+4], offset aD ; "%d\n"
call __printf_chk
mov eax, [esp+14h]
mov eax, [eax+20h]
mov [esp+8], eax
mov dword ptr [esp+4], offset aD ; "%d\n"
call __printf_chk
mov eax, [esp+14h]
test eax, eax
jz short loc_80002AC
mov [esp], eax ; void *
jmp short loc_80002AC
mov ebx, eax

mov edx, [esp+14h]
test edx, edx
jz short loc_80002A4
mov [esp], edx ; void *
loc_80002A4: ; CODE XREF: main+1FE

mov [esp], ebx
call _Unwind_Resume
loc_80002AC: ; CODE XREF: main+1EA

; main+1F4
mov eax, 0
lea esp, [ebp-0Ch]
pop ebx
pop esi
pop edi
pop ebp
locret_80002B8: ; DATA XREF: .eh_frame:08000510

; .eh_frame:080005BC
retn
main endp
.reserve() is inlined as well. It calls new() if the buffer is too small for the new size, calls memmove() to
copy the contents of the buffer, and calls delete() to free the old buffer.
Let’s also see what the compiled program outputs if compiled with GCC:
_Myfirst=0x(nil), _Mylast=0x(nil), _Myend=0x(nil)
size=0, capacity=0
_Myfirst=0x8257008, _Mylast=0x825700c, _Myend=0x825700c
size=1, capacity=1
element 0: 1
_Myfirst=0x8257018, _Mylast=0x8257020, _Myend=0x8257020
size=2, capacity=2
element 0: 1
element 1: 2
size=3, capacity=4
element 0: 1
581
element 1: 2
element 2: 3
size=4, capacity=4
element 0: 1
element 1: 2
element 2: 3
element 3: 4
size=4, capacity=6
element 0: 1
element 1: 2
element 2: 3
element 3: 4
size=5, capacity=6
element 0: 1
element 1: 2
element 2: 3
element 3: 4
element 4: 5
size=6, capacity=6
element 0: 1
element 1: 2
element 2: 3
element 3: 4
element 4: 5
element 5: 6
6
0
We can spot that the buffer size grows in a different way that in MSVC.
Simple experimentation shows that in MSVC’s implementation the buffer grows by ~50% each time it
needs to be enlarged, while GCC’s code enlarges it by 100% each time, i.e., doubles it.
std::map and std::set
The binary tree is another fundamental data structure.

As its name states, this is a tree where each node has at most 2 links to other nodes. Each node has key
and/or value: std::set provides only key at each node, std::map provides both key and value at each
node.
Binary trees are usually the structure used in the implementation of “dictionaries” of key-values (AKA
“associative arrays”).
There are at least three important properties that a binary trees has:
• All keys are always stored in sorted form.
• Keys of any types can be stored easily. Binary tree algorithms are unaware of the key’s type, only a
key comparison function is required.
• Finding a specific key is relatively fast in comparison with lists and arrays.
Here is a very simple example: let’s store these numbers in a binary tree: 0, 1, 2, 3, 5, 6, 9, 10, 11, 12,
20, 99, 100, 101, 107, 1001, 1010.
582
10
1 100
0 5 20 107
3 6 12 99 101 1001
2 9 11 1010
All keys that are smaller than the node key’s value are stored on the left side.
All keys that are bigger than the node key’s value are stored on the right side.
Hence, the lookup algorithm is straightforward: if the value that you are looking for is smaller than the
current node’s key value: move left, if it is bigger: move right, stop if the value required is equal to the
node key’s value.
That is why the searching algorithm may search for numbers, text strings, etc., as long as a key comparison
function is provided.
All keys have unique values.
Having that, one needs ≈ log2 n steps in order to find a key in a balanced binary tree with n keys. This
implies that ≈ 10 steps are needed ≈ 1000 keys, or ≈ 13 steps for ≈ 10000 keys.
Not bad, but the tree has always to be balanced for this: i.e., the keys has to be distributed evenly on all
levels. The insertion and removal operations do some maintenance to keep the tree in a balanced state.
There are several popular balancing algorithms available, including the AVL tree and the red-black tree.
The latter extends each node with a “color” value to simplify the balancing process, hence, each node
may be “red” or “black”.
Both GCC’s and MSVC’s std::map and std::set template implementations use red-black trees.
std::set has only keys. std::map is the “extended” version of std::set: it also has a value at each node.
MSVC
#include <map>
#include <set>
#include <string>
#include <iostream>
// Structure is not packed! Each field occupies 4 bytes.

struct tree_node
{
struct tree_node *Left;
struct tree_node *Parent;
struct tree_node *Right;
char Color; // 0 - Red, 1 - Black
char Isnil;
//std::pair Myval;
unsigned int first; // called Myval in std::set
const char *second; // not present in std::set
};
struct tree_struct
{
struct tree_node *Myhead;
size_t Mysize;
};
583
void dump_tree_node (struct tree_node *n, bool is_set, bool traverse)
{
printf ("ptr=0x%p Left=0x%p Parent=0x%p Right=0x%p Color=%d Isnil=%d\n",
n, n->Left, n->Parent, n->Right, n->Color, n->Isnil);
if (n->Isnil==0)
{
if (is_set)
printf ("first=%d\n", n->first);
else
printf ("first=%d second=[%s]\n", n->first, n->second);
}
if (traverse)
{
if (n->Isnil==1)
dump_tree_node (n->Parent, is_set, true);
else
{
if (n->Left->Isnil==0)
dump_tree_node (n->Left, is_set, true);
if (n->Right->Isnil==0)
dump_tree_node (n->Right, is_set, true);
};
};
};
const char* ALOT_OF_TABS="\t\t\t\t\t\t\t\t\t\t\t";
void dump_as_tree (int tabs, struct tree_node *n, bool is_set)

{
if (is_set)
printf ("%d\n", n->first);
else
printf ("%d [%s]\n", n->first, n->second);
if (n->Left->Isnil==0)
{
printf ("%.*sL-------", tabs, ALOT_OF_TABS);
dump_as_tree (tabs+1, n->Left, is_set);
};
if (n->Right->Isnil==0)
{
printf ("%.*sR-------", tabs, ALOT_OF_TABS);
dump_as_tree (tabs+1, n->Right, is_set);
};
};
void dump_map_and_set(struct tree_struct *m, bool is_set)

{
printf ("ptr=0x%p, Myhead=0x%p, Mysize=%d\n", m, m->Myhead, m->Mysize);
dump_tree_node (m->Myhead, is_set, true);
printf ("As a tree:\n");
printf ("root----");
dump_as_tree (1, m->Myhead->Parent, is_set);
};
int main()
{
// map
std::map<int, const char*> m;
m[10]="ten";
m[20]="twenty";
m[3]="three";
m[101]="one hundred one";
m[100]="one hundred";
m[12]="twelve";
m[107]="one hundred seven";
m[0]="zero";
584
m[1]="one";
m[6]="six";
m[99]="ninety-nine";
m[5]="five";
m[11]="eleven";
m[1001]="one thousand one";
m[1010]="one thousand ten";
m[2]="two";
m[9]="nine";
printf ("dumping m as map:\n");
dump_map_and_set ((struct tree_struct *)(void*)&m, false);
std::map<int, const char*>::iterator it1=m.begin();

printf ("m.begin():\n");
dump_tree_node ((struct tree_node *)*(void**)&it1, false, false);
it1=m.end();
printf ("m.end():\n");
dump_tree_node ((struct tree_node *)*(void**)&it1, false, false);
// set
std::set<int> s;
s.insert(123);
s.insert(456);
s.insert(11);
s.insert(12);
s.insert(100);
s.insert(1001);
printf ("dumping s as set:\n");
dump_map_and_set ((struct tree_struct *)(void*)&s, true);
std::set<int>::iterator it2=s.begin();
printf ("s.begin():\n");
dump_tree_node ((struct tree_node *)*(void**)&it2, true, false);
it2=s.end();
printf ("s.end():\n");
dump_tree_node ((struct tree_node *)*(void**)&it2, true, false);
};

dumping m as map:
ptr=0x0020FE04, Myhead=0x005BB3A0, Mysize=17
ptr=0x005BB3A0 Left=0x005BB4A0 Parent=0x005BB3C0 Right=0x005BB580 Color=1 Isnil=1
ptr=0x005BB3C0 Left=0x005BB4C0 Parent=0x005BB3A0 Right=0x005BB440 Color=1 Isnil=0
first=10 second=[ten]
ptr=0x005BB4C0 Left=0x005BB4A0 Parent=0x005BB3C0 Right=0x005BB520 Color=1 Isnil=0
first=1 second=[one]
ptr=0x005BB4A0 Left=0x005BB3A0 Parent=0x005BB4C0 Right=0x005BB3A0 Color=1 Isnil=0
first=0 second=[zero]
ptr=0x005BB520 Left=0x005BB400 Parent=0x005BB4C0 Right=0x005BB4E0 Color=0 Isnil=0
first=5 second=[five]
ptr=0x005BB400 Left=0x005BB5A0 Parent=0x005BB520 Right=0x005BB3A0 Color=1 Isnil=0
first=3 second=[three]
ptr=0x005BB5A0 Left=0x005BB3A0 Parent=0x005BB400 Right=0x005BB3A0 Color=0 Isnil=0
first=2 second=[two]
ptr=0x005BB4E0 Left=0x005BB3A0 Parent=0x005BB520 Right=0x005BB5C0 Color=1 Isnil=0
first=6 second=[six]
ptr=0x005BB5C0 Left=0x005BB3A0 Parent=0x005BB4E0 Right=0x005BB3A0 Color=0 Isnil=0
first=9 second=[nine]
ptr=0x005BB440 Left=0x005BB3E0 Parent=0x005BB3C0 Right=0x005BB480 Color=1 Isnil=0
first=100 second=[one hundred]
ptr=0x005BB3E0 Left=0x005BB460 Parent=0x005BB440 Right=0x005BB500 Color=0 Isnil=0
first=20 second=[twenty]
ptr=0x005BB460 Left=0x005BB540 Parent=0x005BB3E0 Right=0x005BB3A0 Color=1 Isnil=0
first=12 second=[twelve]
first=11 second=[eleven]
ptr=0x005BB500 Left=0x005BB3A0 Parent=0x005BB3E0 Right=0x005BB3A0 Color=1 Isnil=0
first=99 second=[ninety-nine]
ptr=0x005BB480 Left=0x005BB420 Parent=0x005BB440 Right=0x005BB560 Color=0 Isnil=0
585
first=107 second=[one hundred seven]
first=101 second=[one hundred one]
ptr=0x005BB560 Left=0x005BB3A0 Parent=0x005BB480 Right=0x005BB580 Color=1 Isnil=0
first=1001 second=[one thousand one]
first=1010 second=[one thousand ten]
As a tree:
root----10 [ten]
L-------1 [one]
L-------0 [zero]
R-------5 [five]
L-------3 [three]
L-------2 [two]
R-------6 [six]
R-------9 [nine]
R-------100 [one hundred]
L-------20 [twenty]
L-------12 [twelve]
L-------11 [eleven]
R-------99 [ninety-nine]
R-------107 [one hundred seven]
L-------101 [one hundred one]
R-------1001 [one thousand one]
R-------1010 [one thousand ten]
m.begin():
ptr=0x005BB4A0 Left=0x005BB3A0 Parent=0x005BB4C0 Right=0x005BB3A0 Color=1 Isnil=0
first=0 second=[zero]
m.end():
ptr=0x005BB3A0 Left=0x005BB4A0 Parent=0x005BB3C0 Right=0x005BB580 Color=1 Isnil=1
dumping s as set:
ptr=0x0020FDFC, Myhead=0x005BB5E0, Mysize=6
ptr=0x005BB5E0 Left=0x005BB640 Parent=0x005BB600 Right=0x005BB6A0 Color=1 Isnil=1
ptr=0x005BB600 Left=0x005BB660 Parent=0x005BB5E0 Right=0x005BB620 Color=1 Isnil=0
first=123
ptr=0x005BB660 Left=0x005BB640 Parent=0x005BB600 Right=0x005BB680 Color=1 Isnil=0
first=12
ptr=0x005BB640 Left=0x005BB5E0 Parent=0x005BB660 Right=0x005BB5E0 Color=0 Isnil=0
first=11
first=100
ptr=0x005BB620 Left=0x005BB5E0 Parent=0x005BB600 Right=0x005BB6A0 Color=1 Isnil=0
first=456
ptr=0x005BB6A0 Left=0x005BB5E0 Parent=0x005BB620 Right=0x005BB5E0 Color=0 Isnil=0
first=1001
As a tree:
root----123
L-------12
L-------11
R-------100
R-------456
R-------1001
s.begin():
first=11
s.end():
ptr=0x005BB5E0 Left=0x005BB640 Parent=0x005BB600 Right=0x005BB6A0 Color=1 Isnil=1
The structure is not packed, so both char values occupy 4 bytes each.
As for std::map, first and second can be viewed as a single value of type std::pair. std::set has
only one value at this address in the structure instead.
The current size of the tree is always present, as in the case of the implementation of std::list in MSVC
( 3.21.4 on page 571).
As in the case of std::list, the iterators are just pointers to nodes. The .begin() iterator points to the
minimal key.
That pointer is not stored anywhere (as in lists), the minimal key of the tree is looked up every time.
586
operator-- and operator++ move the current node pointer to the predecessor or successor respectively,
i.e., the nodes which have the previous or next key.
The algorithms for all these operations are explained in [Cormen, Thomas H. and Leiserson, Charles E.
and Rivest, Ronald L. and Stein, Clifford, Introduction to Algorithms, Third Edition, (2009)].
The .end() iterator points to the dummy node, it has 1 in Isnil, which implies that the node has no
key and/or value. It can be viewed as a “landing zone” in HDD33 and often called sentinel [see N. Wirth,
Algorithms and Data Structures, 1985] 34 .
The “parent” field of the dummy node points to the root node, which serves as a vertex of the tree and
contains information.
GCC
#include <stdio.h>
#include <map>
#include <set>
#include <string>
#include <iostream>
struct map_pair
{
int key;
const char *value;
};
struct tree_node
{
int M_color; // 0 - Red, 1 - Black
struct tree_node *M_parent;
struct tree_node *M_left;
struct tree_node *M_right;
};
struct tree_struct
{
int M_key_compare;
struct tree_node M_header;
size_t M_node_count;
};
void dump_tree_node (struct tree_node *n, bool is_set, bool traverse, bool dump_keys_and_values⤦
Ç )
{
printf ("ptr=0x%p M_left=0x%p M_parent=0x%p M_right=0x%p M_color=%d\n",
n, n->M_left, n->M_parent, n->M_right, n->M_color);
void *point_after_struct=((char*)n)+sizeof(struct tree_node);
if (dump_keys_and_values)
{
if (is_set)
printf ("key=%d\n", *(int*)point_after_struct);
else
{
struct map_pair *p=(struct map_pair *)point_after_struct;
printf ("key=%d value=[%s]\n", p->key, p->value);
};
};
if (traverse==false)
return;
if (n->M_left)
dump_tree_node (n->M_left, is_set, traverse, dump_keys_and_values);
if (n->M_right)
33 Hard Disk Drive
34 http://www.ethoberon.ethz.ch/WirthPubl/AD.pdf
587
dump_tree_node (n->M_right, is_set, traverse, dump_keys_and_values);
};
void dump_as_tree (int tabs, struct tree_node *n, bool is_set)

{
if (is_set)
printf ("%d\n", *(int*)point_after_struct);
else
{
struct map_pair *p=(struct map_pair *)point_after_struct;
printf ("%d [%s]\n", p->key, p->value);
}
if (n->M_left)
{
dump_as_tree (tabs+1, n->M_left, is_set);
};
if (n->M_right)
{
dump_as_tree (tabs+1, n->M_right, is_set);
};
};
void dump_map_and_set(struct tree_struct *m, bool is_set)

{
printf ("ptr=0x%p, M_key_compare=0x%x, M_header=0x%p, M_node_count=%d\n",
m, m->M_key_compare, &m->M_header, m->M_node_count);
dump_tree_node (m->M_header.M_parent, is_set, true, true);
printf ("As a tree:\n");
dump_as_tree (1, m->M_header.M_parent, is_set);
};
int main()
{
// map
std::map<int, const char*> m;
m[10]="ten";
m[20]="twenty";
m[3]="three";
m[101]="one hundred one";
m[100]="one hundred";
m[12]="twelve";
m[107]="one hundred seven";
m[0]="zero";
m[1]="one";
m[6]="six";
m[99]="ninety-nine";
m[5]="five";
m[11]="eleven";
m[1001]="one thousand one";
m[1010]="one thousand ten";
m[2]="two";
m[9]="nine";
printf ("dumping m as map:\n");

dump_map_and_set ((struct tree_struct *)(void*)&m, false);
std::map<int, const char*>::iterator it1=m.begin();

printf ("m.begin():\n");
dump_tree_node ((struct tree_node *)*(void**)&it1, false, false, true);
it1=m.end();
588
printf ("m.end():\n");
dump_tree_node ((struct tree_node *)*(void**)&it1, false, false, false);
// set
std::set<int> s;
s.insert(123);
s.insert(456);
s.insert(11);
s.insert(12);
s.insert(100);
s.insert(1001);
printf ("dumping s as set:\n");
dump_map_and_set ((struct tree_struct *)(void*)&s, true);
std::set<int>::iterator it2=s.begin();
printf ("s.begin():\n");
dump_tree_node ((struct tree_node *)*(void**)&it2, true, false, true);
it2=s.end();
printf ("s.end():\n");
dump_tree_node ((struct tree_node *)*(void**)&it2, true, false, false);
};

dumping m as map:
ptr=0x0028FE3C, M_key_compare=0x402b70, M_header=0x0028FE40, M_node_count=17
ptr=0x007A4988 M_left=0x007A4C00 M_parent=0x0028FE40 M_right=0x007A4B80 M_color=1
key=10 value=[ten]
ptr=0x007A4C00 M_left=0x007A4BE0 M_parent=0x007A4988 M_right=0x007A4C60 M_color=1
key=1 value=[one]
ptr=0x007A4BE0 M_left=0x00000000 M_parent=0x007A4C00 M_right=0x00000000 M_color=1
key=0 value=[zero]
ptr=0x007A4C60 M_left=0x007A4B40 M_parent=0x007A4C00 M_right=0x007A4C20 M_color=0
key=5 value=[five]
ptr=0x007A4B40 M_left=0x007A4CE0 M_parent=0x007A4C60 M_right=0x00000000 M_color=1
key=3 value=[three]
ptr=0x007A4CE0 M_left=0x00000000 M_parent=0x007A4B40 M_right=0x00000000 M_color=0
key=2 value=[two]
ptr=0x007A4C20 M_left=0x00000000 M_parent=0x007A4C60 M_right=0x007A4D00 M_color=1
key=6 value=[six]
ptr=0x007A4D00 M_left=0x00000000 M_parent=0x007A4C20 M_right=0x00000000 M_color=0
key=9 value=[nine]
ptr=0x007A4B80 M_left=0x007A49A8 M_parent=0x007A4988 M_right=0x007A4BC0 M_color=1
key=100 value=[one hundred]
ptr=0x007A49A8 M_left=0x007A4BA0 M_parent=0x007A4B80 M_right=0x007A4C40 M_color=0
key=20 value=[twenty]
ptr=0x007A4BA0 M_left=0x007A4C80 M_parent=0x007A49A8 M_right=0x00000000 M_color=1
key=12 value=[twelve]
ptr=0x007A4C80 M_left=0x00000000 M_parent=0x007A4BA0 M_right=0x00000000 M_color=0
key=11 value=[eleven]
ptr=0x007A4C40 M_left=0x00000000 M_parent=0x007A49A8 M_right=0x00000000 M_color=1
key=99 value=[ninety-nine]
ptr=0x007A4BC0 M_left=0x007A4B60 M_parent=0x007A4B80 M_right=0x007A4CA0 M_color=0
key=107 value=[one hundred seven]
ptr=0x007A4B60 M_left=0x00000000 M_parent=0x007A4BC0 M_right=0x00000000 M_color=1
key=101 value=[one hundred one]
ptr=0x007A4CA0 M_left=0x00000000 M_parent=0x007A4BC0 M_right=0x007A4CC0 M_color=1
key=1001 value=[one thousand one]
ptr=0x007A4CC0 M_left=0x00000000 M_parent=0x007A4CA0 M_right=0x00000000 M_color=0
key=1010 value=[one thousand ten]
As a tree:
root----10 [ten]
L-------1 [one]
L-------0 [zero]
R-------5 [five]
L-------3 [three]
L-------2 [two]
R-------6 [six]
R-------9 [nine]
R-------100 [one hundred]
589
L-------20 [twenty]
L-------12 [twelve]
L-------11 [eleven]
R-------99 [ninety-nine]
R-------107 [one hundred seven]
L-------101 [one hundred one]
R-------1001 [one thousand one]
R-------1010 [one thousand ten]
m.begin():
ptr=0x007A4BE0 M_left=0x00000000 M_parent=0x007A4C00 M_right=0x00000000 M_color=1
key=0 value=[zero]
m.end():
ptr=0x0028FE40 M_left=0x007A4BE0 M_parent=0x007A4988 M_right=0x007A4CC0 M_color=0
dumping s as set:
ptr=0x0028FE20, M_key_compare=0x8, M_header=0x0028FE24, M_node_count=6
ptr=0x007A1E80 M_left=0x01D5D890 M_parent=0x0028FE24 M_right=0x01D5D850 M_color=1
key=123
ptr=0x01D5D890 M_left=0x01D5D870 M_parent=0x007A1E80 M_right=0x01D5D8B0 M_color=1
key=12
ptr=0x01D5D870 M_left=0x00000000 M_parent=0x01D5D890 M_right=0x00000000 M_color=0
key=11
ptr=0x01D5D8B0 M_left=0x00000000 M_parent=0x01D5D890 M_right=0x00000000 M_color=0
key=100
ptr=0x01D5D850 M_left=0x00000000 M_parent=0x007A1E80 M_right=0x01D5D8D0 M_color=1
key=456
ptr=0x01D5D8D0 M_left=0x00000000 M_parent=0x01D5D850 M_right=0x00000000 M_color=0
key=1001
As a tree:
root----123
L-------12
L-------11
R-------100
R-------456
R-------1001
s.begin():
ptr=0x01D5D870 M_left=0x00000000 M_parent=0x01D5D890 M_right=0x00000000 M_color=0
key=11
s.end():
ptr=0x0028FE24 M_left=0x01D5D870 M_parent=0x007A1E80 M_right=0x01D5D8D0 M_color=0
GCC’s implementation is very similar 35 . The only difference is the absence of the Isnil field, so the
structure occupies slightly less space in memory than its implementation in MSVC.
The dummy node is also used as a place to point the .end() iterator also has no key and/or value.
Rebalancing demo (GCC)
Here is also a demo showing us how a tree is rebalanced after some insertions.
Listing 3.120: GCC

#include <stdio.h>
#include <map>
#include <set>
#include <string>
#include <iostream>
struct map_pair
{
int key;
const char *value;
};
struct tree_node
{
int M_color; // 0 - Red, 1 - Black
590
struct tree_node *M_parent;
struct tree_node *M_left;
struct tree_node *M_right;
};
struct tree_struct
{
int M_key_compare;
struct tree_node M_header;
size_t M_node_count;
};
void dump_as_tree (int tabs, struct tree_node *n)

{
printf ("%d\n", *(int*)point_after_struct);
if (n->M_left)
{
dump_as_tree (tabs+1, n->M_left);
};
if (n->M_right)
{
dump_as_tree (tabs+1, n->M_right);
};
};
void dump_map_and_set(struct tree_struct *m)

{
dump_as_tree (1, m->M_header.M_parent);
};
int main()
{
std::set<int> s;
s.insert(123);
s.insert(456);
printf ("123, 456 has been inserted\n");
dump_map_and_set ((struct tree_struct *)(void*)&s);
s.insert(11);
s.insert(12);
printf ("\n");
s.insert(100);
s.insert(1001);
printf ("\n");
s.insert(667);
s.insert(1);
s.insert(4);
s.insert(7);
printf ("\n");
printf ("667, 1, 4, 7 has been inserted\n");
printf ("\n");
};

123, 456 has been inserted
root----123
R-------456
591
root----123
L-------11
R-------12
R-------456

root----123
L-------12
L-------11
R-------100
R-------456
R-------1001
667, 1, 4, 7 has been inserted

root----12
L-------4
L-------1
R-------11
L-------7
R-------123
L-------100
R-------667
L-------456
R-------1001
3.21.5 Memory
Sometimes you may hear from C++ programmers “allocate memory on stack” and/or “allocate memory
on heap”.
Allocating object on stack:
void f()
{
...
Class o=Class(...);
...
};
The memory for object (or structure) is allocated in stack, using simple SP shift. The memory is deallocated
upon function exit, or, more precisely, at the end of scope—SP is returning to its state (same as at the
start of function) and destructor of Class is called. In the same manner, memory for allocated structure
in C is deallocated upon function exit.
Allocating object on heap:
void f1()
{
...
Class *o=new Class(...);
...
};
void f2()
{
...
delete o;
...
};
592
This is the same as allocating memory for a structure using malloc() call. In fact, new in C++ is wrapper
for malloc(), and delete is wrapper for free(). Since memory block has been allocated in heap, it must be
deallocated explicitly, using delete. Class destructor will be automatically called right before that moment.
Which method is better? Allocating on stack is very fast, and good for small, short-lived object, which will
be used only in the current function.
Allocating on heap is slower, and better for long-lived object, which will be used across many functions.
Also, objects allocated in heap are prone to memory leakage, because they must to be freed explicitly,
but one can forget about it.
Anyway, this is matter of taste.
3.22 Negative array indices

It’s possible to address the space before an array by supplying a negative index, e.g., array[−1].
3.22.1 Addressing string from the end

Python PL allows to address arrays and strings from the end. For example, string[-1] returns the last
character, string[-2] returns penultimate, etc. Hard to believe, but this is also possible in C/C++:
#include <string.h>
#include <stdio.h>
int main()
{
char *s_end=s+strlen(s);
printf ("last character: %c\n", s_end[-1]);

printf ("penultimate character: %c\n", s_end[-2]);
};
It works, but s_end must always has an address of terminating zero byte at the end of s string. If s string’s
size get changed, s_end must be updated.
The trick is dubious, but again, this is a demonstration of negative indices.
3.22.2 Addressing some kind of block from the end

Let’s first recall why stack grows backwards ( 1.9.1 on page 30). There is some kind of block in memory
and you want to store both heap and stack there, and you are not sure, how big they both can grow during
runtime.
You can set a heap pointer to the beginning of the block, then you can set a stack pointer to the end of the
block (heap + size_of_block), and then you can address nth element of stack like stack[-n]. For example,
stack[-1] for 1st element, stack[-2] for 2nd, etc.
This will work in the same fashion, as our trick of addressing string from the end.
You can easily check if the structures has not begun to overlap each other: just be sure that address of
the last element in heap is below the address of the last element of stack.
Unfortunately, −0 as index will not work, since two’s complement way of representing negative numbers
( 2.2 on page 448) don’t allow negative zero, so it cannot be distinguished from positive zero.
This method is also mentioned in “Transaction processing”, Jim Gray, 1993, “The Tuple-Oriented File Sys-
tem” chapter, p. 755.
3.22.3 Arrays started at 1

Fortran and Mathematica defined first element of array as 1th, probably because this is tradition in math-
ematics. Other PLs like C/C++ defined it as 0th. Which is better? Edsger W. Dijkstra argued that latter is
better 36 .
36 See https://www.cs.utexas.edu/users/EWD/transcriptions/EWD08xx/EWD831.html
593
But programmers may still have a habit after Fortran, so using this little trick, it’s possible to address the
first element in C/C++ using index 1:
#include <stdio.h>
int main()
{
int random_value=0x11223344;
unsigned char array[10];
int i;
unsigned char *fakearray=&array[-1];
for (i=0; i<10; i++)

array[i]=i;
printf ("first element %d\n", fakearray[1]);

printf ("second element %d\n", fakearray[2]);
printf ("last element %d\n", fakearray[10]);
printf ("array[-1]=%02X, array[-2]=%02X, array[-3]=%02X, array[-4]=%02X\n",

array[-1],
array[-2],
array[-3],
array[-4]);
};

1 $SG2751 DB 'first element %d', 0aH, 00H
2 $SG2752 DB 'second element %d', 0aH, 00H
3 $SG2753 DB 'last element %d', 0aH, 00H
4 $SG2754 DB 'array[-1]=%02X, array[-2]=%02X, array[-3]=%02X, array[-4'
5 DB ']=%02X', 0aH, 00H
6
7 _fakearray$ = -24 ; size = 4
8 _random_value$ = -20 ; size = 4
9 _array$ = -16 ; size = 10
10 _i$ = -4 ; size = 4
11 _main PROC
12 push ebp
13 mov ebp, esp
14 sub esp, 24
15 mov DWORD PTR _random_value$[ebp], 287454020 ; 11223344H
16 ; set fakearray[] one byte earlier before array[]
17 lea eax, DWORD PTR _array$[ebp]
18 add eax, -1 ; eax=eax-1
19 mov DWORD PTR _fakearray$[ebp], eax
20 mov DWORD PTR _i$[ebp], 0
21 jmp SHORT $LN3@main
22 ; fill array[] with 0..9
23 $LN2@main:
24 mov ecx, DWORD PTR _i$[ebp]
25 add ecx, 1
26 mov DWORD PTR _i$[ebp], ecx
27 $LN3@main:
28 cmp DWORD PTR _i$[ebp], 10
29 jge SHORT $LN1@main
30 mov edx, DWORD PTR _i$[ebp]
31 mov al, BYTE PTR _i$[ebp]
32 mov BYTE PTR _array$[ebp+edx], al
33 jmp SHORT $LN2@main
34 $LN1@main:
35 mov ecx, DWORD PTR _fakearray$[ebp]
36 ; ecx=address of fakearray[0], ecx+1 is fakearray[1] or array[0]
37 movzx edx, BYTE PTR [ecx+1]
38 push edx
39 push OFFSET $SG2751 ; 'first element %d'
40 call _printf
41 add esp, 8
42 mov eax, DWORD PTR _fakearray$[ebp]
594
43 ; eax=address of fakearray[0], eax+2 is fakearray[2] or array[1]
44 movzx ecx, BYTE PTR [eax+2]
45 push ecx
46 push OFFSET $SG2752 ; 'second element %d'
47 call _printf
48 add esp, 8
49 mov edx, DWORD PTR _fakearray$[ebp]
50 ; edx=address of fakearray[0], edx+10 is fakearray[10] or array[9]
51 movzx eax, BYTE PTR [edx+10]
52 push eax
53 push OFFSET $SG2753 ; 'last element %d'
54 call _printf
55 add esp, 8
56 ; subtract 4, 3, 2 and 1 from pointer to array[0] in order to find values before array[]
57 lea ecx, DWORD PTR _array$[ebp]
58 movzx edx, BYTE PTR [ecx-4]
59 push edx
60 lea eax, DWORD PTR _array$[ebp]
61 movzx ecx, BYTE PTR [eax-3]
62 push ecx
63 lea edx, DWORD PTR _array$[ebp]
64 movzx eax, BYTE PTR [edx-2]
65 push eax
66 lea ecx, DWORD PTR _array$[ebp]
67 movzx edx, BYTE PTR [ecx-1]
68 push edx
69 push OFFSET $SG2754 ;
'array[-1]=%02X, array[-2]=%02X, array[-3]=%02X, array[-4]=%02X'
70 call _printf
71 add esp, 20
72 xor eax, eax
73 mov esp, ebp
74 pop ebp
75 ret 0
76 _main ENDP
So we have array[] of ten elements, filled with 0 . . . 9 bytes.

Then we have the fakearray[] pointer, which points one byte before array[].
fakearray[1] points exactly to array[0]. But we are still curious, what is there before array[]? We have
added random_value before array[] and set it to 0x11223344. The non-optimizing compiler allocated
the variables in the order they were declared, so yes, the 32-bit random_value is right before the array.
We ran it, and:
first element 0
second element 1
last element 9
array[-1]=11, array[-2]=22, array[-3]=33, array[-4]=44
Here is the stack fragment we will copypaste from OllyDbg’s stack window (with comments added by the
author):
CPU Stack
Address Value
001DFBCC /001DFBD3 ; fakearray pointer
001DFBD0 |11223344 ; random_value
001DFBD4 |03020100 ; 4 bytes of array[]
001DFBD8 |07060504 ; 4 bytes of array[]
001DFBDC |00CB0908 ; random garbage + 2 last bytes of array[]
001DFBE0 |0000000A ; last i value after loop was finished
001DFBE4 |001DFC2C ; saved EBP value
001DFBE8 \00CB129D ; Return Address
The pointer to the fakearray[] (0x001DFBD3) is indeed the address of array[] in the stack (0x001DFBD4),
but minus 1 byte.
It’s still very hackish and dubious trick. Doubtfully anyone should use it in production code, but as a
demonstration, it fits perfectly here.
595
3.23 More about pointers
The way C handles pointers, for example,

was a brilliant innovation; it solved a lot of
problems that we had before in data
structuring and made the programs look
good afterwards.
Donald Knuth, interview (1993)
For those, who still have hard time understanding C/C++ pointers, here are more examples. Some of
them are weird and serves only demonstration purpose: use them in production code only if you really
know what you’re doing.
3.23.1 Working with addresses instead of pointers

Pointer is just an address in memory. But why we write char* string instead of something like address
string? Pointer variable is supplied with a type of the value to which pointer points. So then compiler
will be able to catch data typization bugs during compilation.
To be pedantic, data typing in programming languages is all about preventing bugs and self-documentation.
It’s possible to use maybe two of data types like int (or int64_t) and byte—these are the only types which
are available to assembly language programmers. But it’s just very hard task to write big and practical
assembly programs without nasty bugs. Any small typo can lead to hard-to-find bug.
Data type information is absent in a compiled code (and this is one of the main problems for decompilers),
and I can demonstrate this.
This is what sane C/C++ programmer can write:
#include <stdio.h>
#include <stdint.h>
void print_string (char *s)

{
printf ("(address: 0x%llx)\n", s);
printf ("%s\n", s);
};
int main()
{
print_string (s);
};
This is what I can write:

#include <stdio.h>
#include <stdint.h>
void print_string (uint64_t address)

{
printf ("(address: 0x%llx)\n", address);
puts ((char*)address);
};
int main()
{
print_string ((uint64_t)s);
};
I use uint64_t because I run this example on Linux x64. int would work for 32-bit OS-es. First, a pointer to
character (the very first in the greeting string) is casted to uint64_t, then it’s passed further. print_string()
function casts back incoming uint64_t value into pointer to a character.
What is interesting is that GCC 4.8.4 produces identical assembly output for both versions:
596
gcc 1.c -S -masm=intel -O3 -fno-inline
.LC0:
.string "(address: 0x%llx)\n"
print_string:
push rbx
mov rdx, rdi
mov rbx, rdi
mov edi, 1
xor eax, eax
call __printf_chk
mov rdi, rbx
pop rbx
jmp puts
.LC1:
main:
sub rsp, 8
call print_string
add rsp, 8
ret
(I’ve removed all insignificant GCC directives.)

I also tried UNIX diff utility and it shows no differences at all.
Let’s continue to abuse C/C++ programming traditions heavily. Someone may write this:
#include <stdio.h>
#include <stdint.h>
uint8_t load_byte_at_address (uint8_t* address)

{
return *address;
//this is also possible: return address[0];
};
void print_string (char *s)

{
char* current_address=s;
while (1)
{
char current_char=load_byte_at_address(current_address);
if (current_char==0)
break;
printf ("%c", current_char);
current_address++;
};
};
int main()
{
print_string (s);
};
It can be rewritten like this:

#include <stdio.h>
#include <stdint.h>
uint8_t load_byte_at_address (uint64_t address)

{
return *(uint8_t*)address;
};
void print_string (uint64_t address)
597
{
uint64_t current_address=address;
while (1)
{
char current_char=load_byte_at_address(current_address);
if (current_char==0)
break;
printf ("%c", current_char);
current_address++;
};
};
int main()
{
print_string ((uint64_t)s);
};
Both source codes resulting in the same assembly output:

gcc 1.c -S -masm=intel -O3 -fno-inline
load_byte_at_address:
movzx eax, BYTE PTR [rdi]
ret
print_string:
.LFB15:
push rbx
mov rbx, rdi
jmp .L4
.L7:
movsx edi, al
add rbx, 1
call putchar
.L4:
mov rdi, rbx
call load_byte_at_address
test al, al
jne .L7
pop rbx
ret
.LC0:
main:
sub rsp, 8
call print_string
add rsp, 8
ret
(I have also removed all insignificant GCC directives.)

No difference: C/C++ pointers are essentially addresses, but supplied with type information, in order to
prevent possible mistakes at the time of compilation. Types are not checked during runtime—it would be
huge (and unneeded) overhead.
3.23.2 Passing values as pointers; tagged unions

Here is an example on how to pass values in pointers:
#include <stdio.h>
#include <stdint.h>
uint64_t multiply1 (uint64_t a, uint64_t b)

{
return a*b;
};
598
uint64_t* multiply2 (uint64_t *a, uint64_t *b)
{
return (uint64_t*)((uint64_t)a*(uint64_t)b);
};
int main()
{
printf ("%d\n", multiply1(123, 456));
printf ("%d\n", (uint64_t)multiply2((uint64_t*)123, (uint64_t*)456));
};
It works smoothly and GCC 4.8.4 compiles both multiply1() and multiply2() functions identically!
multiply1:
mov rax, rdi
imul rax, rsi
ret
multiply2:
mov rax, rdi
imul rax, rsi
ret
As long as you do not dereference pointer (in other words, you don’t read any data from the address stored
in pointer), everything will work fine. Pointer is a variable which can store anything, like usual variable.
Signed multiplication instruction (IMUL) is used here instead of unsigned one (MUL), read more about it
here: 2.2.1 on page 450.
By the way, it’s well-known hack to abuse pointers a little called tagged pointers. In short, if all your
pointers points to blocks of memory with size of, let’s say, 16 bytes (or it is always aligned on 16-byte
boundary), 4 lowest bits of pointer is always zero bits and this space can be used somehow. It’s very
popular in LISP compilers and interpreters. They store cell/object type in these unused bits, this can save
some memory. Even more, you can judge about cell/object type using just pointer, with no additional
memory access. Read more about it: [Dennis Yurichev, C/C++ programming language notes1.3].
3.23.3 Pointers abuse in Windows kernel

The resource section of PE executable file in Windows OS is a section containing pictures, icons, strings,
etc. Early Windows versions allowed to address resources only by IDs, but then Microsoft added a way to
address them using strings.
So then it would be possible to pass ID or string to FindResource() function. Which is declared like this:
HRSRC WINAPI FindResource(
_In_opt_ HMODULE hModule,
_In_ LPCTSTR lpName,
_In_ LPCTSTR lpType
);
lpName and lpType has char* or wchar* types, and when someone still wants to pass ID, he/she have to
use MAKEINTRESOURCE macro, like this:
result = FindResource(..., MAKEINTRESOURCE(1234), ...);
It’s interesting fact that MAKEINTRESOURCE is merely casting integer to pointer. In MSVC 2013, in the file
Microsoft SDKs\Windows\v7.1A\Include\Ks.h we can find this:
...
#if (!defined( MAKEINTRESOURCE ))

#define MAKEINTRESOURCE( res ) ((ULONG_PTR) (USHORT) res)
#endif
...
Sounds insane. Let’s peek into ancient leaked Windows NT4 source code. In private/windows/base/clien-
t/module.c we can find FindResource() source code:
599
HRSRC
FindResourceA(
HMODULE hModule,
LPCSTR lpName,
LPCSTR lpType
)
...
{
NTSTATUS Status;
ULONG IdPath[ 3 ];
PVOID p;
IdPath[ 0 ] = 0;
IdPath[ 1 ] = 0;
try {
if ((IdPath[ 0 ] = BaseDllMapResourceIdA( lpType )) == -1) {
Status = STATUS_INVALID_PARAMETER;
}
else
if ((IdPath[ 1 ] = BaseDllMapResourceIdA( lpName )) == -1) {
...
Let’s proceed to BaseDllMapResourceIdA() in the same source file:

ULONG
BaseDllMapResourceIdA(
LPCSTR lpId
)
{
NTSTATUS Status;
ULONG Id;
UNICODE_STRING UnicodeString;
ANSI_STRING AnsiString;
PWSTR s;
try {
if ((ULONG)lpId & LDR_RESOURCE_ID_NAME_MASK) {
if (*lpId == '#') {
Status = RtlCharToInteger( lpId+1, 10, &Id );
if (!NT_SUCCESS( Status ) || Id & LDR_RESOURCE_ID_NAME_MASK) {
if (NT_SUCCESS( Status )) {
}
BaseSetLastNTError( Status );
Id = (ULONG)-1;
}
}
else {
RtlInitAnsiString( &AnsiString, lpId );
Status = RtlAnsiStringToUnicodeString( &UnicodeString,
&AnsiString,
TRUE
);
if (!NT_SUCCESS( Status )){
BaseSetLastNTError( Status );
Id = (ULONG)-1;
}
else {
s = UnicodeString.Buffer;
while (*s != UNICODE_NULL) {
*s = RtlUpcaseUnicodeChar( *s );
s++;
}
Id = (ULONG)UnicodeString.Buffer;
}
600
}
}
else {
Id = (ULONG)lpId;
}
}
except (EXCEPTION_EXECUTE_HANDLER) {
BaseSetLastNTError( GetExceptionCode() );
Id = (ULONG)-1;
}
return Id;
}
lpId is ANDed with LDR_RESOURCE_ID_NAME_MASK.

Which we can find in public/sdk/inc/ntldr.h:
...
#define LDR_RESOURCE_ID_NAME_MASK 0xFFFF0000
...
So lpId is ANDed with 0xFFFF0000 and if some bits beyond lowest 16 bits are still present, first half of
function is executed (lpId is treated as an address of string). Otherwise—second half (lpId is treated as
16-bit value).
Still, this code can be found in Windows 7 kernel32.dll file:
....
.text:0000000078D24510 ; __int64 __fastcall BaseDllMapResourceIdA(PCSZ SourceString)

.text:0000000078D24510 BaseDllMapResourceIdA proc near ; CODE XREF: FindResourceExA+34
.text:0000000078D24510 ; FindResourceExA+4B
.text:0000000078D24510
.text:0000000078D24510 var_38 = qword ptr -38h
.text:0000000078D24510 var_30 = qword ptr -30h
.text:0000000078D24510 var_28 = _UNICODE_STRING ptr -28h
.text:0000000078D24510 DestinationString= _STRING ptr -18h
.text:0000000078D24510 arg_8 = dword ptr 10h
.text:0000000078D24510
.text:0000000078D24510 ; FUNCTION CHUNK AT .text:0000000078D42FB4 SIZE 000000D5 BYTES
.text:0000000078D24510
.text:0000000078D24510 push rbx
.text:0000000078D24512 sub rsp, 50h
.text:0000000078D24516 cmp rcx, 10000h
.text:0000000078D2451D jnb loc_78D42FB4
.text:0000000078D24523 mov [rsp+58h+var_38], rcx
.text:0000000078D24528 jmp short $+2
.text:0000000078D2452A ;
---------------------------------------------------------------------------
.text:0000000078D2452A
.text:0000000078D2452A loc_78D2452A: ; CODE XREF:
BaseDllMapResourceIdA+18
.text:0000000078D2452A ; BaseDllMapResourceIdA+1EAD0
.text:0000000078D2452A jmp short $+2
.text:0000000078D2452C ;
---------------------------------------------------------------------------
.text:0000000078D2452C
.text:0000000078D2452C loc_78D2452C: ;
CODE XREF: BaseDllMapResourceIdA:loc_78D2452A
.text:0000000078D2452C ; BaseDllMapResourceIdA+1EB74
.text:0000000078D2452C mov rax, rcx
.text:0000000078D2452F add rsp, 50h
.text:0000000078D24533 pop rbx
.text:0000000078D24534 retn
.text:0000000078D24534 ;
---------------------------------------------------------------------------
.text:0000000078D24535 align 20h
.text:0000000078D24535 BaseDllMapResourceIdA endp
....
601
.text:0000000078D42FB4 loc_78D42FB4: ; CODE XREF:
BaseDllMapResourceIdA+D
.text:0000000078D42FB4 cmp byte ptr [rcx], '#'
.text:0000000078D42FB7 jnz short loc_78D43005
.text:0000000078D42FB9 inc rcx
.text:0000000078D42FBC lea r8, [rsp+58h+arg_8]
.text:0000000078D42FC1 mov edx, 0Ah
.text:0000000078D42FC6 call cs:__imp_RtlCharToInteger
.text:0000000078D42FCC mov ecx, [rsp+58h+arg_8]
.text:0000000078D42FD0 mov [rsp+58h+var_38], rcx
.text:0000000078D42FD5 test eax, eax
.text:0000000078D42FD7 js short loc_78D42FE6
.text:0000000078D42FD9 test rcx, 0FFFFFFFFFFFF0000h
.text:0000000078D42FE0 jz loc_78D2452A
....
If value in input pointer is greater than 0x10000, jump to string processing is occurred. Otherwise, input
value of lpId is returned as is. 0xFFFF0000 mask is not used here any more, because this is 64-bit code
after all, but still, 0xFFFFFFFFFFFF0000 could work here.
Attentive reader may ask, what if address of input string is lower than 0x10000? This code relied on the
fact that in Windows there are nothing on addresses below 0x10000, at least in Win32 realm.
Raymond Chen writes about this:
How does MAKEINTRESOURCE work? It just stashes the integer in the bottom 16 bits of
a pointer, leaving the upper bits zero. This relies on the convention that the first 64KB of
address space is never mapped to valid memory, a convention that is enforced starting in
Windows 7.
In short words, this is dirty hack and probably one should use it only if there is a real necessity. Perhaps,
FindResource() function in past had SHORT type for its arguments, and then Microsoft has added a way
to pass strings there, but older code must also be supported.
Now here is my short distilled example:
#include <stdio.h>
#include <stdint.h>
void f(char* a)
{
if (((uint64_t)a)>0x10000)
printf ("Pointer to string has been passed: %s\n", a);
else
printf ("16-bit value has been passed: %d\n", (uint64_t)a);
};
int main()
{
f("Hello!"); // pass string
f((char*)1234); // pass 16-bit value
};
It works!
Pointers abuse in Linux kernel
As it has been noted in comments on Hacker News, Linux kernel also has something like that.
For example, this function can return both error code and pointer:
struct kernfs_node *kernfs_create_link(struct kernfs_node *parent,
const char *name,
struct kernfs_node *target)
{
struct kernfs_node *kn;
int error;
602
kn = kernfs_new_node(parent, name, S_IFLNK|S_IRWXUGO, KERNFS_LINK);
if (!kn)
return ERR_PTR(-ENOMEM);
if (kernfs_ns_enabled(parent))
kn->ns = target->ns;
kn->symlink.target_kn = target;
kernfs_get(target); /* ref owned by symlink */
error = kernfs_add_one(kn);
if (!error)
return kn;
kernfs_put(kn);
return ERR_PTR(error);
}
( https://github.com/torvalds/linux/blob/fceef393a538134f03b778c5d2519e670269342f/fs/kernfs/
symlink.c#L25 )
ERR_PTR is a macro to cast integer to pointer:
static inline void * __must_check ERR_PTR(long error)
{
return (void *) error;
}
( https://github.com/torvalds/linux/blob/61d0b5a4b2777dcf5daef245e212b3c1fa8091ca/tools/
virtio/linux/err.h )
This header file also has a macro helper to distinguish error code from pointer:
#define IS_ERR_VALUE(x) unlikely((x) >= (unsigned long)-MAX_ERRNO)
This means, error codes are the “pointers” which are very close to -1 and, hopefully, there are nothing in
kernel memory on the addresses like 0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFE, 0xFFFFFFFFFFFFFFFD,
etc.
Much more popular solution is to return NULL in case of error and to pass error code via additional ar-
gument. Linux kernel authors don’t do that, but everyone who use these functions must always keep in
mind that returning pointer must always be checked with IS_ERR_VALUE before dereferencing.
For example:
fman->cam_offset = fman_muram_alloc(fman->muram, fman->cam_size);
if (IS_ERR_VALUE(fman->cam_offset)) {
dev_err(fman->dev, "%s: MURAM alloc for DMA CAM failed\n",
__func__);
return -ENOMEM;
}
( https://github.com/torvalds/linux/blob/aa00edc1287a693eadc7bc67a3d73555d969b35d/drivers/
net/ethernet/freescale/fman/fman.c#L826 )
Pointers abuse in UNIX userland
mmap() function returns -1 in case of error (or MAP_FAILED, which equals to -1). Some people say, mmap()
can map a memory at zeroth address in rare situations, so it can’t use 0 or NULL as error code.
3.23.4 Null pointers

“Null pointer assignment” error of MS-DOS era
Oldschool readers may recall a weird error message of MS-DOS era: “Null pointer assignment”. What
does it mean?
It’s not possible to write a memory at zero address in *NIX and Windows OSes, but it was possible to do
so in MS-DOS due to absence of memory protection whatsoever.
603
So I’ve pulled my ancient Turbo C++ 3.0 (later it was renamed to Borland C++) from early 1990s and
tried to compile this:
#include <stdio.h>
int main()
{
int *ptr=NULL;
*ptr=1234;
printf ("Now let's read at NULL\n");
printf ("%d\n", *ptr);
};
Hard to believe, but it works, with error upon exit, though:
Listing 3.124: Ancient Turbo C 3.0

C:\TC30\BIN\1
Now let's read at NULL
1234
Null pointer assignment
C:\TC30\BIN>_
Let’s dig deeper into the source code of CRT of Borland C++ 3.1, file c0.asm:
; _checknull() check for null pointer zapping copyright message
...
; Check for null pointers before exit
__checknull PROC DIST

PUBLIC __checknull
IF LDATA EQ false
IFNDEF__TINY__
push si
push di
mov es, cs:DGROUP@@
xor ax, ax
mov si, ax
mov cx, lgth_CopyRight
ComputeChecksum label near
add al, es:[si]
adc ah, 0
inc si
loop ComputeChecksum
sub ax, CheckSum
jz @@SumOK
mov cx, lgth_NullCheck
mov dx, offset DGROUP: NullCheck
call ErrorDisplay
@@SumOK: pop di
pop si
ENDIF
ENDIF
_DATA SEGMENT
; Magic symbol used by the debug info to locate the data segment
public DATASEG@
DATASEG@ label byte
; The CopyRight string must NOT be moved or changed without

; changing the null pointer check logic
CopyRight db 4 dup(0)
db 'Borland C++ - Copyright 1991 Borland Intl.',0
lgth_CopyRight equ $ - CopyRight
604
IF LDATA EQ false
IFNDEF __TINY__
CheckSum equ 00D5Ch
NullCheck db 'Null pointer assignment', 13, 10
lgth_NullCheck equ $ - NullCheck
ENDIF
ENDIF
...
The MS-DOS memory model was really weird ( 11.6 on page 986) and probably not worth looking into it
unless you’re fan of retrocomputing or retrogaming. One thing we have to keep in mind is that memory
segment (included data segment) in MS-DOS is a memory segment in which code or data is stored, but
unlike “serious” OSes, it’s started at address 0.
And in Borland C++ CRT, the data segment is started with 4 zero bytes and the copyright string “Borland
C++ - Copyright 1991 Borland Intl.”. The integrity of the 4 zero bytes and text string is checked upon exit,
and if it’s corrupted, the error message is displayed.
But why? Writing at null pointer is common mistake in C/C++, and if you do so in *NIX or Windows, your
application will crash. MS-DOS has no memory protection, so CRT has to check this post-factum and warn
about it upon exit. If you see this message, this means, your program at some point has written at address
0.
Our program did so. And this is why 1234 number has been read correctly: because it was written at the
place of the first 4 zero bytes. Checksum is incorrect upon exit (because the number has been left there),
so error message has been displayed.
Am I right? I’ve rewritten the program to check my assumptions:
#include <stdio.h>
int main()
{
int *ptr=NULL;
*ptr=1234;
printf ("Now let's read at NULL\n");
printf ("%d\n", *ptr);
*ptr=0; // psst, cover our tracks!
};
This program executes without error message upon exit.

Though method to warn about null pointer assignment is relevant for MS-DOS, perhaps, it can still be used
today in low-cost MCUs with no memory protection and/or MMU37 .
Why would anyone write at address 0?
But why would sane programmer write a code which writes something at address 0? It can be done
accidentally: for example, a pointer must be initialized to newly allocated memory block and then passed
to some function which returns data through pointer.
int *ptr=NULL;
... we forgot to allocate memory and initialize ptr
strcpy (ptr, buf); // strcpy() terminates silently because MS-DOS has no memory protection
Even worse:
int *ptr=malloc(1000);
... we forgot to check if memory has been really allocated: this is MS-DOS after all and ⤦
Ç computers had small amount of RAM,
... and RAM shortage was very common.
... if malloc() returned NULL, the ptr will also be NULL.
strcpy (ptr, buf); // strcpy() terminates silently because MS-DOS has no memory protection
37 Memory Management Unit
605
Writing on 0th address on purpose
Here is an example from dmalloc38 , a portable way of generating core dump, if other ways are not avail-
able:
3.4 Generating a Core File on Errors
====================================
If the èrror-abort' debug token has been enabled, when the library
detects any problems with the heap memory, it will immediately attempt
to dump a core file. *Note Debug Tokens::. Core files are a complete
copy of the program and it's state and can be used by a debugger to see
specifically what is going on when the error occurred. *Note Using
With a Debugger::. By default, the low, medium, and high arguments to
the library utility enable the èrror-abort' token. You can disable
this feature by entering `dmalloc -m error-abort' (-m for minus) to
remove the èrror-abort' token and your program will just log errors
and continue. You can also use the èrror-dump' token which tries to
dump core when it sees an error but still continue running. *Note
Debug Tokens::.
When a program dumps core, the system writes the program and all of
its memory to a file on disk usually named `core'. If your program is
called `foo' then your system may dump core as `foo.core'. If you are
not getting a `core' file, make sure that your program has not changed
to a new directory meaning that it may have written the core file in a
different location. Also insure that your program has write privileges
over the directory that it is in otherwise it will not be able to dump
a core file. Core dumps are often security problems since they contain
all program memory so systems often block their being produced. You
will want to check your user and system's core dump size ulimit
settings.
The library by default uses the àbort' function to dump core which
may or may not work depending on your operating system. If the
following program does not dump core then this may be the problem. See
`KILL_PROCESS' definition in `settings.dist'.
main()
{
abort();
}
If àbort' does work then you may want to try the following setting
in `settings.dist'. This code tries to generate a segmentation fault
by dereferencing a `NULL' pointer.
#define KILL_PROCESS { int *_int_p = 0L; *_int_p = 1; }
NULL in C/C++
NULL in C/C++ is just a macro which is often defined like this:

#define NULL ((void*)0)
( libio.h file )
void* is a data type reflecting the fact it’s the pointer, but to a value of unknown data type (void).
NULL is usually used to show absence of an object. For example, you have a single-linked list, and each
node has a value (or pointer to a value) and next pointer. To show that there are no next node, 0 is stored
to next field. (Other solutions are just worse.) Perhaps, you may have some crazy environment where
you need to allocate memory blocks at zero address. How would you indicate absence of the next node?
Some kind of magic number? Maybe -1? Or maybe using additional bit?
In Wikipedia we may find this:
38 http://dmalloc.com/
606
In fact, quite contrary to the zero page’s original preferential use, some modern operating
systems such as FreeBSD, Linux and Microsoft Windows[2] actually make the zero page
inaccessible to trap uses of NULL pointers.
( https://en.wikipedia.org/wiki/Zero_page )
Null pointer to function
It’s possible to call function by its address. For example, I compile this by MSVC 2010 and run it in Windows
7:
#include <stdio.h>
int main()
{
printf ("0x%x\n", &MessageBoxA);
};
The result is 0x7578feae and doesn’t changing after several times I run it, because user32.dll (where
MessageBoxA function resides) is always loads at the same address. And also because ASLR39 is not
enabled (result would be different each time in that case).
Let’s call MessageBoxA() by address:
#include <stdio.h>
typedef int (*msgboxtype)(HWND hWnd, LPCTSTR lpText, LPCTSTR lpCaption, UINT uType);
int main()
{
msgboxtype msgboxaddr=0x7578feae;
// force to load DLL into process memory,

// since our code doesn't use any function from user32.dll,
// and DLL is not imported
LoadLibrary ("user32.dll");
msgboxaddr(NULL, "Hello, world!", "hello", MB_OK);

};
Weird, but works in Windows 7 x86.

This is commonly used in shellcodes, because it’s hard to call DLL functions by name from there. And
ASLR is a countermeasure.
Now what is really weird, some embedded C programmers may be familiar with a code like that:
int reset()
{
void (*foo)(void) = 0;
foo();
};
Who will want to call a function at address 0? This is portable way to jump at zero address. Many low-cost
cheap microcontrollers also have no memory protection or MMU and after reset, they start to execute
code at address 0, where some kind of initialization code is stored. So jumping to address 0 is a way to
reset itself. One could use inline assembly, but if it’s not possible, this portable method can be used.
It even compiles correctly by my GCC 4.8.4 on Linux x64:
reset:
sub rsp, 8
xor eax, eax
call rax
39 Address Space Layout Randomization
607
add rsp, 8
ret
The fact that stack pointer is shifted is not a problem: initialization code in microcontrollers usually com-
pletely ignores registers and RAM state and boots from scratch.
And of course, this code will crash on *NIX or Windows because of memory protection and even in absence
of protection, there are no code at address 0.
GCC even has non-standard extension, allowing to jump to a specific address rather than call a function
there: http://gcc.gnu.org/onlinedocs/gcc/Labels-as-Values.html.
3.23.5 Array as function argument

Someone may ask, what is the difference between declaring function argument type as array and as
pointer?
As it seems, there are no difference at all:
void write_something1(int a[16])
{
a[5]=0;
};
void write_something2(int *a)

{
a[5]=0;
};
int f()
{
int a[16];
write_something1(a);
write_something2(a);
};
Optimizing GCC 4.8.4:

write_something1:
ret
write_something2:
ret
But you may still declare array instead of pointer for self-documenting purposes, if the size of array is
always fixed. And maybe, some static analysis tool will be able to warn you about possible buffer overflow.
Or is it possible with some tools today?
Some people, including Linus Torvalds, criticizes this C/C++ feature: https://lkml.org/lkml/2015/9/
3/428.
C99 standard also have static keyword [ISO/IEC 9899:TC3 (C C99 standard), (2007) 6.7.5.3]:
If the keyword static also appears within the [ and ] of the array type derivation, then
for each call to the function, the value of the corresponding actual argument shall provide
access to the first element of an array with at least as many elements as specified by the
size expression.
3.23.6 Pointer to a function

A function name in C/C++ without brackets, like “printf” is a pointer to function of void (*)() type. Let’s
try to read function’s contents and patch it:
#include <memory.h>
#include <stdio.h>
608
void print_something ()
{
printf ("we are in %s()\n", __FUNCTION__);
};
int main()
{
print_something();
printf ("first 3 bytes: %x %x %x...\n",
*(unsigned char*)print_something,
*((unsigned char*)print_something+1),
*((unsigned char*)print_something+2));
*(unsigned char*)print_something=0xC3; // RET's opcode

printf ("going to call patched print_something():\n");
print_something();
printf ("it must exit at this point\n");
};
It tells, that the first 3 bytes of functions are 55 89 e5. Indeed, these are opcodes of PUSH EBP and MOV
EBP, ESP instructions (these are x86 opcodes). But then our program crashes, because text section is
readonly.
40
We can recompile our example and make text section writable :
gcc --static -g -Wl,--omagic -o example example.c
That works!
we are in print_something()
first 3 bytes: 55 89 e5...
going to call patched print_something():
it must exit at this point
3.23.7 Pointer to a function: copy protection

A software cracker can find a function that checks protection and return true or false. He/she then can
put XOR EAX,EAX / RETN or MOV EAX, 1 / RETN there.
Can you check integrity of the function? As it turns out, this can be done easily.
According to objdump, the first 3 bytes of check_protection() are 0x55 0x89 0xE5 (given the fact this
is non-optimizing GCC):
#include <stdlib.h>
#include <stdio.h>
int check_protection()
{
// do something
return 0;
// or return 1;
};
int main()
{
if (check_protection()==0)
{
printf ("no protection installed\n");
exit(0);
};
// ...and then, at some very important point...

if (*(((unsigned char*)check_protection)+0) != 0x55)
{
printf ("1st byte has been altered\n");
40 http://stackoverflow.com/questions/27581279/make-text-segment-writable-elf
609
// do something mean, add watermark, etc
};
if (*(((unsigned char*)check_protection)+1) != 0x89)
{
printf ("2nd byte has been altered\n");
};
if (*(((unsigned char*)check_protection)+2) != 0xe5)
{
printf ("3rd byte has been altered\n");
};
};
0000054d <check_protection>:
54d: 55 push %ebp
54e: 89 e5 mov %esp,%ebp
550: e8 b7 00 00 00 call 60c <__x86.get_pc_thunk.ax>
555: 05 7f 1a 00 00 add $0x1a7f,%eax
55a: b8 00 00 00 00 mov $0x0,%eax
55f: 5d pop %ebp
560: c3 ret
If someone would patch the beginning of the check_protection() function, your program can do some-
thing mean, maybe exit suddenly. To find such a trick, a cracker can set a memory read breakpoint on
the address of the function’s beginning. (tracer has BPMx options for that.)
3.23.8 Pointer to a function: a common bug (or typo)

A notorious bug/typo:
int expired()
{
// check license key, current date/time, etc
};
int main()
{
if (expired) // must be expired() here
{
print ("expired\n");
exit(0);
}
else
{
// do something
};
};
Since the function’s name alone is interpreted as a pointer to function, or address, the if(function_name)
statement is like if(true).
Unfortunately, a C/C++ compiler wouldn’t issue a warning.
3.23.9 Pointer as object identificator

Both assembly language and C has no OOP features, but it’s possible to write a code in OOP style (just
treat structure as an object).
It’s interesting, that sometimes, pointer to an object (or its address) is called as ID (in sense of data
hiding/encapsulation).
For example, LoadLibrary(), according to MSDN41 , returns “handle to the module” 42 . Then you pass this
“handle” to other functions like GetProcAddress(). But in fact, LoadLibrary() returns pointer to DLL file
41 Microsoft Developer Network
42 https://msdn.microsoft.com/ru-ru/library/windows/desktop/ms684175(v=vs.85).aspx
610
mapped into memory 43 . You can read two bytes from the address LoadLibrary() returns, and that would
be “MZ” (first two bytes of any .EXE/.DLL file in Windows).
Apparently, Microsoft “hides” that fact to provide better forward compatibility. Also, HMODULE and HIN-
STANCE data types had another meaning in 16-bit Windows.
Probably, this is reason why printf() has “%p” modifier, which is used for printing pointers (32-bit inte-
gers on 32-bit architectures, 64-bit on 64-bit, etc) in hexadecimal form. Address of a structure dumped
into debug log may help in finding it in another place of log.
Here is also from SQLite source code:
...
struct Pager {
sqlite3_vfs *pVfs; /* OS functions to use for IO */
u8 exclusiveMode; /* Boolean. True if locking_mode==EXCLUSIVE */
u8 journalMode; /* One of the PAGER_JOURNALMODE_* values */
u8 useJournal; /* Use a rollback journal on this file */
u8 noSync; /* Do not sync the journal if true */
....
static int pagerLockDb(Pager *pPager, int eLock){

int rc = SQLITE_OK;
assert( eLock==SHARED_LOCK || eLock==RESERVED_LOCK || eLock==EXCLUSIVE_LOCK );

if( pPager->eLock<eLock || pPager->eLock==UNKNOWN_LOCK ){
rc = sqlite3OsLock(pPager->fd, eLock);
if( rc==SQLITE_OK && (pPager->eLock!=UNKNOWN_LOCK||eLock==EXCLUSIVE_LOCK) ){
pPager->eLock = (u8)eLock;
IOTRACE(("LOCK %p %d\n", pPager, eLock))
}
}
return rc;
}
...
PAGER_INCR(sqlite3_pager_readdb_count);
PAGER_INCR(pPager->nRead);
IOTRACE(("PGIN %p %d\n", pPager, pgno));
PAGERTRACE(("FETCH %d page %d hash(%08x)\n",
PAGERID(pPager), pgno, pager_pagehash(pPg)));
...
3.23.10 Oracle RDBMS and a simple garbage collector for C/C++

There was a time, when the author of these lines tried to learn more about Oracle RDBMS, searching for
vulnerabilities, etc. This is a huge piece of software, and a typical function can take very large nested
objects as arguments. And I wanted to dump these objects, as trees (or graphs).
Also, I tracked all memory allocations/deallocations by intercepting memory allocating/deallocating func-
tions. And when a function to be intercepted getting a pointer to a block in memory, I search for the block
in a list of blocks allocated. I’m getting its size + short name of block (this is like ”tagging” in Windows
OS kernel44 ).
Given a block, I can scan it for 32-bit words (on 32-bit OS) or for 64-bit words (on 64-bit OS). Each word
can be a pointer to another block. And if it is so (I find this another block in my records), I can process it
recursively.
And then, using GraphViz, I could render such a diagrams:
43 https://blogs.msdn.microsoft.com/oldnewthing/20041025-00/?p=37483
44 Read more about comments in allocated blocks: Dennis Yurichev, C/C++ programming language notes http://yurichev.com/
C-book.html
611
Bigger pictures: 1, 2.
This is quite impressive, given the fact that I had no information about data types of all these structures.
But I could get some information from it.
Now the garbage collector for C/C++: Boehm GC
If you use a block allocated in memory, its address has to be present somewhere, as a pointer in some
structure/array in another allocated block, or in globally allocated structure, or in local variable in stack.
If there are no pointer to a block, you can call it ”orphan”, and it will be a reason of memory leak.
And this is what GC45 does. It scans all blocks (because it keep tabs on all blocks allocated) for pointers.
It’s important to understand, that it has no idea of data types of all these structure fields in blocks—this
is important, GC has no information about types. It just scans blocks for 32-bit of 64-bit words and see, if
they could be a pointers to another block(s). It also scans stack. It treats allocated blocks and stack as
arrays of words, some of which may be pointers. And if it found a block allocated, which is ”orphaned”,
i.e., there are no pointer(s) to it from another block(s) or stack, this block considered unneeded, to be
freed. Scanning process takes time, and this is what for GCs are criticized.
Also, GC like Boehm GC46 (for pure C) has function like GC_malloc_atomic()—using it, you declare that
the block allocated using this function will never contain any pointer(s) to other block(s). Maybe this could
be a text string, or other type of data. (Indeed, GC_strdup() calls GC_malloc_atomic().) GC will not scan
it.
3.24 Loop optimizations

3.24.1 Weird loop optimization
This is a simplest ever memcpy() function implementation:
void memcpy (unsigned char* dst, unsigned char* src, size_t cnt)
{
size_t i;
45 Garbage Collector
46 https://www.hboehm.info/gc/
612
dst[i]=src[i];
};
At least MSVC 6.0 from the end of 1990s till MSVC 2013 can produce a really weird code (this listing is
generated by MSVC 2013 x86):
_dst$ = 8 ; size = 4
_src$ = 12 ; size = 4
_cnt$ = 16 ; size = 4
_memcpy PROC
mov edx, DWORD PTR _cnt$[esp-4]
test edx, edx
je SHORT $LN1@f
mov eax, DWORD PTR _dst$[esp-4]
push esi
mov esi, DWORD PTR _src$[esp]
sub esi, eax
; ESI=src-dst, i.e., pointers difference
$LL8@f:
mov cl, BYTE PTR [esi+eax] ; load byte at "esi+dst" or at "src-dst+dst" at the
beginning or at just "src"
lea eax, DWORD PTR [eax+1] ; dst++
mov BYTE PTR [eax-1], cl ; store the byte at "(dst++)--" or at just "dst" at the
beginning
dec edx ; decrement counter until we finished
jne SHORT $LL8@f
pop esi
$LN1@f:
ret 0
_memcpy ENDP
This is weird, because how humans work with two pointers? They store two addresses in two registers or
two memory cells. MSVC compiler in this case stores two pointers as one pointer (sliding dst in EAX) and
difference between src and dst pointers (left unchanged over the span of loop body execution in ESI). (By
the way, this is a rare case when ptrdiff_t data type can be used.) When it needs to load a byte from src,
it loads it at diff + sliding dst and stores byte at just sliding dst.
This has to be some optimization trick. But I’ve rewritten this function to:
_f2 PROC
mov edx, DWORD PTR _cnt$[esp-4]
test edx, edx
je SHORT $LN1@f
mov eax, DWORD PTR _dst$[esp-4]
push esi
mov esi, DWORD PTR _src$[esp]
; eax=dst; esi=src
$LL8@f:
mov cl, BYTE PTR [esi+edx]
mov BYTE PTR [eax+edx], cl
dec edx
jne SHORT $LL8@f
pop esi
$LN1@f:
ret 0
_f2 ENDP
…and it works as efficient as the optimized version on my Intel Xeon E31220 @ 3.10GHz. Maybe, this
optimization was targeted some older x86 CPUs of 1990s era, since this trick is used at least by ancient
MS VC 6.0?
Any idea?
Hex-Rays 2.2 have a hard time recognizing patterns like that (hopefully, temporary?):
void __cdecl f1(char *dst, char *src, size_t size)
{
size_t counter; // edx@1
char *sliding_dst; // eax@2
char tmp; // cl@3
613
counter = size;
if ( size )
{
sliding_dst = dst;
do
{
tmp = (sliding_dst++)[src - dst]; // difference (src-dst) is calculated once, at
the beginning
*(sliding_dst - 1) = tmp;
--counter;
}
while ( counter );
}
}
Nevertheless, this optimization trick is often used by MSVC (not just in DIY47 homebrew memcpy() routines,
but in many loops which uses two or more arrays), so it’s worth for reverse engineers to keep it in mind.
3.24.2 Another loop optimization

If you process all elements of some array which happens to be located in global memory, compiler can
optimize it. For example, let’s calculate a sum of all elements of array of 128 int’s:
#include <stdio.h>
int a[128];
int sum_of_a()
{
int rt=0;
for (int i=0; i<128; i++)

rt=rt+a[i];
return rt;
};
int main()
{
// initialize
for (int i=0; i<128; i++)
a[i]=i;
// calculate the sum

printf ("%d\n", sum_of_a());
};
Optimizing GCC 5.3.1 (x86) can produce this (IDA):

.text:080484B0 sum_of_a proc near
.text:080484B0 mov edx, offset a
.text:080484B5 xor eax, eax
.text:080484B7 mov esi, esi
.text:080484B9 lea edi, [edi+0]
.text:080484C0
.text:080484C0 loc_80484C0: ; CODE XREF: sum_of_a+1B
.text:080484C0 add eax, [edx]
.text:080484C2 add edx, 4
.text:080484C5 cmp edx, offset __libc_start_main@@GLIBC_2_0
.text:080484CB jnz short loc_80484C0
.text:080484CD rep retn
.text:080484CD sum_of_a endp
.text:080484CD
...
47 Do It Yourself
614
.bss:0804A040 public a
.bss:0804A040 a dd 80h dup(?) ; DATA XREF: main:loc_8048338
.bss:0804A040 ; main+19
.bss:0804A040 _bss ends
.bss:0804A040
extern:0804A240 ; ===========================================================================
extern:0804A240
extern:0804A240 ; Segment type: Externs
extern:0804A240 ; extern
extern:0804A240 extrn __libc_start_main@@GLIBC_2_0:near
extern:0804A240 ; DATA XREF: main+25
extern:0804A240 ; main+5D
extern:0804A244 extrn __printf_chk@@GLIBC_2_3_4:near
extern:0804A248 extrn __libc_start_main:near
extern:0804A248 ; CODE XREF: ___libc_start_main
extern:0804A248 ; DATA XREF: .got.plt:off_804A00C
What the heck is __libc_start_main@@GLIBC_2_0 at 0x080484C5? This is a label just after end of a[]
array. The function can be rewritten like this:
int sum_of_a_v2()
{
int *tmp=a;
int rt=0;
do
{
rt=rt+(*tmp);
tmp++;
}
while (tmp<(a+128));
return rt;
};
First version has i counter, and the address of each element of array is to be calculated at each itera-
tion. The second version is more optimized: the pointer to each element of array is always ready and is
sliding 4 bytes forward at each iteration. How to check if the loop is ended? Just compare the pointer
with the address just behind array’s end, which is, in our case, is happens to be address of imported
__libc_start_main() function from Glibc 2.0. Sometimes code like this is confusing, and this is very
popular optimizing trick, so that’s why I made this example.
My second version is very close to what GCC did, and when I compile it, the code is almost the same as
in first version, but two first instructions are swapped:
.text:080484D0 public sum_of_a_v2
.text:080484D0 sum_of_a_v2 proc near
.text:080484D0 xor eax, eax
.text:080484D2 mov edx, offset a
.text:080484D7 mov esi, esi
.text:080484D9 lea edi, [edi+0]
.text:080484E0
.text:080484E0 loc_80484E0: ; CODE XREF: sum_of_a_v2+1B
.text:080484E0 add eax, [edx]
.text:080484E2 add edx, 4
.text:080484E5 cmp edx, offset __libc_start_main@@GLIBC_2_0
.text:080484EB jnz short loc_80484E0
.text:080484ED rep retn
.text:080484ED sum_of_a_v2 endp
Needless to say, this optimization is possible if the compiler can calculate address of the end of array
during compilation time. This happens if the array is global and it’s size is fixed.
However, if the address of array is unknown during compilation, but size is fixed, address of the label just
behind array’s end can be calculated at the beginning of the loop.
615
3.25 More about structures
3.25.1 Sometimes a C structure can be used instead of array
Arithmetic mean
#include <stdio.h>
int mean(int *a, int len)

{
int sum=0;
for (int i=0; i<len; i++)
sum=sum+a[i];
return sum/len;
};
struct five_ints
{
int a0;
int a1;
int a2;
int a3;
int a4;
};
int main()
{
struct five_ints a;
a.a0=123;
a.a1=456;
a.a2=789;
a.a3=10;
a.a4=100;
printf ("%d\n", mean(&a, 5));
// test: https://www.wolframalpha.com/input/?i=mean(123,456,789,10,100)
};
This works: mean() function will never access behind the end of five_ints structure, because 5 is passed,
meaining, only 5 integers will be accessed.
Putting string into structure
#include <stdio.h>
struct five_chars
{
char a0;
char a1;
char a2;
char a3;
char a4;
} __attribute__ ((aligned (1),packed));
int main()
{
struct five_chars a;
a.a0='h';
a.a1='i';
a.a2='!';
a.a3='\n';
a.a4=0;
printf (&a); // prints "hi!"
};
((aligned (1),packed)) attribute must be used, because otherwise, each structure field will be aligned on
4-byte or 8-byte boundary.
616
Summary
This is just another example of how structures and arrays are stored in memory. Perhaps, no sane pro-
grammer will do something like in this example, except in case of some specific hack. Or maybe in case
of source code obfuscation?
3.25.2 Unsized array in C structure

In some win32 structures we can find ones with last field defined as an array of one element:
typedef struct _SYMBOL_INFO {
ULONG SizeOfStruct;
ULONG TypeIndex;
...
ULONG MaxNameLen;
TCHAR Name[1];
} SYMBOL_INFO, *PSYMBOL_INFO;
( https://msdn.microsoft.com/en-us/library/windows/desktop/ms680686(v=vs.85).aspx )
This is a hack, meaning, the last field is array of unknown size, which is to be calculated at the time of
structure allocation.
Why: Name field may be short, so why to define it with some kind of MAX_NAME constant which can be
128, 256, or even bigger?
Why not to use pointer instead? Then you have to allocate two blocks: one for structure and the other
one for string. This may be slower and may require larger memory overhead. Also, you need dereference
pointer (i.e., read address of the string from the structure)—not a big deal, but some people say this is
still surplus cost.
This is also known as struct hack: http://c-faq.com/struct/structhack.html.
Example:
#include <stdio.h>
struct st
{
int a;
int b;
char s[];
};
void f (struct st *s)

{
printf ("%d %d %s\n", s->a, s->b, s->s);
// f() can't replace s[] with bigger string - size of allocated block is unknown at this
point
};
int main()
{
#define STRING "Hello!"
struct st *s=malloc(sizeof(struct st)+strlen(STRING)+1); // incl. terminating zero
s->a=1;
s->b=2;
strcpy (s->s, STRING);
f(s);
};
In short, it works because C has no array boundary checks. Any array is treated as having infinite size.
Problem: after allocation, the whole size of allocated block for structure is unknown (except for memory
manager), so you can’t just replace string with larger string. You would still be able to do so if the field
would be declared as something like s[MAX_NAME].
In other words, you have a structure plus an array (or string) fused together in the single allocated memory
block. Another problem is what you obviously can’t declare two such arrays in single structure, or to
declare another field after such array.
617
Older compilers require to declare array with at least one element: s[1], newer allows to declare it as
variable-sized array: s[]. This is also called flexible array member in C99 standard.
Read more about it in GCC documentation48 , MSDN documentation49 .
Dennis Ritchie (one of C creators) called this trick “unwarranted chumminess with the C implementation”
(perhaps, acknowledging hackish nature of the trick).
Like it or not, use it or not: it is still another demonstration on how structures are stored in memory, that’s
why I write about it.
3.25.3 Version of C structure

Many Windows programmers have seen this in MSDN:
SizeOfStruct
The size of the structure, in bytes. This member must be set to sizeof(SYMBOL_INFO).
( https://msdn.microsoft.com/en-us/library/windows/desktop/ms680686(v=vs.85).aspx )
Some structures like SYMBOL_INFO has started with this field indeed. Why? This is some kind of structure
version.
Imagine you have a function which draws circle. It takes a single argument—a pointer to a structure with
only three fields: X, Y and radius. And then color displays flooded a market, sometimes in 1980s. And
you want to add color argument to the function. But, let’s say, you cannot add another argument to it (a
lot of software use your API50 and cannot be recompiled). And if the old piece of software uses your API
with color display, let your function draw a circle in (default) black and white colors.
Another day you add another feature: circle now can be filled, and brush type can be set.
Here is one solution to the problem:
#include <stdio.h>
struct ver1
{
size_t SizeOfStruct;
int coord_X;
int coord_Y;
int radius;
};
struct ver2
{
int coord_X;
int coord_Y;
int radius;
int color;
};
struct ver3
{
int coord_X;
int coord_Y;
int radius;
int color;
int fill_brush_type; // 0 - do not fill circle
};
void draw_circle(struct ver3 *s) // latest struct version is used here

{
// we presume SizeOfStruct, coord_X and coord_Y fields are always present
printf ("We are going to draw a circle at %d:%d\n", s->coord_X, s->coord_Y);
if (s->SizeOfStruct>=sizeof(int)*5)
48 https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html
49 https://msdn.microsoft.com/en-us/library/b6fae073.aspx
50 Application Programming Interface
618
{
// this is at least ver2, color field is present
printf ("We are going to set color %d\n", s->color);
}
if (s->SizeOfStruct>=sizeof(int)*6)
{
// this is at least ver3, fill_brush_type field is present
printf ("We are going to fill it using brush type %d\n", s->fill_brush_type);
}
};
// early software version

void call_as_ver1()
{
struct ver1 s;
s.SizeOfStruct=sizeof(s);
s.coord_X=123;
s.coord_Y=456;
s.radius=10;
printf ("** %s()\n", __FUNCTION__);
draw_circle(&s);
};
// next software version

void call_as_ver2()
{
struct ver2 s;
s.coord_X=123;
s.coord_Y=456;
s.radius=10;
s.color=1;
draw_circle(&s);
};
// latest, the most extended version

void call_as_ver3()
{
struct ver3 s;
s.coord_X=123;
s.coord_Y=456;
s.radius=10;
s.color=1;
s.fill_brush_type=3;
draw_circle(&s);
};
int main()
{
call_as_ver1();
call_as_ver2();
call_as_ver3();
};
In other words, SizeOfStruct field takes a role of version of structure field. It could be enumerate type (1,
2, 3, etc.), but to set SizeOfStruct field to sizeof(struct...) is less prone to mistakes/bugs: we just write
s.SizeOfStruct=sizeof(...) in caller’s code.
In C++, this problem is solved using inheritance ( 3.21.1 on page 547). You just extend your base class
(let’s call it Circle), and then you will have ColoredCircle and then FilledColoredCircle, and so on. A current
version of an object (or, more precisely, current type) will be determined using C++ RTTI.
So when you see SizeOfStruct somewhere in MSDN—perhaps this structure was extended at least once
in past.
619
3.25.4 High-score file in “Block out” game and primitive serialization
Many videogames has high-score file, sometimes called “Hall of fame”. Ancient “Block out”51 game (3D
tetris from 1989) isn’t exception, here is what we see at the end:
Figure 3.4: High score table
Now we can see that the file has changed after we added our name is BLSCORE.DAT.
% xxd -g 1 BLSCORE.DAT
00000000: 0a 00 58 65 6e 69 61 2e 2e 2e 2e 2e 00 df 01 00 ..Xenia.........
00000010: 00 30 33 2d 32 37 2d 32 30 31 38 00 50 61 75 6c .03-27-2018.Paul
00000020: 2e 2e 2e 2e 2e 2e 00 61 01 00 00 30 33 2d 32 37 .......a...03-27
00000030: 2d 32 30 31 38 00 4a 6f 68 6e 2e 2e 2e 2e 2e 2e -2018.John......
00000040: 00 46 01 00 00 30 33 2d 32 37 2d 32 30 31 38 00 .F...03-27-2018.
00000050: 4a 61 6d 65 73 2e 2e 2e 2e 2e 00 44 01 00 00 30 James......D...0
00000060: 33 2d 32 37 2d 32 30 31 38 00 43 68 61 72 6c 69 3-27-2018.Charli
00000070: 65 2e 2e 2e 00 ea 00 00 00 30 33 2d 32 37 2d 32 e........03-27-2
00000080: 30 31 38 00 4d 69 6b 65 2e 2e 2e 2e 2e 2e 00 b5 018.Mike........
00000090: 00 00 00 30 33 2d 32 37 2d 32 30 31 38 00 50 68 ...03-27-2018.Ph
000000a0: 69 6c 2e 2e 2e 2e 2e 2e 00 ac 00 00 00 30 33 2d il...........03-
000000b0: 32 37 2d 32 30 31 38 00 4d 61 72 79 2e 2e 2e 2e 27-2018.Mary....
000000c0: 2e 2e 00 7b 00 00 00 30 33 2d 32 37 2d 32 30 31 ...{...03-27-201
000000d0: 38 00 54 6f 6d 2e 2e 2e 2e 2e 2e 2e 00 77 00 00 8.Tom........w..
000000e0: 00 30 33 2d 32 37 2d 32 30 31 38 00 42 6f 62 2e .03-27-2018.Bob.
000000f0: 2e 2e 2e 2e 2e 2e 00 77 00 00 00 30 33 2d 32 37 .......w...03-27
00000100: 2d 32 30 31 38 00 -2018.
All entries are clearly visible. The very first byte is probably number of entries. Second is zero and, in fact,
number of entries can be 16-bit value spanning over first two bytes.
Next, after “Xenia” name we see 0xDF and 0x01 bytes. Xenia has score of 479, and this is exactly 0x1DF
in hexadecimal radix. So a high score value is probably 16-bit integer, or maybe 32-bit integer: there are
two more zero bytes after.
Now let’s think about the fact that both array elements and structure elements are always placed in
memory adjacently to each other. That enables us to write the whole array/structure to the file using
simple write() or fwrite() function, and then restore it using read() or fread(), as simple as that. This is
what is called serialization nowadays.
Read
Now let’s write C program to read highscore file:

51 http://www.bestoldgames.net/eng/old-games/blockout.php
620
#include <assert.h>
#include <stdio.h>
#include <stdint.h>
#include <string.h>
struct entry
{
char name[11]; // incl. terminating zero
uint32_t score;
char date[11]; // incl. terminating zero
} __attribute__ ((aligned (1),packed));
struct highscore_file
{
uint8_t count;
uint8_t unknown;
struct entry entries[10];
} __attribute__ ((aligned (1), packed));
struct highscore_file file;
int main(int argc, char* argv[])

{
FILE* f=fopen(argv[1], "rb");
assert (f!=NULL);
size_t got=fread(&file, 1, sizeof(struct highscore_file), f);
assert (got==sizeof(struct highscore_file));
fclose(f);
for (int i=0; i<file.count; i++)
{
printf ("name=%s score=%d date=%s\n",
file.entries[i].name,
file.entries[i].score,
file.entries[i].date);
};
};
We need GCC ((aligned (1),packed)) attribute so that all structure fields will be packed on 1-byte boundary.
Of course it works:
name=Xenia..... score=479 date=03-27-2018
name=Paul...... score=353 date=03-27-2018
name=John...... score=326 date=03-27-2018
name=James..... score=324 date=03-27-2018
name=Charlie... score=234 date=03-27-2018
name=Mike...... score=181 date=03-27-2018
name=Phil...... score=172 date=03-27-2018
name=Mary...... score=123 date=03-27-2018
name=Tom....... score=119 date=03-27-2018
name=Bob....... score=119 date=03-27-2018
(Needless to say, each name is padded with dots, both on screen and in the file, perhaps, for æsthetical
reasons.)
Write
Let’s check if we right about width of score value. Is it really has 32 bits?
int main(int argc, char* argv[])
{
FILE* f=fopen(argv[1], "rb");
assert (f!=NULL);
size_t got=fread(&file, 1, sizeof(struct highscore_file), f);
fclose(f);
strcpy (file.entries[1].name, "Mallory...");

file.entries[1].score=12345678;
621
strcpy (file.entries[1].date, "08-12-2016");
f=fopen(argv[1], "wb");
assert (f!=NULL);
got=fwrite(&file, 1, sizeof(struct highscore_file), f);
fclose(f);
};
Let’s run Blockout:
First two digits (1 and 2) are truncated: 12345678 becomes 345678. Perhaps, this is formatting issues...
but the number is almost correct. Now I’m changing it to 999999 and run again:
Now it’s correct. Yes, high score value is 32-bit integer.
622
Is it serialization?
…almost. Serialization like this is highly popular in scientific and engineering software, where efficiency
and speed is much more important than converting into XML52 or JSON53 and back.
One important thing is that you obviously cannot serialize pointers, because each time you load the file
into memory, all the structures may be allocated in different places.
But: if you work on some kind of low-cost MCU with simple OS on it and you have your structures allocated
at always same places in memory, perhaps you can save and restore pointers as well.
Random noise
When I prepared this example, I had to run “Block out” many times and played for it a bit to fill high-score
table with random names.
And when there were just 3 entries in the file, I saw this:
00000000: 03 00 54 6f 6d 61 73 2e 2e 2e 2e 2e 00 da 2a 00 ..Tomas.......*.
00000010: 00 30 38 2d 31 32 2d 32 30 31 36 00 43 68 61 72 .08-12-2016.Char
00000020: 6c 69 65 2e 2e 2e 00 8b 1e 00 00 30 38 2d 31 32 lie........08-12
00000030: 2d 32 30 31 36 00 4a 6f 68 6e 2e 2e 2e 2e 2e 2e -2016.John......
00000040: 00 80 00 00 00 30 38 2d 31 32 2d 32 30 31 36 00 .....08-12-2016.
00000050: 00 00 57 c8 a2 01 06 01 ba f9 47 c7 05 00 f8 4f ..W.......G....O
00000060: 06 01 06 01 a6 32 00 00 00 00 00 00 00 00 00 00 .....2..........
00000070: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000080: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000090: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
000000a0: 00 00 00 00 00 00 00 00 00 00 93 c6 a2 01 46 72 ..............Fr
000000b0: 8c f9 f6 c5 05 00 f8 4f 00 02 06 01 a6 32 06 01 .......O.....2..
000000c0: 00 00 98 f9 f2 c0 05 00 f8 4f 00 02 a6 32 a2 f9 .........O...2..
000000d0: 80 c1 a6 32 a6 32 f4 4f aa f9 39 c1 a6 32 06 01 ...2.2.O..9..2..
000000e0: b4 f9 2b c5 a6 32 e1 4f c7 c8 a2 01 82 72 c6 f9 ..+..2.O.....r..
000000f0: 30 c0 05 00 00 00 00 00 00 00 a6 32 d4 f9 76 2d 0..........2..v-
00000100: a6 32 00 00 00 00 .2....
The first byte has value of 3, meaning there are 3 entries. And there are 3 entries present. But then we
see a random noise at the second half of file.
The noise is probably has its origins in uninitialized data. Perhaps, “Block out” allocated memory for 10
entries somewhere in heap, where, obviously, some pseudorandom noise (left from something else) was
present. Then it set first/second byte, fill 3 entries, and then it never touched 7 entries left, so they are
written to the file as is.
When “Block out” loads high score file at the next run, it reads number of entries from the first/second
byte (3) and then completely ignores what is after it.
This is common problem. Not a problem in strict sense: it’s not a bug, but information can be exposed
outwards.
Microsoft Word versions from 1990s has been often left pieces of previously edited texts into the *.doc*
files. It was some kind of amusement back then, to get a .doc file from someone, then open it in a
hexadecimal editor and read something else, what has been edited on that computer before.
The problem can be even much more serious: Heartbleed bug in OpenSSL.
Homework
“Block out” has several polycubes (flat/basic/extended), size of pit can be configured, etc. And it seems,
for each configuration, “Block out” has its own high score table. I’ve noticed that some information is
probably stored in BLSCORE.IDX file. This can be a homework for hardcore “Block out” fans—to understand
its structure as well.
The “Block out” files are here: http://beginners.re/examples/blockout.zip (including the binary high
score files I’ve used in this example). You can use DosBox to run it.
52 Extensible Markup Language
53 JavaScript Object Notation
623
3.26 memmove() and memcpy()
The difference between these standard functions is that memcpy() blindly copies a block to another place,
while memmove() correctly handles overlapping blocks. For example, you want to tug a string two bytes
forward:
`|.|.|h|e|l|l|o|...` -> `|h|e|l|l|o|...`
memcpy() which copies 32-bit or 64-bit words at once, or even SIMD, will obviously fail here, a byte-wise
copy routine must be used instead.
Now even more advanced example, insert two bytes in front of string:
`|h|e|l|l|o|...` -> `|.|.|h|e|l|l|o|...`
Now even byte-wise memory copy routine will fail, you have to copy bytes starting at the end.
That’s a rare case where DF x86 flag is to be set before REP MOVSB instruction: DF defines direction, and
now we must move backwardly.
The typical memmove() routine works like this: 1) if source is below destination, copy forward; 2) if source
is above destination, copy backward.
This is memmove() from uClibc:
void *memmove(void *dest, const void *src, size_t n)
{
int eax, ecx, esi, edi;
__asm__ __volatile__(
" movl %%eax, %%edi\n"
" cmpl %%esi, %%eax\n"
" je 2f\n" /* (optional) src == dest -> NOP */
" jb 1f\n" /* src > dest -> simple copy */
" leal -1(%%esi,%%ecx), %%esi\n"
" leal -1(%%eax,%%ecx), %%edi\n"
" std\n"
"1: rep; movsb\n"
" cld\n"
"2:\n"
: "=&c" (ecx), "=&S" (esi), "=&a" (eax), "=&D" (edi)
: "0" (n), "1" (src), "2" (dest)
: "memory"
);
return (void*)eax;
}
In the first case, REP MOVSB is called with DF flag cleared. In the second, DF is set, then cleared.
More complex algorithm has the following piece in it:
“if difference between source and destination is larger than width of word, copy using words rather than
bytes, and use byte-wise copy to copy unaligned parts”.
This how it happens in Glibc 2.24 in non-optimized C part.
Given all that, memmove() may be slower than memcpy(). But some people, including Linus Torvalds,
argue54 that memcpy() should be an alias (or synonym) of memmove(), and the latter function must
just check at start, if the buffers are overlapping or not, and then behave as memcpy() or memmove().
Nowadays, check for overlapping buffers is very cheap, after all.
3.26.1 Anti-debugging trick

I’ve heard about anti-debugging trick where all you need is just set DF to crash the process: the very
next memcpy() routine will lead to crash because it copies backwardly. But I can’t check this: it seems all
memory copy routines clear/set DF as they want to. On the other hand, memmove() from uClibc I cited
here, has no explicit clear of DF (it assumes DF is always clear?), so it can really crash.
54 https://bugzilla.redhat.com/show_bug.cgi?id=638477#c132
624
3.27 setjmp/longjmp
setjmp/longjmp is a mechanism in C which is very similar to throw/catch mechanism in C++ and other
higher-level PLs. Here is an example from zlib:
...
/* return if bits() or decode() tries to read past available input */

if (setjmp(s.env) != 0) /* if came back here via longjmp(), */
err = 2; /* then skip decomp(), return error */
else
err = decomp(&s); /* decompress */
...
/* load at least need bits into val */

val = s->bitbuf;
while (s->bitcnt < need) {
if (s->left == 0) {
s->left = s->infun(s->inhow, &(s->in));
if (s->left == 0) longjmp(s->env, 1); /* out of input */
...
if (s->left == 0) {
s->left = s->infun(s->inhow, &(s->in));
if (s->left == 0) longjmp(s->env, 1); /* out of input */
( zlib/contrib/blast/blast.c )
Call to setjmp() saves current PC, SP and other registers into env structure, then it returns 0.
In case of error, longjmp() teleporting you into the point after right after setjmp() call, as if setjmp()
call returned non-null value (which was passed to longjmp()). This reminds as fork() syscall in UNIX.
Now let’s take a look on distilled example:
#include <stdio.h>
#include <setjmp.h>
jmp_buf env;
void f2()
{
printf ("%s() begin\n", __FUNCTION__);
// something odd happened here
longjmp (env, 1234);
printf ("%s() end\n", __FUNCTION__);
};
void f1()
{
printf ("%s() begin\n", __FUNCTION__);
f2();
printf ("%s() end\n", __FUNCTION__);
};
int main()
{
int err=setjmp(env);
if (err==0)
{
f1();
}
else
{
printf ("Error %d\n", err);
};
};
If we run it, we will see:
625
f1() begin
f2() begin
Error 1234
jmp_buf structure usually comes undocumented, to preserve forward compatibility.

Let’s see how setjmp() implemented in MSVC 2013 x64:
...
; RCX = address of jmp_buf
mov [rcx], rax

mov [rcx+8], rbx
mov [rcx+18h], rbp
mov [rcx+20h], rsi
mov [rcx+28h], rdi
mov [rcx+30h], r12
mov [rcx+38h], r13
mov [rcx+40h], r14
mov [rcx+48h], r15
lea r8, [rsp+arg_0]
mov [rcx+10h], r8
mov r8, [rsp+0] ; get saved RA from stack
mov [rcx+50h], r8 ; save it
stmxcsr dword ptr [rcx+58h]
fnstcw word ptr [rcx+5Ch]
movdqa xmmword ptr [rcx+60h], xmm6
movdqa xmmword ptr [rcx+0A0h], xmm10
movdqa xmmword ptr [rcx+0B0h], xmm11
movdqa xmmword ptr [rcx+0C0h], xmm12
movdqa xmmword ptr [rcx+0D0h], xmm13
movdqa xmmword ptr [rcx+0E0h], xmm14
movdqa xmmword ptr [rcx+0F0h], xmm15
retn
It just populates jmp_buf structure with current values of almost all registers. Also, current value of RA is
taken from the stack and saved in jmp_buf: it will be used as new value of PC in future.
Now longjmp():
...
; RCX = address of jmp_buf
mov rax, rdx

mov rbx, [rcx+8]
mov rsi, [rcx+20h]
mov rdi, [rcx+28h]
mov r12, [rcx+30h]
mov r13, [rcx+38h]
mov r14, [rcx+40h]
mov r15, [rcx+48h]
ldmxcsr dword ptr [rcx+58h]
fnclex
fldcw word ptr [rcx+5Ch]
movdqa xmm6, xmmword ptr [rcx+60h]
movdqa xmm10, xmmword ptr [rcx+0A0h]
movdqa xmm11, xmmword ptr [rcx+0B0h]
movdqa xmm12, xmmword ptr [rcx+0C0h]
movdqa xmm13, xmmword ptr [rcx+0D0h]
movdqa xmm14, xmmword ptr [rcx+0E0h]
626
movdqa xmm15, xmmword ptr [rcx+0F0h]
mov rdx, [rcx+50h] ; get PC (RIP)
mov rbp, [rcx+18h]
mov rsp, [rcx+10h]
jmp rdx ; jump to saved PC
...
It just restores (almost) all registers, takes RA from structure and jumps there. This effectively works as if
setjmp() returned to caller. Also, RAX is set to be equal to the second argument of longjmp(). This works
as if setjmp() returned non-zero value at first place.
As a side effect of SP restoration, all values in stack which has been set and used between setjmp()
and longjmp() calls are just dropped. They will not be used anymore. Hence, longjmp() usually jumps
backwards 55 .
This implies that, unlike in throw/catch mechanism in C++, no memory will be freed, no destructors will
be called, etc. Hence, this technique sometimes can be dangerous. Nevertheless, it’s still quite popular.
It’s still used in Oracle RDBMS.
It also has unexpected side-effect: if some buffer has been overflown inside of a function (maybe due to
remote attack), and a function wants to report error, and it calls longjmp(), overwritten stack part just
gets unused.
As an exercise, you can try to understand, why not all registers are saved. Why XMM0-XMM5 and other
registers are skipped?
3.28 Other weird stack hacks

3.28.1 Accessing arguments/local variables of caller
From C/C++ basics we know that this is impossible for a function to access arguments of caller function
or its local variables.
Nevertheless, it’s possible using dirty hacks. For example:
#include <stdio.h>
void f(char *text)

{
// print stack
int *tmp=&text;
for (int i=0; i<20; i++)
{
printf ("0x%x\n", *tmp);
tmp++;
};
};
void draw_text(int X, int Y, char* text)

{
f(text);
printf ("We are going to draw [%s] at %d:%d\n", text, X, Y);

};
int main()
{
printf ("address of main()=0x%x\n", &main);
printf ("address of draw_text()=0x%x\n", &draw_text);
draw_text(100, 200, "Hello!");
};
On 32-bit Ubuntu 16.04 and GCC 5.4.0, I got this:

address of main()=0x80484f8
address of draw_text()=0x80484cb
55 However, there are some people who can use it for much more complicated things, imitating coroutines, etc: https://www.
embeddedrelated.com/showarticle/455.php, http://fanf.livejournal.com/105413.html
627
0x8048645 first argument to f()
0x8048628
0xbfd8ab98
0xb7634590
0xb779eddc
0xb77e4918
0xbfd8aba8
0x8048547 return address into the middle of main()
0x64 first argument to draw_text()
0xc8 second argument to draw_text()
0x8048645 third argument to draw_text()
0x8048581
0xb779d3dc
0xbfd8abc0
0x0
0xb7603637
0xb779d000
0xb779d000
0x0
0xb7603637
(Comments are mine.)

Since f() starting to enumerate stack elements at its first argument, the first stack element is indeed a
pointer to “Hello!” string. We see its address is also used as third argument to draw_text() function.
In f() we could read all functions arguments and local variables if we know exact stack layout, but it’s
always changed, from compiler to compiler. Various optimization levels affect stack layout greatly.
But if we can somehow detect information we need, we can use it and even modify it. As an example, I’ll
rework f() function:
void f(char *text)
{
...
// find 100, 200 values pair and modify the second on

tmp=&text;
for (int i=0; i<20; i++)
{
if (*tmp==100 && *(tmp+1)==200)
{
printf ("found\n");
*(tmp+1)=210; // change 200 to 210
break;
};
tmp++;
};
};
Holy moly, it works:

found
We are going to draw [Hello!] at 100:210
Summary
It’s extremely dirty hack, intended to demonstrate stack internals. I never ever seen or heard that anyone
used this in a real code. But still, this is a good example.
Exercise
The example has been compiled without optimization on 32-bit Ubuntu using GCC 5.4.0 and it works. But
when I turn on -O3 maximum optimization, it’s failed. Try to find why.
Use your favorite compiler and OS, try various optimization levels, find if it works and if it doesn’t, find
why.
628
3.28.2 Returning string
This is classic bug from Brian W. Kernighan, Rob Pike, Practice of Programming, (1999):
#include <stdio.h>
char* amsg(int n, char* s)

{
char buf[100];
sprintf (buf, "error %d: %s\n", n, s) ;
return buf;
};
int main()
{
printf ("%s\n", amsg (1234, "something wrong!"));
};
It would crash. First, let’s understand, why.

This is a stack state before amsg() return:
(lower addresses)
...
[amsg(): 100 bytes]

[RA] <- current SP
[two amsg arguments]
[something else]
[main() local variables]
...
(upper addresses)
When amsg() returns control flow to main(), so far so good. But printf() is called from main(), which
is, in turn, use stack for its own needs, zapping 100-byte buffer. A random garbage will be printed at the
best.
Hard to believe, but I know how to fix this problem:
#include <stdio.h>
char* amsg(int n, char* s)

{
char buf[100];
sprintf (buf, "error %d: %s\n", n, s) ;
return buf;
};
char* interim (int n, char* s)

{
char large_buf[8000];
// make use of local array.
// it will be optimized away otherwise, as useless.
large_buf[0]=0;
return amsg (n, s);
};
int main()
{
printf ("%s\n", interim (1234, "something wrong!"));
};
It will work if compiled by MSVC 2013 with no optimizations and with /GS- option56 . MSVC will warn:
56 Turn off buffer security check
629
“warning C4172: returning address of local variable or temporary”, but the code will run and message will
be printed. Let’s see stack state at the moment when amsg() returns control to interim():
(lower addresses)
...
[amsg(): 100 bytes]

[RA] <- current SP
[two amsg() arguments]
[interim() stuff, incl. 8000 bytes]
[something else]
...
(upper addresses)
Now the stack state at the moment when interim() returns control to main():
(lower addresses)
...
[amsg(): 100 bytes]

[RA]
[two amsg() arguments]
[interim() stuff, incl. 8000 bytes]
[something else] <- current SP
...
(upper addresses)
So when main() calls printf(), it uses stack at the place where interim()’s buffer was allocated, and
doesn’t zap 100 bytes with error message inside, because 8000 bytes (or maybe much less) is just enough
for everything printf() and other descending functions do!
It may also work if there are many functions between, like: main() → f1() → f2() → f3() ... → amsg(), and
then the result of amsg() is used in main(). The distance between SP in main() and address of buf[]
must be long enough,
This is why bugs like these are dangerous: sometimes your code works (and bug can be hiding unnoticed),
sometimes not. Bugs like these are jokingly called heisenbugs or schrödinbugs.
3.29 OpenMP
OpenMP is one of the simplest ways to parallelize simple algorithms.
As an example, let’s try to build a program to compute a cryptographic nonce.
In my simplistic example, the nonce is a number added to the plain unencrypted text in order to produce
a hash with some specific features.
For example, at some step, the Bitcoin protocol requires to find such nonce so the resulting hash contains
a specific number of consecutive zeros. This is also called proof of work (i.e., the system proves that it
did some intensive calculations and spent some time for it).
My example is not related to Bitcoin in any way, it will try to add numbers to the “hello, world!_” string in
order to find such number that when “hello, world!_<number>” is hashed with the SHA512 algorithm, it
will contain at least 3 zero bytes.
Let’s limit our brute-force to the interval in 0..INT32_MAX-1 (i.e., 0x7FFFFFFE or 2147483646).
The algorithm is pretty straightforward:
#include <stdio.h>
#include <string.h>
#include <stdlib.h>
630
#include <time.h>
#include "sha512.h"
int found=0;
int32_t checked=0;
int32_t* __min;
int32_t* __max;
time_t start;
#ifdef __GNUC__
#define min(X,Y) ((X) < (Y) ? (X) : (Y))
#define max(X,Y) ((X) > (Y) ? (X) : (Y))
#endif
void check_nonce (int32_t nonce)

{
uint8_t buf[32];
struct sha512_ctx ctx;
uint8_t res[64];
// update statistics
int t=omp_get_thread_num();
if (__min[t]==-1)
__min[t]=nonce;
if (__max[t]==-1)
__max[t]=nonce;
__min[t]=min(__min[t], nonce);
__max[t]=max(__max[t], nonce);
// idle if valid nonce found

if (found)
return;
memset (buf, 0, sizeof(buf));

sprintf (buf, "hello, world!_%d", nonce);
sha512_init_ctx (&ctx);
sha512_process_bytes (buf, strlen(buf), &ctx);
sha512_finish_ctx (&ctx, &res);
if (res[0]==0 && res[1]==0 && res[2]==0)
{
printf ("found (thread %d): [%s]. seconds spent=%d\n", t, buf, time(NULL)-start⤦
Ç );
found=1;
};
#pragma omp atomic
checked++;
#pragma omp critical

if ((checked % 100000)==0)
printf ("checked=%d\n", checked);
};
int main()
{
int32_t i;
int threads=omp_get_max_threads();
printf ("threads=%d\n", threads);
__min=(int32_t*)malloc(threads*sizeof(int32_t));
__max=(int32_t*)malloc(threads*sizeof(int32_t));
for (i=0; i<threads; i++)
__min[i]=__max[i]=-1;
start=time(NULL);
631
#pragma omp parallel for
for (i=0; i<INT32_MAX; i++)
check_nonce (i);
for (i=0; i<threads; i++)

printf ("__min[%d]=0x%08x __max[%d]=0x%08x\n", i, __min[i], i, __max[i]);
free(__min); free(__max);
};
The check_nonce() function just adds a number to the string, hashes it with the SHA512 algorithm and
checks for 3 zero bytes in the result.
A very important part of the code is:
#pragma omp parallel for
for (i=0; i<INT32_MAX; i++)
check_nonce (i);
Yes, that simple, without #pragma we just call check_nonce() for each number from 0 to INT32_MAX
(0x7fffffff or 2147483647). With #pragma, the compiler adds some special code which slices the loop
interval into smaller ones, to run them on all CPU cores available 57 .
58
The example can be compiled in MSVC 2012:
cl openmp_example.c sha512.obj /openmp /O1 /Zi /Faopenmp_example.asm
Or in GCC:
gcc -fopenmp 2.c sha512.c -S -masm=intel
3.29.1 MSVC
Now this is how MSVC 2012 generates the main loop:

push OFFSET _main$omp$1
push 0
push 1
call __vcomp_fork
add esp, 16
All functions prefixed by vcomp are OpenMP-related and are stored in the vcomp*.dll file. So here a group
of threads is started.
Let’s take a look on _main$omp$1:

$T1 = -8 ; size = 4
$T2 = -4 ; size = 4
_main$omp$1 PROC
push ebp
mov ebp, esp
push ecx
push ecx
push esi
push eax
push eax
push 1
push 1
push 2147483646 ; 7ffffffeH
push 0
call __vcomp_for_static_simple_init
57 N.B.: This is intentionally simplest possible example, but in practice, the usage of OpenMP can be harder and more complex
58 sha512.(c|h) and u64.h files can be taken from the OpenSSL library: http://go.yurichev.com/17324
632
mov esi, DWORD PTR $T1[ebp]
add esp, 24
jmp SHORT $LN6@main$omp$1
$LL2@main$omp$1:
push esi
call _check_nonce
pop ecx
inc esi
$LN6@main$omp$1:
cmp esi, DWORD PTR $T2[ebp]
jle SHORT $LL2@main$omp$1
call __vcomp_for_static_end
pop esi
leave
ret 0
_main$omp$1 ENDP
This function is to be started n times in parallel, where n is the number of CPU cores.
vcomp_for_static_simple_init() calculates the interval for the for() construct for the current thread,
depending on the current thread’s number.
The loop’s start and end values are stored in the $T1 and $T2 local variables. You may also notice
7ffffffeh (or 2147483646) as an argument to the vcomp_for_static_simple_init() function—this
is the number of iterations for the whole loop, to be divided evenly.
Then we see a new loop with a call to the check_nonce() function, which does all the work.
Let’s also add some code at the beginning of the check_nonce() function to gather statistics about the
arguments with which the function has been called.
This is what we see when we run it:
threads=4
...
checked=2800000
checked=3000000
checked=3200000
checked=3300000
found (thread 3): [hello, world!_1611446522]. seconds spent=3
__min[0]=0x00000000 __max[0]=0x1fffffff
__min[3]=0x60000000 __max[3]=0x7ffffffe
Yes, the result is correct, the first 3 bytes are zeros:

C:\...\sha512sum test
000000f4a8fac5a4ed38794da4c1e39f54279ad5d9bb3c5465cdf57adaf60403
df6e3fe6019f5764fc9975e505a7395fed780fee50eb38dd4c0279cb114672e2 *test
The running time is ≈ 2..3 seconds on 4-core Intel Xeon E3-1220 3.10 GHz. In the task manager we see 5
threads: 1 main thread + 4 more. No further optimizations are done to keep this example as small and
clear as possible. But probably it can be done much faster. My CPU has 4 cores, that is why OpenMP
started exactly 4 threads.
By looking at the statistics table we can clearly see how the loop has been sliced into 4 even parts. Oh
well, almost even, if we don’t consider the last bit.
There are also pragmas for atomic operations.
Let’s see how this code is compiled:
#pragma omp atomic
checked++;
#pragma omp critical

if ((checked % 100000)==0)
printf ("checked=%d\n", checked);
633
push edi
push OFFSET _checked
call __vcomp_atomic_add_i4
; Line 55
push OFFSET _$vcomp$critsect$
call __vcomp_enter_critsect
add esp, 12
; Line 56
mov ecx, DWORD PTR _checked
mov eax, ecx
cdq
mov esi, 100000 ; 000186a0H
idiv esi
test edx, edx
jne SHORT $LN1@check_nonc
; Line 57
push ecx
push OFFSET ??_C@_0M@NPNHLIOO@checked?$DN?$CFd?6?$AA@
call _printf
pop ecx
pop ecx
$LN1@check_nonc:
push DWORD PTR _$vcomp$critsect$
call __vcomp_leave_critsect
pop ecx
As it turns out, the vcomp_atomic_add_i4() function in the vcomp*.dll is just a tiny function with the LOCK
XADD instruction59 in it.
vcomp_enter_critsect() eventually calling win32 API function
EnterCriticalSection() 60 .
3.29.2 GCC
GCC 4.8.1 produces a program which shows exactly the same statistics table,
so, GCC’s implementation divides the loop in parts in the same fashion.

mov edi, OFFSET FLAT:main._omp_fn.0
call GOMP_parallel_start
mov edi, 0
call main._omp_fn.0
call GOMP_parallel_end
Unlike MSVC’s implementation, what GCC code does is to start 3 threads, and run the fourth in the current
thread. So there are 4 threads instead of the 5 in MSVC.
Here is the main._omp_fn.0 function:

main._omp_fn.0:
push rbp
mov rbp, rsp
push rbx
sub rsp, 40
mov QWORD PTR [rbp-40], rdi
call omp_get_num_threads
mov ebx, eax
call omp_get_thread_num
mov esi, eax
mov eax, 2147483647 ; 0x7FFFFFFF
cdq
idiv ebx
mov ecx, eax
mov eax, 2147483647 ; 0x7FFFFFFF
59 Read more about LOCK prefix: .1.6 on page 1012
60 You can read more about critical sections here: 6.5.4 on page 777
634
cdq
idiv ebx
mov eax, edx
cmp esi, eax
jl .L15
.L18:
imul esi, ecx
mov edx, esi
add eax, edx
lea ebx, [rax+rcx]
cmp eax, ebx
jge .L14
mov DWORD PTR [rbp-20], eax
.L17:
mov edi, eax
call check_nonce
add DWORD PTR [rbp-20], 1
cmp DWORD PTR [rbp-20], ebx
jl .L17
jmp .L14
.L15:
mov eax, 0
add ecx, 1
jmp .L18
.L14:
add rsp, 40
pop rbx
pop rbp
ret
Here we see the division clearly: by calling omp_get_num_threads() and omp_get_thread_num()

we get the number of threads running, and also the current thread’s number, and then determine the
loop’s interval. Then we run check_nonce().
GCC also inserted the LOCK ADD
instruction right in the code, unlike MSVC, which generated a call to a separate DLL function:

lock add DWORD PTR checked[rip], 1
call GOMP_critical_start
mov ecx, DWORD PTR checked[rip]
mov edx, 351843721
mov eax, ecx
imul edx
sar edx, 13
mov eax, ecx
sar eax, 31
sub edx, eax
mov eax, edx
imul eax, eax, 100000
sub ecx, eax
mov eax, ecx
test eax, eax
jne .L7
mov eax, DWORD PTR checked[rip]
mov esi, eax
mov edi, OFFSET FLAT:.LC2 ; "checked=%d\n"
mov eax, 0
call printf
.L7:
call GOMP_critical_end
The functions prefixed with GOMP are from GNU OpenMP library. Unlike vcomp*.dll, its source code is
freely available: GitHub.
635
3.30 Signed division using shifts
Unsigned division by 2n numbers is easy, just use bit shift right by n. Signed division by 2n is easy as well,
but some correction needs to be done before or after shift opeartion.
First, most CPU architectures support two right shift operations: logical and arithmetical. During logical
shift right, free bit(s) at left are set to zero bit(s). This is SHR in x86. During arithmetical shift right, free
bit(s) at left are set equal to the bit which was at the same place. Thus, it preserves sign bit while shifting.
This is SAR in x86.
Interesting to know, there is no special instruction for arithmetical shift left, because it works just as logical
shift left. So, SAL and SHL instructions in x86 are mapped to the same opcode. Many disassemblers don’t
even know about SAL instruction and decode this opcode as SHL.
Hence, arithmetical shift right is used for signed numbers. For example, if you shift -4 (11111100b) by
1 bit right, logical shift right operation will produce 01111110b, which is 126. Arithmetical shift right will
produce 11111110b, which is -2. So far so good.
What if we need to divide -5 by 2? This is -2.5, or just -2 in integer arithmetic. -5 is 11111011b, by shifting
this value by 1 bit right, we’ll get 11111101b, which is -3. This is slightly incorrect.
Another example: − 21 = −0.5 or just 0 in integer arithmetic. But -1 is 11111111b, and 11111111b » 1 =
11111111b, which is -1 again. This is also incorrect.
One solution is to add 1 to the input value if it’s negative.
That is why, if we compile x/2 expression, where x is signed int, GCC 4.8 will produce something like that:
mov eax, edi
shr eax, 31 ; isolate leftmost bit, which is 1 if the number is negative and 0 if
positive
add eax, edi ; add 1 to the input value if it's negative, do nothing otherwise
sar eax ; arithmetical shift right by one bit
ret
If you divide by 4, 3 needs to be added to the input value if it’s negative. So this is what GCC 4.8 does for
x/4:
lea eax, [rdi+3] ; prepare x+3 value ahead of time
test edi, edi
; if the sign is not negative (i.e., positive), move input value to EAX
; if the sign is negative, x+3 value is left in EAX untouched
cmovns eax, edi
; do arithmetical shift right by 2 bits
sar eax, 2
ret
If you divide by 8, 7 will be added to the input value, etc.

MSVC 2013 is slightly different. This is division by 2:
; sign-extend input value to 64-bit value into EDX:EAX
; effectively, that means EDX will be set to 0FFFFFFFFh if the input value is negative
; ... or to 0 if positive
cdq
; subtract -1 from input value if it's negative
; this is the same as adding 1
sub eax, edx
; do arithmetical shift right
sar eax, 1
ret 0
Division by 4 in MSVC 2013 is little more complex:

cdq
; now EDX is 0FFFFFFFFh if input value is negative
; EDX is 0 if it's positive
and edx, 3
; now EDX is 3 if input is negative or 0 otherwise
636
; add 3 to input value if it's negative or do nothing otherwise:
add eax, edx
; do arithmetical shift
sar eax, 2
ret 0
Division by 8 in MSVC 2013 is similar, but 3 bits from EDX is taken instead of 2, producing correction value
of 7 instead of 3.
Sometimes, Hex-Rays 6.8 can’t handle such code correctly, and it may produce something like this:
int v0;
...
__int64 v14
...
v14 = ...;
v0 = ((signed int)v14 - HIDWORD(v14)) >> 1;
... it can be safely rewritten to v0=v14/2.

Hex-Rays 6.8 can also handle signed division by 4 like that:
result = ((BYTE4(v25) & 3) + (signed int)v25) >> 2;
... can be rewritten to v25 / 4.
Also, such correction code is used often when division is replaced by multiplication by magic numbers:
read Mathematics for Programmers61 about multiplicative inverse. And sometimes, additional shifting is
x
used after multiplication. For example, when GCC optimizes 10 , it can’t find multiplicative inverse for 10,
because diophantine equation has no solutions. So it generates code for x5 and then adds arithmetical
shift right operation by 1 bit, to divide the result by 2. Of course, this is true only for signed integers.
So here is division by 10 by GCC 4.8:
mov eax, edi
mov edx, 1717986919 ; magic number
sar edi, 31 ; isolate leftmost bit (which reflects sign)
imul edx ; multiplication by magic number (calculate x/5)
sar edx, 2 ; now calculate (x/5)/2
; subtract -1 (or add 1) if the input value is negative.

; do nothing otherwise:
sub edx, edi
mov eax, edx
ret
Summary: 2n − 1 must be added to input value before arithmetical shift, or 1 must be added to the final
result after shift. Both operations are equivalent to each other, so compiler developers may choose what
is more suitable to them. From the reverse engineer’s point of view, this correction is a clear evidence
that the value has signed type.
3.31 Another heisenbug

Sometimes, array (or buffer) can overflow due to fencepost error:
#include <stdio.h>
int array1[128];
int important_var1;
int important_var2;
int important_var3;
int important_var4;
int important_var5;
61 https://yurichev.com/writings/Math-for-programmers.pdf
637
int main()
{
important_var1=1;
important_var2=2;
important_var3=3;
important_var4=4;
important_var5=5;
array1[0]=123;
array1[128]=456; // BUG
printf ("important_var1=%d\n", important_var1);

};
This is what this program printed in my case (non-optimized GCC 5.4 x86 on Linux):
important_var1=1
important_var2=456
important_var3=3
important_var4=4
important_var5=5
As it happens, important_var2 has been placed by compiler right after array1[]:
Listing 3.131: objdump -x

0804a040 g O .bss 00000200 array1
...
0804a240 g O .bss 00000004 important_var2
...
0804a24c g O .bss 00000004 important_var3
Another compiler can arrange variables in another order, and another variable would be zapped. This
is also heisenbug ( 3.28.2 on page 630)—bug may appear or may left unnoticed depending on compiler
version and optimization switches.
If all variables and arrays are allocated in local stack, stack protection may be triggered, or may not.
However, Valgrind can find bugs like these.
Related example in the book (Angband game): 1.27 on page 301.
3.32 The case of forgotten return

Let’s revisit the “attempt to use the result of a function returning void” part: .
This is a bug I once hit.
And this is also yet another demonstration, how C/C++ places return value into EAX/RAX register.
In the piece of code like that, I forgot to add return:
#include <stdio.h>
#include <stdlib.h>
struct color
{
int R;
int G;
int B;
};
struct color* create_color (int R, int G, int B)

{
638
struct color* rt=(struct color*)malloc(sizeof(struct color));
rt->R=R;
rt->G=G;
rt->B=B;
// must be "return rt;" here
};
int main()
{
struct color* a=create_color(1,2,3);
printf ("%d %d %d\n", a->R, a->G, a->B);
};
Non-optimizing GCC 5.4 silently compiles this with no warnings. And the code works! Let’s see, why:
Listing 3.132: Non-optimizing GCC 5.4

create_color:
push rbp
mov rbp, rsp
sub rsp, 32
mov edi, 12
call malloc
; RAX is pointer to newly allocated buffer
; now fill it with R/G/B:
mov QWORD PTR [rbp-8], rax
mov edx, DWORD PTR [rbp-20]
mov DWORD PTR [rax], edx
mov DWORD PTR [rax+4], edx
nop
leave
; RAX wasn't modified till that point!
ret
If I add return rt;, the only instruction is added at the end, which is redundant:
Listing 3.133: Non-optimizing GCC 5.4

create_color:
push rbp
mov rbp, rsp
sub rsp, 32
mov edi, 12
call malloc
; RAX is pointer to buffer
mov QWORD PTR [rbp-8], rax
mov DWORD PTR [rax], edx
; reload pointer to RAX again, and this is redundant operation...
mov rax, QWORD PTR [rbp-8] ; new instruction
leave
639
ret
Bugs like that are very dangerous, sometimes they appear, sometimes hide. It’s like Heisenbug.
Now I’m trying optimizing GCC:

create_color:
rep ret
main:
xor eax, eax
; as if create_color() was called and returned 0
sub rsp, 8
mov r8d, DWORD PTR ds:8
mov ecx, DWORD PTR [rax+4]
mov edx, DWORD PTR [rax]
mov edi, 1
call __printf_chk
xor eax, eax
add rsp, 8
ret
Compiler deducing that nothing returns from the function, so it optimizes it away. And it assumes, that is
returns 0 by default. The zero is then used as an address to a structure in main().. Of course, this code
crashes.
GCC is C++ mode silent about it as well.
Let’s try non-optimizing MSVC 2015 x86. It warns about the problem:
c:\tmp\3.c(19) : warning C4716: 'create_color': must return a value
And generates crashing code:

_rt$ = -4
_R$ = 8
_G$ = 12
_B$ = 16
_create_color PROC
push ebp
mov ebp, esp
push ecx
push 12
call _malloc
; EAX -> ptr to buffer
add esp, 4
mov ecx, DWORD PTR _R$[ebp]
mov edx, DWORD PTR _rt$[ebp]
mov eax, DWORD PTR _G$[ebp]
; EAX is set to G argument:
mov DWORD PTR [edx+4], eax
mov ecx, DWORD PTR _rt$[ebp]
mov edx, DWORD PTR _B$[ebp]
mov DWORD PTR [ecx+8], edx
mov esp, ebp
pop ebp
; EAX = G at this point:
ret 0
_create_color ENDP
Now optimizing MSVC 2015 x86 generates crashing code as well, but for the different reason:
640
_a$ = -4
_main PROC
; this is inlined optimized version of create_color():
push ecx
push 12
call _malloc
; EAX -> to allocated buffer, and it's filled, OK
; now we reload ptr to buffer, thinking it's in "a" variable
; but inlined function didn't store pointer to "a" variable!
mov eax, DWORD PTR _a$[esp+8]
; EAX = some random garbage at this point
push DWORD PTR [eax]
push OFFSET $SG6074
call _printf
xor eax, eax
add esp, 24
ret 0
_main ENDP
_R$ = 8
_G$ = 12
_B$ = 16
_create_color PROC
push 12
call _malloc
mov ecx, DWORD PTR _R$[esp]
add esp, 4
mov ecx, DWORD PTR _G$[esp-4]
mov ecx, DWORD PTR _B$[esp-4]
; EAX -> to allocated buffer, OK
ret 0
_create_color ENDP
However, non-optimizing MSVC 2015 x64 generates working code:

rt$ = 32
R$ = 64
G$ = 72
B$ = 80
create_color PROC
mov DWORD PTR [rsp+24], r8d
mov DWORD PTR [rsp+16], edx
sub rsp, 56
mov ecx, 12
call malloc
; RAX = allocated buffer
mov QWORD PTR rt$[rsp], rax
mov rax, QWORD PTR rt$[rsp]
mov ecx, DWORD PTR R$[rsp]
mov DWORD PTR [rax], ecx
mov ecx, DWORD PTR G$[rsp]
mov DWORD PTR [rax+4], ecx
mov ecx, DWORD PTR B$[rsp]
mov DWORD PTR [rax+8], ecx
add rsp, 56
; RAX didn't change down to this point
641
ret 0
create_color ENDP
Optimizing MSVC 2015 x64 also inlines the function, as in case of x86, and the resulting code also crashes.
This is a real piece of code from my octothorpe library62 , that worked and all tests passed. It was so,
without return for quite a time...
uint32_t LPHM_u32_hash(void *key)
{
jenkins_one_at_a_time_hash_u32((uint32_t)key);
}
The moral of the story: warnings are very important, use -Wall, etc, etc... When return statement is
absent, compiler can just silently do nothing at that point.
Such a bug left unnoticed can ruin a day.
Also, shotgun debugging is bad, because again, such a bug can left unnoticed (“everything works now, so
be it”).
3.33 Homework: more about function pointers and unions

This code was copypasted from dwm63 , probably, the smallest ever Linux window manager.
The problem: keystrokes from user must be dispatched to various functions inside of dwm. This is usually
solved using a big switch(). Supposedly, dwm’s creators wanted to make the code neat and modifiable
by users:
...
typedef union {
int i;
unsigned int ui;
float f;
const void *v;
} Arg;
...
typedef struct {
unsigned int mod;
KeySym keysym;
void (*func)(const Arg *);
const Arg arg;
} Key;
...
static Key keys[] = {

/* modifier key function argument */
{ MODKEY, XK_p, spawn, {.v = dmenucmd } },
{ MODKEY|ShiftMask, XK_Return, spawn, {.v = termcmd } },
{ MODKEY, XK_b, togglebar, {0} },
{ MODKEY, XK_j, focusstack, {.i = +1 } },
{ MODKEY, XK_k, focusstack, {.i = -1 } },
{ MODKEY, XK_i, incnmaster, {.i = +1 } },
{ MODKEY, XK_d, incnmaster, {.i = -1 } },
{ MODKEY, XK_h, setmfact, {.f = -0.05} },
{ MODKEY, XK_l, setmfact, {.f = +0.05} },
{ MODKEY, XK_Return, zoom, {0} },
{ MODKEY, XK_Tab, view, {0} },
{ MODKEY|ShiftMask, XK_c, killclient, {0} },
62 https://github.com/DennisYurichev/octothorpe
63 https://dwm.suckless.org/
642
{ MODKEY, XK_t, setlayout, {.v = &layouts[0]} },
{ MODKEY, XK_f, setlayout, {.v = &layouts[1]} },
{ MODKEY, XK_m, setlayout, {.v = &layouts[2]} },
...
void
spawn(const Arg *arg)
{
...
void
focusstack(const Arg *arg)
{
...
For each keystroke (or shortcut) a function is defined. Even more: a parameters (or arguments) to be
passed to a function at each case. But parameters can have various type. So union is used here. A value
of needed type is filled in the table. Each function takes what it needs.
As a homework, try to write a code like that, or get into dwm’s and see how union is passed into functions
and handled.
3.34 Windows 16-bit

16-bit Windows programs are rare nowadays, but can be used in the cases of retrocomputing or dongle
hacking ( 8.8 on page 818).
16-bit Windows versions were up to 3.11. 95/98/ME also support 16-bit code, as well as the 32-bit versions
of the Windows NT line. The 64-bit versions of Windows NT line do not support 16-bit executable code at
all.
The code resembles MS-DOS’s one.
Executable files are of type NE-type (so-called “new executable”).
All examples considered here were compiled by the OpenWatcom 1.9 compiler, using these switches:
wcl.exe -i=C:/WATCOM/h/win/ -s -os -bt=windows -bcl=windows example.c
3.34.1 Example#1
int PASCAL WinMain( HINSTANCE hInstance,

HINSTANCE hPrevInstance,
LPSTR lpCmdLine,
int nCmdShow )
{
MessageBeep(MB_ICONEXCLAMATION);
return 0;
};
WinMain proc near

push bp
mov bp, sp
mov ax, 30h ; '0' ; MB_ICONEXCLAMATION constant
push ax
call MESSAGEBEEP
xor ax, ax ; return 0
pop bp
retn 0Ah
WinMain endp
Seems to be easy, so far.
643
3.34.2 Example #2

LPSTR lpCmdLine,
int nCmdShow )
{
MessageBox (NULL, "hello, world", "caption", MB_YESNOCANCEL);
return 0;
};
WinMain proc near

push bp
mov bp, sp
xor ax, ax ; NULL
push ax
push ds
mov ax, offset aHelloWorld ; 0x18. "hello, world"
push ax
push ds
mov ax, offset aCaption ; 0x10. "caption"
push ax
mov ax, 3 ; MB_YESNOCANCEL
push ax
call MESSAGEBOX
pop bp
retn 0Ah
WinMain endp
dseg02:0010 aCaption db 'caption',0

dseg02:0018 aHelloWorld db 'hello, world',0
Couple important things here: the PASCAL calling convention dictates passing the first argument first
(MB_YESNOCANCEL), and the last argument—last (NULL). This convention also tells the callee to restore
the stack pointer: hence the RETN instruction has 0Ah as argument, which implies that the pointer has to
be increased by 10 bytes when the function exits. It is like stdcall ( 6.1.2 on page 725), but the arguments
are passed in “natural” order.
The pointers are passed in pairs: first the data segment is passed, then the pointer inside the segment.
There is only one segment in this example, so DS always points to the data segment of the executable.
3.34.3 Example #3

LPSTR lpCmdLine,
int nCmdShow )
{
int result=MessageBox (NULL, "hello, world", "caption", MB_YESNOCANCEL);
if (result==IDCANCEL)
MessageBox (NULL, "you pressed cancel", "caption", MB_OK);
else if (result==IDYES)
MessageBox (NULL, "you pressed yes", "caption", MB_OK);
else if (result==IDNO)
MessageBox (NULL, "you pressed no", "caption", MB_OK);
return 0;
};
WinMain proc near

push bp
mov bp, sp
644
xor ax, ax ; NULL
push ax
push ds
mov ax, offset aHelloWorld ; "hello, world"
push ax
push ds
mov ax, offset aCaption ; "caption"
push ax
push ax
call MESSAGEBOX
cmp ax, 2 ; IDCANCEL
jnz short loc_2F
xor ax, ax
push ax
push ds
mov ax, offset aYouPressedCanc ; "you pressed cancel"
jmp short loc_49
loc_2F:
cmp ax, 6 ; IDYES
jnz short loc_3D
xor ax, ax
push ax
push ds
mov ax, offset aYouPressedYes ; "you pressed yes"
jmp short loc_49
loc_3D:
cmp ax, 7 ; IDNO
jnz short loc_57
xor ax, ax
push ax
push ds
mov ax, offset aYouPressedNo ; "you pressed no"
loc_49:
push ax
push ds
push ax
xor ax, ax
push ax
call MESSAGEBOX
loc_57:
xor ax, ax
pop bp
retn 0Ah
WinMain endp
Somewhat extended example from the previous section .
3.34.4 Example #4
int PASCAL func1 (int a, int b, int c)

{
return a*b+c;
};
long PASCAL func2 (long a, long b, long c)

{
return a*b+c;
};
long PASCAL func3 (long a, long b, long c, int d)

{
return a*b+c-d;
};
645
LPSTR lpCmdLine,
int nCmdShow )
{
func1 (123, 456, 789);
func2 (600000, 700000, 800000);
func3 (600000, 700000, 800000, 123);
return 0;
};
func1 proc near
c = word ptr 4
b = word ptr 6
a = word ptr 8
push bp
mov bp, sp
mov ax, [bp+a]
imul [bp+b]
add ax, [bp+c]
pop bp
retn 6
func1 endp
func2 proc near
arg_0 = word ptr 4

arg_2 = word ptr 6
arg_4 = word ptr 8
arg_6 = word ptr 0Ah
arg_8 = word ptr 0Ch
arg_A = word ptr 0Eh
push bp
mov bp, sp
mov ax, [bp+arg_8]
mov dx, [bp+arg_A]
mov bx, [bp+arg_4]
mov cx, [bp+arg_6]
call sub_B2 ; long 32-bit multiplication
add ax, [bp+arg_0]
adc dx, [bp+arg_2]
pop bp
retn 12
func2 endp
func3 proc near
arg_0 = word ptr 4

arg_2 = word ptr 6
arg_4 = word ptr 8
arg_8 = word ptr 0Ch
arg_A = word ptr 0Eh
arg_C = word ptr 10h
push bp
mov bp, sp
mov ax, [bp+arg_A]
mov dx, [bp+arg_C]
mov bx, [bp+arg_6]
mov cx, [bp+arg_8]
call sub_B2 ; long 32-bit multiplication
mov cx, [bp+arg_2]
add cx, ax
mov bx, [bp+arg_4]
adc bx, dx ; BX=high part, CX=low part
mov ax, [bp+arg_0]
cwd ; AX=low part d, DX=high part d
646
sub cx, ax
mov ax, cx
sbb bx, dx
mov dx, bx
pop bp
retn 14
func3 endp
WinMain proc near

push bp
mov bp, sp
mov ax, 123
push ax
mov ax, 456
push ax
mov ax, 789
push ax
call func1
mov ax, 9 ; high part of 600000
push ax
mov ax, 27C0h ; low part of 600000
push ax
mov ax, 0Ah ; high part of 700000
push ax
mov ax, 0AE60h ; low part of 700000
push ax
mov ax, 0Ch ; high part of 800000
push ax
mov ax, 3500h ; low part of 800000
push ax
call func2
mov ax, 9 ; high part of 600000
push ax
mov ax, 27C0h ; low part of 600000
push ax
mov ax, 0Ah ; high part of 700000
push ax
mov ax, 0AE60h ; low part of 700000
push ax
mov ax, 0Ch ; high part of 800000
push ax
mov ax, 3500h ; low part of 800000
push ax
mov ax, 7Bh ; 123
push ax
call func3
pop bp
retn 0Ah
WinMain endp
32-bit values (the long data type implies 32 bits, while int is 16-bit) in 16-bit code (both MS-DOS and
Win16) are passed in pairs. It is just like when 64-bit values are used in a 32-bit environment ( 1.34 on
page 392).
sub_B2 here is a library function written by the compiler’s developers that does “long multiplication”, i.e.,
multiplies two 32-bit values. Other compiler functions that do the same are listed here: .5 on page 1029, .4
on page 1029.
The ADD/ADC instruction pair is used for addition of compound values: ADD may set/clear the CF flag, and
ADC uses it after.
The SUB/SBB instruction pair is used for subtraction: SUB may set/clear the CF flag, SBB uses it after.
32-bit values are returned from functions in the DX:AX register pair.
Constants are also passed in pairs in WinMain() here.
The int-typed 123 constant is first converted according to its sign into a 32-bit value using the CWD instruc-
tion.
647
3.34.5 Example #5
int PASCAL string_compare (char *s1, char *s2)

{
while (1)
{
if (*s1!=*s2)
return 0;
if (*s1==0 || *s2==0)
return 1; // end of string
s1++;
s2++;
};
};
int PASCAL string_compare_far (char far *s1, char far *s2)

{
while (1)
{
if (*s1!=*s2)
return 0;
if (*s1==0 || *s2==0)
return 1; // end of string
s1++;
s2++;
};
};
void PASCAL remove_digits (char *s)

{
while (*s)
{
if (*s>='0' && *s<='9')
*s='-';
s++;
};
};
char str[]="hello 1234 world";

LPSTR lpCmdLine,
int nCmdShow )
{
string_compare ("asd", "def");
string_compare_far ("asd", "def");
remove_digits (str);
MessageBox (NULL, str, "caption", MB_YESNOCANCEL);
return 0;
};
string_compare proc near
arg_0 = word ptr 4

arg_2 = word ptr 6
push bp
mov bp, sp
push si
mov si, [bp+arg_0]
mov bx, [bp+arg_2]
loc_12: ; CODE XREF: string_compare+21j

mov al, [bx]
cmp al, [si]
648
jz short loc_1C
xor ax, ax
jmp short loc_2B
loc_1C: ; CODE XREF: string_compare+Ej

test al, al
jz short loc_22
jnz short loc_27

mov ax, 1
jmp short loc_2B

inc bx
inc si
jmp short loc_12
loc_2B: ; CODE XREF: string_compare+12j

; string_compare+1Dj
pop si
pop bp
retn 4
string_compare endp
string_compare_far proc near ; CODE XREF: WinMain+18p
arg_0 = word ptr 4

arg_2 = word ptr 6
arg_4 = word ptr 8
push bp
mov bp, sp
push si
mov si, [bp+arg_0]
mov bx, [bp+arg_4]
loc_3A: ; CODE XREF: string_compare_far+35j

mov es, [bp+arg_6]
mov al, es:[bx]
mov es, [bp+arg_2]
cmp al, es:[si]
jz short loc_4C
xor ax, ax
jmp short loc_67
loc_4C: ; CODE XREF: string_compare_far+16j

mov es, [bp+arg_6]
cmp byte ptr es:[bx], 0
jz short loc_5E
mov es, [bp+arg_2]
cmp byte ptr es:[si], 0
jnz short loc_63
loc_5E: ; CODE XREF: string_compare_far+23j

mov ax, 1
jmp short loc_67
loc_63: ; CODE XREF: string_compare_far+2Cj

inc bx
inc si
jmp short loc_3A
649
loc_67: ; CODE XREF: string_compare_far+1Aj
; string_compare_far+31j
pop si
pop bp
retn 8
string_compare_far endp
remove_digits proc near ; CODE XREF: WinMain+1Fp
arg_0 = word ptr 4
push bp
mov bp, sp
mov bx, [bp+arg_0]
loc_72: ; CODE XREF: remove_digits+18j

mov al, [bx]
test al, al
jz short loc_86
cmp al, 30h ; '0'
jb short loc_83
cmp al, 39h ; '9'
ja short loc_83
mov byte ptr [bx], 2Dh ; '-'
loc_83: ; CODE XREF: remove_digits+Ej

; remove_digits+12j
inc bx
jmp short loc_72
loc_86: ; CODE XREF: remove_digits+Aj

pop bp
retn 2
remove_digits endp
WinMain proc near ; CODE XREF: start+EDp

push bp
mov bp, sp
mov ax, offset aAsd ; "asd"
push ax
mov ax, offset aDef ; "def"
push ax
call string_compare
push ds
mov ax, offset aAsd ; "asd"
push ax
push ds
mov ax, offset aDef ; "def"
push ax
call string_compare_far
mov ax, offset aHello1234World ; "hello 1234 world"
push ax
call remove_digits
xor ax, ax
push ax
push ds
mov ax, offset aHello1234World ; "hello 1234 world"
push ax
push ds
push ax
push ax
call MESSAGEBOX
xor ax, ax
pop bp
retn 0Ah
WinMain endp
650
Here we see a difference between the so-called “near” pointers and the “far” pointers: another weird
artifact of segmented memory in 16-bit 8086.
You can read more about it here: 11.6 on page 986.
“near” pointers are those which point within the current data segment. Hence, the string_compare()
function takes only two 16-bit pointers, and accesses the data from the segment that DS points to (The
mov al, [bx] instruction actually works like mov al, ds:[bx] — DS is implicit here).
“far” pointers are those which may point to data in another memory segment.
Hence string_compare_far() takes the 16-bit pair as a pointer, loads the high part of it in the ES segment
register and accesses the data through it
(mov al, es:[bx]). “far” pointers are also used in my
MessageBox() win16 example: 3.34.2 on page 644. Indeed, the Windows kernel is not aware which data
segment to use when accessing text strings, so it need the complete information.
The reason for this distinction is that a compact program may use just one 64kb data segment, so it
doesn’t need to pass the high part of the address, which is always the same. A bigger program may use
several 64kb data segments, so it needs to specify the segment of the data each time.
It’s the same story for code segments. A compact program may have all executable code within one
64kb-segment, then all functions in it will be called using the CALL NEAR instruction, and the code flow
will be returned using RETN. But if there are several code segments, then the address of the function is to
be specified by a pair, it is to be called using the CALL FAR instruction, and the code flow is to be returned
using RETF.
This is what is set in the compiler by specifying “memory model”.
The compilers targeting MS-DOS and Win16 have specific libraries for each memory model: they differ by
pointer types for code and data.
3.34.6 Example #6
#include <time.h>
#include <stdio.h>
char strbuf[256];

LPSTR lpCmdLine,
int nCmdShow )
{
struct tm *t;
time_t unix_time;
t=localtime (&unix_time);
sprintf (strbuf, "%04d-%02d-%02d %02d:%02d:%02d", t->tm_year+1900, t->tm_mon, t->⤦

Ç tm_mday,
t->tm_hour, t->tm_min, t->tm_sec);
MessageBox (NULL, strbuf, "caption", MB_OK);

return 0;
};
WinMain proc near
var_4 = word ptr -4

var_2 = word ptr -2
push bp
mov bp, sp
push ax
push ax
xor ax, ax
651
call time_
mov [bp+var_4], ax ; low part of UNIX time
mov [bp+var_2], dx ; high part of UNIX time
lea ax, [bp+var_4] ; take a pointer of high part
call localtime_
mov bx, ax ; t
push word ptr [bx] ; second
push word ptr [bx+2] ; minute
push word ptr [bx+4] ; hour
push word ptr [bx+6] ; day
push word ptr [bx+8] ; month
mov ax, [bx+0Ah] ; year
add ax, 1900
push ax
mov ax, offset a04d02d02d02d02 ; "%04d-%02d-%02d %02d:%02d:%02d"
push ax
mov ax, offset strbuf
push ax
call sprintf_
add sp, 10h
xor ax, ax ; NULL
push ax
push ds
push ax
push ds
push ax
xor ax, ax ; MB_OK
push ax
call MESSAGEBOX
xor ax, ax
mov sp, bp
pop bp
retn 0Ah
WinMain endp
UNIX time is a 32-bit value, so it is returned in the DX:AX register pair and stored in two local 16-bit
variables. Then a pointer to the pair is passed to the localtime() function. The localtime() function
has a struct tm allocated somewhere in the guts of the C library, so only a pointer to it is returned.
By the way, this also implies that the function cannot be called again until its results are used.
For the time() and localtime() functions, a Watcom calling convention is used here: the first four
arguments are passed in the AX, DX, BX and CX, registers, and the rest arguments are via the stack.
The functions using this convention are also marked by underscore at the end of their name.
sprintf() does not use the PASCAL calling convention, nor the Watcom one,
so the arguments are passed in the normal cdecl way ( 6.1.1 on page 725).
Global variables
This is the same example, but now these variables are global:
#include <time.h>
#include <stdio.h>
char strbuf[256];
struct tm *t;
time_t unix_time;

LPSTR lpCmdLine,
int nCmdShow )
{
652
t=localtime (&unix_time);
sprintf (strbuf, "%04d-%02d-%02d %02d:%02d:%02d", t->tm_year+1900, t->tm_mon, t->⤦

Ç tm_mday,
t->tm_hour, t->tm_min, t->tm_sec);
MessageBox (NULL, strbuf, "caption", MB_OK);

return 0;
};
unix_time_low dw 0
unix_time_high dw 0
t dw 0
WinMain proc near

push bp
mov bp, sp
xor ax, ax
call time_
mov unix_time_low, ax
mov unix_time_high, dx
mov ax, offset unix_time_low
call localtime_
mov bx, ax
mov t, ax ; will not be used in future...
push word ptr [bx] ; seconds
push word ptr [bx+2] ; minutes
push word ptr [bx+4] ; hour
push word ptr [bx+6] ; day
push word ptr [bx+8] ; month
mov ax, [bx+0Ah] ; year
add ax, 1900
push ax
mov ax, offset a04d02d02d02d02 ; "%04d-%02d-%02d %02d:%02d:%02d"
push ax
push ax
call sprintf_
add sp, 10h
xor ax, ax ; NULL
push ax
push ds
push ax
push ds
push ax
xor ax, ax ; MB_OK
push ax
call MESSAGEBOX
pop bp
retn 0Ah
WinMain endp
t is not to be used, but the compiler emitted the code which stores the value.
Because it is not sure, maybe that value will eventually be used in some other module.
653
Chapter 4
Java
4.1 Java
4.1.1 Introduction
There are some well-known decompilers for Java (or JVM bytecode in general) 1 .
The reason is the decompilation of JVM-bytecode is somewhat easier than for lower level x86 code:
• There is much more information about the data types.
• The JVM memory model is much more rigorous and outlined.
• The Java compiler don’t do any optimizations (the JVM JIT2 does them at runtime), so the bytecode
in the class files is usually pretty readable.
When can the knowledge of JVM be useful?
• Quick-and-dirty patching tasks of class files without the need to recompile the decompiler’s results.
• Analyzing obfuscated code.
• Analyzing of a code generated by newer Java compiler for which no updated decompiler exists yet.
• Building your own obfuscator.
• Building a compiler codegenerator (back-end) targeting JVM (like Scala, Clojure, etc. 3 ).
Let’s start with some simple pieces of code. JDK 1.7 is used everywhere, unless mentioned otherwise.
This is the command used to decompile class files everywhere:
javap -c -verbose.
This is the book I used while preparing all examples: [Tim Lindholm, Frank Yellin, Gilad Bracha, Alex
Buckley, The Java(R) Virtual Machine Specification / Java SE 7 Edition] 4 .
4.1.2 Returning a value

Probably the simplest Java function is the one which returns some value.
Oh, and we must keep in mind that there are no “free” functions in Java in common sense, they are
“methods”.
Each method is related to some class, so it’s not possible to define a method outside of a class.
But we’ll call them “functions” anyway, for simplicity.
public class ret
{
public static int main(String[] args)
{
return 0;
1 For example, JAD: http://varaneckas.com/jad/
2 Just-In-Time compilation
3 Full list: http://en.wikipedia.org/wiki/List_of_JVM_languages
4 Also available as https://docs.oracle.com/javase/specs/jvms/se7/jvms7.pdf; http://docs.oracle.com/javase/specs/
jvms/se7/html/
654
}
}
Let’s compile it:

javac ret.java
…and decompile it using the standard Java utility:

javap -c -verbose ret.class
And we get:
Listing 4.1: JDK 1.7 (excerpt)

public static int main(java.lang.String[]);
flags: ACC_PUBLIC, ACC_STATIC
Code:
stack=1, locals=1, args_size=1
0: iconst_0
1: ireturn
The Java developers decided that 0 is one of the busiest constants in programming, so there is a separate
short one-byte iconst_0 instruction which pushes 0
5
. There are also iconst_1 (which pushes 1), iconst_2, etc., up to iconst_5.
There is also iconst_m1 which pushes -1.
The stack is used in JVM for passing data to called functions and also for return values. So iconst_0
pushes 0 into the stack. ireturn returns an integer value (i in name means integer) from the TOS6 .
Let’s rewrite our example slightly, now we return 1234:
public class ret
{
{
return 1234;
}
}
…we get:
Listing 4.2: JDK 1.7 (excerpt)

Code:
0: sipush 1234
3: ireturn
sipush (short integer) pushes 1234 into the stack. short in name implies a 16-bit value is to be pushed.
The number 1234 indeed fits well in a 16-bit value.
What about larger values?
public class ret
{
{
return 12345678;
}
}
5 Just like in MIPS, where a separate register for zero constant exists: 1.5.4 on page 25.
6 Top of Stack
655
Listing 4.3: Constant pool
...
#2 = Integer 12345678
...

Code:
0: ldc #2 // int 12345678
2: ireturn
It’s not possible to encode a 32-bit number in a JVM instruction opcode, the developers didn’t leave such
possibility.
So the 32-bit number 12345678 is stored in so called “constant pool” which is, let’s say, the library of
most used constants (including strings, objects, etc.).
This way of passing constants is not unique to JVM.
MIPS, ARM and other RISC CPUs also can’t encode a 32-bit number in a 32-bit opcode, so the RISC CPU
code (including MIPS and ARM) has to construct the value in several steps, or to keep it in the data
segment: 1.39.3 on page 438, 1.40.1 on page 441.
MIPS code also traditionally has a constant pool, named “literal pool”, the segments are called “.lit4” (for
32-bit single precision floating point number constants) and “.lit8” (for 64-bit double precision floating
point number constants).
Let’s try some other data types!
Boolean:
public class ret
{
public static boolean main(String[] args)
{
return true;
}
}
public static boolean main(java.lang.String[]);

Code:
0: iconst_1
1: ireturn
This JVM bytecode is no different from one returning integer 1.

32-bit data slots in the stack are also used here for boolean values, like in C/C++.
But one could not use returned boolean value as integer or vice versa — type information is stored in the
class file and checked at runtime.
It’s the same story with a 16-bit short:
public class ret
{
public static short main(String[] args)
{
return 1234;
}
}
public static short main(java.lang.String[]);

Code:
0: sipush 1234
3: ireturn
656
…and char!
public class ret
{
public static char main(String[] args)
{
return 'A';
}
}
public static char main(java.lang.String[]);

Code:
0: bipush 65
2: ireturn
bipush means “push byte”. Needless to say that a char in Java is 16-bit UTF-16 character, and it’s
equivalent to short, but the ASCII code of the “A” character is 65, and it’s possible to use the instruction
for pushing a byte in the stack.
Let’s also try a byte:
public class retc
{
public static byte main(String[] args)
{
return 123;
}
}
public static byte main(java.lang.String[]);

Code:
0: bipush 123
2: ireturn
One may ask, why bother with a 16-bit short data type which internally works as a 32-bit integer?
Why use a char data type if it is the same as a short data type?
The answer is simple: for data type control and source code readability.
A char may essentially be the same as a short, but we quickly grasp that it’s a placeholder for an UTF-16
character, and not for some other integer value.
When using short, we show everyone that the variable’s range is limited by 16 bits.
It’s a very good idea to use the boolean type where needed to, instead of the C-style int.
There is also a 64-bit integer data type in Java:
public class ret3
{
public static long main(String[] args)
{
return 1234567890123456789L;
}
}

...
#2 = Long 1234567890123456789l
...
public static long main(java.lang.String[]);

Code:
657
0: ldc2_w #2 // long 1234567890123456789l
3: lreturn
The 64-bit number is also stored in a constant pool, ldc2_w loads it and lreturn (long return) returns it.
The ldc2_w instruction is also used to load double precision floating point numbers (which also occupy 64
bits) from a constant pool:
public class ret
{
public static double main(String[] args)
{
return 123.456d;
}
}

...
#2 = Double 123.456d
...
public static double main(java.lang.String[]);

Code:
0: ldc2_w #2 // double 123.456d
3: dreturn
dreturn stands for “return double”.

And finally, a single precision floating point number:
public class ret
{
public static float main(String[] args)
{
return 123.456f;
}
}

...
#2 = Float 123.456f
...
public static float main(java.lang.String[]);

Code:
0: ldc #2 // float 123.456f
2: freturn
The ldc instruction used here is the same one as for loading 32-bit integer numbers from a constant pool.
freturn stands for “return float”.
Now what about function that return nothing?
public class ret
{
public static void main(String[] args)
{
return;
}
}
658
public static void main(java.lang.String[]);
Code:
0: return
This means that the return instruction is used to return control without returning an actual value.
Knowing all this, it’s very easy to deduce the function’s (or method’s) returning type from the last instruc-
tion.
4.1.3 Simple calculating functions

Let’s continue with a simple calculating functions.
public class calc
{
public static int half(int a)
{
return a/2;
}
}
Here’s the output when the iconst_2 instruction is used:

public static int half(int);
Code:
0: iload_0
1: iconst_2
2: idiv
3: ireturn
iload_0 takes the zeroth function argument and pushes it to the stack.
iconst_2 pushes 2 in the stack. After the execution of these two instructions, this is how stack looks like:
+---+
TOS ->| 2 |
+---+
| a |
+---+
idiv just takes the two values at the TOS, divides one by the other and leaves the result at TOS:
+--------+
TOS ->| result |
+--------+
ireturn takes it and returns.

Let’s proceed with double precision floating point numbers:
public class calc
{
public static double half_double(double a)
{
return a/2.0;
}
}

...
#2 = Double 2.0d
...
659
public static double half_double(double);
Code:
0: dload_0
1: ldc2_w #2 // double 2.0d
4: ddiv
5: dreturn
It’s the same, but the ldc2_w instruction is used to load the constant 2.0 from the constant pool.
Also, the other three instructions have the d prefix, meaning they work with double data type values.
Let’s now use a function with two arguments:
public class calc
{
public static int sum(int a, int b)
{
return a+b;
}
}
public static int sum(int, int);

Code:
0: iload_0
1: iload_1
2: iadd
3: ireturn
iload_0 loads the first function argument (a), iload_1—second (b).

Here is the stack after the execution of both instructions:
+---+
TOS ->| b |
+---+
| a |
+---+
iadd adds the two values and leaves the result at TOS:
+--------+
TOS ->| result |
+--------+
Let’s extend this example to the long data type:

public static long lsum(long a, long b)
{
return a+b;
}
…we got:
public static long lsum(long, long);
Code:
0: lload_0
1: lload_2
2: ladd
3: lreturn
The second lload instruction takes the second argument from the 2nd slot.
That’s because a 64-bit long value occupies exactly two 32-bit slots.
Slightly more advanced example:
660
public class calc
{
public static int mult_add(int a, int b, int c)
{
return a*b+c;
}
}
public static int mult_add(int, int, int);

Code:
0: iload_0
1: iload_1
2: imul
3: iload_2
4: iadd
5: ireturn
The first step is multiplication. The product is left at the TOS:

+---------+
TOS ->| product |
+---------+
iload_2 loads the third argument (c) in the stack:

+---------+
TOS ->| c |
+---------+
| product |
+---------+
Now the iadd instruction can add the two values.
4.1.4 JVM memory model

x86 and other low-level environments use the stack for argument passing and as a local variables storage.
JVM is slightly different.
It has:
• Local variable array (LVA7 ). Used as storage for incoming function arguments and local variables.
Instructions like iload_0 load values from it.
istore stores values in it. At the beginning the function arguments are stored: starting at 0 or at 1
(if the zeroth argument is occupied by this pointer).
Then the local variables are allocated.
Each slot has size of 32-bit.
Hence, values of long and double data types occupy two slots.
• Operand stack (or just “stack”). It’s used for computations and passing arguments while calling other
functions.
Unlike low-level environments like x86, it’s not possible to access the stack without using instructions
which explicitly pushes or pops values to/from it.
• Heap. It is used as storage for objects and arrays.
These 3 areas are isolated from each other.
7 (Java) Local Variable Array
661
4.1.5 Simple function calling
Math.random() returns a pseudorandom number in range of [0.0 …1.0], but let’s say that for some reason
we need to devise a function that returns a number in range of [0.0 …0.5]:
public class HalfRandom
{
public static double f()
{
return Math.random()/2;
}
}

...
#2 = Methodref #18.#19 // java/lang/Math.random:()D
#3 = Double 2.0d
...
#12 = Utf8 ()D
...
#18 = Class #22 // java/lang/Math
#19 = NameAndType #23:#12 // random:()D
#22 = Utf8 java/lang/Math
#23 = Utf8 random
public static double f();

Code:
0: invokestatic #2 // Method java/lang/Math.random:()D
3: ldc2_w #3 // double 2.0d
6: ddiv
7: dreturn
invokestatic calls the Math.random() function and leaves the result at the TOS.
Then the result is divided by 2.0 and returned.
But how is the function name encoded?
It’s encoded in the constant pool using a Methodref expression.
It defines the class and method names.
The first field of Methodref points to a Class expression which, in turn, points to the usual text string
(“java/lang/Math”).
The second Methodref expression points to a NameAndType expression which also has two links to the
strings.
The first string is “random”, which is the name of the method.
The second string is “()D”, which encodes the function’s type. It means that it returns a double value
(hence the D in the string).
This is the way 1) JVM can check data for type correctness; 2) Java decompilers can restore data types
from a compiled class file.
Now let’s try the “Hello, world!” example:
public class HelloWorld
{
{
System.out.println("Hello, World");
}
}

...
#2 = Fieldref #16.#17 // java/lang/System.out:Ljava/io/PrintStream;
662
#3 = String #18 // Hello, World
#4 = Methodref #19.#20 // java/io/PrintStream.println:(Ljava/lang/String;)V
...
#16 = Class #23 // java/lang/System
#17 = NameAndType #24:#25 // out:Ljava/io/PrintStream;
#18 = Utf8 Hello, World
#19 = Class #26 // java/io/PrintStream
#20 = NameAndType #27:#28 // println:(Ljava/lang/String;)V
...
#23 = Utf8 java/lang/System
#24 = Utf8 out
#25 = Utf8 Ljava/io/PrintStream;
#26 = Utf8 java/io/PrintStream
#27 = Utf8 println
#28 = Utf8 (Ljava/lang/String;)V
...

Code:
0: getstatic #2 // Field java/lang/System.out:Ljava/io/PrintStream;
3: ldc #3 // String Hello, World
5: invokevirtual #4 // Method java/io/PrintStream.println:(Ljava/lang/String;)V
8: return
ldc at offset 3 takes a pointer to the “Hello, World” string in the constant pool and pushes in the stack.
It’s called a reference in the Java world, but it’s rather a pointer, or an address
8
.
The familiar invokevirtual instruction takes the information about the println function (or method)
from the constant pool and calls it.
As we may know, there are several println methods, one for each data type.
Our case is the version of println intended for the String data type.
But what about the first getstatic instruction?
This instruction takes a reference (or address of) a field of the object System.out and pushes it in the
stack.
This value is acts like the this pointer for the println method.
Thus, internally, the println method takes two arguments for input: 1) this, i.e., a pointer to an object;
2) the address of the “Hello, World” string.
Indeed, println() is called as a method within an initialized System.out object.
For convenience, the javap utility writes all this information in the comments.
4.1.6 Calling beep()

This is a simple calling of two functions without arguments:
{
java.awt.Toolkit.getDefaultToolkit().beep();
};

Code:
0: invokestatic #2 // Method java/awt/Toolkit.getDefaultToolkit:()Ljava/awt/⤦
Ç Toolkit;
3: invokevirtual #3 // Method java/awt/Toolkit.beep:()V
6: return
8 About difference in pointers and reference’s in C++ see: 3.21.3 on page 558.
663
First invokestatic at offset 0 calls
java.awt.Toolkit.getDefaultToolkit(), which returns a reference to an object of class Toolkit.
The invokevirtual instruction at offset 3 calls the beep() method of this class.
4.1.7 Linear congruential PRNG

Let’s try a simple pseudorandom numbers generator, which we already considered once in the book ( 1.29
on page 336):
public class LCG
{
public static int rand_state;
public void my_srand (int init)

{
rand_state=init;
}
public static int RNG_a=1664525;

public static int RNG_c=1013904223;
public int my_rand ()

{
}
}
There are couple of class fields which are initialized at start.

But how? In javap output we can find the class constructor:
static {};
flags: ACC_STATIC
Code:
0: ldc #5 // int 1664525
2: putstatic #3 // Field RNG_a:I
5: ldc #6 // int 1013904223
7: putstatic #4 // Field RNG_c:I
10: return
That’s the way variables are initialized.

RNG_a occupies the 3rd slot in the class and RNG_c—4th, and putstatic puts the constants there.
The my_srand() function just stores the input value in rand_state:
public void my_srand(int);
flags: ACC_PUBLIC
Code:
0: iload_1
1: putstatic #2 // Field rand_state:I
4: return
iload_1 takes the input value and pushes it into stack. But why not iload_0?
It’s because this function may use fields of the class, and so this is also passed to the function as a zeroth
argument.
The field rand_state occupies the 2nd slot in the class, so putstatic copies the value from the TOS into
the 2nd slot.
Now my_rand():
public int my_rand();
flags: ACC_PUBLIC
Code:
664
0: getstatic #2 // Field rand_state:I
3: getstatic #3 // Field RNG_a:I
6: imul
13: getstatic #4 // Field RNG_c:I
16: iadd
23: sipush 32767
26: iand
27: ireturn
It just loads all the values from the object’s fields, does the operations and updates rand_state’s value
using the putstatic instruction.
At offset 20, rand_state is reloaded again (because it has been dropped from the stack before, by
putstatic).
This looks like non-efficient code, but be sure, the JVM is usually good enough to optimize such things
really well.
4.1.8 Conditional jumps

Now let’s proceed to conditional jumps.
public class abs
{
public static int abs(int a)
{
if (a<0)
return -a;
return a;
}
}
public static int abs(int);

Code:
0: iload_0
1: ifge 7
4: iload_0
5: ineg
6: ireturn
7: iload_0
8: ireturn
ifge jumps to offset 7 if the value at TOS is greater or equal to 0.

Don’t forget, any ifXX instruction pops the value (to be compared) from the stack.
ineg just negates value at TOS.
Another example:
public static int min (int a, int b)
{
if (a>b)
return b;
return a;
}
We get:
public static int min(int, int);
Code:
0: iload_0
665
1: iload_1
2: if_icmple 7
5: iload_1
6: ireturn
7: iload_0
8: ireturn
if_icmple pops two values and compares them. If the second one is lesser than (or equal to) the first, a
jump to offset 7 is performed.
When we define max() function …
public static int max (int a, int b)
{
if (a>b)
return a;
return b;
}
…the resulting code is the same, but the last two iload instructions (at offsets 5 and 7) are swapped:
public static int max(int, int);
Code:
0: iload_0
1: iload_1
2: if_icmple 7
5: iload_0
6: ireturn
7: iload_1
8: ireturn
A more advanced example:

public class cond
{
public static void f(int i)
{
if (i<100)
System.out.print("<100");
if (i==100)
System.out.print("==100");
if (i>100)
System.out.print(">100");
if (i==0)
System.out.print("==0");
}
}
public static void f(int);

Code:
0: iload_0
1: bipush 100
3: if_icmpge 14
9: ldc #3 // String <100
11: invokevirtual #4 // Method java/io/PrintStream.print:(Ljava/lang/String;)V
14: iload_0
15: bipush 100
17: if_icmpne 28
23: ldc #5 // String ==100
28: iload_0
29: bipush 100
31: if_icmple 42
666
37: ldc #6 // String >100
42: iload_0
43: ifne 54
49: ldc #7 // String ==0
54: return
if_icmpge pops two values and compares them. If the second one is larger or equal than the first, a jump
to offset 14 is performed.
if_icmpne and if_icmple work just the same, but implement different conditions.
There is also a ifne instruction at offset 43.
Its name is misnomer, it would’ve be better to name it ifnz (jump if the value at TOS is not zero).
And that is what it does: it jumps to offset 54 if the input value is not zero.
If zero,the execution flow proceeds to offset 46, where the “==0” string is printed.
N.B.: JVM has no unsigned data types, so the comparison instructions operate only on signed integer
values.
4.1.9 Passing arguments

Let’s extend our min()/max() example:
public class minmax
{
public static int min (int a, int b)
{
if (a>b)
return b;
return a;
}
public static int max (int a, int b)

{
if (a>b)
return a;
return b;
}

{
int a=123, b=456;
int max_value=max(a, b);
int min_value=min(a, b);
System.out.println(min_value);
System.out.println(max_value);
}
}
Here is main() function code:

Code:
0: bipush 123
2: istore_1
3: sipush 456
6: istore_2
7: iload_1
8: iload_2
9: invokestatic #2 // Method max:(II)I
12: istore_3
13: iload_1
14: iload_2
667
15: invokestatic #3 // Method min:(II)I
18: istore 4
23: iload 4
25: invokevirtual #5 // Method java/io/PrintStream.println:(I)V
31: iload_3
35: return
Arguments are passed to the other function in the stack, and the return value is left on TOS.
4.1.10 Bitfields
All bit-wise operations work just like in any other ISA:
public static int set (int a, int b)
{
return a | 1<<b;
}
public static int clear (int a, int b)

{
return a & (~(1<<b));
}
public static int set(int, int);

Code:
0: iload_0
1: iconst_1
2: iload_1
3: ishl
4: ior
5: ireturn
public static int clear(int, int);

Code:
0: iload_0
1: iconst_1
2: iload_1
3: ishl
4: iconst_m1
5: ixor
6: iand
7: ireturn
iconst_m1 loads −1 in the stack, it’s the same as the 0xFFFFFFFF number.
XORing with 0xFFFFFFFF has the same effect of inverting all bits ( 2.6 on page 456).
Let’s extend all data types to 64-bit long:
public static long lset (long a, int b)
{
return a | 1<<b;
}
public static long lclear (long a, int b)

{
return a & (~(1<<b));
}
public static long lset(long, int);

Code:
668
0: lload_0
1: iconst_1
2: iload_2
3: ishl
4: i2l
5: lor
6: lreturn
public static long lclear(long, int);

Code:
0: lload_0
1: iconst_1
2: iload_2
3: ishl
4: iconst_m1
5: ixor
6: i2l
7: land
8: lreturn
The code is the same, but instructions with l prefix are used, which operate on 64-bit values.
Also, the second argument type of the function is still int, and when the 32-bit value in it needs to be
promoted to 64-bit value the i2l instruction is used, which essentially extend the value of an integer type
to a long one.
4.1.11 Loops
public class Loop
{
{
for (int i = 1; i <= 10; i++)
{
System.out.println(i);
}
}
}

Code:
0: iconst_1
1: istore_1
2: iload_1
3: bipush 10
5: if_icmpgt 21
11: iload_1
15: iinc 1, 1
18: goto 2
21: return
iconst_1 loads 1 into TOS, istore_1 stores it in the LVA at slot 1.

Why not the zeroth slot? Because the main() function has one argument (array of String) and a pointer
to it (or reference) is now in the zeroth slot.
So, the i local variable will always be in 1st slot.
Instructions at offsets 3 and 5 compare i with 10.
If i is larger, execution flow passes to offset 21, where the function ends.
If it’s not, println is called.
669
i is then reloaded at offset 11, for println.
By the way, we call the println method for an integer, and we see this in the comments: “(I)V” (I means
integer and V means the return type is void).
When println finishes, i is incremented at offset 15.
The first operand of the instruction is the number of a slot (1), the second is the number (1) to add to the
variable.
goto is just GOTO, it jumps to the beginning of the loop’s body offset 2.
Let’s proceed with a more complex example:
public class Fibonacci
{
{
int limit = 20, f = 0, g = 1;
for (int i = 1; i <= limit; i++)

{
f = f + g;
g = f - g;
System.out.println(f);
}
}
}

Code:
0: bipush 20
2: istore_1
3: iconst_0
4: istore_2
5: iconst_1
6: istore_3
7: iconst_1
8: istore 4
10: iload 4
12: iload_1
13: if_icmpgt 37
16: iload_2
17: iload_3
18: iadd
19: istore_2
20: iload_2
21: iload_3
22: isub
23: istore_3
27: iload_2
31: iinc 4, 1
34: goto 10
37: return
Here is a map of the LVA slots:

• 0 — the sole argument of main()
• 1 — limit, always contains 20
• 2—f
• 3—g
• 4—i
We can see that the Java compiler allocates variables in LVA slots in the same order they were declared
in the source code.
670
There are separate istore instructions for accessing slots 0, 1, 2 and 3, but not for 4 and larger, so there
is istore with an additional operand at offset 8 which takes the slot number as an operand.
It’s the same with iload at offset 10.
But isn’t it dubious to allocate another slot for the limit variable, which always contains 20 (so it’s a
constant in essence), and reload its value so often?
JVM JIT compiler is usually good enough to optimize such things.
Manual intervention in the code is probably not worth it.
4.1.12 switch()
The switch() statement is implemented with the tableswitch instruction:
public static void f(int a)
{
switch (a)
{
case 0: System.out.println("zero"); break;
case 1: System.out.println("one\n"); break;
case 2: System.out.println("two\n"); break;
case 3: System.out.println("three\n"); break;
case 4: System.out.println("four\n"); break;
default: System.out.println("something unknown\n"); break;
};
}
As simple, as possible:
public static void f(int);
Code:
0: iload_0
1: tableswitch { // 0 to 4
0: 36
1: 47
2: 58
3: 69
4: 80
default: 91
}
39: ldc #3 // String zero
44: goto 99
50: ldc #5 // String one\n
55: goto 99
61: ldc #6 // String two\n
66: goto 99
72: ldc #7 // String three\n
77: goto 99
83: ldc #8 // String four\n
88: goto 99
94: ldc #9 // String something unknown\n
99: return
671
4.1.13 Arrays
Simple example
Let’s first create an array of 10 integers and fill it:

{
int a[]=new int[10];
for (int i=0; i<10; i++)
a[i]=i;
dump (a);
}

Code:
0: bipush 10
2: newarray int
4: astore_1
5: iconst_0
6: istore_2
7: iload_2
8: bipush 10
10: if_icmpge 23
13: aload_1
14: iload_2
15: iload_2
16: iastore
17: iinc 2, 1
20: goto 7
23: aload_1
24: invokestatic #4 // Method dump:([I)V
27: return
The newarray instruction creates an array object of 10 int elements.

The array’s size is set with bipush and left at TOS.
The array’s type is set in newarray instruction’s operand.
After newarray’s execution, a reference (or pointer) to the newly created array in the heap is left at the
TOS.
astore_1 stores the reference to the 1st slot in LVA.
The second part of the main() function is the loop which stores i into the corresponding array element.
aload_1 gets a reference of the array and places it in the stack.
iastore then stores the integer value from the stack in the array, reference of which is currently in TOS.
The third part of the main() function calls the dump() function.
An argument for it is prepared by aload_1 (offset 23).
Now let’s proceed to the dump() function:
public static void dump(int a[])
{
for (int i=0; i<a.length; i++)
System.out.println(a[i]);
}
public static void dump(int[]);

Code:
0: iconst_0
1: istore_1
2: iload_1
3: aload_0
672
4: arraylength
5: if_icmpge 23
11: aload_0
12: iload_1
13: iaload
17: iinc 1, 1
20: goto 2
23: return
The incoming reference to the array is in the zeroth slot.

The a.length expression in the source code is converted to an arraylength instruction: it takes a refer-
ence to the array and leaves the array size at TOS.
iaload at offset 13 is used to load array elements, it requires an array reference to be present in the stack
(prepared by aload_0 at 11), and also an index (prepared by iload_1 at offset 12).
Needless to say, instructions prefixed with a may be mistakenly comprehended as array instructions.
It’s not correct. These instructions works with references to objects.
And arrays and strings are objects too.
Summing elements of array
Another example:
public class ArraySum
{
public static int f (int[] a)
{
int sum=0;
for (int i=0; i<a.length; i++)
sum=sum+a[i];
return sum;
}
}
public static int f(int[]);

Code:
0: iconst_0
1: istore_1
2: iconst_0
3: istore_2
4: iload_2
5: aload_0
6: arraylength
7: if_icmpge 22
10: iload_1
11: aload_0
12: iload_2
13: iaload
14: iadd
15: istore_1
16: iinc 2, 1
19: goto 4
22: iload_1
23: ireturn
LVA slot 0 contains a reference to the input array.

LVA slot 1 contains the local variable sum.
The only argument of the main() function is an array too
We’ll be using the only argument of the main() function, which is an array of strings:
673
public class UseArgument
{
{
System.out.print("Hi, ");
System.out.print(args[1]);
System.out.println(". How are you?");
}
}
The zeroth argument is the program’s name (like in C/C++, etc.), so the 1st argument supplied by the
user is 1st.
Code:
3: ldc #3 // String Hi,
11: aload_0
12: iconst_1
13: aaload
20: ldc #5 // String . How are you?
25: return
aload_0 at 11 loads a reference of the zeroth LVA slot (1st and only main() argument).
iconst_1 and aaload at 12 and 13 take a reference to the first (counting at 0) element of array.
The reference to the string object is at TOS at offset 14, and it is taken from there by println method.
Pre-initialized array of strings
class Month
{
public static String[] months =
{
"January",
"February",
"March",
"April",
"May",
"June",
"July",
"August",
"September",
"October",
"November",
"December"
};
public String get_month (int i)

{
return months[i];
};
}
The get_month() function is simple:

public java.lang.String get_month(int);
flags: ACC_PUBLIC
Code:
0: getstatic #2 // Field months:[Ljava/lang/String;
674
3: iload_1
4: aaload
5: areturn
aaload operates on an array of references.

Java String are objects, so the a-instructions are used to operate on them.
areturn returns a reference to a String object.
How is the months[] array initialized?
static {};
flags: ACC_STATIC
Code:
0: bipush 12
2: anewarray #3 // class java/lang/String
5: dup
6: iconst_0
7: ldc #4 // String January
9: aastore
10: dup
11: iconst_1
12: ldc #5 // String February
14: aastore
15: dup
16: iconst_2
17: ldc #6 // String March
19: aastore
20: dup
21: iconst_3
22: ldc #7 // String April
24: aastore
25: dup
26: iconst_4
27: ldc #8 // String May
29: aastore
30: dup
31: iconst_5
32: ldc #9 // String June
34: aastore
35: dup
36: bipush 6
38: ldc #10 // String July
40: aastore
41: dup
42: bipush 7
44: ldc #11 // String August
46: aastore
47: dup
48: bipush 8
50: ldc #12 // String September
52: aastore
53: dup
54: bipush 9
56: ldc #13 // String October
58: aastore
59: dup
60: bipush 10
62: ldc #14 // String November
64: aastore
65: dup
66: bipush 11
68: ldc #15 // String December
70: aastore
71: putstatic #2 // Field months:[Ljava/lang/String;
74: return
anewarray creates a new array of references (hence a prefix).
675
The object’s type is defined in the anewarray’s operand, it is the
“java/lang/String” string.
The bipush 12 before anewarray sets the array’s size.
We see here a new instruction for us: dup.
It’s a standard instruction in stack computers (including the Forth programming language) which just
duplicates the value at TOS.
By the way, FPU 80x87 is also a stack computer and it has similar instruction – FDUP.
It is used here to duplicate a reference to an array, because the aastore instruction pops the reference
to array from the stack, but subsequent aastore will need it again.
The Java compiler concluded that it’s better to generate a dup instead of generating a getstatic instruc-
tion before each array store operation (i.e., 11 times).
aastore puts a reference (to string) into the array at an index which is taken from TOS.
Finally, putstatic puts reference to the newly created array into the second field of our object, i.e.,
months field.
Variadic functions
Variadic functions actually use arrays:

public static void f(int... values)
{
for (int i=0; i<values.length; i++)
System.out.println(values[i]);
}

{
f (1,2,3,4,5);
}
public static void f(int...);

flags: ACC_PUBLIC, ACC_STATIC, ACC_VARARGS
Code:
0: iconst_0
1: istore_1
2: iload_1
3: aload_0
4: arraylength
5: if_icmpge 23
11: aload_0
12: iload_1
13: iaload
17: iinc 1, 1
20: goto 2
23: return
f() just takes an array of integers using aload_0 at offset 3.

Then it gets the array’s size, etc.
Code:
0: iconst_5
1: newarray int
3: dup
4: iconst_0
5: iconst_1
6: iastore
7: dup
676
8: iconst_1
9: iconst_2
10: iastore
11: dup
12: iconst_2
13: iconst_3
14: iastore
15: dup
16: iconst_3
17: iconst_4
18: iastore
19: dup
20: iconst_4
21: iconst_5
22: iastore
23: invokestatic #4 // Method f:([I)V
26: return
The array is constructed in main() using the newarray instruction, then it’s filled, and f() is called.
Oh, by the way, array object is not destroyed at the end of main().
There are no destructors in Java at all, because the JVM has a garbage collector which does this automat-
ically, when it feels it needs to.
What about the format() method?
It takes two arguments at input: a string and an array of objects:
public PrintStream format(String format, Object... args)
( http://docs.oracle.com/javase/tutorial/java/data/numberformat.html )
Let’s see:
{
int i=123;
double d=123.456;
System.out.format("int: %d double: %f.%n", i, d);
}

Code:
0: bipush 123
2: istore_1
3: ldc2_w #2 // double 123.456d
6: dstore_2
10: ldc #5 // String int: %d double: %f.%n
12: iconst_2
13: anewarray #6 // class java/lang/Object
16: dup
17: iconst_0
18: iload_1
19: invokestatic #7 // Method java/lang/Integer.valueOf:(I)Ljava/lang/Integer;
22: aastore
23: dup
24: iconst_1
25: dload_2
26: invokestatic #8 // Method java/lang/Double.valueOf:(D)Ljava/lang/Double;
29: aastore
30: invokevirtual #9 // Method java/io/PrintStream.format:(Ljava/lang/String;[⤦
Ç Ljava/lang/Object;)Ljava/io/PrintStream;
33: pop
34: return
So values of the int and double types are first promoted to Integer and Double objects using the valueOf
methods.
677
The format() method needs objects of type Object at input, and since the Integer and Double classes
are derived from the root Object class, they suitable for elements in the input array.
On the other hand, an array is always homogeneous, i.e., it can’t hold elements of different types, which
makes it impossible to push int and double values in it.
An array of Object objects is created at offset 13, an Integer object is added to the array at offset 22,
and a Double object is added to the array at offset 29.
The penultimate pop instruction discards the element at TOS, so when return is executed, the stack
becomes empty (or balanced).
Two-dimensional arrays
Two-dimensional arrays in Java are just one-dimensional arrays of references to another one-dimensional
arrays.
Let’s create a two-dimensional array:
{
int[][] a = new int[5][10];
a[1][2]=3;
}

Code:
0: iconst_5
1: bipush 10
3: multianewarray #2, 2 // class "[[I"
7: astore_1
8: aload_1
9: iconst_1
10: aaload
11: iconst_2
12: iconst_3
13: iastore
14: return
It’s created using the multianewarray instruction: the object’s type and dimensionality are passed as
operands.
The array’s size (10*5) is left in stack (using the instructions iconst_5 and bipush).
A reference to row #1 is loaded at offset 10 (iconst_1 and aaload).
The column is chosen using iconst_2 at offset 11.
The value to be written is set at offset 12.
iastore at 13 writes the array’s element.
How it is an element accessed?
public static int get12 (int[][] in)
{
return in[1][2];
}
public static int get12(int[][]);

Code:
0: aload_0
1: iconst_1
2: aaload
3: iconst_2
4: iaload
5: ireturn
678
A Reference to the array’s row is loaded at offset 2, the column is set at offset 3, then iaload loads the
array’s element.
Three-dimensional arrays
Three-dimensional arrays are just one-dimensional arrays of references to one-dimensional arrays of ref-
erences to one-dimensional arrays.
{
int[][][] a = new int[5][10][15];
a[1][2][3]=4;
get_elem(a);
}

Code:
0: iconst_5
1: bipush 10
3: bipush 15
5: multianewarray #2, 3 // class "[[[I"
9: astore_1
10: aload_1
11: iconst_1
12: aaload
13: iconst_2
14: aaload
15: iconst_3
16: iconst_4
17: iastore
18: aload_1
19: invokestatic #3 // Method get_elem:([[[I)I
22: pop
23: return
Now it takes two aaload instructions to find right reference:

public static int get_elem (int[][][] a)
{
return a[1][2][3];
}
public static int get_elem(int[][][]);

Code:
0: aload_0
1: iconst_1
2: aaload
3: iconst_2
4: aaload
5: iconst_3
6: iaload
7: ireturn
Summary
Is it possible to do a buffer overflow in Java?

No, because the array’s length is always present in an array object, array bounds are controlled, and an
exception is to be raised in case of out-of-bounds access.
There are no multi-dimensional arrays in Java in the C/C++ sense, so Java is not very suited for fast
scientific computations.
679
4.1.14 Strings
First example
Strings are objects and are constructed in the same way as other objects (and arrays).
{
System.out.println("What is your name?");
String input = System.console().readLine();
System.out.println("Hello, "+input);
}

Code:
3: ldc #3 // String What is your name?
8: invokestatic #5 // Method java/lang/System.console:()Ljava/io/Console;
11: invokevirtual #6 // Method java/io/Console.readLine:()Ljava/lang/String;
14: astore_1
18: new #7 // class java/lang/StringBuilder
21: dup
22: invokespecial #8 // Method java/lang/StringBuilder."<init>":()V
25: ldc #9 // String Hello,
27: invokevirtual #10 // Method java/lang/StringBuilder.append:(Ljava/lang/String⤦
Ç ;)Ljava/lang/StringBuilder;
30: aload_1
34: invokevirtual #11 // Method java/lang/StringBuilder.toString:()Ljava/lang/⤦
Ç String;
40: return
The readLine() method is called at offset 11, a reference to string (which is supplied by the user) is then
stored at TOS.
At offset 14 the reference to string is stored in slot 1 of LVA.
The string the user entered is reloaded at offset 30 and concatenated with the “Hello, ” string using the
StringBuilder class.
The constructed string is then printed using println at offset 37.
Second example
Another example:
public class strings
{
public static char test (String a)
{
return a.charAt(3);
};
public static String concat (String a, String b)

{
return a+b;
}
}
public static char test(java.lang.String);

Code:
0: aload_0
680
1: iconst_3
2: invokevirtual #2 // Method java/lang/String.charAt:(I)C
5: ireturn
The string concatenation is performed using StringBuilder:

public static java.lang.String concat(java.lang.String, java.lang.String);
Code:
3: dup
7: aload_0
11: aload_1
Ç String;
18: areturn
Another example:
{
String s="Hello!";
int n=123;
System.out.println("s=" + s + " n=" + n);
}
And again, the strings are constructed using the StringBuilder class and its append method, then the
constructed string is passed to println:
Code:
0: ldc #2 // String Hello!
2: astore_1
3: bipush 123
5: istore_2
12: dup
16: ldc #6 // String s=
21: aload_1
25: ldc #8 // String n=
30: iload_2
31: invokevirtual #9 // Method java/lang/StringBuilder.append:(I)Ljava/lang/⤦
Ç StringBuilder;
Ç String;
40: return
4.1.15 Exceptions
Let’s rework our Month example ( 4.1.13 on page 674) a bit:
681
Listing 4.10: IncorrectMonthException.java
public class IncorrectMonthException extends Exception
{
private int index;
public IncorrectMonthException(int index)

{
this.index = index;
}
public int getIndex()
{
return index;
}
}
Listing 4.11: Month2.java

class Month2
{
public static String[] months =
{
"January",
"February",
"March",
"April",
"May",
"June",
"July",
"August",
"September",
"October",
"November",
"December"
};
public static String get_month (int i) throws IncorrectMonthException

{
if (i<0 || i>11)
throw new IncorrectMonthException(i);
return months[i];
};
public static void main (String[] args)

{
try
{
System.out.println(get_month(100));
}
catch(IncorrectMonthException e)
{
System.out.println("incorrect month index: "+ e.getIndex());
e.printStackTrace();
}
};
}
Essentially, IncorrectMonthException.class has just an object constructor and one getter method.
The IncorrectMonthException class is derived from Exception, so the IncorrectMonthException con-
structor first calls the constructor of the Exception class, then it puts incoming integer value into the sole
IncorrectMonthException class field:
public IncorrectMonthException(int);
flags: ACC_PUBLIC
Code:
0: aload_0
1: invokespecial #1 // Method java/lang/Exception."<init>":()V
4: aload_0
5: iload_1
682
6: putfield #2 // Field index:I
9: return
getIndex() is just a getter. A reference to IncorrectMonthException is passed in the zeroth LVA slot
(this), aload_0 takes it, getfield loads an integer value from the object, ireturn returns it.
public int getIndex();
flags: ACC_PUBLIC
Code:
0: aload_0
1: getfield #2 // Field index:I
4: ireturn
Now let’s take a look at get_month() in Month2.class:
Listing 4.12: Month2.class

public static java.lang.String get_month(int) throws IncorrectMonthException;
Code:
0: iload_0
1: iflt 10
4: iload_0
5: bipush 11
7: if_icmple 19
10: new #2 // class IncorrectMonthException
13: dup
14: iload_0
15: invokespecial #3 // Method IncorrectMonthException."<init>":(I)V
18: athrow
19: getstatic #4 // Field months:[Ljava/lang/String;
22: iload_0
23: aaload
24: areturn
iflt at offset 1 is if less than.

In case of invalid index, a new object is created using the new instruction at offset 10.
The object’s type is passed as an operand to the instruction (which is IncorrectMonthException).
Then its constructor is called, and index is passed via TOS (offset 15).
When the control flow is offset 18, the object is already constructed, so now the athrow instruction takes
a reference to the newly constructed object and signals to JVM to find the appropriate exception handler.
The athrow instruction doesn’t return the control flow here, so at offset 19 there is another basic block,
not related to exceptions business, where we can get from offset 7.
How do handlers work?
main() in Month2.class:
Listing 4.13: Month2.class

Code:
3: bipush 100
5: invokestatic #6 // Method get_month:(I)Ljava/lang/String;
11: goto 47
14: astore_1
21: dup
25: ldc #10 // String incorrect month index:
683
30: aload_1
31: invokevirtual #12 // Method IncorrectMonthException.getIndex:()I
34: invokevirtual #13 // Method java/lang/StringBuilder.append:(I)Ljava/lang/⤦
Ç StringBuilder;
Ç String;
43: aload_1
44: invokevirtual #15 // Method IncorrectMonthException.printStackTrace:()V
47: return
Exception table:
from to target type
0 11 14 Class IncorrectMonthException
Here is the Exception table, which defines that from offsets 0 to 11 (inclusive) an exception
IncorrectMonthException may happen, and if it does, the control flow is to be passed to offset 14.
Indeed, the main program ends at offset 11.
At offset 14 the handler starts. It’s not possible to get here, there are no conditional/unconditional jumps
to this area.
But JVM will transfer the execution flow here in case of an exception.
The very first astore_1 (at 14) takes the incoming reference to the exception object and stores it in LVA
slot 1.
Later, the getIndex() method (of this exception object) will be called at offset 31.
The reference to the current exception object is passed right before that (offset 30).
The rest of the code is does just string manipulation: first the integer value returned by getIndex() is
converted to string by the toString() method, then it’s concatenated with the “incorrect month index: ”
text string (like we saw before), then println() and printStackTrace() are called.
After printStackTrace() finishes, the exception is handled and we can continue with the normal execu-
tion.
At offset 47 there is a return which finishes the main() function, but there could be any other code which
would execute as if no exceptions were raised.
Here is an example on how IDA shows exception ranges:
Listing 4.14: from some random .class file found on the author’s computer
.catch java/io/FileNotFoundException from met001_335 to met001_360\
using met001_360
using met001_214
using met001_195
using met001_176
.catch java/io/FileNotFoundException from met001_83 to met001_129 using \
met001_129
.catch java/io/FileNotFoundException from met001_42 to met001_66 using \
met001_69
.catch java/io/FileNotFoundException from met001_begin to met001_37\
using met001_37
4.1.16 Classes
Simple class:
Listing 4.15: test.java

public class test
{
public static int a;
private static int b;
684
public test()
{
a=0;
b=0;
}
public static void set_a (int input)
{
a=input;
}
public static int get_a ()
{
return a;
}
public static void set_b (int input)
{
b=input;
}
public static int get_b ()
{
return b;
}
}
The constructor just sets both fields to zero:

public test();
flags: ACC_PUBLIC
Code:
0: aload_0
1: invokespecial #1 // Method java/lang/Object."<init>":()V
4: iconst_0
5: putstatic #2 // Field a:I
8: iconst_0
9: putstatic #3 // Field b:I
12: return
Setter of a:
public static void set_a(int);
Code:
0: iload_0
1: putstatic #2 // Field a:I
4: return
Getter of a:
public static int get_a();
Code:
0: getstatic #2 // Field a:I
3: ireturn
Setter of b:
public static void set_b(int);
Code:
0: iload_0
1: putstatic #3 // Field b:I
4: return
Getter of b:
public static int get_b();
685
Code:
0: getstatic #3 // Field b:I
3: ireturn
There is no difference in the code which works with public and private fields.
But this type information is present in the .class file, and it’s not possible to access private fields from
everywhere.
Let’s create an object and call its method:
Listing 4.16: ex1.java

public class ex1
{
{
test obj=new test();
obj.set_a (1234);
System.out.println(obj.a);
}
}

Code:
0: new #2 // class test
3: dup
4: invokespecial #3 // Method test."<init>":()V
7: astore_1
8: aload_1
9: pop
10: sipush 1234
13: invokestatic #4 // Method test.set_a:(I)V
19: aload_1
20: pop
21: getstatic #6 // Field test.a:I
27: return
The new instruction creates an object, but doesn’t call the constructor (it is called at offset 4).
The set_a() method is called at offset 16.
The a field is accessed using the getstatic instruction at offset 21.
4.1.17 Simple patching

First example
Let’s proceed with a simple code patching task.

public class nag
{
public static void nag_screen()
{
System.out.println("This program is not registered");
};
{
System.out.println("Greetings from the mega-software");
nag_screen();
}
}
How would we remove the printing of “This program is not registered” string?
Let’s load the .class file into IDA:
686
Figure 4.1: IDA
Let’s patch the first byte of the function to 177 (which is the return instruction’s opcode):
Figure 4.2: IDA
But that doesn’t work (JRE 1.7):

Exception in thread "main" java.lang.VerifyError: Expecting a stack map frame
Exception Details:
Location:
nag.nag_screen()V @1: nop
Reason:
Error exists in the bytecode
Bytecode:
0000000: b100 0212 03b6 0004 b1
at java.lang.Class.getDeclaredMethods0(Native Method)
at java.lang.Class.privateGetDeclaredMethods(Class.java:2615)
at java.lang.Class.getMethod0(Class.java:2856)
at java.lang.Class.getMethod(Class.java:1668)
687
at sun.launcher.LauncherHelper.getMainMethod(LauncherHelper.java:494)
at sun.launcher.LauncherHelper.checkAndLoadMain(LauncherHelper.java:486)
Perhaps JVM has some other checks related to the stack maps.
OK, let’s patch it differently by removing the call to nag():
Figure 4.3: IDA
0 is the opcode for NOP.

Now that works!
Second example
Another simple crackme example:

public class password
{
{
System.out.println("Please enter the password");
String input = System.console().readLine();
if (input.equals("secret"))
System.out.println("password is correct");
else
System.out.println("password is not correct");
}
}
Let’s load it in IDA:
688
Figure 4.4: IDA
We see here the ifeq instruction which does the job.

Its name stands for if equal, and this is misnomer, a better name would be ifz (if zero), i.e, if value at
TOS is zero, then do the jump.
In our example, it jumps if the password is not correct (the equals method returns False, which is 0).
The very first idea is to patch this instruction.
There are two bytes in ifeq opcode, which encode the jump offset.
To make this instruction a NOP, we must set the 3rd byte to the value of 3 (because by adding 3 to the
current address we will always jump to the next instruction, since the ifeq instruction’s length is 3 bytes):
689
Figure 4.5: IDA
That doesn’t work (JRE 1.7):

Exception in thread "main" java.lang.VerifyError: Expecting a stackmap frame at branch target ⤦
Ç 24
Exception Details:
Location:
password.main([Ljava/lang/String;)V @21: ifeq
Reason:
Expected stackmap frame at this location.
Bytecode:
0000000: b200 0212 03b6 0004 b800 05b6 0006 4c2b
0000010: 1207 b600 0899 0003 b200 0212 09b6 0004
0000020: a700 0bb2 0002 120a b600 04b1
Stackmap Table:
append_frame(@35,Object[#20])
same_frame(@43)
at java.lang.Class.getDeclaredMethods0(Native Method)
at java.lang.Class.privateGetDeclaredMethods(Class.java:2615)
at java.lang.Class.getMethod0(Class.java:2856)
at java.lang.Class.getMethod(Class.java:1668)
at sun.launcher.LauncherHelper.getMainMethod(LauncherHelper.java:494)
at sun.launcher.LauncherHelper.checkAndLoadMain(LauncherHelper.java:486)
But it must be mentioned that it worked in JRE 1.6.

We can also try to replace to all 3 ifeq opcode bytes with zero bytes (NOP), and it still won’t work.
Seems like there are more stack map checks in JRE 1.7.
OK, we’ll replace the whole call to the equals method with the iconst_1 instruction plus a pack of NOPs:
690
Figure 4.6: IDA
1 needs always to be in the TOS when the ifeq instruction is executed, so ifeq would never jump.
This works.
4.1.18 Summary
What is missing in Java in comparison to C/C++?
• Structures: use classes.
• Unions: use class hierarchies.
• Unsigned data types. By the way, this makes cryptographic algorithms somewhat harder to imple-
ment in Java.
• Function pointers.
691
Chapter 5
Finding important/interesting stuff in

the code
Minimalism it is not a prominent feature of modern software.

But not because the programmers are writing a lot, but because a lot of libraries are commonly linked
statically to executable files. If all external libraries were shifted into an external DLL files, the world would
be different. (Another reason for C++ are the STL and other template libraries.)
Thus, it is very important to determine the origin of a function, if it is from standard library or well-known
library (like Boost1 , libpng2 ), or if it is related to what we are trying to find in the code.
It is just absurd to rewrite all code in C/C++ to find what we’re looking for.
One of the primary tasks of a reverse engineer is to find quickly the code he/she needs, and what is not
that important.
The IDA disassembler allow us to search among text strings, byte sequences and constants. It is even
possible to export the code to .lst or .asm text files and then use grep, awk, etc.
When you try to understand what some code is doing, this easily could be some open-source library like
libpng. So when you see some constants or text strings which look familiar, it is always worth to google
them. And if you find the opensource project where they are used, then it’s enough just to compare the
functions. It may solve some part of the problem.
For example, if a program uses XML files, the first step may be determining which XML library is used for
processing, since the standard (or well-known) libraries are usually used instead of self-made one.
For example, the author of these lines once tried to understand how the compression/decompression of
network packets works in SAP 6.0. It is a huge software, but a detailed .PDB with debugging information
is present, and that is convenient. He finally came to the idea that one of the functions, that was called
CsDecomprLZC, was doing the decompression of network packets. Immediately he tried to google its
name and he quickly found the function was used in MaxDB (it is an open-source SAP project) 3 .
http://www.google.com/search?q=CsDecomprLZC
Astoundingly, MaxDB and SAP 6.0 software shared likewise code for the compression/decompression of
network packets.
5.1 Identification of executable files

5.1.1 Microsoft Visual C++
MSVC versions and DLLs that can be imported:
3 More about it in relevant section ( 8.12.1 on page 860)
692
Marketing ver. Internal ver. CL.EXE ver. DLLs imported Release date
6 6.0 12.00 msvcrt.dll June 1998
msvcp60.dll
.NET (2002) 7.0 13.00 msvcr70.dll February 13, 2002
msvcp70.dll
.NET 2003 7.1 13.10 msvcr71.dll April 24, 2003
msvcp71.dll
2005 8.0 14.00 msvcr80.dll November 7, 2005
msvcp80.dll
2008 9.0 15.00 msvcr90.dll November 19, 2007
msvcp90.dll
2010 10.0 16.00 msvcr100.dll April 12, 2010
msvcp100.dll
2012 11.0 17.00 msvcr110.dll September 12, 2012
msvcp110.dll
2013 12.0 18.00 msvcr120.dll October 17, 2013
msvcp120.dll
msvcp*.dll has C++-related functions, so if it is imported, this is probably a C++ program.
Name mangling
The names usually start with the ? symbol.

You can read more about MSVC’s name mangling here: 3.21.1 on page 542.
5.1.2 GCC
Aside from *NIX targets, GCC is also present in the win32 environment, in the form of Cygwin and MinGW.
Name mangling
Names usually start with the _Z symbols.

You can read more about GCC’s name mangling here: 3.21.1 on page 542.
Cygwin
cygwin1.dll is often imported.
MinGW
msvcrt.dll may be imported.
5.1.3 Intel Fortran

libifcoremd.dll, libifportmd.dll and libiomp5md.dll (OpenMP support) may be imported.
libifcoremd.dll has a lot of functions prefixed with for_, which means Fortran.
5.1.4 Watcom, OpenWatcom

Name mangling
Names usually start with the W symbol.

For example, that is how the method named “method” of the class “class” that does not have any argu-
ments and returns void is encoded:
W?method$_class$n__v
693
5.1.5 Borland
Here is an example of Borland Delphi’s and C++Builder’s name mangling:
@TApplication@IdleAction$qv
@TApplication@ProcessMDIAccels$qp6tagMSG
@TModule@$bctr$qpcpvt1
@TModule@$bdtr$qv
@TModule@ValidWindow$qp14TWindowsObject
@TrueColorTo8BitN$qpviiiiiit1iiiiii
@TrueColorTo16BitN$qpviiiiiit1iiiiii
@DIB24BitTo8BitBitmap$qpviiiiiit1iiiii
@TrueBitmap@$bctr$qpcl
@TrueBitmap@$bctr$qpvl
@TrueBitmap@$bctr$qiilll
The names always start with the @ symbol, then we have the class name came, method name, and encoded
the types of the arguments of the method.
These names can be in the .exe imports, .dll exports, debug data, etc.
Borland Visual Component Libraries (VCL) are stored in .bpl files instead of .dll ones, for example, vcl50.dll,
rtl60.dll.
Another DLL that might be imported: BORLNDMM.DLL.
Delphi
Almost all Delphi executables has the “Boolean” text string at the beginning of the code segment, along
with other type names.
This is a very typical beginning of the CODE segment of a Delphi program, this block came right after the
win32 PE file header:
00000400 04 10 40 00 03 07 42 6f 6f 6c 65 61 6e 01 00 00 |..@...Boolean...|
00000410 00 00 01 00 00 00 00 10 40 00 05 46 61 6c 73 65 |........@..False|
00000420 04 54 72 75 65 8d 40 00 2c 10 40 00 09 08 57 69 |.True.@.,.@...Wi|
00000430 64 65 43 68 61 72 03 00 00 00 00 ff ff 00 00 90 |deChar..........|
00000440 44 10 40 00 02 04 43 68 61 72 01 00 00 00 00 ff |D.@...Char......|
00000450 00 00 00 90 58 10 40 00 01 08 53 6d 61 6c 6c 69 |....X.@...Smalli|
00000460 6e 74 02 00 80 ff ff ff 7f 00 00 90 70 10 40 00 |nt..........p.@.|
00000470 01 07 49 6e 74 65 67 65 72 04 00 00 00 80 ff ff |..Integer.......|
00000480 ff 7f 8b c0 88 10 40 00 01 04 42 79 74 65 01 00 |......@...Byte..|
00000490 00 00 00 ff 00 00 00 90 9c 10 40 00 01 04 57 6f |..........@...Wo|
000004a0 72 64 03 00 00 00 00 ff ff 00 00 90 b0 10 40 00 |rd............@.|
000004b0 01 08 43 61 72 64 69 6e 61 6c 05 00 00 00 00 ff |..Cardinal......|
000004c0 ff ff ff 90 c8 10 40 00 10 05 49 6e 74 36 34 00 |......@...Int64.|
000004d0 00 00 00 00 00 00 80 ff ff ff ff ff ff ff 7f 90 |................|
000004e0 e4 10 40 00 04 08 45 78 74 65 6e 64 65 64 02 90 |..@...Extended..|
000004f0 f4 10 40 00 04 06 44 6f 75 62 6c 65 01 8d 40 00 |..@...Double..@.|
00000500 04 11 40 00 04 08 43 75 72 72 65 6e 63 79 04 90 |..@...Currency..|
00000510 14 11 40 00 0a 06 73 74 72 69 6e 67 20 11 40 00 |..@...string .@.|
00000520 0b 0a 57 69 64 65 53 74 72 69 6e 67 30 11 40 00 |..WideString0.@.|
00000530 0c 07 56 61 72 69 61 6e 74 8d 40 00 40 11 40 00 |..Variant.@.@.@.|
00000540 0c 0a 4f 6c 65 56 61 72 69 61 6e 74 98 11 40 00 |..OleVariant..@.|
00000550 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000560 00 00 00 00 00 00 00 00 00 00 00 00 98 11 40 00 |..............@.|
00000570 04 00 00 00 00 00 00 00 18 4d 40 00 24 4d 40 00 |.........M@.$M@.|
00000580 28 4d 40 00 2c 4d 40 00 20 4d 40 00 68 4a 40 00 |(M@.,M@. M@.hJ@.|
00000590 84 4a 40 00 c0 4a 40 00 07 54 4f 62 6a 65 63 74 |.J@..J@..TObject|
000005a0 a4 11 40 00 07 07 54 4f 62 6a 65 63 74 98 11 40 |..@...TObject..@|
000005b0 00 00 00 00 00 00 00 06 53 79 73 74 65 6d 00 00 |........System..|
000005c0 c4 11 40 00 0f 0a 49 49 6e 74 65 72 66 61 63 65 |..@...IInterface|
000005d0 00 00 00 00 01 00 00 00 00 00 00 00 00 c0 00 00 |................|
000005e0 00 00 00 00 46 06 53 79 73 74 65 6d 03 00 ff ff |....F.System....|
000005f0 f4 11 40 00 0f 09 49 44 69 73 70 61 74 63 68 c0 |..@...IDispatch.|
00000600 11 40 00 01 00 04 02 00 00 00 00 00 c0 00 00 00 |.@..............|
00000610 00 00 00 46 06 53 79 73 74 65 6d 04 00 ff ff 90 |...F.System.....|
00000620 cc 83 44 24 04 f8 e9 51 6c 00 00 83 44 24 04 f8 |..D$...Ql...D$..|
00000630 e9 6f 6c 00 00 83 44 24 04 f8 e9 79 6c 00 00 cc |.ol...D$...yl...|
00000640 cc 21 12 40 00 2b 12 40 00 35 12 40 00 01 00 00 |.!.@.+.@.5.@....|
694
00000650 00 00 00 00 00 00 00 00 00 c0 00 00 00 00 00 00 |................|
00000660 46 41 12 40 00 08 00 00 00 00 00 00 00 8d 40 00 |FA.@..........@.|
00000670 bc 12 40 00 4d 12 40 00 00 00 00 00 00 00 00 00 |..@.M.@.........|
00000680 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
00000690 bc 12 40 00 0c 00 00 00 4c 11 40 00 18 4d 40 00 |..@.....L.@..M@.|
000006a0 50 7e 40 00 5c 7e 40 00 2c 4d 40 00 20 4d 40 00 |P~@.\~@.,M@. M@.|
000006b0 6c 7e 40 00 84 4a 40 00 c0 4a 40 00 11 54 49 6e |l~@..J@..J@..TIn|
000006c0 74 65 72 66 61 63 65 64 4f 62 6a 65 63 74 8b c0 |terfacedObject..|
000006d0 d4 12 40 00 07 11 54 49 6e 74 65 72 66 61 63 65 |..@...TInterface|
000006e0 64 4f 62 6a 65 63 74 bc 12 40 00 a0 11 40 00 00 |dObject..@...@..|
000006f0 00 06 53 79 73 74 65 6d 00 00 8b c0 00 13 40 00 |..System......@.|
00000700 11 0b 54 42 6f 75 6e 64 41 72 72 61 79 04 00 00 |..TBoundArray...|
00000710 00 00 00 00 00 03 00 00 00 6c 10 40 00 06 53 79 |.........l.@..Sy|
00000720 73 74 65 6d 28 13 40 00 04 09 54 44 61 74 65 54 |stem(.@...TDateT|
00000730 69 6d 65 01 ff 25 48 e0 c4 00 8b c0 ff 25 44 e0 |ime..%H......%D.|
The first 4 bytes of the data segment (DATA) can be 00 00 00 00, 32 13 8B C0 or FF FF FF FF.
This information can be useful when dealing with packed/encrypted Delphi executables.
5.1.6 Other known DLLs

• vcomp*.dll—Microsoft’s implementation of OpenMP.
5.2 Communication with outer world (function level)

It’s often advisable to track function arguments and return values in debugger or DBI. For example, the
author once tried to understand meaning of some obscure function, which happens to be incorrectly
implemented bubble sort4 . (It worked correctly, but slower.) Meanwhile, watching inputs and outputs of
this function helps instantly to understand what it does.
Often, when you see division by multiplication ( 3.12 on page 496), but forgot all details about its mechan-
ics, you can just observe input and output and quickly find divisor.
5.3 Communication with the outer world (win32)

Sometimes it’s enough to observe some function’s inputs and outputs in order to understand what it does.
That way you can save time.
Files and registry access: for the very basic analysis, Process Monitor5 utility from SysInternals can help.
For the basic analysis of network accesses, Wireshark6 can be useful.
But then you will have to look inside anyway.
The first thing to look for is which functions from the OS’s APIs and standard libraries are used.
If the program is divided into a main executable file and a group of DLL files, sometimes the names of the
functions in these DLLs can help.
If we are interested in exactly what can lead to a call to MessageBox() with specific text, we can try to
find this text in the data segment, find the references to it and find the points from which the control may
be passed to the MessageBox() call we’re interested in.
If we are talking about a video game and we’re interested in which events are more or less random in it,
we may try to find the rand() function or its replacements (like the Mersenne twister algorithm) and find
the places from which those functions are called, and more importantly, how are the results used. One
example: 8.3 on page 791.
But if it is not a game, and rand() is still used, it is also interesting to know why. There are cases of
unexpected rand() usage in data compression algorithms (for encryption imitation): blog.yurichev.com.
4 https://yurichev.com/blog/weird_sort_KLEE/
695
5.3.1 Often used functions in the Windows API
These functions may be among the imported. It is worth to note that not every function might be used
in the code that was written by the programmer. A lot of functions might be called from library functions
and CRT code.
Some functions may have the -A suffix for the ASCII version and -W for the Unicode version.
• Registry access (advapi32.dll): RegEnumKeyEx, RegEnumValue, RegGetValue, RegOpenKeyEx, Reg-
QueryValueEx.
• Access to text .ini-files (kernel32.dll): GetPrivateProfileString.
• Dialog boxes (user32.dll): MessageBox, MessageBoxEx, CreateDialog, SetDlgItemText, GetDlgItem-
Text.
• Resources access ( 6.5.2 on page 754): (user32.dll): LoadMenu.
• TCP/IP networking (ws2_32.dll): WSARecv, WSASend.
• File access (kernel32.dll): CreateFile, ReadFile, ReadFileEx, WriteFile, WriteFileEx.
• High-level access to the Internet (wininet.dll): WinHttpOpen.
• Checking the digital signature of an executable file (wintrust.dll): WinVerifyTrust.
• The standard MSVC library (if it’s linked dynamically) (msvcr*.dll): assert, itoa, ltoa, open, printf,
read, strcmp, atol, atoi, fopen, fread, fwrite, memcmp, rand, strlen, strstr, strchr.
5.3.2 Extending trial period

Registry access functions are frequent targets for those who try to crack trial period of some software,
which may save installation date/time into registry.
Another popular target are GetLocalTime() and GetSystemTime() functions: a trial software, at each
startup, must check current date/time somehow anyway.
5.3.3 Removing nag dialog box

A popular way to find out what causing popping nag dialog box is intercepting MessageBox(), CreateDia-
log() and CreateWindow() functions.
5.3.4 tracer: Intercepting all functions in specific module

There are INT3 breakpoints in the tracer, that are triggered only once, however, they can be set for all
functions in a specific DLL.
--one-time-INT3-bp:somedll.dll!.*
Or, let’s set INT3 breakpoints on all functions with the xml prefix in their name:
--one-time-INT3-bp:somedll.dll!xml.*
On the other side of the coin, such breakpoints are triggered only once. Tracer will show the call of a
function, if it happens, but only once. Another drawback—it is impossible to see the function’s arguments.
Nevertheless, this feature is very useful when you know that the program uses a DLL, but you do not know
which functions are actually used. And there are a lot of functions.
For example, let’s see, what does the uptime utility from cygwin use:
tracer -l:uptime.exe --one-time-INT3-bp:cygwin1.dll!.*
Thus we may see all that cygwin1.dll library functions that were called at least once, and where from:
One-time INT3 breakpoint: cygwin1.dll!__main (called from uptime.exe!OEP+0x6d (0x40106d))
One-time INT3 breakpoint: cygwin1.dll!_geteuid32 (called from uptime.exe!OEP+0xba3 (0x401ba3))
One-time INT3 breakpoint: cygwin1.dll!_getuid32 (called from uptime.exe!OEP+0xbaa (0x401baa))
One-time INT3 breakpoint: cygwin1.dll!_getegid32 (called from uptime.exe!OEP+0xcb7 (0x401cb7))
One-time INT3 breakpoint: cygwin1.dll!_getgid32 (called from uptime.exe!OEP+0xcbe (0x401cbe))
One-time INT3 breakpoint: cygwin1.dll!sysconf (called from uptime.exe!OEP+0x735 (0x401735))
One-time INT3 breakpoint: cygwin1.dll!setlocale (called from uptime.exe!OEP+0x7b2 (0x4017b2))
696
One-time INT3 breakpoint: cygwin1.dll!_open64 (called from uptime.exe!OEP+0x994 (0x401994))
One-time INT3 breakpoint: cygwin1.dll!_lseek64 (called from uptime.exe!OEP+0x7ea (0x4017ea))
One-time INT3 breakpoint: cygwin1.dll!read (called from uptime.exe!OEP+0x809 (0x401809))
One-time INT3 breakpoint: cygwin1.dll!sscanf (called from uptime.exe!OEP+0x839 (0x401839))
One-time INT3 breakpoint: cygwin1.dll!uname (called from uptime.exe!OEP+0x139 (0x401139))
One-time INT3 breakpoint: cygwin1.dll!time (called from uptime.exe!OEP+0x22e (0x40122e))
One-time INT3 breakpoint: cygwin1.dll!localtime (called from uptime.exe!OEP+0x236 (0x401236))
One-time INT3 breakpoint: cygwin1.dll!sprintf (called from uptime.exe!OEP+0x25a (0x40125a))
One-time INT3 breakpoint: cygwin1.dll!setutent (called from uptime.exe!OEP+0x3b1 (0x4013b1))
One-time INT3 breakpoint: cygwin1.dll!getutent (called from uptime.exe!OEP+0x3c5 (0x4013c5))
One-time INT3 breakpoint: cygwin1.dll!endutent (called from uptime.exe!OEP+0x3e6 (0x4013e6))
One-time INT3 breakpoint: cygwin1.dll!puts (called from uptime.exe!OEP+0x4c3 (0x4014c3))
5.4 Strings
5.4.1 Text strings
C/C++
The normal C strings are zero-terminated (ASCIIZ-strings).

The reason why the C string format is as it is (zero-terminated) is apparently historical. In [Dennis M.
Ritchie, The Evolution of the Unix Time-sharing System, (1979)] we read:
A minor difference was that the unit of I/O was the word, not the byte, because the PDP-7
was a word-addressed machine. In practice this meant merely that all programs dealing
with character streams ignored null characters, because null was used to pad a file to an
even number of characters.
In Hiew or FAR Manager these strings look like this:

int main()
{
printf ("Hello, world!\n");
};
Figure 5.1: Hiew
Borland Delphi
The string in Pascal and Borland Delphi is preceded by an 8-bit or 32-bit string length.
For example:
Listing 5.1: Delphi

CODE:00518AC8 dd 19h
CODE:00518ACC aLoading___Plea db 'Loading... , please wait.',0
...
CODE:00518AFC dd 10h
697
CODE:00518B00 aPreparingRun__ db 'Preparing run...',0
Unicode
Often, what is called Unicode is a methods for encoding strings where each character occupies 2 bytes or
16 bits. This is a common terminological mistake. Unicode is a standard for assigning a number to each
character in the many writing systems of the world, but does not describe the encoding method.
The most popular encoding methods are: UTF-8 (is widespread in Internet and *NIX systems) and UTF-16LE
(is used in Windows).
UTF-8
UTF-8 is one of the most successful methods for encoding characters. All Latin symbols are encoded just
like in ASCII, and the symbols beyond the ASCII table are encoded using several bytes. 0 is encoded as
before, so all standard C string functions work with UTF-8 strings just like any other string.
Let’s see how the symbols in various languages are encoded in UTF-8 and how it looks like in FAR, using
the 437 codepage 7 :
Figure 5.2: FAR: UTF-8
As you can see, the English language string looks the same as it is in ASCII.
The Hungarian language uses some Latin symbols plus symbols with diacritic marks.
These symbols are encoded using several bytes, these are underscored with red. It’s the same story with
the Icelandic and Polish languages.
There is also the “Euro” currency symbol at the start, which is encoded with 3 bytes.
The rest of the writing systems here have no connection with Latin.
7 The example and translations was taken from here: http://go.yurichev.com/17304
698
At least in Russian, Arabic, Hebrew and Hindi we can see some recurring bytes, and that is not surprise:
all symbols from a writing system are usually located in the same Unicode table, so their code begins with
the same numbers.
At the beginning, before the “How much?” string we see 3 bytes, which are in fact the BOM8 . The BOM
defines the encoding system to be used.
UTF-16LE
Many win32 functions in Windows have the suffixes -A and -W. The first type of functions works with
normal strings, the other with UTF-16LE strings (wide).
In the second case, each symbol is usually stored in a 16-bit value of type short.
The Latin symbols in UTF-16 strings look in Hiew or FAR like they are interleaved with zero byte:
int wmain()
{
wprintf (L"Hello, world!\n");
};
Figure 5.3: Hiew
We can see this often in Windows NT system files:
Figure 5.4: Hiew
Strings with characters that occupy exactly 2 bytes are called “Unicode” in IDA:
.data:0040E000 aHelloWorld:
.data:0040E000 unicode 0, <Hello, world!>
.data:0040E000 dw 0Ah, 0
8 Byte Order Mark
699
Here is how the Russian language string is encoded in UTF-16LE:
Figure 5.5: Hiew: UTF-16LE
What we can easily spot is that the symbols are interleaved by the diamond character (which has the
ASCII code of 4). Indeed, the Cyrillic symbols are located in the fourth Unicode plane. Hence, all Cyrillic
symbols in UTF-16LE are located in the 0x400-0x4FF range.
Let’s go back to the example with the string written in multiple languages. Here is how it looks like in
UTF-16LE.
Figure 5.6: FAR: UTF-16LE
Here we can also see the BOM at the beginning. All Latin characters are interleaved with a zero byte.
Some characters with diacritic marks (Hungarian and Icelandic languages) are also underscored in red.
Base64
The base64 encoding is highly popular for the cases when you have to transfer binary data as a text string.
In essence, this algorithm encodes 3 binary bytes into 4 printable characters: all 26 Latin letters (both
lower and upper case), digits, plus sign (“+”) and slash sign (“/”), 64 characters in total.
700
One distinctive feature of base64 strings is that they often (but not always) end with 1 or 2 padding
equality symbol(s) (“=”), for example:
AVjbbVSVfcUMu1xvjaMgjNtueRwBbxnyJw8dpGnLW8ZW8aKG3v4Y0icuQT+qEJAp9lAOuWs=
WVjbbVSVfcUMu1xvjaMgjNtueRwBbxnyJw8dpGnLW8ZW8aKG3v4Y0icuQT+qEJAp9lAOuQ==
The equality sign (“=”) is never encounter in the middle of base64-encoded strings.
Now example of manual encoding. Let’s encode 0x00, 0x11, 0x22, 0x33 hexadecimal bytes into base64
string:
$ echo -n "\x00\x11\x22\x33" | base64
ABEiMw==
Let’s put all 4 bytes in binary form, then regroup them into 6-bit groups:
| 00 || 11 || 22 || 33 || || |
00000000000100010010001000110011????????????????
| A || B || E || i || M || w || = || = |
Three first bytes (0x00, 0x11, 0x22) can be encoded into 4 base64 characters (“ABEi”), but the last one
(0x33) — cannot be, so it’s encoded using two characters (“Mw”) and padding symbol (“=”) is added twice
to pad the last group to 4 characters. Hence, length of all correct base64 strings are always divisible by
4.
Base64 is often used when binary data needs to be stored in XML. “Armored” (i.e., in text form) PGP keys
and signatures are encoded using base64.
Some people tries to use base64 to obfuscate strings: http://blog.sec-consult.com/2016/01/deliberately-
html 9 .
There are utilities for scanning an arbitrary binary files for base64 strings. One such utility is base64scanner10 .
Another encoding system which was much more popular in UseNet and FidoNet is Uuencoding. Binary
files are still encoded in Uuencode format in Phrack magazine. It offers mostly the same features, but is
different from base64 in the sense that file name is also stored in header.
By the way: there is also close sibling to base64: base32, alphabet of which has 10 digits and 26 Latin
characters. One well-known usage of it is onion addresses 11 , like:
http://3g2upl4pq6kufc4m.onion/. URL can’t have mixed-case Latin characters, so apparently, this is
why Tor developers used base32.
5.4.2 Finding strings in binary
Actually, the best form of Unix

documentation is frequently running the
strings command over a program’s object
code. Using strings, you can get a complete
list of the program’s hard-coded file name,
environment variables, undocumented
options, obscure error messages, and so
forth.
The Unix-Haters Handbook
The standard UNIX strings utility is quick-n-dirty way to see strings in file. For example, these are some
strings from OpenSSH 7.2 sshd executable file:
...
0123
0123456789
0123456789abcdefABCDEF.:/
%02x
...
%.100s, line %lu: Bad permitopen specification <%.100s>
9 http://archive.is/nDCas
10 https://github.com/DennisYurichev/base64scanner
11 https://trac.torproject.org/projects/tor/wiki/doc/HiddenServiceNames
701
%.100s, line %lu: invalid criteria
%.100s, line %lu: invalid tun device
...
%.200s/.ssh/environment
...
2886173b9c9b6fdbdeda7a247cd636db38deaa.debug
$2a$06$r3.juUaHZDlIbQaO2dS9FuYxL1W9M81R1Tc92PoSNmzvpEqLkLGrK
...
3des-cbc
...
Bind to port %s on %s.
Bind to port %s on %s failed: %.200s.
/bin/login
/bin/sh
/bin/sh /etc/ssh/sshrc
...
D$4PQWR1
D$4PUj
D$4PV
D$4PVj
D$4PW
D$4PWj
D$4X
D$4XZj
D$4Y
...
diffie-hellman-group-exchange-sha1
diffie-hellman-group-exchange-sha256
digests
D$iPV
direct-streamlocal
direct-streamlocal@openssh.com
...
FFFFFFFFFFFFFFFFC90FDAA22168C234C4C6628B80DC1CD129024E088A6...
...
There are options, error messages, file paths, imported dynamic modules and functions, some other
strange strings (keys?) There is also unreadable noise—x86 code sometimes has chunks consisting of
printable ASCII characters, up to 8 characters.
Of course, OpenSSH is open-source program. But looking at readable strings inside of some unknown
binary is often a first step of analysis.
grep can be applied as well.
Hiew has the same capability (Alt-F6), as well as Sysinternals ProcessMonitor.
5.4.3 Error/debug messages

Debugging messages are very helpful if present. In some sense, the debugging messages are reporting
what’s going on in the program right now. Often these are printf()-like functions, which write to log-files,
or sometimes do not writing anything but the calls are still present since the build is not a debug one but
release one.
If local or global variables are dumped in debug messages, it might be helpful as well since it is possible
to get at least the variable names. For example, one of such function in Oracle RDBMS is ksdwrt().
Meaningful text strings are often helpful. The IDA disassembler may show from which function and from
which point this specific string is used. Funny cases sometimes happen12 .
The error messages may help us as well. In Oracle RDBMS, errors are reported using a group of functions.
You can read more about them here: blog.yurichev.com.
It is possible to find quickly which functions report errors and in which conditions.
By the way, this is often the reason why copy-protection systems use inarticulate cryptic error messages
or just error numbers. No software author is happy if the software cracker can quickly understands copy-
protection’s inner workings judging by error messages it can produce.
One example of encrypted error messages is here: 8.8.2 on page 826.
12 blog.yurichev.com
702
5.4.4 Suspicious magic strings
Some magic strings which are usually used in backdoors look pretty suspicious.
For example, there was a backdoor in the TP-Link WR740 home router13 . The backdoor can activated
using the following URL:
http://192.168.0.1/userRpmNatDebugRpm26525557/start_art.html.
Indeed, the “userRpmNatDebugRpm26525557” string is present in the firmware.

This string was not googleable until the wide disclosure of information about the backdoor.
You would not find this in any RFC14 .
You would not find any computer science algorithm which uses such strange byte sequences.
And it doesn’t look like an error or debugging message.
So it’s a good idea to inspect the usage of such weird strings.
Sometimes, such strings are encoded using base64.

So it’s a good idea to decode them all and to scan them visually, even a glance should be enough.
More precise, this method of hiding backdoors is called “security through obscurity”.
5.5 Calls to assert()

Sometimes the presence of the assert() macro is useful too: commonly this macro leaves source file
name, line number and condition in the code.
The most useful information is contained in the assert’s condition, we can deduce variable names or
structure field names from it. Another useful piece of information are the file names—we can try to
deduce what type of code is there. Also it is possible to recognize well-known open-source libraries by the
file names.
Listing 5.2: Example of informative assert() calls

.text:107D4B29 mov dx, [ecx+42h]
.text:107D4B2D cmp edx, 1
.text:107D4B30 jz short loc_107D4B4A
.text:107D4B32 push 1ECh
.text:107D4B37 push offset aWrite_c ; "write.c"
.text:107D4B3C push offset aTdTd_planarcon ; "td->td_planarconfig == PLANARCONFIG_CON"...
.text:107D4B41 call ds:_assert
...
.text:107D52CA mov edx, [ebp-4]

.text:107D52CD and edx, 3
.text:107D52D0 test edx, edx
.text:107D52D2 jz short loc_107D52E9
.text:107D52D4 push 58h
.text:107D52D6 push offset aDumpmode_c ; "dumpmode.c"
.text:107D52DB push offset aN30 ; "(n & 3) == 0"
.text:107D52E0 call ds:_assert
...
.text:107D6759 mov cx, [eax+6]

.text:107D675D cmp ecx, 0Ch
.text:107D6760 jle short loc_107D677A
.text:107D6762 push 2D8h
.text:107D6767 push offset aLzw_c ; "lzw.c"
.text:107D676C push offset aSpLzw_nbitsBit ; "sp->lzw_nbits <= BITS_MAX"
.text:107D6771 call ds:_assert
13 http://sekurak.pl/tp-link-httptftp-backdoor/
14 Request for Comments
703
It is advisable to “google” both the conditions and file names, which can lead us to an open-source library.
For example, if we “google” “sp->lzw_nbits <= BITS_MAX”, this predictably gives us some open-source
code that’s related to the LZW compression.
5.6 Constants
Humans, including programmers, often use round numbers like 10, 100, 1000, in real life as well as in the
code.
The practicing reverse engineer usually know them well in hexadecimal representation: 10=0xA, 100=0x64,
1000=0x3E8, 10000=0x2710.
The constants 0xAAAAAAAA (0b10101010101010101010101010101010) and
0x55555555 (0b01010101010101010101010101010101) are also popular—those are composed of alter-
nating bits.
That may help to distinguish some signal from a signal where all bits are turned on (0b1111 …) or off
(0b0000 …). For example, the 0x55AA constant is used at least in the boot sector, MBR15 , and in the ROM
of IBM-compatible extension cards.
Some algorithms, especially cryptographical ones use distinct constants, which are easy to find in code
using IDA.
For example, the MD5 algorithm initializes its own internal variables like this:
var int h0 := 0x67452301
var int h1 := 0xEFCDAB89
var int h2 := 0x98BADCFE
var int h3 := 0x10325476
If you find these four constants used in the code in a row, it is highly probable that this function is related
to MD5.
Another example are the CRC16/CRC32 algorithms, whose calculation algorithms often use precomputed
tables like this one:
Listing 5.3: linux/lib/crc16.c
/** CRC table for the CRC-16. The poly is 0x8005 (x^16 + x^15 + x^2 + 1) */
u16 const crc16_table[256] = {
0x0000, 0xC0C1, 0xC181, 0x0140, 0xC301, 0x03C0, 0x0280, 0xC241,
0xC601, 0x06C0, 0x0780, 0xC741, 0x0500, 0xC5C1, 0xC481, 0x0440,
0xCC01, 0x0CC0, 0x0D80, 0xCD41, 0x0F00, 0xCFC1, 0xCE81, 0x0E40,
...
See also the precomputed table for CRC32: 3.8 on page 482.
In tableless CRC algorithms well-known polynomials are used, for example, 0xEDB88320 for CRC32.
5.6.1 Magic numbers

A lot of file formats define a standard file header where a magic number(s) is used, single one or even
several.
For example, all Win32 and MS-DOS executables start with the two characters “MZ”.
At the beginning of a MIDI file the “MThd” signature must be present. If we have a program which uses
MIDI files for something, it’s very likely that it must check the file for validity by checking at least the first
4 bytes.
This could be done like this: (buf points to the beginning of the loaded file in memory)
cmp [buf], 0x6468544D ; "MThd"
jnz _error_not_a_MIDI_file
…or by calling a function for comparing memory blocks like memcmp() or any other equivalent code up to
a CMPSB ( .1.6 on page 1018) instruction.
When you find such point you already can say where the loading of the MIDI file starts, also, we could see
the location of the buffer with the contents of the MIDI file, what is used from the buffer, and how.
15 Master Boot Record
704
Dates
Often, one may encounter number like 0x19870116, which is clearly looks like a date (year 1987, 1th
month (January), 16th day). This may be someone’s birthday date (a programmer, his/her relative, child),
or some other important date. The date may also be written in a reverse order, like 0x16011987. American-
style dates are also popular, like 0x01161987.
Well-known example is 0x19540119 (magic number used in UFS2 superblock structure), which is a birthday
date of Marshall Kirk McKusick, prominent FreeBSD contributor.
Stuxnet uses the number “19790509” (not as 32-bit number, but as string, though), and this led to spec-
ulation that the malware is connected to Israel16 .
Also, numbers like those are very popular in amateur-grade cryptography, for example, excerpt from the
secret function internals from HASP3 dongle 17 :
void xor_pwd(void)
{
int i;
pwd^=0x09071966;
for(i=0;i<8;i++)
{
al_buf[i]= pwd & 7; pwd = pwd >> 3;
}
};
void emulate_func2(unsigned short seed)

{
int i, j;
for(i=0;i<8;i++)
{
ch[i] = 0;
for(j=0;j<8;j++)
{
seed *= 0x1989;
seed += 5;
ch[i] |= (tab[(seed>>9)&0x3f]) << (7-j);
}
}
}
DHCP
This applies to network protocols as well. For example, the DHCP protocol’s network packets contains the
so-called magic cookie: 0x63538263. Any code that generates DHCP packets somewhere must embed
this constant into the packet. If we find it in the code we may find where this happens and, not only
that. Any program which can receive DHCP packet must verify the magic cookie, comparing it with the
constant.
For example, let’s take the dhcpcore.dll file from Windows 7 x64 and search for the constant. And we can
find it, twice: it seems that the constant is used in two functions with descriptive names
DhcpExtractOptionsForValidation() and DhcpExtractFullOptions():
Listing 5.4: dhcpcore.dll (Windows 7 x64)
.rdata:000007FF6483CBE8 dword_7FF6483CBE8 dd 63538263h ; DATA XREF:
DhcpExtractOptionsForValidation+79
.rdata:000007FF6483CBEC dword_7FF6483CBEC dd 63538263h ; DATA XREF:
DhcpExtractFullOptions+97
And here are the places where these constants are accessed:
.text:000007FF6480875F mov eax, [rsi]
.text:000007FF64808761 cmp eax, cs:dword_7FF6483CBE8
.text:000007FF64808767 jnz loc_7FF64817179
16 This is a date of execution of Habib Elghanian, persian jew.

17 https://web.archive.org/web/20160311231616/http://www.woodmann.com/fravia/bayu3.htm
705
And:

.text:000007FF648082C7 mov eax, [r12]
.text:000007FF648082CB cmp eax, cs:dword_7FF6483CBEC
.text:000007FF648082D1 jnz loc_7FF648173AF
5.6.2 Specific constants

Sometimes, there is a specific constant for some type of code. For example, the author once dug into a
code, where number 12 was encountered suspiciously often. Size of many arrays is 12, or multiple of 12
(24, etc). As it turned out, that code takes 12-channel audio file at input and process it.
And vice versa: for example, if a program works with text field which has length of 120 bytes, there has
to be a constant 120 or 119 somewhere in the code. If UTF-16 is used, then 2 ⋅ 120. If a code works with
network packets of fixed size, it’s good idea to search for this constant in the code as well.
This is also true for amateur cryptography (license keys, etc). If encrypted block has size of n bytes, you
may want to try to find occurences of this number throughout the code. Also, if you see a piece of code
which is been repeated n times in loop during execution, this may be encryption/decryption routine.
5.6.3 Searching for constants

It is easy in IDA: Alt-B or Alt-I. And for searching for a constant in a big pile of files, or for searching in
non-executable files, there is a small utility called binary grep18 .
5.7 Finding the right instructions

If the program is utilizing FPU instructions and there are very few of them in the code, one can try to check
each one manually with a debugger.
For example, we may be interested how Microsoft Excel calculates the formulae entered by user. For
example, the division operation.
If we load excel.exe (from Office 2010) version 14.0.4756.1000 into IDA, make a full listing and to find
every FDIV instruction (except the ones which use constants as a second operand—obviously, they do not
suit us):
cat EXCEL.lst | grep fdiv | grep -v dbl_ > EXCEL.fdiv
…then we see that there are 144 of them.

We can enter a string like =(1/3) in Excel and check each instruction.
By checking each instruction in a debugger or tracer (one may check 4 instruction at a time), we get lucky
and the sought-for instruction is just the 14th:
.text:3011E919 DC 33 fdiv qword ptr [ebx]
PID=13944|TID=28744|(0) 0x2f64e919 (Excel.exe!BASE+0x11e919)

EAX=0x02088006 EBX=0x02088018 ECX=0x00000001 EDX=0x00000001
ESI=0x02088000 EDI=0x00544804 EBP=0x0274FA3C ESP=0x0274F9F8
EIP=0x2F64E919
FLAGS=PF IF
FPU ControlWord=IC RC=NEAR PC=64bits PM UM OM ZM DM IM
FPU StatusWord=
FPU ST(0): 1.000000
ST(0) holds the first argument (1) and second one is in [EBX].
The instruction after FDIV (FSTP) writes the result in memory:
18 GitHub
706
.text:3011E91B DD 1E fstp qword ptr [esi]
If we set a breakpoint on it, we can see the result:

PID=32852|TID=36488|(0) 0x2f40e91b (Excel.exe!BASE+0x11e91b)
ESI=0x00598000 EDI=0x00294804 EBP=0x026CF93C ESP=0x026CF8F8
EIP=0x2F40E91B
FLAGS=PF IF
FPU StatusWord=C1 P
FPU ST(0): 0.333333
Also as a practical joke, we can modify it on the fly:

tracer -l:excel.exe bpx=excel.exe!BASE+0x11E91B,set(st0,666)
PID=36540|TID=24056|(0) 0x2f40e91b (Excel.exe!BASE+0x11e91b)

ESI=0x00680000 EDI=0x00395404 EBP=0x0290FD9C ESP=0x0290FD58
EIP=0x2F40E91B
FLAGS=PF IF
FPU StatusWord=C1 P
FPU ST(0): 0.333333
Set ST0 register to 666.000000
Excel shows 666 in the cell, finally convincing us that we have found the right point.
Figure 5.7: The practical joke worked
If we try the same Excel version, but in x64, we will find only 12 FDIV instructions there, and the one we
looking for is the third one.
tracer.exe -l:excel.exe bpx=excel.exe!BASE+0x1B7FCC,set(st0,666)
It seems that a lot of division operations of float and double types, were replaced by the compiler with
SSE instructions like DIVSD (DIVSD is present 268 times in total).
5.8 Suspicious code patterns

5.8.1 XOR instructions
Instructions like XOR op, op (for example, XOR EAX, EAX) are usually used for setting the register value
to zero, but if the operands are different, the “exclusive or” operation is executed.
707
This operation is rare in common programming, but widespread in cryptography, including amateur one.
It’s especially suspicious if the second operand is a big number.
This may point to encrypting/decrypting, checksum computing, etc.
One exception to this observation worth noting is the “canary” ( 1.26.3 on page 279). Its generation and
checking are often done using the XOR instruction.
This AWK script can be used for processing IDA listing (.lst) files:
gawk -e '$2=="xor" { tmp=substr($3, 0, length($3)-1); if (tmp!=$4) if($4!="esp") if ($4!="ebp")⤦
Ç { print $1, $2, tmp, ",", $4 } }' filename.lst
It is also worth noting that this kind of script can also match incorrectly disassembled code ( 5.11.1 on
page 719).
5.8.2 Hand-written assembly code

Modern compilers do not emit the LOOP and RCL instructions. On the other hand, these instructions are
well-known to coders who like to code directly in assembly language. If you spot these, it can be said that
there is a high probability that this fragment of code was hand-written. Such instructions are marked as
(M) in the instructions list in this appendix: .1.6 on page 1012.
Also the function prologue/epilogue are not commonly present in hand-written assembly.
Commonly there is no fixed system for passing arguments to functions in the hand-written code.
Example from the Windows 2003 kernel (ntoskrnl.exe file):
MultiplyTest proc near ; CODE XREF: Get386Stepping
xor cx, cx
loc_620555: ; CODE XREF: MultiplyTest+E
push cx
call Multiply
pop cx
jb short locret_620563
loop loc_620555
clc
locret_620563: ; CODE XREF: MultiplyTest+C
retn
MultiplyTest endp
Multiply proc near ; CODE XREF: MultiplyTest+5

mov ecx, 81h
mov eax, 417A000h
mul ecx
cmp edx, 2
stc
jnz short locret_62057F
cmp eax, 0FE7A000h
stc
jnz short locret_62057F
clc
locret_62057F: ; CODE XREF: Multiply+10
; Multiply+18
retn
Multiply endp
Indeed, if we look in the WRK19 v1.2 source code, this code can be found easily in file
WRK-v1.2\base\ntos\ke\i386\cpu.asm.
As of RCL instruction, I can find it in the ntoskrnl.exe file in Windows 2003 x86 (compiled with MS Visual C
compiler). It is occurred only once there, in RtlExtendedLargeIntegerDivide() function, and this might
be inline assembler code case.
19 Windows Research Kernel
708
5.9 Using magic numbers while tracing
Often, our main goal is to understand how the program uses a value that has been either read from file
or received via network. The manual tracing of a value is often a very labor-intensive task. One of the
simplest techniques for this (although not 100% reliable) is to use your own magic number.
This resembles X-ray computed tomography is some sense: a radiocontrast agent is injected into the
patient’s blood, which is then used to improve the visibility of the patient’s internal structure in to the
X-rays. It is well known how the blood of healthy humans percolates in the kidneys and if the agent is
in the blood, it can be easily seen on tomography, how blood is percolating, and are there any stones or
tumors.
We can take a 32-bit number like 0x0badf00d, or someone’s birth date like 0x11101979 and write this
4-byte number to some point in a file used by the program we investigate.
Then, while tracing this program with tracer in code coverage mode, with the help of grep or just by
searching in the text file (of tracing results), we can easily see where the value has been used and how.
Example of grepable tracer results in cc mode:
0x150bf66 (_kziaia+0x14), e= 1 [MOV EBX, [EBP+8]] [EBP+8]=0xf59c934
0x150bf69 (_kziaia+0x17), e= 1 [MOV EDX, [69AEB08h]] [69AEB08h]=0
0x150bf6f (_kziaia+0x1d), e= 1 [FS: MOV EAX, [2Ch]]
0x150bf75 (_kziaia+0x23), e= 1 [MOV ECX, [EAX+EDX*4]] [EAX+EDX*4]=0xf1ac360
0x150bf78 (_kziaia+0x26), e= 1 [MOV [EBP-4], ECX] ECX=0xf1ac360
This can be used for network packets as well. It is important for the magic number to be unique and not
to be present in the program’s code.
Aside of the tracer, DosBox (MS-DOS emulator) in heavydebug mode is able to write information about all
registers’ states for each executed instruction of the program to a plain text file20 , so this technique may
be useful for DOS programs as well.
5.10 Loops
Whenever your program works with some kind of file, or buffer of some size, it has to be some kind of
decrypting/processing loop inside of the code.
This is a real example of tracer tool output. There was a code which loads some kind of encryted file of
258 bytes. I run it with the intention to get each instruction counts (a DBI tool will serve much better
these days). And I quickly found a piece of code, which executed 259/258 times:
...
0x45a6b5 e= 1 [FS: MOV [0], EAX] EAX=0x218fb08

0x45a6bb e= 1 [MOV [EBP-254h], ECX] ECX=0x218fbd8
0x45a6c1 e= 1 [MOV EAX, [EBP-254h]] [EBP-254h]=0x218fbd8
0x45a6c7 e= 1 [CMP [EAX+14h], 0] [EAX+14h]=0x102
0x45a6cb e= 1 [JZ 45A9F2h] ZF=false
0x45a6d1 e= 1 [MOV [EBP-0Dh], 1]
0x45a6d5 e= 1 [XOR ECX, ECX] ECX=0x218fbd8
0x45a6d7 e= 1 [MOV [EBP-14h], CX] CX=0
0x45a6db e= 1 [MOV [EBP-18h], 0]
0x45a6e2 e= 1 [JMP 45A6EDh]
0x45a6e4 e= 258 [MOV EDX, [EBP-18h]] [EBP-18h]=0..5 (248 items skipped) 0xfd..0x101
0x45a6e7 e= 258 [ADD EDX, 1] EDX=0..5 (248 items skipped) 0xfd..0x101
0x45a6ea e= 258 [MOV [EBP-18h], EDX] EDX=1..6 (248 items skipped) 0xfe..0x102
0x45a6ed e= 259 [MOV EAX, [EBP-254h]] [EBP-254h]=0x218fbd8
0x45a6f3 e= 259 [MOV ECX, [EBP-18h]] [EBP-18h]=0..5 (249 items skipped) 0xfe..0x102
0x45a6f6 e= 259 [CMP ECX, [EAX+14h]] ECX=0..5 (249 items skipped) 0xfe..0x102 [EAX+14h]=0x102
0x45a6f9 e= 259 [JNB 45A727h] CF=false,true
0x45a6fb e= 258 [MOV EDX, [EBP-254h]] [EBP-254h]=0x218fbd8
0x45a701 e= 258 [MOV EAX, [EDX+10h]] [EDX+10h]=0x21ee4c8
0x45a704 e= 258 [MOV ECX, [EBP-18h]] [EBP-18h]=0..5 (248 items skipped) 0xfd..0x101
0x45a707 e= 258 [ADD ECX, 1] ECX=0..5 (248 items skipped) 0xfd..0x101
0x45a70a e= 258 [IMUL ECX, ECX, 1Fh] ECX=1..6 (248 items skipped) 0xfe..0x102
0x45a70d e= 258 [MOV EDX, [EBP-18h]] [EBP-18h]=0..5 (248 items skipped) 0xfd..0x101
20 See also my blog post about this DosBox feature: blog.yurichev.com
709
0x45a710 e= 258 [MOVZX EAX, [EAX+EDX]] [EAX+EDX]=1..6 (156 items skipped) 0xf3, 0xf8, 0xf9, 0⤦
Ç xfc, 0xfd
0x45a714 e= 258 [XOR EAX, ECX] EAX=1..6 (156 items skipped) 0xf3, 0xf8, 0xf9, 0xfc, 0xfd ECX=0⤦
Ç x1f, 0x3e, 0x5d, 0x7c, 0x9b (248 items skipped) 0x1ec2, 0x1ee1, 0x1f00, 0x1f1f, 0x1f3e
0x45a716 e= 258 [MOV ECX, [EBP-254h]] [EBP-254h]=0x218fbd8
0x45a71c e= 258 [MOV EDX, [ECX+10h]] [ECX+10h]=0x21ee4c8
0x45a71f e= 258 [MOV ECX, [EBP-18h]] [EBP-18h]=0..5 (248 items skipped) 0xfd..0x101
0x45a722 e= 258 [MOV [EDX+ECX], AL] AL=0..5 (77 items skipped) 0xe2, 0xee, 0xef, 0xf7, 0xfc
0x45a725 e= 258 [JMP 45A6E4h]
0x45a727 e= 1 [PUSH 5]
0x45a729 e= 1 [MOV ECX, [EBP-254h]] [EBP-254h]=0x218fbd8
0x45a72f e= 1 [CALL 45B500h]
0x45a734 e= 1 [MOV ECX, EAX] EAX=0x218fbd8
0x45a736 e= 1 [CALL 45B710h]
0x45a73b e= 1 [CMP EAX, 5] EAX=5
...
As it turns out, this is the decrypting loop.
5.10.1 Some binary file patterns

All examples here were prepared on the Windows with active code page 437 in console. Binary files
internally may look visually different if another code page is set.
710
Arrays
Sometimes, we can clearly spot an array of 16/32/64-bit values visually, in hex editor.
Here is an example of array of 16-bit values. We see that the first byte in pair is 7 or 8, and the second
looks random:
Figure 5.8: FAR: array of 16-bit values
I used a file containing 12-channel signal digitized using 16-bit ADC21 .
21 Analog-to-Digital Converter
711
And here is an example of very typical MIPS code.
As we may recall, every MIPS (and also ARM in ARM mode or ARM64) instruction has size of 32 bits (or 4
bytes), so such code is array of 32-bit values.
By looking at this screenshot, we may see some kind of pattern.
Vertical red lines are added for clarity:
Figure 5.9: Hiew: very typical MIPS code
Another example of such pattern here is book: 9.5 on page 954.
712
Sparse files
This is sparse file with data scattered amidst almost empty file. Each space character here is in fact zero
byte (which is looks like space). This is a file to program FPGA (Altera Stratix GX device). Of course, files
like these can be compressed easily, but formats like this one are very popular in scientific and engineering
software where efficient access is important while compactness is not.
Figure 5.10: FAR: Sparse file
713
Compressed file
This file is just some compressed archive. It has relatively high entropy and visually looks just chaotic.
This is how compressed and/or encrypted files looks like.
Figure 5.11: FAR: Compressed file
714
CDFS22
OS installations are usually distributed as ISO files which are copies of CD/DVD discs. Filesystem used is
named CDFS, here is you see file names mixed with some additional data. This can be file sizes, pointers
to another directories, file attributes, etc. This is how typical filesystems may look internally.
Figure 5.12: FAR: ISO file: Ubuntu 15 installation CD23
22 Compact Disc File System
715
32-bit x86 executable code
This is how 32-bit x86 executable code looks like. It has not very high entropy, because some bytes
occurred more often than others.
Figure 5.13: FAR: Executable 32-bit x86 code
716
BMP graphics files
BMP files are not compressed, so each byte (or group of bytes) describes each pixel. I’ve found this picture
somewhere inside my installed Windows 8.1:
Figure 5.14: Example picture
You see that this picture has some pixels which unlikely can be compressed very good (around center),
but there are long one-color lines at top and bottom. Indeed, lines like these also looks as lines during
viewing the file:
Figure 5.15: BMP file fragment
5.10.2 Memory “snapshots” comparing

The technique of the straightforward comparison of two memory snapshots in order to see changes was
often used to hack 8-bit computer games and for hacking “high score” files.
For example, if you had a loaded game on an 8-bit computer (there isn’t much memory on these, but the
game usually consumes even less memory) and you know that you have now, let’s say, 100 bullets, you
717
can do a “snapshot” of all memory and back it up to some place. Then shoot once, the bullet count goes
to 99, do a second “snapshot” and then compare both: it must be a byte somewhere which has been 100
at the beginning, and now it is 99.
Considering the fact that these 8-bit games were often written in assembly language and such variables
were global, it can be said for sure which address in memory has holding the bullet count. If you searched
for all references to the address in the disassembled game code, it was not very hard to find a piece of
code decrementing the bullet count, then to write a NOP instruction there, or a couple of NOP-s, and then
have a game with 100 bullets forever. Games on these 8-bit computers were commonly loaded at the
constant address, also, there were not much different versions of each game (commonly just one version
was popular for a long span of time), so enthusiastic gamers knew which bytes must be overwritten (using
the BASIC’s instruction POKE) at which address in order to hack it. This led to “cheat” lists that contained
POKE instructions, published in magazines related to 8-bit games.
Likewise, it is easy to modify “high score” files, this does not work with just 8-bit games. Notice your score
count and back up the file somewhere. When the “high score” count gets different, just compare the two
files, it can even be done with the DOS utility FC24 (“high score” files are often in binary form).
There will be a point where a couple of bytes are different and it is easy to see which ones are holding
the score number. However, game developers are fully aware of such tricks and may defend the program
against it.
Somewhat similar example in this book is: 9.3 on page 942.
A real story from 1999
There was a time of ICQ messenger’s popularity, at least in ex-USSR countries. The messenger had a
peculiarity — some users didn’t want to share their online status with everyone. And you had to ask an
authorization from that user. That user could allow you seeing his/her status, or maybe not.
This is what the author of these lines did:
• Added a user.
• A user appeared in a contact-list, in a “wait for authorization” section.
• Closed ICQ.
• Backed up the ICQ database.
• Loaded ICQ again.
• User authorized.
• Closed ICQ and compared two databases.
It turned out: two database differed by only one byte. In the first version: RESU\x03, in the second:
RESU\x02. (“RESU”, presumably, means “USER”, i.e., a header of a structure where all the information
about user was stored.) That means the information about authorization was stored not at the server, but
at the client. Presumably, 2/3 value reflected authorization status.
Windows registry
It is also possible to compare the Windows registry before and after a program installation.
It is a very popular method of finding which registry elements are used by the program. Perhaps, this is
the reason why the “windows registry cleaner” shareware is so popular.
By the way, this is how to dump Windows registry to text files:
reg export HKLM HKLM.reg
reg export HKCU HKCU.reg
reg export HKCR HKCR.reg
reg export HKU HKU.reg
reg export HKCC HKCC.reg
They can be compared using diff...

24 MS-DOS utility for comparing binary files
718
Engineering software, CADs, etc
If a software uses proprietary files, you can also investigate something here as well. You save file. Then
you add a dot or line or another primitive. Save file, compare. Or move dot, save file, compare.
Blink-comparator
25
Comparison of files or memory snapshots remind us blink-comparator : a device used by astronomers
in past, intended to find moving celestial objects.
Blink-comparator allows to switch quickly between two photographies shot in different time, so astronomer
would spot the difference visually.
By the way, Pluto was discovered by blink-comparator in 1930.
5.11 ISA detection

Often, you can deal with a binary file for an unknown ISA. Perhaps, easiest way to detect ISA is to try
various ones in IDA, objdump or another disassembler.
To achieve this, one should understand a difference between incorrectly disassembled code and correctly
one.
5.11.1 Incorrectly disassembled code

Practicing reverse engineers often have to deal with incorrectly disassembled code.
Disassembling from an incorrect start (x86)
Unlike ARM and MIPS (where any instruction has a length of 2 or 4 bytes), x86 instructions have variable
size, so any disassembler that starts in the middle of a x86 instruction may produce incorrect results.
As an example:
add [ebp-31F7Bh], cl
dec dword ptr [ecx-3277Bh]
dec dword ptr [ebp-2CF7Bh]
inc dword ptr [ebx-7A76F33Ch]
fdiv st(4), st
db 0FFh
dec dword ptr [ecx-21F7Bh]
dec dword ptr [ecx-22373h]
dec dword ptr [ecx-2276Bh]
dec dword ptr [ecx-22B63h]
dec dword ptr [ecx-22F4Bh]
dec dword ptr [ecx-23343h]
jmp dword ptr [esi-74h]
xchg eax, ebp
clc
std
db 0FFh
db 0FFh
mov word ptr [ebp-214h], cs ; <- disassembler finally found right track here
mov word ptr [ebp-238h], ds
mov word ptr [ebp-23Ch], es
mov word ptr [ebp-240h], fs
mov word ptr [ebp-244h], gs
pushf
pop dword ptr [ebp-210h]
mov eax, [ebp+4]
mov [ebp-218h], eax
lea eax, [ebp+4]
mov [ebp-20Ch], eax
mov dword ptr [ebp-2D0h], 10001h
mov eax, [eax-4]
mov [ebp-21Ch], eax
719
mov eax, [ebp+0Ch]
mov [ebp-320h], eax
mov eax, [ebp+10h]
mov [ebp-31Ch], eax
mov eax, [ebp+4]
mov [ebp-314h], eax
call ds:IsDebuggerPresent
mov edi, eax
lea eax, [ebp-328h]
push eax
call sub_407663
pop ecx
test eax, eax
jnz short loc_402D7B
There are incorrectly disassembled instructions at the beginning, but eventually the disassembler gets on
the right track.
How does random noise looks disassembled?
Common properties that can be spotted easily are:

• Unusually big instruction dispersion. The most frequent x86 instructions are PUSH, MOV, CALL, but
here we see instructions from all instruction groups: FPU instructions, IN/OUT instructions, rare and
system instructions, everything mixed up in one single place.
• Big and random values, offsets and immediates.
• Jumps having incorrect offsets, often jumping in the middle of another instructions.
Listing 5.7: random noise (x86)

mov bl, 0Ch
mov ecx, 0D38558Dh
mov eax, ds:2C869A86h
db 67h
mov dl, 0CCh
insb
movsb
push eax
xor [edx-53h], ah
fcom qword ptr [edi-45A0EF72h]
pop esp
pop ss
in eax, dx
dec ebx
push esp
lds esp, [esi-41h]
retf
rcl dword ptr [eax], cl
mov cl, 9Ch
mov ch, 0DFh
push cs
insb
mov esi, 0D9C65E4Dh
imul ebp, [ecx], 66h
pushf
sal dword ptr [ebp-64h], cl
sub eax, 0AC433D64h
out 8Ch, eax
pop ss
sbb [eax], ebx
aas
xchg cl, [ebx+ebx*4+14B31Eh]
jecxz short near ptr loc_58+1
xor al, 0C6h
inc edx
db 36h
pusha
stosb
test [ebx], ebx
720
sub al, 0D3h ; 'L'
pop eax
stosb
loc_58: ; CODE XREF: seg000:0000004A

test [esi], eax
inc ebp
das
db 64h
pop ecx
das
hlt
pop edx
out 0B0h, al
lodsb
push ebx
cdq
out dx, al
sub al, 0Ah
sti
outsd
add dword ptr [edx], 96FCBE4Bh
and eax, 0E537EE4Fh
inc esp
stosd
cdq
push ecx
in al, 0CBh
mov ds:0D114C45Ch, al
mov esi, 659D1985h
Listing 5.8: random noise (x86-64)

lea esi, [rax+rdx*4+43558D29h]
loc_AF3: ; CODE XREF: seg000:0000000000000B46

rcl byte ptr [rsi+rax*8+29BB423Ah], 1
lea ecx, cs:0FFFFFFFFB2A6780Fh
mov al, 96h
mov ah, 0CEh
push rsp
lods byte ptr [esi]
db 2Fh ; /
pop rsp
db 64h
retf 0E993h
cmp ah, [rax+4Ah]

movzx rsi, dword ptr [rbp-25h]
push 4Ah
movzx rdi, dword ptr [rdi+rdx*8]
db 9Ah
rcr byte ptr [rax+1Dh], cl

lodsd
xor [rbp+6CF20173h], edx
xor [rbp+66F8B593h], edx
push rbx
sbb ch, [rbx-0Fh]
stosd
int 87h
db 46h, 4Ch
out 33h, rax
xchg eax, ebp
test ecx, ebp
movsd
721
leave
push rsp
db 16h
xchg eax, esi

pop rdi
loc_B3D: ; CODE XREF: seg000:0000000000000B5F

mov ds:93CA685DF98A90F9h, eax
jnz short near ptr loc_AF3+6
out dx, eax
cwde
mov bh, 5Dh ; ']'
movsb
pop rbp
Listing 5.9: random noise (ARM (ARM mode))

BLNE 0xFE16A9D8
BGE 0x1634D0C
SVCCS 0x450685
STRNVT R5, [PC],#-0x964
LDCGE p6, c14, [R0],#0x168
STCCSL p9, c9, [LR],#0x14C
CMNHIP PC, R10,LSL#22
FLDMIADNV LR!, {D4}
MCR p5, 2, R2,c15,c6, 4
BLGE 0x1139558
BLGT 0xFF9146E4
STRNEB R5, [R4],#0xCA2
STMNEIB R5, {R0,R4,R6,R7,R9-SP,PC}
STMIA R8, {R0,R2-R4,R7,R8,R10,SP,LR}^
STRB SP, [R8],PC,ROR#18
LDCCS p9, c13, [R6,#0x1BC]
LDRGE R8, [R9,#0x66E]
STRNEB R5, [R8],#-0x8C3
STCCSL p15, c9, [R7,#-0x84]
RSBLS LR, R2, R11,ASR LR
SVCGT 0x9B0362
SVCGT 0xA73173
STMNEDB R11!, {R0,R1,R4-R6,R8,R10,R11,SP}
STR R0, [R3],#-0xCE4
LDCGT p15, c8, [R1,#0x2CC]
LDRCCB R1, [R11],-R7,ROR#30
BLLT 0xFED9D58C
BL 0x13E60F4
LDMVSIB R3!, {R1,R4-R7}^
USATNE R10, #7, SP,LSL#11
LDRGEB LR, [R1],#0xE56
STRPLT R9, [LR],#0x567
LDRLT R11, [R1],#-0x29B
SVCNV 0x12DB29
MVNNVS R5, SP,LSL#25
LDCL p8, c14, [R12,#-0x288]
STCNEL p2, c6, [R6,#-0xBC]!
SVCNV 0x2E5A2F
BLX 0x1A8C97E
TEQGE R3, #0x1100000
STMLSIA R6, {R3,R6,R10,R11,SP}
BICPLS R12, R2, #0x5800
BNE 0x7CC408
TEQGE R2, R4,LSL#20
SUBS R1, R11, #0x28C
BICVS R3, R12, R7,ASR R0
LDRMI R7, [LR],R3,LSL#21
BLMI 0x1A79234
STMVCDB R6, {R0-R3,R6,R7,R10,R11}
EORMI R12, R6, #0xC5
MCRRCS p1, 0xF, R1,R3,c2
722
Listing 5.10: random noise (ARM (Thumb mode))
LSRS R3, R6, #0x12
LDRH R1, [R7,#0x2C]
SUBS R0, #0x55 ; 'U'
ADR R1, loc_3C
LDR R2, [SP,#0x218]
CMP R4, #0x86
SXTB R7, R4
LDR R4, [R1,#0x4C]
STR R4, [R4,R2]
STR R0, [R6,#0x20]
BGT 0xFFFFFF72
LDRH R7, [R2,#0x34]
LDRSH R0, [R2,R4]
LDRB R2, [R7,R2]
DCB 0x17
DCB 0xED
STRB R3, [R1,R1]

STR R5, [R0,#0x6C]
LDMIA R3, {R0-R5,R7}
ASRS R3, R2, #3
LDR R4, [SP,#0x2C4]
SVC 0xB5
LDR R6, [R1,#0x40]
LDR R5, =0xB2C5CA32
STMIA R6, {R1-R4,R6}
LDR R1, [R3,#0x3C]
STR R1, [R5,#0x60]
BCC 0xFFFFFF70
LDR R4, [SP,#0x1D4]
STR R5, [R5,#0x40]
ORRS R5, R7
loc_3C ; DATA XREF: ROM:00000006

B 0xFFFFFF98
Listing 5.11: random noise (MIPS little endian)

lw $t9, 0xCB3($t5)
sb $t5, 0x3855($t0)
sltiu $a2, $a0, -0x657A
ldr $t4, -0x4D99($a2)
daddi $s0, $s1, 0x50A4
lw $s7, -0x2353($s4)
bgtzl $a1, 0x17C5C
.byte 0x17
.byte 0xED
.byte 0x4B # K
.byte 0x54 # T
lwc2 $31, 0x66C5($sp)

lwu $s1, 0x10D3($a1)
ldr $t6, -0x204B($zero)
lwc1 $f30, 0x4DBE($s2)
daddiu $t1, $s1, 0x6BD9
lwu $s5, -0x2C64($v1)
cop0 0x13D642D
bne $gp, $t4, 0xFFFF9EF0
lh $ra, 0x1819($s1)
sdl $fp, -0x6474($t8)
jal 0x78C0050
ori $v0, $s2, 0xC634
blez $gp, 0xFFFEA9D4
swl $t8, -0x2CD4($s2)
sltiu $a1, $k0, 0x685
sdc1 $f15, 0x5964($at)
sw $s0, -0x19A6($a1)
723
sltiu $t6, $a3, -0x66AD
lb $t7, -0x4F6($t3)
sd $fp, 0x4B02($a1)
It is also important to keep in mind that cleverly constructed unpacking and decryption code (including
self-modifying) may looks like noise as well, but still execute correctly.
5.11.2 Correctly disassembled code

Each ISA has a dozen of a most used instructions, all the rest are used much less often.
As of x86, it is interesting to know that the fact that function calls (PUSH/CALL/ADD) and MOV instructions
are the most frequently executed pieces of code in almost all programs we use. In other words, CPU is
very busy passing information between levels of abstractions, or, it can be said, it’s very busy switching
between these levels. Regardless type of ISA. This is a cost of splitting problems into several levels of
abstractions (so humans could work with them easier).
5.12 Other things

5.12.1 General idea
A reverse engineer should try to be in programmer’s shoes as often as possible. To take his/her viewpoint
and ask himself, how would one solve some task the specific case.
5.12.2 Order of functions in binary code

All functions located in a single .c or .cpp-file are compiled into corresponding object (.o) file. Later, a linker
puts all object files it needs together, not changing order of functions in them. As a consequence, if you
see two or more consecutive functions, it means, that they were placed together in a single source code
file (unless you’re on border of two object files, of course.) This means these functions have something in
common, that they are from the same API level, from the same library, etc.
This is a real story from practice: once upon a time, the author searched for Twofish-related functions in
a program with CryptoPP library linked, especially encryption/decryption functions.
I found the Twofish::Base::UncheckedSetKey() function, but not others. After peeking into the twofish.cpp
source code 26 , it became clear that all functions are located in one module (twofish.cpp).
So I tried all function that followed Twofish::Base::UncheckedSetKey()—as it happened,
one was Twofish::Enc::ProcessAndXorBlock(), another—Twofish::Dec::ProcessAndXorBlock().
5.12.3 Tiny functions

Tiny functions like empty functions ( 1.3 on page 5) or function which returns just “true” (1) or “false” (0)
( 1.4 on page 7) are very common, and almost all decent compilers tend to put only one such function
into resulting executable code even if there were several similar functions in source code. So, whenever
you see a tiny function consisting just of mov eax, 1 / ret which is referenced (and can be called) from
many places, which are seems unconnected to each other, this may be a result of such optimization.
5.12.4 C++
RTTI ( 3.21.1 on page 557)-data may be also useful for C++ class identification.
5.12.5 Crash on purpose

Often you need to know, which function has been executed, and which is not. You can use a debugger,
but on exotic architectures there may not be the one, so easiest way is to put there an invalid opcode, or
something like INT3 (0xCC). The crash would signal about the very fact this instruction has been executed.
Another example of crashing on purpose: 3.23.4 on page 606.
26 https://github.com/weidai11/cryptopp/blob/b613522794a7633aa2bd81932a98a0b0a51bc04f/twofish.cpp
724
Chapter 6
OS-specific
6.1 Arguments passing methods (calling conventions)

6.1.1 cdecl
This is the most popular method for passing arguments to functions in the C/C++ languages.
The glscaller also must return the value of the stack pointer (ESP) to its initial state after the callee function
exits.
Listing 6.1: cdecl

push arg3
push arg2
push arg1
call function
add esp, 12 ; returns ESP
6.1.2 stdcall
It’s almost the same as cdecl, with the exception that the callee must set ESP to the initial state by
executing the RET x instruction instead of RET,
where x = arguments number * sizeof(int)1 . The caller is not adjusting the stack pointer, there are
no add esp, x instruction.
Listing 6.2: stdcall

push arg3
push arg2
push arg1
call function
function:
... do something ...
ret 12
The method is ubiquitous in win32 standard libraries, but not in win64 (see below about win64).
For example, we can take the function from 1.90 on page 97 and change it slightly by adding the __stdcall
modifier:
int __stdcall f2 (int a, int b, int c)
{
return a*b+c;
};
It is to be compiled in almost the same way as 1.91 on page 97, but you will see RET 12 instead of RET.
SP is not updated in the caller.
As a consequence, the number of function arguments can be easily deduced from the RETN n instruction:
just divide n by 4.
1 The size of an int type variable is 4 in x86 systems and 8 in x64 systems
725
_a$ = 8 ; size = 4
_b$ = 12 ; size = 4
_c$ = 16 ; size = 4
_f2@12 PROC
push ebp
mov ebp, esp
imul eax, DWORD PTR _b$[ebp]
add eax, DWORD PTR _c$[ebp]
pop ebp
ret 12
_f2@12 ENDP
; ...
push 3
push 2
push 1
call _f2@12
push eax
call _printf
add esp, 8
Functions with variable number of arguments
printf()-like functions are, probably, the only case of functions with a variable number of arguments in
C/C++, but it is easy to illustrate an important difference between cdecl and stdcall with their help. Let’s
start with the idea that the compiler knows the argument count of each printf() function call.
However, the called printf(), which is already compiled and located in MSVCRT.DLL (if we talk about Win-
dows), does not have any information about how much arguments were passed, however it can determine
it from the format string.
Thus, if printf() would be a stdcall function and restored stack pointer to its initial state by counting the
number of arguments in the format string, this could be a dangerous situation, when one programmer’s
typo can provoke a sudden program crash. Thus it is not suitable for such functions to use stdcall, cdecl
is better.
6.1.3 fastcall
That’s the general naming for the method of passing some arguments via registers and the rest via the
stack. It worked faster than cdecl/stdcall on older CPUs (because of smaller stack pressure). It may not
help to gain any significant performance on latest (much more complex) CPUs, however.
It is not standardized, so the various compilers can do it differently. It’s a well known caveat: if you have
two DLLs and the one uses another one, and they are built by different compilers with different fastcall
calling conventions, you can expect problems.
Both MSVC and GCC pass the first and second arguments via ECX and EDX and the rest of the arguments
via the stack.
The stack pointer must be restored to its initial state by the callee (like in stdcall).
Listing 6.4: fastcall

push arg3
mov edx, arg2
mov ecx, arg1
call function
function:
.. do something ..
ret 4
For example, we may take the function from 1.90 on page 97 and change it slightly by adding a __fastcall
modifier:
726
int __fastcall f3 (int a, int b, int c)
{
return a*b+c;
};
Here is how it is to be compiled:

_c$ = 8 ; size = 4
@f3@12 PROC
; _a$ = ecx
; _b$ = edx
mov eax, ecx
imul eax, edx
add eax, DWORD PTR _c$[esp-4]
ret 4
@f3@12 ENDP
; ...
mov edx, 2
push 3
lea ecx, DWORD PTR [edx-1]
call @f3@12
push eax
call _printf
add esp, 8
We see that the callee returns SP by using the RETN instruction with an operand.
Which implies that the number of arguments can be deduced easily here as well.
GCC regparm
It is the evolution of fastcall2 in some sense. With the -mregparm option it is possible to set how many
arguments are to be passed via registers (3 is the maximum). Thus, the EAX, EDX and ECX registers are to
be used.
Of course, if the number the of arguments is less than 3, not all 3 registers are to be used.
The caller restores the stack pointer to its initial state.
For example, see ( 1.28.1 on page 305).
Watcom/OpenWatcom
Here it is called “register calling convention”. The first 4 arguments are passed via the EAX, EDX, EBX and
ECX registers. All the rest—via the stack.
These functions have an underscore appended to the function name in order to distinguish them from
those having a different calling convention.
6.1.4 thiscall
This is passing the object’s this pointer to the function-method, in C++.
In MSVC, this is usually passed in the ECX register.
In GCC, the this pointer is passed as the first function-method argument. Thus it will be visible that all
functions in assembly code have an extra argument, in comparison with the source code.
For an example, see ( 3.21.1 on page 542).
727
6.1.5 x86-64
Windows x64
The method of for passing arguments in Win64 somewhat resembles fastcall. The first 4 arguments are
passed via RCX, RDX, R8 and R9, the rest—via the stack. The caller also must prepare space for 32 bytes
or 4 64-bit values, so then the callee can save there the first 4 arguments. Short functions may use the
arguments’ values just from the registers, but larger ones may save their values for further use.
The caller also must return the stack pointer into its initial state.
This calling convention is also used in Windows x86-64 system DLLs (instead of stdcall in win32).
Example:
#include <stdio.h>
void f1(int a, int b, int c, int d, int e, int f, int g)

{
printf ("%d %d %d %d %d %d %d\n", a, b, c, d, e, f, g);
};
int main()
{
f1(1,2,3,4,5,6,7);
};
Listing 6.6: MSVC 2012 /0b

$SG2937 DB '%d %d %d %d %d %d %d', 0aH, 00H
main PROC
sub rsp, 72

mov r9d, 4
mov r8d, 3
mov edx, 2
mov ecx, 1
call f1
xor eax, eax

add rsp, 72
ret 0
main ENDP
a$ = 80
b$ = 88
c$ = 96
d$ = 104
e$ = 112
f$ = 120
g$ = 128
f1 PROC
$LN3:
mov DWORD PTR [rsp+16], edx
sub rsp, 72
mov eax, DWORD PTR g$[rsp]

mov DWORD PTR [rsp+56], eax
mov eax, DWORD PTR f$[rsp]
mov eax, DWORD PTR e$[rsp]
mov eax, DWORD PTR d$[rsp]
mov r9d, DWORD PTR c$[rsp]
728
mov r8d, DWORD PTR b$[rsp]
mov edx, DWORD PTR a$[rsp]
call printf
add rsp, 72
ret 0
f1 ENDP
Here we clearly see how 7 arguments are passed: 4 via registers and the remaining 3 via the stack.
The code of the f1() function’s prologue saves the arguments in the “scratch space”—a space in the stack
intended exactly for this purpose.
This is arranged so because the compiler cannot be sure that there will be enough registers to use without
these 4, which will otherwise be occupied by the arguments until the function’s execution end.
The “scratch space” allocation in the stack is the caller’s duty.
Listing 6.7: Optimizing MSVC 2012 /0b

$SG2777 DB '%d %d %d %d %d %d %d', 0aH, 00H
a$ = 80
b$ = 88
c$ = 96
d$ = 104
e$ = 112
f$ = 120
g$ = 128
f1 PROC
$LN3:
sub rsp, 72
mov eax, DWORD PTR g$[rsp]

mov eax, DWORD PTR f$[rsp]
mov eax, DWORD PTR e$[rsp]
mov r9d, r8d
mov r8d, edx
mov edx, ecx
call printf
add rsp, 72
ret 0
f1 ENDP
main PROC
sub rsp, 72
mov edx, 2
lea ecx, QWORD PTR [rdx-1]
call f1
xor eax, eax

add rsp, 72
ret 0
main ENDP
If we compile the example with optimizations, it is to be almost the same, but the “scratch space” will not
be used, because it won’t be needed.
729
Also take a look on how MSVC 2012 optimizes the loading of primitive values into registers by using LEA
( .1.6 on page 1014). MOV would be 1 byte longer here (5 instead of 4).
Another example of such thing is: 8.2.1 on page 789.
Windows x64: Passing this (C/C++)
The this pointer is passed in RCX, the first argument of the method is in RDX, etc. For an example see: 3.21.1
on page 544.
Linux x64
The way arguments are passed in Linux for x86-64 is almost the same as in Windows, but 6 registers are
used instead of 4 (RDI, RSI, RDX, RCX, R8, R9) and there is no “scratch space”, although the callee may
save the register values in the stack, if it needs/wants to.

.LC0:
.string "%d %d %d %d %d %d %d\n"
f1:
sub rsp, 40
mov eax, DWORD PTR [rsp+48]
mov r9d, ecx
mov DWORD PTR [rsp], r8d
mov ecx, esi
mov r8d, edx
mov edx, edi
mov edi, 1
xor eax, eax
call __printf_chk
add rsp, 40
ret
main:
sub rsp, 24
mov r9d, 6
mov r8d, 5
mov DWORD PTR [rsp], 7
mov ecx, 4
mov edx, 3
mov esi, 2
mov edi, 1
call f1
add rsp, 24
ret
N.B.: here the values are written into the 32-bit parts of the registers (e.g., EAX) but not in the whole
64-bit register (RAX). This is because each write to the low 32-bit part of a register automatically clears
the high 32 bits. Supposedly, it was decided in AMD to do so to simplify porting code to x86-64.
6.1.6 Return values of float and double type

In all conventions except in Win64, the values of type float or double are returned via the FPU register
ST(0).
In Win64, the values of float and double types are returned in the low 32 or 64 bits of the XMM0 register.
6.1.7 Modifying arguments

Sometimes, C/C++ programmers (not limited to these PLs, though), may ask, what can happen if they
modify the arguments?
The answer is simple: the arguments are stored in the stack, that is where the modification takes place.
730
The calling functions is not using them after the callee’s exit (the author of these lines has never seen any
such case in his practice).
#include <stdio.h>
void f(int a, int b)

{
a=a+b;
printf ("%d\n", a);
};

_a$ = 8 ; size = 4
_b$ = 12 ; size = 4
_f PROC
push ebp
mov ebp, esp
mov DWORD PTR _a$[ebp], eax
push ecx
push OFFSET $SG2938 ; '%d', 0aH
call _printf
add esp, 8
pop ebp
ret 0
_f ENDP
So yes, one can modify the arguments easily. Of course, if it is not references in C++ ( 3.21.3 on page 558),
and if you don’t modify data to which a pointer points to, then the effect will not propagate outside the
current function.
Theoretically, after the callee’s return, the caller could get the modified argument and use it somehow.
Maybe if it is written directly in assembly language.
For example, code like this will be generated by usual C/C++ compiler:
push 456 ; will be b
push 123 ; will be a
call f ; f() modifies its first argument
add esp, 2*4
We can rewrite this code like:

push 456 ; will be b
push 123 ; will be a
call f ; f() modifies its first argument
pop eax
add esp, 4
; EAX=1st argument of f() modified in f()
Hard to imagine, why anyone would need this, but this is possible in practice. Nevertheless, the C/C++
languages standards don’t offer any way to do so.
6.1.8 Taking a pointer to function argument

…even more than that, it’s possible to take a pointer to the function’s argument and pass it to another
function:
#include <stdio.h>
// located in some other file

void modify_a (int *a);
void f (int a)
{
modify_a (&a);
printf ("%d\n", a);
731
};
It’s hard to understand how it works until we can see the code:

$SG2796 DB '%d', 0aH, 00H
_a$ = 8
_f PROC
lea eax, DWORD PTR _a$[esp-4] ; just get the address of value in local stack
push eax ; and pass it to modify_a()
call _modify_a
mov ecx, DWORD PTR _a$[esp] ; reload it from the local stack
push ecx ; and pass it to printf()
call _printf
add esp, 12
ret 0
_f ENDP
The address of the place in the stack where a has been passed is just passed to another function. It
modifies the value addressed by the pointer and then printf() prints the modified value.
The observant reader might ask, what about calling conventions where the function’s arguments are
passed in registers?
That’s a situation where the Shadow Space is used.
The input value is copied from the register to the Shadow Space in the local stack, and then this address
is passed to the other function:

$SG2994 DB '%d', 0aH, 00H
a$ = 48
f PROC
mov DWORD PTR [rsp+8], ecx ; save input value in Shadow Space
sub rsp, 40
lea rcx, QWORD PTR a$[rsp] ; get address of value and pass it to modify_a()
call modify_a
mov edx, DWORD PTR a$[rsp] ; reload value from Shadow Space and pass it to
printf()
call printf
add rsp, 40
ret 0
f ENDP
GCC also stores the input value in the local stack:

.LC0:
.string "%d\n"
f:
sub rsp, 24
mov DWORD PTR [rsp+12], edi ; store input value to the local stack
lea rdi, [rsp+12] ; take an address of the value and pass it to
modify_a()
call modify_a
mov edx, DWORD PTR [rsp+12] ; reload value from the local stack and pass it to
printf()
mov esi, OFFSET FLAT:.LC0 ; '%d'
mov edi, 1
xor eax, eax
call __printf_chk
add rsp, 24
ret
GCC for ARM64 does the same, but this space is called Register Save Area here:
732
f:
stp x29, x30, [sp, -32]!
add x29, sp, 0 ; setup FP
add x1, x29, 32 ; calculate address of variable in Register Save Area
str w0, [x1,-4]! ; store input value there
mov x0, x1 ; pass address of variable to the modify_a()
bl modify_a
ldr w1, [x29,28] ; load value from the variable and pass it to printf()
adrp x0, .LC0 ; '%d'
bl printf ; call printf()
ldp x29, x30, [sp], 32
ret
.LC0:
.string "%d\n"
By the way, a similar usage of the Shadow Space is also considered here: 3.17.1 on page 519.
6.1.9 Python ctypes problem (x86 assembly homework)

A Python ctypes module can call external functions in DLLs, .so’s, etc. But calling convention (for 32-bit
environment) must be specified explicitely:
"ctypes" exports the *cdll*, and on Windows *windll* and *oledll*
objects, for loading dynamic link libraries.
You load libraries by accessing them as attributes of these objects.

*cdll* loads libraries which export functions using the standard
"cdecl" calling convention, while *windll* libraries call functions
using the "stdcall" calling convention.
( https://docs.python.org/3/library/ctypes.html )
In fact, we can modify ctypes module (or any other caller code), so that it will successfully call external
cdecl or stdcall functions, without knowledge, which is where. (Number of arguments, however, is to be
specified).
This is possible to solve using maybe 5-10 x86 assembly instructions in caller. Try to find out these.
6.1.10 Cdecl example: a DLL

Let’s back to the fact that this is not very important how to declare the main() function: 1.9.2 on page 32.
This is a real story: once upon a time I wanted to replace an original DLL file in some software by mine.
First I enumerated names of all DLL exports and made a function in my own replacement DLL for each
function in the original DLL, like:
void function1 ()
{
write_to_log ("function1() called\n");
};
I wanted to see, which functions are called during run, and when. However, I was in hurry and had
no time to deduce arguments count for each function, let alone data types. So each function in my
replacement DLL had no argument whatsoever. But everything worked, because all functions had cdecl
calling convention. (It wouldn’t work if functions had stdcall calling convention.) It also worked for x64
version.
And then I did a next step: I deduced argument types for some functions. But I made several mistakes,
for example, the original function took 3 arguments, but I knew only about 2, etc.
Still, it worked. At the beginning, my replacement DLL just ignored all arguments. Then, it ignored the
3rd argument.
733
6.2 Thread Local Storage
TLS is a data area, specific to each thread. Every thread can store what it needs there. One well-known
example is the C standard global variable errno.
Multiple threads may simultaneously call functions which return an error code in errno, so a global variable
will not work correctly here for multi-threaded programs, so errno must be stored in the TLS.
In the C++11 standard, a new thread_local modifier was added, showing that each thread has its own
version of the variable, it can be initialized, and it is located in the TLS 3 :
Listing 6.14: C++11

#include <iostream>
#include <thread>
thread_local int tmp=3;
int main()
{
std::cout << tmp << std::endl;
};
Compiled in MinGW GCC 4.8.1, but not in MSVC 2012.

If we talk about PE files, in the resulting executable file, the tmp variable is to be allocated in the section
devoted to the TLS.
6.2.1 Linear congruential generator revisited

The pseudorandom number generator we considered earlier 1.29 on page 336 has a flaw: it’s not thread-
safe, because it has an internal state variable which can be read and/or modified in different threads
simultaneously.
Win32
Uninitialized TLS data
One solution is to add __declspec( thread ) modifier to the global variable, then it will be allocated in
the TLS (line 9):
1 #include <stdint.h>
2 #include <windows.h>
3 #include <winnt.h>
4
5 // from the Numerical Recipes book:
6 #define RNG_a 1664525
7 #define RNG_c 1013904223
8
9 __declspec( thread ) uint32_t rand_state;
10
11 void my_srand (uint32_t init)
12 {
13 rand_state=init;
14 }
15
16 int my_rand ()
17 {
18 rand_state=rand_state*RNG_a;
19 rand_state=rand_state+RNG_c;
20 return rand_state & 0x7fff;
21 }
22
23 int main()
24 {
25 my_srand(0x12345678);
26 printf ("%d\n", my_rand());
3 C11 also has thread support, optional though
734
27 };
Hiew shows us that there is a new PE section in the executable file: .tls.

_TLS SEGMENT
_TLS ENDS
_DATA SEGMENT
$SG84851 DB '%d', 0aH, 00H
_DATA ENDS
_TEXT SEGMENT
_init$ = 8 ; size = 4
_my_srand PROC
; FS:0=address of TIB
mov eax, DWORD PTR fs:__tls_array ; displayed in IDA as FS:2Ch
; EAX=address of TLS of process
mov ecx, DWORD PTR __tls_index
mov ecx, DWORD PTR [eax+ecx*4]
; ECX=current TLS segment
mov eax, DWORD PTR _init$[esp-4]
mov DWORD PTR _rand_state[ecx], eax
ret 0
_my_srand ENDP
_my_rand PROC
; FS:0=address of TIB
mov eax, DWORD PTR fs:__tls_array ; displayed in IDA as FS:2Ch
; EAX=address of TLS of process
mov ecx, DWORD PTR __tls_index
mov ecx, DWORD PTR [eax+ecx*4]
; ECX=current TLS segment
imul eax, DWORD PTR _rand_state[ecx], 1664525
add eax, 1013904223 ; 3c6ef35fH
mov DWORD PTR _rand_state[ecx], eax
and eax, 32767 ; 00007fffH
ret 0
_my_rand ENDP
_TEXT ENDS
rand_state is now in the TLS segment, and each thread has its own version of this variable.
Here is how it’s accessed: load the address of the TIB from FS:2Ch, then add an additional index (if needed),
then calculate the address of the TLS segment.
Then it’s possible to access the rand_state variable through the ECX register, which points to an unique
area in each thread.
The FS: selector is familiar to every reverse engineer, it is specially used to always point to TIB, so it would
be fast to load the thread-specific data.
The GS: selector is used in Win64 and the address of the TLS is 0x58:

_TLS SEGMENT
rand_state DD 01H DUP (?)
_TLS ENDS
_DATA SEGMENT
$SG85451 DB '%d', 0aH, 00H
_DATA ENDS
_TEXT SEGMENT
init$ = 8
my_srand PROC
mov edx, DWORD PTR _tls_index
735
mov rax, QWORD PTR gs:88 ; 58h
mov r8d, OFFSET FLAT:rand_state
mov rax, QWORD PTR [rax+rdx*8]
mov DWORD PTR [r8+rax], ecx
ret 0
my_srand ENDP
my_rand PROC
mov rax, QWORD PTR gs:88 ; 58h
mov ecx, DWORD PTR _tls_index
mov edx, OFFSET FLAT:rand_state
mov rcx, QWORD PTR [rax+rcx*8]
imul eax, DWORD PTR [rcx+rdx], 1664525 ; 0019660dH
add eax, 1013904223 ; 3c6ef35fH
mov DWORD PTR [rcx+rdx], eax
and eax, 32767 ; 00007fffH
ret 0
my_rand ENDP
_TEXT ENDS
Initialized TLS data
Let’s say, we want to set some fixed value to rand_state, so in case the programmer forgets to, the
rand_state variable would be initialized to some constant anyway (line 9):
1 #include <stdint.h>
2 #include <windows.h>
3 #include <winnt.h>
4
5 // from the Numerical Recipes book:
6 #define RNG_a 1664525
7 #define RNG_c 1013904223
8
9 __declspec( thread ) uint32_t rand_state=1234;
10
11 void my_srand (uint32_t init)
12 {
13 rand_state=init;
14 }
15
16 int my_rand ()
17 {
18 rand_state=rand_state*RNG_a;
19 rand_state=rand_state+RNG_c;
20 return rand_state & 0x7fff;
21 }
22
23 int main()
24 {
25 printf ("%d\n", my_rand());
26 };
The code is not different from what we already saw, but in IDA we see:
.tls:00404000 ; Segment type: Pure data
.tls:00404000 ; Segment permissions: Read/Write
.tls:00404000 _tls segment para public 'DATA' use32
.tls:00404000 assume cs:_tls
.tls:00404000 ;org 404000h
.tls:00404000 TlsStart db 0 ; DATA XREF: .rdata:TlsDirectory
.tls:00404001 db 0
.tls:00404002 db 0
.tls:00404003 db 0
.tls:00404004 dd 1234
.tls:00404008 TlsEnd db 0 ; DATA XREF: .rdata:TlsEnd_ptr
...
736
1234 is there and every time a new thread starts, a new TLS is allocated for it, and all this data, including
1234, will be copied there.
This is a typical scenario:
• Thread A is started. A TLS is created for it, 1234 is copied to rand_state.
• The my_rand() function is called several times in thread A.
rand_state is different from 1234.
• Thread B is started. A TLS is created for it, 1234 is copied to rand_state, while thread A has a
different value in the same variable.
TLS callbacks
But what if the variables in the TLS have to be filled with some data that must be prepared in some unusual
way?
Let’s say, we’ve got the following task: the programmer can forget to call the my_srand() function to
initialize the PRNG, but the generator has to be initialized at start with something truly random, instead
of 1234. This is a case in which TLS callbacks can be used.
The following code is not very portable due to the hack, but nevertheless, you get the idea.
What we do here is define a function (tls_callback()) which is to be called before the process and/or
thread start.
The function initializes the PRNG with the value returned by GetTickCount() function.
#include <stdint.h>
#include <winnt.h>
// from the Numerical Recipes book:

#define RNG_c 1013904223
__declspec( thread ) uint32_t rand_state;
void my_srand (uint32_t init)

{
rand_state=init;
}
void NTAPI tls_callback(PVOID a, DWORD dwReason, PVOID b)

{
my_srand (GetTickCount());
}
#pragma data_seg(".CRT$XLB")
PIMAGE_TLS_CALLBACK p_thread_callback = tls_callback;
#pragma data_seg()
int my_rand ()
{
}
int main()
{
// rand_state is already initialized at the moment (using GetTickCount())
printf ("%d\n", my_rand());
};
Let’s see it in IDA:

.text:00401020 TlsCallback_0 proc near ; DATA XREF: .rdata:TlsCallbacks
.text:00401020 call ds:GetTickCount
737
.text:00401027 call my_srand
.text:0040102C pop ecx
.text:0040102D retn 0Ch
.text:0040102D TlsCallback_0 endp
...
.rdata:004020C0 TlsCallbacks dd offset TlsCallback_0 ; DATA XREF: .rdata:TlsCallbacks_ptr
...
.rdata:00402118 TlsDirectory dd offset TlsStart

.rdata:0040211C TlsEnd_ptr dd offset TlsEnd
.rdata:00402120 TlsIndex_ptr dd offset TlsIndex
.rdata:00402124 TlsCallbacks_ptr dd offset TlsCallbacks
.rdata:00402128 TlsSizeOfZeroFill dd 0
.rdata:0040212C TlsCharacteristics dd 300000h
TLS callback functions are sometimes used in unpacking routines to obscure their processing.
Some people may be confused and be in the dark that some code executed right before the OEP4 .
Linux
Here is how a thread-local global variable is declared in GCC:

__thread uint32_t rand_state=1234;
This is not the standard C/C++ modifier, but a rather GCC-specific one 5 .
The GS: selector is also used to access the TLS, but in a somewhat different way:

.text:08048460 my_srand proc near
.text:08048460
.text:08048460 arg_0 = dword ptr 4
.text:08048460
.text:08048460 mov eax, [esp+arg_0]
.text:08048464 mov gs:0FFFFFFFCh, eax
.text:0804846A retn
.text:0804846A my_srand endp
.text:08048470 my_rand proc near

.text:08048470 imul eax, gs:0FFFFFFFCh, 19660Dh
.text:0804847B add eax, 3C6EF35Fh
.text:08048480 mov gs:0FFFFFFFCh, eax
.text:08048486 and eax, 7FFFh
.text:0804848B retn
.text:0804848B my_rand endp
More about it: [Ulrich Drepper, ELF Handling For Thread-Local Storage, (2013)]6 .
6.3 System calls (syscall-s)

As we know, all running processes inside an OS are divided into two categories: those having full access
to the hardware (“kernel space”) and those that do not (“user space”).
The OS kernel and usually the drivers are in the first category.
All applications are usually in the second category.
For example, Linux kernel is in kernel space, but Glibc in user space.
4 Original Entry Point
738
This separation is crucial for the safety of the OS: it is very important not to give to any process the
possibility to screw up something in other processes or even in the OS kernel. On the other hand, a
failing driver or error inside the OS’s kernel usually leads to a kernel panic or BSOD7 .
The protection in the x86 processors allows to separate everything into 4 levels of protection (rings), but
both in Linux and in Windows only two are used: ring0 (“kernel space”) and ring3 (“user space”).
System calls (syscall-s) are a point where these two areas are connected.
It can be said that this is the main API provided to applications.
As in Windows NT, the syscalls table resides in the SSDT8 .
The usage of syscalls is very popular among shellcode and computer viruses authors, because it is hard
to determine the addresses of needed functions in the system libraries, but it is easier to use syscalls.
However, much more code has to be written due to the lower level of abstraction of the API.
It is also worth noting that the syscall numbers may be different in various OS versions.
6.3.1 Linux
In Linux, a syscall is usually called via int 0x80. The call’s number is passed in the EAX register, and any
other parameters —in the other registers.
Listing 6.19: A simple example of the usage of two syscalls

section .text
global _start
_start:
mov edx,len ; buffer len
mov ecx,msg ; buffer
mov ebx,1 ; file descriptor. 1 is for stdout
mov eax,4 ; syscall number. 4 is for sys_write
int 0x80
mov eax,1 ; syscall number. 1 is for sys_exit

int 0x80
section .data
msg db 'Hello, world!',0xa

len equ $ - msg
Compilation:
nasm -f elf32 1.s
ld 1.o
The full list of syscalls in Linux: http://go.yurichev.com/17319.

For system calls interception and tracing in Linux, strace( 7.2.3 on page 781) can be used.
6.3.2 Windows
Here they are called via int 0x2e or using the special x86 instruction SYSENTER.
The full list of syscalls in Windows: http://go.yurichev.com/17320.
Further reading:
“Windows Syscall Shellcode” by Piotr Bania: http://go.yurichev.com/17321.
6.4 Linux
6.4.1 Position-independent code
While analyzing Linux shared (.so) libraries, one may frequently spot this code pattern:
7 Blue Screen of Death
8 System Service Dispatch Table
739
Listing 6.20: libc-2.17.so x86
.text:0012D5E3 __x86_get_pc_thunk_bx proc near ; CODE XREF: sub_17350+3
.text:0012D5E3 ; sub_173CC+4 ...
.text:0012D5E3 mov ebx, [esp+0]
.text:0012D5E6 retn
.text:0012D5E6 __x86_get_pc_thunk_bx endp
...
.text:000576C0 sub_576C0 proc near ; CODE XREF: tmpfile+73
...
.text:000576C0 push ebp

.text:000576C1 mov ecx, large gs:0
.text:000576C8 push edi
.text:000576C9 push esi
.text:000576CA push ebx
.text:000576CB call __x86_get_pc_thunk_bx
.text:000576D0 add ebx, 157930h
.text:000576D6 sub esp, 9Ch
...
.text:000579F0 lea eax, (a__gen_tempname - 1AF000h)[ebx] ; "__gen_tempname"

.text:000579F6 mov [esp+0ACh+var_A0], eax
.text:000579FA lea eax, (a__SysdepsPosix - 1AF000h)[ebx] ;
"../sysdeps/posix/tempname.c"
.text:00057A00 mov [esp+0ACh+var_A8], eax
.text:00057A04 lea eax, (aInvalidKindIn_ - 1AF000h)[ebx] ;
"! \"invalid KIND in __gen_tempname\""
.text:00057A0A mov [esp+0ACh+var_A4], 14Ah
.text:00057A12 mov [esp+0ACh+var_AC], eax
.text:00057A15 call __assert_fail
All pointers to strings are corrected by some constants and the value in EBX, which is calculated at the
beginning of each function.
This is the so-called PIC, it is intended to be executable if placed at any random point of memory, that is
why it cannot contain any absolute memory addresses.
PIC was crucial in early computer systems and is still crucial today in embedded systems without virtual
memory support (where all processes are placed in a single continuous memory block).
It is also still used in *NIX systems for shared libraries, since they are shared across many processes
while loaded in memory only once. But all these processes can map the same shared library at different
addresses, so that is why a shared library has to work correctly without using any absolute addresses.
Let’s do a simple experiment:
#include <stdio.h>
int global_variable=123;
int f1(int var)

{
int rt=global_variable+var;
printf ("returning %d\n", rt);
return rt;
};
Let’s compile it in GCC 4.7.3 and see the resulting .so file in IDA:
gcc -fPIC -shared -O3 -o 1.so 1.c

.text:00000440 public __x86_get_pc_thunk_bx
.text:00000440 __x86_get_pc_thunk_bx proc near ; CODE XREF: _init_proc+4
.text:00000440 ; deregister_tm_clones+4 ...
.text:00000440 mov ebx, [esp+0]
740
.text:00000443 retn
.text:00000443 __x86_get_pc_thunk_bx endp
.text:00000570 public f1
.text:00000570 f1 proc near
.text:00000570
.text:00000570 var_1C = dword ptr -1Ch
.text:00000570 var_18 = dword ptr -18h
.text:00000570 var_14 = dword ptr -14h
.text:00000570
.text:00000570 sub esp, 1Ch
.text:00000573 mov [esp+1Ch+var_8], ebx
.text:00000577 call __x86_get_pc_thunk_bx
.text:0000057C add ebx, 1A84h
.text:00000582 mov [esp+1Ch+var_4], esi
.text:00000586 mov eax, ds:(global_variable_ptr - 2000h)[ebx]
.text:0000058C mov esi, [eax]
.text:0000058E lea eax, (aReturningD - 2000h)[ebx] ; "returning %d\n"
.text:00000594 add esi, [esp+1Ch+arg_0]
.text:00000598 mov [esp+1Ch+var_18], eax
.text:0000059C mov [esp+1Ch+var_1C], 1
.text:000005A3 mov [esp+1Ch+var_14], esi
.text:000005A7 call ___printf_chk
.text:000005AC mov eax, esi
.text:000005AE mov ebx, [esp+1Ch+var_8]
.text:000005B2 mov esi, [esp+1Ch+var_4]
.text:000005B6 add esp, 1Ch
.text:000005B9 retn
.text:000005B9 f1 endp
That’s it: the pointers to «returning %d\n» and global_variable are to be corrected at each function exe-
cution.
The __x86_get_pc_thunk_bx() function returns in EBX the address of the point after a call to itself (0x57C
here).
That’s a simple way to get the value of the program counter (EIP) at some point. The 0x1A84 constant
is related to the difference between this function’s start and the so-called Global Offset Table Proce-
dure Linkage Table (GOT PLT), the section right after the Global Offset Table (GOT), where the pointer
to global_variable is. IDA shows these offsets in their processed form to make them easier to understand,
but in fact the code is:
.text:00000577 call __x86_get_pc_thunk_bx
.text:0000057C add ebx, 1A84h
.text:00000582 mov [esp+1Ch+var_4], esi
.text:00000586 mov eax, [ebx-0Ch]
.text:0000058C mov esi, [eax]
.text:0000058E lea eax, [ebx-1A30h]
Here EBX points to the GOT PLT section and to calculate a pointer to global_variable (which is stored in
the GOT), 0xC must be subtracted.
To calculate pointer to the «returning %d\n» string, 0x1A30 must be subtracted.
By the way, that is the reason why the AMD64 instruction set supports RIP9 -relative addressing — to
simplify PIC-code.
Let’s compile the same C code using the same GCC version, but for x64.
IDA would simplify the resulting code but would suppress the RIP-relative addressing details, so we are
going to use objdump instead of IDA to see everything:
0000000000000720 <f1>:
720: 48 8b 05 b9 08 20 00 mov rax,QWORD PTR [rip+0x2008b9] ;
200fe0 <_DYNAMIC+0x1d0>
727: 53 push rbx
728: 89 fb mov ebx,edi
9 program counter in AMD64
741
72a: 48 8d 35 20 00 00 00 lea rsi,[rip+0x20] ; 751 <_fini+0x9>
731: bf 01 00 00 00 mov edi,0x1
736: 03 18 add ebx,DWORD PTR [rax]
738: 31 c0 xor eax,eax
73a: 89 da mov edx,ebx
73c: e8 df fe ff ff call 620 <__printf_chk@plt>
741: 89 d8 mov eax,ebx
743: 5b pop rbx
744: c3 ret
0x2008b9 is the difference between the address of the instruction at 0x720 and global_variable, and 0x20
is the difference between the address of the instruction at 0x72A and the «returning %d\n» string.
As you might see, the need to recalculate addresses frequently makes execution slower (it is better in
x64, though).
So it is probably better to link statically if you care about performance [see: Agner Fog, Optimizing software
in C++ (2015)].
Windows
The PIC mechanism is not used in Windows DLLs. If the Windows loader needs to load DLL on another
base address, it “patches” the DLL in memory (at the FIXUP places) in order to correct all addresses.
This implies that several Windows processes cannot share an once loaded DLL at different addresses in
different process’ memory blocks — since each instance that’s loaded in memory is fixed to work only at
these addresses..
6.4.2 LD_PRELOAD hack in Linux

This allows us to load our own dynamic libraries before others, even before system ones, like libc.so.6.
This, in turn, allows us to “substitute” our written functions before the original ones in the system libraries.
For example, it is easy to intercept all calls to time(), read(), write(), etc.
Let’s see if we can fool the uptime utility. As we know, it tells how long the computer has been working.
With the help of strace( 7.2.3 on page 781), it is possible to see that the utility takes this information the
/proc/uptime file:
$ strace uptime
...
open("/proc/uptime", O_RDONLY) = 3
lseek(3, 0, SEEK_SET) = 0
read(3, "416166.86 414629.38\n", 2047) = 20
...
It is not a real file on disk, it is a virtual one and its contents are generated on fly in the Linux kernel. There
are just two numbers:
$ cat /proc/uptime
416690.91 415152.03
10
What we can learn from Wikipedia :
The first number is the total number of seconds the system has been up. The second
number is how much of that time the machine has spent idle, in seconds.
Let’s try to write our own dynamic library with the open(), read(), close() functions working as we need.
At first, our open() will compare the name of the file to be opened with what we need and if it is so, it will
write down the descriptor of the file opened.
Second, read(), if called for this file descriptor, will substitute the output, and in the rest of the cases will
call the original read() from libc.so.6. And also close(), will note if the file we are currently following is to
be closed.
10 https://en.wikipedia.org/wiki/Uptime
742
We are going to use the dlopen() and dlsym() functions to determine the original function addresses in
libc.so.6.
We need them because we must pass control to the “real” functions.
On the other hand, if we intercepted strcmp() and monitored each string comparisons in the program,
then we would have to implement our version of strcmp(), and not use the original function 11 , that would
be easier.
#include <stdio.h>
#include <stdarg.h>
#include <stdlib.h>
#include <stdbool.h>
#include <unistd.h>
#include <dlfcn.h>
#include <string.h>
void *libc_handle = NULL;

int (*open_ptr)(const char *, int) = NULL;
int (*close_ptr)(int) = NULL;
ssize_t (*read_ptr)(int, void*, size_t) = NULL;
bool inited = false;
_Noreturn void die (const char * fmt, ...)

{
va_list va;
va_start (va, fmt);
vprintf (fmt, va);

exit(0);
};
static void find_original_functions ()

{
if (inited)
return;
libc_handle = dlopen ("libc.so.6", RTLD_LAZY);

if (libc_handle==NULL)
die ("can't open libc.so.6\n");
open_ptr = dlsym (libc_handle, "open");

if (open_ptr==NULL)
die ("can't find open()\n");
close_ptr = dlsym (libc_handle, "close");

if (close_ptr==NULL)
die ("can't find close()\n");
read_ptr = dlsym (libc_handle, "read");

if (read_ptr==NULL)
die ("can't find read()\n");
inited = true;
}
static int opened_fd=0;
int open(const char *pathname, int flags)

{
find_original_functions();
int fd=(*open_ptr)(pathname, flags);

if (strcmp(pathname, "/proc/uptime")==0)
opened_fd=fd; // that's our file! record its file descriptor
else
opened_fd=0;
return fd;
11 For example, here is how simple strcmp() interception works in this article 12 written by Yong Huang
743
};
int close(int fd)

{
if (fd==opened_fd)
opened_fd=0; // the file is not opened anymore
return (*close_ptr)(fd);
};
ssize_t read(int fd, void *buf, size_t count)

{
if (opened_fd!=0 && fd==opened_fd)

{
// that's our file!
return snprintf (buf, count, "%d %d", 0x7fffffff, 0x7fffffff)+1;
};
// not our file, go to real read() function
return (*read_ptr)(fd, buf, count);
};
( Source code )
Let’s compile it as common dynamic library:
gcc -fpic -shared -Wall -o fool_uptime.so fool_uptime.c -ldl
Let’s run uptime while loading our library before the others:
LD_PRELOAD=`pwd`/fool_uptime.so uptime
And we see:
01:23:02 up 24855 days, 3:14, 3 users, load average: 0.00, 0.01, 0.05
If the LD_PRELOAD
environment variable always points to the filename and path of our library, it is to be loaded for all starting
programs.
More examples:
• Very simple interception of the strcmp() (Yong Huang) http://go.yurichev.com/17143
• Kevin Pulo—Fun with LD_PRELOAD. A lot of examples and ideas. yurichev.com
• File functions interception for compression/decompression files on fly (zlibc). http://go.yurichev.
com/17146
6.5 Windows NT
6.5.1 CRT (win32)
Does the program execution start right at the main() function? No, it does not.
If we would open any executable file in IDA or HIEW, we can see OEP pointing to some another code block.
This code is doing some maintenance and preparations before passing control flow to our code. It is called
startup-code or CRT code (C RunTime).
The main() function takes an array of the arguments passed on the command line, and also one with
environment variables. But in fact a generic string is passed to the program, the CRT code finds the
spaces in it and cuts it in parts. The CRT code also prepares the environment variables array envp.
As for GUI13 win32 applications, WinMain is used instead of main(), having its own arguments:
13 Graphical User Interface
744
int CALLBACK WinMain(
_In_ HINSTANCE hInstance,
_In_ HINSTANCE hPrevInstance,
_In_ LPSTR lpCmdLine,
_In_ int nCmdShow
);
The CRT code prepares them as well.

Also, the number returned by the main() function is the exit code.
It may be passed in CRT to the ExitProcess() function, which takes the exit code as an argument.
Usually, each compiler has its own CRT code.
Here is a typical CRT code for MSVC 2008.

1 ___tmainCRTStartup proc near
2
3 var_24 = dword ptr -24h
4 var_20 = dword ptr -20h
5 var_1C = dword ptr -1Ch
6 ms_exc = CPPEH_RECORD ptr -18h
7
8 push 14h
9 push offset stru_4092D0
10 call __SEH_prolog4
11 mov eax, 5A4Dh
12 cmp ds:400000h, ax
13 jnz short loc_401096
14 mov eax, ds:40003Ch
15 cmp dword ptr [eax+400000h], 4550h
17 mov ecx, 10Bh
18 cmp [eax+400018h], cx
20 cmp dword ptr [eax+400074h], 0Eh
21 jbe short loc_401096
22 xor ecx, ecx
23 cmp [eax+4000E8h], ecx
24 setnz cl
25 mov [ebp+var_1C], ecx
26 jmp short loc_40109A
27
28
29 loc_401096: ; CODE XREF: ___tmainCRTStartup+18
30 ; ___tmainCRTStartup+29 ...
31 and [ebp+var_1C], 0
32
33 loc_40109A: ; CODE XREF: ___tmainCRTStartup+50
34 push 1
35 call __heap_init
36 pop ecx
37 test eax, eax
38 jnz short loc_4010AE
39 push 1Ch
40 call _fast_error_exit
41 pop ecx
42
43 loc_4010AE: ; CODE XREF: ___tmainCRTStartup+60
44 call __mtinit
45 test eax, eax
46 jnz short loc_4010BF
47 push 10h
48 call _fast_error_exit
49 pop ecx
50
51 loc_4010BF: ; CODE XREF: ___tmainCRTStartup+71
52 call sub_401F2B
53 and [ebp+ms_exc.disabled], 0
745
54 call __ioinit
55 test eax, eax
56 jge short loc_4010D9
57 push 1Bh
58 call __amsg_exit
59 pop ecx
60
61 loc_4010D9: ; CODE XREF: ___tmainCRTStartup+8B
62 call ds:GetCommandLineA
63 mov dword_40B7F8, eax
64 call ___crtGetEnvironmentStringsA
65 mov dword_40AC60, eax
66 call __setargv
67 test eax, eax
68 jge short loc_4010FF
69 push 8
70 call __amsg_exit
71 pop ecx
72
73 loc_4010FF: ; CODE XREF: ___tmainCRTStartup+B1
74 call __setenvp
75 test eax, eax
76 jge short loc_401110
77 push 9
78 call __amsg_exit
79 pop ecx
80
81 loc_401110: ; CODE XREF: ___tmainCRTStartup+C2
82 push 1
83 call __cinit
84 pop ecx
85 test eax, eax
86 jz short loc_401123
87 push eax
88 call __amsg_exit
89 pop ecx
90
91 loc_401123: ; CODE XREF: ___tmainCRTStartup+D6
92 mov eax, envp
93 mov dword_40AC80, eax
94 push eax ; envp
95 push argv ; argv
96 push argc ; argc
97 call _main
98 add esp, 0Ch
99 mov [ebp+var_20], eax
100 cmp [ebp+var_1C], 0
101 jnz short $LN28
102 push eax ; uExitCode
103 call $LN32
104
105 $LN28: ; CODE XREF: ___tmainCRTStartup+105
106 call __cexit
107 jmp short loc_401186
108
109
110 $LN27: ; DATA XREF: .rdata:stru_4092D0
111 mov eax, [ebp+ms_exc.exc_ptr] ; Exception filter 0 for function 401044
112 mov ecx, [eax]
113 mov ecx, [ecx]
114 mov [ebp+var_24], ecx
115 push eax
116 push ecx
117 call __XcptFilter
118 pop ecx
119 pop ecx
120
121 $LN24:
122 retn
123
746
124
125 $LN14: ; DATA XREF: .rdata:stru_4092D0
126 mov esp, [ebp+ms_exc.old_esp] ; Exception handler 0 for function 401044
127 mov eax, [ebp+var_24]
128 mov [ebp+var_20], eax
129 cmp [ebp+var_1C], 0
130 jnz short $LN29
131 push eax ; int
132 call __exit
133
134
135 $LN29: ; CODE XREF: ___tmainCRTStartup+135
136 call __c_exit
137
138 loc_401186: ; CODE XREF: ___tmainCRTStartup+112
139 mov [ebp+ms_exc.disabled], 0FFFFFFFEh
140 mov eax, [ebp+var_20]
141 call __SEH_epilog4
142 retn
Here we can see calls to GetCommandLineA() (line 62), then to setargv() (line 66) and setenvp() (line
74), which apparently fill the global variables argc, argv, envp.
Finally, main() is called with these arguments (line 97).
There are also calls to functions with self-describing names like heap_init() (line 35), ioinit() (line 54).
The heap is indeed initialized in the CRT. If you try to use malloc() in a program without CRT, it will exit
abnormally with the following error:
runtime error R6030
- CRT not initialized
Global object initializations in C++ is also occur in the CRT before the execution of main(): 3.21.4 on
page 564.
The value that main() returns is passed to cexit(), or in $LN32, which in turn calls doexit().
Is it possible to get rid of the CRT? Yes, if you know what you are doing.
The MSVC’s linker has the /ENTRY option for setting an entry point.
int main()
{
MessageBox (NULL, "hello, world", "caption", MB_OK);
};
Let’s compile it in MSVC 2008.

cl no_crt.c user32.lib /link /entry:main
We are getting a runnable .exe with size 2560 bytes, that has a PE header in it, instructions calling
MessageBox, two strings in the data segment, the MessageBox function imported from user32.dll and
nothing else.
This works, but you cannot write WinMain with its 4 arguments instead of main().
To be precise, you can, but the arguments are not prepared at the moment of execution.
By the way, it is possible to make the .exe even shorter by aligning the PE sections at less than the default
4096 bytes.
cl no_crt.c user32.lib /link /entry:main /align:16
Linker says:
LINK : warning LNK4108: /ALIGN specified without /DRIVER; image may not run
747
We get an .exe that’s 720 bytes. It can be executed in Windows 7 x86, but not in x64 (an error message
will be shown when you try to execute it).
With even more efforts, it is possible to make the executable even shorter, but as you can see, compatibility
problems arise quickly.
6.5.2 Win32 PE
PE is an executable file format used in Windows. The difference between .exe, .dll and .sys is that .exe
and .sys usually do not have exports, only imports.
A DLL14 , just like any other PE-file, has an entry point (OEP) (the function DllMain() is located there) but
this function usually does nothing. .sys is usually a device driver. As of drivers, Windows requires the
checksum to be present in the PE file and for it to be correct 15 .
Starting at Windows Vista, a driver’s files must also be signed with a digital signature. It will fail to load
otherwise.
Every PE file begins with tiny DOS program that prints a message like “This program cannot be run in DOS
mode.”—if you run this program in DOS or Windows 3.1 (OS-es which are not aware of the PE format), this
message will be printed.
Terminology
• Module—a separate file, .exe or .dll.

• Process—a program loaded into memory and currently running. Commonly consists of one .exe file
and bunch of .dll files.
• Process memory—the memory a process works with. Each process has its own. There usually are
loaded modules, memory of the stack, heap(s), etc.
• VA16 —an address which is to be used in program while runtime.
• Base address (of module)—the address within the process memory at which the module is to be
loaded. OS loader may change it, if the base address is already occupied by another module just
loaded before.
• RVA17 —the VA-address minus the base address.
Many addresses in PE-file tables use RVA-addresses.
• IAT18 —an array of addresses of imported symbols 19 . Sometimes, the IMAGE_DIRECTORY_ENTRY_IAT
data directory points at the IAT. It is worth noting that IDA (as of 6.1) may allocate a pseudo-section
named .idata for IAT, even if the IAT is a part of another section!
• INT20 —an array of names of symbols to be imported21 .
Base address
The problem is that several module authors can prepare DLL files for others to use and it is not possible
to reach an agreement which addresses is to be assigned to whose modules.
So that is why if two necessary DLLs for a process have the same base address, one of them will be
loaded at this base address, and the other—at some other free space in process memory, and each
virtual addresses in the second DLL will be corrected.
With MSVC the linker often generates the .exe files with a base address of 0x400000 22 , and with the code
section starting at 0x401000. This means that the RVA of the start of the code section is 0x1000.
23
DLLs are often generated by MSVC’s linker with a base address of 0x10000000 .
14 Dynamic-Link Library
15 For example, Hiew( 7.1 on page 780) can calculate it
16 Virtual Address
17 Relative Virtual Address
18 Import Address Table
19 Matt Pietrek, An In-Depth Look into the Win32 Portable Executable File Format, (2002)]
20 Import Name Table
21 Matt Pietrek, An In-Depth Look into the Win32 Portable Executable File Format, (2002)]
22 The origin of this address choice is described here: MSDN
23 This can be changed by the /BASE linker option
748
There is also another reason to load modules at various base addresses, in this case random ones. It is
ASLR.
A shellcode trying to get executed on a compromised system must call system functions, hence, know
their addresses.
In older OS (in Windows NT line: before Windows Vista), system DLL (like kernel32.dll, user32.dll) were
always loaded at known addresses, and if we also recall that their versions rarely changed, the addresses
of functions were fixed and shellcode could call them directly.
In order to avoid this, the ASLR method loads your program and all modules it needs at random base
addresses, different every time.
ASLR support is denoted in a PE file by setting the flag
IMAGE_DLL_CHARACTERISTICS_DYNAMIC_BASE [see Mark Russinovich, Microsoft Windows Internals].
Subsystem
There is also a subsystem field, usually it is:

• native24 (.sys-driver),
• console (console application) or
• GUI (non-console).
OS version
A PE file also specifies the minimal Windows version it needs in order to be loadable.
The table of version numbers stored in the PE file and corresponding Windows codenames is here.
For example, MSVC 2005 compiles .exe files for running on Windows NT4 (version 4.00), but MSVC 2008
does not (the generated files have a version of 5.00, at least Windows 2000 is needed to run them).
MSVC 2012 generates .exe files of version 6.00 by default, targeting at least Windows Vista. However, by
changing the compiler’s options25 , it is possible to force it to compile for Windows XP.
Sections
Division in sections, as it seems, is present in all executable file formats.

It is devised in order to separate code from data, and data—from constant data.
• Either the IMAGE_SCN_CNT_CODE or IMAGE_SCN_MEM_EXECUTE flags will be set on the code section—
this is executable code.
• On data section—IMAGE_SCN_CNT_INITIALIZED_DATA,
IMAGE_SCN_MEM_READ and IMAGE_SCN_MEM_WRITE flags.
• On an empty section with uninitialized data—
IMAGE_SCN_CNT_UNINITIALIZED_DATA, IMAGE_SCN_MEM_READ
and IMAGE_SCN_MEM_WRITE.
• On a constant data section (one that’s protected from writing), the flags
IMAGE_SCN_CNT_INITIALIZED_DATA and IMAGE_SCN_MEM_READ can be set,
but not IMAGE_SCN_MEM_WRITE. A process going to crash if it tries to write to this section.
Each section in PE-file may have a name, however, it is not very important. Often (but not always) the
code section is named .text, the data section—.data, the constant data section — .rdata (readable
data) (perhaps, .rdata means read-only-data). Other popular section names are:
• .idata—imports section. IDA may create a pseudo-section named like this: 6.5.2 on the preceding
page.
• .edata—exports section (rare)
• .pdata—section holding all information about exceptions in Windows NT for MIPS, IA64 and x64: 6.5.3
on page 773
• .reloc—relocs section
24 Meaning, the module use Native API instead of Win32
25 MSDN
749
• .bss—uninitialized data (BSS)
• .tls—thread local storage (TLS)
• .rsrc—resources
• .CRT—may present in binary files compiled by ancient MSVC versions
PE file packers/encryptors often garble section names or replace the names with their own.
26
MSVC allows you to declare data in arbitrarily named section .
Some compilers and linkers can add a section with debugging symbols and other debugging information
(MinGW for instance). However it is not so in latest versions of MSVC (separate PDB files are used there
for this purpose).
That is how a PE section is described in the file:

typedef struct _IMAGE_SECTION_HEADER {
BYTE Name[IMAGE_SIZEOF_SHORT_NAME];
union {
DWORD PhysicalAddress;
DWORD VirtualSize;
} Misc;
DWORD VirtualAddress;
DWORD SizeOfRawData;
DWORD PointerToRawData;
DWORD PointerToRelocations;
DWORD PointerToLinenumbers;
WORD NumberOfRelocations;
WORD NumberOfLinenumbers;
DWORD Characteristics;
} IMAGE_SECTION_HEADER, *PIMAGE_SECTION_HEADER;
27
A word about terminology: PointerToRawData is called “Offset” in Hiew and VirtualAddress is called “RVA”
there.
Data section
Data section in file can be smaller than in memory. For example, some variables can be initialized, some
are not. Compiler and linker will collect them all into one section, but the first part of it is initialized and
allocated in file, while another is absent in file (of course, to make it smaller). VirtualSize will be equal to
the size of section in memory, and SizeOfRawData — to size of section in file.
IDA can show the border between initialized and not initialized parts like that:
...
.data:10017FFA db 0
.data:10017FFB db 0
.data:10017FFC db 0
.data:10017FFD db 0
.data:10017FFE db 0
.data:10017FFF db 0
.data:10018000 db ? ;
.data:10018001 db ? ;
.data:10018002 db ? ;
.data:10018003 db ? ;
.data:10018004 db ? ;
.data:10018005 db ? ;
...
26 MSDN
27 MSDN
750
.rdata — read-only data section
Strings are usually located here (because they have const char* type), other variables marked as const,
imported function names.
See also: 3.3 on page 468.
Relocations (relocs)
AKA FIXUP-s (at least in Hiew).

28
They are also present in almost all executable file formats . Exceptions are shared dynamic libraries
compiled with PIC, or any other PIC-code.
What are they for?
Obviously, modules can be loaded on various base addresses, but how to deal with global variables, for
example? They must be accessed by address. One solution is position-independent code ( 6.4.1 on
page 739). But it is not always convenient.
That is why a relocations table is present. There the addresses of points that must be corrected are
enumerated, in case of loading at a different base address.
For example, there is a global variable at address 0x410000 and this is how it is accessed:
A1 00 00 41 00 mov eax,[000410000]
The base address of the module is 0x400000, the RVA of the global variable is 0x10000.
If the module is loaded at base address 0x500000, the real address of the global variable must be 0x510000.
As we can see, the address of variable is encoded in the instruction MOV, after the byte 0xA1.
That is why the address of the 4 bytes after 0xA1, is written in the relocs table.
If the module is loaded at a different base address, the OS loader enumerates all addresses in the table,
finds each 32-bit word the address points to, subtracts the original base address from it (we get the RVA
here), and adds the new base address to it.
If a module is loaded at its original base address, nothing happens.
All global variables can be treated like that.
Relocs may have various types, however, in Windows for x86 processors, the type is usually
IMAGE_REL_BASED_HIGHLOW.
By the way, relocs are darkened in Hiew, for example: fig.1.22. (You have to circumvent these bytes
during patching.)
OllyDbg underlines the places in memory to which relocs are to be applied, for example: fig.1.53.
Exports and imports
As we all know, any executable program must use the OS’s services and other DLL-libraries somehow.
It can be said that functions from one module (usually DLL) must be connected somehow to the points of
their calls in other modules (.exe-file or another DLL).
For this, each DLL has an “exports” table, which consists of functions plus their addresses in a module.
And every .exe file or DLL has “imports”, a table of functions it needs for execution including list of DLL
filenames.
After loading the main .exe-file, the OS loader processes imports table: it loads the additional DLL-files,
finds function names among the DLL exports and writes their addresses down in the IAT of the main
.exe-module.
As we can see, during loading the loader must compare a lot of function names, but string comparison
is not a very fast procedure, so there is a support for “ordinals” or “hints”, which are function numbers
stored in the table, instead of their names.
That is how they can be located faster when loading a DLL. Ordinals are always present in the “export”
table.
28 Even in .exe files for MS-DOS
751
For example, a program using the MFC29 library usually loads mfc*.dll by ordinals, and in such programs
there are no MFC function names in INT.
When loading such programs in IDA, it will ask for a path to the mfc*.dll files in order to determine the
function names.
If you don’t tell IDA the path to these DLLs, there will be mfc80_123 instead of function names.
Imports section
Often a separate section is allocated for the imports table and everything related to it (with name like
.idata), however, this is not a strict rule.
Imports are also a confusing subject because of the terminological mess. Let’s try to collect all information
in one place.
Figure 6.1: A scheme that unites all PE-file structures related to imports
29 Microsoft Foundation Classes
752
The main structure is the array IMAGE_IMPORT_DESCRIPTOR. Each element for each DLL being imported.
Each element holds the RVA address of the text string (DLL name) (Name).
OriginalFirstThunk is the RVA address of the INT table. This is an array of RVA addresses, each of which
points to a text string with a function name. Each string is prefixed by a 16-bit integer (“hint”)—“ordinal”
of function.
While loading, if it is possible to find a function by ordinal, then the strings comparison will not occur. The
array is terminated by zero.
There is also a pointer to the IAT table named FirstThunk, it is just the RVA address of the place where the
loader writes the addresses of the resolved functions.
The points where the loader writes addresses are marked by IDA like this: __imp_CreateFileA, etc.
There are at least two ways to use the addresses written by the loader.
• The code will have instructions like call __imp_CreateFileA, and since the field with the address of
the imported function is a global variable in some sense, the address of the call instruction (plus 1
or 2) is to be added to the relocs table, for the case when the module is loaded at a different base
address.
But, obviously, this may enlarge relocs table significantly.
Because there are might be a lot of calls to imported functions in the module.
Furthermore, large relocs table slows down the process of loading modules.
• For each imported function, there is only one jump allocated, using the JMP instruction plus a reloc
to it. Such points are also called “thunks”.
All calls to the imported functions are just CALL instructions to the corresponding “thunk”. In this
case, additional relocs are not necessary because these CALL-s have relative addresses and do not
need to be corrected.
These two methods can be combined.
Possible, the linker creates individual “thunk”s if there are too many calls to the function, but not done by
default.
By the way, the array of function addresses to which FirstThunk is pointing is not necessary to be lo-
cated in the IAT section. For example, the author of these lines once wrote the PE_add_import30 utility for
adding imports to an existing .exe-file.
Some time earlier, in the previous versions of the utility, at the place of the function you want to substitute
with a call to another DLL, my utility wrote the following code:
MOV EAX, [yourdll.dll!function]
JMP EAX
FirstThunk points to the first instruction. In other words, when loading yourdll.dll, the loader writes the
address of the function function right in the code.
It also worth noting that a code section is usually write-protected, so my utility adds the
IMAGE_SCN_MEM_WRITE flag for code section. Otherwise, the program to crash while loading with error
code 5 (access denied).
One might ask: what if I supply a program with a set of DLL files which is not supposed to change (including
addresses of all DLL functions), is it possible to speed up the loading process?
Yes, it is possible to write the addresses of the functions to be imported into the FirstThunk arrays in
advance. The Timestamp field is present in the
IMAGE_IMPORT_DESCRIPTOR structure.
If a value is present there, then the loader compares this value with the date-time of the DLL file.
If the values are equal, then the loader does not do anything, and the loading of the process can be faster.
This is called “old-style binding” 31 .
The BIND.EXE utility in Windows SDK is for this. For speeding up the loading of your program, Matt Pietrek
in Matt Pietrek, An In-Depth Look into the Win32 Portable Executable File Format, (2002)]32 , suggests to
30 yurichev.com
31 MSDN. There is also the “new-style binding”.
753
do the binding shortly after your program installation on the computer of the end user.
PE-files packers/encryptors may also compress/encrypt imports table.

In this case, the Windows loader, of course, will not load all necessary DLLs.
Therefore, the packer/encryptor does this on its own, with the help of LoadLibrary() and the GetProcAd-
dress() functions.
That is why these two functions are often present in IAT in packed files.
In the standard DLLs from the Windows installation, IAT often is located right at the beginning of the
PE file. Supposedly, it is made so for optimization.
While loading, the .exe file is not loaded into memory as a whole (recall huge install programs which are
started suspiciously fast), it is “mapped”, and loaded into memory in parts as they are accessed.
Probably, Microsoft developers decided it will be faster.
Resources
Resources in a PE file are just a set of icons, pictures, text strings, dialog descriptions.
Perhaps they were separated from the main code, so all these things could be multilingual, and it would
be simpler to pick text or picture for the language that is currently set in the OS.
As a side effect, they can be edited easily and saved back to the executable file, even if one does not
have special knowledge, by using the ResHack editor, for example ( 6.5.2).
.NET
.NET programs are not compiled into machine code but into a special bytecode. Strictly speaking, there
is bytecode instead of the usual x86 code in the .exe file, however, the entry point (OEP) points to this
tiny fragment of x86 code:
jmp mscoree.dll!_CorExeMain
The .NET loader is located in mscoree.dll, which processes the PE file.

It was so in all pre-Windows XP OSes. Starting from XP, the OS loader is able to detect the .NET file and
run it without executing that JMP instruction 33 .
TLS
This section holds initialized data for the TLS( 6.2 on page 734) (if needed). When a new thread start, its
TLS data is initialized using the data from this section.
Aside from that, the PE file specification also provides initialization of the TLS section, the so-called TLS
callbacks.
If they are present, they are to be called before the control is passed to the main entry point (OEP).
This is used widely in the PE file packers/encryptors.
Tools
• objdump (present in cygwin) for dumping all PE-file structures.

• Hiew( 7.1 on page 780) as editor.
34
• pefile—Python-library for PE-file processing .
• ResHack AKA Resource Hacker—resources editor35 .
• PE_add_import36 — simple tool for adding symbol(s) to PE executable import table.
33 MSDN
754
• PE_patcher37 —simple tool for patching PE executables.
• PE_search_str_refs38 —simple tool for searching for a function in PE executables which use some text
string.
Further reading
39
• Daniel Pistelli—The .NET File Format
6.5.3 Windows SEH

Let’s forget about MSVC
In Windows, the SEH is intended for exceptions handling, nevertheless, it is language-agnostic, not related
to C++ or OOP in any way.
Here we are going to take a look at SEH in its isolated (from C++ and MSVC extensions) form.
Each running process has a chain of SEH handlers, each TIB has the address of the most recently defined
handler.
When an exception occurs (division by zero, incorrect address access, user exception triggered by calling
the RaiseException() function), the OS finds the last handler in the TIB and calls it, passing exception
kind and all information about the CPU state (register values, etc.) at the moment of the exception.
The exception handler considering the exception, does it see something familiar? If so, it handles the
exception.
If not, it signals to the OS that it cannot handle it and the OS calls the next handler in the chain, until a
handler which is able to handle the exception is be found.
At the very end of the chain there a standard handler that shows the well-known dialog box, informing the
user about a process crash, some technical information about the CPU state at the time of the crash, and
offering to collect all information and send it to developers in Microsoft.
Figure 6.2: Windows XP

37 yurichev.com
38 yurichev.com
755
Figure 6.3: Windows XP
Figure 6.4: Windows 7
756
Figure 6.5: Windows 8.1
Earlier, this handler was called Dr. Watson.

By the way, some developers make their own handler that sends information about the program crash
to themselves. It is registered with the help of SetUnhandledExceptionFilter() and to be called if the
OS does not have any other way to handle the exception. An example is Oracle RDBMS—it saves huge
dumps reporting all possible information about the CPU and memory state.
Let’s write our own primitive exception handler. This example is based on the example from [Matt Pietrek,
A Crash Course on the Depths of Win32™ Structured Exception Handling, (1997)]40 . It must be compiled
with the SAFESEH option: cl seh1.cpp /link /safeseh:no. More about SAFESEH here: MSDN.
#include <stdio.h>
DWORD new_value=1234;
EXCEPTION_DISPOSITION __cdecl except_handler(

struct _EXCEPTION_RECORD *ExceptionRecord,
void * EstablisherFrame,
struct _CONTEXT *ContextRecord,
void * DispatcherContext )
{
unsigned i;
printf ("%s\n", __FUNCTION__);

printf ("ExceptionRecord->ExceptionCode=0x%p\n", ExceptionRecord->ExceptionCode);
printf ("ExceptionRecord->ExceptionFlags=0x%p\n", ExceptionRecord->ExceptionFlags);
printf ("ExceptionRecord->ExceptionAddress=0x%p\n", ExceptionRecord->ExceptionAddress);
if (ExceptionRecord->ExceptionCode==0xE1223344)
{
printf ("That's for us\n");
// yes, we "handled" the exception
return ExceptionContinueExecution;
}
else if (ExceptionRecord->ExceptionCode==EXCEPTION_ACCESS_VIOLATION)
{
printf ("ContextRecord->Eax=0x%08X\n", ContextRecord->Eax);
// will it be possible to 'fix' it?
printf ("Trying to fix wrong pointer address\n");
ContextRecord->Eax=(DWORD)&new_value;
// yes, we "handled" the exception
return ExceptionContinueExecution;
}
else
{
printf ("We do not handle this\n");
// someone else's problem
return ExceptionContinueSearch;
};
757
}
int main()
{
DWORD handler = (DWORD)except_handler; // take a pointer to our handler
// install exception handler

__asm
{ // make EXCEPTION_REGISTRATION record:
push handler // address of handler function
push FS:[0] // address of previous handler
mov FS:[0],ESP // add new EXECEPTION_REGISTRATION
}
RaiseException (0xE1223344, 0, 0, NULL);
// now do something very bad

int* ptr=NULL;
int val=0;
val=*ptr;
printf ("val=%d\n", val);
// deinstall exception handler

__asm
{ // remove our EXECEPTION_REGISTRATION record
mov eax,[ESP] // get pointer to previous record
mov FS:[0], EAX // install previous record
add esp, 8 // clean our EXECEPTION_REGISTRATION off stack
}
return 0;
}
The FS: segment register is pointing to the TIB in win32.

The very first element in the TIB is a pointer to the last handler in the chain. We save it in the stack and
store the address of our handler there. The structure is named _EXCEPTION_REGISTRATION, it is a simple
singly-linked list and its elements are stored right in the stack.
Listing 6.22: MSVC/VC/crt/src/exsup.inc

_EXCEPTION_REGISTRATION struc
prev dd ?
handler dd ?
_EXCEPTION_REGISTRATION ends
So each “handler” field points to a handler and an each “prev” field points to the previous record in the
chain of exception handlers. The last record has 0xFFFFFFFF (-1) in the “prev” field.
758
TIB Stack
FS:0 +0: __except_list …
+4: … Prev=0xFFFFFFFF
+8: … Handle handler function
Prev
Handle handler function
Prev
Handle handler function
After our handler is installed, we call RaiseException() 41 . This is an user exception. The handler checks
the code. If the code is 0xE1223344, it returning ExceptionContinueExecution, which means that han-
dler corrected the CPU state (it is usually a correction of the EIP/ESP registers) and the OS can resume the
execution of the thread. If you alter slightly the code so the handler returns ExceptionContinueSearch,
then the OS will call the other handlers, and it’s unlikely that one who can handle it will be found, since no
one will have any information about it (rather about its code). You will see the standard Windows dialog
about a process crash.
What is the difference between a system exceptions and a user one? Here are the system ones:
as defined in WinBase.h as defined in ntstatus.h value
EXCEPTION_ACCESS_VIOLATION STATUS_ACCESS_VIOLATION 0xC0000005
EXCEPTION_DATATYPE_MISALIGNMENT STATUS_DATATYPE_MISALIGNMENT 0x80000002
EXCEPTION_BREAKPOINT STATUS_BREAKPOINT 0x80000003
EXCEPTION_SINGLE_STEP STATUS_SINGLE_STEP 0x80000004
EXCEPTION_ARRAY_BOUNDS_EXCEEDED STATUS_ARRAY_BOUNDS_EXCEEDED 0xC000008C
EXCEPTION_FLT_DENORMAL_OPERAND STATUS_FLOAT_DENORMAL_OPERAND 0xC000008D
EXCEPTION_FLT_DIVIDE_BY_ZERO STATUS_FLOAT_DIVIDE_BY_ZERO 0xC000008E
EXCEPTION_FLT_INEXACT_RESULT STATUS_FLOAT_INEXACT_RESULT 0xC000008F
EXCEPTION_FLT_INVALID_OPERATION STATUS_FLOAT_INVALID_OPERATION 0xC0000090
EXCEPTION_FLT_OVERFLOW STATUS_FLOAT_OVERFLOW 0xC0000091
EXCEPTION_FLT_STACK_CHECK STATUS_FLOAT_STACK_CHECK 0xC0000092
EXCEPTION_FLT_UNDERFLOW STATUS_FLOAT_UNDERFLOW 0xC0000093
EXCEPTION_INT_DIVIDE_BY_ZERO STATUS_INTEGER_DIVIDE_BY_ZERO 0xC0000094
EXCEPTION_INT_OVERFLOW STATUS_INTEGER_OVERFLOW 0xC0000095
EXCEPTION_PRIV_INSTRUCTION STATUS_PRIVILEGED_INSTRUCTION 0xC0000096
EXCEPTION_IN_PAGE_ERROR STATUS_IN_PAGE_ERROR 0xC0000006
EXCEPTION_ILLEGAL_INSTRUCTION STATUS_ILLEGAL_INSTRUCTION 0xC000001D
EXCEPTION_NONCONTINUABLE_EXCEPTION STATUS_NONCONTINUABLE_EXCEPTION 0xC0000025
EXCEPTION_STACK_OVERFLOW STATUS_STACK_OVERFLOW 0xC00000FD
EXCEPTION_INVALID_DISPOSITION STATUS_INVALID_DISPOSITION 0xC0000026
EXCEPTION_GUARD_PAGE STATUS_GUARD_PAGE_VIOLATION 0x80000001
EXCEPTION_INVALID_HANDLE STATUS_INVALID_HANDLE 0xC0000008
EXCEPTION_POSSIBLE_DEADLOCK STATUS_POSSIBLE_DEADLOCK 0xC0000194
CONTROL_C_EXIT STATUS_CONTROL_C_EXIT 0xC000013A
That is how the code is defined:

31 29 28 27 16 15 0
S U 0 Facility code Error code
S is a basic status code: 11—error; 10—warning; 01—informational; 00—success. U—whether the code
is user code.
41 MSDN
759
That is why we chose 0xE1223344—E16 (11102 ) 0xE (1110b) means that it is 1) user exception; 2) error.
But to be honest, this example works fine without these high bits.
Then we try to read a value from memory at address 0.
Of course, there is nothing at this address in win32, so an exception is raised.
The very first handler is to be called—yours, and it will know about it first, by checking the code if it’s
equal to the EXCEPTION_ACCESS_VIOLATION constant.
The code that’s reading from memory at address 0 is looks like this:

...
xor eax, eax
mov eax, DWORD PTR [eax] ; exception will occur here
push eax
push OFFSET msg
call _printf
add esp, 8
...
Will it be possible to fix this error “on the fly” and to continue with program execution?
Yes, our exception handler can fix the EAX value and let the OS execute this instruction once again. So
that is what we do. printf() prints 1234, because after the execution of our handler EAX is not 0, but
contains the address of the global variable new_value. The execution will resume.
That is what is going on: the memory manager in the CPU signals about an error, the CPU suspends the
thread, finds the exception handler in the Windows kernel, which, in turn, starts to call all handlers in the
SEH chain, one by one.
We use MSVC 2010 here, but of course, there is no any guarantee that EAX will be used for this pointer.
This address replacement trick is showy, and we considering it here as an illustration of SEH’s internals.
Nevertheless, it’s hard to recall any case where it is used for “on-the-fly” error fixing.
Why SEH-related records are stored right in the stack instead of some other place?
Supposedly because the OS is not needing to care about freeing this information, these records are simply
disposed when the function finishes its execution. This is somewhat like alloca(): ( 1.9.2 on page 34).
Now let’s get back to MSVC
Supposedly, Microsoft programmers needed exceptions in C, but not in C++ (for use in Windows NT kernel,
which is written in C), so they added a non-standard C extension to MSVC42 . It is not related to C++ PL
exceptions.
__try
{
...
}
__except(filter code)
{
handler code
}
“Finally” block may be instead of handler code:

__try
{
...
}
__finally
{
...
}
42 MSDN
760
The filter code is an expression, telling whether this handler code corresponds to the exception raised.
If your code is too big and cannot fit into one expression, a separate filter function can be defined.
There are a lot of such constructs in the Windows kernel. Here are a couple of examples from there
(WRK):
Listing 6.24: WRK-v1.2/base/ntos/ob/obwait.c

try {
KeReleaseMutant( (PKMUTANT)SignalObject,
MUTANT_INCREMENT,
FALSE,
TRUE );
} except((GetExceptionCode () == STATUS_ABANDONED ||
GetExceptionCode () == STATUS_MUTANT_NOT_OWNED)?
EXCEPTION_EXECUTE_HANDLER :
EXCEPTION_CONTINUE_SEARCH) {
Status = GetExceptionCode();
goto WaitExit;
}
Listing 6.25: WRK-v1.2/base/ntos/cache/cachesub.c

try {
RtlCopyBytes( (PVOID)((PCHAR)CacheBuffer + PageOffset),

UserBuffer,
MorePages ?
(PAGE_SIZE - PageOffset) :
(ReceivedLength - PageOffset) );
} except( CcCopyReadExceptionFilter( GetExceptionInformation(),

&Status ) ) {
Here is also a filter code example:
Listing 6.26: WRK-v1.2/base/ntos/cache/copysup.c

LONG
CcCopyReadExceptionFilter(
IN PEXCEPTION_POINTERS ExceptionPointer,
IN PNTSTATUS ExceptionCode
)
/*++
Routine Description:
This routine serves as an exception filter and has the special job of
extracting the "real" I/O error when Mm raises STATUS_IN_PAGE_ERROR
beneath us.
Arguments:
ExceptionPointer - A pointer to the exception record that contains

the real Io Status.
ExceptionCode - A pointer to an NTSTATUS that is to receive the real

status.
Return Value:
EXCEPTION_EXECUTE_HANDLER
--*/
761
*ExceptionCode = ExceptionPointer->ExceptionRecord->ExceptionCode;
if ( (*ExceptionCode == STATUS_IN_PAGE_ERROR) &&

(ExceptionPointer->ExceptionRecord->NumberParameters >= 3) ) {
*ExceptionCode = (NTSTATUS) ExceptionPointer->ExceptionRecord->ExceptionInformation[2];

}
ASSERT( !NT_SUCCESS(*ExceptionCode) );
return EXCEPTION_EXECUTE_HANDLER;
}
Internally, SEH is an extension of the OS-supported exceptions. But the handler function is _except_handler3
(for SEH3) or _except_handler4 (for SEH4).
The code of this handler is MSVC-related, it is located in its libraries, or in msvcr*.dll. It is very important
to know that SEH is a MSVC thing.
Other win32-compilers may offer something completely different.
SEH3
SEH3 has _except_handler3 as a handler function, and extends the _EXCEPTION_REGISTRATION table,
adding a pointer to the scope table and previous try level variable. SEH4 extends the scope table by 4
values for buffer overflow protection.
The scope table is a table that consists of pointers to the filter and handler code blocks, for each nested
level of try/except.
TIB Stack
scope table Prev
0xFFFFFFFF (-1) Handle handler function

information about
first try/except filter function …
block
handler/finally function Prev
Handle _except_handler3
0
information about scope table
second try/except filter function
block
previous try level
handler/finally function
EBP
1
…
information about
third try/except filter function
block
… more entries …
762
Again, it is very important to understand that the OS takes care only of the prev/handle fields, and nothing
more.
It is the job of the _except_handler3 function to read the other fields and scope table, and decide which
handler to execute and when.
The source code of the _except_handler3 function is closed.

However, Sanos OS, which has a win32 compatibility layer, has the same functions reimplemented, which
are somewhat equivalent to those in Windows 43 . Another reimplementation is present in Wine44 and
ReactOS45 .
If the filter pointer is NULL, the handler pointer is the pointer to the finally code block.
During execution, the previous try level value in the stack changes, so _except_handler3 can get in-
formation about the current level of nestedness, in order to know which scope table entry to use.
SEH3: one try/except block example
#include <stdio.h>
#include <excpt.h>
int main()
{
int* p = NULL;
__try
{
printf("hello #1!\n");
*p = 13; // causes an access violation exception;
printf("hello #2!\n");
}
__except(GetExceptionCode()==EXCEPTION_ACCESS_VIOLATION ?
EXCEPTION_EXECUTE_HANDLER : EXCEPTION_CONTINUE_SEARCH)
{
printf("access violation, can't recover\n");
}
}

$SG74605 DB 'hello #1!', 0aH, 00H
$SG74606 DB 'hello #2!', 0aH, 00H
$SG74608 DB 'access violation, can''t recover', 0aH, 00H
_DATA ENDS
; scope table:
CONST SEGMENT
$T74622 DD 0ffffffffH ; previous try level
DD FLAT:$L74617 ; filter
DD FLAT:$L74618 ; handler
CONST ENDS
_TEXT SEGMENT
$T74621 = -32 ; size = 4
_p$ = -28 ; size = 4
__$SEHRec$ = -24 ; size = 24
_main PROC NEAR
push ebp
mov ebp, esp
push -1 ; previous try level
push OFFSET FLAT:$T74622 ; scope table
push OFFSET FLAT:__except_handler3 ; handler
mov eax, DWORD PTR fs:__except_list
push eax ; prev
44 GitHub
763
mov DWORD PTR fs:__except_list, esp
add esp, -16
; 3 registers to be saved:
push ebx
push esi
push edi
mov DWORD PTR __$SEHRec$[ebp], esp
mov DWORD PTR _p$[ebp], 0
mov DWORD PTR __$SEHRec$[ebp+20], 0 ; previous try level
push OFFSET FLAT:$SG74605 ; 'hello #1!'
call _printf
add esp, 4
mov eax, DWORD PTR _p$[ebp]
push OFFSET FLAT:$SG74606 ; 'hello #2!'
call _printf
add esp, 4
mov DWORD PTR __$SEHRec$[ebp+20], -1 ; previous try level
jmp SHORT $L74616
; filter code:
$L74617:
$L74627:
mov ecx, DWORD PTR __$SEHRec$[ebp+4]
mov DWORD PTR $T74621[ebp], eax
mov eax, DWORD PTR $T74621[ebp]
sub eax, -1073741819; c0000005H
neg eax
sbb eax, eax
inc eax
$L74619:
$L74626:
ret 0
; handler code:
$L74618:
mov esp, DWORD PTR __$SEHRec$[ebp]
push OFFSET FLAT:$SG74608 ; 'access violation, can''t recover'
call _printf
add esp, 4
mov DWORD PTR __$SEHRec$[ebp+20], -1 ; setting previous try level back to -1
$L74616:
xor eax, eax
mov DWORD PTR fs:__except_list, ecx
pop edi
pop esi
pop ebx
mov esp, ebp
pop ebp
ret 0
_main ENDP
_TEXT ENDS
END
Here we see how the SEH frame is constructed in the stack. The scope table is located in the CONST
segment—indeed, these fields are not to be changed. An interesting thing is how the previous try level
variable has changed. The initial value is 0xFFFFFFFF (−1). The moment when the body of the try state-
ment is opened is marked with an instruction that writes 0 to the variable. The moment when the body of
the try statement is closed, −1 is written back to it. We also see the addresses of filter and handler code.
Thus we can easily see the structure of the try/except constructs in the function.
Since the SEH setup code in the function prologue may be shared between many functions, sometimes
the compiler inserts a call to the SEH_prolog() function in the prologue, which does just that.
The SEH cleanup code is in the SEH_epilog() function.
764
Let’s try to run this example in tracer:
tracer.exe -l:2.exe --dump-seh
Listing 6.28: tracer.exe output

EXCEPTION_ACCESS_VIOLATION at 2.exe!main+0x44 (0x401054) ExceptionInformation[0]=1
EAX=0x00000000 EBX=0x7efde000 ECX=0x0040cbc8 EDX=0x0008e3c8
ESI=0x00001db1 EDI=0x00000000 EBP=0x0018feac ESP=0x0018fe80
EIP=0x00401054
FLAGS=AF IF RF
* SEH frame at 0x18fe9c prev=0x18ff78 handler=0x401204 (2.exe!_except_handler3)
SEH3 frame. previous trylevel=0
scopetable entry[0]. previous try level=-1, filter=0x401070 (2.exe!main+0x60) handler=0x401088 ⤦
Ç (2.exe!main+0x78)
* SEH frame at 0x18ff78 prev=0x18ffc4 handler=0x401204 (2.exe!_except_handler3)
scopetable entry[0]. previous try level=-1, filter=0x401531 (2.exe!mainCRTStartup+0x18d) ⤦
Ç handler=0x401545 (2.exe!mainCRTStartup+0x1a1)
* SEH frame at 0x18ffc4 prev=0x18ffe4 handler=0x771f71f5 (ntdll.dll!__except_handler4)
SEH4 header: GSCookieOffset=0xfffffffe GSCookieXOROffset=0x0
EHCookieOffset=0xffffffcc EHCookieXOROffset=0x0
scopetable entry[0]. previous try level=-2, filter=0x771f74d0 (ntdll.dll!⤦
Ç ___safe_se_handler_table+0x20) handler=0x771f90eb (ntdll.dll!_TppTerminateProcess@4+0x43)
* SEH frame at 0x18ffe4 prev=0xffffffff handler=0x77247428 (ntdll.dll!_FinalExceptionHandler@16⤦
Ç )
We see that the SEH chain consists of 4 handlers.
The first two are located in our example. Two? But we made only one? Yes, another one has been
set up in the CRT function _mainCRTStartup(), and as it seems that it handles at least FPU exceptions.
Its source code can be found in the MSVC installation: crt/src/winxfltr.c.
The third is the SEH4 one in ntdll.dll, and the fourth handler is not MSVC-related and is located in nt-
dll.dll, and has a self-describing function name.
As you can see, there are 3 types of handlers in one chain:

one is not related to MSVC at all (the last one) and two MSVC-related: SEH3 and SEH4.
SEH3: two try/except blocks example
#include <stdio.h>
#include <excpt.h>
int filter_user_exceptions (unsigned int code, struct _EXCEPTION_POINTERS *ep)

{
printf("in filter. code=0x%08X\n", code);
if (code == 0x112233)
{
printf("yes, that is our exception\n");
return EXCEPTION_EXECUTE_HANDLER;
}
else
{
printf("not our exception\n");
return EXCEPTION_CONTINUE_SEARCH;
};
}
int main()
{
int* p = NULL;
__try
{
__try
{
765
printf ("hello!\n");
RaiseException (0x112233, 0, 0, NULL);
printf ("0x112233 raised. now let's crash\n");
*p = 13; // causes an access violation exception;
}
__except(GetExceptionCode()==EXCEPTION_ACCESS_VIOLATION ?
EXCEPTION_EXECUTE_HANDLER : EXCEPTION_CONTINUE_SEARCH)
{
printf("access violation, can't recover\n");
}
}
__except(filter_user_exceptions(GetExceptionCode(), GetExceptionInformation()))
{
// the filter_user_exceptions() function answering to the question
// "is this exception belongs to this block?"
// if yes, do the follow:
printf("user exception caught\n");
}
}
Now there are two try blocks. So the scope table now has two entries, one for each block. Previous try
level changes as execution flow enters or exits the try block.

$SG74606 DB 'in filter. code=0x%08X', 0aH, 00H
$SG74608 DB 'yes, that is our exception', 0aH, 00H
$SG74610 DB 'not our exception', 0aH, 00H
$SG74617 DB 'hello!', 0aH, 00H
$SG74619 DB '0x112233 raised. now let''s crash', 0aH, 00H
$SG74623 DB 'user exception caught', 0aH, 00H
_code$ = 8 ; size = 4
_ep$ = 12 ; size = 4
_filter_user_exceptions PROC NEAR
push ebp
mov ebp, esp
mov eax, DWORD PTR _code$[ebp]
push eax
push OFFSET FLAT:$SG74606 ; 'in filter. code=0x%08X'
call _printf
add esp, 8
cmp DWORD PTR _code$[ebp], 1122867; 00112233H
jne SHORT $L74607
push OFFSET FLAT:$SG74608 ; 'yes, that is our exception'
call _printf
add esp, 4
mov eax, 1
jmp SHORT $L74605
$L74607:
push OFFSET FLAT:$SG74610 ; 'not our exception'
call _printf
add esp, 4
xor eax, eax
$L74605:
pop ebp
ret 0
_filter_user_exceptions ENDP
; scope table:
CONST SEGMENT
$T74644 DD 0ffffffffH ; previous try level for outer block
DD FLAT:$L74634 ; outer block filter
DD FLAT:$L74635 ; outer block handler
DD 00H ; previous try level for inner block
DD FLAT:$L74638 ; inner block filter
DD FLAT:$L74639 ; inner block handler
CONST ENDS
766
$T74643 = -36 ; size = 4
$T74642 = -32 ; size = 4
_p$ = -28 ; size = 4
__$SEHRec$ = -24 ; size = 24
_main PROC NEAR
push ebp
mov ebp, esp
push -1 ; previous try level
push OFFSET FLAT:$T74644
push OFFSET FLAT:__except_handler3
mov eax, DWORD PTR fs:__except_list
push eax
mov DWORD PTR fs:__except_list, esp
add esp, -20
push ebx
push esi
push edi
mov DWORD PTR __$SEHRec$[ebp+20], 0 ; outer try block entered. set previous try level to
0
mov DWORD PTR __$SEHRec$[ebp+20], 1 ; inner try block entered. set previous try level to
1
push OFFSET FLAT:$SG74617 ; 'hello!'
call _printf
add esp, 4
push 0
push 0
push 0
push 1122867 ; 00112233H
call DWORD PTR __imp__RaiseException@16
push OFFSET FLAT:$SG74619 ; '0x112233 raised. now let''s crash'
call _printf
add esp, 4
mov DWORD PTR __$SEHRec$[ebp+20], 0 ; inner try block exited. set previous try level back
to 0
jmp SHORT $L74615
; inner block filter:

$L74638:
$L74650:
mov eax, DWORD PTR $T74643[ebp]
sub eax, -1073741819; c0000005H
neg eax
sbb eax, eax
inc eax
$L74640:
$L74648:
ret 0
; inner block handler:

$L74639:
push OFFSET FLAT:$SG74621 ; 'access violation, can''t recover'
call _printf
add esp, 4
mov DWORD PTR __$SEHRec$[ebp+20], 0 ; inner try block exited. set previous try level back
to 0
$L74615:
mov DWORD PTR __$SEHRec$[ebp+20], -1 ; outer try block exited, set previous try level
back to -1
jmp SHORT $L74633
; outer block filter:

$L74634:
767
$L74651:
push ecx
mov edx, DWORD PTR $T74642[ebp]
push edx
call _filter_user_exceptions
add esp, 8
$L74636:
$L74649:
ret 0
; outer block handler:

$L74635:
push OFFSET FLAT:$SG74623 ; 'user exception caught'
call _printf
add esp, 4
mov DWORD PTR __$SEHRec$[ebp+20], -1 ; both try blocks exited. set previous try level
back to -1
$L74633:
xor eax, eax
mov DWORD PTR fs:__except_list, ecx
pop edi
pop esi
pop ebx
mov esp, ebp
pop ebp
ret 0
_main ENDP
If we set a breakpoint on the printf() function, which is called from the handler, we can also see how
yet another SEH handler is added.
Perhaps it’s another machinery inside the SEH handling process. Here we also see our scope table con-
sisting of 2 entries.
tracer.exe -l:3.exe bpx=3.exe!printf --dump-seh
Listing 6.30: tracer.exe output

(0) 3.exe!printf
EAX=0x0000001b EBX=0x00000000 ECX=0x0040cc58 EDX=0x0008e3c8
ESI=0x00000000 EDI=0x00000000 EBP=0x0018f840 ESP=0x0018f838
EIP=0x004011b6
FLAGS=PF ZF IF
* SEH frame at 0x18f88c prev=0x18fe9c handler=0x771db4ad (ntdll.dll!ExecuteHandler2@20+0x3a)
* SEH frame at 0x18fe9c prev=0x18ff78 handler=0x4012e0 (3.exe!_except_handler3)
scopetable entry[0]. previous try level=-1, filter=0x401120 (3.exe!main+0xb0) handler=0x40113b ⤦
Ç (3.exe!main+0xcb)
scopetable entry[1]. previous try level=0, filter=0x4010e8 (3.exe!main+0x78) handler=0x401100 ⤦
Ç (3.exe!main+0x90)
* SEH frame at 0x18ff78 prev=0x18ffc4 handler=0x4012e0 (3.exe!_except_handler3)
scopetable entry[0]. previous try level=-1, filter=0x40160d (3.exe!mainCRTStartup+0x18d) ⤦
Ç handler=0x401621 (3.exe!mainCRTStartup+0x1a1)
* SEH frame at 0x18ffc4 prev=0x18ffe4 handler=0x771f71f5 (ntdll.dll!__except_handler4)
SEH4 header: GSCookieOffset=0xfffffffe GSCookieXOROffset=0x0
EHCookieOffset=0xffffffcc EHCookieXOROffset=0x0
scopetable entry[0]. previous try level=-2, filter=0x771f74d0 (ntdll.dll!⤦
Ç ___safe_se_handler_table+0x20) handler=0x771f90eb (ntdll.dll!_TppTerminateProcess@4+0x43)
* SEH frame at 0x18ffe4 prev=0xffffffff handler=0x77247428 (ntdll.dll!_FinalExceptionHandler@16⤦
Ç )
768
SEH4
During a buffer overflow ( 1.26.2 on page 271) attack, the address of the scope table can be rewritten, so
starting from MSVC 2005, SEH3 was upgraded to SEH4 in order to have buffer overflow protection. The
pointer to the scope table is now xored with a security cookie. The scope table was extended to have a
header consisting of two pointers to security cookies.
Each element has an offset inside the stack of another value: the address of the stack frame (EBP) xored
with the security_cookie , placed in the stack.
This value will be read during exception handling and checked for correctness. The security cookie in the
stack is random each time, so hopefully a remote attacker can’t predict it.
The initial previous try level is −2 in SEH4 instead of −1.
TIB Stack
scope table Prev
GS Cookie Offset Handle handler function
GS Cookie XOR Offset …
EH Cookie Offset Prev
EH Cookie XOR Offset Handle _except_handler4
scope table
0xFFFFFFFF (-1) ⊕security_cookie
information about previous try level

first try/except filter function
block
EBP
…
0
EBP⊕security_cookie
information about
second try/except filter function
block …
1
information about
third try/except filter function
block
… more entries …
Here are both examples compiled in MSVC 2012 with SEH4:
Listing 6.31: MSVC 2012: one try block example

$SG85485 DB 'hello #1!', 0aH, 00H
$SG85486 DB 'hello #2!', 0aH, 00H
769
; scope table:
xdata$x SEGMENT
__sehtable$_main DD 0fffffffeH ; GS Cookie Offset
DD 00H ; GS Cookie XOR Offset
DD 0ffffffccH ; EH Cookie Offset
DD 00H ; EH Cookie XOR Offset
DD 0fffffffeH ; previous try level
DD FLAT:$LN12@main ; filter
DD FLAT:$LN8@main ; handler
xdata$x ENDS
$T2 = -36 ; size = 4

_p$ = -32 ; size = 4
tv68 = -28 ; size = 4
__$SEHRec$ = -24 ; size = 24
_main PROC
push ebp
mov ebp, esp
push -2
push OFFSET __sehtable$_main
push OFFSET __except_handler4
mov eax, DWORD PTR fs:0
push eax
add esp, -20
push ebx
push esi
push edi
mov eax, DWORD PTR ___security_cookie
xor DWORD PTR __$SEHRec$[ebp+16], eax ; xored pointer to scope table
xor eax, ebp
push eax ; ebp ^ security_cookie
lea eax, DWORD PTR __$SEHRec$[ebp+8] ; pointer to VC_EXCEPTION_REGISTRATION_RECORD
mov DWORD PTR fs:0, eax
mov DWORD PTR __$SEHRec$[ebp+20], 0 ; previous try level
push OFFSET $SG85485 ; 'hello #1!'
call _printf
add esp, 4
push OFFSET $SG85486 ; 'hello #2!'
call _printf
add esp, 4
jmp SHORT $LN6@main
; filter:
$LN7@main:
$LN12@main:
cmp DWORD PTR $T2[ebp], -1073741819 ; c0000005H
jne SHORT $LN4@main
mov DWORD PTR tv68[ebp], 1
jmp SHORT $LN5@main
$LN4@main:
$LN5@main:
$LN9@main:
$LN11@main:
ret 0
; handler:
$LN8@main:
770
push OFFSET $SG85488 ; 'access violation, can''t recover'
call _printf
add esp, 4
$LN6@main:
xor eax, eax
mov DWORD PTR fs:0, ecx
pop ecx
pop edi
pop esi
pop ebx
mov esp, ebp
pop ebp
ret 0
_main ENDP
Listing 6.32: MSVC 2012: two try blocks example

xdata$x SEGMENT
__sehtable$_main DD 0fffffffeH ; GS Cookie Offset
DD 00H ; GS Cookie XOR Offset
DD 0ffffffc8H ; EH Cookie Offset
DD 00H ; EH Cookie Offset
DD 0fffffffeH ; previous try level for outer block
DD FLAT:$LN19@main ; outer block filter
DD FLAT:$LN9@main ; outer block handler
DD 00H ; previous try level for inner block
DD FLAT:$LN18@main ; inner block filter
DD FLAT:$LN13@main ; inner block handler
xdata$x ENDS
$T2 = -40 ; size = 4

$T3 = -36 ; size = 4
_p$ = -32 ; size = 4
tv72 = -28 ; size = 4
__$SEHRec$ = -24 ; size = 24
_main PROC
push ebp
mov ebp, esp
push -2 ; initial previous try level
push OFFSET __sehtable$_main
push OFFSET __except_handler4
mov eax, DWORD PTR fs:0
push eax ; prev
add esp, -24
push ebx
push esi
push edi
mov eax, DWORD PTR ___security_cookie
xor DWORD PTR __$SEHRec$[ebp+16], eax ; xored pointer to scope table
xor eax, ebp ; ebp ^ security_cookie
push eax
lea eax, DWORD PTR __$SEHRec$[ebp+8] ; pointer to VC_EXCEPTION_REGISTRATION_RECORD
mov DWORD PTR fs:0, eax
mov DWORD PTR __$SEHRec$[ebp+20], 0 ; entering outer try block, setting previous try
level=0
mov DWORD PTR __$SEHRec$[ebp+20], 1 ; entering inner try block, setting previous try
level=1
push OFFSET $SG85497 ; 'hello!'
771
call _printf
add esp, 4
push 0
push 0
push 0
push 1122867 ; 00112233H
call DWORD PTR __imp__RaiseException@16
push OFFSET $SG85499 ; '0x112233 raised. now let''s crash'
call _printf
add esp, 4
mov DWORD PTR __$SEHRec$[ebp+20], 0 ; exiting inner try block, set previous try level
back to 0
jmp SHORT $LN2@main
; inner block filter:

$LN12@main:
$LN18@main:
cmp DWORD PTR $T3[ebp], -1073741819 ; c0000005H
jne SHORT $LN5@main
jmp SHORT $LN6@main
$LN5@main:
$LN6@main:
$LN14@main:
$LN16@main:
ret 0
; inner block handler:

$LN13@main:
push OFFSET $SG85501 ; 'access violation, can''t recover'
call _printf
add esp, 4
mov DWORD PTR __$SEHRec$[ebp+20], 0 ; exiting inner try block, setting previous try level
back to 0
$LN2@main:
mov DWORD PTR __$SEHRec$[ebp+20], -2 ; exiting both blocks, setting previous try level
back to -2
jmp SHORT $LN7@main
; outer block filter:

$LN8@main:
$LN19@main:
push ecx
mov edx, DWORD PTR $T2[ebp]
push edx
call _filter_user_exceptions
add esp, 8
$LN10@main:
$LN17@main:
ret 0
; outer block handler:

$LN9@main:
push OFFSET $SG85503 ; 'user exception caught'
call _printf
add esp, 4
772
mov DWORD PTR __$SEHRec$[ebp+20], -2 ; exiting both blocks, setting previous try level
back to -2
$LN7@main:
xor eax, eax
mov DWORD PTR fs:0, ecx
pop ecx
pop edi
pop esi
pop ebx
mov esp, ebp
pop ebp
ret 0
_main ENDP
_code$ = 8 ; size = 4
_ep$ = 12 ; size = 4
_filter_user_exceptions PROC
push ebp
mov ebp, esp
mov eax, DWORD PTR _code$[ebp]
push eax
push OFFSET $SG85486 ; 'in filter. code=0x%08X'
call _printf
add esp, 8
cmp DWORD PTR _code$[ebp], 1122867 ; 00112233H
jne SHORT $LN2@filter_use
push OFFSET $SG85488 ; 'yes, that is our exception'
call _printf
add esp, 4
mov eax, 1
jmp SHORT $LN3@filter_use
jmp SHORT $LN3@filter_use
$LN2@filter_use:
push OFFSET $SG85490 ; 'not our exception'
call _printf
add esp, 4
xor eax, eax
$LN3@filter_use:
pop ebp
ret 0
_filter_user_exceptions ENDP
Here is the meaning of the cookies: Cookie Offset is the difference between the address of the saved
EBP value in the stack and the EBP ⊕security_cookie value in the stack. Cookie XOR Offset is an additional
difference between the EBP ⊕ security_cookie value and what is stored in the stack.
If this equation is not true, the process is to halt due to stack corruption:
security_cookie⊕(CookieXOROf f set+address_of _saved_EBP ) == stack[address_of _saved_EBP +CookieOf f set]
If Cookie Offset is −2, this implies that it is not present.
Cookies checking is also implemented in my tracer, see GitHub for details.
It is still possible to fall back to SEH3 in the compilers after (and including) MSVC 2005 by setting the
/GS- option, however, the CRT code use SEH4 anyway.
Windows x64
As you might think, it is not very fast to set up the SEH frame at each function prologue. Another perfor-
mance problem is changing the previous try level value many times during the function’s execution.
So things are changed completely in x64: now all pointers to try blocks, filter and handler functions are
stored in another PE segment .pdata, and from there the OS’s exception handler takes all the information.
Here are the two examples from the previous section compiled for x64:

$SG86276 DB 'hello #1!', 0aH, 00H
$SG86277 DB 'hello #2!', 0aH, 00H
773
pdata SEGMENT
$pdata$main DD imagerel $LN9
DD imagerel $LN9+61
DD imagerel $unwind$main
pdata ENDS
pdata SEGMENT
$pdata$main$filt$0 DD imagerel main$filt$0
DD imagerel main$filt$0+32
DD imagerel $unwind$main$filt$0
pdata ENDS
xdata SEGMENT
$unwind$main DD 020609H
DD 030023206H
DD imagerel __C_specific_handler
DD 01H
DD imagerel $LN9+8
DD imagerel $LN9+40
DD imagerel main$filt$0
DD imagerel $LN9+40
$unwind$main$filt$0 DD 020601H
DD 050023206H
xdata ENDS
_TEXT SEGMENT
main PROC
$LN9:
push rbx
sub rsp, 32
xor ebx, ebx
lea rcx, OFFSET FLAT:$SG86276 ; 'hello #1!'
call printf
mov DWORD PTR [rbx], 13
lea rcx, OFFSET FLAT:$SG86277 ; 'hello #2!'
call printf
jmp SHORT $LN8@main
$LN6@main:
lea rcx, OFFSET FLAT:$SG86279 ; 'access violation, can''t recover'
call printf
$LN8@main:
xor eax, eax
add rsp, 32
pop rbx
ret 0
main ENDP
_TEXT ENDS
text$x SEGMENT
main$filt$0 PROC
push rbp
sub rsp, 32
mov rbp, rdx
$LN5@main$filt$:
mov rax, QWORD PTR [rcx]
xor ecx, ecx
cmp DWORD PTR [rax], -1073741819; c0000005H
sete cl
mov eax, ecx
$LN7@main$filt$:
add rsp, 32
pop rbp
ret 0
int 3
main$filt$0 ENDP
text$x ENDS
774
pdata SEGMENT
$pdata$filter_user_exceptions DD imagerel $LN6
DD imagerel $LN6+73
DD imagerel $unwind$filter_user_exceptions
$pdata$main DD imagerel $LN14
DD imagerel $LN14+95
DD imagerel $unwind$main
pdata ENDS
pdata SEGMENT
pdata ENDS
xdata SEGMENT
$unwind$filter_user_exceptions DD 020601H
DD 030023206H
$unwind$main DD 020609H
DD 030023206H
DD imagerel __C_specific_handler
DD 02H
DD imagerel $LN14+8
DD imagerel $LN14+8
DD 050023206H
DD 050023206H
xdata ENDS
_TEXT SEGMENT
main PROC
$LN14:
push rbx
sub rsp, 32
xor ebx, ebx
lea rcx, OFFSET FLAT:$SG86288 ; 'hello!'
call printf
xor r9d, r9d
xor r8d, r8d
xor edx, edx
mov ecx, 1122867 ; 00112233H
call QWORD PTR __imp_RaiseException
lea rcx, OFFSET FLAT:$SG86290 ; '0x112233 raised. now let''s crash'
call printf
mov DWORD PTR [rbx], 13
jmp SHORT $LN13@main
$LN11@main:
lea rcx, OFFSET FLAT:$SG86292 ; 'access violation, can''t recover'
call printf
$LN13@main:
jmp SHORT $LN9@main
$LN7@main:
775
lea rcx, OFFSET FLAT:$SG86294 ; 'user exception caught'
call printf
$LN9@main:
xor eax, eax
add rsp, 32
pop rbx
ret 0
main ENDP
text$x SEGMENT
main$filt$0 PROC
push rbp
sub rsp, 32
mov rbp, rdx
$LN10@main$filt$:
xor ecx, ecx
cmp DWORD PTR [rax], -1073741819; c0000005H
sete cl
mov eax, ecx
$LN12@main$filt$:
add rsp, 32
pop rbp
ret 0
int 3
main$filt$0 ENDP
main$filt$1 PROC
push rbp
sub rsp, 32
mov rbp, rdx
$LN6@main$filt$:
mov rdx, rcx
mov ecx, DWORD PTR [rax]
call filter_user_exceptions
$LN8@main$filt$:
add rsp, 32
pop rbp
ret 0
int 3
main$filt$1 ENDP
text$x ENDS
_TEXT SEGMENT
code$ = 48
ep$ = 56
filter_user_exceptions PROC
$LN6:
push rbx
sub rsp, 32
mov ebx, ecx
mov edx, ecx
lea rcx, OFFSET FLAT:$SG86277 ; 'in filter. code=0x%08X'
call printf
cmp ebx, 1122867; 00112233H
jne SHORT $LN2@filter_use
lea rcx, OFFSET FLAT:$SG86279 ; 'yes, that is our exception'
call printf
mov eax, 1
add rsp, 32
pop rbx
ret 0
$LN2@filter_use:
lea rcx, OFFSET FLAT:$SG86281 ; 'not our exception'
call printf
xor eax, eax
add rsp, 32
776
pop rbx
ret 0
filter_user_exceptions ENDP
_TEXT ENDS
46
Read [Igor Skochinsky, Compiler Internals: Exceptions and RTTI, (2012)] for more detailed information
about this.
Aside from exception information, .pdata is a section that contains the addresses of almost all function
starts and ends, hence it may be useful for a tools targeted at automated analysis.
Read more about SEH
[Matt Pietrek, A Crash Course on the Depths of Win32™ Structured Exception Handling, (1997)]47 , [Igor
Skochinsky, Compiler Internals: Exceptions and RTTI, (2012)] 48 .
6.5.4 Windows NT: Critical section

Critical sections in any OS are very important in multithreaded environment, mostly for giving a guarantee
that only one thread can access some data in a single moment of time, while blocking other threads and
interrupts.
That is how a CRITICAL_SECTION structure is declared in Windows NT line OS:
Listing 6.35: (Windows Research Kernel v1.2) public/sdk/inc/nturtl.h

typedef struct _RTL_CRITICAL_SECTION {
PRTL_CRITICAL_SECTION_DEBUG DebugInfo;
//
// The following three fields control entering and exiting the critical
// section for the resource
//
LONG LockCount;
LONG RecursionCount;
HANDLE OwningThread; // from the thread's ClientId->UniqueThread
HANDLE LockSemaphore;
ULONG_PTR SpinCount; // force size on 64-bit systems when packed
} RTL_CRITICAL_SECTION, *PRTL_CRITICAL_SECTION;
That’s is how EnterCriticalSection() function works:
Listing 6.36: Windows 2008/ntdll.dll/x86 (begin)

_RtlEnterCriticalSection@4

arg_0 = dword ptr 8
mov edi, edi

push ebp
mov ebp, esp
sub esp, 0Ch
push esi
push edi
mov edi, [ebp+arg_0]
lea esi, [edi+4] ; LockCount
mov eax, esi
lock btr dword ptr [eax], 0
jnb wait ; jump if CF=0
loc_7DE922DD:
mov eax, large fs:18h
777
mov ecx, [eax+24h]
mov [edi+0Ch], ecx
mov dword ptr [edi+8], 1
pop edi
xor eax, eax
pop esi
mov esp, ebp
pop ebp
retn 4
... skipped
The most important instruction in this code fragment is BTR (prefixed with LOCK):
the zeroth bit is stored in the CF flag and cleared in memory. This is an atomic operation,
blocking all other CPUs’ access to this piece of memory (see the LOCK prefix before the BTR instruction).
If the bit at LockCount is 1,
fine, reset it and return from the function: we are in a critical section.
If not—the critical section is already occupied by other thread, so wait.
The wait is performed there using WaitForSingleObject().
And here is how the LeaveCriticalSection() function works:
Listing 6.37: Windows 2008/ntdll.dll/x86 (begin)

_RtlLeaveCriticalSection@4 proc near
arg_0 = dword ptr 8
mov edi, edi

push ebp
mov ebp, esp
push esi
mov esi, [ebp+arg_0]
add dword ptr [esi+8], 0FFFFFFFFh ; RecursionCount
jnz short loc_7DE922B2
push ebx
push edi
lea edi, [esi+4] ; LockCount
mov dword ptr [esi+0Ch], 0
mov ebx, 1
mov eax, edi
lock xadd [eax], ebx
inc ebx
cmp ebx, 0FFFFFFFFh
jnz loc_7DEA8EB7
loc_7DE922B0:
pop edi
pop ebx
loc_7DE922B2:
xor eax, eax
pop esi
pop ebp
retn 4
... skipped
XADD is “exchange and add”.

In this case, it adds 1 to LockCount, meanwhile saves initial value of LockCount in the EBX register. How-
ever, value in EBX is to incremented with a help of subsequent INC EBX, and it also will be equal to the
updated value of LockCount.
This operation is atomic since it is prefixed by LOCK as well, meaning that all other CPUs or CPU cores in
system are blocked from accessing this point in memory.
The LOCK prefix is very important:
778
without it two threads, each of which works on separate CPU or CPU core can try to enter a critical section
and to modify the value in memory, which will result in non-deterministic behavior.
779
Chapter 7
Tools
Now that Dennis Yurichev has made this book

free (libre), it is a contribution to the world of
free knowledge and free education. However,
for our freedom’s sake, we need free (libre)
reverse engineering tools to replace the
proprietary tools described in this book.
Richard M. Stallman
7.1 Binary analysis

Tools you use when you don’t run any process.
• (Free, open-source) ent 1 : entropy analyzing tool. Read more about entropy: 9.2 on page 929.
• Hiew2 : for small modifications of code in binary files. Has assembler/disassembler.
• (Free, open-source) GHex3 : simple hexadecimal editor for Linux.
• (Free, open-source) xxd and od: standard UNIX utilities for dumping.
• (Free, open-source) strings: *NIX tool for searching for ASCII strings in binary files, including exe-
cutable ones. Sysinternals has alternative4 supporting wide char strings (UTF-16, widely used in
Windows).
• (Free, open-source) Binwalk 5 : analyzing firmware images.
• (Free, open-source) binary grep: a small utility for searching any byte sequence in a big pile of files,
including non-executable ones: GitHub. There is also rafind2 in rada.re for the same purpose.
7.1.1 Disassemblers
• IDA. An older freeware version is available for download 6 . Hot-keys cheatsheet: .6.1 on page 1029
• Ghidra7 — free alternative to IDA from NSA.
• Binary Ninja8
• (Free, open-source) zynamics BinNavi9
• (Free, open-source) objdump: simple command-line utility for dumping and disassembling.
• (Free, open-source) readelf 10 : dump information about ELF file.
1 http://www.fourmilab.ch/random/
2 hiew.ru
3 https://wiki.gnome.org/Apps/Ghex
4 https://technet.microsoft.com/en-us/sysinternals/strings
5 http://binwalk.org/
6 hex-rays.com/products/ida/support/download_freeware.shtml
7 https://ghidra-sre.org/
8 http://binary.ninja/
9 https://www.zynamics.com/binnavi.html
10 https://sourceware.org/binutils/docs/binutils/readelf.html
780
7.1.2 Decompilers
The best known is Hex-Rays: http://hex-rays.com/products/decompiler/. Read more about it: 11.8
on page 988.
There is also a free alternative from NSA: Ghidra11 .
7.1.3 Patch comparison/diffing

You may want to use it when you compare original version of some executable and patched one, in order
to find what has been patched and why.
• (Free) zynamics BinDiff 12
• (Free, open-source) Diaphora13
7.2 Live analysis

Tools you use on a live system or during running of a process.
7.2.1 Debuggers
• (Free) OllyDbg. Very popular user-mode win32 debugger14 . Hot-keys cheatsheet: .6.2 on page 1030
• (Free, open-source) GDB. Not quite popular debugger among reverse engineers, because it’s in-
tended mostly for programmers. Some commands: .6.5 on page 1031. There is a visual interface for
GDB, “GDB dashboard”15 .
• (Free, open-source) LLDB16 .
• WinDbg17 : kernel debugger for Windows.
• IDA has internal debugger.
• (Free, open-source) Radare AKA rada.re AKA r218 . A GUI also exists: ragui19 .
20
• (Free, open-source) tracer. The author often uses tracer instead of a debugger.
The author of these lines stopped using a debugger eventually, since all he needs from it is to spot
function arguments while executing, or registers state at some point. Loading a debugger each time
is too much, so a small utility called tracer was born. It works from command line, allows intercepting
function execution, setting breakpoints at arbitrary places, reading and changing registers state, etc.
N.B.: the tracer isn’t evolving, because it was developed as a demonstration tool for this book, not
as everyday tool.
7.2.2 Library calls tracing

ltrace21 .
7.2.3 System calls tracing

strace / dtruss
It shows which system calls (syscalls( 6.3 on page 738)) are called by a process right now.
For example:
11 https://ghidra-sre.org/
12 https://www.zynamics.com/software.html
13 https://github.com/joxeankoret/diaphora
14 ollydbg.de
15 https://github.com/cyrus-and/gdb-dashboard
16 http://lldb.llvm.org/
17 https://developer.microsoft.com/en-us/windows/hardware/windows-driver-kit
18 http://rada.re/r/
19 http://radare.org/ragui/
20 yurichev.com
21 http://www.ltrace.org/
781
# strace df -h
...
access("/etc/ld.so.nohwcap", F_OK) = -1 ENOENT (No such file or directory)

open("/lib/i386-linux-gnu/libc.so.6", O_RDONLY|O_CLOEXEC) = 3
read(3, "\177ELF\1\1\1\0\0\0\0\0\0\0\0\0\3\0\3\0\1\0\0\0\220\232\1\0004\0\0\0"..., 512) = 512
fstat64(3, {st_mode=S_IFREG|0755, st_size=1770984, ...}) = 0
mmap2(NULL, 1780508, PROT_READ|PROT_EXEC, MAP_PRIVATE|MAP_DENYWRITE, 3, 0) = 0xb75b3000
Mac OS X has dtruss for doing the same.

Cygwin also has strace, but as far as it’s known, it works only for .exe-files compiled for the cygwin
environment itself.
7.2.4 Network sniffing

Sniffing is intercepting some information you may be interested in.
(Free, open-source) Wireshark 22 for network sniffing. It has also capability for USB sniffing23 .
Wireshark has a younger (or older) brother tcpdump24 , simpler command-line tool.
7.2.5 Sysinternals
(Free) Sysinternals (developed by Mark Russinovich) 25 . At least these tools are important and worth
studying: Process Explorer, Handle, VMMap, TCPView, Process Monitor.
7.2.6 Valgrind
(Free, open-source) a powerful tool for detecting memory leaks: http://valgrind.org/. Due to its pow-
erful JIT mechanism, Valgrind is used as a framework for other tools.
7.2.7 Emulators
• (Free, open-source) QEMU26 : emulator for various CPUs and architectures.
• (Free, open-source) DosBox27 : MS-DOS emulator, mostly used for retrogaming.
• (Free, open-source) SimH28 : emulator of ancient computers, mainframes, etc.
7.3 Other tools

29
Microsoft Visual Studio Express : Stripped-down free version of Visual Studio, convenient for simple
experiments.
Some useful options: .6.3 on page 1030.
There is a website named “Compiler Explorer”, allowing to compile small code snippets and see output in
various GCC versions and architectures (at least x86, ARM, MIPS): http://godbolt.org/—I would have
used it myself for the book if I would know about it!
7.3.1 SMT solvers

From the reverse engineer’s perspective, SMT solvers are used when dealing with amateur cryptography,
symbolic/concolic execution, ROP chains generation.
For more information, read: https://yurichev.com/writings/SAT_SMT_by_example.pdf.
22 https://www.wireshark.org/
23 https://wiki.wireshark.org/CaptureSetup/USB
24 http://www.tcpdump.org/
25 https://technet.microsoft.com/en-us/sysinternals/bb842062
26 http://qemu.org
27 https://www.dosbox.com/
28 http://simh.trailing-edge.com/
29 visualstudio.com/en-US/products/visual-studio-express-vs
782
7.3.2 Calculators
Good calculator for reverse engineer’s needs should support at least decimal, hexadecimal and binary
bases, as well as many important operations like XOR and shifts.
• IDA has built-in calculator (“?”).
• rada.re has rax2.
• https://yurichev.com/progcalc/
• As a last resort, standard calculator in Windows has programmer’s mode.
7.4 Do You Think Something Is Missing Here?

If you know a great tool not listed here, please drop a note:
<first_name @ last_name . com> or <first_name . last_name @ gmail . com>.
783
Chapter 8
Case studies
Instead of epigraph:
Peter Seibel: How do you tackle reading source code? Even reading something in a
programming language you already know is a tricky problem.
Donald Knuth: But it’s really worth it for what it builds in your brain. So how do I do it?
There was a machine called the Bunker Ramo 300 and somebody told me that the Fortran
compiler for this machine was really amazingly fast, but nobody had any idea why it worked.
I got a copy of the source-code listing for it. I didn’t have a manual for the machine, so I
wasn’t even sure what the machine language was.
But I took it as an interesting challenge. I could figure out BEGIN and then I would start to
decode. The operation codes had some two-letter mnemonics and so I could start to figure
out “This probably was a load instruction, this probably was a branch.” And I knew it was a
Fortran compiler, so at some point it looked at column seven of a card, and that was where
it would tell if it was a comment or not.
After three hours I had figured out a little bit about the machine. Then I found these big,
branching tables. So it was a puzzle and I kept just making little charts like I’m working at
a security agency trying to decode a secret code. But I knew it worked and I knew it was a
Fortran compiler—it wasn’t encrypted in the sense that it was intentionally obscure; it was
only in code because I hadn’t gotten the manual for the machine.
Eventually I was able to figure out why this compiler was so fast. Unfortunately it wasn’t
because the algorithms were brilliant; it was just because they had used unstructured pro-
gramming and hand optimized the code to the hilt.
It was just basically the way you solve some kind of an unknown puzzle— make tables
and charts and get a little more information here and make a hypothesis. In general when
I’m reading a technical paper, it’s the same challenge. I’m trying to get into the author’s
mind, trying to figure out what the concept is. The more you learn to read other people’s
stuff, the more able you are to invent your own in the future, it seems to me.
( Peter Seibel — Coders at Work: Reflections on the Craft of Programming )
8.1 Mahjong solitaire prank (Windows 7)

Mahjong solitaire is a great game, but can we make it harder, by disabling the Hint menu item?
In my Windows 7, I can find Mahjong.dll and Mahjong.exe in:
C:\Windows\winsxs\
x86_microsoft-windows-s..inboxgames-shanghai_31bf3856ad364e35_6.1.7600.16385_none\
c07a51d9507d9398.
Also, the Mahjong.exe.mui file in:
C:\Windows\winsxs\
x86_microsoft-windows-s..-shanghai.resources_31bf3856ad364e35_6.1.7600.16385_en-us
_c430954533c66bf3
and
C:\Windows\winsxs\
x86_microsoft-windows-s..-shanghai.resources_31bf3856ad364e35_6.1.7600.16385_ru-ru
784
_0d51acf984cb679a.
I’m using English Windows, but with Russian language support, so there are might be resource files for
two languages. By opening Mahjong.exe.mui in Resource Hacker, there we can see a menu definition:
Listing 8.1: Menu resources from Mahjong.exe.mui
103 MENU
LANGUAGE LANG_ENGLISH, SUBLANG_ENGLISH_US
{
POPUP "&Game"
{
MENUITEM "&New Game\tF2", 40000
MENUITEM SEPARATOR
MENUITEM "&Undo\tCtrl+Z", 40001
MENUITEM "&Hint\tH", 40002
MENUITEM SEPARATOR
MENUITEM "&Statistics\tF4", 40003
MENUITEM "&Options\tF5", 40004
MENUITEM "Change &Appearance\tF7", 40005
MENUITEM SEPARATOR
MENUITEM "E&xit", 40006
}
POPUP "&Help"
{
MENUITEM "&View Help\tF1", 40015
MENUITEM "&About Mahjong Titans", 40016
MENUITEM SEPARATOR
MENUITEM "Get &More Games Online", 40020
}
}
The Hint submenu has the 40002 code. Now I’m opening Mahjong.exe in IDA and find the 40002 value.
(I’m writing this in November 2019. Somehow, IDA can’t get PDBs from Microsoft servers. Maybe Windows
7 is unsupported anymore? Anyway, I can’t get function names...)
Listing 8.2: Mahjong.exe
.text:010205C8 6A 03 push 3
.text:010205CA 85 FF test edi, edi
.text:010205CC 5B pop ebx
...
.text:01020625 57 push edi ; uIDEnableItem

.text:01020626 FF 35 C8 97 08 01 push hmenu ; hMenu
.text:0102062C FF D6 call esi ; EnableMenuItem
.text:0102062E 83 7D 08 01 cmp [ebp+arg_0], 1
.text:01020632 BF 42 9C 00 00 mov edi, 40002
.text:01020637 75 18 jnz short loc_1020651 ; must be always
.text:01020639 6A 00 push 0 ; uEnable
.text:0102063B 57 push edi ; uIDEnableItem
.text:0102063C FF 35 B4 8B 08 01 push hMenu ; hMenu
.text:01020642 FF D6 call esi ; EnableMenuItem
.text:01020644 6A 00 push 0 ; uEnable
.text:01020647 FF 35 C8 97 08 01 push hmenu ; hMenu
.text:0102064D FF D6 call esi ; EnableMenuItem
.text:0102064F EB 1A jmp short loc_102066B
.text:01020651 ;
---------------------------------------------------------------------------
.text:01020651
.text:01020651 loc_1020651: ; CODE XREF:
sub_1020581+B6
.text:01020651 53 push ebx ; 3
.text:01020653 FF 35 B4 8B 08 01 push hMenu ; hMenu
.text:0102065B 53 push ebx ; 3
.text:0102065C 57 push edi ; uIDEnableItem
.text:0102065D FF 35 C8 97 08 01 push hmenu ; hMenu
785
This piece of code enables or disables the Hint menu item.
And according to MSDN1 :
MF_DISABLED | MF_GRAYED = 3 and MF_ENABLED = 0.
I think, this function enables or disables several menu items (Hint, Undo, etc), according to the value in
arg_0. Because at start, when a user choose solitaire type, Hint and Undo are disabled. They are enabled
when the game has begun.
So I’m patching the Mahjong.exe file at 0x01020637 by replacing 0x75 with 0xEB byte, making this JNZ
jump working always. Effectively, this will make calling EnableMenuItem(..., ..., 3) always. Now the
Hint submenu is always disabled.
Also, somehow, EnableMenuItem() called twice, for hMenu and for hmenu. Perhaps, the program has two
menus, and maybe switching them?
As a homework, try to disable Undo menu item, to make the game even harder.
8.2 Task manager practical joke (Windows Vista)

Let’s see if it’s possible to hack Task Manager slightly so it would detect more CPU cores.
Let us first think, how does the Task Manager know the number of cores?
There is the GetSystemInfo() win32 function present in win32 userspace which can tell us this. But it’s
not imported in taskmgr.exe.
There is, however, another one in NTAPI, NtQuerySystemInformation(), which is used in taskmgr.exe
in several places.
To get the number of cores, one has to call this function with the SystemBasicInformation constant as a
first argument (which is zero 2 ).
The second argument has to point to the buffer which is getting all the information.
So we have to find all calls to the
NtQuerySystemInformation(0, ?, ?, ?) function. Let’s open taskmgr.exe in IDA.
What is always good about Microsoft executables is that IDA can download the corresponding PDB file for
this executable and show all function names.
It is visible that Task Manager is written in C++ and some of the function names and classes are really
speaking for themselves. There are classes CAdapter, CNetPage, CPerfPage, CProcInfo, CProcPage, CSvc-
Page, CTaskPage, CUserPage.
Apparently, each class corresponds to each tab in Task Manager.
Let’s visit each call and add comment with the value which is passed as the first function argument. We
will write “not zero” at some places, because the value there was clearly not zero, but something really
different (more about this in the second part of this chapter).
And we are looking for zero passed as argument, after all.
1 https://docs.microsoft.com/en-us/windows/win32/api/winuser/nf-winuser-enablemenuitem
2 MSDN
786
Figure 8.1: IDA: cross references to NtQuerySystemInformation()
Yes, the names are really speaking for themselves.

When we closely investigate each place where
NtQuerySystemInformation(0, ?, ?, ?) is called, we quickly find what we need in the InitPerfInfo()
function:
Listing 8.3: taskmgr.exe (Windows Vista)

.text:10000B4B3 xor r9d, r9d
.text:10000B4B6 lea rdx, [rsp+0C78h+var_C58] ; buffer
.text:10000B4BB xor ecx, ecx
.text:10000B4BD lea ebp, [r9+40h]
.text:10000B4C1 mov r8d, ebp
.text:10000B4C4 call cs:__imp_NtQuerySystemInformation ; 0
.text:10000B4CA xor ebx, ebx
.text:10000B4CC cmp eax, ebx
.text:10000B4CE jge short loc_10000B4D7
.text:10000B4D0
.text:10000B4D0 loc_10000B4D0: ; CODE XREF: InitPerfInfo(void)+97
.text:10000B4D0 ; InitPerfInfo(void)+AF
.text:10000B4D0 xor al, al
.text:10000B4D2 jmp loc_10000B5EA
.text:10000B4D7 ; ---------------------------------------------------------------------------
.text:10000B4D7
.text:10000B4D7 loc_10000B4D7: ; CODE XREF: InitPerfInfo(void)+36
.text:10000B4D7 mov eax, [rsp+0C78h+var_C50]
.text:10000B4DB mov esi, ebx
.text:10000B4DD mov r12d, 3E80h
.text:10000B4E3 mov cs:?g_PageSize@@3KA, eax ; ulong g_PageSize
.text:10000B4E9 shr eax, 0Ah
.text:10000B4EC lea r13, __ImageBase
.text:10000B4F3 imul eax, [rsp+0C78h+var_C4C]
.text:10000B4F8 cmp [rsp+0C78h+var_C20], bpl
.text:10000B4FD mov cs:?g_MEMMax@@3_JA, rax ; __int64 g_MEMMax
.text:10000B504 movzx eax, [rsp+0C78h+var_C20] ; number of CPUs
.text:10000B509 cmova eax, ebp
.text:10000B50C cmp al, bl
.text:10000B50E mov cs:?g_cProcessors@@3EA, al ; uchar g_cProcessors
g_cProcessors is a global variable, and this name has been assigned by IDA according to the PDB loaded
from Microsoft’s symbol server.
787
The byte is taken from var_C20. And var_C58 is passed to
NtQuerySystemInformation() as a pointer to the receiving buffer. The difference between 0xC20 and
0xC58 is 0x38 (56).
Let’s take a look at format of the return structure, which we can find in MSDN:
typedef struct _SYSTEM_BASIC_INFORMATION {
BYTE Reserved1[24];
PVOID Reserved2[4];
CCHAR NumberOfProcessors;
} SYSTEM_BASIC_INFORMATION;
This is a x64 system, so each PVOID takes 8 bytes.

All reserved fields in the structure take 24 + 4 ∗ 8 = 56 bytes.
Oh yes, this implies that var_C20 is the local stack is exactly the NumberOfProcessors field of the
SYSTEM_BASIC_INFORMATION structure.
Let’s check our guess. Copy taskmgr.exe from C:\Windows\System32 to some other folder (so the Win-
dows Resource Protection will not try to restore the patched taskmgr.exe).
Let’s open it in Hiew and find the place:
Figure 8.2: Hiew: find the place to be patched
Let’s replace the MOVZX instruction with ours. Let’s pretend we’ve got 64 CPU cores.
Add one additional NOP (because our instruction is shorter than the original one):
Figure 8.3: Hiew: patch it
And it works! Of course, the data in the graphs is not correct.

At times, Task Manager even shows an overall CPU load of more than 100%.
788
Figure 8.4: Fooled Windows Task Manager
The biggest number Task Manager does not crash with is 64.
Apparently, Task Manager in Windows Vista was not tested on computers with a large number of cores.
So there are probably some static data structure(s) inside it limited to 64 cores.
8.2.1 Using LEA to load values

Sometimes, LEA is used in taskmgr.exe instead of MOV to set the first argument of
NtQuerySystemInformation():
Listing 8.4: taskmgr.exe (Windows Vista)

xor r9d, r9d
div dword ptr [rsp+4C8h+WndClass.lpfnWndProc]
lea rdx, [rsp+4C8h+VersionInformation]
lea ecx, [r9+2] ; put 2 to ECX
mov r8d, 138h
mov ebx, eax
; ECX=SystemPerformanceInformation
call cs:__imp_NtQuerySystemInformation ; 2
...
mov r8d, 30h

lea r9, [rsp+298h+var_268]
lea rdx, [rsp+298h+var_258]
lea ecx, [r8-2Dh] ; put 3 to ECX
; ECX=SystemTimeOfDayInformation
call cs:__imp_NtQuerySystemInformation ; not zero
789
...
mov rbp, [rsi+8]

mov r8d, 20h
lea r9, [rsp+98h+arg_0]
lea rdx, [rsp+98h+var_78]
lea ecx, [r8+2Fh] ; put 0x4F to ECX
mov [rsp+98h+var_60], ebx
mov [rsp+98h+var_68], rbp
; ECX=SystemSuperfetchInformation
call cs:__imp_NtQuerySystemInformation ; not zero
Perhaps MSVC did so because machine code of LEA is shorter than MOV REG, 5 (would be 5 instead of 4).
LEA with offset in −128..127 range (offset will occupy 1 byte in opcode) with 32-bit registers is even shorter
(for lack of REX prefix)—3 bytes.
Another example of such thing is: 6.1.5 on page 729.
790
8.3 Color Lines game practical joke
This is a very popular game with several implementations in existence. We can take one of them, called
BallTriX, from 1997, available freely at http://go.yurichev.com/17311 3 . Here is how it looks:
Figure 8.5: This is how the game is usually looks like
3 Or at http://go.yurichev.com/17365 or http://go.yurichev.com/17366.
791
So let’s see, is it be possible to find the random generator and do some trick with it. IDA quickly recognize
the standard _rand function in balltrix.exe at 0x00403DA0. IDA also shows that it is called only from
one place:
.text:00402C9C sub_402C9C proc near ; CODE XREF: sub_402ACA+52
.text:00402C9C ; sub_402ACA+64 ...
.text:00402C9C
.text:00402C9C arg_0 = dword ptr 8
.text:00402C9C
.text:00402C9C push ebp
.text:00402C9D mov ebp, esp
.text:00402C9F push ebx
.text:00402CA0 push esi
.text:00402CA1 push edi
.text:00402CA2 mov eax, dword_40D430
.text:00402CA7 imul eax, dword_40D440
.text:00402CAE add eax, dword_40D5C8
.text:00402CB4 mov ecx, 32000
.text:00402CB9 cdq
.text:00402CBA idiv ecx
.text:00402CBC mov dword_40D440, edx
.text:00402CC2 call _rand
.text:00402CC7 cdq
.text:00402CC8 idiv [ebp+arg_0]
.text:00402CCB mov dword_40D430, edx
.text:00402CD1 mov eax, dword_40D430
.text:00402CD6 jmp $+5
.text:00402CDB pop edi
.text:00402CDC pop esi
.text:00402CDD pop ebx
.text:00402CDE leave
.text:00402CDF retn
.text:00402CDF sub_402C9C endp
We’ll call it “random”. Let’s not to dive into this function’s code yet.
This function is referred from 3 places.
Here are the first two:
.text:00402B16 mov eax, dword_40C03C ; 10 here
.text:00402B1B push eax
.text:00402B1C call random
.text:00402B21 add esp, 4
.text:00402B24 inc eax
.text:00402B25 mov [ebp+var_C], eax
.text:00402B28 mov eax, dword_40C040 ; 10 here
.text:00402B2D push eax
.text:00402B2E call random
Here is the third one:

.text:00402BBB mov eax, dword_40C058 ; 5 here
.text:00402BC0 push eax
.text:00402BC1 call random
.text:00402BC6 add esp, 4
.text:00402BC9 inc eax
So the function has only one argument.

10 is passed in first two cases and 5 in third. We can also notice that the board has a size of 10*10 and
there are 5 possible colors. This is it! The standard rand() function returns a number in the 0..0x7FFF
range and this is often inconvenient, so many programmers implement their own random functions which
returns a random number in a specified range. In our case, the range is 0..n − 1 and n is passed as the sole
argument of the function. We can quickly check this in any debugger.
So let’s fix the third function call to always return zero. First, we will replace three instructions (PUSH/CALL/ADD)
by NOPs. Then we’ll add XOR EAX, EAX instruction, to clear the EAX register.
792
.00402BB8: 83C410 add esp,010
.00402BBB: A158C04000 mov eax,[00040C058]
.00402BC0: 31C0 xor eax,eax
.00402BC2: 90 nop
.00402BC3: 90 nop
.00402BC4: 90 nop
.00402BC5: 90 nop
.00402BC6: 90 nop
.00402BC7: 90 nop
.00402BC8: 90 nop
.00402BC9: 40 inc eax
.00402BCA: 8B4DF8 mov ecx,[ebp][-8]
.00402BCD: 8D0C49 lea ecx,[ecx][ecx]*2
.00402BD0: 8B15F4D54000 mov edx,[00040D5F4]
So what we did is we replaced a call to the random() function by a code which always returns zero.
793
Let’s run it now:
Figure 8.6: Practical joke works
Oh yes, it works4 .
But why are the arguments to the random() functions global variables? That’s just because it’s possible
to change the board size in the game’s settings, so these values are not hardcoded. The 10 and 5 values
are just defaults.
8.4 Minesweeper (Windows XP)

For those who are not very good at playing Minesweeper, we could try to reveal the hidden mines in the
debugger.
As we know, Minesweeper places mines randomly, so there has to be some kind of random number
generator or a call to the standard rand() C-function.
What is really cool about reversing Microsoft products is that there are PDB file with symbols (function
names, etc). When we load winmine.exe into IDA, it downloads the PDB file exactly for this executable
and shows all names.
So here it is, the only call to rand() is this function:
.text:01003940 ; __stdcall Rnd(x)
.text:01003940 _Rnd@4 proc near ; CODE XREF: StartGame()+53
.text:01003940 ; StartGame()+61
.text:01003940
.text:01003940
.text:01003940 call ds:__imp__rand
.text:01003946 cdq
.text:01003947 idiv [esp+arg_0]
.text:0100394B mov eax, edx
.text:0100394D retn 4
.text:0100394D _Rnd@4 endp
IDA named it so, and it was the name given to it by Minesweeper’s developers.
The function is very simple:
int Rnd(int limit)
{
4 Author of this book once did this as a joke for his coworkers with the hope that they would stop playing. They didn’t.
794
return rand() % limit;
};
(There is no “limit” name in the PDB file; we manually named this argument like this.)
So it returns a random value from 0 to a specified limit.
Rnd() is called only from one place, a function called StartGame(), and as it seems, this is exactly the
code which place the mines:
.text:010036C7 push _xBoxMac
.text:010036CD call _Rnd@4 ; Rnd(x)
.text:010036D2 push _yBoxMac
.text:010036D8 mov esi, eax
.text:010036DA inc esi
.text:010036DB call _Rnd@4 ; Rnd(x)
.text:010036E0 inc eax
.text:010036E1 mov ecx, eax
.text:010036E3 shl ecx, 5 ; ECX=ECX*32
.text:010036E6 test _rgBlk[ecx+esi], 80h
.text:010036EE jnz short loc_10036C7
.text:010036F0 shl eax, 5 ; EAX=EAX*32
.text:010036F3 lea eax, _rgBlk[eax+esi]
.text:010036FA or byte ptr [eax], 80h
.text:010036FD dec _cBombStart
.text:01003703 jnz short loc_10036C7
Minesweeper allows you to set the board size, so the X (xBoxMac) and Y (yBoxMac) of the board are global
variables. They are passed to Rnd() and random coordinates are generated. A mine is placed by the OR
instruction at 0x010036FA. And if it has been placed before (it’s possible if the pair of Rnd() generates
a coordinates pair which has been already generated), then TEST and JNZ at 0x010036E6 jumps to the
generation routine again.
cBombStart is the global variable containing total number of mines. So this is loop.
The width of the array is 32 (we can conclude this by looking at the SHL instruction, which multiplies one
of the coordinates by 32).
The size of the rgBlk global array can be easily determined by the difference between the rgBlk label in
the data segment and the next known one. It is 0x360 (864):
.data:01005340 _rgBlk db 360h dup(?) ; DATA XREF: MainWndProc(x,x,x,x)+574
.data:01005340 ; DisplayBlk(x,x)+23
.data:010056A0 _Preferences dd ? ; DATA XREF: FixMenus()+2
...
864/32 = 27.
So the array size is 27 ∗ 32? It is close to what we know: when we try to set board size to 100 ∗ 100 in
Minesweeper settings, it fallbacks to a board of size 24 ∗ 30. So this is the maximal board size here. And
the array has a fixed size for any board size.
So let’s see all this in OllyDbg. We will ran Minesweeper, attaching OllyDbg to it and now we can see the
memory dump at the address of the rgBlk array (0x01005340) 5 .
So we got this memory dump of the array:
Address Hex dump
01005340 10 10 10 10|10 10 10 10|10 10 10 0F|0F 0F 0F 0F|
01005350 0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|
01005360 10 0F 0F 0F|0F 0F 0F 0F|0F 0F 10 0F|0F 0F 0F 0F|
01005380 10 0F 0F 0F|0F 0F 0F 0F|0F 0F 10 0F|0F 0F 0F 0F|
010053A0 10 0F 0F 0F|0F 0F 0F 0F|8F 0F 10 0F|0F 0F 0F 0F|
010053B0 0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|
010053C0 10 0F 0F 0F|0F 0F 0F 0F|0F 0F 10 0F|0F 0F 0F 0F|
010053D0 0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|
010053E0 10 0F 0F 0F|0F 0F 0F 0F|0F 0F 10 0F|0F 0F 0F 0F|
010053F0 0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|
5 All addresses here are for Minesweeper for Windows XP SP3 English. They may differ for other service packs.
795
01005400 10 0F 0F 8F|0F 0F 8F 0F|0F 0F 10 0F|0F 0F 0F 0F|
01005420 10 8F 0F 0F|8F 0F 0F 0F|0F 0F 10 0F|0F 0F 0F 0F|
01005440 10 8F 0F 0F|0F 0F 8F 0F|0F 8F 10 0F|0F 0F 0F 0F|
01005460 10 0F 0F 0F|0F 8F 0F 0F|0F 8F 10 0F|0F 0F 0F 0F|
01005480 10 10 10 10|10 10 10 10|10 10 10 0F|0F 0F 0F 0F|
010054A0 0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|
010054B0 0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|
010054C0 0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|0F 0F 0F 0F|
OllyDbg, like any other hexadecimal editor, shows 16 bytes per line. So each 32-byte array row occupies
exactly 2 lines here.
This is beginner level (9*9 board).
There is some square structure can be seen visually (0x10 bytes).
We will click “Run” in OllyDbg to unfreeze the Minesweeper process, then we’ll clicked randomly at the
Minesweeper window and trapped into mine, but now all mines are visible:
Figure 8.7: Mines
By comparing the mine places and the dump, we can conclude that 0x10 stands for border, 0x0F—empty
block, 0x8F—mine. Perhaps, 0x10 is just a sentinel value.
Now we’ll add comments and also enclose all 0x8F bytes into square brackets:
border:
01005340 10 10 10 10 10 10 10 10 10 10 10 0F 0F 0F 0F 0F
01005350 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F
line #1:
01005360 10 0F 0F 0F 0F 0F 0F 0F 0F 0F 10 0F 0F 0F 0F 0F
line #2:
01005380 10 0F 0F 0F 0F 0F 0F 0F 0F 0F 10 0F 0F 0F 0F 0F
line #3:
010053A0 10 0F 0F 0F 0F 0F 0F 0F[8F]0F 10 0F 0F 0F 0F 0F
010053B0 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F
line #4:
010053C0 10 0F 0F 0F 0F 0F 0F 0F 0F 0F 10 0F 0F 0F 0F 0F
010053D0 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F
line #5:
010053E0 10 0F 0F 0F 0F 0F 0F 0F 0F 0F 10 0F 0F 0F 0F 0F
796
010053F0 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F 0F
line #6:
01005400 10 0F 0F[8F]0F 0F[8F]0F 0F 0F 10 0F 0F 0F 0F 0F
line #7:
01005420 10[8F]0F 0F[8F]0F 0F 0F 0F 0F 10 0F 0F 0F 0F 0F
line #8:
01005440 10[8F]0F 0F 0F 0F[8F]0F 0F[8F]10 0F 0F 0F 0F 0F
line #9:
01005460 10 0F 0F 0F 0F[8F]0F 0F 0F[8F]10 0F 0F 0F 0F 0F
border:
01005480 10 10 10 10 10 10 10 10 10 10 10 0F 0F 0F 0F 0F
Now we’ll remove all border bytes (0x10) and what’s beyond those:
0F 0F 0F 0F 0F 0F 0F 0F 0F
0F 0F 0F 0F 0F 0F 0F 0F 0F
0F 0F 0F 0F 0F 0F 0F[8F]0F
0F 0F 0F 0F 0F 0F 0F 0F 0F
0F 0F 0F 0F 0F 0F 0F 0F 0F
0F 0F[8F]0F 0F[8F]0F 0F 0F
[8F]0F 0F[8F]0F 0F 0F 0F 0F
[8F]0F 0F 0F 0F[8F]0F 0F[8F]
0F 0F 0F 0F[8F]0F 0F 0F[8F]
Yes, these are mines, now it can be clearly seen and compared with the screenshot.
797
What is interesting is that we can modify the array right in OllyDbg. We can remove all mines by changing
all 0x8F bytes by 0x0F, and here is what we’ll get in Minesweeper:
Figure 8.8: All mines are removed in debugger
We can also move all of them to the first line:
Figure 8.9: Mines set in debugger
Well, the debugger is not very convenient for eavesdropping (which is our goal anyway), so we’ll write a
small utility to dump the contents of the board:
// Windows XP MineSweeper cheater
// written by dennis(a)yurichev.com for http://beginners.re/ book
#include <assert.h>
#include <stdio.h>
int main (int argc, char * argv[])

{
int i, j;
HANDLE h;
DWORD PID, address, rd;
BYTE board[27][32];
798
if (argc!=3)
{
printf ("Usage: %s <PID> <address>\n", argv[0]);
return 0;
};
assert (argv[1]!=NULL);
assert (argv[2]!=NULL);
assert (sscanf (argv[1], "%d", &PID)==1);

assert (sscanf (argv[2], "%x", &address)==1);
h=OpenProcess (PROCESS_VM_OPERATION | PROCESS_VM_READ | PROCESS_VM_WRITE, FALSE, PID);
if (h==NULL)
{
DWORD e=GetLastError();
printf ("OpenProcess error: %08X\n", e);
return 0;
};
if (ReadProcessMemory (h, (LPVOID)address, board, sizeof(board), &rd)!=TRUE)

{
printf ("ReadProcessMemory() failed\n");
return 0;
};
for (i=1; i<26; i++)

{
if (board[i][0]==0x10 && board[i][1]==0x10)
break; // end of board
for (j=1; j<31; j++)
{
if (board[i][j]==0x10)
break; // board border
if (board[i][j]==0x8F)
printf ("*");
else
printf (" ");
};
printf ("\n");
};
CloseHandle (h);
};
Just set the PID6 7

and the address of the array (0x01005340 for Windows XP SP3 English) and it will dump
it 8 .
It attaches itself to a win32 process by PID and just reads process memory at the address.
8.4.1 Finding grid automatically

This is kind of nuisance to set address each time when we run our utility. Also, various Minesweeper
versions may have the array on different address. Knowing the fact that there is always a border (0x10
bytes), we can just find it in memory:
// find frame to determine the address
process_mem=(BYTE*)malloc(process_mem_size);
assert (process_mem!=NULL);
if (ReadProcessMemory (h, (LPVOID)start_addr, process_mem, process_mem_size, &rd)!=TRUE⤦

Ç )
{
6 Program/process ID
7 PIDit can be seen in Task Manager (enable it in “View → Select Columns”)
8 The compiled executable is here: beginners.re
799
printf ("ReadProcessMemory() failed\n");
return 0;
};
// for 9*9 grid.

// FIXME: slow!
for (i=0; i<process_mem_size; i++)
{
if (memcmp(process_mem+i, "\x10\x10\x10\x10\x10\x10\x10\x10\x10\x10\x10\x0F\x0F⤦
Ç \x0F\x0F\x0F\x0F\x0F\x0F\x0F\x0F\x0F\x0F\x0F\x0F\x0F\x0F\x0F\x0F\x0F\x0F\x0F\x10", 32)⤦
Ç ==0)
{
// found
address=start_addr+i;
break;
};
};
if (address==0)
{
printf ("Can't determine address of frame (and grid)\n");
return 0;
}
else
{
printf ("Found frame and grid at 0x%x\n", address);
};
Full source code: https://beginners.re/current-tree/examples/minesweeper/minesweeper_cheater2.

c.
8.4.2 Exercises
• Why do the border bytes (or sentinel values) (0x10) exist in the array?
What they are for if they are not visible in Minesweeper’s interface? How could it work without them?
• As it turns out, there are more values possible (for open blocks, for flagged by user, etc). Try to find
the meaning of each one.
• Modify my utility so it can remove all mines or set them in a fixed pattern that you want in the
Minesweeper process currently running.
8.5 Hacking Windows clock

Sometimes I do some kind of first April prank for my coworkers.
Let’s find, if we could do something with Windows clock? Can we force to go clock hands backwards?
First of all, when you click on date/time in status bar,
a C:\WINDOWS\SYSTEM32\TIMEDATE.CPL module gets executed, which is usual executable PE-file.
Let’s see, how it draw hands? When I open the file (from Windows 7) in Resource Hacker, there are clock
faces, but with no hands:
800
Figure 8.10: Resource Hacker
OK, what we know? How to draw a clock hand? All they are started at the middle of circle, ending with its
border. Hence, we must calculate coordinates of a point on circle’s border. From school-level mathematics
we may recall that we have to use sine/cosine functions to draw circle, or at least square root. There are
no such things in TIMEDATE.CPL, at least at first glance. But, thanks to Microsoft debugging PDB files, I
can find a function named CAnalogClock::DrawHand(), which calls Gdiplus::Graphics::DrawLine() at least
twice.
Here is its code:
.text:6EB9DBC7 ; private: enum Gdiplus::Status __thiscall CAnalogClock::_DrawHand(class
Gdiplus::Graphics *, int, struct ClockHand const &, class Gdiplus::Pen *)
.text:6EB9DBC7 ?_DrawHand@CAnalogClock@@AAE?⤦
Ç AW4Status@Gdiplus@@PAVGraphics@3@HABUClockHand@@PAVPen@3@@Z proc near
.text:6EB9DBC7 ; CODE XREF: CAnalogClock::_ClockPaint(HDC__ *)+163
.text:6EB9DBC7 ; CAnalogClock::_ClockPaint(HDC__ *)+18B
.text:6EB9DBC7
.text:6EB9DBC7 var_10 = dword ptr -10h
.text:6EB9DBC7 var_C = dword ptr -0Ch
.text:6EB9DBC7 var_8 = dword ptr -8
.text:6EB9DBC7 var_4 = dword ptr -4
.text:6EB9DBC7 arg_0 = dword ptr 8
.text:6EB9DBC7 arg_4 = dword ptr 0Ch
.text:6EB9DBC7 arg_8 = dword ptr 10h
.text:6EB9DBC7 arg_C = dword ptr 14h
.text:6EB9DBC7
.text:6EB9DBC7 mov edi, edi
.text:6EB9DBC9 push ebp
.text:6EB9DBCA mov ebp, esp
.text:6EB9DBCC sub esp, 10h
.text:6EB9DBCF mov eax, [ebp+arg_4]
.text:6EB9DBD2 push ebx
.text:6EB9DBD3 push esi
.text:6EB9DBD4 push edi
801
.text:6EB9DBD5 cdq
.text:6EB9DBD6 push 3Ch
.text:6EB9DBD8 mov esi, ecx
.text:6EB9DBDA pop ecx
.text:6EB9DBDB idiv ecx
.text:6EB9DBDD push 2
.text:6EB9DBDF lea ebx, table[edx*8]
.text:6EB9DBE6 lea eax, [edx+1Eh]
.text:6EB9DBE9 cdq
.text:6EB9DBEA idiv ecx
.text:6EB9DBEC mov ecx, [ebp+arg_0]
.text:6EB9DBEF mov [ebp+var_4], ebx
.text:6EB9DBF2 lea eax, table[edx*8]
.text:6EB9DBF9 mov [ebp+arg_4], eax
.text:6EB9DBFC call ?SetInterpolationMode@Graphics@Gdiplus@@QAE?⤦
Ç AW4Status@2@W4InterpolationMode@2@@Z ;
Gdiplus::Graphics::SetInterpolationMode(Gdiplus::InterpolationMode)
.text:6EB9DC01 mov eax, [esi+70h]
.text:6EB9DC04 mov edi, [ebp+arg_8]
.text:6EB9DC07 mov [ebp+var_10], eax
.text:6EB9DC0A mov eax, [esi+74h]
.text:6EB9DC0D mov [ebp+var_C], eax
.text:6EB9DC10 mov eax, [edi]
.text:6EB9DC12 sub eax, [edi+8]
.text:6EB9DC15 push 8000 ; nDenominator
.text:6EB9DC1A push eax ; nNumerator
.text:6EB9DC1B push dword ptr [ebx+4] ; nNumber
.text:6EB9DC1E mov ebx, ds:__imp__MulDiv@12 ; MulDiv(x,x,x)
.text:6EB9DC24 call ebx ; MulDiv(x,x,x) ; MulDiv(x,x,x)
.text:6EB9DC26 add eax, [esi+74h]
.text:6EB9DC2E mov [ebp+arg_8], eax
.text:6EB9DC31 mov eax, [edi]
.text:6EB9DC33 sub eax, [edi+8]
.text:6EB9DC36 push eax ; nNumerator
.text:6EB9DC37 mov eax, [ebp+var_4]
.text:6EB9DC3A push dword ptr [eax] ; nNumber
.text:6EB9DC3C call ebx ; MulDiv(x,x,x) ; MulDiv(x,x,x)
.text:6EB9DC3E add eax, [esi+70h]
.text:6EB9DC41 mov ecx, [ebp+arg_0]
.text:6EB9DC47 mov eax, [ebp+arg_8]
.text:6EB9DC4A mov [ebp+var_4], eax
.text:6EB9DC4D lea eax, [ebp+var_8]
.text:6EB9DC50 push eax
.text:6EB9DC51 lea eax, [ebp+var_10]
.text:6EB9DC54 push eax
.text:6EB9DC55 push [ebp+arg_C]
.text:6EB9DC58 call ?DrawLine@Graphics@Gdiplus@@QAE?⤦
Ç AW4Status@2@PBVPen@2@ABVPoint@2@1@Z ; Gdiplus::Graphics::DrawLine(Gdiplus::Pen const
*,Gdiplus::Point const &,Gdiplus::Point const &)
.text:6EB9DC5D mov ecx, [edi+8]
.text:6EB9DC60 test ecx, ecx
.text:6EB9DC62 jbe short loc_6EB9DCAA
.text:6EB9DC64 test eax, eax
.text:6EB9DC66 jnz short loc_6EB9DCAA
.text:6EB9DC6B push 8000 ; nDenominator
.text:6EB9DC70 push ecx ; nNumerator
.text:6EB9DC71 push dword ptr [eax+4] ; nNumber
.text:6EB9DC76 add eax, [esi+74h]
.text:6EB9DC7E push dword ptr [edi+8] ; nNumerator
.text:6EB9DC81 mov [ebp+arg_8], eax
.text:6EB9DC87 push dword ptr [eax] ; nNumber
.text:6EB9DC8B add eax, [esi+70h]
.text:6EB9DC8E mov ecx, [ebp+arg_0]
802
.text:6EB9DC9A lea eax, [ebp+var_8]
.text:6EB9DC9D push eax
.text:6EB9DC9E lea eax, [ebp+var_10]
.text:6EB9DCA1 push eax
.text:6EB9DCA2 push [ebp+arg_C]
.text:6EB9DCA5 call ?DrawLine@Graphics@Gdiplus@@QAE?⤦
Ç AW4Status@2@PBVPen@2@ABVPoint@2@1@Z ; Gdiplus::Graphics::DrawLine(Gdiplus::Pen const
*,Gdiplus::Point const &,Gdiplus::Point const &)
.text:6EB9DCAA
.text:6EB9DCAA loc_6EB9DCAA: ; CODE XREF: CAnalogClock::_DrawHand(Gdiplus::Graphics
*,int,ClockHand const &,Gdiplus::Pen *)+9B
.text:6EB9DCAA ; CAnalogClock::_DrawHand(Gdiplus::Graphics *,int,ClockHand const
&,Gdiplus::Pen *)+9F
.text:6EB9DCAA pop edi
.text:6EB9DCAB pop esi
.text:6EB9DCAC pop ebx
.text:6EB9DCAD leave
.text:6EB9DCAE retn 10h
.text:6EB9DCAE ?_DrawHand@CAnalogClock@@AAE?⤦
Ç AW4Status@Gdiplus@@PAVGraphics@3@HABUClockHand@@PAVPen@3@@Z endp
.text:6EB9DCAE
We can see that DrawLine() arguments are dependent on result of MulDiv() function and a table[] table
(name is mine), which has 8-byte elements (look at LEA’s second operand).
What is inside of table[]?
.text:6EB87890 ; int table[]
.text:6EB87890 table dd 0
.text:6EB87894 dd 0FFFFE0C1h
.text:6EB87898 dd 344h
.text:6EB8789C dd 0FFFFE0ECh
.text:6EB878A0 dd 67Fh
.text:6EB878A4 dd 0FFFFE16Fh
.text:6EB878A8 dd 9A8h
.text:6EB878AC dd 0FFFFE248h
.text:6EB878B0 dd 0CB5h
.text:6EB878B4 dd 0FFFFE374h
.text:6EB878B8 dd 0F9Fh
.text:6EB878BC dd 0FFFFE4F0h
.text:6EB878C0 dd 125Eh
.text:6EB878C4 dd 0FFFFE6B8h
.text:6EB878C8 dd 14E9h
...
It’s referenced only from DrawHand() function. It has 120 32-bit words or 60 32-bit pairs... wait, 60? Let’s
take a closer look at these values. First of all, I’ll zap 6 pairs or 12 32-bit words with zeros, and then I’ll put
patched TIMEDATE.CPL into C:\WINDOWS\SYSTEM32. (You may need to set owner of the *TIMEDATE.CPL*
file to your primary user account (instead of TrustedInstaller), and also, boot in safe mode with command
prompt so you can copy the file, which is usually locked.)
803
Figure 8.11: Attempt to run
Now when any hand is located at 0..5 seconds/minutes, it’s invisible! However, opposite (shorter) part of
second hand is visible and moving. When any hand is outside of this area, hand is visible as usual.
Let’s take even closer look at the table in Mathematica. I have copypasted table from the TIMEDATE.CPL
to a tbl file (480 bytes). We will take for granted the fact that these are signed values, because half of
elements are below zero (0FFFFE0C1h, etc.). If these values would be unsigned, they would be suspiciously
huge.
In[]:= tbl = BinaryReadList["~/.../tbl", "Integer32"]
Out[]= {0, -7999, 836, -7956, 1663, -7825, 2472, -7608, 3253, -7308, 3999, \
-6928, 4702, -6472, 5353, -5945, 5945, -5353, 6472, -4702, 6928, \
-4000, 7308, -3253, 7608, -2472, 7825, -1663, 7956, -836, 8000, 0, \
7956, 836, 7825, 1663, 7608, 2472, 7308, 3253, 6928, 4000, 6472, \
4702, 5945, 5353, 5353, 5945, 4702, 6472, 3999, 6928, 3253, 7308, \
2472, 7608, 1663, 7825, 836, 7956, 0, 7999, -836, 7956, -1663, 7825, \
-2472, 7608, -3253, 7308, -4000, 6928, -4702, 6472, -5353, 5945, \
-5945, 5353, -6472, 4702, -6928, 3999, -7308, 3253, -7608, 2472, \
-7825, 1663, -7956, 836, -7999, 0, -7956, -836, -7825, -1663, -7608, \
-2472, -7308, -3253, -6928, -4000, -6472, -4702, -5945, -5353, -5353, \
-5945, -4702, -6472, -3999, -6928, -3253, -7308, -2472, -7608, -1663, \
-7825, -836, -7956}
In[]:= Length[tbl]
Out[]= 120
Let’s treat two consecutive 32-bit values as pair:

In[]:= pairs = Partition[tbl, 2]
Out[]= {{0, -7999}, {836, -7956}, {1663, -7825}, {2472, -7608}, \
{3253, -7308}, {3999, -6928}, {4702, -6472}, {5353, -5945}, {5945, \
-5353}, {6472, -4702}, {6928, -4000}, {7308, -3253}, {7608, -2472}, \
{7825, -1663}, {7956, -836}, {8000, 0}, {7956, 836}, {7825,
1663}, {7608, 2472}, {7308, 3253}, {6928, 4000}, {6472,
4702}, {5945, 5353}, {5353, 5945}, {4702, 6472}, {3999,
6928}, {3253, 7308}, {2472, 7608}, {1663, 7825}, {836, 7956}, {0,
7999}, {-836, 7956}, {-1663, 7825}, {-2472, 7608}, {-3253,
7308}, {-4000, 6928}, {-4702, 6472}, {-5353, 5945}, {-5945,
5353}, {-6472, 4702}, {-6928, 3999}, {-7308, 3253}, {-7608,
2472}, {-7825, 1663}, {-7956, 836}, {-7999,
0}, {-7956, -836}, {-7825, -1663}, {-7608, -2472}, {-7308, -3253}, \
{-6928, -4000}, {-6472, -4702}, {-5945, -5353}, {-5353, -5945}, \
{-4702, -6472}, {-3999, -6928}, {-3253, -7308}, {-2472, -7608}, \
{-1663, -7825}, {-836, -7956}}
In[]:= Length[pairs]
804
Out[]= 60
Let’s try to treat each pair as X/Y coordinate and draw all 60 pairs, and also first 15 pairs:
Figure 8.12: Mathematica
1
Now this is something! Each pair is just coordinate. First 15 pairs are coordinates for 4
of circle.
Perhaps, Microsoft developers precalculated all coordinates and put them into table. This is widespread,
though somewhat old school practice – precalculated table access is faster than calling relatively slow
sine/cosine functions9 . Sine/cosine operations are not that expensive anymore...
Now I can understand why when I zapped first 6 pairs, hands were invisible at that area: in fact, hands
were drawn, they just had zero length, because hand started at 0:0 coordinate and ended there.
9 Today this is known as memoization
805
The prank (practical joke)
Given all that, how would we force hands to go counterclockwise? In fact, this is simple, we need just to
rotate the table, so each hand, instead of drawing at place of zeroth second, would be drawing at place
of 59th second.
I made the patcher a long time ago, at the very beginning of 2000s, for Windows 2000. Hard to believe,
it still works for Windows 7, perhaps, the table hasn’t been changed since then!
The patcher source code: https://beginners.re/current-tree/examples/timedate/time_pt.c.
Now I can see all hands goes backwards:
Figure 8.13: Now it works
Well, there is no animation in this book, but if you look closer, you can see, that hands are in fact shows
correct time, but the whole clock face is rotated vertically, like we see it from the inside of clock.
Windows 2000 leaked source code
So I did the patcher and then Windows 2000 source code has been leaked (I can’t force you to trust me,
though). Let’s take a look on source code if that function and table.
The file is win2k/private/shell/cpls/utc/clock.c:
//
// Array containing the sine and cosine values for hand positions.
//
POINT rCircleTable[] =
{
{ 0, -7999},
{ 836, -7956},
{ 1663, -7825},
{ 2472, -7608},
{ 3253, -7308},
...
{ -4702, -6472},
{ -3999, -6928},
{ -3253, -7308},
{ -2472, -7608},
{ -1663, -7825},
{ -836 , -7956},
};
////////////////////////////////////////////////////////////////////////////
//
// DrawHand
//
// Draws the hands of the clock.
806
//
////////////////////////////////////////////////////////////////////////////
void DrawHand(
HDC hDC,
int pos,
HPEN hPen,
int scale,
int patMode,
PCLOCKSTR np)
{
LPPOINT lppt;
int radius;
MoveTo(hDC, np->clockCenter.x, np->clockCenter.y);

radius = MulDiv(np->clockRadius, scale, 100);
lppt = rCircleTable + pos;
SetROP2(hDC, patMode);
SelectObject(hDC, hPen);
LineTo( hDC,
np->clockCenter.x + MulDiv(lppt->x, radius, 8000),
np->clockCenter.y + MulDiv(lppt->y, radius, 8000) );
}
Now it’s clear: coordinates has been precalculated as if clock face has height and width of 2 ⋅ 8000, and
then it’s rescaled to current clock face radius using MulDiv() function.
POINT structure10 is a structure of two 32-bit values, first is x, second is y.
8.6 (Windows 7) Solitaire: practical jokes

8.6.1 51 cards
This is a joke I did once for my coworkers who played a Solitaire game too much. I was wondering, if it’s
possible to remove some cards, or maybe add (duplicates).
I opened Solitaire.exe in IDA disassembler, which asked if to download PDB file for it from Microsoft servers.
This is usually a rule for many Windows executables and DLLs. At least, PDB has all function names.
Then I tried to find a 52 number in all functions (because this card game uses 52 cards). As it turned out,
only 2 functions has it.
The first is:
.text:00000001000393B4 ; __int64 __fastcall SolitaireGame::OnMoveComplete(SolitaireGame *this)
.text:00000001000393B4 ?OnMoveComplete@SolitaireGame@@QEAAHXZ proc near
...
The second is the function with self-describing name (name pulled from PDB by IDA): InitialDeal():
.text:00000001000365F8 ; void __fastcall SolitaireGame::InitialDeal(SolitaireGame *__hidden ⤦
Ç this)
.text:00000001000365F8 ?InitialDeal@SolitaireGame@@QEAAXXZ proc near
.text:00000001000365F8
.text:00000001000365F8 var_58 = byte ptr -58h
.text:00000001000365F8 var_48 = qword ptr -48h
.text:00000001000365F8 var_40 = dword ptr -40h
.text:00000001000365F8 var_3C = dword ptr -3Ch
.text:00000001000365F8 var_38 = dword ptr -38h
.text:00000001000365F8 var_30 = qword ptr -30h
.text:00000001000365F8 var_28 = xmmword ptr -28h
.text:00000001000365F8 var_18 = byte ptr -18h
.text:00000001000365F8
.text:00000001000365F8 ; FUNCTION CHUNK AT .text:00000001000A55C2 SIZE 00000018 BYTES
.text:00000001000365F8
.text:00000001000365F8 ; __unwind { // __CxxFrameHandler3
10 https://msdn.microsoft.com/en-us/library/windows/desktop/dd162805(v=vs.85).aspx
807
.text:00000001000365F8 mov rax, rsp
.text:00000001000365FB push rdi
.text:00000001000365FC push r12
.text:00000001000365FE push r13
.text:0000000100036600 sub rsp, 60h
.text:0000000100036604 mov [rsp+78h+var_48], 0FFFFFFFFFFFFFFFEh
.text:000000010003660D mov [rax+8], rbx
.text:0000000100036611 mov [rax+10h], rbp
.text:0000000100036615 mov [rax+18h], rsi
.text:0000000100036619 movaps xmmword ptr [rax-28h], xmm6
.text:000000010003661D mov rsi, rcx
.text:0000000100036620 xor edx, edx ; struct Card *
.text:0000000100036622 call ?⤦
Ç SetSelectedCard@SolitaireGame@@QEAAXPEAVCard@@@Z ; SolitaireGame::SetSelectedCard(Card *)
.text:0000000100036627 and qword ptr [rsi+0F0h], 0
.text:000000010003662F mov rax, cs:?⤦
Ç g_pSolitaireGame@@3PEAVSolitaireGame@@EA ; SolitaireGame * g_pSolitaireGame
.text:0000000100036636 mov rdx, [rax+48h]
.text:000000010003663A cmp byte ptr [rdx+51h], 0
.text:000000010003663E jz short loc_10003664E
.text:0000000100036640 xor r8d, r8d ; bool
.text:0000000100036643 mov dl, 1 ; int
.text:0000000100036645 lea ecx, [r8+3] ; this
.text:0000000100036649 call ?PlaySoundProto@GameAudio@@YA_NH_NPEAI@Z ; ⤦
Ç GameAudio::PlaySoundProto(int,bool,uint *)
.text:000000010003664E
.text:000000010003664E loc_10003664E: ; CODE XREF: SolitaireGame::⤦
Ç InitialDeal(void)+46
.text:000000010003664E mov rbx, [rsi+88h]
.text:0000000100036655 mov r8d, 4
.text:000000010003665B lea rdx, aCardstackCreat ; "CardStack::CreateDeck()⤦
Ç ::uiNumSuits == "...
.text:0000000100036662 mov ebp, 10000h
.text:0000000100036667 mov ecx, ebp ; unsigned int
.text:0000000100036669 call ?Log@@YAXIPEBGZZ ; Log(uint,ushort const *,...)
.text:000000010003666E mov r8d, 52 ; ---
.text:0000000100036674 lea rdx, aCardstackCreat_0 ; "CardStack::CreateDeck⤦
Ç ()::uiNumCards == "...
.text:000000010003667B mov ecx, ebp ; unsigned int
.text:000000010003667D call ?Log@@YAXIPEBGZZ ; Log(uint,ushort const *,...)
.text:0000000100036682 xor edi, edi
.text:0000000100036684 loc_100036684: ; CODE XREF: SolitaireGame::⤦

Ç InitialDeal(void)+C0
.text:0000000100036684 mov eax, 4EC4EC4Fh
.text:0000000100036689 mul edi
.text:000000010003668B mov r8d, edx
.text:000000010003668E shr r8d, 4 ; unsigned int
.text:0000000100036692 mov eax, r8d
.text:0000000100036695 imul eax, 52 ; ---
.text:0000000100036698 mov edx, edi
.text:000000010003669A sub edx, eax ; unsigned int
.text:000000010003669C mov rcx, [rbx+128h] ; this
.text:00000001000366A3 call ?CreateCard@CardTable@@IEAAPEAVCard@@II@Z ; ⤦
Ç CardTable::CreateCard(uint,uint)
.text:00000001000366A8 mov rdx, rax ; struct Card *
.text:00000001000366AB mov rcx, rbx ; this
.text:00000001000366AE call ?Push@CardStack@@QEAAXPEAVCard@@@Z ; CardStack::⤦
Ç Push(Card *)
.text:00000001000366B3 inc edi
.text:00000001000366B5 cmp edi, 52 ; ---
.text:00000001000366B8 jb short loc_100036684
.text:00000001000366BA xor r8d, r8d ; bool

.text:00000001000366BD xor edx, edx ; bool
.text:00000001000366BF mov rcx, rbx ; this
.text:00000001000366C2 call ?Arrange@CardStack@@QEAAX_N0@Z ; CardStack::⤦
Ç Arrange(bool,bool)
.text:00000001000366C7 mov r13, [rsi+88h]
.text:00000001000366CE lea rdx, aCardstackShuff ; "CardStack::Shuffle()"
808
.text:00000001000366D5 mov ecx, ebp ; unsigned int
.text:00000001000366D7 call ?Log@@YAXIPEBGZZ ; Log(uint,ushort const *,...)
.text:00000001000366DC and [rsp+78h+var_40], 0
.text:00000001000366E1 and [rsp+78h+var_3C], 0
.text:00000001000366E6 mov [rsp+78h+var_38], 10h
.text:00000001000366EE xor ebx, ebx
.text:00000001000366F0 mov [rsp+78h+var_30], rbx
...
Anyway, we clearly see a loop of 52 iterations. A loop body has calls to CardTable()::CreateCard() and
CardStack::Push().
The CardTable::CreateCard() eventually calls Card::Init() with values in 0..51 range, as one of its
arguments. This can be easily checked using debugger.
So I tried just to change the 52 (0x34) number to 51 (0x33) in the cmp edi, 52 instruction at 0x1000366B5
and run it. At first glance, nothing happened, but I noticed that now it’s hard to solve the game. I spent
almost an hour to reach this position:
Ace of hearts is missing. Perhaps, internally, this card is numbered as 51th (if to number them from zero).
In the other place I found all card names. Maybe names to be used to fetch card graphics from resources?
.data:00000001000B6970 ?CARD_NAME@Card@@2PAPEBGA dq offset aTwoofclubs
.data:00000001000B6970 ; "TwoOfClubs"
.data:00000001000B6978 dq offset aThreeofclubs ; "ThreeOfClubs"
.data:00000001000B6980 dq offset aFourofclubs ; "FourOfClubs"
.data:00000001000B6988 dq offset aFiveofclubs ; "FiveOfClubs"
.data:00000001000B6990 dq offset aSixofclubs ; "SixOfClubs"
.data:00000001000B6998 dq offset aSevenofclubs ; "SevenOfClubs"
.data:00000001000B69A0 dq offset aEightofclubs ; "EightOfClubs"
.data:00000001000B69A8 dq offset aNineofclubs ; "NineOfClubs"
.data:00000001000B69B0 dq offset aTenofclubs ; "TenOfClubs"
.data:00000001000B69B8 dq offset aJackofclubs ; "JackOfClubs"
809
.data:00000001000B69C0 dq offset aQueenofclubs ; "QueenOfClubs"
.data:00000001000B69C8 dq offset aKingofclubs ; "KingOfClubs"
.data:00000001000B69D0 dq offset aAceofclubs ; "AceOfClubs"
.data:00000001000B69D8 dq offset aTwoofdiamonds ; "TwoOfDiamonds"
.data:00000001000B69E0 dq offset aThreeofdiamond ; "ThreeOfDiamonds"
.data:00000001000B69E8 dq offset aFourofdiamonds ; "FourOfDiamonds"
.data:00000001000B69F0 dq offset aFiveofdiamonds ; "FiveOfDiamonds"
.data:00000001000B69F8 dq offset aSixofdiamonds ; "SixOfDiamonds"
.data:00000001000B6A00 dq offset aSevenofdiamond ; "SevenOfDiamonds"
.data:00000001000B6A08 dq offset aEightofdiamond ; "EightOfDiamonds"
.data:00000001000B6A10 dq offset aNineofdiamonds ; "NineOfDiamonds"
.data:00000001000B6A18 dq offset aTenofdiamonds ; "TenOfDiamonds"
.data:00000001000B6A20 dq offset aJackofdiamonds ; "JackOfDiamonds"
.data:00000001000B6A28 dq offset aQueenofdiamond ; "QueenOfDiamonds"
.data:00000001000B6A30 dq offset aKingofdiamonds ; "KingOfDiamonds"
.data:00000001000B6A38 dq offset aAceofdiamonds ; "AceOfDiamonds"
.data:00000001000B6A40 dq offset aTwoofspades ; "TwoOfSpades"
.data:00000001000B6A48 dq offset aThreeofspades ; "ThreeOfSpades"
.data:00000001000B6A50 dq offset aFourofspades ; "FourOfSpades"
.data:00000001000B6A58 dq offset aFiveofspades ; "FiveOfSpades"
.data:00000001000B6A60 dq offset aSixofspades ; "SixOfSpades"
.data:00000001000B6A68 dq offset aSevenofspades ; "SevenOfSpades"
.data:00000001000B6A70 dq offset aEightofspades ; "EightOfSpades"
.data:00000001000B6A78 dq offset aNineofspades ; "NineOfSpades"
.data:00000001000B6A80 dq offset aTenofspades ; "TenOfSpades"
.data:00000001000B6A88 dq offset aJackofspades ; "JackOfSpades"
.data:00000001000B6A90 dq offset aQueenofspades ; "QueenOfSpades"
.data:00000001000B6A98 dq offset aKingofspades ; "KingOfSpades"
.data:00000001000B6AA0 dq offset aAceofspades ; "AceOfSpades"
.data:00000001000B6AA8 dq offset aTwoofhearts ; "TwoOfHearts"
.data:00000001000B6AB0 dq offset aThreeofhearts ; "ThreeOfHearts"
.data:00000001000B6AB8 dq offset aFourofhearts ; "FourOfHearts"
.data:00000001000B6AC0 dq offset aFiveofhearts ; "FiveOfHearts"
.data:00000001000B6AC8 dq offset aSixofhearts ; "SixOfHearts"
.data:00000001000B6AD0 dq offset aSevenofhearts ; "SevenOfHearts"
.data:00000001000B6AD8 dq offset aEightofhearts ; "EightOfHearts"
.data:00000001000B6AE0 dq offset aNineofhearts ; "NineOfHearts"
.data:00000001000B6AE8 dq offset aTenofhearts ; "TenOfHearts"
.data:00000001000B6AF0 dq offset aJackofhearts ; "JackOfHearts"
.data:00000001000B6AF8 dq offset aQueenofhearts ; "QueenOfHearts"
.data:00000001000B6B00 dq offset aKingofhearts ; "KingOfHearts"
.data:00000001000B6B08 dq offset aAceofhearts ; "AceOfHearts"
.data:00000001000B6B10 ; public: static unsigned short const * near * Card::CARD_HUMAN_NAME

.data:00000001000B6B10 ?CARD_HUMAN_NAME@Card@@2PAPEBGA dq offset a54639Cardnames
.data:00000001000B6B10 ; "|54639|CardNames|Two Of Clubs⤦
Ç "
.data:00000001000B6B18 dq offset a64833Cardnames ; "|64833|CardNames|Three Of ⤦
Ç Clubs"
.data:00000001000B6B20 dq offset a62984Cardnames ; "|62984|CardNames|Four Of ⤦
Ç Clubs"
.data:00000001000B6B28 dq offset a65200Cardnames ; "|65200|CardNames|Five Of ⤦
Ç Clubs"
.data:00000001000B6B30 dq offset a52967Cardnames ; "|52967|CardNames|Six Of ⤦
Ç Clubs"
.data:00000001000B6B38 dq offset a42781Cardnames ; "|42781|CardNames|Seven Of ⤦
Ç Clubs"
.data:00000001000B6B40 dq offset a49217Cardnames ; "|49217|CardNames|Eight Of ⤦
Ç Clubs"
.data:00000001000B6B48 dq offset a44682Cardnames ; "|44682|CardNames|Nine Of ⤦
Ç Clubs"
.data:00000001000B6B50 dq offset a51853Cardnames ; "|51853|CardNames|Ten Of ⤦
Ç Clubs"
.data:00000001000B6B58 dq offset a46368Cardnames ; "|46368|CardNames|Jack Of ⤦
Ç Clubs"
.data:00000001000B6B60 dq offset a61344Cardnames ; "|61344|CardNames|Queen Of ⤦
Ç Clubs"
.data:00000001000B6B68 dq offset a65017Cardnames ; "|65017|CardNames|King Of ⤦
Ç Clubs"
.data:00000001000B6B70 dq offset a57807Cardnames ; "|57807|CardNames|Ace Of ⤦
810
Ç Clubs"
.data:00000001000B6B78 dq offset a48455Cardnames ; "|48455|CardNames|Two Of ⤦
Ç Diamonds"
.data:00000001000B6B80 dq offset a44156Cardnames ; "|44156|CardNames|Three Of ⤦
Ç Diamonds"
.data:00000001000B6B88 dq offset a51672Cardnames ; "|51672|CardNames|Four Of ⤦
Ç Diamonds"
.data:00000001000B6B90 dq offset a45972Cardnames ; "|45972|CardNames|Five Of ⤦
Ç Diamonds"
.data:00000001000B6B98 dq offset a47206Cardnames ; "|47206|CardNames|Six Of ⤦
Ç Diamonds"
.data:00000001000B6BA0 dq offset a48399Cardnames ; "|48399|CardNames|Seven Of ⤦
Ç Diamonds"
.data:00000001000B6BA8 dq offset a47847Cardnames ; "|47847|CardNames|Eight Of ⤦
Ç Diamonds"
.data:00000001000B6BB0 dq offset a48606Cardnames ; "|48606|CardNames|Nine Of ⤦
Ç Diamonds"
.data:00000001000B6BB8 dq offset a61278Cardnames ; "|61278|CardNames|Ten Of ⤦
Ç Diamonds"
.data:00000001000B6BC0 dq offset a52038Cardnames ; "|52038|CardNames|Jack Of ⤦
Ç Diamonds"
.data:00000001000B6BC8 dq offset a54643Cardnames ; "|54643|CardNames|Queen Of ⤦
Ç Diamonds"
.data:00000001000B6BD0 dq offset a48902Cardnames ; "|48902|CardNames|King Of ⤦
Ç Diamonds"
.data:00000001000B6BD8 dq offset a46672Cardnames ; "|46672|CardNames|Ace Of ⤦
Ç Diamonds"
.data:00000001000B6BE0 dq offset a41049Cardnames ; "|41049|CardNames|Two Of ⤦
Ç Spades"
.data:00000001000B6BE8 dq offset a49327Cardnames ; "|49327|CardNames|Three Of ⤦
Ç Spades"
.data:00000001000B6BF0 dq offset a51933Cardnames ; "|51933|CardNames|Four Of ⤦
Ç Spades"
.data:00000001000B6BF8 dq offset a42651Cardnames ; "|42651|CardNames|Five Of ⤦
Ç Spades"
.data:00000001000B6C00 dq offset a65342Cardnames ; "|65342|CardNames|Six Of ⤦
Ç Spades"
.data:00000001000B6C08 dq offset a53644Cardnames ; "|53644|CardNames|Seven Of ⤦
Ç Spades"
.data:00000001000B6C10 dq offset a54466Cardnames ; "|54466|CardNames|Eight Of ⤦
Ç Spades"
.data:00000001000B6C18 dq offset a56874Cardnames ; "|56874|CardNames|Nine Of ⤦
Ç Spades"
.data:00000001000B6C20 dq offset a46756Cardnames ; "|46756|CardNames|Ten Of ⤦
Ç Spades"
.data:00000001000B6C28 dq offset a62876Cardnames ; "|62876|CardNames|Jack Of ⤦
Ç Spades"
.data:00000001000B6C30 dq offset a64633Cardnames ; "|64633|CardNames|Queen Of ⤦
Ç Spades"
.data:00000001000B6C38 dq offset a46215Cardnames ; "|46215|CardNames|King Of ⤦
Ç Spades"
.data:00000001000B6C40 dq offset a60450Cardnames ; "|60450|CardNames|Ace Of ⤦
Ç Spades"
.data:00000001000B6C48 dq offset a51010Cardnames ; "|51010|CardNames|Two Of ⤦
Ç Hearts"
.data:00000001000B6C50 dq offset a64948Cardnames ; "|64948|CardNames|Three Of ⤦
Ç Hearts"
.data:00000001000B6C58 dq offset a43079Cardnames ; "|43079|CardNames|Four Of ⤦
Ç Hearts"
.data:00000001000B6C60 dq offset a57131Cardnames ; "|57131|CardNames|Five Of ⤦
Ç Hearts"
.data:00000001000B6C68 dq offset a58953Cardnames ; "|58953|CardNames|Six Of ⤦
Ç Hearts"
.data:00000001000B6C70 dq offset a45105Cardnames ; "|45105|CardNames|Seven Of ⤦
Ç Hearts"
.data:00000001000B6C78 dq offset a47775Cardnames ; "|47775|CardNames|Eight Of ⤦
Ç Hearts"
.data:00000001000B6C80 dq offset a41825Cardnames ; "|41825|CardNames|Nine Of ⤦
Ç Hearts"
.data:00000001000B6C88 dq offset a41501Cardnames ; "|41501|CardNames|Ten Of ⤦
811
Ç Hearts"
.data:00000001000B6C90 dq offset a47108Cardnames ; "|47108|CardNames|Jack Of ⤦
Ç Hearts"
.data:00000001000B6C98 dq offset a55659Cardnames ; "|55659|CardNames|Queen Of ⤦
Ç Hearts"
.data:00000001000B6CA0 dq offset a44572Cardnames ; "|44572|CardNames|King Of ⤦
Ç Hearts"
.data:00000001000B6CA8 dq offset a44183Cardnames ; "|44183|CardNames|Ace Of ⤦
Ç Hearts"
If you want to do this to someone, be sure his/her mental health is stable.

Aside of function names from PDB file, there are lots of Log() function calls that can help significantly,
because the Solitaire game reports about what it’s doing right now.
Homework: try to remove several cards or two of clubs. And what if to swap card names in arrays of
strings?
I also tried to pass a numbers like 0, 0..50 to Card:Init() (so to have 2 zeroes in a list of 52 numbers).
Then I saw two two of clubs cards at one moment, but Solitaire behaves erratically.
This is patched Windows 7 Solitare: Solitaire51.exe.
8.6.2 53 cards
Now take a look on the first part of the loop:
.text:0000000100036684 loc_100036684: ; CODE XREF: SolitaireGame::⤦
Ç InitialDeal(void)+↓C0j
.text:0000000100036684 mov eax, 4EC4EC4Fh
.text:0000000100036689 mul edi
.text:000000010003668B mov r8d, edx
.text:000000010003668E shr r8d, 4 ; unsigned int
.text:0000000100036692 mov eax, r8d
.text:0000000100036695 imul eax, 52
.text:0000000100036698 mov edx, edi
.text:000000010003669A sub edx, eax ; unsigned int
.text:000000010003669C mov rcx, [rbx+128h] ; this
.text:00000001000366A3 call ?CreateCard@CardTable@@IEAAPEAVCard@@II@Z ; ⤦
Ç CardTable::CreateCard(uint,uint)
.text:00000001000366A8 mov rdx, rax ; struct Card *
.text:00000001000366AB mov rcx, rbx ; this
.text:00000001000366AE call ?Push@CardStack@@QEAAXPEAVCard@@@Z ; CardStack::⤦
Ç Push(Card *)
.text:00000001000366B3 inc edi
.text:00000001000366B5 cmp edi, 52
.text:00000001000366B8 jb short loc_100036684
What is with multiplication by 4EC4EC4Fh? Surely, this is division by multiplication. And what Hex-Rays
can say:
v5 = 0;
do
{
v6 = CardTable::CreateCard(v4[37], v5 % 0x34, v5 / 0x34);
CardStack::Push((CardStack *)v4, v6);
++v5;
}
while ( v5 < 0x34 );
Somehow, CreateCard() functions takes two arguments: iterator divided by 52 and a remainder of the
division operation. Hard to say, why they did so. Solitaire can’t allow more than 52 cards, so the last
argument is senseless, it’s always zero.
But when I patch cmp edi, 52 instruction at 0x1000366B5 to be cmp edi, 53, I found that there are now
53 cards. The last one is two of clubs, because it’s numbered as 0th card.
During the last iteration, 0x52 is divided by 0x52, remainder is zero, so 0th card is added twice.
What a frustration, there are two two of clubs:
812
This is patched Windows 7 Solitare: Solitaire53.
8.7 FreeCell prank (Windows 7)

This is a prank I once played for my coworkers who played FreeCell solitaire too much. Can we make
FreeCell always deal the same game each time? Like, you know, in “Groundhog Day” movie?
(I’m writing this in November 2019. Somehow, IDA can’t get PDBs from Microsoft servers. Maybe Windows
7 is unsupported anymore? Anyway, I can’t get function names...)
8.7.1 Part I
So I loaded FreeCell.exe into IDA and found that both rand(), srand() and time() are imported from msvcrt.dll.
time() is indeed used as a seed for srand():
.text:01029612 sub_1029612 proc near ; CODE XREF:
sub_102615C+149
.text:01029612 ; sub_1029DA6+67
.text:01029612 8B FF mov edi, edi
.text:01029614 56 push esi
.text:01029615 57 push edi
.text:01029616 6A 00 push 0 ; Time
.text:01029618 8B F9 mov edi, ecx
.text:0102961A FF 15 80 16 00+ call ds:time
.text:01029620 50 push eax ; Seed
.text:01029621 FF 15 84 16 00+ call ds:srand
.text:01029627 8B 35 AC 16 00+ mov esi, ds:rand
.text:0102962D 59 pop ecx
.text:0102962E 59 pop ecx
.text:0102962F FF D6 call esi ; rand
.text:01029631 FF D6 call esi ; rand
813
.text:01029633
.text:01029633 loc_1029633: ; CODE XREF:
sub_1029612+26
.text:01029633 ; sub_1029612+2D
.text:01029633 FF D6 call esi ; rand
.text:01029635 83 F8 01 cmp eax, 1
.text:01029638 7C F9 jl short loc_1029633
.text:0102963A 3D 40 42 0F 00 cmp eax, 1000000
.text:0102963F 7F F2 jg short loc_1029633
.text:01029641 6A 01 push 1
.text:01029643 50 push eax
.text:01029644 8B CF mov ecx, edi
.text:01029646 E8 2D F8 FF FF call sub_1028E78
.text:0102964B 5F pop edi
.text:0102964C 5E pop esi
.text:0102964D C3 retn
.text:0102964D sub_1029612 endp
Several (redundant) calls to rand() are funny, that reminds me:

“In the morning you will send for a hansom, desiring your man to take neither the first nor the second
which may present itself.” ( The Memoirs of Sherlock Holmes, by Arthur Conan Doyle11 )
There is another call of time() and srand() pair, but my tracer showed that this is the point of our interest:
tracer.exe -l:FreeCell.exe bpf=msvcrt.dll!time bpf=msvcrt.dll!srand,args:1
...
TID=5340|(0) msvcrt.dll!time() (called from FreeCell.exe!BASE+0x29620 (0x209620))

TID=5340|(0) msvcrt.dll!time() -> 0x5ddb68aa
TID=5340|(1) msvcrt.dll!srand(0x5ddb68aa) (called from FreeCell.exe!BASE+0x29627 (0x209627))
TID=5340|(1) msvcrt.dll!srand() -> 0x5507e0
TID=5340|(1) msvcrt.dll!srand(0x399f) (called from FreeCell.exe!BASE+0x27d3a (0x207d3a))
You see, the time() function returned 0x5ddb68aa and the very same value is used as an argument for
srand().
Let’s try to force time() to always return 0:
tracer.exe -l:FreeCell.exe bpf=msvcrt.dll!time,rt:0 bpf=msvcrt.dll!srand,args:1
...
TID=2104|(0) msvcrt.dll!time() (called from FreeCell.exe!BASE+0x29620 (0xb19620))

TID=2104|(0) msvcrt.dll!time() -> 0x5ddb68f6
TID=2104|(0) Modifying EAX register to 0x0
TID=2104|(1) msvcrt.dll!srand(0x0) (called from FreeCell.exe!BASE+0x29627 (0xb19627))
TID=2104|(1) msvcrt.dll!srand(0x52f6) (called from FreeCell.exe!BASE+0x27d3a (0xb17d3a))
Now I’m seeing the same game each time I’m running FreeCell using tracer:
11 http://www.gutenberg.org/files/834/834-0.txt
814
Now how to patch the executable?
We want to pass 0 as an argument to srand() at 0x01029620. But there is a one-byte instruction: PUSH
EAX. PUSH 0 is two-byte instruction. How to squeeze it into?
What is in other registers at this moment? Using tracer I’m dumping all them:
tracer.exe -l:FreeCell.exe bpx=FreeCell.exe!0x01029620
...
TID=4448|(0) FreeCell.exe!0x1029620
EAX=0x5ddb6ac4 EBX=0x00000000 ECX=0x00000000 EDX=0x00000000
ESI=0x054732d0 EDI=0x054732d0 EBP=0x0020f2bc ESP=0x0020f298
EIP=0x00899620
FLAGS=PF ZF IF
EAX=0x5ddb6ac8 EBX=0x00000002 ECX=0x00000000 EDX=0x00000000
ESI=0xffffff11 EDI=0x054732d0 EBP=0x0020da78 ESP=0x0020d9d4
EIP=0x00899620
FLAGS=PF ZF IF
EAX=0x5ddb6aca EBX=0x00000002 ECX=0x00000000 EDX=0x00000000
ESI=0x7740c460 EDI=0x054732d0 EBP=0x0020da78 ESP=0x0020d9d4
EIP=0x00899620
FLAGS=PF ZF IF
...
No matter how often I restart the game, ECX and EDX are seems to be always 0. So I patching PUSH EAX
at 0x01029620 to PUSH EDX (also one-byte instruction), and now FreeCell always shows the same game
to the player.
However, other options could exist. As a matter of fact, we don’t need to pass 0 to srand(). Rather, we
815
want to pass a constant to srand() to make game the same each time. As we can see, EDI’s value hasn’t
been changing. Maybe we could try it as well.
Now a bit harder patching. Let’s open FreeCell.exe in Hiew:
We have no space to replace one-byte PUSH EAX with two-byte PUSH 0. And we can’t simply fill CALL ds:time
with NOPs, because there is a FIXUP (address of time() function in msvcrt.dll). (Hiew marked these 4 bytes
are gray bytes.) So what I’m doing: patching first 2 bytes to EB 04. This is a JMP to bypass 4 FIXUP-ed
bytes:
Then I replace PUSH EAX with NOP. So that srand() would have its zeroes arguments from PUSH 0 above.
Also, I patch one of POP ECX to NOP, because I removed one PUSH.
816
Now Windows loader will write 4-byte FIXUP at 0x0102961C, but we don’t care: time()’s address will not
be used anymore.
8.7.2 Part II: breaking the Select Game submenu

The user can still choose different game in the menu. Let’s see if srand() is still called. I’m trying to enter
1/2/3 in ”Select Game” dialog box:
tracer.exe -l:FreeCell.exe bpf=msvcrt.dll!srand,args:1
...
TID=4936|(0) msvcrt.dll!srand(0x5ddb6df9) (called from FreeCell.exe!BASE+0x29627 (0xb49627))

TID=4936|(0) msvcrt.dll!srand(0x2b40) (called from FreeCell.exe!BASE+0x27d3a (0xb47d3a))
TID=4936|(0) msvcrt.dll!srand(0x1) (called from FreeCell.exe!BASE+0x27d3a (0xb47d3a))
Yes, the number user enters is just an argument for srand(). Where it is called?
.text:01027CBA loc_1027CBA: ; CODE XREF: sub\_1027AC6⤦
Ç +179
.text:01027CBA 83 FF FC cmp edi, 0FFFFFFFCh
.text:01027CBD 75 74 jnz short loc_1027D33
...
.text:01027D33 loc_1027D33: ; CODE XREF: sub\_1027AC6⤦

Ç +1F7
.text:01027D33 57 push edi ; Seed
.text:01027D34 FF 15 84 16 00+ call ds:srand
.text:01027D3A 59 pop ecx
.text:01027D3B 6A 34 push 34h
.text:01027D3D 5B pop ebx
.text:01027D3E 33 C0 xor eax, eax
I couldn’t patch one-byte PUSH EDI to two-byte PUSH 0. But I see that there is only one single jump to
loc_1027D33 from the above.
I’m patching CMP EDI, ... to XOR EDI, EDI, padding the 3rd byte to NOP. I’m patching also JNZ to JMP,
so that jump will always occur.
Now FreeCell ignores the number user enters, but suddenly, there is also the same game at start:
817
It seems that the code we patched in part I is somehow connected to a code after 0x01027CBD, that
executes if EDI==0xFFFFFFFC. Anyway, our goal is accomplished — the game is always the same at the
start and the user can’t choose another using the menu.
8.8 Dongles
The author of these lines, occasionally did software copy-protection dongle replacements, or “dongle
emulators” and here are couple examples of how it’s happening.
About one of the cases about Rocket and Z3 that is not present here, you can read here: http://yurichev.
com/tmp/SAT_SMT_DRAFT.pdf.
8.8.1 Example #1: MacOS Classic and PowerPC

Here is an example of a program for MacOS Classic 12 , for PowerPC. The company who developed the
software product has disappeared a long time ago, so the (legal) customer was afraid of physical dongle
damage.
While running without a dongle connected, a message box with the text ”Invalid Security Device” ap-
peared.
Luckily, this text string could easily be found in the executable binary file.
Let’s pretend we are not very familiar both with Mac OS Classic and PowerPC, but will try anyway.
IDA opened the executable file smoothly, reported its type as ”PEF (Mac OS or Be OS executable)” (indeed,
it is a standard Mac OS Classic file format).
By searching for the text string with the error message, we’ve got into this code fragment:
12 pre-UNIX MacOS
818
...
seg000:000C87FC 38 60 00 01 li %r3, 1
seg000:000C8800 48 03 93 41 bl check1
seg000:000C8804 60 00 00 00 nop
seg000:000C8808 54 60 06 3F clrlwi. %r0, %r3, 24
seg000:000C880C 40 82 00 40 bne OK
seg000:000C8810 80 62 9F D8 lwz %r3, TC_aInvalidSecurityDevice
...
Yes, this is PowerPC code.

The CPU is a very typical 32-bit RISC of 1990s era.
Each instruction occupies 4 bytes (just as in MIPS and ARM) and the names somewhat resemble MIPS
instruction names.
check1() is a function name we’ll give to it later. BL is Branch Link instruction, e.g., intended for calling
subroutines.
The crucial point is the BNE instruction which jumps if the dongle protection check passes or not if an error
occurs: then the address of the text string gets loaded into the r3 register for the subsequent passing into
a message box routine.
From the [Steve Zucker, SunSoft and Kari Karhi, IBM, SYSTEM V APPLICATION BINARY INTERFACE: PowerPC
Processor Supplement, (1995)]13 we will found out that the r3 register is used for return values (and r4, in
case of 64-bit values).
Another yet unknown instruction is CLRLWI. From [PowerPC(tm) Microprocessor Family: The Programming
Environments for 32-Bit Microprocessors, (2000)]14 we’ll learn that this instruction does both clearing and
loading. In our case, it clears the 24 high bits from the value in r3 and puts them in r0, so it is analogical
to MOVZX in x86 ( 1.23.1 on page 201), but it also sets the flags, so BNE can check them afterwards.
Let’s take a look into the check1() function:
seg000:00101B40 check1: # CODE XREF: seg000:00063E7Cp
seg000:00101B40 # sub_64070+160p ...
seg000:00101B40
seg000:00101B40 .set arg_8, 8
seg000:00101B40
seg000:00101B40 7C 08 02 A6 mflr %r0
seg000:00101B44 90 01 00 08 stw %r0, arg_8(%sp)
seg000:00101B48 94 21 FF C0 stwu %sp, -0x40(%sp)
seg000:00101B4C 48 01 6B 39 bl check2
seg000:00101B50 60 00 00 00 nop
seg000:00101B54 80 01 00 48 lwz %r0, 0x40+arg_8(%sp)
seg000:00101B58 38 21 00 40 addi %sp, %sp, 0x40
seg000:00101B5C 7C 08 03 A6 mtlr %r0
seg000:00101B60 4E 80 00 20 blr
seg000:00101B60 # End of function check1
As you can see in IDA, that function is called from many places in the program, but only the r3 register’s
value is checked after each call.
All this function does is to call the other function, so it is a thunk function: there are function prologue and
epilogue, but the r3 register is not touched, so checkl() returns what check2() returns.
BLR15 looks like the return from the function, but since IDA does the function layout, we probably do not
need to care about this.
Since it is a typical RISC, it seems that subroutines are called using a link register, just like in ARM.
The check2() function is more complex:
seg000:00118684 check2: # CODE XREF: check1+Cp
seg000:00118684
seg000:00118684 .set var_18, -0x18
13 Alsoavailable as http://yurichev.com/mirrors/PowerPC/elfspec_ppc.pdf
14 Alsoavailable as http://yurichev.com/mirrors/PowerPC/6xx_pem.pdf
15 (PowerPC) Branch to Link Register
819
seg000:00118684 .set var_C, -0xC
seg000:00118684 .set var_8, -8
seg000:00118684 .set var_4, -4
seg000:00118684 .set arg_8, 8
seg000:00118684
seg000:00118684 93 E1 FF FC stw %r31, var_4(%sp)
seg000:00118688 7C 08 02 A6 mflr %r0
seg000:0011868C 83 E2 95 A8 lwz %r31, off_1485E8 # dword_24B704
seg000:00118690 .using dword_24B704, %r31
seg000:00118690 93 C1 FF F8 stw %r30, var_8(%sp)
seg000:00118694 93 A1 FF F4 stw %r29, var_C(%sp)
seg000:00118698 7C 7D 1B 78 mr %r29, %r3
seg000:0011869C 90 01 00 08 stw %r0, arg_8(%sp)
seg000:001186A0 54 60 06 3E clrlwi %r0, %r3, 24
seg000:001186A4 28 00 00 01 cmplwi %r0, 1
seg000:001186A8 94 21 FF B0 stwu %sp, -0x50(%sp)
seg000:001186AC 40 82 00 0C bne loc_1186B8
seg000:001186B0 38 60 00 01 li %r3, 1
seg000:001186B4 48 00 00 6C b exit
seg000:001186B8
seg000:001186B8 loc_1186B8: # CODE XREF: check2+28j
seg000:001186B8 48 00 03 D5 bl sub_118A8C
seg000:001186BC 60 00 00 00 nop
seg000:001186C0 3B C0 00 00 li %r30, 0
seg000:001186C4
seg000:001186C4 skip: # CODE XREF: check2+94j
seg000:001186C4 57 C0 06 3F clrlwi. %r0, %r30, 24
seg000:001186C8 41 82 00 18 beq loc_1186E0
seg000:001186CC 38 61 00 38 addi %r3, %sp, 0x50+var_18
seg000:001186D0 80 9F 00 00 lwz %r4, dword_24B704
seg000:001186D4 48 00 C0 55 bl .RBEFINDNEXT
seg000:001186D8 60 00 00 00 nop
seg000:001186DC 48 00 00 1C b loc_1186F8
seg000:001186E0
seg000:001186E0 loc_1186E0: # CODE XREF: check2+44j
seg000:001186E0 80 BF 00 00 lwz %r5, dword_24B704
seg000:001186E4 38 81 00 38 addi %r4, %sp, 0x50+var_18
seg000:001186E8 38 60 08 C2 li %r3, 0x1234
seg000:001186EC 48 00 BF 99 bl .RBEFINDFIRST
seg000:001186F0 60 00 00 00 nop
seg000:001186F4 3B C0 00 01 li %r30, 1
seg000:001186F8
seg000:001186F8 loc_1186F8: # CODE XREF: check2+58j
seg000:001186F8 54 60 04 3F clrlwi. %r0, %r3, 16
seg000:001186FC 41 82 00 0C beq must_jump
seg000:00118700 38 60 00 00 li %r3, 0 # error
seg000:00118704 48 00 00 1C b exit
seg000:00118708
seg000:00118708 must_jump: # CODE XREF: check2+78j
seg000:00118708 7F A3 EB 78 mr %r3, %r29
seg000:0011870C 48 00 00 31 bl check3
seg000:00118710 60 00 00 00 nop
seg000:00118714 54 60 06 3F clrlwi. %r0, %r3, 24
seg000:00118718 41 82 FF AC beq skip
seg000:0011871C 38 60 00 01 li %r3, 1
seg000:00118720
seg000:00118720 exit: # CODE XREF: check2+30j
seg000:00118720 # check2+80j
seg000:00118720 80 01 00 58 lwz %r0, 0x50+arg_8(%sp)
seg000:00118724 38 21 00 50 addi %sp, %sp, 0x50
seg000:00118728 83 E1 FF FC lwz %r31, var_4(%sp)
seg000:0011872C 7C 08 03 A6 mtlr %r0
seg000:00118730 83 C1 FF F8 lwz %r30, var_8(%sp)
seg000:00118734 83 A1 FF F4 lwz %r29, var_C(%sp)
seg000:00118738 4E 80 00 20 blr
seg000:00118738 # End of function check2
We are lucky again: some function names are left in the executable (debug symbols section?
Hard to say while we are not very familiar with the file format, maybe it is some kind of PE exports?
820
(6.5.2)),
like .RBEFINDNEXT() and .RBEFINDFIRST().
Eventually these functions call other functions with names like .GetNextDeviceViaUSB(), .USBSendPKT(),
so these are clearly dealing with an USB device.
There is even a function named .GetNextEve3Device()—sounds familiar, there was a Sentinel Eve3 don-
gle for ADB port (present on Macs) in 1990s.
Let’s first take a look on how the r3 register is set before return, while ignoring everything else.
We know that a “good” r3 value has to be non-zero, zero r3 leads the execution flow to the message box
with an error message.
There are two li %r3, 1 instructions present in the function and one li %r3, 0 (Load Immediate, i.e.,
loading a value into a register). The first instruction is at 0x001186B0—and frankly speaking, it’s hard to
say what it means.
What we see next is, however, easier to understand: .RBEFINDFIRST() is called: if it fails, 0 is written into
r3 and we jump to exit, otherwise another function is called (check3())—if it fails too, .RBEFINDNEXT() is
called, probably in order to look for another USB device.
N.B.: clrlwi. %r0, %r3, 16 it is analogical to what we already saw, but it clears 16 bits, i.e.,
.RBEFINDFIRST() probably returns a 16-bit value.
B (stands for branch) unconditional jump.
BEQ is the inverse instruction of BNE.
Let’s see check3():
seg000:0011873C check3: # CODE XREF: check2+88p
seg000:0011873C
seg000:0011873C .set var_18, -0x18
seg000:0011873C .set var_C, -0xC
seg000:0011873C .set var_8, -8
seg000:0011873C .set var_4, -4
seg000:0011873C .set arg_8, 8
seg000:0011873C
seg000:0011873C 93 E1 FF FC stw %r31, var_4(%sp)
seg000:00118740 7C 08 02 A6 mflr %r0
seg000:00118744 38 A0 00 00 li %r5, 0
seg000:00118748 93 C1 FF F8 stw %r30, var_8(%sp)
seg000:0011874C 83 C2 95 A8 lwz %r30, off_1485E8 # dword_24B704
seg000:00118750 .using dword_24B704, %r30
seg000:00118750 93 A1 FF F4 stw %r29, var_C(%sp)
seg000:00118754 3B A3 00 00 addi %r29, %r3, 0
seg000:00118758 38 60 00 00 li %r3, 0
seg000:0011875C 90 01 00 08 stw %r0, arg_8(%sp)
seg000:00118760 94 21 FF B0 stwu %sp, -0x50(%sp)
seg000:00118764 80 DE 00 00 lwz %r6, dword_24B704
seg000:00118768 38 81 00 38 addi %r4, %sp, 0x50+var_18
seg000:0011876C 48 00 C0 5D bl .RBEREAD
seg000:00118770 60 00 00 00 nop
seg000:00118774 54 60 04 3F clrlwi. %r0, %r3, 16
seg000:00118778 41 82 00 0C beq loc_118784
seg000:0011877C 38 60 00 00 li %r3, 0
seg000:00118780 48 00 02 F0 b exit
seg000:00118784
seg000:00118784 loc_118784: # CODE XREF: check3+3Cj
seg000:00118784 A0 01 00 38 lhz %r0, 0x50+var_18(%sp)
seg000:00118788 28 00 04 B2 cmplwi %r0, 0x1100
seg000:0011878C 41 82 00 0C beq loc_118798
seg000:00118790 38 60 00 00 li %r3, 0
seg000:00118794 48 00 02 DC b exit
seg000:00118798
seg000:00118798 loc_118798: # CODE XREF: check3+50j
seg000:0011879C 38 81 00 38 addi %r4, %sp, 0x50+var_18
seg000:001187A0 38 60 00 01 li %r3, 1
seg000:001187A4 38 A0 00 00 li %r5, 0
seg000:001187A8 48 00 C0 21 bl .RBEREAD
seg000:001187AC 60 00 00 00 nop
821
seg000:001187B0 54 60 04 3F clrlwi. %r0, %r3, 16
seg000:001187B4 41 82 00 0C beq loc_1187C0
seg000:001187B8 38 60 00 00 li %r3, 0
seg000:001187BC 48 00 02 B4 b exit
seg000:001187C0
seg000:001187C0 loc_1187C0: # CODE XREF: check3+78j
seg000:001187C0 A0 01 00 38 lhz %r0, 0x50+var_18(%sp)
seg000:001187C4 28 00 06 4B cmplwi %r0, 0x09AB
seg000:001187C8 41 82 00 0C beq loc_1187D4
seg000:001187CC 38 60 00 00 li %r3, 0
seg000:001187D0 48 00 02 A0 b exit
seg000:001187D4
seg000:001187D4 loc_1187D4: # CODE XREF: check3+8Cj
seg000:001187D4 4B F9 F3 D9 bl sub_B7BAC
seg000:001187D8 60 00 00 00 nop
seg000:001187DC 54 60 06 3E clrlwi %r0, %r3, 24
seg000:001187E0 2C 00 00 05 cmpwi %r0, 5
seg000:001187E4 41 82 01 00 beq loc_1188E4
seg000:001187E8 40 80 00 10 bge loc_1187F8
seg000:001187EC 2C 00 00 04 cmpwi %r0, 4
seg000:001187F0 40 80 00 58 bge loc_118848
seg000:001187F4 48 00 01 8C b loc_118980
seg000:001187F8
seg000:001187F8 loc_1187F8: # CODE XREF: check3+ACj
seg000:001187F8 2C 00 00 0B cmpwi %r0, 0xB
seg000:001187FC 41 82 00 08 beq loc_118804
seg000:00118800 48 00 01 80 b loc_118980
seg000:00118804
seg000:00118804 loc_118804: # CODE XREF: check3+C0j
seg000:00118808 38 81 00 38 addi %r4, %sp, 0x50+var_18
seg000:0011880C 38 60 00 08 li %r3, 8
seg000:00118810 38 A0 00 00 li %r5, 0
seg000:00118814 48 00 BF B5 bl .RBEREAD
seg000:00118818 60 00 00 00 nop
seg000:0011881C 54 60 04 3F clrlwi. %r0, %r3, 16
seg000:00118820 41 82 00 0C beq loc_11882C
seg000:00118824 38 60 00 00 li %r3, 0
seg000:00118828 48 00 02 48 b exit
seg000:0011882C
seg000:0011882C loc_11882C: # CODE XREF: check3+E4j
seg000:0011882C A0 01 00 38 lhz %r0, 0x50+var_18(%sp)
seg000:00118830 28 00 11 30 cmplwi %r0, 0xFEA0
seg000:00118834 41 82 00 0C beq loc_118840
seg000:00118838 38 60 00 00 li %r3, 0
seg000:0011883C 48 00 02 34 b exit
seg000:00118840
seg000:00118840 loc_118840: # CODE XREF: check3+F8j
seg000:00118840 38 60 00 01 li %r3, 1
seg000:00118844 48 00 02 2C b exit
seg000:00118848
seg000:00118848 loc_118848: # CODE XREF: check3+B4j
seg000:00118850 38 60 00 0A li %r3, 0xA
seg000:00118854 38 A0 00 00 li %r5, 0
seg000:00118858 48 00 BF 71 bl .RBEREAD
seg000:0011885C 60 00 00 00 nop
seg000:00118860 54 60 04 3F clrlwi. %r0, %r3, 16
seg000:00118864 41 82 00 0C beq loc_118870
seg000:00118868 38 60 00 00 li %r3, 0
seg000:0011886C 48 00 02 04 b exit
seg000:00118870
seg000:00118870 A0 01 00 38 lhz %r0, 0x50+var_18(%sp)
seg000:00118874 28 00 03 F3 cmplwi %r0, 0xA6E1
seg000:00118878 41 82 00 0C beq loc_118884
seg000:0011887C 38 60 00 00 li %r3, 0
seg000:00118880 48 00 01 F0 b exit
seg000:00118884
822
seg000:00118884 loc_118884: # CODE XREF: check3+13Cj
seg000:00118884 57 BF 06 3E clrlwi %r31, %r29, 24
seg000:00118888 28 1F 00 02 cmplwi %r31, 2
seg000:0011888C 40 82 00 0C bne loc_118898
seg000:00118890 38 60 00 01 li %r3, 1
seg000:00118894 48 00 01 DC b exit
seg000:00118898
seg000:001188A0 38 60 00 0B li %r3, 0xB
seg000:001188A4 38 A0 00 00 li %r5, 0
seg000:001188A8 48 00 BF 21 bl .RBEREAD
seg000:001188AC 60 00 00 00 nop
seg000:001188B0 54 60 04 3F clrlwi. %r0, %r3, 16
seg000:001188B4 41 82 00 0C beq loc_1188C0
seg000:001188B8 38 60 00 00 li %r3, 0
seg000:001188BC 48 00 01 B4 b exit
seg000:001188C0
seg000:001188C0 A0 01 00 38 lhz %r0, 0x50+var_18(%sp)
seg000:001188C4 28 00 23 1C cmplwi %r0, 0x1C20
seg000:001188C8 41 82 00 0C beq loc_1188D4
seg000:001188CC 38 60 00 00 li %r3, 0
seg000:001188D0 48 00 01 A0 b exit
seg000:001188D4
seg000:001188D4 loc_1188D4: # CODE XREF: check3+18Cj
seg000:001188D4 28 1F 00 03 cmplwi %r31, 3
seg000:001188D8 40 82 01 94 bne error
seg000:001188DC 38 60 00 01 li %r3, 1
seg000:001188E0 48 00 01 90 b exit
seg000:001188E4
seg000:001188E4 loc_1188E4: # CODE XREF: check3+A8j
seg000:001188E4 80 DE 00 00 lwz %r6, dword_24B704
seg000:001188E8 38 81 00 38 addi %r4, %sp, 0x50+var_18
seg000:001188EC 38 60 00 0C li %r3, 0xC
seg000:001188F0 38 A0 00 00 li %r5, 0
seg000:001188F4 48 00 BE D5 bl .RBEREAD
seg000:001188F8 60 00 00 00 nop
seg000:001188FC 54 60 04 3F clrlwi. %r0, %r3, 16
seg000:00118900 41 82 00 0C beq loc_11890C
seg000:00118904 38 60 00 00 li %r3, 0
seg000:00118908 48 00 01 68 b exit
seg000:0011890C
seg000:0011890C loc_11890C: # CODE XREF: check3+1C4j
seg000:00118910 28 00 1F 40 cmplwi %r0, 0x40FF
seg000:00118914 41 82 00 0C beq loc_118920
seg000:00118918 38 60 00 00 li %r3, 0
seg000:0011891C 48 00 01 54 b exit
seg000:00118920
seg000:00118920 loc_118920: # CODE XREF: check3+1D8j
seg000:00118920 57 BF 06 3E clrlwi %r31, %r29, 24
seg000:00118924 28 1F 00 02 cmplwi %r31, 2
seg000:00118928 40 82 00 0C bne loc_118934
seg000:0011892C 38 60 00 01 li %r3, 1
seg000:00118930 48 00 01 40 b exit
seg000:00118934
seg000:00118934 loc_118934: # CODE XREF: check3+1ECj
seg000:00118938 38 81 00 38 addi %r4, %sp, 0x50+var_18
seg000:0011893C 38 60 00 0D li %r3, 0xD
seg000:00118940 38 A0 00 00 li %r5, 0
seg000:00118944 48 00 BE 85 bl .RBEREAD
seg000:00118948 60 00 00 00 nop
seg000:0011894C 54 60 04 3F clrlwi. %r0, %r3, 16
seg000:00118950 41 82 00 0C beq loc_11895C
seg000:00118954 38 60 00 00 li %r3, 0
seg000:00118958 48 00 01 18 b exit
seg000:0011895C
823
seg000:0011895C loc_11895C: # CODE XREF: check3+214j
seg000:00118960 28 00 07 CF cmplwi %r0, 0xFC7
seg000:00118964 41 82 00 0C beq loc_118970
seg000:00118968 38 60 00 00 li %r3, 0
seg000:0011896C 48 00 01 04 b exit
seg000:00118970
seg000:00118970 28 1F 00 03 cmplwi %r31, 3
seg000:00118974 40 82 00 F8 bne error
seg000:00118978 38 60 00 01 li %r3, 1
seg000:0011897C 48 00 00 F4 b exit
seg000:00118980
seg000:00118980 loc_118980: # CODE XREF: check3+B8j
seg000:00118980 # check3+C4j
seg000:00118984 38 81 00 38 addi %r4, %sp, 0x50+var_18
seg000:00118988 3B E0 00 00 li %r31, 0
seg000:0011898C 38 60 00 04 li %r3, 4
seg000:00118990 38 A0 00 00 li %r5, 0
seg000:00118994 48 00 BE 35 bl .RBEREAD
seg000:00118998 60 00 00 00 nop
seg000:0011899C 54 60 04 3F clrlwi. %r0, %r3, 16
seg000:001189A0 41 82 00 0C beq loc_1189AC
seg000:001189A4 38 60 00 00 li %r3, 0
seg000:001189A8 48 00 00 C8 b exit
seg000:001189AC
seg000:001189AC loc_1189AC: # CODE XREF: check3+264j
seg000:001189AC A0 01 00 38 lhz %r0, 0x50+var_18(%sp)
seg000:001189B0 28 00 1D 6A cmplwi %r0, 0xAED0
seg000:001189B4 40 82 00 0C bne loc_1189C0
seg000:001189B8 3B E0 00 01 li %r31, 1
seg000:001189BC 48 00 00 14 b loc_1189D0
seg000:001189C0
seg000:001189C0 28 00 18 28 cmplwi %r0, 0x2818
seg000:001189C4 41 82 00 0C beq loc_1189D0
seg000:001189C8 38 60 00 00 li %r3, 0
seg000:001189CC 48 00 00 A4 b exit
seg000:001189D0
seg000:001189D0 loc_1189D0: # CODE XREF: check3+280j
seg000:001189D0 # check3+288j
seg000:001189D0 57 A0 06 3E clrlwi %r0, %r29, 24
seg000:001189D4 28 00 00 02 cmplwi %r0, 2
seg000:001189D8 40 82 00 20 bne loc_1189F8
seg000:001189DC 57 E0 06 3F clrlwi. %r0, %r31, 24
seg000:001189E0 41 82 00 10 beq good2
seg000:001189E4 48 00 4C 69 bl sub_11D64C
seg000:001189E8 60 00 00 00 nop
seg000:001189EC 48 00 00 84 b exit
seg000:001189F0
seg000:001189F0 good2: # CODE XREF: check3+2A4j
seg000:001189F0 38 60 00 01 li %r3, 1
seg000:001189F4 48 00 00 7C b exit
seg000:001189F8
seg000:001189F8 loc_1189F8: # CODE XREF: check3+29Cj
seg000:001189F8 80 DE 00 00 lwz %r6, dword_24B704
seg000:001189FC 38 81 00 38 addi %r4, %sp, 0x50+var_18
seg000:00118A00 38 60 00 05 li %r3, 5
seg000:00118A04 38 A0 00 00 li %r5, 0
seg000:00118A08 48 00 BD C1 bl .RBEREAD
seg000:00118A0C 60 00 00 00 nop
seg000:00118A10 54 60 04 3F clrlwi. %r0, %r3, 16
seg000:00118A14 41 82 00 0C beq loc_118A20
seg000:00118A18 38 60 00 00 li %r3, 0
seg000:00118A1C 48 00 00 54 b exit
seg000:00118A20
seg000:00118A20 loc_118A20: # CODE XREF: check3+2D8j
seg000:00118A20 A0 01 00 38 lhz %r0, 0x50+var_18(%sp)
seg000:00118A24 28 00 11 D3 cmplwi %r0, 0xD300
824
seg000:00118A28 40 82 00 0C bne loc_118A34
seg000:00118A2C 3B E0 00 01 li %r31, 1
seg000:00118A30 48 00 00 14 b good1
seg000:00118A34
seg000:00118A34 loc_118A34: # CODE XREF: check3+2ECj
seg000:00118A34 28 00 1A EB cmplwi %r0, 0xEBA1
seg000:00118A38 41 82 00 0C beq good1
seg000:00118A3C 38 60 00 00 li %r3, 0
seg000:00118A40 48 00 00 30 b exit
seg000:00118A44
seg000:00118A44 good1: # CODE XREF: check3+2F4j
seg000:00118A44 # check3+2FCj
seg000:00118A44 57 A0 06 3E clrlwi %r0, %r29, 24
seg000:00118A48 28 00 00 03 cmplwi %r0, 3
seg000:00118A4C 40 82 00 20 bne error
seg000:00118A50 57 E0 06 3F clrlwi. %r0, %r31, 24
seg000:00118A54 41 82 00 10 beq good
seg000:00118A58 48 00 4B F5 bl sub_11D64C
seg000:00118A5C 60 00 00 00 nop
seg000:00118A60 48 00 00 10 b exit
seg000:00118A64
seg000:00118A64 good: # CODE XREF: check3+318j
seg000:00118A64 38 60 00 01 li %r3, 1
seg000:00118A68 48 00 00 08 b exit
seg000:00118A6C
seg000:00118A6C error: # CODE XREF: check3+19Cj
seg000:00118A6C # check3+238j ...
seg000:00118A6C 38 60 00 00 li %r3, 0
seg000:00118A70
seg000:00118A70 exit: # CODE XREF: check3+44j
seg000:00118A70 # check3+58j ...
seg000:00118A70 80 01 00 58 lwz %r0, 0x50+arg_8(%sp)
seg000:00118A74 38 21 00 50 addi %sp, %sp, 0x50
seg000:00118A78 83 E1 FF FC lwz %r31, var_4(%sp)
seg000:00118A7C 7C 08 03 A6 mtlr %r0
seg000:00118A80 83 C1 FF F8 lwz %r30, var_8(%sp)
seg000:00118A84 83 A1 FF F4 lwz %r29, var_C(%sp)
seg000:00118A88 4E 80 00 20 blr
seg000:00118A88 # End of function check3
There are a lot of calls to .RBEREAD().

Perhaps, the function returns some values from the dongle, so they are compared here with some hard-
coded variables using CMPLWI.
We also see that the r3 register is also filled before each call to .RBEREAD() with one of these values: 0,
1, 8, 0xA, 0xB, 0xC, 0xD, 4, 5. Probably a memory address or something like that?
Yes, indeed, by googling these function names it is easy to find the Sentinel Eve3 dongle manual!
Perhaps we don’t even have to learn any other PowerPC instructions: all this function does is just call
.RBEREAD(), compare its results with the constants and returns 1 if the comparisons are fine or 0 other-
wise.
OK, all we’ve got is that check1() has always to return 1 or any other non-zero value.
But since we are not very confident in our knowledge of PowerPC instructions, we are going to be careful:
we will patch the jumps in check2() at 0x001186FC and 0x00118718.
At 0x001186FC we’ll write bytes 0x48 and 0 thus converting the BEQ instruction in an B (unconditional
jump): we can spot its opcode in the code without even referring to [PowerPC(tm) Microprocessor Family:
The Programming Environments for 32-Bit Microprocessors, (2000)]16 .
At 0x00118718 we’ll write 0x60 and 3 zero bytes, thus converting it to a NOP instruction: Its opcode we
could spot in the code too.
And now it all works without a dongle connected.
In summary, such small modifications can be done with IDA and minimal assembly language knowledge.
16 Also available as http://yurichev.com/mirrors/PowerPC/6xx_pem.pdf
825
8.8.2 Example #2: SCO OpenServer
An ancient software for SCO OpenServer from 1997 developed by a company that disappeared a long
time ago.
There is a special dongle driver to be installed in the system, that contains the following text strings:
“Copyright 1989, Rainbow Technologies, Inc., Irvine, CA” and “Sentinel Integrated Driver Ver. 3.0 ”.
After the installation of the driver in SCO OpenServer, these device files appear in the /dev filesystem:
/dev/rbsl8
/dev/rbsl9
/dev/rbsl10
The program reports an error without dongle connected, but the error string cannot be found in the exe-
cutables.
Thanks to IDA, it is easy to load the COFF executable used in SCO OpenServer.
Let’s also try to find “rbsl” string and indeed, found it in this code fragment:
.text:00022AB8 public SSQC
.text:00022AB8 SSQC proc near ; CODE XREF: SSQ+7p
.text:00022AB8
.text:00022AB8 var_44 = byte ptr -44h
.text:00022AB8 var_29 = byte ptr -29h
.text:00022AB8 arg_0 = dword ptr 8
.text:00022AB8
.text:00022AB8 push ebp
.text:00022AB9 mov ebp, esp
.text:00022ABB sub esp, 44h
.text:00022ABE push edi
.text:00022ABF mov edi, offset unk_4035D0
.text:00022AC4 push esi
.text:00022AC5 mov esi, [ebp+arg_0]
.text:00022AC8 push ebx
.text:00022AC9 push esi
.text:00022ACA call strlen
.text:00022ACF add esp, 4
.text:00022AD2 cmp eax, 2
.text:00022AD7 jnz loc_22BA4
.text:00022ADD inc esi
.text:00022ADE mov al, [esi-1]
.text:00022AE1 movsx eax, al
.text:00022AE4 cmp eax, '3'
.text:00022AE9 jz loc_22B84
.text:00022AEF cmp eax, '4'
.text:00022AF4 jz loc_22B94
.text:00022AFA cmp eax, '5'
.text:00022AFF jnz short loc_22B6B
.text:00022B01 movsx ebx, byte ptr [esi]
.text:00022B04 sub ebx, '0'
.text:00022B07 mov eax, 7
.text:00022B0C add eax, ebx
.text:00022B0E push eax
.text:00022B0F lea eax, [ebp+var_44]
.text:00022B12 push offset aDevSlD ; "/dev/sl%d"
.text:00022B17 push eax
.text:00022B18 call nl_sprintf
.text:00022B1D push 0 ; int
.text:00022B1F push offset aDevRbsl8 ; char *
.text:00022B24 call _access
.text:00022B29 add esp, 14h
.text:00022B2C cmp eax, 0FFFFFFFFh
.text:00022B31 jz short loc_22B48
.text:00022B33 lea eax, [ebx+7]
.text:00022B36 push eax
.text:00022B37 lea eax, [ebp+var_44]
.text:00022B3A push offset aDevRbslD ; "/dev/rbsl%d"
.text:00022B3F push eax
.text:00022B40 call nl_sprintf
.text:00022B45 add esp, 0Ch
826
.text:00022B48
.text:00022B48 loc_22B48: ; CODE XREF: SSQC+79j
.text:00022B48 mov edx, [edi]
.text:00022B4A test edx, edx
.text:00022B4C jle short loc_22B57
.text:00022B4E push edx ; int
.text:00022B4F call _close
.text:00022B57
.text:00022B57 push 2 ; int
.text:00022B59 lea eax, [ebp+var_44]
.text:00022B5C push eax ; char *
.text:00022B5D call _open
.text:00022B65 test eax, eax
.text:00022B67 mov [edi], eax
.text:00022B69 jge short loc_22B78
.text:00022B6B
.text:00022B6B loc_22B6B: ; CODE XREF: SSQC+47j
.text:00022B6B mov eax, 0FFFFFFFFh
.text:00022B70 pop ebx
.text:00022B71 pop esi
.text:00022B72 pop edi
.text:00022B73 mov esp, ebp
.text:00022B75 pop ebp
.text:00022B76 retn
.text:00022B78
.text:00022B78 loc_22B78: ; CODE XREF: SSQC+B1j
.text:00022B7A pop edi
.text:00022B7B xor eax, eax
.text:00022B7D mov esp, ebp
.text:00022B7F pop ebp
.text:00022B80 retn
.text:00022B84
.text:00022B84 mov al, [esi]
.text:00022B89 mov ds:byte_407224, al
.text:00022B8E mov esp, ebp
.text:00022B90 xor eax, eax
.text:00022B92 pop ebp
.text:00022B93 retn
.text:00022B94
.text:00022B94 loc_22B94: ; CODE XREF: SSQC+3Cj
.text:00022B94 mov al, [esi]
.text:00022B99 mov ds:byte_407225, al
.text:00022B9E mov esp, ebp
.text:00022BA0 xor eax, eax
.text:00022BA2 pop ebp
.text:00022BA3 retn
.text:00022BA4
.text:00022BA4 loc_22BA4: ; CODE XREF: SSQC+1Fj
.text:00022BA4 movsx eax, ds:byte_407225
.text:00022BAB push esi
.text:00022BAC push eax
.text:00022BAD movsx eax, ds:byte_407224
.text:00022BB4 push eax
.text:00022BB5 lea eax, [ebp+var_44]
.text:00022BB8 push offset a46CCS ; "46%c%c%s"
.text:00022BBD push eax
.text:00022BBE call nl_sprintf
.text:00022BC3 lea eax, [ebp+var_44]
827
.text:00022BC6 push eax
.text:00022BC7 call strlen
.text:00022BCC add esp, 18h
.text:00022BCF cmp eax, 1Bh
.text:00022BD4 jle short loc_22BDA
.text:00022BD6 mov [ebp+var_29], 0
.text:00022BDA
.text:00022BDA loc_22BDA: ; CODE XREF: SSQC+11Cj
.text:00022BDA lea eax, [ebp+var_44]
.text:00022BDD push eax
.text:00022BDE call strlen
.text:00022BE3 push eax ; unsigned int
.text:00022BE4 lea eax, [ebp+var_44]
.text:00022BE7 push eax ; void *
.text:00022BE8 mov eax, [edi]
.text:00022BEA push eax ; int
.text:00022BEB call _write
.text:00022BF0 add esp, 10h
.text:00022BF3 pop ebx
.text:00022BF4 pop esi
.text:00022BF5 pop edi
.text:00022BF6 mov esp, ebp
.text:00022BF8 pop ebp
.text:00022BF9 retn
.text:00022BFA db 0Eh dup(90h)
.text:00022BFA SSQC endp
Yes, indeed, the program needs to communicate with the driver somehow.
The only place where the SSQC() function is called is the thunk function:
.text:0000DBE8 public SSQ
.text:0000DBE8 SSQ proc near ; CODE XREF: sys_info+A9p
.text:0000DBE8 ; sys_info+CBp ...
.text:0000DBE8
.text:0000DBE8 arg_0 = dword ptr 8
.text:0000DBE8
.text:0000DBE8 push ebp
.text:0000DBE9 mov ebp, esp
.text:0000DBEB mov edx, [ebp+arg_0]
.text:0000DBEE push edx
.text:0000DBEF call SSQC
.text:0000DBF4 add esp, 4
.text:0000DBF7 mov esp, ebp
.text:0000DBF9 pop ebp
.text:0000DBFA retn
.text:0000DBFB SSQ endp
SSQ() can be called from at least 2 functions.

One of these is:
.data:0040169C _51_52_53 dd offset aPressAnyKeyT_0 ; DATA XREF: init_sys+392r
.data:0040169C ; sys_info+A1r
.data:0040169C ; "PRESS ANY KEY TO CONTINUE: "
.data:004016A0 dd offset a51 ; "51"
...
.data:004016B8 _3C_or_3E dd offset a3c ; DATA XREF: sys_info:loc_D67Br

.data:004016B8 ; "3C"
.data:004016BC dd offset a3e ; "3E"
; these names we gave to the labels:

.data:004016C0 answers1 dd 6B05h ; DATA XREF: sys_info+E7r
.data:004016C4 dd 3D87h
.data:004016C8 answers2 dd 3Ch ; DATA XREF: sys_info+F2r
.data:004016CC dd 832h
.data:004016D0 _C_and_B db 0Ch ; DATA XREF: sys_info+BAr
828
.data:004016D0 ; sys_info:OKr
.data:004016D1 byte_4016D1 db 0Bh ; DATA XREF: sys_info+FDr
.data:004016D2 db 0
...
.text:0000D652 xor eax, eax

.text:0000D654 mov al, ds:ctl_port
.text:0000D659 mov ecx, _51_52_53[eax*4]
.text:0000D660 push ecx
.text:0000D661 call SSQ
.text:0000D666 add esp, 4
.text:0000D669 cmp eax, 0FFFFFFFFh
.text:0000D66E jz short loc_D6D1
.text:0000D670 xor ebx, ebx
.text:0000D672 mov al, _C_and_B
.text:0000D677 test al, al
.text:0000D679 jz short loc_D6C0
.text:0000D67B
.text:0000D67B loc_D67B: ; CODE XREF: sys_info+106j
.text:0000D67B mov eax, _3C_or_3E[ebx*4]
.text:0000D682 push eax
.text:0000D688 push offset a4g ; "4G"
.text:0000D68D call SSQ
.text:0000D692 push offset a0123456789 ; "0123456789"
.text:0000D69C add esp, 0Ch
.text:0000D69F mov edx, answers1[ebx*4]
.text:0000D6A6 cmp eax, edx
.text:0000D6A8 jz short OK
.text:0000D6AA mov ecx, answers2[ebx*4]
.text:0000D6B1 cmp eax, ecx
.text:0000D6B3 jz short OK
.text:0000D6B5 mov al, byte_4016D1[ebx]
.text:0000D6BB inc ebx
.text:0000D6BC test al, al
.text:0000D6BE jnz short loc_D67B
.text:0000D6C0
.text:0000D6C0 loc_D6C0: ; CODE XREF: sys_info+C1j
.text:0000D6C0 inc ds:ctl_port
.text:0000D6C6 xor eax, eax
.text:0000D6C8 mov al, ds:ctl_port
.text:0000D6CD cmp eax, edi
.text:0000D6CF jle short loc_D652
.text:0000D6D1
.text:0000D6D1 loc_D6D1: ; CODE XREF: sys_info+98j
.text:0000D6D1 ; sys_info+B6j
.text:0000D6D1 mov edx, [ebp+var_8]
.text:0000D6D4 inc edx
.text:0000D6D5 mov [ebp+var_8], edx
.text:0000D6D8 cmp edx, 3
.text:0000D6DB jle loc_D641
.text:0000D6E1
.text:0000D6E1 loc_D6E1: ; CODE XREF: sys_info+16j
.text:0000D6E1 ; sys_info+51j ...
.text:0000D6E1 pop ebx
.text:0000D6E2 pop edi
.text:0000D6E3 mov esp, ebp
.text:0000D6E5 pop ebp
.text:0000D6E6 retn
.text:0000D6E8 OK: ; CODE XREF: sys_info+F0j
.text:0000D6E8 ; sys_info+FBj
.text:0000D6E8 mov al, _C_and_B[ebx]
.text:0000D6EE pop ebx
.text:0000D6EF pop edi
.text:0000D6F0 mov ds:ctl_model, al
.text:0000D6F5 mov esp, ebp
.text:0000D6F7 pop ebp
.text:0000D6F8 retn
829
.text:0000D6F8 sys_info endp
“3C” and “3E” sound familiar: there was a Sentinel Pro dongle by Rainbow with no memory, providing only
one crypto-hashing secret function.
You can read a short description of what hash function is here: 2.11 on page 462.
But let’s get back to the program.
So the program can only check the presence or absence of a connected dongle.
No other information can be written to such dongle, as it has no memory. The two-character codes are
commands (we can see how the commands are handled in the SSQC() function) and all other strings are
hashed inside the dongle, being transformed into a 16-bit number. The algorithm was secret, so it was not
possible to write a driver replacement or to remake the dongle hardware that would emulate it perfectly.
However, it is always possible to intercept all accesses to it and to find what constants the hash function
results are compared to.
But we need to say that it is possible to build a robust software copy protection scheme based on secret
cryptographic hash-function: let it encrypt/decrypt the data files your software uses.
But let’s get back to the code.
Codes 51/52/53 are used for LPT printer port selection. 3x/4x are used for “family” selection (that’s how
Sentinel Pro dongles are differentiated from each other: more than one dongle can be connected to a LPT
port).
The only non-2-character string passed to the hashing function is ”0123456789”.
Then, the result is compared against the set of valid results.
If it is correct, 0xC or 0xB is to be written into the global variable ctl_model.
Another text string that gets passed is ”PRESS ANY KEY TO CONTINUE: ”, but the result is not checked.
Hard to say why, probably by mistake 17 .
Let’s see where the value from the global variable ctl_model is used.
One such place is:
.text:0000D708 prep_sys proc near ; CODE XREF: init_sys+46Ap
.text:0000D708
.text:0000D708 var_14 = dword ptr -14h
.text:0000D708 var_10 = byte ptr -10h
.text:0000D708 var_8 = dword ptr -8
.text:0000D708 var_2 = word ptr -2
.text:0000D708
.text:0000D708 push ebp
.text:0000D709 mov eax, ds:net_env
.text:0000D70E mov ebp, esp
.text:0000D710 sub esp, 1Ch
.text:0000D713 test eax, eax
.text:0000D715 jnz short loc_D734
.text:0000D717 mov al, ds:ctl_model
.text:0000D71C test al, al
.text:0000D71E jnz short loc_D77E
.text:0000D720 mov [ebp+var_8], offset aIeCvulnvvOkgT_ ; "Ie-cvulnvV\\\bOKG]T_"
.text:0000D727 mov edx, 7
.text:0000D72C jmp loc_D7E7
...
.text:0000D7E7 loc_D7E7: ; CODE XREF: prep_sys+24j

.text:0000D7E7 ; prep_sys+33j
.text:0000D7E7 push edx
.text:0000D7E8 mov edx, [ebp+var_8]
.text:0000D7EB push 20h
.text:0000D7ED push edx
.text:0000D7EE push 16h
.text:0000D7F0 call err_warn
.text:0000D7F5 push offset station_sem
17 What a strange feeling: to find bugs in such ancient software.
830
.text:0000D7FA call ClosSem
.text:0000D7FF call startup_err
If it is 0, an encrypted error message is passed to a decryption routine and printed.

The error string decryption routine seems a simple xoring:
.text:0000A43C err_warn proc near ; CODE XREF: prep_sys+E8p
.text:0000A43C ; prep_sys2+2Fp ...
.text:0000A43C
.text:0000A43C var_55 = byte ptr -55h
.text:0000A43C var_54 = byte ptr -54h
.text:0000A43C arg_0 = dword ptr 8
.text:0000A43C arg_4 = dword ptr 0Ch
.text:0000A43C arg_8 = dword ptr 10h
.text:0000A43C arg_C = dword ptr 14h
.text:0000A43C
.text:0000A43C push ebp
.text:0000A43D mov ebp, esp
.text:0000A43F sub esp, 54h
.text:0000A442 push edi
.text:0000A443 mov ecx, [ebp+arg_8]
.text:0000A446 xor edi, edi
.text:0000A448 test ecx, ecx
.text:0000A44A push esi
.text:0000A44B jle short loc_A466
.text:0000A44D mov esi, [ebp+arg_C] ; key
.text:0000A450 mov edx, [ebp+arg_4] ; string
.text:0000A453
.text:0000A453 loc_A453: ; CODE XREF: err_warn+28j
.text:0000A453 xor eax, eax
.text:0000A455 mov al, [edx+edi]
.text:0000A458 xor eax, esi
.text:0000A45A add esi, 3
.text:0000A45D inc edi
.text:0000A45E cmp edi, ecx
.text:0000A460 mov [ebp+edi+var_55], al
.text:0000A464 jl short loc_A453
.text:0000A466
.text:0000A466 loc_A466: ; CODE XREF: err_warn+Fj
.text:0000A466 mov [ebp+edi+var_54], 0
.text:0000A46B mov eax, [ebp+arg_0]
.text:0000A46E cmp eax, 18h
.text:0000A473 jnz short loc_A49C
.text:0000A475 lea eax, [ebp+var_54]
.text:0000A478 push eax
.text:0000A479 call status_line
.text:0000A47E add esp, 4
.text:0000A481
.text:0000A481 loc_A481: ; CODE XREF: err_warn+72j
.text:0000A481 push 50h
.text:0000A483 push 0
.text:0000A485 lea eax, [ebp+var_54]
.text:0000A489 call memset
.text:0000A48E call pcv_refresh
.text:0000A493 add esp, 0Ch
.text:0000A496 pop esi
.text:0000A497 pop edi
.text:0000A498 mov esp, ebp
.text:0000A49A pop ebp
.text:0000A49B retn
.text:0000A49C
.text:0000A49C loc_A49C: ; CODE XREF: err_warn+37j
.text:0000A49C push 0
.text:0000A49E lea eax, [ebp+var_54]
.text:0000A4A1 mov edx, [ebp+arg_0]
.text:0000A4A4 push edx
.text:0000A4A5 push eax
.text:0000A4A6 call pcv_lputs
831
.text:0000A4AB add esp, 0Ch
.text:0000A4AE jmp short loc_A481
.text:0000A4AE err_warn endp
That’s why we were unable to find the error messages in the executable files, because they are encrypted
(which is is popular practice).
Another call to the SSQ() hashing function passes the “offln” string to it and compares the result with
0xFE81 and 0x12A9.
If they don’t match, it works with some timer() function (maybe waiting for a poorly connected dongle
to be reconnected and check again?) and then decrypts another error message to dump.
.text:0000DA55 loc_DA55: ; CODE XREF: sync_sys+24Cj
.text:0000DA55 push offset aOffln ; "offln"
.text:0000DA5A call SSQ
.text:0000DA5F add esp, 4
.text:0000DA62 mov dl, [ebx]
.text:0000DA64 mov esi, eax
.text:0000DA66 cmp dl, 0Bh
.text:0000DA69 jnz short loc_DA83
.text:0000DA6B cmp esi, 0FE81h
.text:0000DA71 jz OK
.text:0000DA77 cmp esi, 0FFFFF8EFh
.text:0000DA7D jz OK
.text:0000DA83
.text:0000DA83 loc_DA83: ; CODE XREF: sync_sys+201j
.text:0000DA83 mov cl, [ebx]
.text:0000DA85 cmp cl, 0Ch
.text:0000DA88 jnz short loc_DA9F
.text:0000DA8A cmp esi, 12A9h
.text:0000DA96 cmp esi, 0FFFFFFF5h
.text:0000DA9F
.text:0000DA9F loc_DA9F: ; CODE XREF: sync_sys+220j
.text:0000DA9F mov eax, [ebp+var_18]
.text:0000DAA2 test eax, eax
.text:0000DAA4 jz short loc_DAB0
.text:0000DAA6 push 24h
.text:0000DAA8 call timer
.text:0000DAAD add esp, 4
.text:0000DAB0
.text:0000DAB0 loc_DAB0: ; CODE XREF: sync_sys+23Cj
.text:0000DAB0 inc edi
.text:0000DAB1 cmp edi, 3
.text:0000DAB4 jle short loc_DA55
.text:0000DAB6 mov eax, ds:net_env
.text:0000DABB test eax, eax
.text:0000DABD jz short error
...
.text:0000DAF7 error: ; CODE XREF: sync_sys+255j

.text:0000DAF7 ; sync_sys+274j ...
.text:0000DAF7 mov [ebp+var_8], offset encrypted_error_message2
.text:0000DAFE mov [ebp+var_C], 17h ; decrypting key
.text:0000DB05 jmp decrypt_end_print_message
...
; this name we gave to label:

.text:0000D9B6 decrypt_end_print_message: ; CODE XREF: sync_sys+29Dj
.text:0000D9B6 ; sync_sys+2ABj
.text:0000D9B6 mov eax, [ebp+var_18]
.text:0000D9B9 test eax, eax
.text:0000D9BB jnz short loc_D9FB
.text:0000D9BD mov edx, [ebp+var_C] ; key
.text:0000D9C0 mov ecx, [ebp+var_8] ; string
.text:0000D9C3 push edx
832
.text:0000D9C4 push 20h
.text:0000D9C6 push ecx
.text:0000D9C9 call err_warn
.text:0000D9CE push 0Fh
.text:0000D9D0 push 190h
.text:0000D9D5 call sound
.text:0000D9DA mov [ebp+var_18], 1
.text:0000D9E1 add esp, 18h
.text:0000D9E4 call pcv_kbhit
.text:0000D9E9 test eax, eax
.text:0000D9EB jz short loc_D9FB
...
; this name we gave to label:

.data:00401736 encrypted_error_message2 db 74h, 72h, 78h, 43h, 48h, 6, 5Ah, 49h, 4Ch, 2 dup(47h⤦
Ç )
.data:00401736 db 51h, 4Fh, 47h, 61h, 20h, 22h, 3Ch, 24h, 33h, 36h, 76h
.data:00401736 db 3Ah, 33h, 31h, 0Ch, 0, 0Bh, 1Fh, 7, 1Eh, 1Ah
Bypassing the dongle is pretty straightforward: just patch all jumps after the relevant CMP instructions.
Another option is to write our own SCO OpenServer driver, containing a table of questions and answers,
all of those which present in the program.
Decrypting error messages
By the way, we can also try to decrypt all error messages. The algorithm that is located in the err_warn()
function is very simple, indeed:
Listing 8.5: Decryption function

.text:0000A44D mov esi, [ebp+arg_C] ; key
.text:0000A450 mov edx, [ebp+arg_4] ; string
.text:0000A453 loc_A453:
.text:0000A455 mov al, [edx+edi] ; load encrypted byte
.text:0000A458 xor eax, esi ; decrypt it
.text:0000A45A add esi, 3 ; change key for the next byte
.text:0000A45D inc edi
.text:0000A45E cmp edi, ecx
.text:0000A460 mov [ebp+edi+var_55], al
.text:0000A464 jl short loc_A453
As we can see, not just string is supplied to the decryption function, but also the key:
.text:0000DAF7 error: ; CODE XREF: sync_sys+255j
.text:0000DAF7 ; sync_sys+274j ...
.text:0000DAF7 mov [ebp+var_8], offset encrypted_error_message2
.text:0000DAFE mov [ebp+var_C], 17h ; decrypting key
.text:0000DB05 jmp decrypt_end_print_message
...
; this name we gave to label manually:

.text:0000D9B6 decrypt_end_print_message: ; CODE XREF: sync_sys+29Dj
.text:0000D9B6 ; sync_sys+2ABj
.text:0000D9B6 mov eax, [ebp+var_18]
.text:0000D9B9 test eax, eax
.text:0000D9BB jnz short loc_D9FB
.text:0000D9BD mov edx, [ebp+var_C] ; key
.text:0000D9C0 mov ecx, [ebp+var_8] ; string
.text:0000D9C3 push edx
.text:0000D9C6 push ecx
.text:0000D9C9 call err_warn
833
The algorithm is a simple xoring: each byte is xored with a key, but the key is increased by 3 after the
processing of each byte.
We can write a simple Python script to check our hypothesis:
Listing 8.6: Python 3.x

#!/usr/bin/python
import sys
msg=[0x74, 0x72, 0x78, 0x43, 0x48, 0x6, 0x5A, 0x49, 0x4C, 0x47, 0x47,
0x51, 0x4F, 0x47, 0x61, 0x20, 0x22, 0x3C, 0x24, 0x33, 0x36, 0x76,
0x3A, 0x33, 0x31, 0x0C, 0x0, 0x0B, 0x1F, 0x7, 0x1E, 0x1A]
key=0x17
tmp=key
for i in msg:
sys.stdout.write ("%c" % (i^tmp))
tmp=tmp+3
sys.stdout.flush()
And it prints: “check security device connection”. So yes, this is the decrypted message.
There are also other encrypted messages with their corresponding keys. But needless to say, it is possible
to decrypt them without their keys. First, we can see that the key is in fact a byte. It is because the core
decryption instruction (XOR) works on byte level. The key is located in the ESI register, but only one byte
part of ESI is used. Hence, a key may be greater than 255, but its value is always to be rounded.
As a consequence, we can just try brute-force, trying all possible keys in the 0..255 range. We are also
going to skip the messages that has unprintable characters.
Listing 8.7: Python 3.x

#!/usr/bin/python
import sys, curses.ascii
msgs=[
[0x74, 0x72, 0x78, 0x43, 0x48, 0x6, 0x5A, 0x49, 0x4C, 0x47, 0x47,
0x51, 0x4F, 0x47, 0x61, 0x20, 0x22, 0x3C, 0x24, 0x33, 0x36, 0x76,
0x3A, 0x33, 0x31, 0x0C, 0x0, 0x0B, 0x1F, 0x7, 0x1E, 0x1A],
[0x49, 0x65, 0x2D, 0x63, 0x76, 0x75, 0x6C, 0x6E, 0x76, 0x56, 0x5C,
8, 0x4F, 0x4B, 0x47, 0x5D, 0x54, 0x5F, 0x1D, 0x26, 0x2C, 0x33,
0x27, 0x28, 0x6F, 0x72, 0x75, 0x78, 0x7B, 0x7E, 0x41, 0x44],
[0x45, 0x61, 0x31, 0x67, 0x72, 0x79, 0x68, 0x52, 0x4A, 0x52, 0x50,
0x0C, 0x4B, 0x57, 0x43, 0x51, 0x58, 0x5B, 0x61, 0x37, 0x33, 0x2B,
0x39, 0x39, 0x3C, 0x38, 0x79, 0x3A, 0x30, 0x17, 0x0B, 0x0C],
[0x40, 0x64, 0x79, 0x75, 0x7F, 0x6F, 0x0, 0x4C, 0x40, 0x9, 0x4D, 0x5A,
0x46, 0x5D, 0x57, 0x49, 0x57, 0x3B, 0x21, 0x23, 0x6A, 0x38, 0x23,
0x36, 0x24, 0x2A, 0x7C, 0x3A, 0x1A, 0x6, 0x0D, 0x0E, 0x0A, 0x14,
0x10],
[0x72, 0x7C, 0x72, 0x79, 0x76, 0x0,

0x50, 0x43, 0x4A, 0x59, 0x5D, 0x5B, 0x41, 0x41, 0x1B, 0x5A,
0x24, 0x32, 0x2E, 0x29, 0x28, 0x70, 0x20, 0x22, 0x38, 0x28, 0x36,
0x0D, 0x0B, 0x48, 0x4B, 0x4E]]
def is_string_printable(s):
return all(list(map(lambda x: curses.ascii.isprint(x), s)))
cnt=1
for msg in msgs:
print ("message #%d" % cnt)
for key in range(0,256):
result=[]
tmp=key
for i in msg:
result.append (i^tmp)
tmp=tmp+3
834
if is_string_printable (result):
print ("key=", key, "value=", "".join(list(map(chr, result))))
cnt=cnt+1
And we get:
Listing 8.8: Results

message #1
key= 20 value= èb^h%|``hudw|_af{n~f%ljmSbnwlpk
key= 21 value= ajc]i"}cawtgv{^bgto}g"millcmvkqh
key= 22 value= bkd\j#rbbvsfuz!cduh|d#bhomdlujni
key= 23 value= check security device connection
key= 24 value= lifbl!pd|tqhsx#ejwjbb!`nQofbshlo
message #2
key= 7 value= No security device found
key= 8 value= An#rbbvsVuz!cduhld#ghtme?!#!'!#!
message #3
key= 7 value= Bk<waoqNUpu$`yreoa\wpmpusj,bkIjh
key= 8 value= Mj?vfnrOjqv%gxqd``_vwlstlk/clHii
key= 9 value= Lm>ugasLkvw&fgpgagûvcrwml.`mwhj
key= 10 value= Ol!td`tMhwx'efwfbf!tubuvnm!anvok
key= 11 value= No security device station found
key= 12 value= In#rjbvsnuz!{duhdd#r{`whho#gPtme
message #4
key= 14 value= Number of authorized users exceeded
key= 15 value= Ovlmdq!hg#`juknuhydk!vrbsp!Zy`dbefe
message #5
key= 17 value= check security device station
key= 18 value= ìjbh!td`tmhwx'efwfbf!tubuVnm!'!
There is some garbage, but we can quickly find the English-language messages!
By the way, since the algorithm is a simple xoring encryption, the very same function can be used to
encrypt messages. If needed, we can encrypt our own messages, and patch the program by inserting
them.
8.8.3 Example #3: MS-DOS

Another very old software for MS-DOS from 1995 also developed by a company that disappeared a long
time ago.
In the pre-DOS extenders era, all the software for MS-DOS mostly relied on 16-bit 8086 or 80286 CPUs, so
the code was 16-bit en masse.
The 16-bit code is mostly same as you already saw in this book, but all registers are 16-bit and there are
less instructions available.
The MS-DOS environment has no system drivers, and any program can deal with the bare hardware via
ports, so here you can see the OUT/IN instructions, which are present in mostly in drivers in our times (it
is impossible to access ports directly in user mode on all modern OSes).
Given that, the MS-DOS program which works with a dongle has to access the LPT printer port directly.
So we can just search for such instructions. And yes, here they are:
seg030:0034 out_port proc far ; CODE XREF: sent_pro+22p
seg030:0034 ; sent_pro+2Ap ...
seg030:0034
seg030:0034 arg_0 = byte ptr 6
seg030:0034
seg030:0034 55 push bp
seg030:0035 8B EC mov bp, sp
seg030:0037 8B 16 7E E7 mov dx, _out_port ; 0x378
seg030:003B 8A 46 06 mov al, [bp+arg_0]
seg030:003E EE out dx, al
seg030:003F 5D pop bp
seg030:0040 CB retf
seg030:0040 out_port endp
835
(All label names in this example were given by me).
out_port() is referenced only in one function:
seg030:0041 sent_pro proc far ; CODE XREF: check_dongle+34p
seg030:0041
seg030:0041 var_3 = byte ptr -3
seg030:0041 var_2 = word ptr -2
seg030:0041 arg_0 = dword ptr 6
seg030:0041
seg030:0041 C8 04 00 00 enter 4, 0
seg030:0045 56 push si
seg030:0046 57 push di
seg030:0047 8B 16 82 E7 mov dx, _in_port_1 ; 0x37A
seg030:004B EC in al, dx
seg030:004C 8A D8 mov bl, al
seg030:004E 80 E3 FE and bl, 0FEh
seg030:0051 80 CB 04 or bl, 4
seg030:0054 8A C3 mov al, bl
seg030:0056 88 46 FD mov [bp+var_3], al
seg030:0059 80 E3 1F and bl, 1Fh
seg030:005C 8A C3 mov al, bl
seg030:005E EE out dx, al
seg030:005F 68 FF 00 push 0FFh
seg030:0062 0E push cs
seg030:0063 E8 CE FF call near ptr out_port
seg030:0066 59 pop cx
seg030:0067 68 D3 00 push 0D3h
seg030:006A 0E push cs
seg030:006B E8 C6 FF call near ptr out_port
seg030:006E 59 pop cx
seg030:006F 33 F6 xor si, si
seg030:0071 EB 01 jmp short loc_359D4
seg030:0073
seg030:0073 loc_359D3: ; CODE XREF: sent_pro+37j
seg030:0073 46 inc si
seg030:0074
seg030:0074 loc_359D4: ; CODE XREF: sent_pro+30j
seg030:0074 81 FE 96 00 cmp si, 96h
seg030:0078 7C F9 jl short loc_359D3
seg030:007A 68 C3 00 push 0C3h
seg030:007D 0E push cs
seg030:007E E8 B3 FF call near ptr out_port
seg030:0081 59 pop cx
seg030:0082 68 C7 00 push 0C7h
seg030:0086 E8 AB FF call near ptr out_port
seg030:0089 59 pop cx
seg030:008A 68 D3 00 push 0D3h
seg030:008E E8 A3 FF call near ptr out_port
seg030:0091 59 pop cx
seg030:0092 68 C3 00 push 0C3h
seg030:0096 E8 9B FF call near ptr out_port
seg030:0099 59 pop cx
seg030:009A 68 C7 00 push 0C7h
seg030:009E E8 93 FF call near ptr out_port
seg030:00A1 59 pop cx
seg030:00A2 68 D3 00 push 0D3h
seg030:00A5 0E push cs
seg030:00A6 E8 8B FF call near ptr out_port
seg030:00A9 59 pop cx
seg030:00AA BF FF FF mov di, 0FFFFh
seg030:00AD EB 40 jmp short loc_35A4F
seg030:00AF
seg030:00AF loc_35A0F: ; CODE XREF: sent_pro+BDj
seg030:00AF BE 04 00 mov si, 4
seg030:00B2
seg030:00B2 loc_35A12: ; CODE XREF: sent_pro+ACj
836
seg030:00B2 D1 E7 shl di, 1
seg030:00B4 8B 16 80 E7 mov dx, _in_port_2 ; 0x379
seg030:00B8 EC in al, dx
seg030:00B9 A8 80 test al, 80h
seg030:00BB 75 03 jnz short loc_35A20
seg030:00BD 83 CF 01 or di, 1
seg030:00C0
seg030:00C0 loc_35A20: ; CODE XREF: sent_pro+7Aj
seg030:00C0 F7 46 FE 08+ test [bp+var_2], 8
seg030:00C5 74 05 jz short loc_35A2C
seg030:00C7 68 D7 00 push 0D7h ; '+'
seg030:00CA EB 0B jmp short loc_35A37
seg030:00CC
seg030:00CC loc_35A2C: ; CODE XREF: sent_pro+84j
seg030:00CC 68 C3 00 push 0C3h
seg030:00CF 0E push cs
seg030:00D0 E8 61 FF call near ptr out_port
seg030:00D3 59 pop cx
seg030:00D4 68 C7 00 push 0C7h
seg030:00D7
seg030:00D7 loc_35A37: ; CODE XREF: sent_pro+89j
seg030:00D7 0E push cs
seg030:00D8 E8 59 FF call near ptr out_port
seg030:00DB 59 pop cx
seg030:00DC 68 D3 00 push 0D3h
seg030:00DF 0E push cs
seg030:00E0 E8 51 FF call near ptr out_port
seg030:00E3 59 pop cx
seg030:00E4 8B 46 FE mov ax, [bp+var_2]
seg030:00E7 D1 E0 shl ax, 1
seg030:00E9 89 46 FE mov [bp+var_2], ax
seg030:00EC 4E dec si
seg030:00ED 75 C3 jnz short loc_35A12
seg030:00EF
seg030:00EF loc_35A4F: ; CODE XREF: sent_pro+6Cj
seg030:00EF C4 5E 06 les bx, [bp+arg_0]
seg030:00F2 FF 46 06 inc word ptr [bp+arg_0]
seg030:00F5 26 8A 07 mov al, es:[bx]
seg030:00F8 98 cbw
seg030:00F9 89 46 FE mov [bp+var_2], ax
seg030:00FC 0B C0 or ax, ax
seg030:00FE 75 AF jnz short loc_35A0F
seg030:0100 68 FF 00 push 0FFh
seg030:0104 E8 2D FF call near ptr out_port
seg030:0107 59 pop cx
seg030:010C EC in al, dx
seg030:010D 8A C8 mov cl, al
seg030:010F 80 E1 5F and cl, 5Fh
seg030:0112 8A C1 mov al, cl
seg030:0114 EE out dx, al
seg030:0115 EC in al, dx
seg030:0116 8A C8 mov cl, al
seg030:0118 F6 C1 20 test cl, 20h
seg030:011B 74 08 jz short loc_35A85
seg030:011D 8A 5E FD mov bl, [bp+var_3]
seg030:0120 80 E3 DF and bl, 0DFh
seg030:0123 EB 03 jmp short loc_35A88
seg030:0125
seg030:0125 loc_35A85: ; CODE XREF: sent_pro+DAj
seg030:0125 8A 5E FD mov bl, [bp+var_3]
seg030:0128
seg030:0128 loc_35A88: ; CODE XREF: sent_pro+E2j
seg030:0128 F6 C1 80 test cl, 80h
seg030:012B 74 03 jz short loc_35A90
seg030:012D 80 E3 7F and bl, 7Fh
seg030:0130
seg030:0130 loc_35A90: ; CODE XREF: sent_pro+EAj
837
seg030:0134 8A C3 mov al, bl
seg030:0136 EE out dx, al
seg030:0137 8B C7 mov ax, di
seg030:0139 5F pop di
seg030:013A 5E pop si
seg030:013B C9 leave
seg030:013C CB retf
seg030:013C sent_pro endp
This is again a Sentinel Pro “hashing” dongle as in the previous example. It is noticeably because text
strings are passed here, too, and 16 bit values are returned and compared with others.
So that is how Sentinel Pro is accessed via ports.
The output port address is usually 0x378, i.e., the printer port, where the data to the old printers in pre-USB
era was passed to.
The port is uni-directional, because when it was developed, no one imagined that someone will need to
transfer information from the printer 18 .
The only way to get information from the printer is a status register on port 0x379, which contains such
bits as “paper out”, “ack”, “busy”—thus the printer may signal to the host computer if it is ready or not
and if paper is present in it.
So the dongle returns information from one of these bits, one bit at each iteration.
_in_port_2 contains the address of the status word (0x379) and _in_port_1 contains the control register
address (0x37A).
It seems that the dongle returns information via the “busy” flag at seg030:00B9: each bit is stored in the
DI register, which is returned at the end of the function.
What do all these bytes sent to output port mean? Hard to say. Perhaps, commands to the dongle.
But generally speaking, it is not necessary to know: it is easy to solve our task without that knowledge.
Here is the dongle checking routine:
00000000 struct_0 struc ; (sizeof=0x1B)
00000000 field_0 db 25 dup(?) ; string(C)
00000019 _A dw ?
0000001B struct_0 ends
dseg:3CBC 61 63 72 75+_Q struct_0 <'hello', 01122h>

dseg:3CBC 6E 00 00 00+ ; DATA XREF: check_dongle+2Eo
... skipped ...
dseg:3E00 63 6F 66 66+ struct_0 <'coffee', 7EB7h>

dseg:3E1B 64 6F 67 00+ struct_0 <'dog', 0FFADh>
dseg:3E36 63 61 74 00+ struct_0 <'cat', 0FF5Fh>
dseg:3E51 70 61 70 65+ struct_0 <'paper', 0FFDFh>
dseg:3E6C 63 6F 6B 65+ struct_0 <'coke', 0F568h>
dseg:3E87 63 6C 6F 63+ struct_0 <'clock', 55EAh>
dseg:3EA2 64 69 72 00+ struct_0 <'dir', 0FFAEh>
dseg:3EBD 63 6F 70 79+ struct_0 <'copy', 0F557h>
seg030:0145 check_dongle proc far ; CODE XREF: sub_3771D+3EP

seg030:0145
seg030:0145 var_6 = dword ptr -6
seg030:0145 var_2 = word ptr -2
seg030:0145
seg030:0145 C8 06 00 00 enter 6, 0
seg030:0149 56 push si
seg030:014A 66 6A 00 push large 0 ; newtime
seg030:014D 6A 00 push 0 ; cmd
seg030:014F 9A C1 18 00+ call _biostime
seg030:0154 52 push dx
seg030:0155 50 push ax
seg030:0156 66 58 pop eax
18 If we consider Centronics only. The following IEEE 1284 standard allows the transfer of information from the printer.
838
seg030:0158 83 C4 06 add sp, 6
seg030:015B 66 89 46 FA mov [bp+var_6], eax
seg030:015F 66 3B 06 D8+ cmp eax, _expiration
seg030:0164 7E 44 jle short loc_35B0A
seg030:0166 6A 14 push 14h
seg030:0168 90 nop
seg030:016A E8 52 00 call near ptr get_rand
seg030:016D 59 pop cx
seg030:016E 8B F0 mov si, ax
seg030:0170 6B C0 1B imul ax, 1Bh
seg030:0173 05 BC 3C add ax, offset _Q
seg030:0176 1E push ds
seg030:0177 50 push ax
seg030:0179 E8 C5 FE call near ptr sent_pro
seg030:017C 83 C4 04 add sp, 4
seg030:017F 89 46 FE mov [bp+var_2], ax
seg030:0182 8B C6 mov ax, si
seg030:0184 6B C0 12 imul ax, 18
seg030:0187 66 0F BF C0 movsx eax, ax
seg030:018B 66 8B 56 FA mov edx, [bp+var_6]
seg030:018F 66 03 D0 add edx, eax
seg030:0192 66 89 16 D8+ mov _expiration, edx
seg030:0197 8B DE mov bx, si
seg030:0199 6B DB 1B imul bx, 27
seg030:019C 8B 87 D5 3C mov ax, _Q._A[bx]
seg030:01A0 3B 46 FE cmp ax, [bp+var_2]
seg030:01A3 74 05 jz short loc_35B0A
seg030:01A5 B8 01 00 mov ax, 1
seg030:01A8 EB 02 jmp short loc_35B0C
seg030:01AA
seg030:01AA loc_35B0A: ; CODE XREF: check_dongle+1Fj
seg030:01AA ; check_dongle+5Ej
seg030:01AA 33 C0 xor ax, ax
seg030:01AC
seg030:01AC loc_35B0C: ; CODE XREF: check_dongle+63j
seg030:01AC 5E pop si
seg030:01AD C9 leave
seg030:01AE CB retf
seg030:01AE check_dongle endp
Since the routine can be called very frequently, e.g., before the execution of each important software
feature, and accessing the dongle is generally slow (because of the slow printer port and also slow MCU
in the dongle), they probably added a way to skip some dongle checks, by checking the current time in
the biostime() function.
The get_rand() function uses the standard C function:
seg030:01BF get_rand proc far ; CODE XREF: check_dongle+25p
seg030:01BF
seg030:01BF arg_0 = word ptr 6
seg030:01BF
seg030:01BF 55 push bp
seg030:01C0 8B EC mov bp, sp
seg030:01C2 9A 3D 21 00+ call _rand
seg030:01C7 66 0F BF C0 movsx eax, ax
seg030:01CB 66 0F BF 56+ movsx edx, [bp+arg_0]
seg030:01D0 66 0F AF C2 imul eax, edx
seg030:01D4 66 BB 00 80+ mov ebx, 8000h
seg030:01DA 66 99 cdq
seg030:01DC 66 F7 FB idiv ebx
seg030:01DF 5D pop bp
seg030:01E0 CB retf
seg030:01E0 get_rand endp
So the text string is selected randomly, passed into the dongle, and then the result of the hashing is
compared with the correct value.
The text strings seem to be constructed randomly as well, during software development.
839
And this is how the main dongle checking function is called:
seg033:087B 9A 45 01 96+ call check_dongle
seg033:0880 0B C0 or ax, ax
seg033:0882 74 62 jz short OK
seg033:0884 83 3E 60 42+ cmp word_620E0, 0
seg033:0889 75 5B jnz short OK
seg033:088B FF 06 60 42 inc word_620E0
seg033:088F 1E push ds
seg033:0890 68 22 44 push offset aTrupcRequiresA ;
"This Software Requires a Software Lock\n"
seg033:0894 68 60 E9 push offset byte_6C7E0 ; dest
seg033:0897 9A 79 65 00+ call _strcpy
seg033:089C 83 C4 08 add sp, 8
seg033:08A0 68 42 44 push offset aPleaseContactA ; "Please Contact ..."
seg033:08A3 1E push ds
seg033:08A4 68 60 E9 push offset byte_6C7E0 ; dest
seg033:08A7 9A CD 64 00+ call _strcat
Bypassing the dongle is easy, just force the check_dongle() function to always return 0.
For example, by inserting this code at its beginning:
mov ax,0
retf
The observant reader might recall that the strcpy() C function usually requires two pointers in its argu-
ments, but we see that 4 values are passed:
seg033:0890 68 22 44 push offset aTrupcRequiresA ;
"This Software Requires a Software Lock\n"
seg033:0894 68 60 E9 push offset byte_6C7E0 ; dest
seg033:0897 9A 79 65 00+ call _strcpy
seg033:089C 83 C4 08 add sp, 8
This is related to MS-DOS’ memory model. You can read more about it here: 11.6 on page 986.
So as you may see, strcpy() and any other function that take pointer(s) in arguments work with 16-bit
pairs.
Let’s get back to our example. DS is currently set to the data segment located in the executable, that is
where the text string is stored.
In the sent_pro() function, each byte of the string is loaded at
seg030:00EF: the LES instruction loads the ES:BX pair simultaneously from the passed argument.
The MOV at seg030:00F5 loads the byte from the memory at which the ES:BX pair points.
8.9 Encrypted database case #1

(This part has been first appeared in my blog at 26-Aug-2015. Some discussion: https://news.ycombinator.
com/item?id=10128684.)
8.9.1 Base64 and entropy

I’ve got the XML file containing some encrypted data. Perhaps, it’s related to some orders and/or cus-
tomers information.
<?xml version = "1.0" encoding = "UTF-8"?>
<Orders>
<Order>
<OrderID>1</OrderID>
<Data>yjmxhXUbhB/5MV45chPsXZWAJwIh1S0aD9lFn3XuJMSxJ3/E+UE3hsnH</Data>
</Order>
<Order>
840
<Data>0KGe/wnypFBjsy+U0C2P9fC5nDZP3XDZLMPCRaiBw9OjIk6Tu5U=</Data>
</Order>
<Order>
<Data>mqkXfdzvQKvEArdzh+zD9oETVGBFvcTBLs2ph1b5bYddExzp</Data>
</Order>
<Order>
<Data>FCx6JhIDqnESyT3HAepyE1BJ3cJd7wCk+APCRUeuNtZdpCvQ2MR/7kLXtfUHuA==</Data>
</Order>
...
The file is available here.

This is clearly base64-encoded data, because all strings consisting of Latin characters, digits, plus (+) and
slash (/) symbols. There can be 1 or 2 padding symbols (=), but they are never occurred in the middle of
string. Keeping in mind these base64 properties, it’s very easy to recognize them.
Let’s decode them and calculate entropies ( 9.2 on page 929) of these blocks in Wolfram Mathematica:
In[]:= ListOfBase64Strings =
Map[First[#[[3]]] &, Cases[Import["encrypted.xml"], XMLElement["Data", _, _], Infinity]];
In[]:= BinaryStrings =
Map[ImportString[#, {"Base64", "String"}] &, ListOfBase64Strings];
In[]:= Entropies = Map[N[Entropy[2, #]] &, BinaryStrings];
In[]:= Variance[Entropies]
Out[]= 0.0238614
Variance is low. This means the entropy values are not very different from each other. This is visible on
graph:
In[]:= ListPlot[Entropies]
Most values are between 5.0 and 5.4. This is a sign that the data is compressed and/or encrypted.
841
To understand variance, let’s calculate entropies of all lines in Conan Doyle’s The Hound of the Baskervilles
book:
In[]:= BaskervillesLines = Import["http://www.gutenberg.org/cache/epub/2852/pg2852.txt", "List⤦
Ç "];
In[]:= EntropiesT = Map[N[Entropy[2, #]] &, BaskervillesLines];
In[]:= Variance[EntropiesT]
Out[]= 2.73883
In[]:= ListPlot[EntropiesT]
Most values are gathered around value of 4, but there are also values which are smaller, and they are
influenced final variance value.
Perhaps, shortest strings has smaller entropy, let’s take short string from the Conan Doyle’s book:
In[]:= Entropy[2, "Yes, sir."] // N
Out[]= 2.9477
Let’s try even shorter:

In[]:= Entropy[2, "Yes"] // N
Out[]= 1.58496
In[]:= Entropy[2, "No"] // N

Out[]= 1.
8.9.2 Is data compressed?

OK, so our data is compressed and/or encrypted. Is it compressed? Almost all data compressors put some
header at the start, signature, or something like that. As we can see, there are no consistent data at the
start of each block. It’s still possible that this is a handmade data compressor, but they are very rare. On
the other hand, handmade cryptoalgorithms are much more popular, because it’s very easy to make it
842
work. Even primitive keyless cryptosystems like memfrob()19 and ROT13 works fine without errors. It’s
a serious challenge to write data compressor from scratch using only fantasy and imagination in a way
so it will have no evident bugs. Some programmers implements data compression functions by reading
textbooks, but this is also rare. The most popular two ways are: 1) just take open-source library like zlib; 2)
copy&paste something from somewhere. Open-source data compressions algorithms usually puts some
kind of header, and so do algorithms from sites like http://www.codeproject.com/.
8.9.3 Is data encrypted?

Major data encryption algorithms process data in blocks. DES—8 bytes, AES—16 bytes. If the input buffer
is not divided evenly by block size, it’s padded by zeroes (or something else), so encrypted data will be
aligned by cryptoalgorithm’s block size. This is not our case.
Using Wolfram Mathematica, I analyzed block’s lengths:
In[]:= Counts[Map[StringLength[#] &, BinaryStrings]]
Out[]= <|42 -> 1858, 38 -> 1235, 36 -> 699, 46 -> 1151, 40 -> 1784,
44 -> 1558, 50 -> 366, 34 -> 291, 32 -> 74, 56 -> 15, 48 -> 716,
30 -> 13, 52 -> 156, 54 -> 71, 60 -> 3, 58 -> 6, 28 -> 4|>
1858 blocks has size of 42 bytes, 1235 blocks has size of 38 bytes, etc.
I made a graph:
ListPlot[Counts[Map[StringLength[#] &, BinaryStrings]]]
So, most blocks has size between ~36 and ~48. There is also another thing to notice: all block sizes are
even. No single block with odd size.
There are, however, stream ciphers which can operate on byte level or even on bit level.
8.9.4 CryptoPP
The program which can browse this encrypted database is written C# and the .NET code is heavily obfus-
cated. Nevertheless, there is DLL with x86 code, which, after brief examination, has parts of the CryptoPP
19 http://linux.die.net/man/3/memfrob
843
popular open-source library! (I just spotted “CryptoPP” strings inside.) Now it’s very easy to find all
functions inside of DLL because CryptoPP library is open-source.
CryptoPP library has a lot of crypto-functions, including AES (AKA Rijndael). Newer x86 CPUs has AES
helper instructions like AESENC, AESDEC and AESKEYGENASSIST 20 . They are not performing encryption/de-
cryption completely, but they do significant amount of job. And newer CryptoPP versions use them. For
example, here: 1, 2. To my surprise, during decryption, AESENC gets executed, while AESDEC is not (I just
checked with my tracer utility, but any debugger can be used). I checked, if my CPU really supports AES
instructions. Some Intel i3 CPUs are not. And if not, CryptoPP library falling back to AES functions imple-
mented in old way 21 . But my CPU supports them. Why AESDEC is still not executed? Why the program
use AES encryption in order to decrypt database?
OK, it’s not a problem to find a function which encrypts block. It is called
CryptoPP::Rijndael::Enc::ProcessAndXorBlock: src, and it can call another function:
Rijndael::Enc::AdvancedProcessBlocks() src, which, in turn, can call two other functions ( AESNI_Enc_Block
and AESNI_Enc_4_Blocks ) which has AESENC instructions.
So, judging by CryptoPP internals,
CryptoPP::Rijndael::Enc::ProcessAndXorBlock() encrypts one 16-byte block. Let’s set breakpoint on it and
see, what happens during decryption. I use my simple tracer tool again. The software must decrypt
first data block now. Oh, by the way, here is the first data block converted from base64 encoding to
hexadecimal data, let’s have it at hand:
00000000: CA 39 B1 85 75 1B 84 1F F9 31 5E 39 72 13 EC 5D .9..u....1^9r..]
00000010: 95 80 27 02 21 D5 2D 1A 0F D9 45 9F 75 EE 24 C4 ..'.!.-...E.u.$.
00000020: B1 27 7F 84 FE 41 37 86 C9 C0 .'...A7...
These are also arguments of the function from CryptoPP source files:
size_t Rijndael::Enc::AdvancedProcessBlocks(const byte *inBlocks, const byte *xorBlocks, byte *⤦
Ç outBlocks, size_t length, word32 flags);
So it has 5 arguments. Possible flags are:

enum {BT_InBlockIsCounter=1, BT_DontIncrementInOutPointers=2, BT_XorInput=4, ⤦
Ç BT_ReverseDirection=8, BT_AllowParallel=16} FlagsForAdvancedProcessBlocks;
OK, run tracer on ProcessAndXorBlock() function:

... tracer.exe -l:filename.exe bpf=filename.exe!0x4339a0,args:5,dump_args:0x10
Warning: no tracer.cfg file.

PID=1984|New process software.exe
no module registered with image base 0x77320000
no module registered with image base 0x76e20000
Warning: unknown (to us) INT3 breakpoint at ntdll.dll!LdrVerifyImageMatchesChecksum+0x96c (0⤦
Ç x776c103b)
(0) software.exe!0x4339a0(0x38b920, 0x0, 0x38b978, 0x10, 0x0) (called from software.exe!.text+0⤦
Ç x33c0d (0x13e4c0d))
Argument 1/5
0038B920: 01 00 00 00 FF FF FF FF-79 C1 69 0B 67 C1 04 7D "........y.i.g..}"
Argument 3/5
0038B978: CD CD CD CD CD CD CD CD-CD CD CD CD CD CD CD CD "................"
(0) software.exe!0x4339a0() -> 0x0
Argument 3/5 difference
00000000: C7 39 4E 7B 33 1B D6 1F-B8 31 10 39 39 13 A5 5D ".9N{3....1.99..]"
(0) software.exe!0x4339a0(0x38a828, 0x38a838, 0x38bb40, 0x0, 0x8) (called from software.exe!.⤦
Ç text+0x3a407 (0x13eb407))
Argument 1/5
0038A828: 95 80 27 02 21 D5 2D 1A-0F D9 45 9F 75 EE 24 C4 "..'.!.-...E.u.$."
Argument 2/5
0038A838: B1 27 7F 84 FE 41 37 86-C9 C0 00 CD CD CD CD CD ".'...A7........."
Argument 3/5
0038BB40: CD CD CD CD CD CD CD CD-CD CD CD CD CD CD CD CD "................"
20 https://en.wikipedia.org/wiki/AES_instruction_set
21 https://github.com/mmoss/cryptopp/blob/2772f7b57182b31a41659b48d5f35a7b6cedd34d/src/rijndael.cpp#L355
844
(0) software.exe!0x4339a0(0x38b920, 0x38a828, 0x38bb30, 0x10, 0x0) (called from software.exe!.⤦
Ç text+0x33c0d (0x13e4c0d))
Argument 1/5
0038B920: CA 39 B1 85 75 1B 84 1F-F9 31 5E 39 72 13 EC 5D ".9..u....1^9r..]"
Argument 2/5
0038A828: 95 80 27 02 21 D5 2D 1A-0F D9 45 9F 75 EE 24 C4 "..'.!.-...E.u.$."
Argument 3/5
0038BB30: CD CD CD CD CD CD CD CD-CD CD CD CD CD CD CD CD "................"
00000000: 45 00 20 00 4A 00 4F 00-48 00 4E 00 53 00 00 00 "E. .J.O.H.N.S..."
(0) software.exe!0x4339a0(0x38b920, 0x0, 0x38b978, 0x10, 0x0) (called from software.exe!.text+0⤦
Ç x33c0d (0x13e4c0d))
Argument 1/5
0038B920: 95 80 27 02 21 D5 2D 1A-0F D9 45 9F 75 EE 24 C4 "..'.!.-...E.u.$."
Argument 3/5
0038B978: 95 80 27 02 21 D5 2D 1A-0F D9 45 9F 75 EE 24 C4 "..'.!.-...E.u.$."
00000000: B1 27 7F E4 9F 01 E3 81-CF C6 12 FB B9 7C F1 BC ".'...........|.."
PID=1984|Process software.exe exited. ExitCode=0 (0x0)
Here we can see inputs to the ProcessAndXorBlock() function, and outputs from it.
This is output from the function during first call:
00000000: C7 39 4E 7B 33 1B D6 1F-B8 31 10 39 39 13 A5 5D ".9N{3....1.99..]"
Then the ProcessAndXorBlock() is called with zero-length block, but with 8 flag (BT_ReverseDirection).
Second call:
00000000: 45 00 20 00 4A 00 4F 00-48 00 4E 00 53 00 00 00 "E. .J.O.H.N.S..."
Wow, there is some string familiar to us!

Third call:
00000000: B1 27 7F E4 9F 01 E3 81-CF C6 12 FB B9 7C F1 BC ".'...........|.."
The first output is very similar to the first 16 bytes of the encrypted buffer.
Output of the first call of ProcessAndXorBlock():
00000000: C7 39 4E 7B 33 1B D6 1F-B8 31 10 39 39 13 A5 5D ".9N{3....1.99..]"
First 16 bytes of encrypted buffer:

00000000: CA 39 B1 85 75 1B 84 1F F9 31 5E 39 72 13 EC 5D .9..u....1^9r..]
There are too much equal bytes! How come AES encryption result can be very similar to the encrypted
buffer while this is not encryption but rather decryption?!
8.9.5 Cipher Feedback mode

The answer is CFB22 : in this mode, AES algorithm used not as encryption algorithm, but as a device which
generates cryptographically secure random data. The actual encryption is happening using simple XOR
operation.
Here is encryption algorithm (images are taken from Wikipedia):
22 Cipher Feedback
845
And decryption:
Now let’s see: AES encryption operation generates 16 bytes (or 128 bits) of random data to be used while
XOR-ing, who forces us to use all 16 bytes? If at the last iteration we’ve got 1 byte of data, let’s xor 1 byte
of data with 1 byte of generated random data? This leads to important property of CFB mode: data can
be not padded, data of arbitrary size can be encrypted and decrypted.
Oh, that’s why all encrypted blocks are not padded. And that’s why AESDEC instruction is never called.
Let’s try to decrypt first block manually, using Python. CFB mode also use IV, as a seed for CSPRNG23 . In
our case, IV is the block which is encrypted at first iteration:
0038B920: 01 00 00 00 FF FF FF FF-79 C1 69 0B 67 C1 04 7D "........y.i.g..}"
Oh, and we also have to recover encryption key. There is AESKEYGENASSIST is DLL, and it is called, and it
is used in the
Rijndael::Base::UncheckedSetKey() function: src. It’s easy to find it in IDA and set breakpoint. Let’s see:
... tracer.exe -l:filename.exe bpf=filename.exe!0x435c30,args:3,dump_args:0x10
Warning: no tracer.cfg file.

PID=2068|New process software.exe
no module registered with image base 0x76e20000
23 Cryptographically Secure Pseudorandom Number Generator
846
Warning: unknown (to us) INT3 breakpoint at ntdll.dll!LdrVerifyImageMatchesChecksum+0x96c (0⤦
Ç x776c103b)
(0) software.exe!0x435c30(0x15e8000, 0x10, 0x14f808) (called from software.exe!.text+0x22fa1 (0⤦
Ç x13d3fa1))
Argument 1/3
015E8000: CD C5 7E AD 28 5F 6D E1-CE 8F CC 29 B1 21 88 8E "..~.(_m....).!.."
Argument 3/3
0014F808: 38 82 58 01 C8 B9 46 00-01 D1 3C 01 00 F8 14 00 "8.X...F...<....."
Argument 3/3 +0x0: software.exe!.rdata+0x5238
Argument 3/3 +0x8: software.exe!.text+0x1c101
(0) software.exe!0x435c30() -> 0x13c2801
PID=2068|Process software.exe exited. ExitCode=0 (0x0)
So this is the key: CD C5 7E AD 28 5F 6D E1-CE 8F CC 29 B1 21 88 8E.

During manual decryption we’ve got this:
00000000: 0D 00 FF FE 46 00 52 00 41 00 4E 00 4B 00 49 00 ....F.R.A.N.K.I.
00000010: 45 00 20 00 4A 00 4F 00 48 00 4E 00 53 00 66 66 E. .J.O.H.N.S.ff
00000020: 66 66 66 9E 61 40 D4 07 06 01 fff.a@....
Now this is something readable! And now we can see why there were so many equal bytes at the first
decryption iteration: because plaintext has so many zero bytes!
Let’s decrypt the second block:
00000000: 17 98 D0 84 3A E9 72 4F DB 82 3F AD E9 3E 2A A8 ....:.rO..?..>*.
00000010: 41 00 52 00 52 00 4F 00 4E 00 CD CC CC CC CC CC A.R.R.O.N.......
00000020: 1B 40 D4 07 06 01 .@....
Third, fourth and fifth:

00000000: 5D 90 59 06 EF F4 96 B4 7C 33 A7 4A BE FF 66 AB ].Y.....|3.J..f.
00000010: 49 00 47 00 47 00 53 00 00 00 00 00 00 C0 65 40 I.G.G.S.......e@
00000020: D4 07 06 01 ....
00000000: D3 15 34 5D 21 18 7C 6E AA F8 2D FE 38 F9 D7 4E ..4]!.|n..-.8..N
00000010: 41 00 20 00 44 00 4F 00 48 00 45 00 52 00 54 00 A. .D.O.H.E.R.T.
00000020: 59 00 48 E1 7A 14 AE FF 68 40 D4 07 06 02 Y.H.z...h@....
00000000: 1E 8B 90 0A 17 7B C5 52 31 6C 4E 2F DE 1B 27 19 .....{.R1lN...'.
00000010: 41 00 52 00 43 00 55 00 53 00 00 00 00 00 00 60 A.R.C.U.S.......
00000020: 66 40 D4 07 06 03 f@....
All blocks decrypted seems correct except of first 16 bytes part.
8.9.6 Initializing Vector

What can affect first 16 bytes?
Let’s back to CFB decryption algorithm again: 8.9.5 on the previous page.
We can see that IV can affect to first block decryption operation, but not the second, because during the
second iteration, ciphertext from the first iteration is used, and in case of decryption, it’s the same, no
matter what IV has!
So probably, IV is different each time. Using my tracer, I’ll take a look at the first input during decryption
of the second block of XML file:
0038B920: 02 00 00 00 FE FF FF FF-79 C1 69 0B 67 C1 04 7D "........y.i.g..}"
…third:
0038B920: 03 00 00 00 FD FF FF FF-79 C1 69 0B 67 C1 04 7D "........y.i.g..}"
It seems, first and fifth byte are changed each time. I finally concluded that the first 32-bit integer is
just OrderID from the XML file, and the second 32-bit integer is also OrderID, but negated. All other 8
847
bytes are same for each operation. Now I have decrypted the whole database: https://beginners.re/
current-tree/examples/encrypted_DB1/decrypted.full.txt.
The Python script used for this is: https://beginners.re/current-tree/examples/encrypted_DB1/
decrypt_blocks.py.
Perhaps, the author wanted each block encrypted differently, so he/she used OrderID as part of key. It
would be also possible to make different AES key instead of IV.
So now we know that IV only affects first block during decryption in CFB mode, this is feature of it. All
other blocks can be decrypted without knowledge IV, but using the key.
OK, so why CFB mode? Apparently, because the very first AES example on CryptoPP wiki uses CFB mode:
http://www.cryptopp.com/wiki/Advanced_Encryption_Standard#Encrypting_and_Decrypting_Using_
AES. Supposedly, developer choose it for simplicity: the example can encrypt/decrypt text strings with
arbitrary lengths, without padding.
It is very likely, program’s author(s) just copypasted the example from CryptoPP wiki page. Many pro-
grammers do so.
The only difference that IV is chosen randomly in CryptoPP wiki example, while this indeterminism wasn’t
allowable to programmers of the software we are dissecting now, so they choose to initialize IV using
Order ID.
Now we can proceed to analyzing matter of each byte in the decrypted block.
8.9.7 Structure of the buffer

Let’s take first four decrypted blocks:
00000000: 0D 00 FF FE 46 00 52 00 41 00 4E 00 4B 00 49 00 ....F.R.A.N.K.I.
00000010: 45 00 20 00 4A 00 4F 00 48 00 4E 00 53 00 66 66 E. .J.O.H.N.S.ff
00000020: 66 66 66 9E 61 40 D4 07 06 01 fff.a@....
00000000: 0B 00 FF FE 4C 00 4F 00 52 00 49 00 20 00 42 00 ....L.O.R.I. .B.

00000020: 1B 40 D4 07 06 01 .@....
00000000: 0A 00 FF FE 47 00 41 00 52 00 59 00 20 00 42 00 ....G.A.R.Y. .B.

00000010: 49 00 47 00 47 00 53 00 00 00 00 00 00 C0 65 40 I.G.G.S.......e@
00000020: D4 07 06 01 ....
00000000: 0F 00 FF FE 4D 00 45 00 4C 00 49 00 4E 00 44 00 ....M.E.L.I.N.D.
00000010: 41 00 20 00 44 00 4F 00 48 00 45 00 52 00 54 00 A. .D.O.H.E.R.T.
00000020: 59 00 48 E1 7A 14 AE FF 68 40 D4 07 06 02 Y.H.z...h@....
UTF-16 encoded text strings are clearly visible, these are names and surnames. The first byte (or 16-bit
word) is seems string length, we can visually check it. FF FE is seems Unicode BOM.
There are 12 more bytes after each string.
Using this script (https://beginners.re/current-tree/examples/encrypted_DB1/dump_buffer_rest.
py) I’ve got random selection of the tails:
dennis@...:$ python decrypt.py encrypted.xml | shuf | head -20
00000000: 48 E1 7A 14 AE 5F 62 40 DD 07 05 08 H.z.._b@....
00000000: 00 00 00 00 00 40 5A 40 DC 07 08 18 .....@Z@....
00000000: 00 00 00 00 00 80 56 40 D7 07 0B 04 ......V@....
00000000: 00 00 00 00 00 60 61 40 D7 07 0C 1C ......a@....
00000000: 00 00 00 00 00 20 63 40 D9 07 05 18 ..... c@....
00000000: 3D 0A D7 A3 70 FD 34 40 D7 07 07 11 =...p.4@....
00000000: 00 00 00 00 00 A0 63 40 D5 07 05 19 ......c@....
00000000: CD CC CC CC CC 3C 5C 40 D7 07 08 11 .......@....
00000000: 66 66 66 66 66 FE 62 40 D4 07 06 05 fffff.b@....
00000000: 1F 85 EB 51 B8 FE 40 40 D6 07 09 1E ...Q..@@....
00000000: 00 00 00 00 00 40 5F 40 DC 07 02 18 .....@_@....
00000000: 48 E1 7A 14 AE 9F 67 40 D8 07 05 12 H.z...g@....
00000000: CD CC CC CC CC 3C 5E 40 DC 07 01 07 ......^@....
00000000: 00 00 00 00 00 00 67 40 D4 07 0B 0E ......g@....
00000000: 00 00 00 00 00 40 51 40 DC 07 04 0B .....@Q@....
00000000: 00 00 00 00 00 40 56 40 D7 07 07 0A .....@V@....
00000000: 8F C2 F5 28 5C 7F 55 40 DB 07 01 16 ...(..U@....
848
00000000: 00 00 00 00 00 00 32 40 DB 07 06 09 ......2@....
00000000: 66 66 66 66 66 7E 66 40 D9 07 0A 06 fffff~f@....
00000000: 48 E1 7A 14 AE DF 68 40 D5 07 07 16 H.z...h@....
We first see the 0x40 and 0x07 bytes present in each tail. The very last byte s always in 1..0x1F (1..31)
range, I’ve checked. The penultimate byte is always in 1..0xC (1..12) range. Wow, that looks like a date!
Year can be represented as 16-bit value, and maybe last 4 bytes is date (16 bits for year, 8 bits for month
and 8 more for day)? 0x7DD is 2013, 0x7D5 is 2005, etc. Seems fine. This is a date. There are 8 more
bytes. Judging by the fact this is database named orders, maybe some kind of sum is present here? I
made attempt to interpret it as double-precision IEEE 754 floating point and dump all values!
Some are:
71.0
134.0
51.95
53.0
121.99
96.95
98.95
15.95
85.95
184.99
94.95
29.95
85.0
36.0
130.99
115.95
87.99
127.95
114.0
150.95
Looks like real!

Now we can dump names, sums and dates.
plain:
00000000: 0D 00 FF FE 46 00 52 00 41 00 4E 00 4B 00 49 00 ....F.R.A.N.K.I.
00000010: 45 00 20 00 4A 00 4F 00 48 00 4E 00 53 00 66 66 E. .J.O.H.N.S.ff
00000020: 66 66 66 9E 61 40 D4 07 06 01 fff.a@....
OrderID= 1 name= FRANKIE JOHNS sum= 140.95 date= 2004 / 6 / 1
plain:
00000000: 0B 00 FF FE 4C 00 4F 00 52 00 49 00 20 00 42 00 ....L.O.R.I. .B.
00000020: 1B 40 D4 07 06 01 .@....
OrderID= 2 name= LORI BARRON sum= 6.95 date= 2004 / 6 / 1
plain:
00000000: 0A 00 FF FE 47 00 41 00 52 00 59 00 20 00 42 00 ....G.A.R.Y. .B.
00000010: 49 00 47 00 47 00 53 00 00 00 00 00 00 C0 65 40 I.G.G.S.......e@
00000020: D4 07 06 01 ....
OrderID= 3 name= GARY BIGGS sum= 174.0 date= 2004 / 6 / 1
plain:
00000000: 0F 00 FF FE 4D 00 45 00 4C 00 49 00 4E 00 44 00 ....M.E.L.I.N.D.
00000010: 41 00 20 00 44 00 4F 00 48 00 45 00 52 00 54 00 A. .D.O.H.E.R.T.
00000020: 59 00 48 E1 7A 14 AE FF 68 40 D4 07 06 02 Y.H.z...h@....
OrderID= 4 name= MELINDA DOHERTY sum= 199.99 date= 2004 / 6 / 2
plain:
00000000: 0B 00 FF FE 4C 00 45 00 4E 00 41 00 20 00 4D 00 ....L.E.N.A. .M.
00000010: 41 00 52 00 43 00 55 00 53 00 00 00 00 00 00 60 A.R.C.U.S.......
00000020: 66 40 D4 07 06 03 f@....
OrderID= 5 name= LENA MARCUS sum= 179.0 date= 2004 / 6 / 3
See more: https://beginners.re/current-tree/examples/encrypted_DB1/decrypted.full.with_data.

txt. Or filtered: https://beginners.re/current-tree/examples/encrypted_DB1/decrypted.short.
849
txt. Seems correct.
This is some kind of OOP serialization, i.e., packing differently typed values into binary buffer for storing
and/or transmitting.
8.9.8 Noise at the end

The only question remaining is that sometimes, tail is bigger:
00000000: 0E 00 FF FE 54 00 48 00 45 00 52 00 45 00 53 00 ....T.H.E.R.E.S.
00000010: 45 00 20 00 54 00 55 00 54 00 54 00 4C 00 45 00 E. .T.U.T.T.L.E.
00000020: 66 66 66 66 66 1E 63 40 D4 07 07 1A 00 07 07 19 fffff.c@........
OrderID= 172 name= THERESE TUTTLE sum= 152.95 date= 2004 / 7 / 26
(00 07 07 19 bytes are not used and is ballast.)

00000000: 0C 00 FF FE 4D 00 45 00 4C 00 41 00 4E 00 49 00 ....M.E.L.A.N.I.
00000010: 45 00 20 00 4B 00 49 00 52 00 4B 00 00 00 00 00 E. .K.I.R.K.....
00000020: 00 20 64 40 D4 07 09 02 00 02 . d@......
OrderID= 286 name= MELANIE KIRK sum= 161.0 date= 2004 / 9 / 2
(00 02 are not used.)

After close examination, we can see, that the noise at the end of tail is just left from previous encryption!
Here are two subsequent buffers:
00000000: 10 00 FF FE 42 00 4F 00 4E 00 4E 00 49 00 45 00 ....B.O.N.N.I.E.
00000010: 20 00 47 00 4F 00 4C 00 44 00 53 00 54 00 45 00 .G.O.L.D.S.T.E.
00000020: 49 00 4E 00 9A 99 99 99 99 79 46 40 D4 07 07 19 I.N......yF@....
OrderID= 171 name= BONNIE GOLDSTEIN sum= 44.95 date= 2004 / 7 / 25
00000000: 0E 00 FF FE 54 00 48 00 45 00 52 00 45 00 53 00 ....T.H.E.R.E.S.
00000010: 45 00 20 00 54 00 55 00 54 00 54 00 4C 00 45 00 E. .T.U.T.T.L.E.
00000020: 66 66 66 66 66 1E 63 40 D4 07 07 1A 00 07 07 19 fffff.c@........
OrderID= 172 name= THERESE TUTTLE sum= 152.95 date= 2004 / 7 / 26
(The last 07 07 19 bytes are copied from the previous plaintext buffer.)
Another two subsequent buffers:
00000000: 0D 00 FF FE 4C 00 4F 00 52 00 45 00 4E 00 45 00 ....L.O.R.E.N.E.
00000010: 20 00 4F 00 54 00 4F 00 4F 00 4C 00 45 00 CD CC .O.T.O.O.L.E...
00000020: CC CC CC 3C 5E 40 D4 07 09 02 ...<^@....
OrderID= 285 name= LORENE OTOOLE sum= 120.95 date= 2004 / 9 / 2
00000000: 0C 00 FF FE 4D 00 45 00 4C 00 41 00 4E 00 49 00 ....M.E.L.A.N.I.
00000010: 45 00 20 00 4B 00 49 00 52 00 4B 00 00 00 00 00 E. .K.I.R.K.....
00000020: 00 20 64 40 D4 07 09 02 00 02 . d@......
OrderID= 286 name= MELANIE KIRK sum= 161.0 date= 2004 / 9 / 2
The last 02 byte has been copied from the previous plaintext buffer.
It’s possible if the buffer used while encrypting is global and/or isn’t clearing before each encryption. The
final buffer size is also chaotic, nevertheless, the bug left uncaught because it doesn’t affect decrypting
process, which just ignores noise at the end. This is common mistake. It’s been present in OpenSSL
(Heartbleed bug).
8.9.9 Conclusion
Summary: every practicing reverse engineer should be familiar with major crypto algorithms and also
major cryptographical modes. Some books about it: 12.1.10 on page 1000.
Encrypted database contents has been artificially constructed by me for the sake of demonstration. I’ve
got most popular USA names and surnames from there: http://stackoverflow.com/questions/1803628/
raw-list-of-person-names, and combined them randomly. Dates and sums were also generated ran-
domly.
All files used in this part are here: https://beginners.re/current-tree/examples/encrypted_DB1.
Nevertheless, many features like these I’ve observed in real-world software applications. This example is
based on them.
850
8.9.10 Post Scriptum: brute-forcing IV
The case you have just seen has been artificially constructed, but is based on a real application I’ve reverse
engineered. When I’ve been working on it, I first noticed that IV has been generating using some 32-bit
number, and I wasn’t able to find a link between this value and OrderID. So I prepared to use brute-force,
which is indeed possible here.
It’s not a problem to enumerate all 32-bit values and try each as a base for IV. Then you decrypt the first
16-byte block and check for zero bytes, which are always at fixed places.
8.10 Overclocking Cointerra Bitcoin miner

There was Cointerra Bitcoin miner, looking like that:
Figure 8.14: Board
And there was also (possibly leaked) utility24 which can set clock rate for the board. It runs on additional
BeagleBone Linux ARM board (small board at bottom of the picture).
And the author was once asked, is it possible to hack this utility to see, which frequency can be set and
which are not. And it is possible to tweak it?
The utility must be executed like that: ./cointool-overclock 0 0 900, where 900 is frequency in MHz.
If the frequency is too high, utility will print “Error with arguments” and exit.
This is a fragment of code around reference to “Error with arguments” text string:
...
24 Can be downloaded here: https://beginners.re/current-tree/examples/bitcoin_miner/files/cointool-overclock
851
.text:0000ABC4 STR R3, [R11,#var_28]
.text:0000ABC8 MOV R3, #optind
.text:0000ABD0 LDR R3, [R3]
.text:0000ABD4 ADD R3, R3, #1
.text:0000ABD8 MOV R3, R3,LSL#2
.text:0000ABDC LDR R2, [R11,#argv]
.text:0000ABE0 ADD R3, R2, R3
.text:0000ABE4 LDR R3, [R3]
.text:0000ABE8 MOV R0, R3 ; nptr
.text:0000ABEC MOV R1, #0 ; endptr
.text:0000ABF0 MOV R2, #0 ; base
.text:0000ABF4 BL strtoll
.text:0000ABF8 MOV R2, R0
.text:0000ABFC MOV R3, R1
.text:0000AC00 MOV R3, R2
.text:0000AC04 STR R3, [R11,#var_2C]
.text:0000AC08 MOV R3, #optind
.text:0000AC10 LDR R3, [R3]
.text:0000AC14 ADD R3, R3, #2
.text:0000AC18 MOV R3, R3,LSL#2
.text:0000AC1C LDR R2, [R11,#argv]
.text:0000AC20 ADD R3, R2, R3
.text:0000AC24 LDR R3, [R3]
.text:0000AC28 MOV R0, R3 ; nptr
.text:0000AC2C MOV R1, #0 ; endptr
.text:0000AC30 MOV R2, #0 ; base
.text:0000AC34 BL strtoll
.text:0000AC3C MOV R3, R1
.text:0000AC44 STR R3, [R11,#third_argument]
.text:0000AC48 LDR R3, [R11,#var_28]
.text:0000AC4C CMP R3, #0
.text:0000AC50 BLT errors_with_arguments
.text:0000AC54 LDR R3, [R11,#var_28]
.text:0000AC58 CMP R3, #1
.text:0000AC5C BGT errors_with_arguments
.text:0000AC60 LDR R3, [R11,#var_2C]
.text:0000AC68 BLT errors_with_arguments
.text:0000AC6C LDR R3, [R11,#var_2C]
.text:0000AC74 BGT errors_with_arguments
.text:0000AC78 LDR R3, [R11,#third_argument]
.text:0000AC7C CMP R3, #0x31
.text:0000AC80 BLE errors_with_arguments
.text:0000AC88 MOV R3, #950
.text:0000AC8C CMP R2, R3
.text:0000AC90 BGT errors_with_arguments
.text:0000AC98 MOV R3, #0x51EB851F
.text:0000ACA0 SMULL R1, R3, R3, R2
.text:0000ACA4 MOV R1, R3,ASR#4
.text:0000ACA8 MOV R3, R2,ASR#31
.text:0000ACAC RSB R3, R3, R1
.text:0000ACB0 MOV R1, #50
.text:0000ACB4 MUL R3, R1, R3
.text:0000ACB8 RSB R3, R3, R2
.text:0000ACBC CMP R3, #0
.text:0000ACC0 BEQ loc_ACEC
.text:0000ACC4
.text:0000ACC4 errors_with_arguments
.text:0000ACC4
.text:0000ACC4 LDR R3, [R11,#argv]
.text:0000ACC8 LDR R3, [R3]
.text:0000ACCC MOV R0, R3 ; path
.text:0000ACD0 BL __xpg_basename
.text:0000ACD4 MOV R3, R0
852
.text:0000ACD8 MOV R0, #aSErrorWithArgu ; format
.text:0000ACE0 MOV R1, R3
.text:0000ACE4 BL printf
.text:0000ACE8 B loc_ADD4
.text:0000ACEC ; ------------------------------------------------------------
.text:0000ACEC
.text:0000ACEC loc_ACEC ; CODE XREF: main+66C
.text:0000ACEC LDR R2, [R11,#third_argument]
.text:0000ACF0 MOV R3, #499
.text:0000ACF4 CMP R2, R3
.text:0000ACF8 BGT loc_AD08
.text:0000ACFC MOV R3, #0x64
.text:0000AD00 STR R3, [R11,#unk_constant]
.text:0000AD04 B jump_to_write_power
.text:0000AD08 ; ------------------------------------------------------------
.text:0000AD08
.text:0000AD08 loc_AD08 ; CODE XREF: main+6A4
.text:0000AD08 LDR R2, [R11,#third_argument]
.text:0000AD0C MOV R3, #799
.text:0000AD10 CMP R2, R3
.text:0000AD14 BGT loc_AD24
.text:0000AD18 MOV R3, #0x5F
.text:0000AD1C STR R3, [R11,#unk_constant]
.text:0000AD24 ; ------------------------------------------------------------
.text:0000AD24
.text:0000AD24 loc_AD24 ; CODE XREF: main+6C0
.text:0000AD28 MOV R3, #899
.text:0000AD2C CMP R2, R3
.text:0000AD30 BGT loc_AD40
.text:0000AD34 MOV R3, #0x5A
.text:0000AD3C B jump_to_write_power
.text:0000AD40 ; ------------------------------------------------------------
.text:0000AD40
.text:0000AD40 loc_AD40 ; CODE XREF: main+6DC
.text:0000AD44 MOV R3, #999
.text:0000AD4C BGT loc_AD5C
.text:0000AD50 MOV R3, #0x55
.text:0000AD5C ; ------------------------------------------------------------
.text:0000AD5C
.text:0000AD5C loc_AD5C ; CODE XREF: main+6F8
.text:0000AD5C LDR R2, [R11,#third_argument]
.text:0000AD60 MOV R3, #1099
.text:0000AD68 BGT jump_to_write_power
.text:0000AD6C MOV R3, #0x50
.text:0000AD74
.text:0000AD74 jump_to_write_power ; CODE XREF: main+6B0
.text:0000AD74 ; main+6CC ...
.text:0000AD74 LDR R3, [R11,#var_28]
.text:0000AD78 UXTB R1, R3
.text:0000AD7C LDR R3, [R11,#var_2C]
.text:0000AD84 LDR R3, [R11,#unk_constant]
.text:0000AD8C LDR R0, [R11,#third_argument]
.text:0000AD90 UXTH R0, R0
.text:0000AD94 STR R0, [SP,#0x44+var_44]
.text:0000AD98 LDR R0, [R11,#var_24]
.text:0000AD9C BL write_power
.text:0000ADA0 LDR R0, [R11,#var_24]
.text:0000ADA4 MOV R1, #0x5A
.text:0000ADA8 BL read_loop
853
.text:0000ADAC B loc_ADD4
...
.rodata:0000B378 aSErrorWithArgu DCB "%s: Error with arguments",0xA,0 ; DATA XREF: main+684
...
Function names were present in debugging information of the original binary, like write_power, read_loop.
But labels inside functions were named by me.
optind name looks familiar. It is from getopt *NIX library intended for command-line parsing—well, this is
exactly what happens inside. Then, the 3rd argument (where frequency value is to be passed) is converted
from a string to a number using a call to strtoll() function.
The value is then checked against various constants. At 0xACEC, it’s checked, if it is lesser or equal to
499, and if it is so, 0x64 is to be passed to write_power() function (which sends a command through
USB using send_msg()). If it is greater than 499, jump to 0xAD08 is occurred.
At 0xAD08 it’s checked, if it’s lesser or equal to 799. 0x5F is then passed to write_power() function in
case of success.
There are more checks: for 899 at 0xAD24, for 0x999 at 0xAD40 and finally, for 1099 at 0xAD5C. If the
input frequency is lesser or equal to 1099, 0x50 will be passed (at 0xAD6C) to write_power() function.
And there is some kind of bug. If the value is still greater than 1099, the value itself is passed into
write_power() function. Oh, it’s not a bug, because we can’t get here: value is checked first against 950
at 0xAC88, and if it is greater, error message will be displayed and the utility will finish.
Now the table between frequency in MHz and value passed to write_power() function:
MHz hexadecimal decimal

499MHz 0x64 100
799MHz 0x5f 95
899MHz 0x5a 90
999MHz 0x55 85
1099MHz 0x50 80
As it seems, a value passed to the board is gradually decreasing during frequency increasing.
Now we see that value of 950MHz is a hardcoded limit, at least in this utility. Can we trick it?
Let’s back to this piece of code:
.text:0000AC88 MOV R3, #950
.text:0000AC8C CMP R2, R3
.text:0000AC90 BGT errors_with_arguments ; I've patched here to 00 00 00 00
We must disable BGT branch instruction at 0xAC90 somehow. And this is ARM in ARM mode, because, as
we see, all addresses are increasing by 4, i.e., each instruction has size of 4 bytes. NOP (no operation)
instruction in ARM mode is just four zero bytes: 00 00 00 00. So by writing four zeros at 0xAC90 address
(or physical offset in file 0x2C90) we can disable the check.
Now it’s possible to set frequencies up to 1050MHz. Even more is possible, but due to the bug, if input
value is greater than 1099, a value as is in MHz will be passed to the board, which is incorrect.
I didn’t go further, but if I had to, I would try to decrease a value which is passed to write_power()
function.
Now the scary piece of code which I skipped at first:
.text:0000AC98 MOV R3, #0x51EB851F
.text:0000ACA0 SMULL R1, R3, R3, R2 ; R3=3rg_arg/3.125
.text:0000ACA4 MOV R1, R3,ASR#4 ; R1=R3/16=3rg_arg/50
.text:0000ACA8 MOV R3, R2,ASR#31 ; R3=MSB(3rg_arg)
.text:0000ACAC RSB R3, R3, R1 ; R3=3rd_arg/50
.text:0000ACB0 MOV R1, #50
.text:0000ACB4 MUL R3, R1, R3 ; R3=50*(3rd_arg/50)
854
.text:0000ACB8 RSB R3, R3, R2
.text:0000ACBC CMP R3, #0
.text:0000ACC0 BEQ loc_ACEC
.text:0000ACC4
.text:0000ACC4 errors_with_arguments
Division via multiplication is used here, and constant is 0x51EB851F. I wrote a simple programmer’s cal-
culator25 for myself. And I have there a feature to calculate modulo inverse.
modinv32(0x51EB851F)
Warning, result is not integer: 3.125000
(unsigned) dec: 3 hex: 0x3 bin: 11
That means that SMULL instruction at 0xACA0 is basically divides 3rd argument by 3.125. In fact, all
modinv32() function in my calculator does, is this:
1 232
input
=
232
input
Then there are additional shifts and now we see than 3rg argument is just divided by 50. And then it’s
multiplied by 50 again. Why? This is simplest check, if the input value is can be divided by 50 evenly. If
the value of this expression is non-zero, x can’t be divided by 50 evenly:
x
x − (( ) ⋅ 50)
50
This is in fact simple way to calculate remainder of division.

And then, if the remainder is non-zero, error message is displayed. So this utility takes frequency values
in form like 850, 900, 950, 1000, etc., but not 855 or 911.
That’s it! If you do something like that, please be warned that you may damage your board, just as in
case of overclocking other devices like CPUs, GPU26 s, etc. If you have a Cointerra board, do this on your
own risk!
8.11 Breaking simple executable cryptor

I’ve got an executable file which is encrypted by relatively simple encryption. Here is it (only executable
section is left here).
First, all encryption function does is just adds number of position in buffer to the byte. Here is how this
can be encoded in Python:
Listing 8.9: Python script

#!/usr/bin/env python
def e(i, k):
return chr ((ord(i)+k) % 256)
def encrypt(buf):
return e(buf[0], 0)+ e(buf[1], 1)+ e(buf[2], 2) + e(buf[3], 3)+ e(buf[4], 4)+ e(buf[5], 5)+⤦
Ç e(buf[6], 6)+ e(buf[7], 7)+
e(buf[8], 8)+ e(buf[9], 9)+ e(buf[10], 10)+ e(buf[11], 11)+ e(buf[12], 12)+ e(buf⤦
Ç [13], 13)+ e(buf[14], 14)+ e(buf[15], 15)
Hence, if you encrypt buffer with 16 zeros, you’ll get 0, 1, 2, 3 ... 12, 13, 14, 15.
Propagating Cipher Block Chaining (PCBC) is also used, here is how it works:
25 https://yurichev.com/progcalc/
26 Graphics Processing Unit
855
Figure 8.15: Propagating Cipher Block Chaining encryption (image is taken from Wikipedia article)
The problem is that it’s too boring to recover IV (Initialization Vector) each time. Brute-force is also not
an option, because IV is too long (16 bytes). Let’s see, if it’s possible to recover IV for arbitrary encrypted
executable file?
Let’s try simple frequency analysis. This is 32-bit x86 executable code, so let’s gather statistics about most
frequent bytes and opcodes. I tried huge oracle.exe file from Oracle RDBMS version 11.2 for windows x86
and I’ve found that the most frequent byte (no surprise) is zero ( 10%). The next most frequent byte is
(again, no surprise) 0xFF ( 5%). The next is 0x8B ( 5%).
0x8B is opcode for MOV, this is indeed one of the most busy x86 instructions. Now what about popularity
of zero byte? If compiler needs to encode value bigger than 127, it has to use 32-bit displacement instead
of 8-bit one, but large values are very rare, so it is padded by zeros. This is at least in LEA, MOV, PUSH,
CALL.
For example:
8D B0 28 01 00 00 lea esi, [eax+128h]
8D BF 40 38 00 00 lea edi, [edi+3840h]
Displacements bigger than 127 are very popular, but they are rarely exceeds 0x10000 (indeed, such large
memory buffers/structures are also rare).
Same story with MOV, large constants are rare, the most heavily used are 0, 1, 10, 100, 2n , and so on.
Compiler has to pad small constants by zeros to represent them as 32-bit values:
BF 02 00 00 00 mov edi, 2
BF 01 00 00 00 mov edi, 1
Now about 00 and FF bytes combined: jumps (including conditional) and calls can pass execution flow
forward or backwards, but very often, within the limits of the current executable module. If forward,
displacement is not very big and also padded with zeros. If backwards, displacement is represented as
negative value, so padded with FF bytes. For example, transfer execution flow forward:
E8 43 0C 00 00 call _function1
E8 5C 00 00 00 call _function2
0F 84 F0 0A 00 00 jz loc_4F09A0
0F 84 EB 00 00 00 jz loc_4EFBB8
Backwards:
E8 79 0C FE FF call _function1
E8 F4 16 FF FF call _function2
0F 84 F8 FB FF FF jz loc_8212BC
0F 84 06 FD FF FF jz loc_FF1E7D
FF byte is also very often occurred in negative displacements like these:
856
8D 85 1E FF FF FF lea eax, [ebp-0E2h]
8D 95 F8 5C FF FF lea edx, [ebp-0A308h]
So far so good. Now we have to try various 16-byte keys, decrypt executable section and measure how
often 00, FF and 8B bytes are occurred. Let’s also keep in sight how PCBC decryption works:
Figure 8.16: Propagating Cipher Block Chaining decryption (image is taken from Wikipedia article)
The good news is that we don’t really have to decrypt whole piece of data, but only slice by slice, this is
exactly how I did in my previous example: 9.1.5 on page 924.
Now I’m trying all possible bytes (0..255) for each byte in key and just pick the byte producing maximal
amount of 00/FF/8B bytes in a decrypted slice:
import sys, hexdump, array, string, operator
KEY_LEN=16
def chunks(l, n):

# split n by l-byte chunks
# https://stackoverflow.com/q/312443
n = max(1, n)
return [l[i:i + n] for i in range(0, len(l), n)]
def read_file(fname):
file=open(fname, mode='rb')
content=file.read()
file.close()
return content
def decrypt_byte (c, key):

return chr((ord(c)-key) % 256)
def XOR_PCBC_step (IV, buf, k):

prev=IV
rt=""
for c in buf:
new_c=decrypt_byte(c, k)
plain=chr(ord(new_c)ôrd(prev))
prev=chr(ord(c)ôrd(plain))
rt=rt+plain
return rt
each_Nth_byte=[""]*KEY_LEN
content=read_file(sys.argv[1])
857
# split input by 16-byte chunks:
all_chunks=chunks(content, KEY_LEN)
for c in all_chunks:
for i in range(KEY_LEN):
each_Nth_byte[i]=each_Nth_byte[i] + c[i]
# try each byte of key

for N in range(KEY_LEN):
print "N=", N
stat={}
for i in range(256):
tmp_key=chr(i)
tmp=XOR_PCBC_step(tmp_key,each_Nth_byte[N], N)
# count 0, FFs and 8Bs in decrypted buffer:
important_bytes=tmp.count('\x00')+tmp.count('\xFF')+tmp.count('\x8B')
stat[i]=important_bytes
sorted_stat = sorted(stat.iteritems(), key=operator.itemgetter(1), reverse=True)
print sorted_stat[0]
(Source code can be downloaded here.)

I run it and here is a key for which 00/FF/8B bytes presence in decrypted buffer is maximal:
N= 0
(147, 1224)
N= 1
(94, 1327)
N= 2
(252, 1223)
N= 3
(218, 1266)
N= 4
(38, 1209)
N= 5
(192, 1378)
N= 6
(199, 1204)
N= 7
(213, 1332)
N= 8
(225, 1251)
N= 9
(112, 1223)
N= 10
(143, 1177)
N= 11
(108, 1286)
N= 12
(10, 1164)
N= 13
(3, 1271)
N= 14
(128, 1253)
N= 15
(232, 1330)
Let’s write decryption utility with the key we got:

import sys, hexdump, array
def xor_strings(s,t):
# https://en.wikipedia.org/wiki/XOR_cipher#Example_implementation
"""xor two strings together"""
return "".join(chr(ord(a)ôrd(b)) for a,b in zip(s,t))
IV=array.array('B', [147, 94, 252, 218, 38, 192, 199, 213, 225, 112, 143, 108, 10, 3, 128, ⤦
Ç 232]).tostring()
def chunks(l, n):
858
n = max(1, n)
return [l[i:i + n] for i in range(0, len(l), n)]
def read_file(fname):
file=open(fname, mode='rb')
content=file.read()
file.close()
return content
def decrypt_byte(i, k):

return chr ((ord(i)-k) % 256)
def decrypt(buf):
return "".join(decrypt_byte(buf[i], i) for i in range(16))
fout=open(sys.argv[2], mode='wb')
prev=IV
tmp=chunks(content, 16)
for c in tmp:
new_c=decrypt(c)
p=xor_strings (new_c, prev)
prev=xor_strings(c, p)
fout.write(p)
fout.close()
(Source code can be downloaded here.)

Let’s check resulting file:
$ objdump -b binary -m i386 -D decrypted.bin
...
5: 8b ff mov %edi,%edi
7: 55 push %ebp
8: 8b ec mov %esp,%ebp
a: 51 push %ecx
b: 53 push %ebx
c: 33 db xor %ebx,%ebx
e: 43 inc %ebx
f: 84 1d a0 e2 05 01 test %bl,0x105e2a0
15: 75 09 jne 0x20
17: ff 75 08 pushl 0x8(%ebp)
1a: ff 15 b0 13 00 01 call *0x10013b0
20: 6a 6c push $0x6c
22: ff 35 54 d0 01 01 pushl 0x101d054
28: ff 15 b4 13 00 01 call *0x10013b4
2e: 89 45 fc mov %eax,-0x4(%ebp)
31: 85 c0 test %eax,%eax
33: 0f 84 d9 00 00 00 je 0x112
39: 56 push %esi
3a: 57 push %edi
3b: 6a 00 push $0x0
3d: 50 push %eax
3e: ff 15 b8 13 00 01 call *0x10013b8
44: 8b 35 bc 13 00 01 mov 0x10013bc,%esi
4a: 8b f8 mov %eax,%edi
4c: a1 e0 e2 05 01 mov 0x105e2e0,%eax
51: 3b 05 e4 e2 05 01 cmp 0x105e2e4,%eax
57: 75 12 jne 0x6b
59: 53 push %ebx
5a: 6a 03 push $0x3
5c: 57 push %edi
5d: ff d6 call *%esi
...
Yes, this is seems correctly disassembled piece of x86 code. The whole decryped file can be downloaded
859
here.
In fact, this is text section from regedit.exe from Windows 7. But this example is based on a real case I
encountered, so just executable is different (and key), algorithm is the same.
8.11.1 Other ideas to consider

What if I would fail with such simple frequency analysis? There are other ideas on how to measure cor-
rectness of decrypted/decompressed x86 code:
• Many modern compilers aligns functions on 0x10 border. So the space left before is filled with NOPs
(0x90) or other NOP instructions with known opcodes: .1.7 on page 1024.
• Perhaps, the most frequent pattern in any assembly language is function call:
PUSH chain / CALL / ADD ESP, X. This sequence can easily detected and found. I’ve even gath-
ered statistics about average number of function arguments: 11.2 on page 980. (Hence, this is
average length of PUSH chain.)
Read more about incorrectly/correctly disassembled code: 5.11 on page 719.
8.12 SAP
8.12.1 About SAP client network traffic compression
(This article first appeared in my blog, in 13-Jul-2010.)
(Tracing the connection between the TDW_NOCOMPRESS SAPGUI27 environment variable and the pesky
annoying pop-up window and the actual data compression routine.)
It is known that the network traffic between SAPGUI and SAP is not encrypted by default, but compressed
(see here28 and here29 ).
It is also known that by setting the environment variable TDW_NOCOMPRESS to 1, it is possible to turn
the network packet compression off.
But you will see an annoying pop-up window that cannot be closed:
27 SAP GUI client
29 blog.yurichev.com
860
Figure 8.17: Screenshot
Let’s see if we can remove the window somehow.

But before this, let’s see what we already know.
First: we know that the environment variable TDW_NOCOMPRESS is checked somewhere inside the SAPGUI
client.
Second: a string like “data compression switched off” must be present somewhere in it.
With the help of the FAR file manager30 we can found that both of these strings are stored in the SAPguilib.dll
file.
So let’s open SAPguilib.dll in IDA and search for the TDW_NOCOMPRESS string. Yes, it is present and there
is only one reference to it.
We see the following fragment of code (all file offsets are valid for SAPGUI 720 win32, SAPguilib.dll file
version 7200,1,0,9009):
.text:6440D51B lea eax, [ebp+2108h+var_211C]
.text:6440D51E push eax ; int
.text:6440D51F push offset aTdw_nocompress ; "TDW_NOCOMPRESS"
.text:6440D524 mov byte ptr [edi+15h], 0
.text:6440D528 call chk_env
.text:6440D52D pop ecx
.text:6440D52E pop ecx
.text:6440D52F push offset byte_64443AF8
.text:6440D534 lea ecx, [ebp+2108h+var_211C]
; demangled name: int ATL::CStringT::Compare(char const *)const

.text:6440D537 call ds:mfc90_1603
.text:6440D53D test eax, eax
.text:6440D53F jz short loc_6440D55A
.text:6440D541 lea ecx, [ebp+2108h+var_211C]
; demangled name: const char* ATL::CSimpleStringT::operator PCXSTR

861
.text:6440D54A push eax ; Str
.text:6440D54B call ds:atoi
.text:6440D551 test eax, eax
.text:6440D553 setnz al
.text:6440D556 pop ecx
.text:6440D557 mov [edi+15h], al
The string returned by chk_env() via its second argument is then handled by the MFC string functions
and then atoi()31 is called. After that, the numerical value is stored in edi+15h.
Also take a look at the chk_env() function (we gave this name to it manually):
.text:64413F20 ; int __cdecl chk_env(char *VarName, int)
.text:64413F20 chk_env proc near
.text:64413F20
.text:64413F20 DstSize = dword ptr -0Ch
.text:64413F20 var_8 = dword ptr -8
.text:64413F20 DstBuf = dword ptr -4
.text:64413F20 VarName = dword ptr 8
.text:64413F20 arg_4 = dword ptr 0Ch
.text:64413F20
.text:64413F20 push ebp
.text:64413F21 mov ebp, esp
.text:64413F23 sub esp, 0Ch
.text:64413F26 mov [ebp+DstSize], 0
.text:64413F2D mov [ebp+DstBuf], 0
.text:64413F34 push offset unk_6444C88C
.text:64413F39 mov ecx, [ebp+arg_4]
; (demangled name) ATL::CStringT::operator=(char const *)

.text:64413F3C call ds:mfc90_820
.text:64413F42 mov eax, [ebp+VarName]
.text:64413F45 push eax ; VarName
.text:64413F46 mov ecx, [ebp+DstSize]
.text:64413F49 push ecx ; DstSize
.text:64413F4A mov edx, [ebp+DstBuf]
.text:64413F4D push edx ; DstBuf
.text:64413F4E lea eax, [ebp+DstSize]
.text:64413F51 push eax ; ReturnSize
.text:64413F52 call ds:getenv_s
.text:64413F58 add esp, 10h
.text:64413F5B mov [ebp+var_8], eax
.text:64413F5E cmp [ebp+var_8], 0
.text:64413F62 jz short loc_64413F68
.text:64413F64 xor eax, eax
.text:64413F66 jmp short loc_64413FBC
.text:64413F68
.text:64413F68 loc_64413F68:
.text:64413F68 cmp [ebp+DstSize], 0
.text:64413F6C jnz short loc_64413F72
.text:64413F6E xor eax, eax
.text:64413F70 jmp short loc_64413FBC
.text:64413F72
.text:64413F72 loc_64413F72:
.text:64413F72 mov ecx, [ebp+DstSize]
.text:64413F75 push ecx
.text:64413F76 mov ecx, [ebp+arg_4]
; demangled name: ATL::CSimpleStringT<char, 1>::Preallocate(int)

.text:64413F79 call ds:mfc90_2691
.text:64413F7F mov [ebp+DstBuf], eax
.text:64413F82 mov edx, [ebp+VarName]
.text:64413F85 push edx ; VarName
.text:64413F86 mov eax, [ebp+DstSize]
.text:64413F89 push eax ; DstSize
.text:64413F8A mov ecx, [ebp+DstBuf]
.text:64413F8D push ecx ; DstBuf
.text:64413F8E lea edx, [ebp+DstSize]
.text:64413F91 push edx ; ReturnSize
31 standard C library function that converts the digits in a string to a number
862
.text:64413F92 call ds:getenv_s
.text:64413F98 add esp, 10h
.text:64413F9B mov [ebp+var_8], eax
.text:64413F9E push 0FFFFFFFFh
.text:64413FA0 mov ecx, [ebp+arg_4]
; demangled name: ATL::CSimpleStringT::ReleaseBuffer(int)

.text:64413FA3 call ds:mfc90_5835
.text:64413FA9 cmp [ebp+var_8], 0
.text:64413FAD jz short loc_64413FB3
.text:64413FAF xor eax, eax
.text:64413FB1 jmp short loc_64413FBC
.text:64413FB3
.text:64413FB3 loc_64413FB3:
.text:64413FB3 mov ecx, [ebp+arg_4]

.text:64413FB6 call ds:mfc90_910
.text:64413FBC
.text:64413FBC loc_64413FBC:
.text:64413FBC
.text:64413FBC mov esp, ebp
.text:64413FBE pop ebp
.text:64413FBF retn
.text:64413FBF chk_env endp
Yes. The getenv_s()32

function is a Microsoft security-enhanced version of getenv()33 .
There are also some MFC string manipulations.
Lots of other environment variables are checked as well. Here is a list of all variables that are being
checked and what SAPGUI would write to its trace log when logging is turned on:
DPTRACE “GUI-OPTION: Trace set to %d”
TDW_HEXDUMP “GUI-OPTION: Hexdump enabled”
TDW_WORKDIR “GUI-OPTION: working directory ‘%s’́’
TDW_SPLASHSRCEENOFF “GUI-OPTION: Splash Screen Off”
“GUI-OPTION: Splash Screen On”
TDW_REPLYTIMEOUT “GUI-OPTION: reply timeout %d milliseconds”
TDW_PLAYBACKTIMEOUT “GUI-OPTION: PlaybackTimeout set to %d milliseconds”
TDW_NOCOMPRESS “GUI-OPTION: no compression read”
TDW_EXPERT “GUI-OPTION: expert mode”
TDW_PLAYBACKPROGRESS “GUI-OPTION: PlaybackProgress”
TDW_PLAYBACKNETTRAFFIC “GUI-OPTION: PlaybackNetTraffic”
TDW_PLAYLOG “GUI-OPTION: /PlayLog is YES, file %s”
TDW_PLAYTIME “GUI-OPTION: /PlayTime set to %d milliseconds”
TDW_LOGFILE “GUI-OPTION: TDW_LOGFILE ‘%s’́’
TDW_WAN “GUI-OPTION: WAN - low speed connection enabled”
TDW_FULLMENU “GUI-OPTION: FullMenu enabled”
SAP_CP / SAP_CODEPAGE “GUI-OPTION: SAP_CODEPAGE ‘%d’́’
UPDOWNLOAD_CP “GUI-OPTION: UPDOWNLOAD_CP ‘%d’́’
SNC_PARTNERNAME “GUI-OPTION: SNC name ‘%s’́’
SNC_QOP “GUI-OPTION: SNC_QOP ‘%s’́’
SNC_LIB “GUI-OPTION: SNC is set to: %s”
SAPGUI_INPLACE “GUI-OPTION: environment variable SAPGUI_INPLACE is on”
The settings for each variable are written in the array via a pointer in the EDI register. EDI is set before
the function call:
.text:6440EE00 lea edi, [ebp+2884h+var_2884] ; options here like +0x15...
.text:6440EE03 lea ecx, [esi+24h]
.text:6440EE06 call load_command_line
.text:6440EE0B mov edi, eax
.text:6440EE0D xor ebx, ebx
.text:6440EE0F cmp edi, ebx
.text:6440EE11 jz short loc_6440EE42
.text:6440EE13 push edi
32 MSDN
33 Standard C library returning environment variable
863
.text:6440EE14 push offset aSapguiStoppedA ; "Sapgui stopped after
commandline interp"...
.text:6440EE19 push dword_644F93E8
.text:6440EE1F call FEWTraceError
Now, can we find the data record mode switched on string?

Yes, and the only reference is in
CDwsGui::PrepareInfoWindow().
How do we get know the class/method names? There are a lot of special debugging calls that write to the
log files, like:
.text:64405160 push dword ptr [esi+2854h]
.text:64405166 push offset aCdwsguiPrepare ;
"\nCDwsGui::PrepareInfoWindow: sapgui env"...
.text:6440516B push dword ptr [esi+2848h]
.text:64405171 call dbg
.text:64405176 add esp, 0Ch
…or:
.text:6440237A push eax
.text:6440237B push offset aCclientStart_6 ; "CClient::Start: set shortcut
user to '%"...
.text:64402380 push dword ptr [edi+4]
It is very useful.
So let’s see the contents of this pesky annoying pop-up window’s function:
.text:64404F4F CDwsGui__PrepareInfoWindow proc near
.text:64404F4F
.text:64404F4F pvParam = byte ptr -3Ch
.text:64404F4F var_38 = dword ptr -38h
.text:64404F4F rc = tagRECT ptr -2Ch
.text:64404F4F cy = dword ptr -1Ch
.text:64404F4F h = dword ptr -18h
.text:64404F4F var_4 = dword ptr -4
.text:64404F4F
.text:64404F4F push 30h
.text:64404F51 mov eax, offset loc_64438E00
.text:64404F56 call __EH_prolog3
.text:64404F5B mov esi, ecx ; ECX is pointer to object
.text:64404F5D xor ebx, ebx
.text:64404F5F lea ecx, [ebp+var_14]
.text:64404F62 mov [ebp+var_10], ebx
; demangled name: ATL::CStringT(void)

.text:64404F65 call ds:mfc90_316
.text:64404F6B mov [ebp+var_4], ebx
.text:64404F6E lea edi, [esi+2854h]
.text:64404F74 push offset aEnvironmentInf ; "Environment information:\n"
.text:64404F79 mov ecx, edi
; demangled name: ATL::CStringT::operator=(char const *)

.text:64404F7B call ds:mfc90_820
.text:64404F81 cmp [esi+38h], ebx
.text:64404F84 mov ebx, ds:mfc90_2539
.text:64404F8A jbe short loc_64404FA9
.text:64404F8C push dword ptr [esi+34h]
.text:64404F8F lea eax, [ebp+var_14]
.text:64404F92 push offset aWorkingDirecto ; "working directory: '%s'\n"
.text:64404F97 push eax
; demangled name: ATL::CStringT::Format(char const *,...)
864
.text:64404F98 call ebx ; mfc90_2539
.text:64404F9A add esp, 0Ch
.text:64404F9D lea eax, [ebp+var_14]
.text:64404FA0 push eax
.text:64404FA1 mov ecx, edi
; demangled name: ATL::CStringT::operator+=(class ATL::CSimpleStringT<char, 1> const &)

.text:64404FA3 call ds:mfc90_941
.text:64404FA9
.text:64404FA9 loc_64404FA9:
.text:64404FA9 mov eax, [esi+38h]
.text:64404FAC test eax, eax
.text:64404FAE jbe short loc_64404FD3
.text:64404FB0 push eax
.text:64404FB1 lea eax, [ebp+var_14]
.text:64404FB4 push offset aTraceLevelDAct ; "trace level %d activated\n"
.text:64404FB9 push eax
; demangled name: ATL::CStringT::Format(char const *,...)

.text:64404FBA call ebx ; mfc90_2539
.text:64404FBC add esp, 0Ch
.text:64404FBF lea eax, [ebp+var_14]
.text:64404FC2 push eax
.text:64404FC3 mov ecx, edi
; demangled name: ATL::CStringT::operator+=(class ATL::CSimpleStringT<char, 1> const &)

.text:64404FC5 call ds:mfc90_941
.text:64404FCB xor ebx, ebx
.text:64404FCD inc ebx
.text:64404FCE mov [ebp+var_10], ebx
.text:64404FD1 jmp short loc_64404FD6
.text:64404FD3
.text:64404FD3 loc_64404FD3:
.text:64404FD3 xor ebx, ebx
.text:64404FD5 inc ebx
.text:64404FD6
.text:64404FD6 loc_64404FD6:
.text:64404FD6 cmp [esi+38h], ebx
.text:64404FD9 jbe short loc_64404FF1
.text:64404FDB cmp dword ptr [esi+2978h], 0
.text:64404FE2 jz short loc_64404FF1
.text:64404FE4 push offset aHexdumpInTrace ; "hexdump in trace activated\n"
.text:64404FE9 mov ecx, edi
; demangled name: ATL::CStringT::operator+=(char const *)

.text:64404FEB call ds:mfc90_945
.text:64404FF1
.text:64404FF1 loc_64404FF1:
.text:64404FF1
.text:64404FF1 cmp byte ptr [esi+78h], 0
.text:64404FF5 jz short loc_64405007
.text:64404FF7 push offset aLoggingActivat ; "logging activated\n"
.text:64404FFC mov ecx, edi

.text:64404FFE call ds:mfc90_945
.text:64405004 mov [ebp+var_10], ebx
.text:64405007
.text:64405007 loc_64405007:
.text:64405007 cmp byte ptr [esi+3Dh], 0
.text:6440500B jz short bypass
.text:6440500D push offset aDataCompressio ;
"data compression switched off\n"

.text:64405014 call ds:mfc90_945
.text:6440501A mov [ebp+var_10], ebx
.text:6440501D
.text:6440501D bypass:
865
.text:6440501D mov eax, [esi+20h]
.text:64405020 test eax, eax
.text:64405022 jz short loc_6440503A
.text:64405024 cmp dword ptr [eax+28h], 0
.text:64405028 jz short loc_6440503A
.text:6440502A push offset aDataRecordMode ; "data record mode switched on\n"
.text:6440502F mov ecx, edi

.text:64405037 mov [ebp+var_10], ebx
.text:6440503A
.text:6440503A loc_6440503A:
.text:6440503A
.text:6440503A mov ecx, edi
.text:6440503C cmp [ebp+var_10], ebx
.text:6440503F jnz loc_64405142
.text:64405045 push offset aForMaximumData ;
"\nFor maximum data security delete\nthe s"...

.text:6440504A call ds:mfc90_945
.text:64405050 xor edi, edi
.text:64405052 push edi ; fWinIni
.text:64405053 lea eax, [ebp+pvParam]
.text:64405056 push eax ; pvParam
.text:64405057 push edi ; uiParam
.text:64405058 push 30h ; uiAction
.text:6440505A call ds:SystemParametersInfoA
.text:64405063 cmp eax, 1600
.text:64405068 jle short loc_64405072
.text:6440506A cdq
.text:6440506B sub eax, edx
.text:6440506D sar eax, 1
.text:6440506F mov [ebp+var_34], eax
.text:64405072
.text:64405072 loc_64405072:
.text:64405072 push edi ; hWnd
.text:64405073 mov [ebp+cy], 0A0h
.text:6440507A call ds:GetDC
.text:64405080 mov [ebp+var_10], eax
.text:64405083 mov ebx, 12Ch
.text:64405088 cmp eax, edi
.text:6440508A jz loc_64405113
.text:64405090 push 11h ; i
.text:64405092 call ds:GetStockObject
.text:64405098 mov edi, ds:SelectObject
.text:6440509E push eax ; h
.text:6440509F push [ebp+var_10] ; hdc
.text:644050A2 call edi ; SelectObject
.text:644050A4 and [ebp+rc.left], 0
.text:644050A8 and [ebp+rc.top], 0
.text:644050AC mov [ebp+h], eax
.text:644050AF push 401h ; format
.text:644050B4 lea eax, [ebp+rc]
.text:644050B7 push eax ; lprc
.text:644050B8 lea ecx, [esi+2854h]
.text:644050BE mov [ebp+rc.right], ebx
.text:644050C1 mov [ebp+rc.bottom], 0B4h
; demangled name: ATL::CSimpleStringT::GetLength(void)

.text:644050C8 call ds:mfc90_3178
.text:644050CE push eax ; cchText
.text:644050CF lea ecx, [esi+2854h]

.text:644050DB push eax ; lpchText
.text:644050DC push [ebp+var_10] ; hdc
866
.text:644050DF call ds:DrawTextA
.text:644050E5 push 4 ; nIndex
.text:644050E7 call ds:GetSystemMetrics
.text:644050ED mov ecx, [ebp+rc.bottom]
.text:644050F0 sub ecx, [ebp+rc.top]
.text:644050F3 cmp [ebp+h], 0
.text:644050F7 lea eax, [eax+ecx+28h]
.text:644050FB mov [ebp+cy], eax
.text:644050FE jz short loc_64405108
.text:64405100 push [ebp+h] ; h
.text:64405103 push [ebp+var_10] ; hdc
.text:64405106 call edi ; SelectObject
.text:64405108
.text:64405108 loc_64405108:
.text:64405108 push [ebp+var_10] ; hDC
.text:6440510B push 0 ; hWnd
.text:6440510D call ds:ReleaseDC
.text:64405113
.text:64405113 loc_64405113:
.text:64405116 push 80h ; uFlags
.text:6440511B push [ebp+cy] ; cy
.text:6440511E inc eax
.text:6440511F push ebx ; cx
.text:64405120 push eax ; Y
.text:64405124 add eax, 0FFFFFED4h
.text:64405129 cdq
.text:6440512A sub eax, edx
.text:6440512C sar eax, 1
.text:6440512E push eax ; X
.text:6440512F push 0 ; hWndInsertAfter
.text:64405131 push dword ptr [esi+285Ch] ; hWnd
.text:64405137 call ds:SetWindowPos
.text:6440513D xor ebx, ebx
.text:6440513F inc ebx
.text:64405140 jmp short loc_6440514D
.text:64405142
.text:64405142 loc_64405142:
.text:64405142 push offset byte_64443AF8
; demangled name: ATL::CStringT::operator=(char const *)

.text:6440514D
.text:6440514D loc_6440514D:
.text:6440514D cmp dword_6450B970, ebx
.text:64405153 jl short loc_64405188
.text:64405155 call sub_6441C910
.text:6440515A mov dword_644F858C, ebx
.text:64405160 push dword ptr [esi+2854h]
.text:64405166 push offset aCdwsguiPrepare ;
"\nCDwsGui::PrepareInfoWindow: sapgui env"...
.text:6440516B push dword ptr [esi+2848h]
.text:64405179 mov dword_644F858C, 2
.text:64405183 call sub_6441C920
.text:64405188
.text:64405188 loc_64405188:
.text:64405188 or [ebp+var_4], 0FFFFFFFFh
.text:6440518C lea ecx, [ebp+var_14]
; demangled name: ATL::CStringT:: CStringT()

.text:6440518F call ds:mfc90_601
.text:64405195 call __EH_epilog3
.text:6440519A retn
.text:6440519A CDwsGui__PrepareInfoWindow endp
At the start of the function ECX has a pointer to the object (since it is a thiscall ( 3.21.1 on page 542)-type
of function). In our case, the object obviously has class type of CDwsGui. Depending on the option turned
867
on in the object, a specific message part is to be concatenated with the resulting message.
If the value at address this+0x3D is not zero, the compression is off:
.text:64405007 loc_64405007:
.text:64405007 cmp byte ptr [esi+3Dh], 0
.text:6440500B jz short bypass
.text:6440500D push offset aDataCompressio ;
"data compression switched off\n"

.text:6440501A mov [ebp+var_10], ebx
.text:6440501D
.text:6440501D bypass:
It is interesting that finally the var_10 variable state defines whether the message is to be shown at all:
.text:6440503C cmp [ebp+var_10], ebx

.text:6440503F jnz exit ; bypass drawing
; add strings "For maximum data security delete" / "the setting(s) as soon as possible !":
.text:64405045 push offset aForMaximumData ;

"\nFor maximum data security delete\nthe s"...
.text:6440504A call ds:mfc90_945 ; ATL::CStringT::operator+=(char const *)
.text:64405050 xor edi, edi
.text:64405052 push edi ; fWinIni
.text:64405053 lea eax, [ebp+pvParam]
.text:64405056 push eax ; pvParam
.text:64405057 push edi ; uiParam
.text:64405058 push 30h ; uiAction
.text:6440505A call ds:SystemParametersInfoA
.text:64405063 cmp eax, 1600
.text:64405068 jle short loc_64405072
.text:6440506A cdq
.text:6440506B sub eax, edx
.text:6440506D sar eax, 1
.text:6440506F mov [ebp+var_34], eax
.text:64405072
.text:64405072 loc_64405072:
start drawing:
.text:64405072 push edi ; hWnd

.text:64405073 mov [ebp+cy], 0A0h
.text:6440507A call ds:GetDC
Let’s check our theory on practice.

JNZ at this line …
.text:6440503F jnz exit ; bypass drawing
…replace it with just JMP, and we get SAPGUI working without the pesky annoying pop-up window appear-
ing!
Now let’s dig deeper and find a connection between the 0x15 offset in the load_command_line() (we gave
it this name) function and the this+0x3D variable in CDwsGui::PrepareInfoWindow. Are we sure the value
is the same?
We are starting to search for all occurrences of the 0x15 value in code. For a small programs like SAPGUI,
it sometimes works. Here is the first occurrence we’ve got:
.text:64404C19 sub_64404C19 proc near
.text:64404C19
.text:64404C19 arg_0 = dword ptr 4
.text:64404C19
868
.text:64404C19 push ebx
.text:64404C1A push ebp
.text:64404C1B push esi
.text:64404C1C push edi
.text:64404C1D mov edi, [esp+10h+arg_0]
.text:64404C21 mov eax, [edi]
.text:64404C23 mov esi, ecx ; ESI/ECX are pointers to some unknown object.
.text:64404C25 mov [esi], eax
.text:64404C27 mov eax, [edi+4]
.text:64404C2A mov [esi+4], eax
.text:64404C2D mov eax, [edi+8]
.text:64404C30 mov [esi+8], eax
.text:64404C33 lea eax, [edi+0Ch]
.text:64404C36 push eax
.text:64404C37 lea ecx, [esi+0Ch]
; demangled name: ATL::CStringT::operator=(class ATL::CStringT ... &)

.text:64404C3A call ds:mfc90_817
.text:64404C40 mov eax, [edi+10h]
.text:64404C43 mov [esi+10h], eax
.text:64404C46 mov al, [edi+14h]
.text:64404C49 mov [esi+14h], al
.text:64404C4C mov al, [edi+15h] ; copy byte from 0x15 offset
.text:64404C4F mov [esi+15h], al ; to 0x15 offset in CDwsGui object
The function has been called from the function named CDwsGui::CopyOptions! And thanks again for
debugging information.
But the real answer is in CDwsGui::Init():
.text:6440B0BF loc_6440B0BF:
.text:6440B0BF mov eax, [ebp+arg_0]
.text:6440B0C2 push [ebp+arg_4]
.text:6440B0C5 mov [esi+2844h], eax
.text:6440B0CB lea eax, [esi+28h] ; ESI is pointer to CDwsGui object
.text:6440B0CE push eax
.text:6440B0CF call CDwsGui__CopyOptions
Finally, we understand: the array filled in the load_command_line() function is actually placed in the
CDwsGui class, but at address this+0x28. 0x15 + 0x28 is exactly 0x3D. OK, we found the point where
the value is copied to.
Let’s also find the rest of the places where the 0x3D offset is used. Here is one of them in the CDws-
Gui::SapguiRun function (again, thanks to the debugging calls):
.text:64409D58 cmp [esi+3Dh], bl ; ESI is pointer to CDwsGui object
.text:64409D5B lea ecx, [esi+2B8h]
.text:64409D61 setz al
.text:64409D64 push eax ; arg_10 of CConnectionContext::CreateNetwork
.text:64409D65 push dword ptr [esi+64h]

.text:64409D68 ; no arguments
.text:64409D6E push eax
.text:64409D6F lea ecx, [esi+2BCh]

.text:64409D75 ; no arguments
.text:64409D7B push eax
.text:64409D7C push esi
.text:64409D7D lea ecx, [esi+8]
.text:64409D80 call CConnectionContext__CreateNetwork
Let’s check our findings.

Replace the setz al here with the xor eax, eax / nop instructions, clear the TDW_NOCOMPRESS en-
vironment variable and run SAPGUI. Wow! There pesky annoying window is no more (just as expected,
because the variable is not set) but in Wireshark we can see that the network packets are not compressed
869
anymore! Obviously, this is the point where the compression flag is to be set in the CConnectionContext
object.
So, the compression flag is passed in the 5th argument of CConnectionContext::CreateNetwork. Inside
the function, another one is called:
...
.text:64403476 push [ebp+compression]
.text:64403479 push [ebp+arg_C]
.text:6440347C push [ebp+arg_8]
.text:6440347F push [ebp+arg_4]
.text:64403482 push [ebp+arg_0]
.text:64403485 call CNetwork__CNetwork
The compression flag is passed here in the 5th argument to the CNetwork::CNetwork constructor.
And here is how the CNetwork constructor sets the flag in the CNetwork object according to its 5th argu-
ment and another variable which probably could also affect network packets compression.
.text:64411DF1 cmp [ebp+compression], esi
.text:64411DF7 jz short set_EAX_to_0
.text:64411DF9 mov al, [ebx+78h] ; another value may affect compression?
.text:64411DFC cmp al, '3'
.text:64411DFE jz short set_EAX_to_1
.text:64411E00 cmp al, '4'
.text:64411E02 jnz short set_EAX_to_0
.text:64411E04
.text:64411E04 set_EAX_to_1:
.text:64411E04 xor eax, eax
.text:64411E06 inc eax ; EAX -> 1
.text:64411E07 jmp short loc_64411E0B
.text:64411E09
.text:64411E09 set_EAX_to_0:
.text:64411E09
.text:64411E09 xor eax, eax ; EAX -> 0
.text:64411E0B
.text:64411E0B loc_64411E0B:
.text:64411E0B mov [ebx+3A4h], eax ; EBX is pointer to CNetwork object
At this point we know the compression flag is stored in the CNetwork class at address this+0x3A4.
Now let’s dig through SAPguilib.dll for the 0x3A4 value. And here is the second occurrence in CDws-
Gui::OnClientMessageWrite (endless thanks for the debugging information):
.text:64406F76 loc_64406F76:
.text:64406F76 mov ecx, [ebp+7728h+var_7794]
.text:64406F79 cmp dword ptr [ecx+3A4h], 1
.text:64406F80 jnz compression_flag_is_zero
.text:64406F86 mov byte ptr [ebx+7], 1
.text:64406F8A mov eax, [esi+18h]
.text:64406F8D mov ecx, eax
.text:64406F8F test eax, eax
.text:64406F91 ja short loc_64406FFF
.text:64406F93 mov ecx, [esi+14h]
.text:64406F96 mov eax, [esi+20h]
.text:64406F99
.text:64406F99 loc_64406F99:
.text:64406F99 push dword ptr [edi+2868h] ; int
.text:64406F9F lea edx, [ebp+7728h+var_77A4]
.text:64406FA2 push edx ; int
.text:64406FA3 push 30000 ; int
.text:64406FA8 lea edx, [ebp+7728h+Dst]
.text:64406FAB push edx ; Dst
.text:64406FAC push ecx ; int
.text:64406FAD push eax ; Src
.text:64406FAE push dword ptr [edi+28C0h] ; int
.text:64406FB4 call sub_644055C5 ; actual compression routine
.text:64406FB9 add esp, 1Ch
.text:64406FBC cmp eax, 0FFFFFFF6h
.text:64406FBF jz short loc_64407004
.text:64406FC1 cmp eax, 1
870
.text:64406FC4 jz loc_6440708C
.text:64406FCA cmp eax, 2
.text:64406FCD jz short loc_64407004
.text:64406FCF push eax
.text:64406FD0 push offset aCompressionErr ;
"compression error [rc = %d]- program wi"...
.text:64406FD5 push offset aGui_err_compre ; "GUI_ERR_COMPRESS"
.text:64406FDA push dword ptr [edi+28D0h]
.text:64406FE0 call SapPcTxtRead
Let’s take a look in sub_644055C5. In it we can only see the call to memcpy() and another function named
(by IDA) sub_64417440.
And, let’s take a look inside sub_64417440. What we see is:
.text:6441747C push offset aErrorCsrcompre ;
"\nERROR: CsRCompress: invalid handle"
.text:64417481 call eax ; dword_644F94C8
34
Voilà! We’ve found the function that actually compresses the data. As it was shown in past ,
this function is used in SAP and also the open-source MaxDB project. So it is available in source form.
Doing the last check here:
.text:64406F79 cmp dword ptr [ecx+3A4h], 1
.text:64406F80 jnz compression_flag_is_zero
Replace JNZ here for an unconditional JMP. Remove the environment variable TDW_NOCOMPRESS. Voilà!
In Wireshark we see that the client messages are not compressed. The server responses, however, are
compressed.
So we found exact connection between the environment variable and the point where data compression
routine can be called or bypassed.
8.12.2 SAP 6.0 password checking functions

One time when the author of this book have returned again to his SAP 6.0 IDES installed in a VMware box,
he figured out that he forgot the password for the SAP* account, then he have recalled it, but then he got
this error message «Password logon no longer possible - too many failed attempts», since he’ve made all
these attempts in attempt to recall it.
The first extremely good news was that the full disp+work.pdb PDB file is supplied with SAP, and it contain
almost everything: function names, structures, types, local variable and argument names, etc. What a
lavish gift!
There is TYPEINFODUMP35 utility for converting PDB files into something readable and grepable.
Here is an example of a function information + its arguments + its local variables:
FUNCTION ThVmcSysEvent
Address: 10143190 Size: 675 bytes Index: 60483 TypeIndex: 60484
Type: int NEAR_C ThVmcSysEvent (unsigned int, unsigned char, unsigned short*)
Flags: 0
PARAMETER events
Address: Reg335+288 Size: 4 bytes Index: 60488 TypeIndex: 60489
Type: unsigned int
Flags: d0
PARAMETER opcode
Type: unsigned char
Flags: d0
PARAMETER serverName
Type: unsigned short*
Flags: d0
STATIC_LOCAL_VAR func
871
Address: 12274af0 Size: 8 bytes Index: 60495 TypeIndex: 60496
Type: wchar_t*
Flags: 80
LOCAL_VAR admhead
Type: unsigned char*
Flags: 90
LOCAL_VAR record
Type: AD_RECORD
Flags: 90
LOCAL_VAR adlen
Type: int
Flags: 90
And here is an example of some structure:

STRUCT DBSL_STMTID
Size: 120 Variables: 4 Functions: 0 Base classes: 0
MEMBER moduletype
Type: DBSL_MODULETYPE
Offset: 0 Index: 3 TypeIndex: 38653
MEMBER module
Type: wchar_t module[40]
MEMBER stmtnum
Type: long
MEMBER timestamp
Type: wchar_t timestamp[15]
Wow!
Another good news: debugging calls (there are plenty of them) are very useful.
Here you can also notice the ct_level global variable36 , that reflects the current trace level.
There are a lot of debugging inserts in the disp+work.exe file:
cmp cs:ct_level, 1
jl short loc_1400375DA
call DpLock
lea rcx, aDpxxtool4_c ; "dpxxtool4.c"
mov edx, 4Eh ; line
call CTrcSaveLocation
mov r8, cs:func_48
mov rcx, cs:hdl ; hdl
lea rdx, aSDpreadmemvalu ; "%s: DpReadMemValue (%d)"
mov r9d, ebx
call DpTrcErr
call DpUnlock
If the current trace level is bigger or equal to threshold defined in the code here, a debugging message is
to be written to the log files like dev_w0, dev_disp, and other dev* files.
Let’s try grepping in the file that we have got with the help of the TYPEINFODUMP utility:
cat "disp+work.pdb.d" | grep FUNCTION | grep -i password
We have got:
FUNCTION rcui::AgiPassword::DiagISelection
FUNCTION ssf_password_encrypt
FUNCTION ssf_password_decrypt
FUNCTION password_logon_disabled
FUNCTION dySignSkipUserPassword
FUNCTION migrate_password_history
36 More about trace level: http://go.yurichev.com/17039
872
FUNCTION password_is_initial
FUNCTION rcui::AgiPassword::IsVisible
FUNCTION password_distance_ok
FUNCTION get_password_downwards_compatibility
FUNCTION dySignUnSkipUserPassword
FUNCTION rcui::AgiPassword::GetTypeName
FUNCTION `rcui::AgiPassword::AgiPassword'::`1'::dtor$2
FUNCTION usm_set_password
FUNCTION rcui::AgiPassword::TraceTo
FUNCTION days_since_last_password_change
FUNCTION rsecgrp_generate_random_password
FUNCTION rcui::AgiPassword::`scalar deleting destructor'
FUNCTION password_attempt_limit_exceeded
FUNCTION handle_incorrect_password
FUNCTION `rcui::AgiPassword::`scalar deleting destructor''::`1'::dtor$1
FUNCTION calculate_new_password_hash
FUNCTION shift_password_to_history
FUNCTION rcui::AgiPassword::GetType
FUNCTION found_password_in_history
FUNCTION `rcui::AgiPassword::`scalar deleting destructor''::`1'::dtor$0
FUNCTION rcui::AgiObj::IsaPassword
FUNCTION password_idle_check
FUNCTION SlicHwPasswordForDay
FUNCTION rcui::AgiPassword::IsaPassword
FUNCTION rcui::AgiPassword::AgiPassword
FUNCTION delete_user_password
FUNCTION usm_set_user_password
FUNCTION Password_API
FUNCTION get_password_change_for_SSO
FUNCTION password_in_USR40
FUNCTION rsec_agrp_abap_generate_random_password
Let’s also try to search for debug messages which contain the words «password» and «locked». One of
them is the string «user was locked by subsequently failed password logon attempts» , referenced in
function password_attempt_limit_exceeded().
Other strings that this function can write to a log file are: «password logon attempt will be rejected
immediately (preventing dictionary attacks)», «failed-logon lock: expired (but not removed due to ’read-
only’ operation)», «failed-logon lock: expired => removed».
After playing for a little with this function, we noticed that the problem is exactly in it. It is called from the
chckpass() function —one of the password checking functions.
First, we would like to make sure that we are at the correct point:
Run tracer:
tracer64.exe -a:disp+work.exe bpf=disp+work.exe!chckpass,args:3,unicode
PID=2236|TID=2248|(0) disp+work.exe!chckpass (0x202c770, L"Brewered1 ⤦

Ç ", 0x41) (called from 0x1402f1060 (disp+work.exe!usrexist+0x3c0))
PID=2236|TID=2248|(0) disp+work.exe!chckpass -> 0x35
The call path is: syssigni() -> DyISigni() -> dychkusr() -> usrexist() -> chckpass().
The number 0x35 is an error returned in chckpass() at that point:
.text:00000001402ED567 loc_1402ED567: ; CODE XREF: chckpass+B4
.text:00000001402ED567 mov rcx, rbx ; usr02
.text:00000001402ED56A call password_idle_check
.text:00000001402ED56F cmp eax, 33h
.text:00000001402ED572 jz loc_1402EDB4E
.text:00000001402ED578 cmp eax, 36h
.text:00000001402ED57B jz loc_1402EDB3D
.text:00000001402ED581 xor edx, edx ; usr02_readonly
.text:00000001402ED583 mov rcx, rbx ; usr02
.text:00000001402ED586 call password_attempt_limit_exceeded
.text:00000001402ED58B test al, al
.text:00000001402ED58D jz short loc_1402ED5A0
873
.text:00000001402ED58F mov eax, 35h
.text:00000001402ED594 add rsp, 60h
.text:00000001402ED598 pop r14
.text:00000001402ED59A pop r12
.text:00000001402ED59C pop rdi
.text:00000001402ED59D pop rsi
.text:00000001402ED59E pop rbx
.text:00000001402ED59F retn
Fine, let’s check:

tracer64.exe -a:disp+work.exe bpf=disp+work.exe!password_attempt_limit_exceeded,args:4,unicode,⤦
Ç rt:0
PID=2744|TID=360|(0) disp+work.exe!password_attempt_limit_exceeded (0x202c770, 0, 0x257758, 0) ⤦

Ç (called from 0x1402ed58b (disp+work.exe!chckpass+0xeb))
PID=2744|TID=360|(0) disp+work.exe!password_attempt_limit_exceeded -> 1
PID=2744|TID=360|We modify return value (EAX/RAX) of this function to 0
PID=2744|TID=360|(0) disp+work.exe!password_attempt_limit_exceeded (0x202c770, 0, 0, 0) (called⤦
Ç from 0x1402e9794 (disp+work.exe!chngpass+0xe4))
PID=2744|TID=360|(0) disp+work.exe!password_attempt_limit_exceeded -> 1
Excellent! We can successfully login now.

By the way, we can pretend we forgot the password, fixing the chckpass() function to return a value of 0
is enough to bypass the check:
tracer64.exe -a:disp+work.exe bpf=disp+work.exe!chckpass,args:3,unicode,rt:0
PID=2744|TID=360|(0) disp+work.exe!chckpass (0x202c770, L"bogus ⤦

Ç ", 0x41) (called from 0x1402f1060 (disp+work.exe!usrexist+0x3c0))
PID=2744|TID=360|(0) disp+work.exe!chckpass -> 0x35
What also can be said while analyzing the

password_attempt_limit_exceeded() function is that at the very beginning of it, this call can be seen:
lea rcx, aLoginFailed_us ; "login/failed_user_auto_unlock"
call sapgparam
test rax, rax
jz short loc_1402E19DE
movzx eax, word ptr [rax]
cmp ax, 'N'
jz short loc_1402E19D4
cmp ax, 'n'
jz short loc_1402E19D4
cmp ax, '0'
jnz short loc_1402E19DE
Obviously, function sapgparam() is used to query the value of some configuration parameter. This function
can be called from 1768 different places. It seems that with the help of this information, we can easily
find the places in code, the control flow of which can be affected by specific configuration parameters.
It is really sweet. The function names are very clear, much clearer than in the Oracle RDBMS.
It seems that the disp+work process is written in C++. Has it been rewritten some time ago?
8.13 Oracle RDBMS

8.13.1 V$VERSION table in the Oracle RDBMS
Oracle RDBMS 11.2 is a huge program, its main module oracle.exe contain approx. 124,000 functions.
For comparison, the Windows 7 x86 kernel (ntoskrnl.exe) contains approx. 11,000 functions and the Linux
3.9.8 kernel (with default drivers compiled)—31,000 functions.
Let’s start with an easy question. Where does Oracle RDBMS get all this information, when we execute
this simple statement in SQL*Plus:
874
SQL> select * from V$VERSION;
And we get:
BANNER
--------------------------------------------------------
Oracle Database 11g Enterprise Edition Release 11.2.0.1.0 - Production

PL/SQL Release 11.2.0.1.0 - Production
CORE 11.2.0.1.0 Production
TNS for 32-bit Windows: Version 11.2.0.1.0 - Production
NLSRTL Version 11.2.0.1.0 - Production
Let’s start. Where in the Oracle RDBMS can we find the string V$VERSION?
In the win32-version, oracle.exe file contains the string, it’s easy to see. But we can also use the object
(.o) files from the Linux version of Oracle RDBMS since, unlike the win32 version oracle.exe, the function
names (and global variables as well) are preserved there.
So, the kqf.o file contains the V$VERSION string. The object file is in the main Oracle-library libserver11.a.
A reference to this text string can find in the kqfviw table stored in the same file, kqf.o:
Listing 8.10: kqf.o
.rodata:0800C4A0 kqfviw dd 0Bh ; DATA XREF: kqfchk:loc_8003A6D
.rodata:0800C4A0 ; kqfgbn+34
.rodata:0800C4A4 dd offset _2__STRING_10102_0 ; "GV$WAITSTAT"
.rodata:0800C4A8 dd 4
.rodata:0800C4AC dd offset _2__STRING_10103_0 ; "NULL"
.rodata:0800C4B0 dd 3
.rodata:0800C4B4 dd 0
.rodata:0800C4B8 dd 195h
.rodata:0800C4BC dd 4
.rodata:0800C4C0 dd 0
.rodata:0800C4C4 dd 0FFFFC1CBh
.rodata:0800C4C8 dd 3
.rodata:0800C4CC dd 0
.rodata:0800C4D0 dd 0Ah
.rodata:0800C4D4 dd offset _2__STRING_10104_0 ; "V$WAITSTAT"
.rodata:0800C4D8 dd 4
.rodata:0800C4DC dd offset _2__STRING_10103_0 ; "NULL"
.rodata:0800C4E0 dd 3
.rodata:0800C4E4 dd 0
.rodata:0800C4E8 dd 4Eh
.rodata:0800C4EC dd 3
.rodata:0800C4F0 dd 0
.rodata:0800C4F4 dd 0FFFFC003h
.rodata:0800C4F8 dd 4
.rodata:0800C4FC dd 0
.rodata:0800C500 dd 5
.rodata:0800C504 dd offset _2__STRING_10105_0 ; "GV$BH"
.rodata:0800C50C dd offset _2__STRING_10103_0 ; "NULL"
.rodata:0800C518 dd 269h
.rodata:0800C51C dd 15h
.rodata:0800C524 dd 0FFFFC1EDh
.rodata:0800C52C dd 0
.rodata:0800C534 dd offset _2__STRING_10106_0 ; "V$BH"
.rodata:0800C53C dd offset _2__STRING_10103_0 ; "NULL"
.rodata:0800C548 dd 0F5h
.rodata:0800C54C dd 14h
875
.rodata:0800C554 dd 0FFFFC1EEh
.rodata:0800C55C dd 0
By the way, often, while analyzing Oracle RDBMS’s internals, you may ask yourself, why are the names
of the functions and global variable so weird.
Probably, because Oracle RDBMS is a very old product and was developed in C in the 1980s.
And that was a time when the C standard guaranteed that the function names/variables can support only
up to 6 characters inclusive: «6 significant initial characters in an external identifier»37
Probably, the table kqfviw contains most (maybe even all) views prefixed with V$, these are fixed views,
present all the time. Superficially, by noticing the cyclic recurrence of data, we can easily see that each
kqfviw table element has 12 32-bit fields. It is very simple to create a 12-elements structure in IDA and
apply it to all table elements. As of Oracle RDBMS version 11.2, there are 1023 table elements, i.e., in it
are described 1023 of all possible fixed views.
We are going to return to this number later.
As we can see, there is not much information in these numbers in the fields. The first field is always
equals to the name of the view (without the terminating zero). This is correct for each element. But this
information is not very useful.
We also know that the information about all fixed views can be retrieved from a fixed view named
V$FIXED_VIEW_DEFINITION (by the way, the information for this view is also taken from the kqfviw and
kqfvip tables.) Incidentally, there are 1023 elements in those too. Coincidence? No.
SQL> select * from V$FIXED_VIEW_DEFINITION where view_name='V$VERSION';
VIEW_NAME
------------------------------
VIEW_DEFINITION
------------------------------
V$VERSION
select BANNER from GV$VERSION where inst_id = USERENV('Instance')
So, V$VERSION is some kind of a thunk view for another view, named GV$VERSION, which is, in turn:
SQL> select * from V$FIXED_VIEW_DEFINITION where view_name='GV$VERSION';
VIEW_NAME
------------------------------
VIEW_DEFINITION
------------------------------
GV$VERSION
select inst_id, banner from x$version
The tables prefixed with X$ in the Oracle RDBMS are service tables too, undocumented, cannot be changed
by the user and are refreshed dynamically.
If we search for the text
select BANNER from GV$VERSION where inst_id =

USERENV('Instance')
... in the kqf.o file, we find it in the kqfvip table:
Listing 8.11: kqf.o

.rodata:080185A0 kqfvip dd offset _2__STRING_11126_0 ; DATA XREF: kqfgvcn+18
.rodata:080185A0 ; kqfgvt+F
.rodata:080185A0 ; "select inst_id,decode(indx,1,'data bloc"...
.rodata:080185A4 dd offset kqfv459_c_0
.rodata:080185A8 dd 0
.rodata:080185AC dd 0
37 Draft ANSI C Standard (ANSI X3J11/88-090) (May 13, 1988) (yurichev.com)
876
...
.rodata:08019570 dd offset _2__STRING_11378_0 ;

"select BANNER from GV$VERSION where in"...
.rodata:08019574 dd offset kqfv133_c_0
.rodata:08019578 dd 0
.rodata:0801957C dd 0
"select inst_id,decode(bitand(cfflg,1),0"...
.rodata:08019588 dd 0
.rodata:0801958C dd 0
"select STATUS , NAME, IS_RECOVERY_DEST"...
The table appear to have 4 fields in each element. By the way, there are 1023 elements in it, again, the
number we already know.
The second field points to another table that contains the table fields for this fixed view. As for V$VERSION,
this table has only two elements, the first is 6 and the second is the BANNER string (the number 6 is this
string’s length) and after, a terminating element that contains 0 and a null C string:
Listing 8.12: kqf.o

.rodata:080BBAC4 kqfv133_c_0 dd 6 ; DATA XREF: .rodata:08019574
.rodata:080BBAC8 dd offset _2__STRING_5017_0 ; "BANNER"
.rodata:080BBACC dd 0
.rodata:080BBAD0 dd offset _2__STRING_0_0
By joining data from both kqfviw and kqfvip tables, we can get the SQL statements which are executed
when the user wants to query information from a specific fixed view.
So we can write an oracle tables38 program, to gather all this information from Oracle RDBMS for Linux’s
object files. For V$VERSION, we find this:
Listing 8.13: Result of oracle tables

kqfviw_element.viewname: [V$VERSION] ?: 0x3 0x43 0x1 0xffffc085 0x4
kqfvip_element.statement: [select BANNER from GV$VERSION where inst_id = USERENV('Instance')]
kqfvip_element.params:
[BANNER]
And:

kqfviw_element.viewname: [GV$VERSION] ?: 0x3 0x26 0x2 0xffffc192 0x1
kqfvip_element.statement: [select inst_id, banner from x$version]
kqfvip_element.params:
[INST_ID] [BANNER]
The GV$VERSION fixed view is different from V$VERSION only in that it has one more field with the identifier
instance.
Anyway, we are going to stick with the X$VERSION table. Just like any other X$-table, it is undocumented,
however, we can query it:
SQL> select * from x$version;
ADDR INDX INST_ID

-------- ---------- ----------
BANNER
------------------------------
0DBAF574 0 1
Oracle Database 11g Enterprise Edition Release 11.2.0.1.0 - Production
...
38 yurichev.com
877
This table has some additional fields, like ADDR and INDX.
While scrolling kqf.o in IDA we can spot another table that contains a pointer to the X$VERSION string,
this is kqftab:
Listing 8.15: kqf.o

.rodata:0803CAC0 dd 9 ; element number 0x1f6
.rodata:0803CAC4 dd offset _2__STRING_13113_0 ; "X$VERSION"
.rodata:0803CAC8 dd 4
.rodata:0803CACC dd offset _2__STRING_13114_0 ; "kqvt"
.rodata:0803CAD0 dd 4
.rodata:0803CADC dd 4
.rodata:0803CAE0 dd 0Ch
.rodata:0803CAE4 dd 0FFFFC075h
.rodata:0803CAE8 dd 3
.rodata:0803CAEC dd 0
.rodata:0803CAF0 dd 7
.rodata:0803CAF4 dd offset _2__STRING_13115_0 ; "X$KQFSZ"
.rodata:0803CAF8 dd 5
.rodata:0803CAFC dd offset _2__STRING_13116_0 ; "kqfsz"
.rodata:0803CB00 dd 1
.rodata:0803CB04 dd 38h
.rodata:0803CB0C dd 7
.rodata:0803CB14 dd 0FFFFC09Dh
.rodata:0803CB1C dd 0
There are a lot of references to the X$-table names, apparently, to all Oracle RDBMS 11.2 X$-tables. But
again, we don’t have enough information.
It’s not clear what does the kqvt string stands for.
The kq prefix may mean kernel or query.
v apparently stands for version and t—type? Hard to say.
A table with a similar name can be found in kqf.o:
Listing 8.16: kqf.o

.rodata:0808C360 kqvt_c_0 kqftap_param <4, offset _2__STRING_19_0, 917h, 0, 0, 0, 4, 0, 0>
.rodata:0808C360 ; DATA XREF: .rodata:08042680
.rodata:0808C360 ; "ADDR"
.rodata:0808C384 kqftap_param <4, offset _2__STRING_20_0, 0B02h, 0, 0, 0, 4, 0, 0> ;
"INDX"
.rodata:0808C3A8 kqftap_param <7, offset _2__STRING_21_0, 0B02h, 0, 0, 0, 4, 0, 0> ;
"INST_ID"
.rodata:0808C3CC kqftap_param <6, offset _2__STRING_5017_0, 601h, 0, 0, 0, 50h, 0, 0> ⤦
Ç ;
"BANNER"
.rodata:0808C3F0 kqftap_param <0, offset _2__STRING_0_0, 0, 0, 0, 0, 0, 0, 0>
It contains information about all fields in the X$VERSION table. The only reference to this table is in the
kqftap table:
Listing 8.17: kqf.o

.rodata:08042680 kqftap_element <0, offset kqvt_c_0, offset kqvrow, 0> ; ⤦
Ç element 0x1f6
It is interesting that this element here is 0x1f6th (502nd), just like the pointer to the X$VERSION string in
the kqftab table.
Probably, the kqftap and kqftab tables complement each other, just like kqfvip and kqfviw.
We also see a pointer to the kqvrow() function. Finally, we got something useful!
So we will add these tables to our oracle tables39 utility too. For X$VERSION we get:
39 yurichev.com
878
kqftab_element.name: [X$VERSION] ?: [kqvt] 0x4 0x4 0x4 0xc 0xffffc075 0x3
kqftap_param.name=[ADDR] ?: 0x917 0x0 0x0 0x0 0x4 0x0 0x0
kqftap_param.name=[INDX] ?: 0xb02 0x0 0x0 0x0 0x4 0x0 0x0
kqftap_param.name=[INST_ID] ?: 0xb02 0x0 0x0 0x0 0x4 0x0 0x0
kqftap_param.name=[BANNER] ?: 0x601 0x0 0x0 0x0 0x50 0x0 0x0
kqftap_element.fn1=kqvrow
kqftap_element.fn2=NULL
With the help of tracer, it is easy to check that this function is called 6 times in row (from the qerfxFetch()
function) while querying the X$VERSION table.
Let’s run tracer in cc mode (it comments each executed instruction):
tracer -a:oracle.exe bpf=oracle.exe!_kqvrow,trace:cc
_kqvrow_ proc near
var_7C = byte ptr -7Ch

Dest = dword ptr -10h
; FUNCTION CHUNK AT .text1:056C11A0 SIZE 00000049 BYTES
push ebp
mov ebp, esp
sub esp, 7Ch
mov eax, [ebp+arg_14] ; [EBP+1Ch]=1
mov ecx, TlsIndex ; [69AEB08h]=0
mov edx, large fs:2Ch
mov edx, [edx+ecx*4] ; [EDX+ECX*4]=0xc98c938
cmp eax, 2 ; EAX=1
mov eax, [ebp+arg_8] ; [EBP+10h]=0xcdfe554
jz loc_2CE1288
mov ecx, [eax] ; [EAX]=0..5
mov [ebp+var_4], edi ; EDI=0xc98c938
loc_2CE10F6: ; CODE XREF: _kqvrow_+10A

; _kqvrow_+1A9
cmp ecx, 5 ; ECX=0..5
ja loc_56C11C7
mov edi, [ebp+arg_18] ; [EBP+20h]=0
mov [ebp+var_14], edx ; EDX=0xc98c938
mov [ebp+var_8], ebx ; EBX=0
mov ebx, eax ; EAX=0xcdfe554
mov [ebp+var_C], esi ; ESI=0xcdfe248
loc_2CE110D: ; CODE XREF: _kqvrow_+29E00E6

mov edx, ds:off_628B09C[ecx*4] ; [ECX*4+628B09Ch]=0x2ce1116, 0x2ce11ac, 0x2ce11db,
0x2ce11f6, 0x2ce1236, 0x2ce127a
jmp edx ; EDX=0x2ce1116, 0x2ce11ac, 0x2ce11db, 0x2ce11f6, 0x2ce1236,
0x2ce127a
loc_2CE1116: ; DATA XREF: .rdata:off_628B09C

push offset aXKqvvsnBuffer ; "x$kqvvsn buffer"
mov ecx, [ebp+arg_C] ; [EBP+14h]=0x8a172b4
xor edx, edx
mov esi, [ebp+var_14] ; [EBP-14h]=0xc98c938
push edx ; EDX=0
push edx ; EDX=0
push 50h
push ecx ; ECX=0x8a172b4
879
push dword ptr [esi+10494h] ; [ESI+10494h]=0xc98cd58
call _kghalf ; tracing nested maximum level (1) reached, skipping this CALL
mov esi, ds:__imp__vsnnum ; [59771A8h]=0x61bc49e0
mov [ebp+Dest], eax ; EAX=0xce2ffb0
mov [ebx+8], eax ; EAX=0xce2ffb0
mov [ebx+4], eax ; EAX=0xce2ffb0
mov edi, [esi] ; [ESI]=0xb200100
mov esi, ds:__imp__vsnstr ; [597D6D4h]=0x65852148, "- Production"
push esi ; ESI=0x65852148, "- Production"
mov ebx, edi ; EDI=0xb200100
shr ebx, 18h ; EBX=0xb200100
mov ecx, edi ; EDI=0xb200100
shr ecx, 14h ; ECX=0xb200100
and ecx, 0Fh ; ECX=0xb2
mov edx, edi ; EDI=0xb200100
shr edx, 0Ch ; EDX=0xb200100
movzx edx, dl ; DL=0
mov eax, edi ; EDI=0xb200100
shr eax, 8 ; EAX=0xb200100
and eax, 0Fh ; EAX=0xb2001
and edi, 0FFh ; EDI=0xb200100
push edi ; EDI=0
mov edi, [ebp+arg_18] ; [EBP+20h]=0
push eax ; EAX=1
mov eax, ds:__imp__vsnban ;
[597D6D8h]=0x65852100, "Oracle Database 11g Enterprise Edition Release %d.%d.%d.%d.%d %s"
push edx ; EDX=0
push ecx ; ECX=2
push ebx ; EBX=0xb
mov ebx, [ebp+arg_8] ; [EBP+10h]=0xcdfe554
push eax ;
EAX=0x65852100, "Oracle Database 11g Enterprise Edition Release %d.%d.%d.%d.%d %s"
mov eax, [ebp+Dest] ; [EBP-10h]=0xce2ffb0
push eax ; EAX=0xce2ffb0
call ds:__imp__sprintf ; op1=MSVCR80.dll!sprintf tracing nested maximum level (1)
reached, skipping this CALL
add esp, 38h
mov dword ptr [ebx], 1
loc_2CE1192: ; CODE XREF: _kqvrow_+FB

; _kqvrow_+128 ...
test edi, edi ; EDI=0
jnz __VInfreq__kqvrow
mov esi, [ebp+var_C] ; [EBP-0Ch]=0xcdfe248
mov edi, [ebp+var_4] ; [EBP-4]=0xc98c938
mov eax, ebx ; EBX=0xcdfe554
mov ebx, [ebp+var_8] ; [EBP-8]=0
lea eax, [eax+4] ; [EAX+4]=0xce2ffb0, "NLSRTL Version 11.2.0.1.0 - Production",
"Oracle Database 11g Enterprise Edition Release 11.2.0.1.0 - Production", "PL/SQL Release
11.2.0.1.0 - Production", "TNS for 32-bit Windows: Version 11.2.0.1.0 - Production"
loc_2CE11A8: ; CODE XREF: _kqvrow_+29E00F6

mov esp, ebp
pop ebp
retn ; EAX=0xcdfe558
loc_2CE11AC: ; DATA XREF: .rdata:0628B0A0

mov edx, [ebx+8] ; [EBX+8]=0xce2ffb0, "Oracle Database 11g Enterprise Edition
Release 11.2.0.1.0 - Production"
mov [ebx+4], edx ; EDX=0xce2ffb0, "Oracle Database 11g Enterprise Edition
push edx ; EDX=0xce2ffb0, "Oracle Database 11g Enterprise Edition
call _kkxvsn ; tracing nested maximum level (1) reached, skipping this CALL
pop ecx
mov edx, [ebx+4] ; [EBX+4]=0xce2ffb0, "PL/SQL Release 11.2.0.1.0 - Production"
movzx ecx, byte ptr [edx] ; [EDX]=0x50
test ecx, ecx ; ECX=0x50
jnz short loc_2CE1192
mov edx, [ebp+var_14]
mov esi, [ebp+var_C]
880
mov eax, ebx
mov ebx, [ebp+var_8]
mov ecx, [eax]
jmp loc_2CE10F6
loc_2CE11DB: ; DATA XREF: .rdata:0628B0A4

push 0
push 50h
mov edx, [ebx+8] ; [EBX+8]=0xce2ffb0, "PL/SQL Release 11.2.0.1.0 - Production"
mov [ebx+4], edx ; EDX=0xce2ffb0, "PL/SQL Release 11.2.0.1.0 - Production"
push edx ; EDX=0xce2ffb0, "PL/SQL Release 11.2.0.1.0 - Production"
call _lmxver ; tracing nested maximum level (1) reached, skipping this CALL
add esp, 0Ch
jmp short loc_2CE1192
loc_2CE11F6: ; DATA XREF: .rdata:0628B0A8

mov edx, [ebx+8] ; [EBX+8]=0xce2ffb0
mov [ebp+var_18], 50h
mov [ebx+4], edx ; EDX=0xce2ffb0
push 0
call _npinli ; tracing nested maximum level (1) reached, skipping this CALL
pop ecx
test eax, eax ; EAX=0
jnz loc_56C11DA
mov ecx, [ebp+var_14] ; [EBP-14h]=0xc98c938
lea edx, [ebp+var_18] ; [EBP-18h]=0x50
push edx ; EDX=0xd76c93c
push dword ptr [ebx+8] ; [EBX+8]=0xce2ffb0
push dword ptr [ecx+13278h] ; [ECX+13278h]=0xacce190
call _nrtnsvrs ; tracing nested maximum level (1) reached, skipping this CALL
add esp, 0Ch
loc_2CE122B: ; CODE XREF: _kqvrow_+29E0118

jmp loc_2CE1192
loc_2CE1236: ; DATA XREF: .rdata:0628B0AC

lea edx, [ebp+var_7C] ; [EBP-7Ch]=1
push edx ; EDX=0xd76c8d8
push 0
mov esi, [ebx+8] ; [EBX+8]=0xce2ffb0, "TNS for 32-bit Windows: Version
11.2.0.1.0 - Production"
mov [ebx+4], esi ; ESI=0xce2ffb0, "TNS for 32-bit Windows: Version 11.2.0.1.0 -
Production"
mov ecx, 50h
mov [ebp+var_18], ecx ; ECX=0x50
push ecx ; ECX=0x50
push esi ; ESI=0xce2ffb0, "TNS for 32-bit Windows: Version 11.2.0.1.0 -
Production"
call _lxvers ; tracing nested maximum level (1) reached, skipping this CALL
add esp, 10h
mov edx, [ebp+var_18] ; [EBP-18h]=0x50
test edx, edx ; EDX=0x50
jnz loc_2CE1192
mov edx, [ebp+var_14]
mov esi, [ebp+var_C]
mov eax, ebx
mov ecx, 5
jmp loc_2CE10F6
loc_2CE127A: ; DATA XREF: .rdata:0628B0B0

mov edx, [ebp+var_14] ; [EBP-14h]=0xc98c938
mov esi, [ebp+var_C] ; [EBP-0Ch]=0xcdfe248
mov edi, [ebp+var_4] ; [EBP-4]=0xc98c938
mov eax, ebx ; EBX=0xcdfe554
mov ebx, [ebp+var_8] ; [EBP-8]=0
loc_2CE1288: ; CODE XREF: _kqvrow_+1F
881
mov eax, [eax+8] ; [EAX+8]=0xce2ffb0, "NLSRTL Version 11.2.0.1.0 - Production"
test eax, eax ; EAX=0xce2ffb0, "NLSRTL Version 11.2.0.1.0 - Production"
jz short loc_2CE12A7
push offset aXKqvvsnBuffer ; "x$kqvvsn buffer"
push eax ; EAX=0xce2ffb0, "NLSRTL Version 11.2.0.1.0 - Production"
mov eax, [ebp+arg_C] ; [EBP+14h]=0x8a172b4
push eax ; EAX=0x8a172b4
push dword ptr [edx+10494h] ; [EDX+10494h]=0xc98cd58
call _kghfrf ; tracing nested maximum level (1) reached, skipping this CALL
add esp, 10h
loc_2CE12A7: ; CODE XREF: _kqvrow_+1C1

xor eax, eax
mov esp, ebp
pop ebp
retn ; EAX=0
_kqvrow_ endp
Now it is easy to see that the row number is passed from outside. The function returns the string, con-
structing it as follows:
String 1 Using vsnstr, vsnnum, vsnban global variables.
Calls sprintf().
String 2 Calls kkxvsn().
String 3 Calls lmxver().
String 4 Calls npinli(), nrtnsvrs().
String 5 Calls lxvers().
That’s how the corresponding functions are called for determining each module’s version.
8.13.2 X$KSMLRU table in Oracle RDBMS

There is a mention of a special table in the Diagnosing and Resolving Error ORA-04031 on the Shared Pool
or Other Memory Pools [Video] [ID 146599.1] note:
There is a fixed table called X$KSMLRU that tracks allocations in the shared pool that
cause other objects in the shared pool to be aged out. This fixed table can be used to
identify what is causing the large allocation.
If many objects are being periodically flushed from the shared pool then this will cause
response time problems and will likely cause library cache latch contention problems when
the objects are reloaded into the shared pool.
One unusual thing about the X$KSMLRU fixed table is that the contents of the fixed table
are erased whenever someone selects from the fixed table. This is done since the fixed
table stores only the largest allocations that have occurred. The values are reset after being
selected so that subsequent large allocations can be noted even if they were not quite as
large as others that occurred previously. Because of this resetting, the output of selecting
from this table should be carefully kept since it cannot be retrieved back after the query is
issued.
However, as it can be easily checked, the contents of this table are cleared each time it’s queried. Are we
able to find why? Let’s get back to tables we already know: kqftab and kqftap which were generated with
oracle tables40 ’s help, that has all information about the X$-tables. We can see here that the ksmlrs()
function is called to prepare this table’s elements:

kqftab_element.name: [X$KSMLRU] ?: [ksmlr] 0x4 0x64 0x11 0xc 0xffffc0bb 0x5
kqftap_param.name=[INDX] ?: 0xb02 0x0 0x0 0x0 0x4 0x0 0x0
kqftap_param.name=[KSMLRIDX] ?: 0xb02 0x0 0x0 0x0 0x4 0x0 0x0
kqftap_param.name=[KSMLRDUR] ?: 0xb02 0x0 0x0 0x0 0x4 0x4 0x0
kqftap_param.name=[KSMLRSHRPOOL] ?: 0xb02 0x0 0x0 0x0 0x4 0x8 0x0
kqftap_param.name=[KSMLRCOM] ?: 0x501 0x0 0x0 0x0 0x14 0xc 0x0
kqftap_param.name=[KSMLRSIZ] ?: 0x2 0x0 0x0 0x0 0x4 0x20 0x0
40 yurichev.com
882
kqftap_param.name=[KSMLRNUM] ?: 0x2 0x0 0x0 0x0 0x4 0x24 0x0
kqftap_param.name=[KSMLRHON] ?: 0x501 0x0 0x0 0x0 0x20 0x28 0x0
kqftap_param.name=[KSMLROHV] ?: 0xb02 0x0 0x0 0x0 0x4 0x48 0x0
kqftap_param.name=[KSMLRSES] ?: 0x17 0x0 0x0 0x0 0x4 0x4c 0x0
kqftap_param.name=[KSMLRADU] ?: 0x2 0x0 0x0 0x0 0x4 0x50 0x0
kqftap_param.name=[KSMLRNID] ?: 0x2 0x0 0x0 0x0 0x4 0x54 0x0
kqftap_param.name=[KSMLRNSD] ?: 0x2 0x0 0x0 0x0 0x4 0x58 0x0
kqftap_param.name=[KSMLRNCD] ?: 0x2 0x0 0x0 0x0 0x4 0x5c 0x0
kqftap_param.name=[KSMLRNED] ?: 0x2 0x0 0x0 0x0 0x4 0x60 0x0
kqftap_element.fn1=ksmlrs
Indeed, with tracer’s help it is easy to see that this function is called each time we query the X$KSMLRU
table.
Here we see a references to the ksmsplu_sp() and ksmsplu_jp() functions, each of them calls the
ksmsplu() at the end. At the end of the ksmsplu() function we see a call to memset():
Listing 8.20: ksm.o

...
.text:00434C50 loc_434C50: ; DATA XREF: .rdata:off_5E50EA8

.text:00434C50 mov edx, [ebp-4]
.text:00434C53 mov [eax], esi
.text:00434C55 mov esi, [edi]
.text:00434C57 mov [eax+4], esi
.text:00434C5A mov [edi], eax
.text:00434C5C add edx, 1
.text:00434C5F mov [ebp-4], edx
.text:00434C62 jnz loc_434B7D
.text:00434C68 mov ecx, [ebp+14h]
.text:00434C6B mov ebx, [ebp-10h]
.text:00434C6E mov esi, [ebp-0Ch]
.text:00434C71 mov edi, [ebp-8]
.text:00434C74 lea eax, [ecx+8Ch]
.text:00434C7A push 370h ; Size
.text:00434C7F push 0 ; Val
.text:00434C81 push eax ; Dst
.text:00434C82 call __intel_fast_memset
.text:00434C87 add esp, 0Ch
.text:00434C8A mov esp, ebp
.text:00434C8C pop ebp
.text:00434C8D retn
.text:00434C8D _ksmsplu endp
Constructions like memset (block, 0, size) are often used just to zero memory block. What if we take
a risk, block the memset() call and see what happens?
Let’s run tracer with the following options: set breakpoint at 0x434C7A (the point where the arguments to
memset() are to be passed), so that tracer will set program counter EIP to the point where the arguments
passed to memset() are to be cleared (at 0x434C8A) It can be said that we just simulate an unconditional
jump from address 0x434C7A to 0x434C8A.
tracer -a:oracle.exe bpx=oracle.exe!0x00434C7A,set(eip,0x00434C8A)
(Important: all these addresses are valid only for the win32 version of Oracle RDBMS 11.2)
Indeed, now we can query the X$KSMLRU table as many times as we want and it is not being cleared
anymore!
Just in case, do not try this on your production servers.
It is probably not a very useful or desired system behavior, but as an experiment for locating a piece of
code that we need, it perfectly suits our needs!
8.13.3 V$TIMER table in Oracle RDBMS

V$TIMER is another fixed view that reflects a rapidly changing value:
883
V$TIMER displays the elapsed time in hundredths of a second. Time is measured since
the beginning of the epoch, which is operating system specific, and wraps around to 0 again
whenever the value overflows four bytes (roughly 497 days).
41
(From Oracle RDBMS documentation )
It is interesting that the periods are different for Oracle for win32 and for Linux. Will we be able to find
the function that generates this value?
As we can see, this information is finally taken from the X$KSUTM table.
SQL> select * from V$FIXED_VIEW_DEFINITION where view_name='V$TIMER';
VIEW_NAME
------------------------------
VIEW_DEFINITION
------------------------------
V$TIMER
select HSECS from GV$TIMER where inst_id = USERENV('Instance')
SQL> select * from V$FIXED_VIEW_DEFINITION where view_name='GV$TIMER';
VIEW_NAME
------------------------------
VIEW_DEFINITION
------------------------------
GV$TIMER
select inst_id,ksutmtim from x$ksutm
Now we are stuck in a small problem, there are no references to value generating function(s) in the tables
kqftab/kqftap:

kqftab_element.name: [X$KSUTM] ?: [ksutm] 0x1 0x4 0x4 0x0 0xffffc09b 0x3
kqftap_param.name=[INDX] ?: 0x20b02 0x0 0x0 0x0 0x4 0x0 0x0
kqftap_param.name=[KSUTMTIM] ?: 0x1302 0x0 0x0 0x0 0x4 0x0 0x1e
When we try to find the string KSUTMTIM, we see it in this function:

kqfd_DRN_ksutm_c proc near ; DATA XREF: .rodata:0805B4E8
arg_0 = dword ptr 8

push ebp
mov ebp, esp
push [ebp+arg_C]
push offset ksugtm
push offset _2__STRING_1263_0 ; "KSUTMTIM"
push [ebp+arg_8]
push [ebp+arg_0]
call kqfd_cfui_drain
add esp, 14h
mov esp, ebp
pop ebp
retn
kqfd_DRN_ksutm_c endp
884
The kqfd_DRN_ksutm_c() function is mentioned in the
kqfd_tab_registry_0 table:
dd offset _2__STRING_62_0 ; "X$KSUTM"
dd offset kqfd_OPN_ksutm_c
dd offset kqfd_tabl_fetch
dd 0
dd 0
dd offset kqfd_DRN_ksutm_c
There is a function ksugtm() referenced here. Let’s see what’s in it (Linux x86):
Listing 8.22: ksu.o

ksugtm proc near
var_1C = byte ptr -1Ch

push ebp
mov ebp, esp
sub esp, 1Ch
lea eax, [ebp+var_1C]
push eax
call slgcs
pop ecx
mov [edx], eax
mov eax, 4
mov esp, ebp
pop ebp
retn
ksugtm endp
The code in the win32 version is almost the same.

Is this the function we are looking for? Let’s see:
tracer -a:oracle.exe bpf=oracle.exe!_ksugtm,args:2,dump_args:0x4
Let’s try again:

SQL> select * from V$TIMER;
HSECS
----------
27294929
HSECS
----------
27295006
HSECS
----------
27295167
Listing 8.23: tracer output

TID=2428|(0) oracle.exe!_ksugtm (0x0, 0xd76c5f0) (called from oracle.exe!__VInfreq__qerfxFetch⤦
Ç +0xfad (0x56bb6d5))
Argument 2/2
0D76C5F0: 38 C9 "8. "
TID=2428|(0) oracle.exe!_ksugtm () -> 0x4 (0x4)
00000000: D1 7C A0 01 ".|.. "
885
Argument 2/2
0D76C5F0: 38 C9 "8. "
00000000: 1E 7D A0 01 ".}.. "
Argument 2/2
0D76C5F0: 38 C9 "8. "
00000000: BF 7D A0 01 ".}.. "
Indeed—the value is the same we see in SQL*Plus and it is returned via the second argument.
Let’s see what is in slgcs() (Linux x86):
slgcs proc near

arg_0 = dword ptr 8
push ebp
mov ebp, esp
push esi
mov [ebp+var_4], ebx
call $+5
pop ebx
nop ; PIC mode
mov ebx, offset _GLOBAL_OFFSET_TABLE_
call sltrgatime64 ; PIC mode
push 0
push 0Ah
push edx
push eax
call __udivdi3 ; PIC mode
add esp, 10h
mov esp, ebp
pop ebp
retn
slgcs endp
(it is just a call to sltrgatime64()

and division of its result by 10 ( 3.12 on page 496))
And win32-version:
_slgcs proc near ; CODE XREF: _dbgefgHtElResetCount+15
; _dbgerRunActions+1528
db 66h
nop
push ebp
mov ebp, esp
mov eax, [ebp+8]
call ds:__imp__GetTickCount@0 ; GetTickCount()
mov edx, eax
mov eax, 0CCCCCCCDh
mul edx
shr edx, 3
mov eax, edx
mov esp, ebp
pop ebp
retn
_slgcs endp
886
42
It is just the result of GetTickCount() divided by 10 ( 3.12 on page 496).
Voilà! That’s why the win32 version and the Linux x86 version show different results, because they are
generated by different OS functions.
Drain apparently implies connecting a specific table column to a specific function.
We will add support of the table kqfd_tab_registry_0 to oracle tables43 , now we can see how the table
column’s variables are connected to a specific functions:
[X$KSUTM] [kqfd_OPN_ksutm_c] [kqfd_tabl_fetch] [NULL] [NULL] [kqfd_DRN_ksutm_c]
[X$KSUSGIF] [kqfd_OPN_ksusg_c] [kqfd_tabl_fetch] [NULL] [NULL] [kqfd_DRN_ksusg_c]
OPN, apparently stands for, open, and DRN, apparently, for drain.
8.14 Handwritten assembly code

8.14.1 EICAR test file
This .COM-file is intended for testing antivirus software, it is possible to run in in MS-DOS and it prints this
string: “EICAR-STANDARD-ANTIVIRUS-TEST-FILE!”.
Its important property is that it consists entirely of printable ASCII-symbols, which, in turn, makes it pos-
sible to create it in any text editor:
X5O!P%@AP[4\PZX54(P^)7CC)7}$EICAR-STANDARD-ANTIVIRUS-TEST-FILE!$H+H*
Let’s decompile it:

; initial conditions: SP=0FFFEh, SS:[SP]=0
0100 58 pop ax
; AX=0, SP=0
0101 35 4F 21 xor ax, 214Fh
; AX = 214Fh and SP = 0
0104 50 push ax
; AX = 214Fh, SP = FFFEh and SS:[FFFE] = 214Fh
0105 25 40 41 and ax, 4140h
; AX = 140h, SP = FFFEh and SS:[FFFE] = 214Fh
0108 50 push ax
; AX = 140h, SP = FFFCh, SS:[FFFC] = 140h and SS:[FFFE] = 214Fh
0109 5B pop bx
; AX = 140h, BX = 140h, SP = FFFEh and SS:[FFFE] = 214Fh
010A 34 5C xor al, 5Ch
; AX = 11Ch, BX = 140h, SP = FFFEh and SS:[FFFE] = 214Fh
010C 50 push ax
010D 5A pop dx
; AX = 11Ch, BX = 140h, DX = 11Ch, SP = FFFEh and SS:[FFFE] = 214Fh
010E 58 pop ax
; AX = 214Fh, BX = 140h, DX = 11Ch and SP = 0
010F 35 34 28 xor ax, 2834h
; AX = 97Bh, BX = 140h, DX = 11Ch and SP = 0
0112 50 push ax
0113 5E pop si
; AX = 97Bh, BX = 140h, DX = 11Ch, SI = 97Bh and SP = 0
0114 29 37 sub [bx], si
0116 43 inc bx
0117 43 inc bx
0118 29 37 sub [bx], si
011A 7D 24 jge short near ptr word_10140
011C 45 49 43 ... db 'EICAR-STANDARD-ANTIVIRUS-TEST-FILE!$'
0140 48 2B word_10140 dw 2B48h ; CD 21 (INT 21) will be here
0142 48 2A dw 2A48h ; CD 20 (INT 20) will be here
0144 0D db 0Dh
0145 0A db 0Ah
We will add comments about the registers and stack after each instruction.
Essentially, all these instructions are here only to execute this code:
42 MSDN
43 yurichev.com
887
B4 09 MOV AH, 9
BA 1C 01 MOV DX, 11Ch
CD 21 INT 21h
CD 20 INT 20h
INT 21h with 9th function (passed in AH) just prints a string, the address of which is passed in DS:DX. By
the way, the string has to be terminated with the ’$’ sign. Apparently, it’s inherited from CP/M and this
function was left in DOS for compatibility. INT 20h exits to DOS.
But as we can see, these instruction’s opcodes are not strictly printable. So the main part of EICAR file is:
• preparing the register (AH and DX) values that we need;
• preparing INT 21 and INT 20 opcodes in memory;
• executing INT 21 and INT 20.
By the way, this technique is widely used in shellcode construction, when one have to pass x86 code in
string form.
Here is also a list of all x86 instructions which have printable opcodes: .1.6 on page 1023.
8.15 Demos
Demos (or demomaking) were an excellent exercise in mathematics, computer graphics programming
and very tight x86 hand coding.
8.15.1 10 PRINT CHR$(205.5+RND(1)); : GOTO 10

All examples here are MS-DOS .COM files.
44
In [Nick Montfort et al, 10 PRINT CHR$(205.5+RND(1)); : GOTO 10, (The MIT Press:2012)]
we can read about one of the most simple possible random maze generators.
It just prints a slash or backslash characters randomly and endlessly, resulting in something like this:
There are a few known implementations for 16-bit x86.
Trixter’s 42 byte version
The listing was taken from his website45 , but the comments are mine.
888
00000000: B001 mov al,1 ; set 40x25 video mode
00000002: CD10 int 010
00000004: 30FF xor bh,bh ; set video page for int 10h call
00000006: B9D007 mov cx,007D0 ; 2000 characters to output
00000009: 31C0 xor ax,ax
0000000B: 9C pushf ; push flags
; get random value from timer chip
0000000C: FA cli ; disable interrupts
0000000D: E643 out 043,al ; write 0 to port 43h
; read 16-bit value from port 40h
0000000F: E440 in al,040
00000011: 88C4 mov ah,al
00000013: E440 in al,040
00000015: 9D popf ; enable interrupts by restoring IF flag
00000016: 86C4 xchg ah,al
; here we have 16-bit pseudorandom value
00000018: D1E8 shr ax,1
0000001A: D1E8 shr ax,1
; CF currently have second bit from the value
0000001C: B05C mov al,05C ;'́
; if CF=1, skip the next instruction
0000001E: 7202 jc 000000022
; if CF=0, reload AL register with another character
00000020: B02F mov al,02F ;'/'
; output character
00000022: B40E mov ah,00E
00000024: CD10 int 010
00000026: E2E1 loop 000000009 ; loop 2000 times
00000028: CD20 int 020 ; exit to DOS
The pseudo-random value here is in fact the time that has passed from the system’s boot, taken from the
8253 time chip, the value increases by one 18.2 times per second.
By writing zero to port 43h, we send the command “select counter 0”, ”counter latch”, ”binary counter”
(not a BCD value).
The interrupts are enabled back with the POPF instruction, which restores the IF flag as well.
It is not possible to use the IN instruction with registers other than AL, hence the shuffling.
My attempt to reduce Trixter’s version: 27 bytes
We can say that since we use the timer not to get a precise time value, but a pseudo-random one, we do
not need to spend time (and code) to disable the interrupts.
Another thing we can say is that we need only one bit from the low 8-bit part, so let’s read only it.
We can reduced the code slightly and we’ve got 27 bytes:
00000000: B9D007 mov cx,007D0 ; limit output to 2000 characters
00000003: 31C0 xor ax,ax ; command to timer chip
00000005: E643 out 043,al
00000007: E440 in al,040 ; read 8-bit of timer
00000009: D1E8 shr ax,1 ; get second bit to CF flag
0000000B: D1E8 shr ax,1
0000000D: B05C mov al,05C ; prepare '\'
0000000F: 7202 jc 000000013
00000011: B02F mov al,02F ; prepare '/'
; output character to screen
00000013: B40E mov ah,00E
00000015: CD10 int 010
00000017: E2EA loop 000000003
; exit to DOS
00000019: CD20 int 020
Taking random memory garbage as a source of randomness
Since it is MS-DOS, there is no memory protection at all, we can read from whatever address we want.
Even more than that: a simple LODSB instruction reads a byte from the DS:SI address, but it’s not a
889
problem if the registers’ values are not set up, let it read 1) random bytes; 2) from a random place in
memory!
It is suggested in Trixter’s webpage46 to use LODSB without any setup.
It is also suggested that the SCASB instruction can be used instead, because it sets a flag according to the
byte it reads.
Another idea to minimize the code is to use the INT 29h DOS syscall, which just prints the character stored
in the AL register.
47
That is what Peter Ferrie did :
Listing 8.24: Peter Ferrie: 10 bytes

; AL is random at this point
00000000: AE scasb
; CF is set according subtracting random memory byte from AL.
; so it is somewhat random at this point
00000001: D6 setalc
; AL is set to 0xFF if CF=1 or to 0 if otherwise
00000002: 242D and al,02D ;'-'
; AL here is 0x2D or 0
00000004: 042F add al,02F ;'/'
; AL here is 0x5C or 0x2F
00000006: CD29 int 029 ; output AL to screen
00000008: EBF6 jmps 000000000 ; loop endlessly
So it is possible to get rid of conditional jumps at all. The ASCII code of backslash (“\”) is 0x5C and 0x2F
for slash (“/”). So we have to convert one (pseudo-random) bit in the CF flag to a value of 0x5C or 0x2F.
This is done easily: by AND-ing all bits in AL (where all 8 bits are set or cleared) with 0x2D we have just 0
or 0x2D.
By adding 0x2F to this value, we get 0x5C or 0x2F.
Then we just output it to the screen.
Conclusion
It is also worth mentioning that the result may be different in DOSBox, Windows NT and even MS-DOS,
due to different conditions: the timer chip can be emulated differently and the initial register contents
may be different as well.
890
8.15.2 Mandelbrot set
You know, if you magnify the coastline, it still
looks like a coastline, and a lot of other
things have this property. Nature has
recursive algorithms that it uses to generate
clouds and Swiss cheese and things like that.
Donald Knuth, interview (1993)
Mandelbrot set is a fractal, which exhibits self-similarity.

When you increase scale, you see that this characteristic pattern repeating infinitely.
Here is a demo48 written by “Sir_Lagsalot” in 2009, that draws the Mandelbrot set, which is just a x86
program with executable file size of only 64 bytes. There are only 30 16-bit x86 instructions.
Here it is what it draws:
Let’s try to understand how it works.
Theory
A word about complex numbers
A complex number is a number that consists of two parts—real (Re) and imaginary (Im).
The complex plane is a two-dimensional plane where any complex number can be placed: the real part is
one coordinate and the imaginary part is the other.
Some basic rules we have to keep in mind:
• Addition: (a + bi) + (c + di) = (a + c) + (b + d)i
In other words:
Re(sum) = Re(a) + Re(b)
48 Download it here,
891
Im(sum) = Im(a) + Im(b)
• Multiplication: (a + bi)(c + di) = (ac − bd) + (bc + ad)i
In other words:
Re(product) = Re(a) ⋅ Re(c) − Re(b) ⋅ Re(d)
Im(product) = Im(b) ⋅ Im(c) + Im(a) ⋅ Im(d)
• Square: (a + bi)2 = (a + bi)(a + bi) = (a2 − b2 ) + (2ab)i
In other words:
Re(square) = Re(a)2 − Im(a)2
Im(square) = 2 ⋅ Re(a) ⋅ Im(a)
How to draw the Mandelbrot set
The Mandelbrot set is a set of points for which the zn+1 = zn 2 + c recursive sequence (where z and c are
complex numbers and c is the starting value) does not approach infinity.
In plain English language:

• Enumerate all points on screen.
• Check if the specific point is in the Mandelbrot set.
• Here is how to check it:
– Represent the point as a complex number.
– Calculate the square of it.
– Add the starting value of the point to it.
– Does it go off limits? If yes, break.
– Move the point to the new place at the coordinates we just calculated.
– Repeat all this for some reasonable number of iterations.
• The point is still in limits? Then draw the point.
• The point has eventually gone off limits?
– (For a black-white image) do not draw anything.
– (For a colored image) transform the number of iterations to some color. So the color shows the
speed with which point has gone off limits.
Here is Pythonesque algorithm for both complex and integer number representations:
Listing 8.25: For complex numbers

def check_if_is_in_set(P):
P_start=P
iterations=0
while True:
if (P>bounds):
break
P=P^2+P_start
if iterations > max_iterations:
break
iterations++
return iterations
# black-white
for each point on screen P:
if check_if_is_in_set (P) < max_iterations:
draw point
# colored
for each point on screen P:
892
iterations = if check_if_is_in_set (P)
map iterations to color
draw color point
The integer version is where the operations on complex numbers are replaced with integer operations
according to the rules which were explained above.
Listing 8.26: For integer numbers

def check_if_is_in_set(X, Y):
X_start=X
Y_start=Y
iterations=0
while True:
if (X^2 + Y^2 > bounds):
break
new_X=X^2 - Y^2 + X_start
new_Y=2*X*Y + Y_start
if iterations > max_iterations:
break
iterations++
return iterations
# black-white
for X = min_X to max_X:
for Y = min_Y to max_Y:
if check_if_is_in_set (X,Y) < max_iterations:
draw point at X, Y
# colored
for X = min_X to max_X:
for Y = min_Y to max_Y:
iterations = if check_if_is_in_set (X,Y)
map iterations to color
draw color point at X,Y
Here is also a C# source which is present in the Wikipedia article49 , but we’ll modify it so it will print the
iteration numbers instead of some symbol 50 :
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
namespace Mnoj
{
class Program
{
static void Main(string[] args)
{
double realCoord, imagCoord;
double realTemp, imagTemp, realTemp2, arg;
int iterations;
for (imagCoord = 1.2; imagCoord >= -1.2; imagCoord -= 0.05)
{
for (realCoord = -0.6; realCoord <= 1.77; realCoord += 0.03)
{
iterations = 0;
realTemp = realCoord;
imagTemp = imagCoord;
arg = (realCoord * realCoord) + (imagCoord * imagCoord);
while ((arg < 2*2) && (iterations < 40))
{
realTemp2 = (realTemp * realTemp) - (imagTemp * imagTemp) - realCoord;
imagTemp = (2 * realTemp * imagTemp) - imagCoord;
49 wikipedia
50 Here is also the executable file: beginners.re
893
realTemp = realTemp2;
arg = (realTemp * realTemp) + (imagTemp * imagTemp);
iterations += 1;
}
Console.Write("{0,2:D} ", iterations);
}
Console.Write("\n");
}
Console.ReadKey();
}
}
}
Here is the resulting file, which is too wide to be included here:

beginners.re.
The maximal number of iterations is 40, so when you see 40 in this dump, it means that this point has
been wandering for 40 iterations but never got off limits.
A number n less than 40 means that point remained inside the bounds only for n iterations, then it went
outside them.
894
There is a cool demo available at http://go.yurichev.com/17309, which shows visually how the point
moves on the plane at each iteration for some specific point. Here are two screenshots.
First, we’ve clicked inside the yellow area and saw that the trajectory (green line) eventually swirls at
some point inside:
Figure 8.18: Click inside yellow area
This implies that the point we’ve clicked belongs to the Mandelbrot set.
895
Then we’ve clicked outside the yellow area and saw a much more chaotic point movement, which quickly
went off bounds:
Figure 8.19: Click outside yellow area
This means the point doesn’t belong to Mandelbrot set.

Another good demo is available here: http://go.yurichev.com/17310.
896
Let’s get back to the demo
The demo, although very tiny (just 64 bytes or 30 instructions), implements the common algorithm de-
scribed here, but using some coding tricks.
The source code is easily downloadable, so here is it, but let’s also add comments:
Listing 8.27: Commented source code

1 ; X is column on screen
2 ; Y is row on screen
3
4
5 ; X=0, Y=0 X=319, Y=0
6 ; +------------------------------->
7 ; |
8 ; |
9 ; |
10 ; |
11 ; |
12 ; |
13 ; v
14 ; X=0, Y=199 X=319, Y=199
15
16
17 ; switch to VGA 320*200*256 graphics mode
18 mov al,13h
19 int 10h
20 ; initial BX is 0
21 ; initial DI is 0xFFFE
22 ; DS:BX (or DS:0) is pointing to Program Segment Prefix at this moment
23 ; ... first 4 bytes of which are CD 20 FF 9F
24 les ax,[bx]
25 ; ES:AX=9FFF:20CD
26
27 FillLoop:
28 ; set DX to 0. CWD works as: DX:AX = sign_extend(AX).
29 ; AX here 0x20CD (at startup) or less then 320 (when getting back after loop),
30 ; so DX will always be 0.
31 cwd
32 mov ax,di
33 ; AX is current pointer within VGA buffer
34 ; divide current pointer by 320
35 mov cx,320
36 div cx
37 ; DX (start_X) - remainder (column: 0..319); AX - result (row: 0..199)
38 sub ax,100
39 ; AX=AX-100, so AX (start_Y) now is in range -100..99
40 ; DX is in range 0..319 or 0x0000..0x013F
41 dec dh
42 ; DX now is in range 0xFF00..0x003F (-256..63)
43
44 xor bx,bx
45 xor si,si
46 ; BX (temp_X)=0; SI (temp_Y)=0
47
48 ; get maximal number of iterations
49 ; CX is still 320 here, so this is also maximal number of iteration
50 MandelLoop:
51 mov bp,si ; BP = temp_Y
52 imul si,bx ; SI = temp_X*temp_Y
53 add si,si ; SI = SI*2 = (temp_X*temp_Y)*2
54 imul bx,bx ; BX = BX^2 = temp_X^2
55 jo MandelBreak ; overflow?
56 imul bp,bp ; BP = BP^2 = temp_Y^2
58 add bx,bp ; BX = BX+BP = temp_X^2 + temp_Y^2
60 sub bx,bp ; BX = BX-BP = temp_X^2 + temp_Y^2 - temp_Y^2 = temp_X^2
61 sub bx,bp ; BX = BX-BP = temp_X^2 - temp_Y^2
62
897
63 ; correct scale:
64 sar bx,6 ; BX=BX/64
65 add bx,dx ; BX=BX+start_X
66 ; now temp_X = temp_X^2 - temp_Y^2 + start_X
67 sar si,6 ; SI=SI/64
68 add si,ax ; SI=SI+start_Y
69 ; now temp_Y = (temp_X*temp_Y)*2 + start_Y
70
71 loop MandelLoop
72
73 MandelBreak:
74 ; CX=iterations
75 xchg ax,cx
76 ; AX=iterations. store AL to VGA buffer at ES:[DI]
77 stosb
78 ; stosb also increments DI, so DI now points to the next point in VGA buffer
79 ; jump always, so this is eternal loop here
80 jmp FillLoop
Algorithm:
• Switch to 320*200 VGA video mode, 256 colors. 320 ∗ 200 = 64000 (0xFA00).
Each pixel is encoded by one byte, so the buffer size is 0xFA00 bytes. It is addressed using the ES:DI
registers pair.
ES must be 0xA000 here, because this is the segment address of the VGA video buffer, but storing
0xA000 to ES requires at least 4 bytes (PUSH 0A000h / POP ES). You can read more about the 16-bit
MS-DOS memory model here: 11.6 on page 986.
Assuming that BX is zero here, and the Program Segment Prefix is at the zeroth address, the 2-byte
LES AX,[BX] instruction stores 0x20CD to AX and 0x9FFF to ES.
So the program starts to draw 16 pixels (or bytes) before the actual video buffer. But this is MS-DOS,
there is no memory protection, so a write happens into the very end of conventional memory, and
usually, there is nothing important. That’s why you see a red strip 16 pixels wide at the right side.
The whole picture is shifted left by 16 pixels. This is the price of saving 2 bytes.
• An infinite loop processes each pixel.
Probably, the most common way to enumerate all pixels on the screen is with two loops: one for
the X coordinate, another for the Y coordinate. But then you’ll need to multiply the coordinates to
address a byte in the VGA video buffer.
The author of this demo decided to do it otherwise: enumerate all bytes in the video buffer by using
one single loop instead of two, and get the coordinates of the current point using division. The
resulting coordinates are: X in the range of −256..63 and Y in the range of −100..99. You can see on the
screenshot that the picture is somewhat shifted to the right part of screen.
That’s because the biggest heart-shaped black hole usually appears on coordinates 0,0 and these
are shifted here to right. Could the author just subtract 160 from the value to get X in the range of
−160..159? Yes, but the instruction SUB DX, 160 takes 4 bytes, while DEC DH—2 bytes (which subtracts
0x100 (256) from DX). So the whole picture is shifted for the cost of another 2 bytes of saved space.
– Check, if the current point is inside the Mandelbrot set. The algorithm is the one that has been
described here.
– The loop is organized using the LOOP instruction, which uses the CX register as counter.
The author could set the number of iterations to some specific number, but he didn’t: 320 is
already present in CX (has been set at line 35), and this is good maximal iteration number
anyway. We save here some space by not the reloading CX register with another value.
– IMUL is used here instead of MUL, because we work with signed values: keep in mind that the
0,0 coordinates has to be somewhere near the center of the screen.
It’s the same with SAR (arithmetic shift for signed values): it’s used instead of SHR.
– Another idea is to simplify the bounds check. We must check a coordinate pair, i.e., two variables.
What the author does is to checks thrice for overflow: two squaring operations and one addition.
898
Indeed, we use 16-bit registers, which hold signed values in the range of -32768..32767, so if any
of the coordinates is greater than 32767 during the signed multiplication, this point is definitely
out of bounds: we jump to the MandelBreak label.
– There is also a division by 64 (SAR instruction). 64 sets scale.
Try to increase the value and you can get a closer look, or to decrease if for a more distant look.
• We are at the MandelBreak label, there are two ways of getting here: the loop ended with CX=0 (
the point is inside the Mandelbrot set); or because an overflow has happened (CX still holds some
value). Now we write the low 8-bit part of CX (CL) to the video buffer.
The default palette is rough, nevertheless, 0 is black: hence we see black holes in the places where
the points are in the Mandelbrot set. The palette can be initialized at the program’s start, but keep
in mind, this is only a 64 bytes program!
• The program runs in an infinite loop, because an additional check where to stop, or any user interface
will result in additional instructions.
Some other optimization tricks:
• The 1-byte CWD is used here for clearing DX instead of the 2-byte XOR DX, DX or even the 3-byte
MOV DX, 0.
• The 1-byte XCHG AX, CX is used instead of the 2-byte MOV AX,CX. The current value of AX is not
needed here anyway.
51
• DI (position in video buffer) is not initialized, and it is 0xFFFE at the start .
That’s OK, because the program works for all DI in the range of 0..0xFFFF eternally, and the user
can’t notice that it is started off the screen (the last pixel of a 320*200 video buffer is at address
0xF9FF). So some work is actually done off the limits of the screen.
Otherwise, you’ll need an additional instructions to set DI to 0 and check for the video buffer’s end.
My “fixed” version
Listing 8.28: My “fixed” version

1 org 100h
2 mov al,13h
3 int 10h
4
5 ; set palette
6 mov dx, 3c8h
7 mov al, 0
8 out dx, al
9 mov cx, 100h
10 inc dx
11 l00:
12 mov al, cl
13 shl ax, 2
14 out dx, al ; red
15 out dx, al ; green
16 out dx, al ; blue
17 loop l00
18
19 push 0a000h
20 pop es
21
22 xor di, di
23
24 FillLoop:
25 cwd
26 mov ax,di
27 mov cx,320
28 div cx
29 sub ax,100
30 sub dx,160
31
32 xor bx,bx
51 More information about initial register values: http://go.yurichev.com/17004
899
33 xor si,si
34
35 MandelLoop:
36 mov bp,si
37 imul si,bx
38 add si,si
39 imul bx,bx
40 jo MandelBreak
41 imul bp,bp
42 jo MandelBreak
43 add bx,bp
44 jo MandelBreak
45 sub bx,bp
46 sub bx,bp
47
48 sar bx,6
49 add bx,dx
50 sar si,6
51 add si,ax
52
53 loop MandelLoop
54
55 MandelBreak:
56 xchg ax,cx
57 stosb
58 cmp di, 0FA00h
59 jb FillLoop
60
61 ; wait for keypress
62 xor ax,ax
63 int 16h
64 ; set text video mode
65 mov ax, 3
66 int 10h
67 ; exit
68 int 20h
The author of these lines made an attempt to fix all these oddities: now the palette is smooth grayscale,
the video buffer is at the correct place (lines 19..20), the picture is drawn on center of the screen (line 30),
the program eventually ends and waits for the user’s keypress (lines 58..68).
52
But now it’s much bigger: 105 bytes (or 54 instructions) .
52 You can experiment by yourself: get DosBox and NASM and compile it as: nasm file.asm -fbin -o file.com
900
Figure 8.20: My “fixed” version
See also: small C program printing Mandelbrot set in ASCII: https://people.sc.fsu.edu/~jburkardt/

c_src/mandelbrot_ascii/mandelbrot_ascii.html
https://miyuki.github.io/2017/10/04/gcc-archaeology-1.html.
8.16 A nasty bug in MSVCRT.DLL

This is the bug that costed me several hours of debugging.
In 2013 I was using MinGW, my C project seems to be very unstable and I saw the “Invalid parameter
passed to C runtime function.” error message in debugger.
The error message was also visible using Sysinternals DebugView. And my project has no such error mes-
sages or strings. So I started to search it in the whole Windows and found in MSVCRT.DLL file. (Needless
to say I was using Windows 7.)
So here it is, the error message in MSVCRT.DLL file supplied with Windows 7:
.text:6FFB69D0 OutputString db 'Invalid parameter passed to C runtime function.',0Ah,0
.text:6FFB69D0 ; DATA XREF: sub_6FFB6930+83
Where it is referenced?
.text:6FFB6930 sub_6FFB6930 proc near ; CODE XREF: _wfindfirst64+203FC
.text:6FFB6930 ; sub_6FF62563+319AD
.text:6FFB6930
.text:6FFB6930 var_2D0 = dword ptr -2D0h
.text:6FFB6930 var_244 = word ptr -244h
.text:6FFB6930 var_23C = word ptr -23Ch
.text:6FFB6930 var_234 = dword ptr -234h
901
.text:6FFB6930 var_22C = dword ptr -22Ch
.text:6FFB6930 var_4 = dword ptr -4
.text:6FFB6930
.text:6FFB6930 mov edi, edi
.text:6FFB6932 push ebp
.text:6FFB6933 mov ebp, esp
.text:6FFB6935 sub esp, 2D0h
.text:6FFB693B mov eax, ___security_cookie
.text:6FFB6940 xor eax, ebp
.text:6FFB6942 mov [ebp+var_4], eax
.text:6FFB694B mov [ebp+var_224], ecx
.text:6FFB6951 mov [ebp+var_228], edx
.text:6FFB6957 mov [ebp+var_22C], ebx
.text:6FFB695D mov [ebp+var_230], esi
.text:6FFB6963 mov [ebp+var_234], edi
.text:6FFB6969 mov [ebp+var_208], ss
.text:6FFB696F mov [ebp+var_214], cs
.text:6FFB6975 mov [ebp+var_238], ds
.text:6FFB697B mov [ebp+var_23C], es
.text:6FFB6981 mov [ebp+var_240], fs
.text:6FFB6987 mov [ebp+var_244], gs
.text:6FFB698D pushf
.text:6FFB698E pop [ebp+var_210]
.text:6FFB6994 mov eax, [ebp+4]
.text:6FFB699D lea eax, [ebp+4]
.text:6FFB69A0 mov [ebp+var_2D0], 10001h
.text:6FFB69AA mov [ebp+var_20C], eax
.text:6FFB69B0 mov eax, [eax-4]
.text:6FFB69B3 push offset OutputString ; "Invalid parameter passed to C
runtime f"...
.text:6FFB69B8 mov [ebp+var_21C], eax
.text:6FFB69BE call ds:OutputDebugStringA
.text:6FFB69C4 mov ecx, [ebp+var_4]
.text:6FFB69C7 xor ecx, ebp
.text:6FFB69C9 call @__security_check_cookie@4 ; __security_check_cookie(x)
.text:6FFB69CE leave
.text:6FFB69CF retn
.text:6FFB69CF sub_6FFB6930 endp
The string it reported into debugger or DebugView utility using the standard OutputDebugStringA() func-
tion. How the sub_6FFB6930 can be called? IDA shows at least 280 references.
Using my tracer, I set a breakpoint at sub_6FFB6930 to see, when it’s called in my case:
tracer.exe -l:1.exe bpf=msvcrt.dll!0x6FFB6930 -s
...
PID=3560|New process 1.exe

(0) msvcrt.dll!0x6ffb6930() (called from msvcrt.dll!_ftol2_sse_excpt+0x1b467 (0x759ed222))
Call stack:
return address=0x401010 (1.exe!.text+0x10), arguments in stack: 0x12ff14, 0x401010, 0x403010("⤦
Ç asd"), 0x0, 0x12ff88, 0x4010f8
return address=0x4010f8 (1.exe!OEP+0xe3), arguments in stack: 0x12ff88, 0x4010f8, 0x1, 0x2e0ea8⤦
Ç , 0x2e1640, 0x403000
return address=0x75b6ef3c (KERNEL32.dll!BaseThreadInitThunk+0x12), arguments in stack: 0x12ff94⤦
Ç , 0x75b6ef3c, 0x7ffdf000, 0x12ffd4, 0x77523688, 0x7ffdf000
return address=0x77523688 (ntdll.dll!RtlInitializeExceptionChain+0xef), arguments in stack: 0⤦
Ç x12ffd4, 0x77523688, 0x7ffdf000, 0x74117ec7, 0x0, 0x0
902
return address=0x7752365b (ntdll.dll!RtlInitializeExceptionChain+0xc2), arguments in stack: 0⤦
Ç x12ffec, 0x7752365b, 0x401015, 0x7ffdf000, 0x0, 0x0
(0) msvcrt.dll!0x6ffb6930() -> 0x12f94c
PID=3560|Process 1.exe exited. ExitCode=2147483647 (0x7fffffff)
I found that my code was calling stricmp() function with NULL as one argument. In fact, I made up this
example when writing this:
#include <stdio.h>
#include <string.h>
int main()
{
stricmp ("asd", NULL);
};
If this piece of code is compiled using old MinGW or old MSVC 6.0, it is linked against MSVCRT.DLL file.
Which, as of Windows 7, silently sends the “Invalid parameter passed to C runtime function.” error mes-
sage to the debugger and then does nothing!
Let’s see how stricmp() is implemented in MSVCRT.DLL:
.text:6FF5DB38 ; Exported entry 855. _strcmpi
.text:6FF5DB38 ; Exported entry 863. _stricmp
.text:6FF5DB38
.text:6FF5DB38 ; =============== S U B R O U T I N E =======================================
.text:6FF5DB38
.text:6FF5DB38 ; Attributes: bp-based frame
.text:6FF5DB38
.text:6FF5DB38 ; int __cdecl strcmpi(const char *, const char *)
.text:6FF5DB38 public _strcmpi
.text:6FF5DB38 _strcmpi proc near ; CODE XREF: LocaleEnumProc-2B
.text:6FF5DB38 ; LocaleEnumProc+5E
.text:6FF5DB38
.text:6FF5DB38 arg_0 = dword ptr 8
.text:6FF5DB38 arg_4 = dword ptr 0Ch
.text:6FF5DB38
.text:6FF5DB38 ; FUNCTION CHUNK AT .text:6FF68CFD SIZE 00000012 BYTES
.text:6FF5DB38 ; FUNCTION CHUNK AT .text:6FF9D20D SIZE 00000022 BYTES
.text:6FF5DB38
.text:6FF5DB38 mov edi, edi ; _strcmpi
.text:6FF5DB3A push ebp
.text:6FF5DB3B mov ebp, esp
.text:6FF5DB3D push esi
.text:6FF5DB3E xor esi, esi
.text:6FF5DB40 cmp dword_6FFF0000, esi
.text:6FF5DB46 jnz loc_6FF68CFD
.text:6FF5DB4C cmp [ebp+arg_0], esi ; is arg_0==NULL?
.text:6FF5DB4F jz loc_6FF9D20D
.text:6FF5DB55 cmp [ebp+arg_4], esi ; is arg_0==NULL?
.text:6FF5DB58 jz loc_6FF9D20D
.text:6FF5DB5E pop esi
.text:6FF5DB5F pop ebp
.text:6FF5DB5F _strcmpi endp ; sp-analysis failed
Actual strings comparison here:

.text:6FF5DB60 sub_6FF5DB60 proc near ; CODE XREF: _stricmp_l+16C7F
.text:6FF5DB60 ; sub_6FFD19CD+229
.text:6FF5DB60
.text:6FF5DB60 arg_0 = dword ptr 8
.text:6FF5DB60 arg_4 = dword ptr 0Ch
.text:6FF5DB60
.text:6FF5DB60 push ebp
.text:6FF5DB61 mov ebp, esp
.text:6FF5DB63 push edi
.text:6FF5DB64 push esi
.text:6FF5DB65 push ebx
.text:6FF5DB66 mov esi, [ebp+arg_4]
.text:6FF5DB69 mov edi, [ebp+arg_0]
903
.text:6FF5DB6C mov al, 0FFh
.text:6FF5DB6E mov edi, edi
.text:6FF5DB70
.text:6FF5DB70 loc_6FF5DB70: ; CODE XREF: sub_6FF5DB60+20
.text:6FF5DB70 ; sub_6FF5DB60+40
.text:6FF5DB70 or al, al
.text:6FF5DB72 jz short loc_6FF5DBA6
.text:6FF5DB74 mov al, [esi]
.text:6FF5DB76 add esi, 1
.text:6FF5DB79 mov ah, [edi]
.text:6FF5DB7B add edi, 1
.text:6FF5DB7E cmp ah, al
.text:6FF5DB80 jz short loc_6FF5DB70
.text:6FF5DB82 sub al, 41h
.text:6FF5DB84 cmp al, 1Ah
.text:6FF5DB86 sbb cl, cl
.text:6FF5DB88 and cl, 20h
.text:6FF5DB8B add al, cl
.text:6FF5DB8D add al, 41h
.text:6FF5DB8F xchg ah, al
.text:6FF5DB91 sub al, 41h
.text:6FF5DB93 cmp al, 1Ah
.text:6FF5DB95 sbb cl, cl
.text:6FF5DB97 and cl, 20h
.text:6FF5DB9A add al, cl
.text:6FF5DB9C add al, 41h
.text:6FF5DB9E cmp al, ah
.text:6FF5DBA0 jz short loc_6FF5DB70
.text:6FF5DBA2 sbb al, al
.text:6FF5DBA4 sbb al, 0FFh
.text:6FF5DBA6
.text:6FF5DBA6 loc_6FF5DBA6: ; CODE XREF: sub_6FF5DB60+12
.text:6FF5DBA6 movsx eax, al
.text:6FF5DBA9 pop ebx
.text:6FF5DBAA pop esi
.text:6FF5DBAB pop edi
.text:6FF5DBAC leave
.text:6FF5DBAD retn
.text:6FF5DBAD sub_6FF5DB60 endp
.text:6FF68D0C loc_6FF68D0C: ; CODE XREF: _strcmpi+3F6F2

.text:6FF68D0C pop esi
.text:6FF68D0D pop ebp
.text:6FF68D0E retn
.text:6FF9D20D loc_6FF9D20D: ; CODE XREF: _strcmpi+17

.text:6FF9D20D ; _strcmpi+20
.text:6FF9D20D call near ptr _errno
.text:6FF9D212 push esi
.text:6FF9D217 mov dword ptr [eax], 16h
.text:6FF9D21D call _invalid_parameter
.text:6FF9D222 add esp, 14h
.text:6FF9D225 mov eax, 7FFFFFFFh
.text:6FF9D22A jmp loc_6FF68D0C
Now the invalid_parameter() function:

.text:6FFB6A06 public _invalid_parameter
.text:6FFB6A06 _invalid_parameter proc near ; CODE XREF: sub_6FF5B494:loc_6FF5B618
.text:6FFB6A06 ; sub_6FF5CCFD:loc_6FF5C8A2
.text:6FFB6A06 mov edi, edi
.text:6FFB6A08 push ebp
.text:6FFB6A09 mov ebp, esp
.text:6FFB6A0B pop ebp
.text:6FFB6A0C jmp sub_6FFB6930
.text:6FFB6A0C _invalid_parameter endp
904
The invalid_parameter() function eventually calls the function with OutputDebugStringA(): 8.16 on
page 901.
You see, the stricmp() code is like:
int stricmp(const char *s1, const char *s2, size_t len)
{
if (s1==NULL || s2==NULL)
{
// print error message AND exit:
return 0x7FFFFFFFh;
};
// do comparison
};
How come this error is rare? Because newer MSVC versions links against MSVCR120.DLL file, etc (where
120 is version number).
Let’s peek inside the newer MSVCR120.DLL from Windows 7:
.text:1002A0D4 public _stricmp_l
.text:1002A0D4 _stricmp_l proc near ; CODE XREF: _stricmp+18
.text:1002A0D4 ; _mbsicmp_l+47
.text:1002A0D4 ; DATA XREF: ...
.text:1002A0D4
.text:1002A0D4 var_10 = dword ptr -10h
.text:1002A0D4 var_8 = dword ptr -8
.text:1002A0D4 var_4 = byte ptr -4
.text:1002A0D4 arg_0 = dword ptr 8
.text:1002A0D4 arg_4 = dword ptr 0Ch
.text:1002A0D4 arg_8 = dword ptr 10h
.text:1002A0D4
.text:1002A0D4 ; FUNCTION CHUNK AT .text:1005AA7B SIZE 0000002A BYTES
.text:1002A0D4
.text:1002A0D4 push ebp
.text:1002A0D5 mov ebp, esp
.text:1002A0D7 sub esp, 10h
.text:1002A0DA lea ecx, [ebp+var_10]
.text:1002A0DD push ebx
.text:1002A0DE push esi
.text:1002A0DF push edi
.text:1002A0E0 push [ebp+arg_8]
.text:1002A0E3 call sub_1000F764
.text:1002A0E8 mov edi, [ebp+arg_0] ; arg==NULL?
.text:1002A0EB test edi, edi
.text:1002A0ED jz loc_1005AA7B
.text:1002A0F3 mov ebx, [ebp+arg_4] ; arg==NULL?
.text:1002A0F6 test ebx, ebx
.text:1002A0F8 jz loc_1005AA7B
.text:1002A0FE mov eax, [ebp+var_10]
.text:1002A101 cmp dword ptr [eax+0A8h], 0
.text:1002A108 jz loc_1005AA95
.text:1002A10E sub edi, ebx
...
.text:1005AA7B loc_1005AA7B: ; CODE XREF: _stricmp_l+19

.text:1005AA7B ; _stricmp_l+24
.text:1005AA7B call _errno
.text:1005AA80 mov dword ptr [eax], 16h
.text:1005AA86 call _invalid_parameter_noinfo
.text:1005AA8B mov esi, 7FFFFFFFh
.text:1005AA90 jmp loc_1002A13B
...
.text:100A4670 _invalid_parameter_noinfo proc near ; CODE XREF: sub_10013BEC-10F

.text:100A4670 ; sub_10016C0F-10F
905
.text:100A4677 call _invalid_parameter
.text:100A467C add esp, 14h
.text:100A467F retn
.text:100A467F _invalid_parameter_noinfo endp
...
.text:100A4645 _invalid_parameter proc near ; CODE XREF: _invalid_parameter(ushort

const *,ushort const *,ushort const *,uint,uint)
.text:100A4645 ; _invalid_parameter_noinfo+7
.text:100A4645
.text:100A4645 arg_0 = dword ptr 8
.text:100A4645 arg_4 = dword ptr 0Ch
.text:100A4645 arg_8 = dword ptr 10h
.text:100A4645 arg_C = dword ptr 14h
.text:100A4645 arg_10 = dword ptr 18h
.text:100A4645
.text:100A4645 push ebp
.text:100A4646 mov ebp, esp
.text:100A4648 push dword_100E0ED8 ; Ptr
.text:100A464E call ds:DecodePointer
.text:100A4654 test eax, eax
.text:100A4656 jz short loc_100A465B
.text:100A4658 pop ebp
.text:100A4659 jmp eax
.text:100A465B ; ---------------------------------------------------------------------------
.text:100A465B
.text:100A465B loc_100A465B: ; CODE XREF: _invalid_parameter+11
.text:100A465B push [ebp+arg_10]
.text:100A465E push [ebp+arg_C]
.text:100A466A call _invoke_watson
.text:100A466F int 3 ; Trap to Debugger
.text:100A466F _invalid_parameter endp
.text:100A469B _invoke_watson proc near ; CODE XREF: sub_1002CDB0+27068

.text:100A469B ; sub_10029704+2A792
.text:100A469B push 17h ; ProcessorFeature
.text:100A469D call IsProcessorFeaturePresent
.text:100A46A2 test eax, eax
.text:100A46A4 jz short loc_100A46AB
.text:100A46A6 push 5
.text:100A46A8 pop ecx
.text:100A46A9 int 29h ; Win8: RtlFailFast(ecx)
.text:100A46AB ; ---------------------------------------------------------------------------
.text:100A46AB
.text:100A46AB loc_100A46AB: ; CODE XREF: _invoke_watson+9
.text:100A46AB push esi
.text:100A46AC push 1
.text:100A46AE mov esi, 0C0000417h
.text:100A46B3 push esi
.text:100A46B4 push 2
.text:100A46B6 call sub_100A4519
.text:100A46BB push esi ; uExitCode
.text:100A46BC call __crtTerminateProcess
.text:100A46C1 add esp, 10h
.text:100A46C4 pop esi
.text:100A46C5 retn
.text:100A46C5 _invoke_watson endp
Now the invalid_parameter() function is rewritten in newer MSVCR*.DLL version, it shows the message
box, if you want to kill the process or call debugger. Of course, this is much better than silently return.
Perhaps, Microsoft forgot to fix MSVCRT.DLL since then.
But how it was working in the era of Windows XP? It wasn’t: MSVCRT.DLL from Windows XP doesn’t check
906
arguments against NULL. So under Windows XP my stricmp ("asd", NULL) code will crash, and this is
good.
My hypothesis: Microsoft upgraded MSVCR*.DLL files (including MSVCRT.DLL) for Windows 7 by adding
sanitizing checks everywhere. However, since MSVCRT.DLL wasn’t used much since MSVS .NET (year
2002), it wasn’t properly tested and the bug left here. But compilers like MinGW can still use this DLL.
What would I do without my reverse engineering skills?
The MSVCRT.DLL from Windows 8.1 has the same bug.
8.17 Other examples

An example about Z3 and manual decompilation was here. It is moved there: https://yurichev.com/
writings/SAT_SMT_by_example.pdf.
907
Chapter 9
Examples of reversing proprietary

file formats
9.1 Primitive XOR-encryption

9.1.1 Simplest ever XOR encryption
I once saw a software where all debugging messages has been encrypted using XOR by value of 3. In
other words, two lowest bits of all characters has been flipped.
“Hello, world” would become “Kfool/#tlqog”:
#!/usr/bin/python
msg="Hello, world!"
print "".join(map(lambda x: chr(ord(x)^3), msg))
This is quite interesting encryption (or rather obfuscation), because it has two important properties: 1)
single function for encryption/decryption, just apply it again; 2) resulting characters are also printable, so
the whole string can be used in source code without escaping characters.
The second property exploits the fact that all printable characters organized in rows: 0x2x-0x7x, and
when you flip two lowest bits, character moving 1 or 3 characters left or right, but never moved to another
(maybe non-printable) row:
…with a single exception of 0x7F character.

For example, let’s encrypt characters in A-Z range:
#!/usr/bin/python
msg="@ABCDEFGHIJKLMNO"
print "".join(map(lambda x: chr(ord(x)^3), msg))
Result:
908
CBA@GFEDKJIHONML
It’s like “@” and “C” characters has been swapped, and so are “B” and “a”.
Yet again, this is interesting example of exploiting XOR properties, rather than encryption: the very same
effect of preserving printableness can be achieved while flipping any of lowest 4 bits, in any combination.
909
9.1.2 Norton Guide: simplest possible 1-byte XOR encryption
Norton Guide was popular in the epoch of MS-DOS, it was a resident program that worked as a hypertext
reference manual.
Norton Guide’s databases are files with the extension .ng, the contents of which look encrypted:
Figure 9.2: Very typical look
Why did we think that it’s encrypted but not compressed?

We see that the 0x1A byte (looking like “→”) occurs often, it would not be possible in a compressed file.
We also see long parts that consist only of Latin letters, and they look like strings in an unknown language.
910
Since the 0x1A byte occurs so often, we can try to decrypt the file, assuming that it’s encrypted by the
simplest XOR-encryption.
If we apply XOR with the 0x1A constant to each byte in Hiew, we can see familiar English text strings:
Figure 9.3: Hiew XORing with 0x1A
XOR encryption with one single constant byte is the simplest possible encryption method, which is, nev-
ertheless, encountered sometimes.
Now we understand why the 0x1A byte is occurring so often: because there are so many zero bytes and
they were replaced by 0x1A in encrypted form.
But the constant might be different. In this case, we could try every constant in the 0..255 range and look
for something familiar in the decrypted file. 256 is not so much.
More about Norton Guide’s file format: http://go.yurichev.com/17317.
Entropy
A very important property of such primitive encryption systems is that the information entropy of the
encrypted/decrypted block is the same.
Here is my analysis in Wolfram Mathematica 10.
Listing 9.1: Wolfram Mathematica 10
911
In[1]:= input = BinaryReadList["X86.NG"];
In[2]:= Entropy[2, input] // N

Out[2]= 5.62724
In[3]:= decrypted = Map[BitXor[#, 16^^1A] &, input];
In[4]:= Export["X86_decrypted.NG", decrypted, "Binary"];
In[5]:= Entropy[2, decrypted] // N

Out[5]= 5.62724
In[6]:= Entropy[2, ExampleData[{"Text", "ShakespearesSonnets"}]] // N

Out[6]= 4.42366
What we do here is load the file, get its entropy, decrypt it, save it and get the entropy again (the same!).
Mathematica also offers some well-known English language texts for analysis.
So we also get the entropy of Shakespeare’s sonnets, and it is close to the entropy of the file we just
analyzed.
The file we analyzed consists of English language sentences, which are close to the language of Shake-
speare.
And the XOR-ed bitwise English language text has the same entropy.
However, this is not true when the file is XOR-ed with a pattern larger than one byte.
The file we analyzed can be downloaded here: http://beginners.re/examples/norton_guide/X86.NG.
One more word about base of entropy
Wolfram Mathematica calculates entropy with base of e (base of the natural logarithm), and the UNIX ent
utility1 uses base 2.
So we set base 2 explicitly in Entropy command, so Mathematica will give us the same results as the ent
utility.
912
9.1.3 Simplest possible 4-byte XOR encryption
If a longer pattern was used for XOR-encryption, for example a 4 byte pattern, it’s easy to spot as well.
For example, here is the beginning of the kernel32.dll file (32-bit version from Windows Server 2008):
Figure 9.4: Original file
913
Here it is “encrypted” with a 4-byte key:
Figure 9.5: “Encrypted” file
It’s very easy to spot the recurring 4 symbols.

Indeed, the header of a PE-file has a lot of long zero areas, which are the reason for the key to become
visible.
914
Here is the beginning of a PE-header in hexadecimal form:
Figure 9.6: PE-header
915
Here it is “encrypted”:
Figure 9.7: “Encrypted” PE-header
It’s easy to spot that the key is the following 4 bytes: 8C 61 D2 63.
With this information, it’s easy to decrypt the whole file.
So it is important to keep in mind these properties of PE-files: 1) PE-header has many zero-filled areas;
2) all PE-sections are padded with zeros at a page boundary (4096 bytes), so long zero areas are usually
present after each section.
Some other file formats may contain long zero areas.
It’s typical for files used by scientific and engineering software.
For those who want to inspect these files on their own, they are downloadable here: http://go.yurichev.
com/17352.
Exercise
916
9.1.4 Simple encryption using XOR mask
I’ve found an old interactive fiction game while diving deep into if-archive2 :
The New Castle v3.5 - Text/Adventure Game
in the style of the original Infocom (tm)
type games, Zork, Collosal Cave (Adventure),
etc. Can you solve the mystery of the
abandoned castle?
Shareware from Software Customization.
Software Customization [ASP] Version 3.5 Feb. 2000
It’s downloadable here: https://beginners.re/current-tree/ff/XOR/mask_1/files/newcastle.tgz.

There is a file inside (named castle.dbf ) which is clearly encrypted, but not by a real crypto algorithm,
nor it’s compressed, this is something rather simpler. I wouldn’t even measure entropy level ( 9.2 on
page 929) of the file, because I’m sure it’s low. Here is how it looks like in Midnight Commander:
Figure 9.8: Encrypted file in Midnight Commander
The encrypted file can be downloaded here: https://beginners.re/current-tree/ff/XOR/mask_1/

files/castle.dbf.bz2.
Will it be possible to decrypt it without accessing to the program, using just this file?
There is a clearly visible pattern of repeating string. If a simple encryption by XOR mask was applied, such
repeating strings is a prominent signature, because, probably, there were a long lacunas3 of zero bytes,
which, in turn, are present in many executable files as well as in binary data files.
Here I’ll dump the file’s beginning using xxd UNIX utility:
...
0000030: 09 61 0d 63 0f 77 14 69 75 62 67 76 01 7e 1d 61 .a.c.w.iubgv.~.a
0000040: 7a 11 0f 72 6e 03 05 7d 7d 63 7e 77 66 1e 7a 02 z..rn..}}c~wf.z.
0000050: 75 50 02 4a 31 71 31 33 5c 27 08 5c 51 74 3e 39 uP.J1q13\'.\Qt>9
0000060: 50 2e 28 72 24 4b 38 21 4c 09 37 38 3b 51 41 2d P.(r$K8!L.78;QA-
0000070: 1c 3c 37 5d 27 5a 1c 7c 6a 10 14 68 77 08 6d 1a .<7]'Z.|j..hw.m.
0000080: 6a 09 61 0d 63 0f 77 14 69 75 62 67 76 01 7e 1d j.a.c.w.iubgv.~.
0000090: 61 7a 11 0f 72 6e 03 05 7d 7d 63 7e 77 66 1e 7a az..rn..}}c~wf.z
2 http://www.ifarchive.org/
3 As in https://en.wikipedia.org/wiki/Lacuna_(manuscripts)
917
00000a0: 02 75 50 64 02 74 71 66 76 19 63 08 13 17 74 7d .uPd.tqfv.c...t}
00000b0: 6b 19 63 6d 72 66 0e 79 73 1f 09 75 71 6f 05 04 k.cmrf.ys..uqo..
00000c0: 7f 1c 7a 65 08 6e 0e 12 7c 6a 10 14 68 77 08 6d ..ze.n..|j..hw.m
00000d0: 1a 6a 09 61 0d 63 0f 77 14 69 75 62 67 76 01 7e .j.a.c.w.iubgv.~
00000e0: 1d 61 7a 11 0f 72 6e 03 05 7d 7d 63 7e 77 66 1e .az..rn..}}c~wf.
00000f0: 7a 02 75 50 01 4a 3b 71 2d 38 56 34 5b 13 40 3c z.uP.J;q-8V4[.@<
0000100: 3c 3f 19 26 3b 3b 2a 0e 35 26 4d 42 26 71 26 4b <?.&;;*.5&MB&q&K
0000110: 04 2b 54 3f 65 40 2b 4f 40 28 39 10 5b 2e 77 45 .+T?e@+O@(9.[.wE
0000120: 28 54 75 09 61 0d 63 0f 77 14 69 75 62 67 76 01 (Tu.a.c.w.iubgv.
0000130: 7e 1d 61 7a 11 0f 72 6e 03 05 7d 7d 63 7e 77 66 ~.az..rn..}}c~wf
0000140: 1e 7a 02 75 50 02 4a 31 71 15 3e 58 27 47 44 17 .z.uP.J1q.>X'GD.
0000150: 3f 33 24 4e 30 6c 72 66 0e 79 73 1f 09 75 71 6f ?3$N0lrf.ys..uqo
0000160: 05 04 7f 1c 7a 65 08 6e 0e 12 7c 6a 10 14 68 77 ....ze.n..|j..hw
...
Let’s stick at visible repeating iubgv string. By looking at this dump, we can clearly see that the period
of the string occurrence is 0x51 or 81. Probably, 81 is size of block? The size of the file is 1658961, and
it can be divided evenly by 81 (and there are 20481 blocks then).
Now I’ll use Mathematica to analyze, are there repeating 81-byte blocks in the file? I’ll split input file by
81-byte blocks and then I’ll use Tally[]4 function which just counts, how many times some item has been
occurred in the input list. Tally’s output is not sorted, so I also add Sort[] function to sort it by number of
occurrences in descending order.
input = BinaryReadList["/home/dennis/.../castle.dbf"];
blocks = Partition[input, 81];
stat = Sort[Tally[blocks], #1[[2]] > #2[[2]] &]
And here is output:

{{{80, 103, 2, 116, 113, 102, 118, 25, 99, 8, 19, 23, 116, 125, 107,
25, 99, 109, 114, 102, 14, 121, 115, 31, 9, 117, 113, 111, 5, 4,
127, 28, 122, 101, 8, 110, 14, 18, 124, 106, 16, 20, 104, 119, 8,
109, 26, 106, 9, 97, 13, 99, 15, 119, 20, 105, 117, 98, 103, 118,
1, 126, 29, 97, 122, 17, 15, 114, 110, 3, 5, 125, 125, 99, 126,
119, 102, 30, 122, 2, 117}, 1739},
{{80, 100, 2, 116, 113, 102, 118, 25, 99, 8, 19, 23, 116,
125, 107, 25, 99, 109, 114, 102, 14, 121, 115, 31, 9, 117, 113,
111, 5, 4, 127, 28, 122, 101, 8, 110, 14, 18, 124, 106, 16, 20,
104, 119, 8, 109, 26, 106, 9, 97, 13, 99, 15, 119, 20, 105, 117,
98, 103, 118, 1, 126, 29, 97, 122, 17, 15, 114, 110, 3, 5, 125,
125, 99, 126, 119, 102, 30, 122, 2, 117}, 1422},
{{80, 101, 2, 116, 113, 102, 118, 25, 99, 8, 19, 23, 116,
125, 107, 25, 99, 109, 114, 102, 14, 121, 115, 31, 9, 117, 113,
111, 5, 4, 127, 28, 122, 101, 8, 110, 14, 18, 124, 106, 16, 20,
104, 119, 8, 109, 26, 106, 9, 97, 13, 99, 15, 119, 20, 105, 117,
98, 103, 118, 1, 126, 29, 97, 122, 17, 15, 114, 110, 3, 5, 125,
125, 99, 126, 119, 102, 30, 122, 2, 117}, 1012},
{{80, 120, 2, 116, 113, 102, 118, 25, 99, 8, 19, 23, 116,
125, 107, 25, 99, 109, 114, 102, 14, 121, 115, 31, 9, 117, 113,
111, 5, 4, 127, 28, 122, 101, 8, 110, 14, 18, 124, 106, 16, 20,
104, 119, 8, 109, 26, 106, 9, 97, 13, 99, 15, 119, 20, 105, 117,
98, 103, 118, 1, 126, 29, 97, 122, 17, 15, 114, 110, 3, 5, 125,
125, 99, 126, 119, 102, 30, 122, 2, 117}, 377},
...
{{80, 2, 74, 49, 113, 21, 62, 88, 39, 71, 68, 23, 63, 51, 36, 78, 48,
108, 114, 102, 14, 121, 115, 31, 9, 117, 113, 111, 5, 4, 127, 28,
122, 101, 8, 110, 14, 18, 124, 106, 16, 20, 104, 119, 8, 109, 26,
106, 9, 97, 13, 99, 15, 119, 20, 105, 117, 98, 103, 118, 1, 126,
29, 97, 122, 17, 15, 114, 110, 3, 5, 125, 125, 99, 126, 119, 102,
30, 122, 2, 117}, 1},
4 https://reference.wolfram.com/language/ref/Tally.html
918
{{80, 1, 74, 59, 113, 45, 56, 86, 52, 91, 19, 64, 60, 60, 63,
25, 38, 59, 59, 42, 14, 53, 38, 77, 66, 38, 113, 38, 75, 4, 43, 84,
63, 101, 64, 43, 79, 64, 40, 57, 16, 91, 46, 119, 69, 40, 84, 117,
9, 97, 13, 99, 15, 119, 20, 105, 117, 98, 103, 118, 1, 126, 29,
97, 122, 17, 15, 114, 110, 3, 5, 125, 125, 99, 126, 119, 102, 30,
122, 2, 117}, 1},
{{80, 2, 74, 49, 113, 49, 51, 92, 39, 8, 92, 81, 116, 62, 57,
80, 46, 40, 114, 36, 75, 56, 33, 76, 9, 55, 56, 59, 81, 65, 45, 28,
60, 55, 93, 39, 90, 28, 124, 106, 16, 20, 104, 119, 8, 109, 26,
106, 9, 97, 13, 99, 15, 119, 20, 105, 117, 98, 103, 118, 1, 126,
29, 97, 122, 17, 15, 114, 110, 3, 5, 125, 125, 99, 126, 119, 102,
30, 122, 2, 117}, 1}}
Tally’s output is a list of pairs, each pair has 81-byte block and number of times it has been occurred in
the file. We see that the most frequent block is the first, it has been occurred 1739 times. The second
one has been occurred 1422 times. There are others: 1012 times, 377 times, etc. 81-byte blocks which
has been occurred just once are at the end of output.
Let’s try to compare these blocks. The first and the second. Is there a function in Mathematica which
compares lists/arrays? Certainly is, but for educational purposes, I’ll use XOR operation for comparison.
Indeed: if bytes in two input arrays are identical, XOR result is 0. If they are non-equal, result will be
non-zero.
Let’s compare first block (occurred 1739 times) and the second (occurred 1422 times):
In[]:= BitXor[stat[[1]][[1]], stat[[2]][[1]]]
Out[]= {0, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
They are differ only in the second byte.

Let’s compare the second block (occurred 1422 times) and the third (occurred 1012 times):
In[]:= BitXor[stat[[2]][[1]], stat[[3]][[1]]]
Out[]= {0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
They are also differ only in the second byte.

Anyway, let’s try to use the most occurred block as a XOR key and try to decrypt four first 81-byte blocks
in the file:
In[]:= key = stat[[1]][[1]]
Out[]= {80, 103, 2, 116, 113, 102, 118, 25, 99, 8, 19, 23, 116, \
125, 107, 25, 99, 109, 114, 102, 14, 121, 115, 31, 9, 117, 113, 111, \
5, 4, 127, 28, 122, 101, 8, 110, 14, 18, 124, 106, 16, 20, 104, 119, \
8, 109, 26, 106, 9, 97, 13, 99, 15, 119, 20, 105, 117, 98, 103, 118, \
1, 126, 29, 97, 122, 17, 15, 114, 110, 3, 5, 125, 125, 99, 126, 119, \
102, 30, 122, 2, 117}
In[]:= ToASCII[val_] := If[val == 0, " ", FromCharacterCode[val, "PrintableASCII"]]
In[]:= DecryptBlockASCII[blk_] := Map[ToASCII[#] &, BitXor[key, blk]]
In[]:= DecryptBlockASCII[blocks[[1]]]
Out[]= {" ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " \
", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " \
", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " \
", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " \
", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " \
", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " "}
Out[]= {" ", "e", "H", "E", " ", "W", "E", "E", "D", " ", "O", \
"F", " ", "C", "R", "I", "M", "E", " ", "B", "E", "A", "R", "S", " ", \
"B", "I", "T", "T", "E", "R", " ", "F", "R", "U", "I", "T", "?", \
" ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", \
919
" ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", \
" ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", \
" "}
Out[]= {" ", "?", " ", " ", " ", " ", " ", " ", " ", " ", " \
", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " \
", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " \
", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " \
", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " \
", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " \
"}
Out[]= {" ", "f", "H", "O", " ", "K", "N", "O", "W", "S", " ", \
"W", "H", "A", "T", " ", "E", "V", "I", "L", " ", "L", "U", "R", "K", \
"S", " ", "I", "N", " ", "T", "H", "E", " ", "H", "E", "A", "R", "T", \
"S", " ", "O", "F", " ", "M", "E", "N", "?", " ", " ", " ", " ", \
" ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", \
" ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", " ", \
" "}
(I’ve replaced unprintable characters by “?”.)

So we see that the first and the third blocks are empty (or almost empty), but the second and the fourth
has clearly visible English language words/phrases. It seems that our assumption about key is correct
(at least partially). This means that the most occurred 81-byte block in the file can be found at places of
lacunas of zero bytes or something like that.
Let’s try to decrypt the whole file:
DecryptBlock[blk_] := BitXor[key, blk]
decrypted = Map[DecryptBlock[#] &, blocks];
BinaryWrite["/home/dennis/.../tmp", Flatten[decrypted]]
Close["/home/dennis/.../tmp"]
920
Figure 9.9: Decrypted file in Midnight Commander, 1st attempt
Looks like some kind of English phrases from some game, but something wrong. First of all, cases are
inverted: phrases and some words are started with lowercase characters, while other characters are in
upper case. Also, some phrases started with wrong letters. Take a look at the very first phrase: “eHE
WEED OF CRIME BEARS BITTER FRUIT”. What is “eHE”? Isn’t “tHE” have to be here? Is it possible that our
decryption key has wrong byte at this place?
Let’s look again at the second block in the file, at key and at decryption result:
In[]:= blocks[[2]]
Out[]= {80, 2, 74, 49, 113, 49, 51, 92, 39, 8, 92, 81, 116, 62, \
57, 80, 46, 40, 114, 36, 75, 56, 33, 76, 9, 55, 56, 59, 81, 65, 45, \
28, 60, 55, 93, 39, 90, 28, 124, 106, 16, 20, 104, 119, 8, 109, 26, \
106, 9, 97, 13, 99, 15, 119, 20, 105, 117, 98, 103, 118, 1, 126, 29, \
97, 122, 17, 15, 114, 110, 3, 5, 125, 125, 99, 126, 119, 102, 30, \
122, 2, 117}
In[]:= key
Out[]= {80, 103, 2, 116, 113, 102, 118, 25, 99, 8, 19, 23, 116, \
125, 107, 25, 99, 109, 114, 102, 14, 121, 115, 31, 9, 117, 113, 111, \
5, 4, 127, 28, 122, 101, 8, 110, 14, 18, 124, 106, 16, 20, 104, 119, \
8, 109, 26, 106, 9, 97, 13, 99, 15, 119, 20, 105, 117, 98, 103, 118, \
1, 126, 29, 97, 122, 17, 15, 114, 110, 3, 5, 125, 125, 99, 126, 119, \
102, 30, 122, 2, 117}
In[]:= BitXor[key, blocks[[2]]]

Out[]= {0, 101, 72, 69, 0, 87, 69, 69, 68, 0, 79, 70, 0, 67, 82, \
73, 77, 69, 0, 66, 69, 65, 82, 83, 0, 66, 73, 84, 84, 69, 82, 0, 70, \
82, 85, 73, 84, 14, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \
0, 0, 0, 0}
921
Encrypted byte is 2, the byte from the key is 103, 2⊕103 = 101 and 101 is ASCII code for “e” character. What
byte of a key must be equal to, so the resulting ASCII code will be 116 (for “t” character)? 2 ⊕ 116 = 118,
let’s put 118 in key at the second byte …
key = {80, 118, 2, 116, 113, 102, 118, 25, 99, 8, 19, 23, 116, 125,
107, 25, 99, 109, 114, 102, 14, 121, 115, 31, 9, 117, 113, 111, 5,
4, 127, 28, 122, 101, 8, 110, 14, 18, 124, 106, 16, 20, 104, 119, 8,
109, 26, 106, 9, 97, 13, 99, 15, 119, 20, 105, 117, 98, 103, 118,
1, 126, 29, 97, 122, 17, 15, 114, 110, 3, 5, 125, 125, 99, 126, 119,
102, 30, 122, 2, 117}
…and decrypt the whole file again.
Figure 9.10: Decrypted file in Midnight Commander, 2nd attempt
Wow, now the grammar is correct, all phrases started with correct letters. But still, case inversion is
suspicious. Why would game’s developer write them in such a manner? Maybe our key is still incorrect?
While observing ASCII table we can notice that uppercase and lowercase letter’s ASCII codes are differ in
just one bit (6th bit starting at 1st, 0b100000):
6th bit set in a zero byte has decimal form of 32. But 32 is ASCII code for space!
922
Indeed, one can switch case just by XOR-ing ASCII character code with 32 (more about it: 3.19.3 on
page 536).
It is possible that the empty lacunas in the file are not zero bytes, but rather spaces? Let’s modify XOR
key one more time (I’ll XOR each byte of key by 32):
(* "32" is scalar and "key" is vector, but that's OK *)
In[]:= key3 = BitXor[32, key]

Out[]= {112, 86, 34, 84, 81, 70, 86, 57, 67, 40, 51, 55, 84, 93, 75, \
57, 67, 77, 82, 70, 46, 89, 83, 63, 41, 85, 81, 79, 37, 36, 95, 60, \
90, 69, 40, 78, 46, 50, 92, 74, 48, 52, 72, 87, 40, 77, 58, 74, 41, \
65, 45, 67, 47, 87, 52, 73, 85, 66, 71, 86, 33, 94, 61, 65, 90, 49, \
47, 82, 78, 35, 37, 93, 93, 67, 94, 87, 70, 62, 90, 34, 85}
In[]:= DecryptBlock[blk_] := BitXor[key3, blk]
Let’s decrypt the input file again:
Figure 9.12: Decrypted file in Midnight Commander, final attempt
(Decrypted file is available here: https://beginners.re/current-tree/ff/XOR/mask_1/files/decrypted.

dat.bz2.)
This is undoubtedly a correct source file. Oh, and we see numbers at the start of each block. It has to be a
source of our erroneous XOR key. As it seems, the most occurred 81-byte block in the file is a block filled
with spaces and containing “1” character at the place of second byte. Indeed, somehow, many blocks
here are interleaved with this one. Maybe it’s some kind of padding for short phrases/messages? Other
frequently occurred 81-byte blocks are also space-filled blocks, but with different digit, hence, they are
differ only at the second byte.
That’s all! Now we can write an utility to encrypt the file back, and maybe modify it before.
923
Mathematica notebook file is downloadable here:
https://beginners.re/current-tree/ff/XOR/mask_1/files/XOR_mask_1.nb.
Summary: XOR encryption like that is not robust at all. It has been intended by game’s developer(s), prob-
ably, just to prevent gamer(s) to peek into internals of game, nothing else more serious. Still, encryption
like that is extremely popular due to its simplicity and many reverse engineers are usually familiar with it.
9.1.5 Simple encryption using XOR mask, case II

I’ve got another encrypted file, which is clearly encrypted by something simple, like XOR-ing:
Figure 9.13: Encrypted file in Midnight Commander
The encrypted file can be downloaded here.

ent Linux utility reports about ~7.5 bits per byte, and this is high level of entropy ( 9.2 on page 929), close
to compressed or properly encrypted file. But still, we clearly see some pattern, there are some blocks
with size of 17 bytes, and you can see some kind of ladder, shifting by 1 byte at each 16-byte line.
It’s also known that the plain text is just English language text.
Now let’s assume that this piece of text is encrypted by simple XOR-ing with 17-byte key.
I tried to find some repeating 17-byte blocks using Mathematica, like I did before in my previous example
( 9.1.4 on page 917):
Listing 9.2: Mathematica

In[]:=input = BinaryReadList["/home/dennis/tmp/cipher.txt"];
924
In[]:=blocks = Partition[input, 17];
In[]:=Sort[Tally[blocks], #1[[2]] > #2[[2]] &]
Out[]:={{{248,128,88,63,58,175,159,154,232,226,161,50,97,127,3,217,80},1},
{{226,207,67,60,42,226,219,150,246,163,166,56,97,101,18,144,82},1},
{{228,128,79,49,59,250,137,154,165,236,169,118,53,122,31,217,65},1},
{{252,217,1,39,39,238,143,223,241,235,170,91,75,119,2,152,82},1},
{{244,204,88,112,59,234,151,147,165,238,170,118,49,126,27,144,95},1},
{{241,196,78,112,54,224,142,223,242,236,186,58,37,50,17,144,95},1},
{{176,201,71,112,56,230,143,151,234,246,187,118,44,125,8,156,17},1},
...
{{255,206,82,112,56,231,158,145,165,235,170,118,54,115,9,217,68},1},
{{249,206,71,34,42,254,142,154,235,247,239,57,34,113,27,138,88},1},
{{157,170,84,32,32,225,219,139,237,236,188,51,97,124,21,141,17},1},
{{248,197,1,61,32,253,149,150,235,228,188,122,97,97,27,143,84},1},
{{252,217,1,38,42,253,130,223,233,226,187,51,97,123,20,217,69},1},
{{245,211,13,112,56,231,148,223,242,226,188,118,52,97,15,152,93},1},
{{221,210,15,112,28,231,158,141,233,236,172,61,97,90,21,149,92},1}}
No luck, each 17-byte block is unique within the file and occurred only once. Perhaps, there are no 17-byte
zero lacunas, or lacunas containing only spaces. It is possible indeed: such long space indentation and
padding may be absent in tightly typeset text.
The first idea is to try all possible 17-byte keys and find those, which will result in readable text after
decryption. Bruteforce is not an option, because there are 25617 possible keys (~1040 ), that’s too much.
But there are good news: who said we have to test 17-byte key as a whole, why can’t we test each byte
of key separately? It is possible indeed.
Now the algorithm is:
• try all 256 bytes for 1st byte of key;
• decrypt 1st byte of each 17-byte blocks in the file;
• are all decrypted bytes we got are printable? keep tabs on it;
• do so for all 17 bytes of key.
I’ve written the following Python script to check this idea:


print "N=", N
possible_keys=[]
tmp_key=chr(i)*len(each_Nth_byte[N])
tmp=xor_strings(tmp_key,each_Nth_byte[N])
# are all characters in tmp[] are printable?
if is_string_printable(tmp)==False:
continue
possible_keys.append(i)
print possible_keys, "len=", len(possible_keys)
(Full version of the source code is here.)

Here is its output:
N= 0
[144, 145, 151] len= 3
N= 1
925
[160, 161] len= 2
N= 2
[32, 33, 38] len= 3
N= 3
[80, 81, 87] len= 3
N= 4
[78, 79] len= 2
N= 5
[142, 143] len= 2
N= 6
[250, 251] len= 2
N= 7
[254, 255] len= 2
N= 8
[130, 132, 133] len= 3
N= 9
[130, 131] len= 2
N= 10
[206, 207] len= 2
N= 11
[81, 86, 87] len= 3
N= 12
[64, 65] len= 2
N= 13
[18, 19] len= 2
N= 14
[122, 123] len= 2
N= 15
[248, 249] len= 2
N= 16
[48, 49] len= 2
So there are 2 or 3 possible bytes for each byte of 17-byte key. This is much better than 256 possible
bytes for each byte, but still too much. There are up to 1 million of possible keys:

In[]:= 3*2*3*3*2*2*2*2*3*2*2*3*2*2*2*2*2
Out[]= 995328
It’s possible to check all of them, but then we must check visually, if the decrypted text is looks like English
language text.
Let’s also take into consideration the fact that we deal with 1) natural language; 2) English language.
Natural languages has some prominent statistical features. First of all, punctuation and word lengths.
What is average word length in English language? Let’s just count spaces in some well-known English
language texts using Mathematica.
Here is “The Complete Works of William Shakespeare” text file from Gutenberg Library:

In[]:= input = BinaryReadList["/home/dennis/tmp/pg100.txt"];
In[]:= Tally[input]
Out[]= {{239, 1}, {187, 1}, {191, 1}, {84, 39878}, {104,
218875}, {101, 406157}, {32, 1285884}, {80, 12038}, {114,
209907}, {111, 282560}, {106, 2788}, {99, 67194}, {116,
291243}, {71, 11261}, {117, 115225}, {110, 216805}, {98,
46768}, {103, 57328}, {69, 42703}, {66, 15450}, {107, 29345}, {102,
69103}, {67, 21526}, {109, 95890}, {112, 46849}, {108, 146532}, {87,
16508}, {115, 215605}, {105, 199130}, {97, 245509}, {83,
34082}, {44, 83315}, {121, 85549}, {13, 124787}, {10, 124787}, {119,
73155}, {100, 134216}, {118, 34077}, {46, 78216}, {89, 9128}, {45,
8150}, {76, 23919}, {42, 73}, {79, 33268}, {82, 29040}, {73,
55893}, {72, 18486}, {68, 15726}, {58, 1843}, {65, 44560}, {49,
982}, {50, 373}, {48, 325}, {91, 2076}, {35, 3}, {93, 2068}, {74,
2071}, {57, 966}, {52, 107}, {70, 11770}, {85, 14169}, {78,
27393}, {75, 6206}, {77, 15887}, {120, 4681}, {33, 8840}, {60,
468}, {86, 3587}, {51, 343}, {88, 608}, {40, 643}, {41, 644}, {62,
440}, {39, 31077}, {34, 488}, {59, 17199}, {126, 1}, {95, 71}, {113,
926
2414}, {81, 1179}, {63, 10476}, {47, 48}, {55, 45}, {54, 73}, {64,
3}, {53, 94}, {56, 47}, {122, 1098}, {90, 532}, {124, 33}, {38,
21}, {96, 1}, {125, 2}, {37, 1}, {36, 2}}
In[]:= Length[input]/1285884 // N
Out[]= 4.34712
There are 1285884 spaces in the whole file, and the frequency of space occurrence is 1 space per ~4.3
characters.
Now here is Alice’s Adventures in Wonderland, by Lewis Carroll from the same library:

In[]:= input = BinaryReadList["/home/dennis/tmp/pg11.txt"];
In[]:= Tally[input]
Out[]= {{239, 1}, {187, 1}, {191, 1}, {80, 172}, {114, 6398}, {111,
9243}, {106, 222}, {101, 15082}, {99, 2815}, {116, 11629}, {32,
27964}, {71, 193}, {117, 3867}, {110, 7869}, {98, 1621}, {103,
2750}, {39, 2885}, {115, 6980}, {65, 721}, {108, 5053}, {105,
7802}, {100, 5227}, {118, 911}, {87, 256}, {97, 9081}, {44,
2566}, {121, 2442}, {76, 158}, {119, 2696}, {67, 185}, {13,
3735}, {10, 3735}, {84, 571}, {104, 7580}, {66, 125}, {107,
1202}, {102, 2248}, {109, 2245}, {46, 1206}, {89, 142}, {112,
1796}, {45, 744}, {58, 255}, {68, 242}, {74, 13}, {50, 12}, {53,
13}, {48, 22}, {56, 10}, {91, 4}, {69, 313}, {35, 1}, {49, 68}, {93,
4}, {82, 212}, {77, 222}, {57, 11}, {52, 10}, {42, 88}, {83,
288}, {79, 234}, {70, 134}, {72, 309}, {73, 831}, {85, 111}, {78,
182}, {75, 88}, {86, 52}, {51, 13}, {63, 202}, {40, 76}, {41,
76}, {59, 194}, {33, 451}, {113, 135}, {120, 170}, {90, 1}, {122,
79}, {34, 135}, {95, 4}, {81, 85}, {88, 6}, {47, 24}, {55, 6}, {54,
7}, {37, 1}, {64, 2}, {36, 2}}
In[]:= Length[input]/27964 // N
Out[]= 5.99049
The result is different probably because of different formatting of these texts (maybe indentation and/or
padding).
OK, so let’s assume the average frequency of space in English language is 1 space per 4..7 characters.
Now the good news again: we can measure frequency of spaces while decrypting our file gradually. Now
I count spaces in each slice and throw away 1-byte keys which produce results with too small number of
spaces (or too large, but this is almost impossible given so short key):


print "N=", N
possible_keys=[]
tmp_key=chr(i)*len(each_Nth_byte[N])
tmp=xor_strings(tmp_key,each_Nth_byte[N])
# are all characters in tmp[] are printable?
if is_string_printable(tmp)==False:
continue
# count spaces in decrypted buffer:
spaces=tmp.count(' ')
if spaces==0:
continue
927
spaces_ratio=len(tmp)/spaces
if spaces_ratio<4:
continue
if spaces_ratio>7:
continue
possible_keys.append(i)
print possible_keys, "len=", len(possible_keys)
(Full version of the source code is here.)

This reports just one single possible byte for each byte of key:
N= 0
[144] len= 1
N= 1
[160] len= 1
N= 2
[33] len= 1
N= 3
[80] len= 1
N= 4
[79] len= 1
N= 5
[143] len= 1
N= 6
[251] len= 1
N= 7
[255] len= 1
N= 8
[133] len= 1
N= 9
[131] len= 1
N= 10
[207] len= 1
N= 11
[86] len= 1
N= 12
[65] len= 1
N= 13
[18] len= 1
N= 14
[122] len= 1
N= 15
[249] len= 1
N= 16
[49] len= 1
Let’s check this key in Mathematica:

In[]:= input = BinaryReadList["/home/dennis/tmp/cipher.txt"];
In[]:= blocks = Partition[input, 17];
In[]:= key = {144, 160, 33, 80, 79, 143, 251, 255, 133, 131, 207, 86, 65, 18, 122, 249, 49};
In[]:= EncryptBlock[blk_] := BitXor[key, blk]
In[]:= encrypted = Map[EncryptBlock[#] &, blocks];
In[]:= BinaryWrite["/home/dennis/tmp/plain2.txt", Flatten[encrypted]]
In[]:= Close["/home/dennis/tmp/plain2.txt"]
And the plain text is:

Mr. Sherlock Holmes, who was usually very late in the mornings, save
upon those not infrequent occasions when he was up all night, was seated
at the breakfast table. I stood upon the hearth-rug and picked up the
stick which our visitor had left behind him the night before. It was a
928
fine, thick piece of wood, bulbous-headed, of the sort which is known as
a "Penang lawyer." Just under the head was a broad silver band nearly
an inch across. "To James Mortimer, M.R.C.S., from his friends of the
C.C.H.," was engraved upon it, with the date "1884." It was just such a
stick as the old-fashioned family practitioner used to carry--dignified,
solid, and reassuring.
"Well, Watson, what do you make of it?"
Holmes was sitting with his back to me, and I had given him no sign of
my occupation.
...
(Full version of the text is here.)

The text looks correct. Yes, I made up this example and choose well-known text of Conan Doyle, but it’s
very close to what I had in my practice some time ago.
Other ideas to consider
If we would fail with space counting, there are other ideas to try:
• Take into consideration the fact that lowercase letters are much more frequent than uppercase ones.
• Frequency analysis.
• There is also a good technique to detect language of a text: trigrams. Each language has some
very frequent letter triplets, these may be “the” and “tha” for English. Read more about it: N-
Gram-Based Text Categorization, http://code.activestate.com/recipes/326576/. Interestingly
enough, trigrams detection can be used when you decrypt a ciphertext gradually, like in this example
(you just have to test 3 adjacent decrypted characters).
For non-Latin writing systems encoded in UTF-8, things may be easier. For example, Russian text
encoded in UTF-8 has each byte interleaved with 0xD0/0xD1 byte. It is because Cyrillic characters
are placed in 4th block of Unicode table. Other writing systems has their own blocks.
9.1.6 Homework
An ancient text adventure for MS-DOS, developed in the end of 1980’s. To conceal game information from
player, data files, most likely, XOR-ed with something: https://beginners.re/homework/XOR_crypto_
1/destiny.zip. Try to get into...
9.2 Information entropy
Entropy: The quantitative measure of

disorder, which in turn relates to the
thermodynamic functions, temperature, and
heat.
Dictionary of Applied Math for Engineers and

Scientists
For the sake of simplification, I would say, information entropy is a measure, how tightly some piece of
data can be compressed. For example, it is usually not possible to compress already compressed archive
file, so it has high entropy. On the other hand, 1MiB of zero bytes can be compressed to a tiny output file.
Indeed, in plain English language, one million of zeros can be described just as “resulting file is one million
zero bytes”. Compressed files are usually a list of instructions to decompressor, like this: “put 1000 zeros,
then 0x23 byte, then 0x45 byte, then put a block of size 10 bytes which we’ve seen 500 bytes back, etc.”
Texts written in natural languages are also can be compressed tightly, because natural languages has
a lot of redundancy (otherwise, a tiny typo will always lead to misunderstanding, like any toggled bit in
compressed archive make decompression nearly impossible), some words are used very often, etc. In
everyday speech, it’s possible to drop up to half of words and it still be recognizable.
Code for CPUs is also can be compressed, because some ISA instructions are used much more often than
others. In x86, most used instructions are MOV/PUSH/CALL ( 5.11.2 on page 724).
929
Data compressors and ciphers tend to produce very high entropy results. Good PRNG also produce data
which cannot be compressed (it is possible to measure their quality by this sign).
So, in other words, entropy is a measure which can help to probe contents of unknown data block.
9.2.1 Analyzing entropy in Mathematica

(This part has been first appeared in my blog at 13-May-2015. Some discussion: https://news.ycombinator.
com/item?id=9545276.)
It is possible to slice a file by blocks, calculate entropy of each and draw a graph. I did this in Wolfram
Mathematica for demonstration and here is a source code (Mathematica 10):
(* loading the file *)
input=BinaryReadList["file.bin"];
(* setting block sizes *)

BlockSize=4096;BlockSizeToShow=256;
(* slice blocks by 4k *)
blocks=Partition[input,BlockSize];
(* how many blocks we've got? *)

Length[blocks]
(* calculate entropy for each block. 2 in Entropy[] (base) is set with the intention so Entropy⤦
Ç []
function will produce the same results as Linux ent utility does *)
entropies=Map[N[Entropy[2,#]]&,blocks];
(* helper functions *)
fBlockToShow[input_,offset_]:=Take[input,{1+offset,1+offset+BlockSizeToShow}]
fToASCII[val_]:=FromCharacterCode[val,"PrintableASCII"]
fToHex[val_]:=IntegerString[val,16]
fPutASCIIWindow[data_]:=Framed[Grid[Partition[Map[fToASCII,data],16]]]
fPutHexWindow[data_]:=Framed[Grid[Partition[Map[fToHex,data],16],Alignment->Right]]
(* that will be the main knob here *)

{Slider[Dynamic[offset],{0,Length[input]-BlockSize,BlockSize}],Dynamic[BaseForm[offset,16]]}
(* main UI part *)
Dynamic[{ListLinePlot[entropies,GridLines->{{-1,offset/BlockSize,1}},Filling->Axis,AxesLabel->{⤦
Ç "offset","entropy"}],
CurrentBlock=fBlockToShow[input,offset];
fPutHexWindow[CurrentBlock],
fPutASCIIWindow[CurrentBlock]}]
GeoIP ISP database
Let’s start with the GeoIP file (which assigns ISP to the block of IP addresses). This binary file GeoIPISP.dat
has some tables (which are IP address ranges perhaps) plus some text blob at the end of the file (containing
ISP names).
When I load it to Mathematica, I see this:
930
There are two parts in graph: first is somewhat chaotic, second is more steady.
0 in horizontal axis in graph means lowest entropy (the data which can be compressed very tightly, ordered
in other words) and 8 is highest (cannot be compressed at all, chaotic or random in other words). Why 0
and 8? 0 means 0 bits per byte (byte as a container is not filled at all) and 8 means 8 bits per byte, i.e.,
the whole byte container is filled with the information tightly.
So I put slider to point in the middle of the first block, and I clearly see some array of 32-bit integers. Now
I put slider in the middle of the second block and I see English text:
931
Indeed, this are names of ISPs. So, entropy of English text is 4.5-5.5 bits per byte? Yes, something like
this. Wolfram Mathematica has some well-known English literature corpus embedded, and we can see
entropy of Shakespeare’s sonnets:
In[]:= Entropy[2,ExampleData[{"Text","ShakespearesSonnets"}]]//N
Out[]= 4.42366
4.4 is close to what we’ve got (4.7-5.3). Of course, classic English literature texts are somewhat different
from ISP names and other English texts we can find in binary files (debugging/logging/error messages),
but this value is close.
TP-Link WR941 firmware
Next example. I’ve got firmware for TP-Link WR941 router:
932
We see here 3 blocks with empty lacunas. Then the first block with high entropy (started at address 0) is
small, second (address somewhere at 0x22000) is bigger and third (address 0x123000) is biggest. I can’t
be sure about exact entropy of the first block, but 2nd and 3rd has very high entropy, meaning that these
blocks are either compressed and/or encrypted.
I tried binwalk for this firmware file:
DECIMAL HEXADECIMAL DESCRIPTION
--------------------------------------------------------------------------------
0 0x0 TP-Link firmware header, firmware version: 0.-15221.3, image ⤦
Ç version: "", product ID: 0x0, product version: 155254789, kernel load address: 0x0, ⤦
Ç kernel entry point: 0x-7FFFE000, kernel offset: 4063744, kernel length: 512, rootfs ⤦
Ç offset: 837431, rootfs length: 1048576, bootloader offset: 2883584, bootloader length: 0
14832 0x39F0 U-Boot version string, "U-Boot 1.1.4 (Jun 27 2014 - 14:56:49)"
14880 0x3A20 CRC32 polynomial table, big endian
16176 0x3F30 uImage header, header size: 64 bytes, header CRC: 0x3AC66E95, ⤦
Ç created: 2014-06-27 06:56:50, image size: 34587 bytes, Data Address: 0x80010000, Entry ⤦
Ç Point: 0x80010000, data CRC: 0xDF2DBA0B, OS: Linux, CPU: MIPS, image type: Firmware Image⤦
Ç , compression type: lzma, image name: "u-boot image"
16240 0x3F70 LZMA compressed data, properties: 0x5D, dictionary size: 33554432⤦
Ç bytes, uncompressed size: 90000 bytes
131584 0x20200 TP-Link firmware header, firmware version: 0.0.3, image version: ⤦
Ç "", product ID: 0x0, product version: 155254789, kernel load address: 0x0, kernel entry ⤦
Ç point: 0x-7FFFE000, kernel offset: 3932160, kernel length: 512, rootfs offset: 837431, ⤦
Ç rootfs length: 1048576, bootloader offset: 2883584, bootloader length: 0
132096 0x20400 LZMA compressed data, properties: 0x5D, dictionary size: 33554432⤦
Ç bytes, uncompressed size: 2388212 bytes
1180160 0x120200 Squashfs filesystem, little endian, version 4.0, compression:lzma⤦
Ç , size: 2548511 bytes, 536 inodes, blocksize: 131072 bytes, created: 2014-06-27 07:06:52
Indeed: there are some stuff at the beginning, but two large LZMA compressed blocks are started at
0x20400 and 0x120200. These are roughly addresses we have seen in Mathematica. Oh, and by the way,
binwalk can show entropy information as well (-E option):
DECIMAL HEXADECIMAL ENTROPY
--------------------------------------------------------------------------------
0 0x0 Falling entropy edge (0.419187)
16384 0x4000 Rising entropy edge (0.988639)
51200 0xC800 Falling entropy edge (0.000000)
968704 0xEC800 Falling entropy edge (0.508720)
3727360 0x38E000 Falling entropy edge (0.732390)
933
Rising edges are corresponding to rising edges of block on our graph. Falling edges are the points where
empty lacunas are started.
Binwalk can also generate PNG graphs (-E -J):
What can we say about lacunas? By looking in hex editor, we see that these are just filled with 0xFF bytes.
Why developers put them? Perhaps, because they weren’t able to calculate precise compressed blocks
sizes, so they allocated space for them with some reserve.
Notepad
Another example is notepad.exe I’ve picked in Windows 8.1:
934
There is cavity at ≈ 0x19000 (absolute file offset). I’ve opened the executable file in hex editor and found
imports table there (which has lower entropy than x86-64 code in the first half of graph).
There are also high entropy block started at ≈ 0x20000:
935
In hex editor I can see PNG file here, embedded in the PE file resource section (it is a large image of
notepad icon). PNG files are compressed, indeed.
Unnamed dashcam
Now the most advanced example in this part is the firmware of some unnamed dashcam I’ve received
from a friend:
936
The cavity at the very beginning is an English text: debugging messages. I checked various ISAs and I
found that the first third of the whole file (with the text segment inside) is in fact MIPS (little-endian) code.
For instance, this is very distinctive MIPS function epilogue:
ROM:000013B0 move $sp, $fp
ROM:000013B4 lw $ra, 0x1C($sp)
ROM:000013B8 lw $fp, 0x18($sp)
ROM:000013BC lw $s1, 0x14($sp)
ROM:000013C0 lw $s0, 0x10($sp)
ROM:000013C4 jr $ra
ROM:000013C8 addiu $sp, 0x20
From our graph we can see that MIPS code has entropy of 5-6 bits per byte. Indeed, I once measured
various ISAs entropy and I’ve got these values:
• x86: .text section of ntoskrnl.exe file from Windows 2003: 6.6
937
• x64: .text section of ntoskrnl.exe file from Windows 7 x64: 6.5
• ARM (thumb mode), Angry Birds Classic: 7.05
• ARM (ARM mode) Linux Kernel 3.8.0: 6.03
• MIPS (little endian), .text section of user32.dll from Windows NT 4: 6.09
So the entropy of executable code is higher than of English text, but still can be compressed.
Now the second third is started at 0xF5000. I don’t know what this is. I tried different ISAs but without
success. The entropy of the block is looks even steadier than for executable one. Maybe some kind of
data?
There is also a spike at ≈ 0x213000. I checked it in hex editor and I found JPEG file there (which, of course,
compressed)! I also don’t know what is at the end. Let’s try Binwalk for this file:
% binwalk FW96650A.bin
DECIMAL HEXADECIMAL DESCRIPTION

--------------------------------------------------------------------------------
167698 0x28F12 Unix path: /15/20/24/25/30/60/120/240fps can be served..
280286 0x446DE Copyright string: "Copyright (c) 2012 Novatek Microelectronic ⤦
Ç Corp."
2169199 0x21196F JPEG image data, JFIF standard 1.01
2300847 0x231BAF MySQL MISAM compressed data file Version 3
% binwalk -E FW96650A.bin
DECIMAL HEXADECIMAL ENTROPY

--------------------------------------------------------------------------------
2596864 0x27A000 Falling entropy edge (0.712814)
2600960 0x27B000 Rising entropy edge (0.968167)
2607104 0x27C800 Rising entropy edge (0.958582)
2609152 0x27D000 Falling entropy edge (0.760989)
2670592 0x28C000 Rising entropy edge (0.967883)
2676736 0x28D800 Rising entropy edge (0.975779)
2684928 0x28F800 Falling entropy edge (0.744369)
Yes, it found JPEG file and even MySQL data! But I’m not sure if it’s true—I didn’t check it yet.
It’s also interesting to try clusterization in Mathematica:
938
Here is an example of how Mathematica grouped various entropy values into distinctive groups. Indeed,
there is something credible. Blue dots in range of 5.0-5.5 are supposedly related to English text. Yellow
dots in 5.5-6 are MIPS code. A lot of green dots in 6.0-6.5 is the unknown second third. Orange dots close
to 8.0 are related to compressed JPEG file. Other orange dots are supposedly related to the end of the
firmware (unknown to us data).
Links
Binary files used in this part:

https://beginners.re/current-tree/ff/entropy/files/.
Wolfram Mathematica notebook file:
https://beginners.re/current-tree/ff/entropy/files/binary_file_entropy.nb
(all cells must be evaluated to start things working).
9.2.2 Conclusion
Information entropy can be used as a quick-n-dirty method for inspecting unknown binary files. In partic-
ular, it is a very quick way to find compressed/encrypted pieces of data. Someone say it’s possible to find
RSA5 (and other asymmetric cryptographic algorithms) public/private keys in executable code (keys has
high entropy as well), but I didn’t try this myself.
9.2.3 Tools
Handy Linux ent utility to measure entropy of a file6 .
There is a great online entropy visualizer made by Aldo Cortesi, which I tried to mimic using Mathematica:
http://binvis.io. His articles about entropy visualization are worth reading: http://corte.si/posts/
visualisation/entropy/index.html, http://corte.si/posts/visualisation/malware/index.html,
http://corte.si/posts/visualisation/binvis/index.html.
radare2 framework has #entropy command for this.
A tool for IDA: IDAtropy7 .
5 Rivest Shamir Adleman
7 https://github.com/danigargu/IDAtropy
939
9.2.4 A word about primitive encryption like XORing
It’s interesting that simple XOR encryption doesn’t affect entropy of data. I’ve shown this in Norton Guide
example in the book ( 9.1.2 on page 910).
Generalizing: encryption by substitution cipher also doesn’t affect entropy of data (and XOR can be viewed
as substitution cipher). The reason of that is because entropy calculation algorithm view data on byte-
level. On the other hand, the data encrypted by 2 or 4-byte XOR pattern will result in another level of
entropy.
Nevertheless, low entropy is usually a good sign of weak amateur cryptography (which is also used in
license keys/files, etc.).
9.2.5 More about entropy of executable code

It is quickly noticeable that probably a biggest source of high-entropy in executable code are relative
offsets encoded in opcodes. For example, these two consequent instructions will have different relative
offsets in their opcodes, while they are in fact pointing to the same function:
function proc
...
function endp
...
CALL function
...
CALL function
Ideal executable code compressor would encode information like this: there is a CALL to a “function” at
address X and the same CALL at address Y without necessity to encode address of the function twice.
To deal with this, executable compressors are sometimes able to reduce entropy here. One example is
UPX: http://sourceforge.net/p/upx/code/ci/default/tree/doc/filter.txt.
9.2.6 PRNG
When I run GnuPG to generate new private (secret) key, it asking for some entropy …
We need to generate a lot of random bytes. It is a good idea to perform
some other action (type on the keyboard, move the mouse, utilize the
disks) during the prime generation; this gives the random number
generator a better chance to gain enough entropy.
Not enough random bytes available. Please do some other work to give
the OS a chance to collect more entropy! (Need 169 more bytes)
This means that good a PRNG produces long high-entropy results, and this is what the secret asymmetrical
cryptographical key needs. But CPRNG8 is tricky (because computer is highly deterministic device itself),
so the GnuPG asking for some additional randomness from the user.
9.2.7 More examples

Here is a case where I try to calculate entropy of some blocks with unknown contents: 8.9 on page 840.
9.2.8 Entropy of various files

Entropy of random data is close to 8:
% dd bs=1M count=1 if=/dev/urandom | ent
Entropy = 7.999803 bits per byte.
This means, almost all available space inside of byte is filled with information.
256 bytes in range of 0..255 gives exact value of 8:
8 Cryptographically secure PseudoRandom Number Generator
940
import sys
sys.stdout.write(chr(i))
% python 1.py | ent

Order of bytes doesn’t matter. This means, all available space inside of byte is filled.
Entropy of any block filled with zero bytes is 0:
% dd bs=1M count=1 if=/dev/zero | ent
Entropy of a string constisting of a single (any) byte is 0:

% echo -n "aaaaaaaaaaaaaaaaaaa" | ent
3
Entropy of base64 string is the same as entropy of source data, but multiplied by 4
. This is because
base64 encoding uses 64 symbols instead of 256.
% dd bs=1M count=1 if=/dev/urandom | base64 | ent
Perhaps, 6.02 is slightly bigger than 6 because padding symbols (=) spoils our statistics for a little.
Uuencode also uses 64 symbols:
% dd bs=1M count=1 if=/dev/urandom | uuencode - | ent
This means, any base64 and Uuencode strings can be transmitted using 6-bit bytes or characters.
Any random information in hexadecimal form has entropy of 4 bits per byte:
% openssl rand -hex $\$$(( 2**16 )) | ent
Entropy of randomly picked English language text from Gutenberg library has entropy ≈ 4.5. The reason of
this is because English texts uses mostly 26 symbols, and log2 (26) =≈ 4.7, i.e., you would need 5-bit bytes
to transmit uncompressed English texts, that would be enough (it was indeed so in teletype era).
Randomly chosen Russian language text from http://lib.ru library is F.M.Dostoevsky “Idiot”9 , internally
encoded in CP1251 encoding.
And this file has entropy of ≈ 4.98. Russian language has 33 characters, and log2 (33) =≈ 5.04. But it has
unpopular and rare “ё” character. And log2 (32) = 5 (Russian alphabet without this rare character)—now
this close to what we’ve got.
However, the text we studying uses “ё” letter, but, probably, it’s still rarely used there.
The very same file transcoded from CP1251 to UTF-8 gave entropy of ≈ 4.23. Each Cyrillic character
encoded in UTF-8 is usually encoded as a pair, and the first byte is always one of: 0xD0 or 0xD1. Perhaps,
this caused bias.
Let’s generate random bits and output them as “T” and “F” characters:
import random, sys
rt=""
if random.randint(0,1)==1:
rt=rt+"T"
else:
rt=rt+"F"
print rt
9 http://az.lib.ru/d/dostoewskij_f_m/text_0070.shtml
941
Sample: ...TTTFTFTTTFFFTTTFTTTTTTFTTFFTTTFTFTTFTTFFFFFF....
Entropy is very close to 1 (i.e., 1 bit per byte).
Let’s generate random decimal digits:
import random, sys
rt=""
rt=rt+"%d" % random.randint(0,9)
print rt
Sample: ...52203466119390328807552582367031963888032....
Entropy will be close to 3.32, indeed, this is log2 (10).
9.2.9 Making lower level of entropy

The author of these lines once saw a software which stored each byte of encrypted data in 3 bytes: each
has ≈ 3 value, so reconstructing encrypted byte back involving summing up 3 consecutive bytes. Looks
byte
absurdly.
But some people say this was done in order to conceal the very fact the data has something encrypted
inside: measuring entropy of such block will show much lower level of it.
9.3 Millenium game save file

The “Millenium Return to Earth” is an ancient DOS game (1991), that allows you to mine resources, build
ships, equip them and send them on other planets, and so on10 .
Like many other games, it allows you to save all game state into a file.
Let’s see if we can find something in it.
10 It can be downloaded for free here
942
So there is a mine in the game. Mines at some planets work faster, or slower on others. The set of
resources is also different.
Here we can see what resources are mined at the time:
Figure 9.14: Mine: state 1
Let’s save a game state. This is a file of size 9538 bytes.

Let’s wait some “days” here in the game, and now we’ve got more resources from the mine:
943
Figure 9.15: Mine: state 2
Let’s save game state again.

Now let’s try to just do binary comparison of the save files using the simple DOS/Windows FC utility:
...> FC /b 2200save.i.v1 2200SAVE.I.V2
Comparing files 2200save.i.v1 and 2200SAVE.I.V2

00000016: 0D 04
00000017: 03 04
0000001C: 1F 1E
00000146: 27 3B
00000BDA: 0E 16
00000BDC: 66 9B
00000BDE: 0E 16
00000BE0: 0E 16
00000BE6: DB 4C
00000BE7: 00 01
00000BE8: 99 E8
00000BEC: A1 F3
00000BEE: 83 C7
00000BFB: A8 28
00000BFD: 98 18
00000BFF: A8 28
00000C01: A8 28
00000C07: D8 58
00000C09: E4 A4
00000C0D: 38 B8
00000C0F: E8 68
...
The output is incomplete here, there are more differences, but we will cut result to show the most inter-
esting.
In the first state, we have 14 “units” of hydrogen and 102 “units” of oxygen.
944
We have 22 and 155 “units” respectively in the second state. If these values are saved into the save file,
we would see this in the difference. And indeed we do. There is 0x0E (14) at position 0xBDA and this
value is 0x16 (22) in the new version of the file. This is probably hydrogen. There is 0x66 (102) at position
0xBDC in the old version and 0x9B (155) in the new version of the file. This seems to be the oxygen.
Both files are available on the website for those who wants to inspect them (or experiment) more: begin-
ners.re.
945
Here is the new version of file opened in Hiew, we marked the values related to the resources mined in
the game:
Figure 9.16: Hiew: state 1
Let’s check each of them.

These are clearly 16-bit values: not a strange thing for 16-bit DOS software where the int type has 16-bit
width.
946
Let’s check our assumptions. We will write the 1234 (0x4D2) value at the first position (this must be
hydrogen):
Figure 9.17: Hiew: let’s write 1234 (0x4D2) there
Then we will load the changed file in the game and took a look at mine statistics:
Figure 9.18: Let’s check for hydrogen value
So yes, this is it.
947
Now let’s try to finish the game as soon as possible, set the maximal values everywhere:
Figure 9.19: Hiew: let’s set maximal values
0xFFFF is 65535, so yes, we now have a lot of resources:
Figure 9.20: All resources are 65535 (0xFFFF) indeed
948
Let’s skip some “days” in the game and oops! We have a lower amount of some resources:
Figure 9.21: Resource variables overflow
That’s just overflow.

The game’s developer supposedly didn’t think about such high amounts of resources, so there are probably
no overflow checks, but the mine is “working” in the game, resources are added, hence the overflows.
Apparently, it is a bad idea to be that greedy.
There are probably a lot of more values saved in this file.
So this is very simple method of cheating in games. High score files often can be easily patched like that.
More about files and memory snapshots comparing: 5.10.2 on page 717.
9.4 fortune program indexing file

(This part was first appeared in my blog at 25-Apr-2015.)
fortune is well-known UNIX program which shows random phrase from a collection. Some geeks are often
set up their system in such way, so fortune can be called after logging on. fortune takes phrases from the
text files laying in /usr/share/games/fortunes (as of Ubuntu Linux). Here is example (“fortunes” text file):
A day for firm decisions!!!!! Or is it?
%
A few hours grace before the madness begins again.
%
A gift of a flower will soon be made to you.
%
A long-forgotten loved one will appear soon.
Buy the negatives at any price.

%
A tall, dark stranger will have more fun than you.
949
%
...
So it is just phrases, sometimes multiline ones, divided by percent sign. The task of fortune program is to
find random phrase and to print it. In order to achieve this, it must scan the whole text file, count phrases,
choose random and print it. But the text file can get bigger, and even on modern computers, this naive
algorithm is a bit uneconomical to computer resources. The straightforward way is to keep binary index
file containing offset of each phrase in text file. With index file, fortune program can work much faster:
just to choose random index element, take offset from there, set offset in text file and read phrase from it.
This is actually done in fortune program. Let’s inspect what is in its index file inside (these are .dat files
in the same directory) in hexadecimal editor. This program is open-source of course, but intentionally, I
will not peek into its source code.
% od -t x1 --address-radix=x fortunes.dat
000000 00 00 00 02 00 00 01 af 00 00 00 bb 00 00 00 0f
000010 00 00 00 00 25 00 00 00 00 00 00 00 00 00 00 2b
000020 00 00 00 60 00 00 00 8f 00 00 00 df 00 00 01 14
000030 00 00 01 48 00 00 01 7c 00 00 01 ab 00 00 01 e6
000040 00 00 02 20 00 00 02 3b 00 00 02 7a 00 00 02 c5
000050 00 00 03 04 00 00 03 3d 00 00 03 68 00 00 03 a7
000060 00 00 03 e1 00 00 04 19 00 00 04 2d 00 00 04 7f
000070 00 00 04 ad 00 00 04 d5 00 00 05 05 00 00 05 3b
000080 00 00 05 64 00 00 05 82 00 00 05 ad 00 00 05 ce
000090 00 00 05 f7 00 00 06 1c 00 00 06 61 00 00 06 7a
0000a0 00 00 06 d1 00 00 07 0a 00 00 07 53 00 00 07 9a
0000b0 00 00 07 f8 00 00 08 27 00 00 08 59 00 00 08 8b
0000c0 00 00 08 a0 00 00 08 c4 00 00 08 e1 00 00 08 f9
0000d0 00 00 09 27 00 00 09 43 00 00 09 79 00 00 09 a3
0000e0 00 00 09 e3 00 00 0a 15 00 00 0a 4d 00 00 0a 5e
0000f0 00 00 0a 8a 00 00 0a a6 00 00 0a bf 00 00 0a ef
000100 00 00 0b 18 00 00 0b 43 00 00 0b 61 00 00 0b 8e
000110 00 00 0b cf 00 00 0b fa 00 00 0c 3b 00 00 0c 66
000120 00 00 0c 85 00 00 0c b9 00 00 0c d2 00 00 0d 02
000130 00 00 0d 3b 00 00 0d 67 00 00 0d ac 00 00 0d e0
000140 00 00 0e 1e 00 00 0e 67 00 00 0e a5 00 00 0e da
000150 00 00 0e ff 00 00 0f 43 00 00 0f 8a 00 00 0f bc
000160 00 00 0f e5 00 00 10 1e 00 00 10 63 00 00 10 9d
000170 00 00 10 e3 00 00 11 10 00 00 11 46 00 00 11 6c
000180 00 00 11 99 00 00 11 cb 00 00 11 f5 00 00 12 32
000190 00 00 12 61 00 00 12 8c 00 00 12 ca 00 00 13 87
0001a0 00 00 13 c4 00 00 13 fc 00 00 14 1a 00 00 14 6f
0001b0 00 00 14 ae 00 00 14 de 00 00 15 1b 00 00 15 55
0001c0 00 00 15 a6 00 00 15 d8 00 00 16 0f 00 00 16 4e
...
Without any special aid we could see that there are four 4-byte elements on each 16-byte line. Perhaps,
it’s our index array. I’m trying to load the whole file in Wolfram Mathematica as 32-bit integer array:
In[]:= BinaryReadList["c:/tmp1/fortunes.dat", "UnsignedInteger32"]
Out[]= {33554432, 2936078336, 3137339392, 251658240, 0, 37, 0, \

721420288, 1610612736, 2399141888, 3741319168, 335609856, 1208025088, \
2080440320, 2868969472, 3858825216, 537001984, 989986816, 2046951424, \
3305242624, 67305472, 1023606784, 1745027072, 2801991680, 3775070208, \
419692544, 755236864, 2130968576, 2902720512, 3573809152, 84213760, \
990183424, 1678049280, 2181365760, 2902786048, 3456434176, \
4144300032, 470155264, 1627783168, 2047213568, 3506831360, 168230912, \
1392967680, 2584150016, 4161208320, 654835712, 1493696512, \
2332557312, 2684878848, 3288858624, 3775397888, 4178051072, \
...
Nope, something wrong. Numbers are suspiciously big. But let’s back to od output: each 4-byte element
has two zero bytes and two non-zero bytes, so the offsets (at least at the beginning of the file) are 16-
bit at maximum. Probably different endianness is used in the file? Default endiannes in Mathematica is
little-endian, as used in Intel CPUs. Now I’m changing it to big-endian:
In[]:= BinaryReadList["c:/tmp1/fortunes.dat", "UnsignedInteger32",
ByteOrdering -> 1]
950
Out[]= {2, 431, 187, 15, 0, 620756992, 0, 43, 96, 143, 223, 276, \
328, 380, 427, 486, 544, 571, 634, 709, 772, 829, 872, 935, 993, \
1049, 1069, 1151, 1197, 1237, 1285, 1339, 1380, 1410, 1453, 1486, \
1527, 1564, 1633, 1658, 1745, 1802, 1875, 1946, 2040, 2087, 2137, \
2187, 2208, 2244, 2273, 2297, 2343, 2371, 2425, 2467, 2531, 2581, \
2637, 2654, 2698, 2726, 2751, 2799, 2840, 2883, 2913, 2958, 3023, \
3066, 3131, 3174, 3205, 3257, 3282, 3330, 3387, 3431, 3500, 3552, \
...
Yes, this is something readable. I choose random element (3066) which is 0xBFA in hexadecimal form. I’m
opening ’fortunes’ text file in hex editor, I’m setting 0xBFA as offset and I see this phrase:
% od -t x1 -c --skip-bytes=0xbfa --address-radix=x fortunes
000bfa 44 6f 20 77 68 61 74 20 63 6f 6d 65 73 20 6e 61
D o w h a t c o m e s n a
000c0a 74 75 72 61 6c 6c 79 2e 20 20 53 65 65 74 68 65
t u r a l l y . S e e t h e
000c1a 20 61 6e 64 20 66 75 6d 65 20 61 6e 64 20 74 68
a n d f u m e a n d t h
....
Or:
Do what comes naturally. Seethe and fume and throw a tantrum.
%
Other offset are also can be checked, yes, they are valid offsets.
I can also check in Mathematica that each subsequent element is bigger than previous. I.e., elements of
array are ascending. In mathematics lingo, this is called strictly increasing monotonic function.
In[]:= Differences[input]
Out[]= {429, -244, -172, -15, 620756992, -620756992, 43, 53, 47, \
80, 53, 52, 52, 47, 59, 58, 27, 63, 75, 63, 57, 43, 63, 58, 56, 20, \
82, 46, 40, 48, 54, 41, 30, 43, 33, 41, 37, 69, 25, 87, 57, 73, 71, \
94, 47, 50, 50, 21, 36, 29, 24, 46, 28, 54, 42, 64, 50, 56, 17, 44, \
28, 25, 48, 41, 43, 30, 45, 65, 43, 65, 43, 31, 52, 25, 48, 57, 44, \
69, 52, 62, 73, 62, 53, 37, 68, 71, 50, 41, 57, 69, 58, 70, 45, 54, \
38, 45, 50, 42, 61, 47, 43, 62, 189, 61, 56, 30, 85, 63, 48, 61, 58, \
81, 50, 55, 63, 83, 80, 49, 42, 94, 54, 67, 81, 52, 57, 68, 43, 28, \
120, 64, 53, 81, 33, 82, 88, 29, 61, 32, 75, 63, 70, 47, 101, 60, 79, \
33, 48, 65, 35, 59, 47, 55, 22, 43, 35, 102, 53, 80, 65, 45, 31, 29, \
69, 32, 25, 38, 34, 35, 49, 59, 39, 41, 18, 43, 41, 83, 37, 31, 34, \
59, 72, 72, 81, 77, 53, 53, 50, 51, 45, 53, 39, 70, 54, 103, 33, 70, \
51, 95, 67, 54, 55, 65, 61, 54, 54, 53, 45, 100, 63, 48, 65, 71, 23, \
28, 43, 51, 61, 101, 65, 39, 78, 66, 43, 36, 56, 40, 67, 92, 65, 61, \
31, 45, 52, 94, 82, 82, 91, 46, 76, 55, 19, 58, 68, 41, 75, 30, 67, \
92, 54, 52, 108, 60, 56, 76, 41, 79, 54, 65, 74, 112, 76, 47, 53, 61, \
66, 53, 28, 41, 81, 75, 69, 89, 63, 60, 18, 18, 50, 79, 92, 37, 63, \
88, 52, 81, 60, 80, 26, 46, 80, 64, 78, 70, 75, 46, 91, 22, 63, 46, \
34, 81, 75, 59, 62, 66, 74, 76, 111, 55, 73, 40, 61, 55, 38, 56, 47, \
78, 81, 62, 37, 41, 60, 68, 40, 33, 54, 34, 41, 36, 49, 44, 68, 51, \
50, 52, 36, 53, 66, 46, 41, 45, 51, 44, 44, 33, 72, 40, 71, 57, 55, \
39, 66, 40, 56, 68, 43, 88, 78, 30, 54, 64, 36, 55, 35, 88, 45, 56, \
76, 61, 66, 29, 76, 53, 96, 36, 46, 54, 28, 51, 82, 53, 60, 77, 21, \
84, 53, 43, 104, 85, 50, 47, 39, 66, 78, 81, 94, 70, 49, 67, 61, 37, \
51, 91, 99, 58, 51, 49, 46, 68, 72, 40, 56, 63, 65, 41, 62, 47, 41, \
43, 30, 43, 67, 78, 80, 101, 61, 73, 70, 41, 82, 69, 45, 65, 38, 41, \
57, 82, 66}
As we can see, except of the very first 6 values (which is probably belongs to index file header), all numbers
are in fact length of all text phrases (offset of the next phrase minus offset of the current phrase is in fact
length of the current phrase).
It’s very important to keep in mind that bit-endiannes can be confused with incorrect array start. Indeed,
from od output we see that each element started with two zeros. But when shifted by two bytes in either
side, we can interpret this array as little-endian:
951
% od -t x1 --address-radix=x --skip-bytes=0x32 fortunes.dat
000032 01 48 00 00 01 7c 00 00 01 ab 00 00 01 e6 00 00
000042 02 20 00 00 02 3b 00 00 02 7a 00 00 02 c5 00 00
000052 03 04 00 00 03 3d 00 00 03 68 00 00 03 a7 00 00
000062 03 e1 00 00 04 19 00 00 04 2d 00 00 04 7f 00 00
000072 04 ad 00 00 04 d5 00 00 05 05 00 00 05 3b 00 00
000082 05 64 00 00 05 82 00 00 05 ad 00 00 05 ce 00 00
000092 05 f7 00 00 06 1c 00 00 06 61 00 00 06 7a 00 00
0000a2 06 d1 00 00 07 0a 00 00 07 53 00 00 07 9a 00 00
0000b2 07 f8 00 00 08 27 00 00 08 59 00 00 08 8b 00 00
0000c2 08 a0 00 00 08 c4 00 00 08 e1 00 00 08 f9 00 00
0000d2 09 27 00 00 09 43 00 00 09 79 00 00 09 a3 00 00
0000e2 09 e3 00 00 0a 15 00 00 0a 4d 00 00 0a 5e 00 00
...
If we would interpret this array as little-endian, the first element is 0x4801, second is 0x7C01, etc. High
8-bit part of each of these 16-bit values are seems random to us, and the lowest 8-bit part is seems
ascending.
But I’m sure that this is big-endian array, because the very last 32-bit element of the file is big-endian (00
00 5f c4 here):
...
000660 00 00 59 0d 00 00 59 55 00 00 59 7d 00 00 59 b5
000670 00 00 59 f4 00 00 5a 35 00 00 5a 5e 00 00 5a 9c
000680 00 00 5a cb 00 00 5a f4 00 00 5b 1f 00 00 5b 3d
000690 00 00 5b 68 00 00 5b ab 00 00 5b f9 00 00 5c 49
0006a0 00 00 5c ae 00 00 5c eb 00 00 5d 34 00 00 5d 7a
0006b0 00 00 5d a3 00 00 5d f5 00 00 5e 3a 00 00 5e 67
0006c0 00 00 5e a8 00 00 5e ce 00 00 5e f7 00 00 5f 30
0006d0 00 00 5f 82 00 00 5f c4
0006d8
Perhaps, fortune program developer had big-endian computer or maybe it was ported from something
like it.
OK, so the array is big-endian, and, judging by common sense, the very first phrase in the text file must
be started at zeroth offset. So zero value should be present in the array somewhere at the very beginning.
We’ve got couple of zero elements at the beginning. But the second is most appealing: 43 is going right
after it and 43 is valid offset to valid English phrase in the text file.
The last array element is 0x5FC4, and there are no such byte at this offset in the text file. So the last
array element is pointing behind the end of file. It’s supposedly done because phrase length is calculated
as difference between offset to the current phrase and offset to the next phrase. This can be faster than
traversing phrase string for percent character. But this wouldn’t work for the last element. So the dummy
element is also added at the end of array.
So the first 6 32-bit integer values are supposedly some kind of header.
Oh, I forgot to count phrases in text file:
% cat fortunes | grep % | wc -l
432
The number of phrases can be present in index, but may be not. In case of very simple index files, number
of elements can be easily deduced from index file size. Anyway, there are 432 phrases in the text file. And
we see something very familiar at the second element (value 431). I’ve checked other files (literature.dat
and riddles.dat in Ubuntu Linux) and yes, the second 32-bit element is indeed number of phrases minus 1.
Why minus 1? Perhaps, this is not number of phrases, but rather the number of the last phrase (starting
at zero)?
And there are some other elements in the header. In Mathematica, I’m loading each of three available
files and I’m taking a look on the header:
952
I have no idea what other values mean, except the size of index file. Some fields are the same for all files,
some are not. From my own experience, there could be:
• file signature;
• file version;
• checksum;
• some flags;
• maybe even text language identifier;
• text file timestamp, so the fortune program will regenerate index file if a user modified text file.
For example, Oracle .SYM files ( 9.5 on the next page) which contain symbols table for DLL files, also
contain timestamp of corresponding DLL file, so to be sure it is still valid.
On the other hand, text file and index file timestamps can gone out of sync after archiving/unarchiving/in-
stalling/deploying/etc.
But there are no timestamp, in my opinion. The most compact way of representing date and time is UNIX
time value, which is big 32-bit number. We don’t see any of such here. Other ways of representation are
even less compact.
So here is algorithm, how fortune supposedly works:
• take number of last phrase from the second element;
• generate random number in range of 0..number_of_last_phrase;
• find corresponding element in array of offsets, take also following offset;
• output to stdout all characters from the text file starting at the offset until the next offset minus 2
(so to ignore terminating percent sign and character of the following phrase).
9.4.1 Hacking
Let’s try to check some of our assumptions. I will create this text file under the path and name /us-
r/share/games/fortunes/fortunes:
Phrase one.
%
Phrase two.
%
Then this fortunes.dat file. I take header from the original fortunes.dat, I changed second field (count of
all phrases) to zero and I left two elements in the array: 0 and 0x1c, because the whole length of the text
fortunes file is 28 (0x1c) bytes:
953
000000 00 00 00 02 00 00 00 00 00 00 00 bb 00 00 00 0f
000010 00 00 00 00 25 00 00 00 00 00 00 00 00 00 00 1c
Now I run it:

% /usr/games/fortune
fortune: no fortune found
Something wrong. Let’s change the second field to 1:

000000 00 00 00 02 00 00 00 01 00 00 00 bb 00 00 00 0f
000010 00 00 00 00 25 00 00 00 00 00 00 00 00 00 00 1c
Now it works. It’s always shows only the first phrase:

% /usr/games/fortune
Phrase one.
Hmmm. Let’s leave only one element in array (0) without terminating one:
000000 00 00 00 02 00 00 00 01 00 00 00 bb 00 00 00 0f
000010 00 00 00 00 25 00 00 00 00 00 00 00
00001c
Fortune program always shows only first phrase.

From this experiment we got to know that percent sign in text file is parsed and the size is not calculated
as I deduced before, perhaps, even terminal array element is not used. However, it still can be used. And
probably it was used in past?
9.4.2 The files

For the sake of demonstration, I still didn’t take a look in fortune source code. If you want to try to
understand meaning of other values in index file header, you may try to achieve it without looking into
source code as well. Files I took from Ubuntu Linux 14.04 are here: http://beginners.re/examples/
fortune/, hacked files are also here.
Oh, and I took the files from x64 version of Ubuntu, but array elements are still has size of 32 bit. It is
because fortune text files are probably never exceeds 4GiB11 size. But if it will, all elements must have
size of 64 bit so to be able to store offset to the text file larger than 4GiB.
For impatient readers, the source code of fortune is here: https://launchpad.net/ubuntu/+source/
fortune-mod/1:1.99.1-3.1ubuntu4.
9.5 Oracle RDBMS: .SYM-files

When an Oracle RDBMS process experiences some kind of crash, it writes a lot of information into log files,
including stack trace, like this:
----- Call Stack Trace -----
calling call entry argument values in hex
location type point (? means dubious value)
-------------------- -------- -------------------- ----------------------------
_kqvrow() 00000000
_opifch2()+2729 CALLptr 00000000 23D4B914 E47F264 1F19AE2
EB1C8A8 1
_kpoal8()+2832 CALLrel _opifch2() 89 5 EB1CC74
_opiodr()+1248 CALLreg 00000000 5E 1C EB1F0A0
_ttcpip()+1051 CALLreg 00000000 5E 1C EB1F0A0 0
_opitsk()+1404 CALL??? 00000000 C96C040 5E EB1F0A0 0 EB1ED30
EB1F1CC 53E52E 0 EB1F1F8
_opiino()+980 CALLrel _opitsk() 0 0
11 Gibibyte
954
_opiodr()+1248 CALLreg 00000000 3C 4 EB1FBF4
_opidrv()+1201 CALLrel _opiodr() 3C 4 EB1FBF4 0
_sou2o()+55 CALLrel _opidrv() 3C 4 EB1FBF4
_opimai_real()+124 CALLrel _sou2o() EB1FC04 3C 4 EB1FBF4
_opimai()+125 CALLrel _opimai_real() 2 EB1FC2C
_OracleThreadStart@ CALLrel _opimai() 2 EB1FF6C 7C88A7F4 EB1FC34 0
4()+830 EB1FD04
77E6481C CALLreg 00000000 E41FF9C 0 0 E41FF9C 0 EB1FFC4
00000000 CALL??? 00000000
But of course, Oracle RDBMS’s executables must have some kind of debug information or map files with
symbol information included or something like that.
Windows NT Oracle RDBMS has symbol information in files with .SYM extension, but the format is propri-
etary. (Plain text files are good, but needs additional parsing, hence offer slower access.)
Let’s see if we can understand its format.
12
We will pick the shortest orawtc8.sym file that comes with the orawtc8.dll file in Oracle 8.1.7 .
12 We can chose an ancient Oracle RDBMS version intentionally due to the smaller size of its modules
955
Here is the file opened in Hiew:
Figure 9.22: The whole file in Hiew
By comparing the file with other .SYM files, we can quickly see that OSYM is always header (and footer),
so this is maybe the file’s signature.
We also see that basically, the file format is: OSYM + some binary data + zero delimited text strings +
OSYM. The strings are, obviously, function and global variable names.
956
We will mark the OSYM signatures and strings here:
Figure 9.23: OSYM signature and text strings
Well, let’s see. In Hiew, we will mark the whole strings block (except the trailing OSYM signatures) and
put it into a separate file. Then we run UNIX strings and wc utilities to count the text strings:
strings strings_block | wc -l
66
So there are 66 text strings. Please note that number.

We can say, in general, as a rule, the number of anything is often stored separately in binary files.
It’s indeed so, we can find the 66 value (0x42) at the file’s start, right after the OSYM signature:
$ hexdump -C orawtc8.sym
00000000 4f 53 59 4d 42 00 00 00 00 10 00 10 80 10 00 10 |OSYMB...........|
00000010 f0 10 00 10 50 11 00 10 60 11 00 10 c0 11 00 10 |....P...`.......|
00000020 d0 11 00 10 70 13 00 10 40 15 00 10 50 15 00 10 |....p...@...P...|
00000030 60 15 00 10 80 15 00 10 a0 15 00 10 a6 15 00 10 |`...............|
....
Of course, 0x42 here is not a byte, but most likely a 32-bit value packed as little-endian, hence we see
0x42 and then at least 3 zero bytes.
Why do we believe it’s 32-bit? Because, Oracle RDBMS’s symbol files may be pretty big.
The oracle.sym file for the main oracle.exe (version 10.2.0.4) executable contains 0x3A38E (238478) sym-
bols. A 16-bit value isn’t enough here.
We can check other .SYM files like this and it proves our guess: the value after the 32-bit OSYM signature
always reflects the number of text strings in the file.
957
It’s a general feature of almost all binary files: a header with a signature plus some other information
about the file.
Now let’s investigate closer what this binary block is.
Using Hiew again, we put the block starting at address 8 (i.e., after the 32-bit count value) ending at the
strings block, into a separate binary file.
958
Let’s see the binary block in Hiew:
Figure 9.24: Binary block
There is a clear pattern in it.
959
We will add red lines to divide the block:
Figure 9.25: Binary block patterns
Hiew, like almost any other hexadecimal editor, shows 16 bytes per line. So the pattern is clearly visible:
there are 4 32-bit values per line.
The pattern is visually visible because some values here (till address 0x104) are always in 0x1000xxxx
form, started with 0x10 and zero bytes.
Other values (starting at 0x108) are in 0x0000xxxx form, so always started with two zero bytes.
Let’s dump the block as an array of 32-bit values:
Listing 9.9: first column is address

$ od -v -t x4 binary_block
0000000 10001000 10001080 100010f0 10001150
0000020 10001160 100011c0 100011d0 10001370
0000040 10001540 10001550 10001560 10001580
0000060 100015a0 100015a6 100015ac 100015b2
0000100 100015b8 100015be 100015c4 100015ca
0000120 100015d0 100015e0 100016b0 10001760
0000140 10001766 1000176c 10001780 100017b0
0000160 100017d0 100017e0 10001810 10001816
0000200 10002000 10002004 10002008 1000200c
0000220 10002010 10002014 10002018 1000201c
960
0000240 10002020 10002024 10002028 1000202c
0000260 10002030 10002034 10002038 1000203c
0000300 10002040 10002044 10002048 1000204c
0000320 10002050 100020d0 100020e4 100020f8
0000340 1000210c 10002120 10003000 10003004
0000360 10003008 1000300c 10003098 1000309c
0000400 100030a0 100030a4 00000000 00000008
0000420 00000012 0000001b 00000025 0000002e
0000440 00000038 00000040 00000048 00000051
0000460 0000005a 00000064 0000006e 0000007a
0000500 00000088 00000096 000000a4 000000ae
0000520 000000b6 000000c0 000000d2 000000e2
0000540 000000f0 00000107 00000110 00000116
0000560 00000121 0000012a 00000132 0000013a
0000600 00000146 00000153 00000170 00000186
0000620 000001a9 000001c1 000001de 000001ed
0000640 000001fb 00000207 0000021b 0000022a
0000660 0000023d 0000024e 00000269 00000277
0000700 00000287 00000297 000002b6 000002ca
0000720 000002dc 000002f0 00000304 00000321
0000740 0000033e 0000035d 0000037a 00000395
0000760 000003ae 000003b6 000003be 000003c6
0001000 000003ce 000003dc 000003e9 000003f8
0001020
There are 132 values, that’s 66*2. Probably, there are two 32-bit values for each symbol, but maybe there
are two arrays? Let’s see.
Values starting with 0x1000 may be addresses.
This is a .SYM file for a DLL after all, and the default base address of win32 DLLs is 0x10000000, and the
code usually starts at 0x10001000.
When we open the orawtc8.dll file in IDA, the base address is different, but nevertheless, the first function
is:
.text:60351000 sub_60351000 proc near
.text:60351000
.text:60351000 arg_4 = dword ptr 0Ch
.text:60351000 arg_8 = dword ptr 10h
.text:60351000
.text:60351003 mov eax, dword_60353014
.text:60351008 cmp eax, 0FFFFFFFFh
.text:6035100B jnz short loc_6035104F
.text:6035100D mov ecx, hModule
.text:60351015 cmp ecx, 0FFFFFFFFh
.text:60351018 mov dword_60353014, eax
.text:6035101D jnz short loc_60351031
.text:6035101F call sub_603510F0
.text:60351024 mov ecx, eax
.text:6035102B mov hModule, ecx
.text:60351031
.text:60351031 loc_60351031: ; CODE XREF: sub_60351000+1D
.text:60351031 test ecx, ecx
.text:60351033 jbe short loc_6035104F
.text:60351035 push offset ProcName ; "ax_reg"
.text:6035103A push ecx ; hModule
.text:6035103B call ds:GetProcAddress
...
Wow, “ax_reg” string sounds familiar.

It’s indeed the first string in the strings block! So the name of this function seems to be “ax_reg”.
The second function is:
961
.text:60351080 sub_60351080 proc near
.text:60351080
.text:60351080 arg_4 = dword ptr 0Ch
.text:60351080
.text:60351088 cmp eax, 0FFFFFFFFh
.text:6035108B jnz short loc_603510CF
.text:6035108D mov ecx, hModule
.text:60351095 cmp ecx, 0FFFFFFFFh
.text:60351098 mov dword_60353018, eax
.text:6035109D jnz short loc_603510B1
.text:6035109F call sub_603510F0
.text:603510A4 mov ecx, eax
.text:603510A6 mov eax, dword_60353018
.text:603510AB mov hModule, ecx
.text:603510B1
.text:603510B1 loc_603510B1: ; CODE XREF: sub_60351080+1D
.text:603510B1 test ecx, ecx
.text:603510B3 jbe short loc_603510CF
.text:603510B5 push offset aAx_unreg ; "ax_unreg"
.text:603510BA push ecx ; hModule
.text:603510BB call ds:GetProcAddress
...
The “ax_unreg” string is also the second string in the strings block!
The starting address of the second function is 0x60351080, and the second value in the binary block is
10001080. So this is the address, but for a DLL with the default base address.
We can quickly check and be sure that the first 66 values in the array (i.e., the first half of the array) are
just function addresses in the DLL, including some labels, etc. Well, what’s the other part of array then?
The other 66 values that start with 0x0000? These seem to be in range [0...0x3F8]. And they do not
look like bitfields: the series of numbers is increasing.
The last hexadecimal digit seems to be random, so, it’s unlikely the address of something (it would be
divisible by 4 or maybe 8 or 0x10 otherwise).
Let’s ask ourselves: what else Oracle RDBMS’s developers would save here, in this file?
Quick wild guess: it could be the address of the text string (function name).
It can be quickly checked, and yes, each number is just the position of the first character in the strings
block.
This is it! All done.
We will write an utility to convert these .SYM files into IDA script, so we can load the .idc script and it sets
the function names:
#include <stdio.h>
#include <stdint.h>
#include <io.h>
#include <assert.h>
#include <malloc.h>
#include <fcntl.h>
#include <string.h>
int main (int argc, char *argv[])

{
uint32_t sig, cnt, offset;
uint32_t *d1, *d2;
int h, i, remain, file_len;
char *d3;
uint32_t array_size_in_bytes;
assert (argv[1]); // file name

assert (argv[2]); // additional offset (if needed)
962
// additional offset
assert (sscanf (argv[2], "%X", &offset)==1);
// get file length

assert ((h=open (argv[1], _O_RDONLY | _O_BINARY, 0))!=-1);
assert ((file_len=lseek (h, 0, SEEK_END))!=-1);
assert (lseek (h, 0, SEEK_SET)!=-1);
// read signature
assert (read (h, &sig, 4)==4);
// read count
assert (read (h, &cnt, 4)==4);
assert (sig==0x4D59534F); // OSYM
// skip timedatestamp (for 11g)

//_lseek (h, 4, 1);
array_size_in_bytes=cnt*sizeof(uint32_t);
// load symbol addresses array

d1=(uint32_t*)malloc (array_size_in_bytes);
assert (d1);
assert (read (h, d1, array_size_in_bytes)==array_size_in_bytes);
// load string offsets array

d2=(uint32_t*)malloc (array_size_in_bytes);
assert (d2);
assert (read (h, d2, array_size_in_bytes)==array_size_in_bytes);
// calculate strings block size

remain=file_len-(8+4)-(cnt*8);
// load strings block

assert (d3=(char*)malloc (remain));
assert (read (h, d3, remain)==remain);
printf ("#include <idc.idc>\n\n");

printf ("static main() {\n");

printf ("\tMakeName(0x%08X, \"%s\");\n", offset + d1[i], &d3[d2[i]]);
printf ("}\n");
close (h);
free (d1); free (d2); free (d3);
};
Here is an example of its work:

#include <idc.idc>
static main() {
MakeName(0x60351000, "_ax_reg");
MakeName(0x60351080, "_ax_unreg");
MakeName(0x603510F0, "_loaddll");
MakeName(0x60351150, "_wtcsrin0");
MakeName(0x60351160, "_wtcsrin");
MakeName(0x603511C0, "_wtcsrfre");
MakeName(0x603511D0, "_wtclkm");
MakeName(0x60351370, "_wtcstu");
...
}
The example files were used in this example are here: beginners.re.
963
Oh, let’s also try Oracle RDBMS for win64. There has to be 64-bit addresses instead, right?
The 8-byte pattern is visible even easier here:
Figure 9.26: .SYM-file example from Oracle RDBMS for win64
So yes, all tables now have 64-bit elements, even string offsets!
The signature is now OSYMAM64, to distinguish the target platform, apparently.
This is it!
Here is also library which has functions to access Oracle RDBMS.SYM-files: GitHub.
9.6 Oracle RDBMS: .MSB-files

When working toward the solution of a
problem, it always helps if you know the
answer.
Murphy’s Laws, Rule of Accuracy
This is a binary file that contains error messages with their corresponding numbers. Let’s try to understand
its format and find a way to unpack it.
964
There are Oracle RDBMS error message files in text form, so we can compare the text and packed binary
files 13 .
This is the beginning of the ORAUS.MSG text file with some irrelevant comments stripped:
Listing 9.10: Beginning of ORAUS.MSG file without comments

00000, 00000, "normal, successful completion"
00001, 00000, "unique constraint (%s.%s) violated"
00017, 00000, "session requested to set trace event"
00018, 00000, "maximum number of sessions exceeded"
00019, 00000, "maximum number of session licenses exceeded"
00020, 00000, "maximum number of processes (%s) exceeded"
00021, 00000, "session attached to some other process; cannot switch session"
00022, 00000, "invalid session ID; access denied"
00023, 00000, "session references process private memory; cannot detach session"
00024, 00000, "logins from more than one process not allowed in single-process mode"
00025, 00000, "failed to allocate %s"
00026, 00000, "missing or invalid session ID"
00027, 00000, "cannot kill current session"
00028, 00000, "your session has been killed"
00029, 00000, "session is not a user session"
00030, 00000, "User session ID does not exist."
00031, 00000, "session marked for kill"
...
The first number is the error code. The second is perhaps maybe some additional flags.
13 Open-source text files don’t exist in Oracle RDBMS for every .MSB file, so that’s why we will work on their file format
965
Now let’s open the ORAUS.MSB binary file and find these text strings. And there are:
Figure 9.27: Hiew: first block
We see the text strings (including those from the beginning of the ORAUS.MSG file) interleaved with some
binary values. By quick investigation, we can see that main part of the binary file is divided by blocks of
size 0x200 (512) bytes.
966
Let’s see the contents of the first block:
Figure 9.28: Hiew: first block
Here we see the texts of the first messages errors. What we also see is that there are no zero bytes
between the error messages. This implies that these are not null-terminated C strings. As a consequence,
the length of each error message must be encoded somehow. Let’s also try to find the error numbers.
The ORAUS.MSG files starts with these: 0, 1, 17 (0x11), 18 (0x12), 19 (0x13), 20 (0x14), 21 (0x15), 22
(0x16), 23 (0x17), 24 (0x18)... We will find these numbers at the beginning of the block and mark them
with red lines. The period between error codes is 6 bytes.
This implies that there are probably 6 bytes of information allocated for each error message.
The first 16-bit value (0xA here or 10) means the number of messages in each block: this can be checked
by investigating other blocks. Indeed: the error messages have arbitrary size. Some are longer, some are
shorter. But block size is always fixed, hence, you never know how many text messages can be packed
in each block.
As we already noted, since these are not null-terminated C strings, their size must be encoded somewhere.
The size of the first string “normal, successful completion” is 29 (0x1D) bytes. The size of the second string
“unique constraint (%s.%s) violated” is 34 (0x22) bytes. We can’t find these values (0x1D or/and 0x22)
in the block.
There is also another thing. Oracle RDBMS has to determine the position of the string it needs to load in the
block, right? The first string “normal, successful completion” starts at position 0x1444 (if we count starting
at the beginning of the file) or at 0x44 (from the block’s start). The second string “unique constraint
967
(%s.%s) violated” starts at position 0x1461 (from the file’s start) or at 0x61 (from the at the block’s start).
These numbers (0x44 and 0x61) are familiar somehow! We can clearly see them at the start of the block.
So, each 6-byte block is:
• 16-bit error number;
• 16-bit zero (maybe additional flags);
• 16-bit starting position of the text string within the current block.
We can quickly check the other values and be sure our guess is correct. And there is also the last “dummy”
6-byte block with an error number of zero and starting position beyond the last error message’s last
character. Probably that’s how text message length is determined? We just enumerate 6-byte blocks to
find the error number we need, then we get the text string’s position, then we get the position of the text
string by looking at the next 6-byte block! This way we determine the string’s boundaries! This method
allows to save some space by not saving the text string’s size in the file!
It’s not possible to say it saves a lot of space, but it’s a clever trick.
968
Let’s back to the header of .MSB-file:
Figure 9.29: Hiew: file header
Now we can quickly find the number of blocks in the file (marked by red). We can checked other .MSB-files
and we see that it’s true for all of them.
There are a lot of other values, but we will not investigate them, since our job (an unpacking utility) is
done.
If we have to write a .MSB file packer, we would probably have to understand the meaning of the other
values.
969
There is also a table that came after the header which probably contains 16-bit values:
Figure 9.30: Hiew: last_errnos table
Their size can be determined visually (red lines are drawn here).
While dumping these values, we have found that each 16-bit number is the last error code for each block.
So that’s how Oracle RDBMS quickly finds the error message:
• load a table we will call last_errnos (that contains the last error number for each block);
• find a block that contains the error code we need, assuming all error codes increase across each
block and across the file as well;
• load the specific block;
• enumerate the 6-byte structures until the specific error number is found;
• get the position of the first character from the current 6-byte block;
• get the position of the last character from the next 6-byte block;
• load all characters of the message in this range.
This is C program that we wrote which unpacks .MSB-files: beginners.re.
There are also the two files which were used in the example (Oracle RDBMS 11.1.0.6): beginners.re,
beginners.re.
970
9.6.1 Summary
The method is probably too old-school for modern computers. Supposedly, this file format was developed
in the mid-80’s by someone who also coded for big iron with memory/disk space economy in mind. Nev-
ertheless, it has been an interesting and yet easy task to understand a proprietary file format without
looking into Oracle RDBMS’s code.
9.7 Exercises
Try to reverse engineer of any binary files of your favorite game, including high-score files, resources, etc.
There are also binary files with known structure: utmp/wtmp files, try to understand its structure without
documentation.
The EXIF header in JPEG file is documented, but you can try to understand its structure without help, just
shoot photos at various date/time, places, and try to find date/time and GPS location in EXIF. Try to patch
GPS location, upload JPEG file to Facebook and see, how it will put your picture on the map.
Try to patch any information in MP3 file and see how your favorite MP3-player will react.
9.8 Further reading

Pierre Capillon – Black-box cryptanalysis of home-made encryption algorithms: a practical case study.
How to Hack an Expensive Camera and Not Get Killed by Your Wife.
971
Chapter 10
Dynamic binary instrumentation
DBI tools can be viewed as highly advanced and fast debuggers.
10.1 Using PIN DBI for XOR interception

PIN from Intel is a DBI tool. That means, it takes compiled binary and inserts your instructions in it, where
you want.
Let’s try to intercept all XOR instructions. These are heavily used in cryptography, and we can try to
run WinRAR archiver in encryption mode with a hope that some XOR instruction is indeed is used while
encryption.
Here is the source code of my PIN tool: https://beginners.re/current-tree/DBI/XOR/files//XOR_
ins.cpp.
The code is almost self-explanatory: it scans input executable file for all XOR/PXOR instructions and inserts
a call to our function before each. log_info() function first checks, if operands are different (since XOR is
often used just to clear register, like XOR EAX, EAX), and if they are different, it increments a counter at
this EIP/RIP, so the statistics will be gathered.
I have prepared two files for test: test1.bin (30720 bytes) and test2.bin (5547752 bytes), I’ll compress
them by RAR with password and see difference in statistics.
You’ll also need to turn off ASLR 1 , so the PIN tool will report the same RIPs as in RAR executable.
Now let’s run it:
c:\pin-3.2-81205-msvc-windows\pin.exe -t XOR_ins.dll -- rar a -pLongPassword tmp.rar test1.bin
c:\pin-3.2-81205-msvc-windows\pin.exe -t XOR_ins.dll -- rar a -pLongPassword tmp.rar test2.bin
Now here is statistics for the test1.bin:

https://beginners.re/current-tree/DBI/XOR/files//XOR_ins.out.test1. ... and for test2.bin:
https://beginners.re/current-tree/DBI/XOR/files//XOR_ins.out.test2. So far, you can ignore
all addresses other than ip=0x1400xxxxx, which are in other DLLs.
Now let’s see a difference: https://beginners.re/current-tree/DBI/XOR/files//XOR_ins.diff.
Some XOR instructions are executed more often for test2.bin (which is bigger) than for test1.bin (which is
smaller). So these are clearly related to file size!
The first block of differences is:
< ip=0x140017b21 count=0xd84
< ip=0x140017b48 count=0x81f
< ip=0x140017b59 count=0x858
< ip=0x140017b6a count=0xc13
< ip=0x140017b7b count=0xefc
< ip=0x140017b8a count=0xefd
< ip=0x140017b92 count=0xb86
< ip=0x140017ba1 count=0xf01
---
> ip=0x140017b21 count=0x9eab5
1 https://stackoverflow.com/q/9560993
972
> ip=0x140017b48 count=0x79863
> ip=0x140017b59 count=0x862e8
> ip=0x140017b6a count=0x99495
> ip=0x140017b7b count=0xa891c
> ip=0x140017b8a count=0xa89f4
> ip=0x140017b92 count=0x8ed72
> ip=0x140017ba1 count=0xa8a8a
This is indeed some kind of loop inside of RAR.EXE:

.text:0000000140017B21 loc_140017B21:
.text:0000000140017B21 xor r11d, [rbx]
.text:0000000140017B24 mov r9d, [rbx+4]
.text:0000000140017B28 add rbx, 8
.text:0000000140017B2C mov eax, r9d
.text:0000000140017B2F shr eax, 18h
.text:0000000140017B32 movzx edx, al
.text:0000000140017B35 mov eax, r9d
.text:0000000140017B38 shr eax, 10h
.text:0000000140017B3B movzx ecx, al
.text:0000000140017B3E mov eax, r9d
.text:0000000140017B41 shr eax, 8
.text:0000000140017B44 mov r8d, [rsi+rdx*4]
.text:0000000140017B48 xor r8d, [rsi+rcx*4+400h]
.text:0000000140017B50 movzx ecx, al
.text:0000000140017B53 mov eax, r11d
.text:0000000140017B56 shr eax, 18h
.text:0000000140017B59 xor r8d, [rsi+rcx*4+800h]
.text:0000000140017B64 mov eax, r11d
.text:0000000140017B67 shr eax, 10h
.text:0000000140017B6A xor r8d, [rsi+rcx*4+1000h]
.text:0000000140017B75 mov eax, r11d
.text:0000000140017B78 shr eax, 8
.text:0000000140017B7B xor r8d, [rsi+rcx*4+1400h]
.text:0000000140017B86 movzx eax, r9b
.text:0000000140017B8A xor r8d, [rsi+rcx*4+1800h]
.text:0000000140017B92 xor r8d, [rsi+rax*4+0C00h]
.text:0000000140017B9A movzx eax, r11b
.text:0000000140017B9E mov r11d, r8d
.text:0000000140017BA1 xor r11d, [rsi+rax*4+1C00h]
.text:0000000140017BA9 sub rdi, 1
.text:0000000140017BAD jnz loc_140017B21
What does it do? No idea yet.

The next:
< ip=0x14002c4f1 count=0x4fce
---
> ip=0x14002c4f1 count=0x4463be
0x4fce is 20430, which is close to size of test1.bin (30720 bytes). 0x4463be is 4481982 which is close to
size of test2.bin (5547752 bytes). Not equal, but close.
This is a piece of code with that XOR instruction:
.text:000000014002C4EA loc_14002C4EA:
.text:000000014002C4EA movzx eax, byte ptr [r8]
.text:000000014002C4EE shl ecx, 5
.text:000000014002C4F1 xor ecx, eax
.text:000000014002C4F3 and ecx, 7FFFh
.text:000000014002C4F9 cmp [r11+rcx*4], esi
.text:000000014002C4FD jb short loc_14002C507
.text:000000014002C4FF cmp [r11+rcx*4], r10d
.text:000000014002C503 ja short loc_14002C507
.text:000000014002C505 inc ebx
973
Loop body can be written as:
state = input_byte ^ (state<<5) & 0x7FFF}.
state is then used as index in some table. Is this some kind of CRC2 ? I don’t know, but this could be a
checksumming routine. Or maybe optimized CRC routine? Any ideas?
The next block:
< ip=0x14004104a count=0x367
< ip=0x140041057 count=0x367
---
> ip=0x14004104a count=0x24193
> ip=0x140041057 count=0x24193
.text:0000000140041039 loc_140041039:
.text:0000000140041039 mov rax, r10
.text:000000014004103C add r10, 10h
.text:0000000140041040 cmp byte ptr [rcx+1], 0
.text:0000000140041044 movdqu xmm0, xmmword ptr [rax]
.text:0000000140041048 jz short loc_14004104E
.text:000000014004104A pxor xmm0, xmm1
.text:000000014004104E
.text:000000014004104E loc_14004104E:
.text:000000014004104E movdqu xmm1, xmmword ptr [rcx+18h]
.text:0000000140041053 movsxd r8, dword ptr [rcx+4]
.text:0000000140041057 pxor xmm1, xmm0
.text:000000014004105B cmp r8d, 1
.text:000000014004105F jle short loc_14004107C
.text:0000000140041061 lea rdx, [rcx+28h]
.text:0000000140041065 lea r9d, [r8-1]
.text:0000000140041069
.text:0000000140041069 loc_140041069:
.text:0000000140041069 movdqu xmm0, xmmword ptr [rdx]
.text:000000014004106D lea rdx, [rdx+10h]
.text:0000000140041071 aesenc xmm1, xmm0
.text:0000000140041076 sub r9, 1
.text:000000014004107A jnz short loc_140041069
.text:000000014004107C
This piece has both PXOR and AESENC instructions (the last is AES3 encryption instruction). So yes, we
found encryption function, RAR uses AES.
There is also another big block of almost contiguous XOR instructions:
< ip=0x140043e10 count=0x23006
---
> ip=0x140043e10 count=0x23004
499c510
< ip=0x140043e56 count=0x22ffd
---
> ip=0x140043e56 count=0x23002
But, its count is not very different during compressing/encrypting test1.bin/test2.bin. What is on these
addresses?
.text:0000000140043E07 xor ecx, r9d
.text:0000000140043E0A mov r11d, eax
.text:0000000140043E0D and ecx, r10d
.text:0000000140043E10 xor ecx, r8d
.text:0000000140043E13 rol eax, 8
.text:0000000140043E16 and eax, esi
.text:0000000140043E18 ror r11d, 8
.text:0000000140043E1C add edx, 5A827999h
.text:0000000140043E22 ror r10d, 2
.text:0000000140043E26 add r8d, 5A827999h
.text:0000000140043E2D and r11d, r12d
2 Cyclic redundancy check
3 Advanced Encryption Standard
974
.text:0000000140043E30 or r11d, eax
.text:0000000140043E33 mov eax, ebx
Let’s google 5A827999h constant... this looks like SHA-1! But why would RAR use SHA-1 during encryp-
tion?
Here is the answer:
In comparison, WinRAR uses its own key derivation scheme that requires (password length * 2 + ⤦
Ç 11)*4096 SHA-1 transformations. ’Thats why it takes longer to brute-force attack ⤦
Ç encrypted WinRAR archives.
( http://www.tomshardware.com/reviews/password-recovery-gpu,2945-8.html )
This is key scheduling: input password hashed many times and the hash is then used as AES key. This is
why we see the count of XOR instruction is almost unchanged during we switched to bigger test file.
This is it, it took couple of hours for me to write this tool and to get at least 3 points: 1) probably check-
summing; 2) AES encryption; 3) SHA-1 calculation. The first function is still unknown for me.
Still, this is impressive, because I didn’t dig into RAR code (which is proprietary, of course). I didn’t even
peek into UnRAR source code (which is available).
The files, including test files and RAR executable I’ve used (win64, 5.40):
https://beginners.re/current-tree/DBI/XOR/files/.
10.2 Cracking Minesweeper with PIN

In this book, I wrote about cracking Minesweeper for Windows XP: 8.4 on page 794.
The Minesweeper in Windows Vista and 7 is different: probably it was (re)written to C++, and a cell
information is now stored not in global array, but rather in malloc’ed heap blocks.
This is a case when we can try PIN DBI tool.
10.2.1 Intercepting all rand() calls

First, since Minesweeper places mines randomly, it has to call rand() or similar function. Let’s intercept
all rand() calls: https://beginners.re/current-tree/DBI/minesweeper/minesweeper1.cpp.
Now we can run it:
c:\pin-3.2-81205-msvc-windows\pin.exe -t minesweeper1.dll -- C:\PATH\TO\MineSweeper.exe
During startup, PIN searches for all calls to rand() function and adds a hook right after each call. The hook
is the RandAfter() function we defined: it is logging about return value and also about return address. Here
is a log I got during run of standard 9*9 configuration (10 mines): https://beginners.re/current-tree/
DBI/minesweeper/minesweeper1.out.10mines. The rand() function was called many times from several
places, but was called from 0x10002770d just 10 times. I switched Minesweeper to 16*16 configuration
(40 mines) and rand() was called from 0x10002770d 40 times. So yes, this is our point. When I load
minesweeper.exe (from Windows 7) into IDA and PDB from Microsoft website is fetched, the function
which calls rand() at 0x10002770d called Board::placeMines().
10.2.2 Replacing rand() calls with our function

Let’s now try to replace rand() function with our version, let it always return zero: https://beginners.
re/current-tree/DBI/minesweeper/minesweeper2.cpp. During startup, PIN replaces all calls to rand()
to calls to our function, which writes to log and returns zero. OK, I run it, and clicked on leftmost/topmost
cell:
975
Yes, unlike Minesweeper from Windows XP, mines are places randomly after user’s click on cell, so to
guarantee there is no mine at the cell user first clicked. So Minesweeper placed mines on cells other than
leftmost/topmost (where I clicked).
Now I clicked on rightmost/topmost cell:
This can be some kind of practical joke? I don’t know.

I clicked on 5th cell (right at the middle) at the 1st row:
This is nice, because Minesweeper can do some correct placement even with such a broken PRNG!
976
10.2.3 Peeking into placement of mines
How can we get information about where mines are placed? rand()’s result is seems to be useless: it
returned zero all the time, but Minesweeper somehow managed to place mines in different cells, though,
lined up.
This Minesweeper also written in C++ tradition, so it has no global arrays.
Let us put ourselves in the position of programmer. It has to be loop like:
for (int i; i<mines_total; i++)
{
// get coordinates using rand()
// put a cell: in other words, modify a block allocated in heap
};
How can we get information about heap block which gets modified at the 2nd step? What we need to do:
1) track all heap allocations by intercepting malloc()/realloc()/free(). 2) track all memory writes (slow). 3)
intercept calls to rand().
Now the algorithm: 1) mark all heap blocks gets modified between 1st and 2nd call to rand() from
0x10002770d; 2) whenever heap block gets freed, dump its contents.
Tracking all memory writes is slow, but after 2nd call to rand(), we don’t need to track it (since we’ve got
already a list of blocks of interest at this point), so we turn it off.
Now the code: https://beginners.re/current-tree/DBI/minesweeper/minesweeper3.cpp.
As it turns out, only 4 heap blocks gets modified between first two rand() calls, this is how they looks like:
free(0x20aa6360)
free(): we have this block in our records, size=0x28
0x20AA6360: 36 00 00 00 4E 00 00 00-2D 00 00 00 29 00 00 00 "6...N...-...)..."
0x20AA6370: 06 00 00 00 37 00 00 00-35 00 00 00 19 00 00 00 "....7...5......."
0x20AA6380: 46 00 00 00 0B 00 00 00- "F....... "
...
free(0x20af9d10)
0x20AF9D10: 0A 00 00 00 0A 00 00 00-0A 00 00 00 00 00 00 00 "................"
0x20AF9D20: 60 63 AA 20 00 00 00 00- "`c. .... "
...
free(0x20b28b20)
0x20B28B20: 02 00 00 00 03 00 00 00-04 00 00 00 05 00 00 00 "................"
0x20B28B30: 07 00 00 00 08 00 00 00-0C 00 00 00 0D 00 00 00 "................"
0x20B28B40: 0E 00 00 00 0F 00 00 00-10 00 00 00 11 00 00 00 "................"
0x20B28B50: 12 00 00 00 13 00 00 00-14 00 00 00 15 00 00 00 "................"
0x20B28B60: 16 00 00 00 17 00 00 00-18 00 00 00 1A 00 00 00 "................"
0x20B28B70: 1B 00 00 00 1C 00 00 00-1D 00 00 00 1E 00 00 00 "................"
0x20B28B80: 1F 00 00 00 20 00 00 00-21 00 00 00 22 00 00 00 ".... ...!..."..."
0x20B28B90: 23 00 00 00 24 00 00 00-25 00 00 00 26 00 00 00 "#...$...%...&..."
0x20B28BA0: 27 00 00 00 28 00 00 00-2A 00 00 00 2B 00 00 00 "'...(...*...+..."
0x20B28BB0: 2C 00 00 00 2E 00 00 00-2F 00 00 00 30 00 00 00 ",......./...0..."
0x20B28BC0: 31 00 00 00 32 00 00 00-33 00 00 00 34 00 00 00 "1...2...3...4..."
0x20B28BD0: 38 00 00 00 39 00 00 00-3A 00 00 00 3B 00 00 00 "8...9...:...;..."
0x20B28BE0: 3C 00 00 00 3D 00 00 00-3E 00 00 00 3F 00 00 00 "<...=...>...?..."
0x20B28BF0: 40 00 00 00 41 00 00 00-42 00 00 00 43 00 00 00 "@...A...B...C..."
0x20B28C00: 44 00 00 00 45 00 00 00-47 00 00 00 48 00 00 00 "D...E...G...H..."
0x20B28C10: 49 00 00 00 4A 00 00 00-4B 00 00 00 4C 00 00 00 "I...J...K...L..."
0x20B28C20: 4D 00 00 00 4F 00 00 00-50 00 00 00 50 00 00 00 "M...O...P...P..."
0x20B28C30: 50 00 00 00 50 00 00 00-50 00 00 00 50 00 00 00 "P...P...P...P..."
0x20B28C40: 50 00 00 00 50 00 00 00-50 00 00 00 50 00 00 00 "P...P...P...P..."
0x20B28C50: 50 00 00 00 00 00 00 00-00 00 00 00 00 00 00 00 "P..............."
...
free(0x20af9cf0)
977
0x20AF9CF0: 43 00 00 00 50 00 00 00-10 00 00 00 20 00 74 00 "C...P....... .t."
0x20AF9D00: 20 8B B2 20 00 00 00 00- " .. .... "
We can easily see that the biggest blocks (with size 0x28 and 0x140) are just arrays of values up to ≈
0x50. Wait... 0x50 is 80 in decimal representation. And 9*9=81 (standard minesweeper configuration).
After quick investigation, I’ve found that each 32-bit element is indeed cell coordinate. A cell is represented
using a single number, it’s a number inside of 2D-array. Row and column of each mine is decoded like
that: row=n / WIDTH; col=n % HEIGHT;
So when I tried to decode these two biggest blocks, I’ve got these cell maps:
try_to_dump_cells(). unique elements=0xa
......*..
..*......
.......*.
.........
.....*...
*.......*
**.......
.......*.
......*..
...
try_to_dump_cells(). unique elements=0x44

*.****.**
...******
*******.*
*********
*****.***
.*******.
..*******
*******.*
******.**
It seems that the first block is just a list of mines placed, while the second block is a list of free cells, but,
the second is somewhat out of sync with the first one, and it’s negative version of the first one coincides
only partially. Nevertheless, the first map is correct - we can peek into it in log file when Minesweeper is
still loaded and almost all cells are hidden, and click safely on cells marked as dots here.
So it seems, when user first clicked somewhere, Minesweeper places 10 mines, than destroys the block
with a list of it (perhaps, it copies all the data to another block before?), so we can see it during free() call.
Another fact: the method Array<NodeType>::Add(NodeType) modifies blocks we observed, and is called
from various places, including Board::placeMines(). But what is cool: I never got into its details, everything
has been resolved using just PIN.
The files: https://beginners.re/current-tree/DBI/minesweeper.
10.2.4 Exercise
Try to understand how rand()’s result being converted into coordinate(s). As a practical joke, make rand()
to output such results, so mines will be placed in shape of some symbol or figure.
10.3 Building Pin

Building Pin for Windows may be tricky. This is my working recipe.
• Unpack latest Pin to, say, C:\pin-3.7\
• Install latest Cygwin, to, say, c:\cygwin64
• Install MSVC 2015 or newer.
• Open file C:\pin-3.7\source\tools\Config\makefile.default.rules, replace mkdir -p $@ to
/bin/mkdir -p $@
• (If needed) in C:\pin-3.7\source\tools\SimpleExamples\makefile.rules, add your pintool to
the TEST_TOOL_ROOTS list.
978
• Open ”VS2015 x86 Native Tools Command Prompt”. Type:
cd c:\pin-3.7\source\tools\SimpleExamples
c:\cygwin64\bin\make all TARGET=ia32
Now pintools are in c:\pin-3.7\source\tools\SimpleExamples\obj-ia32

• For winx64, use ”x64 Native Tools Command Prompt” and run:
c:\cygwin64\bin\make all TARGET=intel64
• Run pintool:
c:\pin-3.7\pin.exe -t C:\pin-3.7\source\tools\SimpleExamples\obj-ia32\XOR_ins.dll -- ⤦
Ç program.exe arguments
10.4 Why “instrumentation”?

Perhaps, this is a term of code profiling. There are at least two methods: 1) ”sampling”: you break
into running code as many times as possible (hundreds per second), and see, where it is executed at
the moment; 2) ”instrumentation”: compiled code is interleaved with other code, which can increment
counters, etc.
Perhaps, DBI tools inherited the term?
979
Chapter 11
Other things
11.1 Executable files patching

11.1.1 x86 code
Frequent patching tasks are:
• One of the most frequent jobs is to disable some instruction. It is often done by filling it using byte
0x90 (NOP).
• Conditional jumps, which have an opcode like 74 xx (JZ), can be filled with two NOPs.
It is also possible to disable a conditional jump by writing 0 at the second byte (jump offset).
• Another frequent job is to make a conditional jump to always trigger: this can be done by writing
0xEB instead of the opcode, which stands for JMP.
• A function’s execution can be disabled by writing RETN (0xC3) at its beginning. This is true for all
functions excluding stdcall ( 6.1.2 on page 725). While patching stdcall functions, one has to
determine the number of arguments (for example, by finding RETN in this function), and use RETN
with a 16-bit argument (0xC2).
• Sometimes, a disabled functions has to return 0 or 1. This can be done by MOV EAX, 0 or MOV EAX,
1, but it’s slightly verbose.
A better way is XOR EAX, EAX (2 bytes 0x31 0xC0) or XOR EAX, EAX / INC EAX (3 bytes 0x31 0xC0
0x40).
A software may be protected against modifications.
This protection is often done by reading the executable code and calculating a checksum. Therefore, the
code must be read before protection is triggered.
This can be determined by setting a breakpoint on reading memory.
tracer has the BPM option for this.
PE executable file relocs ( 6.5.2 on page 751) must not to be touched while patching, because the Windows
loader may overwrite your new code. (They are grayed in Hiew, for example: fig.1.22).
As a last resort, it is possible to write jumps that circumvent the relocs, or you will have to edit the relocs
table.
11.2 Function arguments number statistics

I’ve always been interesting in what is average number of function arguments.
I’ve analyzed many Windows 7 32-bit DLLs
(crypt32.dll, mfc71.dll, msvcr100.dll, shell32.dll, user32.dll, d3d11.dll, mshtml.dll, msxml6.dll, sqlncli11.dll,
wininet.dll, mfc120.dll, msvbvm60.dll, ole32.dll, themeui.dll, wmp.dll) (because they use stdcall conven-
tion, and so it is easy to grep disassembly output just by RETN X).
• no arguments: ≈ 29%
• 1 argument: ≈ 23%
980
• 2 arguments: ≈ 20%
Figure 11.1: Function arguments number statistics
This is heavily dependent on programming style and may be very different for other software products.
11.3 Compiler intrinsic

A function specific to a compiler which is not an usual library function. The compiler generates a specific
machine code instead of a call to it. It is often a pseudofunction for specific CPU instruction.
For example, there are no cyclic shift operations in C/C++ languages, but they are present in most CPUs.
1
For programmer’s convenience, at least MSVC has pseudofunctions _rotl() and _rotr() which are trans-
lated by the compiler directly to the ROL/ROR x86 instructions.
Another example are functions to generate SSE-instructions right in the code.

Full list of MSVC intrinsics: MSDN.
11.4 Compiler’s anomalies

11.4.1 Oracle RDBMS 11.2 and Intel C++ 10.1
Intel C++ 10.1, which was used for Oracle RDBMS 11.2 Linux86 compilation, may emit two JZ in row, and
there are no references to the second JZ. The second JZ is thus meaningless.
Listing 11.1: kdli.o from libserver11.a

.text:08114CF1 loc_8114CF1: ; CODE XREF: __PGOSF539_kdlimemSer+89A
.text:08114CF1 ; __PGOSF539_kdlimemSer+3994
.text:08114CF1 8B 45 08 mov eax, [ebp+arg_0]
1
MSDN
981
.text:08114CF4 0F B6 50 14 movzx edx, byte ptr [eax+14h]
.text:08114CF8 F6 C2 01 test dl, 1
.text:08114CFB 0F 85 17 08 00 00 jnz loc_8115518
.text:08114D01 85 C9 test ecx, ecx
.text:08114D03 0F 84 8A 00 00 00 jz loc_8114D93
.text:08114D09 0F 84 09 08 00 00 jz loc_8115518
.text:08114D0F 8B 53 08 mov edx, [ebx+8]
.text:08114D12 89 55 FC mov [ebp+var_4], edx
.text:08114D15 31 C0 xor eax, eax
.text:08114D17 89 45 F4 mov [ebp+var_C], eax
.text:08114D1A 50 push eax
.text:08114D1B 52 push edx
.text:08114D1C E8 03 54 00 00 call len2nbytes
.text:08114D21 83 C4 08 add esp, 8
Listing 11.2: from the same code

.text:0811A2A5 loc_811A2A5: ; CODE XREF: kdliSerLengths+11C
.text:0811A2A5 ; kdliSerLengths+1C1
.text:0811A2A5 8B 7D 08 mov edi, [ebp+arg_0]
.text:0811A2A8 8B 7F 10 mov edi, [edi+10h]
.text:0811A2AB 0F B6 57 14 movzx edx, byte ptr [edi+14h]
.text:0811A2AF F6 C2 01 test dl, 1
.text:0811A2B2 75 3E jnz short loc_811A2F2
.text:0811A2B4 83 E0 01 and eax, 1
.text:0811A2B7 74 1F jz short loc_811A2D8
.text:0811A2B9 74 37 jz short loc_811A2F2
.text:0811A2BB 6A 00 push 0
.text:0811A2BD FF 71 08 push dword ptr [ecx+8]
.text:0811A2C0 E8 5F FE FF FF call len2nbytes
It is supposedly a code generator bug that was not found by tests, because resulting code works correctly
anyway.
Another example from Oracle RDBMS 11.1.0.6.0 for win32.
.text:0051FBF8 85 C0 test eax, eax
.text:0051FBFA 0F 84 8F 00 00 00 jz loc_51FC8F
.text:0051FC00 74 1D jz short loc_51FC1F
11.4.2 MSVC 6.0

Just found in some old code:
fabs
fild [esp+50h+var_34]
fabs
fxch st(1) ; first instruction
fxch st(1) ; second instruction
faddp st(1), st
fcomp [esp+50h+var_3C]
fnstsw ax
test ah, 41h
jz short loc_100040B7
The first FXCH instruction swaps ST(0) and ST(1), the second do the same, so both do nothing. This is a
program uses MFC42.dll, so it could be MSVC 6.0, 5.0 or maybe even MSVC 4.2 from 1990s.
This pair do nothing, so it probably wasn’t caught by MSVC compiler tests. Or maybe I wrong?
11.4.3 ftol2() in MSVC 2012

Just found this in ftol2() standard C/C++ library function (float-to-long conversion routine) in Microsoft
Visual Studio 2012.
.text:00000036 public __ftol2
.text:00000036 __ftol2 proc near ; CODE XREF: $$000000+7
.text:00000036 ; __ftol2_sse_excpt+7
982
.text:00000039 sub esp, 20h
.text:0000003C and esp, 0FFFFFFF0h
.text:0000003F fld st
.text:00000041 fst dword ptr [esp+18h]
.text:00000045 fistp qword ptr [esp+10h]
.text:00000049 fild qword ptr [esp+10h]
.text:0000004D mov edx, [esp+18h]
.text:00000051 mov eax, [esp+10h]
.text:00000055 test eax, eax
.text:00000057 jz short integer_QnaN_or_zero
.text:00000059
.text:00000059 arg_is_not_integer_QnaN: ; CODE XREF: __ftol2+69
.text:00000059 fsubp st(1), st
.text:0000005B test edx, edx
.text:0000005D jns short positive
.text:0000005F fstp dword ptr [esp]
.text:00000062 mov ecx, [esp]
.text:00000065 xor ecx, 80000000h
.text:0000006B add ecx, 7FFFFFFFh
.text:00000071 adc eax, 0
.text:00000074 mov edx, [esp+14h]
.text:00000078 adc edx, 0
.text:0000007B jmp short localexit
.text:0000007D ; ---------------------------------------------------------------------------
.text:0000007D
.text:0000007D positive: ; CODE XREF: __ftol2+27
.text:0000007D fstp dword ptr [esp]
.text:00000080 mov ecx, [esp]
.text:00000083 add ecx, 7FFFFFFFh
.text:00000089 sbb eax, 0
.text:0000008C mov edx, [esp+14h]
.text:00000090 sbb edx, 0
.text:00000093 jmp short localexit
.text:00000095 ; ---------------------------------------------------------------------------
.text:00000095
.text:00000095 integer_QnaN_or_zero: ; CODE XREF: __ftol2+21
.text:00000095 mov edx, [esp+14h]
.text:00000099 test edx, 7FFFFFFFh
.text:0000009F jnz short arg_is_not_integer_QnaN
.text:000000A1 fstp dword ptr [esp+18h] ; first
.text:000000A5 fstp dword ptr [esp+18h] ; second
.text:000000A9
.text:000000A9 localexit: ; CODE XREF: __ftol2+45
.text:000000A9 ; __ftol2+5D
.text:000000A9 leave
.text:000000AA retn
.text:000000AA __ftol2 endp
Note two identical FSTP-s (float store with pop) at the end. First I thought it was compiler anomaly
(I’m collecting such cases just as someone do with butterflies), but it seems, it’s handwritten assem-
bler piece, in msvcrt.lib there is an object file with this function in it, and we can find this string in it:
f:\dd\vctools\crt_bld\SELF_X86\crt\prebuild\tran\i386\ftol2.asm — that was probably a path
to the file on developer’s computer where msvcrt.lib was built.
So, bug, text editor-induced typo, or feature?
11.4.4 Summary
Other compiler anomalies here in this book: 1.28.2 on page 313, 3.10.3 on page 493, 3.18.7 on page 531, 1.26.7
on page 299, 1.18.4 on page 146, 1.28.5 on page 330.
Such cases are demonstrated here in this book, to show that such compilers errors are possible and
sometimes one should not to rack one’s brain while thinking why did the compiler generate such strange
code.
983
11.5 Itanium
Although almost failed, Intel Itanium (IA64) is a very interesting architecture.
While OOE CPUs decides how to rearrange their instructions and execute them in parallel, EPIC2 was an
attempt to shift these decisions to the compiler: to let it group the instructions at the compile stage.
This resulted in notoriously complex compilers.
Here is one sample of IA64 code: simple cryptographic algorithm from the Linux kernel:
Listing 11.3: Linux kernel 3.2.0.4

#define TEA_ROUNDS 32
#define TEA_DELTA 0x9e3779b9
static void tea_encrypt(struct crypto_tfm *tfm, u8 *dst, const u8 *src)

{
u32 y, z, n, sum = 0;
u32 k0, k1, k2, k3;
struct tea_ctx *ctx = crypto_tfm_ctx(tfm);
const __le32 *in = (const __le32 *)src;
__le32 *out = (__le32 *)dst;
y = le32_to_cpu(in[0]);
z = le32_to_cpu(in[1]);
k0 = ctx->KEY[0];
k1 = ctx->KEY[1];
k2 = ctx->KEY[2];
k3 = ctx->KEY[3];
n = TEA_ROUNDS;
while (n-- > 0) {

sum += TEA_DELTA;
y += ((z << 4) + k0) ^ (z + sum) ^ ((z >> 5) + k1);
z += ((y << 4) + k2) ^ (y + sum) ^ ((y >> 5) + k3);
}
out[0] = cpu_to_le32(y);
out[1] = cpu_to_le32(z);
}
Here is how it was compiled:
Listing 11.4: Linux Kernel 3.2.0.4 for Itanium 2 (McKinley)

0090| tea_encrypt:
0090|08 80 80 41 00 21 adds r16 = 96, r32 // ptr to ctx->KEY[2]
0096|80 C0 82 00 42 00 adds r8 = 88, r32 // ptr to ctx->KEY[0]
009C|00 00 04 00 nop.i 0
00A0|09 18 70 41 00 21 adds r3 = 92, r32 // ptr to ctx->KEY[1]
00A6|F0 20 88 20 28 00 ld4 r15 = [r34], 4 // load z
00AC|44 06 01 84 adds r32 = 100, r32;; // ptr to ctx->KEY[3]
00B0|08 98 00 20 10 10 ld4 r19 = [r16] // r19=k2
00B6|00 01 00 00 42 40 mov r16 = r0 // r0 always contain zero
00BC|00 08 CA 00 mov.i r2 = ar.lc // save lc register
00C0|05 70 00 44 10 10
9E FF FF FF 7F 20 ld4 r14 = [r34] // load y
00CC|92 F3 CE 6B movl r17 = 0xFFFFFFFF9E3779B9;; // TEA_DELTA
00D0|08 00 00 00 01 00 nop.m 0
00D6|50 01 20 20 20 00 ld4 r21 = [r8] // r21=k0
00DC|F0 09 2A 00 mov.i ar.lc = 31 // TEA_ROUNDS is 32
00E0|0A A0 00 06 10 10 ld4 r20 = [r3];; // r20=k1
00E6|20 01 80 20 20 00 ld4 r18 = [r32] // r18=k3
00EC|00 00 04 00 nop.i 0
00F0|
00F0| loc_F0:
00F0|09 80 40 22 00 20 add r16 = r16, r17 // r16=sum, r17=TEA_DELTA
2 Explicitly Parallel Instruction Computing
984
00F6|D0 71 54 26 40 80 shladd r29 = r14, 4, r21 // r14=y, r21=k0
00FC|A3 70 68 52 extr.u r28 = r14, 5, 27;;
0100|03 F0 40 1C 00 20 add r30 = r16, r14
0106|B0 E1 50 00 40 40 add r27 = r28, r20;; // r20=k1
010C|D3 F1 3C 80 xor r26 = r29, r30;;
0110|0B C8 6C 34 0F 20 xor r25 = r27, r26;;
0116|F0 78 64 00 40 00 add r15 = r15, r25 // r15=z
011C|00 00 04 00 nop.i 0;;
0120|00 00 00 00 01 00 nop.m 0
0126|80 51 3C 34 29 60 extr.u r24 = r15, 5, 27
012C|F1 98 4C 80 shladd r11 = r15, 4, r19 // r19=k2
0130|0B B8 3C 20 00 20 add r23 = r15, r16;;
0136|A0 C0 48 00 40 00 add r10 = r24, r18 // r18=k3
013C|00 00 04 00 nop.i 0;;
0140|0B 48 28 16 0F 20 xor r9 = r10, r11;;
0146|60 B9 24 1E 40 00 xor r22 = r23, r9
014C|00 00 04 00 nop.i 0;;
0150|11 00 00 00 01 00 nop.m 0
0156|E0 70 58 00 40 A0 add r14 = r14, r22
015C|A0 FF FF 48 br.cloop.sptk.few loc_F0;;
0160|09 20 3C 42 90 15 st4 [r33] = r15, 4 // store z
0166|00 00 00 02 00 00 nop.m 0
016C|20 08 AA 00 mov.i ar.lc = r2;; // restore lc register
0170|11 00 38 42 90 11 st4 [r33] = r14 // store y
0176|00 00 00 02 00 80 nop.i 0
017C|08 00 84 00 br.ret.sptk.many b0;;
First of all, all IA64 instructions are grouped into 3-instruction bundles.
Each bundle has a size of 16 bytes (128 bits) and consists of template code (5 bits) + 3 instructions (41
bits for each).
IDA shows the bundles as 6+6+4 bytes —you can easily spot the pattern.
All 3 instructions from each bundle usually executes simultaneously, unless one of instructions has a “stop
bit”.
Supposedly, Intel and HP engineers gathered statistics on most frequent instruction patterns and decided
to bring bundle types (AKA “templates”): a bundle code defines the instruction types in the bundle. There
are 12 of them.
For example, the zeroth bundle type is MII, which implies the first instruction is Memory (load or store),
the second and third ones are I (integer instructions).
Another example is the bundle of type 0x1d: MFB: the first instruction is Memory (load or store), the second
one is Float (FPU instruction), and the third is Branch (branch instruction).
If the compiler cannot pick a suitable instruction for the relevant bundle slot, it may insert a NOP: you
can see here the nop.i instructions (NOP at the place where the integer instruction might be) or nop.m
(a memory instruction might be at this slot).
NOPs are inserted automatically when one uses assembly language manually.
And that is not all. Bundles are also grouped.
Each bundle may have a “stop bit”, so all the consecutive bundles with a terminating bundle which has
the “stop bit” can be executed simultaneously.
In practice, Itanium 2 can execute 2 bundles at once, resulting in the execution of 6 instructions at once.
So all instructions inside a bundle and a bundle group cannot interfere with each other (i.e., must not have
data hazards).
If they do, the results are to be undefined.
Each stop bit is marked in assembly language as two semicolons (;;) after the instruction.
So, the instructions at [90-ac] may be executed simultaneously: they do not interfere. The next group is
[b0-cc].
We also see a stop bit at 10c. The next instruction at 110 has a stop bit too.
This implies that these instructions must be executed isolated from all others (as in CISC).
985
Indeed: the next instruction at 110 uses the result from the previous one (the value in register r26), so
they cannot be executed at the same time.
Apparently, the compiler was not able to find a better way to parallelize the instructions, in other words,
to load CPU as much as possible, hence too much stop bits and NOPs.
Manual assembly programming is a tedious job as well: the programmer has to group the instructions
manually.
The programmer is still able to add stop bits to each instructions, but this will degrade the performance
that Itanium was made for.
An interesting examples of manual IA64 assembly code can be found in the Linux kernel’s sources:
http://go.yurichev.com/17322.
Another introductory paper on Itanium assembly: [Mike Burrell, Writing Efficient Itanium 2 Assembly Code
(2010)]3 , [papasutra of haquebright, WRITING SHELLCODE FOR IA-64 (2001)]4 .
Another very interesting Itanium feature is the speculative execution and the NaT (“not a thing”) bit,
somewhat resembling NaN numbers:
MSDN.
11.6 8086 memory model

When dealing with 16-bit programs for MS-DOS or Win16 ( 8.8.3 on page 835 or 3.34.5 on page 648), we
can see that the pointers consist of two 16-bit values. What do they mean? Oh yes, that is another weird
MS-DOS and 8086 artifact.
8086/8088 was a 16-bit CPU, but was able to address 20-bit address in RAM (thus being able to access
1MB of external memory).
The external memory address space was divided between RAM (640KB max), ROM, windows for video
memory, EMS cards, etc.
Let’s also recall that 8086/8088 was in fact an inheritor of the 8-bit 8080 CPU.
The 8080 has a 16-bit memory space, i.e., it was able to address only 64KB.
And probably because of reason of old software porting5 , 8086 can support many 64KB windows simulta-
neously, placed within the 1MB address space.
This is some kind of a toy-level virtualization.
All 8086 registers are 16-bit, so to address more, special segment registers (CS, DS, ES, SS) were intro-
duced.
Each 20-bit pointer is calculated using the values from a segment register and an address register pair
(e.g. DS:BX) as follows:
real_address = (segment_register ≪ 4) + address_register
For example, the graphics (EGA6 , VGA7 ) video RAM window on old IBM PC-compatibles has a size of 64KB.
To access it, a value of 0xA000 has to be stored in one of the segment registers, e.g. into DS.
Then DS:0 will address the first byte of video RAM and DS:0xFFFF — the last byte of RAM.
The real address on the 20-bit address bus, however, will range from 0xA0000 to 0xAFFFF.
The program may contain hard-coded addresses like 0x1234, but the OS may need to load the program
at arbitrary addresses, so it recalculates the segment register values in a way that the program does not
have to care where it’s placed in the RAM.
So, any pointer in the old MS-DOS environment in fact consisted of the segment address and the address
inside segment, i.e., two 16-bit values. 20-bit was enough for that, though, but we needed to recalculate
the addresses very often: passing more information on the stack seemed a better space/convenience
balance.
3 Alsoavailable as http://yurichev.com/mirrors/RE/itanium.pdf
4 Alsoavailable as http://phrack.org/issues/57/5.html
5 The author is not 100% sure here
6 Enhanced Graphics Adapter
7 Video Graphics Array
986
By the way, because of all this it was not possible to allocate a memory block larger than 64KB.
The segment registers were reused at 80286 as selectors, serving a different function.
When the 80386 CPU and computers with bigger RAM were introduced, MS-DOS was still popular, so the
DOS extenders emerged: these were in fact a step toward a “serious” OS, switching the CPU in protected
mode and providing much better memory APIs for the programs which still needed to run under MS-DOS.
Widely popular examples include DOS/4GW (the DOOM video game was compiled for it), Phar Lap, PMODE.
By the way, the same way of addressing memory was used in the 16-bit line of Windows 3.x, before Win32.
11.7 Basic blocks reordering

11.7.1 Profile-guided optimization
This optimization method can move some basic blocks to another section of the executable binary file.
Obviously, there are parts of a function which are executed more frequently (e.g., loop bodies) and less
often (e.g., error reporting code, exception handlers).
The compiler adds instrumentation code into the executable, then the developer runs it with a lot of tests
to collect statistics.
Then the compiler, with the help of the statistics gathered, prepares final the executable file with all
infrequently executed code moved into another section.
As a result, all frequently executed function code is compacted, and that is very important for execution
speed and cache usage.
An example from Oracle RDBMS code, which was compiled with Intel C++:
Listing 11.5: orageneric11.dll (win32)

public _skgfsync
_skgfsync proc near
; address 0x6030D86A
db 66h
nop
push ebp
mov ebp, esp
mov edx, [ebp+0Ch]
test edx, edx
mov eax, [edx+30h]
test eax, 400h
jnz __VInfreq__skgfsync ; write to log
continue:
mov eax, [ebp+8]
mov edx, [ebp+10h]
lea eax, [edx+0Fh]
and eax, 0FFFFFFFCh
mov ecx, [eax]
cmp ecx, 45726963h
jnz error ; exit with error
mov esp, ebp
pop ebp
retn
_skgfsync endp
...
; address 0x60B953F0
__VInfreq__skgfsync:
mov eax, [edx]
test eax, eax
jz continue
mov ecx, [ebp+10h]
987
push ecx
mov ecx, [ebp+8]
push edx
push ecx
push offset ... ; "skgfsync(se=0x%x, ctx=0x%x, iov=0x%x)\n"
push dword ptr [edx+4]
call dword ptr [eax] ; write to log
add esp, 14h
jmp continue
error:
mov edx, [ebp+8]
mov dword ptr [edx], 69AAh ; 27050 "function called with invalid FIB/IOV
structure"
mov eax, [eax]
mov [edx+4], eax
mov dword ptr [edx+8], 0FA4h ; 4004
mov esp, ebp
pop ebp
retn
; END OF FUNCTION CHUNK FOR _skgfsync
The distance of addresses between these two code fragments is almost 9 MB.
All infrequently executed code was placed at the end of the code section of the DLL file, among all function
parts.
This part of the function was marked by the Intel C++ compiler with the VInfreq prefix.
Here we see that a part of the function that writes to a log file (presumably in case of error or warning
or something like that) which was probably not executed very often when Oracle’s developers gathered
statistics (if it was executed at all).
The writing to log basic block eventually returns the control flow to the “hot” part of the function.
Another “infrequent” part is the basic block returning error code 27050.
In Linux ELF files, all infrequently executed code is moved by Intel C++ into the separate text.unlikely
section, leaving all “hot” code in the text.hot section.
From a reverse engineer’s perspective, this information may help to split the function into its core and
error handling parts.
11.8 My experience with Hex-Rays 2.2.0

11.8.1 Bugs
There are couple of bugs.
First of all, Hex-Rays is getting lost when FPU instructions are interleaved (by compiler codegenerator)
with others.
For example, this:
f proc near
lea eax, [esp+4]

fild dword ptr [eax]
lea eax, [esp+8]
fild dword ptr [eax]
fabs
fcompp
fnstsw ax
test ah, 1
jz l01
mov eax, 1
retn
l01:
mov eax, 2
retn
988
f endp
…will be correcly decompiled to:

signed int __cdecl f(signed int a1, signed int a2)
{
signed int result; // eax@2
if ( fabs((double)a2) >= (double)a1 )

result = 2;
else
result = 1;
return result;
}
But let’s comment one of the instructions at the end:

...
l01:
;mov eax, 2
retn
...
…we getting an obvious bug:

void __cdecl f(char a1, char a2)
{
fabs((double)a2);
}
This is another bug:

extrn f1:dword
extrn f2:dword
f proc near
fld dword ptr [esp+4]

fadd dword ptr [esp+8]
fst dword ptr [esp+12]
fcomp ds:const_100
fld dword ptr [esp+16] ; comment this instruction and it will be OK
fnstsw ax
test ah, 1
jnz short l01
call f1
retn
l01:
call f2
retn
f endp
...
const_100 dd 42C80000h ; 100.0
Result:
int __cdecl f(float a1, float a2, float a3, float a4)
{
double v5; // st7@1
char v6; // c0@1
int result; // eax@2
v5 = a4;
if ( v6 )
989
result = f2(v5);
else
result = f1(v5);
return result;
}
v6 variable has char type and if you’ll try to compile this code, compiler will warn you about variable
usage before assignment.
Another bug: FPATAN instruction is correctly decompiled into atan2(), but arguments are swapped.
11.8.2 Odd peculiarities

Hex-Rays too often promotes 32-bit int to 64-bit one. Here is example:
f proc near
mov eax, [esp+4]

cdq
xor eax, edx
sub eax, edx
; EAX=abs(a1)
sub eax, [esp+8]

; EAX=EAX-a2
; EAX at this point somehow gets promoted to 64-bit (RAX)
cdq
xor eax, edx
sub eax, edx
; EAX=abs(abs(a1)-a2)
retn
f endp
Result:
int __cdecl f(int a1, int a2)
{
__int64 v2; // rax@1
v2 = abs(a1) - a2;
return (HIDWORD(v2) ^ v2) - HIDWORD(v2);
}
Perhaps, this is result of CDQ instruction? I’m not sure. Anyway, whenever you see __int64 type in 32-bit
code, pay attention.
This is also weird:
f proc near
mov esi, [esp+4]
lea ebx, [esi+10h]

cmp esi, ebx
jge short l00
cmp esi, 1000

jg short l00
mov eax, 2
retn
l00:
mov eax, 1
retn
990
f endp
Result:
signed int __cdecl f(signed int a1)
{
if ( __OFSUB__(a1, a1 + 16) ^ 1 && a1 <= 1000 )

result = 2;
else
result = 1;
return result;
}
The code is correct, but needs manual intervention.

Sometimes, Hex-Rays doesn’t fold (or reduce) division by multiplication code:
f proc near
mov eax, [esp+4]

mov edx, 2AAAAAABh
imul edx
mov eax, edx
retn
f endp
Result:
int __cdecl f(int a1)
{
return (unsigned __int64)(715827883i64 * a1) >> 32;
}
This can be folded (rewritten) manually.

Many of these peculiarities can be solved by manual reordering of instructions, recompiling assembly
code, and then feeding it to Hex-Rays again.
11.8.3 Silence
extrn some_func:dword
f proc near
mov ecx, [esp+4]

mov eax, [esp+8]
push eax
call some_func
add esp, 4
; use ECX
mov eax, ecx
retn
f endp
Result:
int __cdecl f(int a1, int a2)
{
int v2; // ecx@1
some_func(a2);
return v2;
}
991
v2 variable (from ECX) is lost …Yes, this code is incorrect (ECX value doesn’t saved during call to another
function), but it would be good for Hex-Rays to give a warning.
Another one:
extrn some_func:dword
f proc near
call some_func
jnz l01
mov eax, 1
retn
l01:
mov eax, 2
retn
f endp
Result:
signed int f()
{
char v0; // zf@1
some_func();
if ( v0 )
result = 1;
else
result = 2;
return result;
}
Again, warning would be great.

Anyway, whenever you see variable of char type, or variable which is used without initialization, this is
clear sign that something went wrong and needs manual intervention.
11.8.4 Comma
Comma in C/C++ has a bad fame, because it can lead to a confusing code.
Quick quiz, what does this C/C++ function return?
int f()
{
return 1, 2;
};
It’s 2: when compiler encounters comma-expression, it generates code which executes all sub-expressions,
and returns value of the last sub-expression.
I’ve seen something like that in production code:
if (cond)
return global_var=123, 456; // 456 is returned
else
return global_var=789, 321; // 321 is returned
Apparently, programmer wanted to make code slightly shorter without additional curly brackets. In other
words, comma allows to pack couple of expressions into one, without forming statement/code block inside
of curly brackets.
Comma in C/C++ is close to begin in Scheme/Racket: https://docs.racket-lang.org/guide/begin.
html.
Perhaps, the only widely accepted usage of comma is in for() statements:
992
char *s="hello, world";
for(int i=0; *s; s++, i++);
// i = string length
Both s++ and i++ are executed at each loop iteration.

Read more: https://stackoverflow.com/q/52550.
I’m writing all this because Hex-Rays produces (at least in my case) code which is rich with both commas
and short-circuit expressions. For example, this is real output from Hex-Rays:
if ( a >= b || (c = a, (d[a] - e) >> 2 > f) )
{
...
This is correct, it compiles and works, and let god help you to understand it. Here is it rewritten:
if (cond1 || (comma_expr, cond2))
{
...
Short-circuit is effective here: first cond1 is checked, if it’s true, if() body is executed, the rest of if()
expression is ignored completely. If cond1 is false, comma_expr is executed (in the previous example,
a gets copied to c), then cond2 is checked. If cond2 is true, if() body gets executed, or not. In other
words, if() body gets executed if cond1 is true or cond2 is true, but if the latter is true, comma_expr is
also executed.
Now you can see why comma is so notorious.
A word about short-circuit. A common beginner’s misconception is that sub-conditions are checked in
some unspecified order, which is not true. In a | b | c expression, a, b and c gets evaluated in unspecified
order, so that is why || has also been added to C/C++, to apply short-circuit explicitly.
11.8.5 Data types

Data types is a problem for decompilers.
Hex-Rays can be blind to arrays in local stack, if they weren’t set correctly before decompilation. Same
story about global arrays.
Another problem is too big functions, where a single slot in local stack can be used by several variables
across function’s execution. It’s not a rare case when a slot is used for int-variable, then for pointer, then
for float-variable. Hex-Rays correctly decompiles it: it creates a variable with some type, then cast it to
another type in various parts of functions. This problem has been solved by me by manual splitting big
function into several smaller. Just make local variables as global ones, etc, etc. And don’t forget about
tests.
11.8.6 Long and messed expressions

Sometimes, during rewriting, you can end up with long and hard to understand expressions in if() con-
structs, like:
if ((! (v38 && v30 <= 5 && v27 != -1)) && ((! (v38 && v30 <= 5) && v27 != -1) || (v24 >= 5 || ⤦
Ç v26)) && v25)
{
...
}
Wolfram Mathematica can minimize some of them, using BooleanMinimize[] function:

In[1]:= BooleanMinimize[(! (v38 && v30 <= 5 && v27 != -1)) && v38 && v30 <= 5 && v25 == 0]
Out[1]:= v38 && v25 == 0 && v27 == -1 && v30 <= 5
There is even better way, to find common subexpressions:
993
In[2]:= ExperimentalÒptimizeExpression[(! (v38 && v30 <= 5 &&
v27 != -1)) && ((! (v38 && v30 <= 5) &&
v27 != -1) || (v24 >= 5 || v26)) && v25]
Out[2]= ExperimentalÒptimizedExpression[
Block[{Compile`$1, Compile`$2}, Compile`$1 = v30 <= 5;
Compile`$2 =
v27 != -1; ! (v38 && Compile`$1 &&
Compile`$2) && ((! (v38 && Compile`$1) && Compile`$2) ||
v24 >= 5 || v26) && v25]]
Mathematica has added two new variables: Compile`$1 and Compile`$2, values of which are to be used
several times in expression. So we can add two additional variables.
11.8.7 De Morgan’s laws and decompilation

Sometimes a compiler’s optimizer can use De Morgan’s laws to make code shorter/faster.
For example, this:
void f(int a, int b, int c, int d)
{
if (a>0 && b>0)
printf ("both a and b are positive\n");
else if (c>0 && d>0)
printf ("both c and d are positive\n");
else
printf ("something else\n");
};
... looks pretty innocent, when compiled by optimizing GCC 5.4.0 x64:
; int __fastcall f(int a, int b, int c, int d)
public f
f proc near
test edi, edi
jle short loc_8
test esi, esi
jg short loc_30
loc_8:
test edx, edx
jle short loc_20
test ecx, ecx
jle short loc_20
mov edi, offset s ; "both c and d are positive"
jmp puts
loc_20:
mov edi, offset aSomethingElse ; "something else"
jmp puts
loc_30:
mov edi, offset aAAndBPositive ; "both a and b are positive"
loc_35:
jmp puts
f endp
... also looks innocent, but Hex-Rays 2.2.0 cannot clearly see that both AND operations were actually used
in the source code:
int __fastcall f(int a, int b, int c, int d)
{
int result;
if ( a > 0 && b > 0 )

{
result = puts("both a and b are positive");
994
}
else if ( c <= 0 || d <= 0 )
{
result = puts("something else");
}
else
{
result = puts("both c and d are positive");
}
return result;
}
The c <= 0 || d <= 0 expression is inversion of c>0 && d>0 since A ∪ B = A∩B and A ∩ B = A∪B, in other
words, !(cond1 || cond2) == !cond1 && !cond2 and !(cond1 && cond2) == !cond1 || !cond2.
These rules are worth to be kept in mind, since this compiler optimization is used heavily almost every-
where.
Sometimes it’s good idea to invert a condition, in order to understand a code better. This is a piece of a
real code decompiled by Hex-Rays:
for (int i=0; i<12; i++)
{
if (v1[i-12] != 0.0 || v1[i] != 0.0)
{
v108=min(v108, (float)v0[i*24 -2]);
v113=max(v113, (float)v0[i*24]);
};
}
... it can be rewritten like:

for (int i=0; i<12; i++)
{
if (v1[i-12] == 0.0 && v1[i] == 0.0)
continue;
v108=min(v108, (float)v0[i*24 -2]);

v113=max(v113, (float)v0[i*24]);
}
Which is better? I don’t know yet, but for better understanding, it’s great to take a look on both.
11.8.8 My plan
• Split big functions (and don’t forget about tests). Sometimes it’s very helpful to form new functions
out of big loop bodies.
• Check/set data type of variables, arrays, etc.
• If you see odd result, dangling variable (which used before initialization), try to swap instructions
manually, recompile it and feed to Hex-Rays again.
11.8.9 Summary
Nevertheless, quality of Hex-Rays 2.2.0 is very, very good. It makes life way easier.
11.9 Cyclomatic complexity

The term is used to measure complexity of a function. Complex functions are usually evil, because they
are hard to maintain, hard to test, etc.
There are several heuristics to measure it.
For example, we can find in Linux kernel coding style8 :
8 https://www.kernel.org/doc/html/v4.10/process/coding-style.html
995
Now, some people will claim that having 8-character indentations makes the code move
too far to the right, and makes it hard to read on a 80-character terminal screen. The answer
to that is that if you need more than 3 levels of indentation, you’re screwed anyway, and
should fix your program.
...
Functions should be short and sweet, and do just one thing. They should fit on one or
two screenfuls of text (the ISO/ANSI screen size is 80x24, as we all know), and do one thing
and do that well.
The maximum length of a function is inversely proportional to the complexity and in-
dentation level of that function. So, if you have a conceptually simple function that is just
one long (but simple) case-statement, where you have to do lots of small things for a lot of
different cases, it’s OK to have a longer function.
However, if you have a complex function, and you suspect that a less-than-gifted first-
year high-school student might not even understand what the function is all about, you
should adhere to the maximum limits all the more closely. Use helper functions with descrip-
tive names (you can ask the compiler to in-line them if you think it’s performance-critical,
and it will probably do a better job of it than you would have done).
Another measure of the function is the number of local variables. They shouldn’t exceed
5-10, or you’re doing something wrong. Re-think the function, and split it into smaller pieces.
A human brain can generally easily keep track of about 7 different things, anything more
and it gets confused. You know you’re brilliant, but maybe you’d like to understand what
you did 2 weeks from now.
In JPL Institutional Coding Standard for the C Programming Language 9 :
Functions should be no longer than 60 lines of text and define no more than 6 parameters.
A function should not be longer than what can be printed on a single sheet of paper
in a standard reference format with one line per statement and one line per declaration.
Typically, this means no more than about 60 lines of code per function. Long lists of function
parameters similarly compromise code clarity and should be avoided.
Each function should be a logical unit in the code that is understandable and verifiable
as a unit. It is much harder to understand a logical unit that spans multiple screens on a
computer display or multiple pages when printed. Excessively long functions are often a
sign of poorly structured code.
Now let’s back to cyclomatic complexity.

Without diving deep into graph theory: there are basic blocks and links between them. For example, this
is how IDA shows BB10 s and links (as arrows). Just click space and you’ll see this: 1.18 on page 88. Each
BB is also called vertex or node in graph theory. Each link - edge.
There are at least two popular ways to calculate cyclomatic complexity: 1) edges - nodes + 2 2) edges -
nodes + number of exits (RET instructions)
As of IDA example below, there are 4 BBs, so that is 4 nodes. But there are also 4 links and 1 return
instruction. By 1st rule, this is 2, by the second: 1.
The bigger the number, the more complex your function and things go from bad to worse. As you can see,
additional exit (return instructions) make things even worse, as well as additional links between nodes
(including additional goto’s).
I wrote the simple IDAPython script (https://beginners.re/current-tree/other/cyclomatic/cyclomatic.
py) to measure it. Here is result for Linux kernel 4.11 (most complex functions in it):
1829c0 do_check edges=937 nodes=574 rets=1 E-N+2=365 E-N+rets=364
2effe0 ext4_fill_super edges=862 nodes=568 rets=1 E-N+2=296 E-N+rets=295
5d92e0 wm5110_readable_register edges=661 nodes=369 rets=2 E-N+2=294 E-N+rets=294
277650 do_blockdev_direct_IO edges=771 nodes=507 rets=1 E-N+2=266 E-N+rets=265
10f7c0 load_module edges=711 nodes=465 rets=1 E-N+2=248 E-N+rets=247
787730 dev_ethtool edges=559 nodes=315 rets=1 E-N+2=246 E-N+rets=245
84e440 do_ipv6_setsockopt edges=468 nodes=237 rets=1 E-N+2=233 E-N+rets=232
72c3c0 mmc_init_card edges=593 nodes=365 rets=1 E-N+2=230 E-N+rets=229
...
9 https://yurichev.com/mirrors/C/JPL_Coding_Standard_C.pdf
10 Basic Block
996
( Full list: https://beginners.re/current-tree/other/cyclomatic/linux_4.11_sorted.txt )
This is source code of some of them: do_check(), ext4_fill_super(), do_blockdev_direct_IO(), do_jit().
Most complex functions in Windows 7 ntoskrnl.exe file:
140569400 sub_140569400 edges=3070 nodes=1889 rets=1 E-N+2=1183 E-N+rets=1182
14007c640 MmAccessFault edges=2256 nodes=1424 rets=1 E-N+2=834 E-N+rets=833
1401a0410 FsRtlMdlReadCompleteDevEx edges=1241 nodes=752 rets=1 E-N+2=491 E-N+rets=490
14008c190 MmProbeAndLockPages edges=983 nodes=623 rets=1 E-N+2=362 E-N+rets=361
14037fd10 ExpQuerySystemInformation edges=995 nodes=671 rets=1 E-N+2=326 E-N+rets=325
140197260 MmProbeAndLockSelectedPages edges=875 nodes=551 rets=1 E-N+2=326 E-N+rets=325
140362a50 NtSetInformationProcess edges=880 nodes=586 rets=1 E-N+2=296 E-N+rets=295
....
( Full list: https://beginners.re/current-tree/other/cyclomatic/win7_ntoskrnl_sorted.txt )

From a bug hunter’s standpoint, complex functions are prone to have bugs, so an attention should be paid
to them.
Read more about it: https://en.wikipedia.org/wiki/Cyclomatic_complexity, http://wiki.c2.com/
?CyclomaticComplexityMetric.
Measuring cyclomatic complexity in MSVS (C#): https://blogs.msdn.microsoft.com/zainnab/2011/
05/17/code-metrics-cyclomatic-complexity/.
Couple of other Python scripts for measuring cyclomatic complexity in IDA: http://www.openrce.org/
articles/full_view/11, https://github.com/mxmssh/IDAmetrics (incl. other metrics).
GCC plugin: https://github.com/ephox-gcc-plugins/cyclomatic_complexity.
997
Chapter 12
Books/blogs worth reading
12.1 Books and other materials

12.1.1 Reverse Engineering
• Eldad Eilam, Reversing: Secrets of Reverse Engineering, (2005)
• Bruce Dang, Alexandre Gazet, Elias Bachaalany, Sebastien Josse, Practical Reverse Engineering: x86,
x64, ARM, Windows Kernel, Reversing Tools, and Obfuscation, (2014)
• Michael Sikorski, Andrew Honig, Practical Malware Analysis: The Hands-On Guide to Dissecting Mali-
cious Software, (2012)
• Chris Eagle, IDA Pro Book, (2011)
• Reginald Wong, Mastering Reverse Engineering: Re-engineer your ethical hacking skills, (2018)
(Outdated, but still interesting) Pavol Cerven, Crackproof Your Software: Protect Your Software Against
Crackers, (2002).
Also, Kris Kaspersky’s books.
12.1.2 Windows
• Mark Russinovich, Microsoft Windows Internals
• Peter Ferrie – The “Ultimate” Anti-Debugging Reference1
Blogs:
• Microsoft: Raymond Chen
• nynaeve.net
12.1.3 C/C++
• Brian W. Kernighan, Dennis M. Ritchie, The C Programming Language, 2ed, (1988)
• ISO/IEC 9899:TC3 (C C99 standard), (2007)2
• Bjarne Stroustrup, The C++ Programming Language, 4th Edition, (2013)
• C++11 standard3
• Agner Fog, Optimizing software in C++ (2015)4
• Marshall Cline, C++ FAQ5
• Dennis Yurichev, C/C++ programming language notes6
1 http://pferrie.host22.com/papers/antidebug.pdf
3 Also available as http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2013/n3690.pdf.
4 Also available as http://agner.org/optimize/optimizing_cpp.pdf.
6 Also available as http://yurichev.com/C-book.html
998
• JPL Institutional Coding Standard for the C Programming Language7
12.1.4 x86 / x86-64

• Intel manuals8
• AMD manuals9
• Agner Fog, The microarchitecture of Intel, AMD and VIA CPUs, (2016)10
• Agner Fog, Calling conventions (2015)11
• Intel® 64 and IA-32 Architectures Optimization Reference Manual, (2014)
• Software Optimization Guide for AMD Family 16h Processors, (2013)
Somewhat outdated, but still interesting to read:
Michael Abrash, Graphics Programming Black Book, 199712 (he is known for his work on low-level opti-
mization for such projects as Windows NT 3.1 and id Quake).
12.1.5 ARM
• ARM manuals13
• ARM(R) Architecture Reference Manual, ARMv7-A and ARMv7-R edition, (2012)
• [ARM Architecture Reference Manual, ARMv8, for ARMv8-A architecture profile, (2013)]14
• Advanced RISC Machines Ltd, The ARM Cookbook, (1994)15
12.1.6 Assembly language

Richard Blum — Professional Assembly Language.
12.1.7 Java
[Tim Lindholm, Frank Yellin, Gilad Bracha, Alex Buckley, The Java(R) Virtual Machine Specification / Java
SE 7 Edition] 16 .
12.1.8 UNIX
Eric S. Raymond, The Art of UNIX Programming, (2003)
12.1.9 Programming in general

• Brian W. Kernighan, Rob Pike, Practice of Programming, (1999)
• Henry S. Warren, Hacker’s Delight, (2002). Some people say tricks and hacks from the book are
not relevant today because they were good only for RISC CPUs, where branching instructions are
expensive. Nevertheless, these can help immensely to understand boolean algebra and what all the
mathematics near it.
7 Also available as https://yurichev.com/mirrors/C/JPL_Coding_Standard_C.pdf
8 Also available as http://www.intel.com/content/www/us/en/processors/architectures-software-developer-manuals.
html
9 Also available as http://developer.amd.com/resources/developer-guides-manuals/
10 Also available as http://agner.org/optimize/microarchitecture.pdf
11 Also available as http://www.agner.org/optimize/calling_conventions.pdf
12 Also available as https://github.com/jagregory/abrash-black-book
13 Also available as http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.subset.architecture.reference/
index.html
14 Also available as http://yurichev.com/mirrors/ARMv8-A_Architecture_Reference_Manual_(Issue_A.a).pdf
16 Also available as https://docs.oracle.com/javase/specs/jvms/se7/jvms7.pdf; http://docs.oracle.com/javase/specs/
jvms/se7/html/
999
12.1.10 Cryptography
• Bruce Schneier, Applied Cryptography, (John Wiley & Sons, 1994)
• (Free) lvh, Crypto 10117
• (Free) Dan Boneh, Victor Shoup, A Graduate Course in Applied Cryptography18 .
17 Also available as https://www.crypto101.io/

18 Also available as https://crypto.stanford.edu/~dabo/cryptobook/
1000
Chapter 13
Communities
There are two excellent RE1 -related subreddits on reddit.com: reddit.com/r/ReverseEngineering/ and red-
dit.com/r/remath (on the topics for the intersection of RE and mathematics).
There is also a RE part of the Stack Exchange website: reverseengineering.stackexchange.com.
On IRC there are ##re and ##asm channels on FreeNode2 .
1 Reverse Engineering
2 freenode.net
1001
Afterword
1002
13.1 Questions?
Do not hesitate to mail any questions to the author:
<first_name @ last_name . com> or <first_name . last_name @ gmail . com>. Do you have any
suggestion on new content for to the book? Please do not hesitate to send any corrections (including
grammar (you see how horrible my English is?)), etc.
The author is working on the book a lot, so the page and listing numbers, etc., are changing very rapidly.
Please do not refer to page and listing numbers in your emails to me. There is a much simpler method:
make a screenshot of the page, in a graphics editor underline the place where you see the error, and send
it to the author. He’ll fix it much faster. And if you familiar with git and LATEX you can fix the error right in
the source code:
GitHub.
Do not worry to bother me while writing me about any petty mistakes you found, even if you are not very
confident. I’m writing for beginners, after all, so beginners’ opinions and comments are crucial for my
job.
1003
Appendix
1004
.1 x86
.1.1 Terminology
Common for 16-bit (8086/80286), 32-bit (80386, etc.), 64-bit.
byte 8-bit. The DB assembly directive is used for defining variables and arrays of bytes. Bytes are passed
in the 8-bit part of registers: AL/BL/CL/DL/AH/BH/CH/DH/SIL/DIL/R*L.
word 16-bit. DW assembly directive —”—. Words are passed in the 16-bit part of the registers:
AX/BX/CX/DX/SI/DI/R*W.
double word (“dword”) 32-bit. DD assembly directive —”—. Double words are passed in registers (x86)
or in the 32-bit part of registers (x64). In 16-bit code, double words are passed in 16-bit register
pairs.
quad word (“qword”) 64-bit. DQ assembly directive —”—. In 32-bit environment, quad words are passed
in 32-bit register pairs.
tbyte (10 bytes) 80-bit or 10 bytes (used for IEEE 754 FPU registers).
paragraph (16 bytes)—term was popular in MS-DOS environment.
Data types of the same width (BYTE, WORD, DWORD) are also the same in Windows API.
.1.2 General purpose registers

It is possible to access many registers by byte or 16-bit word parts. .
It is all inheritance from older Intel CPUs (up to the 8-bit 8080) still supported for backward compatibility.
Older 8-bit CPUs (8080) had 16-bit registers divided by two.
Programs written for 8080 could access the low byte part of 16-bit registers, high byte part or the whole
16-bit register.
Perhaps, this feature was left in 8086 as a helper for easier porting.
This feature is usually not present in RISC CPUs.
Registers prefixed with R- appeared in x86-64, and those prefixed with E-—in 80386.
Thus, R-registers are 64-bit, and E-registers—32-bit.
8 more GPR’s were added in x86-86: R8-R15. .
N.B.: In the Intel manuals the byte parts of these registers are prefixed by L, e.g.: R8L, but IDA names
these registers by adding the B suffix, e.g.: R8B.
RAX/EAX/AX/AL
Byte number:
RAXx64
EAX
AX
AH AL
AKA accumulator. The result of a function is usually returned via this register.
RBX/EBX/BX/BL
Byte number:
RBXx64
EBX
BX
BH BL
1005
RCX/ECX/CX/CL
Byte number:
RCXx64
ECX
CX
CH CL
AKA counter: in this role it is used in REP prefixed instructions and also in shift instructions (SHL/SHR/RxL/RxR).
RDX/EDX/DX/DL
Byte number:
RDXx64
EDX
DX
DH DL
RSI/ESI/SI/SIL
Byte number:
RSIx64
ESI
SI
SILx64
AKA “source index”. Used as source in the instructions REP MOVSx, REP CMPSx.
RDI/EDI/DI/DIL
Byte number:
RDIx64
EDI
DI
DILx64
AKA “destination index”. Used as a pointer to the destination in the instructions REP MOVSx, REP STOSx.
R8/R8D/R8W/R8L
Byte number:
R8
R8D
R8W
R8L
R9/R9D/R9W/R9L
Byte number:
R9
R9D
R9W
R9L
1006
R10/R10D/R10W/R10L
Byte number:
R10
R10D
R10W
R10L
R11/R11D/R11W/R11L
Byte number:
R11
R11D
R11W
R11L
R12/R12D/R12W/R12L
Byte number:
R12
R12D
R12W
R12L
R13/R13D/R13W/R13L
Byte number:
R13
R13D
R13W
R13L
R14/R14D/R14W/R14L
Byte number:
R14
R14D
R14W
R14L
R15/R15D/R15W/R15L
Byte number:
R15
R15D
R15W
R15L
RSP/ESP/SP/SPL
Byte number:
RSP
ESP
SP
SPL
AKA stack pointer. Usually points to the current stack except in those cases when it is not yet initialized.
1007
RBP/EBP/BP/BPL
Byte number:
RBP
EBP
BP
BPL
AKA frame pointer. Usually used for local variables and accessing the arguments of the function. More
about it: ( 1.12.1 on page 67).
RIP/EIP/IP
Byte number:
RIPx64
EIP
IP
AKA “instruction pointer” 3 . Usually always points to the instruction to be executed right now. Cannot be
modified, however, it is possible to do this (which is equivalent):
MOV EAX, ...
JMP EAX
Or:
PUSH value
RET
CS/DS/ES/SS/FS/GS
16-bit registers containing code selector (CS), data selector (DS), stack selector (SS).
FS in win32 points to TLS, GS took this role in Linux. It is made so for faster access to the TLS and
other structures like the TIB.
In the past, these registers were used as segment registers ( 11.6 on page 986).
Flags register
AKA EFLAGS.
3 Sometimes also called “program counter”
1008
Bit (mask) Abbreviation (meaning) Description
0 (1) CF (Carry)
The CLC/STC/CMC instructions are used
for setting/resetting/toggling this flag
2 (4) PF (Parity) ( 1.25.7 on page 231).
4 (0x10) AF (Adjust) Exist solely for work with BCD-numbers
6 (0x40) ZF (Zero) Setting to 0
if the last operation’s result is equal to 0.
7 (0x80) SF (Sign)
8 (0x100) TF (Trap) Used for debugging.
If turned on, an exception is to be
generated after each instruction’s execution.
9 (0x200) IF (Interrupt enable) Are interrupts enabled.
The CLI/STI instructions are used
for setting/resetting the flag
10 (0x400) DF (Direction) A direction is set for the
REP MOVSx/CMPSx/LODSx/SCASx instructions.
The CLD/STD instructions are used
for setting/resetting the flag
See also: 3.26 on page 624.
11 (0x800) OF (Overflow)
12, 13 (0x3000) IOPL (I/O privilege level)i286
14 (0x4000) NT (Nested task)i286
16 (0x10000) RF (Resume)i386 Used for debugging.
The CPU ignores the hardware
breakpoint in DRx if the flag is set.
17 (0x20000) VM (Virtual 8086 mode)i386
18 (0x40000) AC (Alignment check)i486
19 (0x80000) VIF (Virtual interrupt)i586
20 (0x100000) VIP (Virtual interrupt pending)i586
21 (0x200000) ID (Identification)i586
All the rest flags are reserved.
.1.3 FPU registers

8 80-bit registers working as a stack: ST(0)-ST(7). N.B.: IDA calls ST(0) as just ST. Numbers are stored in
the IEEE 754 format.
long double value format:
79 78 64 63 62 0
S exponent I mantissa or fraction
( S — sign, I — integer part )
Control Word
Register controlling the behavior of the FPU.
1009
Bit Abbreviation (meaning) Description
0 IM (Invalid operation Mask)
1 DM (Denormalized operand Mask)
2 ZM (Zero divide Mask)
3 OM (Overflow Mask)
4 UM (Underflow Mask)
5 PM (Precision Mask)
7 IEM (Interrupt Enable Mask) Exceptions enabling, 1 by default (disabled)
8, 9 PC (Precision Control)
00 — 24 bits (REAL4)
10 — 53 bits (REAL8)
11 — 64 bits (REAL10)
10, 11 RC (Rounding Control)
00 — (by default) round to nearest
01 — round toward −∞
10 — round toward +∞
11 — round toward 0
12 IC (Infinity Control) 0 — (by default) treat +∞ and −∞ as unsigned
1 — respect both +∞ and −∞
The PM, UM, OM, ZM, DM, IM flags define if to generate exception in the case of a corresponding error.
Status Word
Read-only register.
Bit Abbreviation (meaning) Description
15 B (Busy) Is FPU do something (1) or results are ready (0)
14 C3
13, 12, 11 TOP points to the currently zeroth register
10 C2
9 C1
8 C0
7 IR (Interrupt Request)
6 SF (Stack Fault)
5 P (Precision)
4 U (Underflow)
3 O (Overflow)
2 Z (Zero)
1 D (Denormalized)
0 I (Invalid operation)
The SF, P, U, O, Z, D, I bits signal about exceptions.

About the C3, C2, C1, C0 you can read more here: ( 1.25.7 on page 231).
N.B.: When ST(x) is used, the FPU adds x to TOP (by modulo 8) and that is how it gets the internal register’s
number.
Tag Word
The register has current information about the usage of numbers registers.
Bit Abbreviation (meaning)
15, 14 Tag(7)
13, 12 Tag(6)
11, 10 Tag(5)
9, 8 Tag(4)
7, 6 Tag(3)
5, 4 Tag(2)
3, 2 Tag(1)
1, 0 Tag(0)
Each tag contains information about a physical FPU register (R(x)), not logical (ST(x)).
For each tag:
• 00 — The register contains a non-zero value
1010
• 01 — The register contains 0
• 10 — The register contains a special value (NAN4 , ∞, or denormal)
• 11 — The register is empty
.1.4 SIMD registers

MMX registers
8 64-bit registers: MM0..MM7.
SSE and AVX registers
SSE: 8 128-bit registers: XMM0..XMM7. In the x86-64 8 more registers were added: XMM8..XMM15.
AVX is the extension of all these registers to 256 bits .
.1.5 Debugging registers

Used for hardware breakpoints control.
• DR0 — address of breakpoint #1
• DR6 — a cause of break is reflected here
• DR7 — breakpoint types are set here
DR6
Bit (mask) Description
0 (1) B0 — breakpoint #1 has been triggered
13 (0x2000) BD — modification attempt of one of the DRx registers.
may be raised if GD is enabled
14 (0x4000) BS — single step breakpoint (TF flag has been set in EFLAGS).
Highest priority. Other bits may also be set.
15 (0x8000) BT (task switch flag)
N.B. A single step breakpoint is a breakpoint which occurs after each instruction. It can be enabled by
setting TF in EFLAGS ( .1.2 on page 1008).
DR7
Breakpoint types are set here.

4 Not a Number
1011
Bit (mask) Description
0 (1) L0 — enable breakpoint #1 for the current task
1 (2) G0 — enable breakpoint #1 for all tasks
2 (4) L1 — enable breakpoint #2 for the current task
3 (8) G1 — enable breakpoint #2 for all tasks
4 (0x10) L2 — enable breakpoint #3 for the current task
5 (0x20) G2 — enable breakpoint #3 for all tasks
6 (0x40) L3 — enable breakpoint #4 for the current task
7 (0x80) G3 — enable breakpoint #4 for all tasks
8 (0x100) LE — not supported since P6
9 (0x200) GE — not supported since P6
13 (0x2000) GD — exception is to be raised if any MOV instruction
tries to modify one of the DRx registers
16,17 (0x30000) breakpoint #1: R/W — type
18,19 (0xC0000) breakpoint #1: LEN — length
The breakpoint type is to be set as follows (R/W):
• 00 — instruction execution
• 01 — data writes
• 10 — I/O reads or writes (not available in user-mode)
• 11 — on data reads or writes
N.B.: breakpoint type for data reads is absent, indeed.
Breakpoint length is to be set as follows (LEN):

• 00 — one-byte
• 01 — two-byte
• 10 — undefined for 32-bit mode, eight-byte in 64-bit mode
• 11 — four-byte
.1.6 Instructions
Instructions marked as (M) are not usually generated by the compiler: if you see one of them, it is probably
a hand-written piece of assembly code, or a compiler intrinsic ( 11.3 on page 981).
Only the most frequently used instructions are listed here. You can read 12.1.4 on page 999 for a full
documentation.
Do you have to know all instruction’s opcodes by heart? No, only those which are used for code patching
( 11.1.1 on page 980). All the rest of the opcodes don’t need to be memorized.
Prefixes
LOCK forces CPU to make exclusive access to the RAM in multiprocessor environment. For the sake of
simplification, it can be said that when an instruction with this prefix is executed, all other CPUs in a
multiprocessor system are stopped. Most often it is used for critical sections, semaphores, mutexes.
Commonly used with ADD, AND, BTR, BTS, CMPXCHG, OR, XADD, XOR. You can read more about
critical sections here ( 6.5.4 on page 777).
REP is used with the MOVSx and STOSx instructions: execute the instruction in a loop, the counter is
located in the CX/ECX/RCX register. For a detailed description, read more about the MOVSx ( .1.6
on page 1015) and STOSx ( .1.6 on page 1016) instructions.
The instructions prefixed by REP are sensitive to the DF flag, which is used to set the direction.
1012
REPE/REPNE (AKA REPZ/REPNZ) used with CMPSx and SCASx instructions: execute the last instruction
in a loop, the count is set in the CX/ECX/RCX register. It terminates prematurely if ZF is 0 (REPE) or if
ZF is 1 (REPNE).
For a detailed description, you can read more about the CMPSx ( .1.6 on page 1018) and SCASx ( .1.6
on page 1016) instructions.
Instructions prefixed by REPE/REPNE are sensitive to the DF flag, which is used to set the direction.
Most frequently used instructions
These can be memorized in the first place.

ADC (add with carry) add values, increment the result if the CF flag is set. ADC is often used for the
addition of large values, for example, to add two 64-bit values in a 32-bit environment using two
ADD and ADC instructions. For example:
; work with 64-bit values: add val1 to val2.
; .lo means lowest 32 bits, .hi means highest.
ADD val1.lo, val2.lo
ADC val1.hi, val2.hi ; use CF that was set or cleared at the previous instruction
One more example: 1.34 on page 392.

ADD add two values
AND logical “and”
CALL call another function:
PUSH address_after_CALL_instruction; JMP label
CMP compare values and set flags, the same as SUB but without writing the result
DEC decrement. Unlike other arithmetic instructions, DEC doesn’t modify CF flag.
IMUL signed multiply IMUL often used instead of MUL, read more about it: 2.2.1 on page 450.
INC increment. Unlike other arithmetic instructions, INC doesn’t modify CF flag.
JCXZ, JECXZ, JRCXZ (M) jump if CX/ECX/RCX=0
JMP jump to another address. The opcode has a jump offset.
Jcc (where cc — condition code)
A lot of these instructions have synonyms (denoted with AKA), this was done for convenience. Syn-
onymous instructions are translated into the same opcode. The opcode has a jump offset.
JAE AKA JNC: jump if above or equal (unsigned): CF=0
JA AKA JNBE: jump if greater (unsigned): CF=0 and ZF=0
JBE jump if lesser or equal (unsigned): CF=1 or ZF=1
JB AKA JC: jump if below (unsigned): CF=1
JC AKA JB: jump if CF=1
JE AKA JZ: jump if equal or zero: ZF=1
JGE jump if greater or equal (signed): SF=OF
JG jump if greater (signed): ZF=0 and SF=OF
JLE jump if lesser or equal (signed): ZF=1 or SF≠OF
JL jump if lesser (signed): SF≠OF
JNAE AKA JC: jump if not above or equal (unsigned) CF=1
JNA jump if not above (unsigned) CF=1 and ZF=1
JNBE jump if not below or equal (unsigned): CF=0 and ZF=0
JNB AKA JNC: jump if not below (unsigned): CF=0
JNC AKA JAE: jump CF=0 synonymous to JNB.
JNE AKA JNZ: jump if not equal or not zero: ZF=0
1013
JNGE jump if not greater or equal (signed): SF≠OF
JNG jump if not greater (signed): ZF=1 or SF≠OF
JNLE jump if not lesser (signed): ZF=0 and SF=OF
JNL jump if not lesser (signed): SF=OF
JNO jump if not overflow: OF=0
JNS jump if SF flag is cleared
JNZ AKA JNE: jump if not equal or not zero: ZF=0
JO jump if overflow: OF=1
JPO jump if PF flag is cleared (Jump Parity Odd)
JP AKA JPE: jump if PF flag is set
JS jump if SF flag is set
JZ AKA JE: jump if equal or zero: ZF=1
LAHF copy some flag bits to AH:
7 6 4 2 0
SF ZF AF PF CF
This instruction is often used in FPU-related code.

LEAVE equivalent of the MOV ESP, EBP and POP EBP instruction pair — in other words, this instruction
sets the stack pointer (ESP) back and restores the EBP register to its initial state.
LEA (Load Effective Address) form an address
This instruction was intended not for summing values and multiplication but for forming an address,
e.g., for calculating the address of an array element by adding the array address, element index,
with multiplication of element size5 .
So, the difference between MOV and LEA is that MOV forms a memory address and loads a value from
memory or stores it there, but LEA just forms an address.
But nevertheless, it is can be used for any other calculations.
LEA is convenient because the computations performed by it does not alter CPU flags. This may be
very important for OOE processors (to create less data dependencies).
Aside from this, starting at least at Pentium, LEA instruction is executed in 1 cycle.
int f(int a, int b)
{
return a*8+b;
};
Listing 1: Optimizing MSVC 2010

_a$ = 8 ; size = 4
_b$ = 12 ; size = 4
_f PROC
mov eax, DWORD PTR _b$[esp-4]
lea eax, DWORD PTR [eax+ecx*8]
ret 0
_f ENDP
Intel C++ uses LEA even more:

int f1(int a)
{
return a*13;
};
5 See also: wikipedia
1014
Listing 2: Intel C++ 2011
_f1 PROC NEAR
mov ecx, DWORD PTR [4+esp] ; ecx = a
lea edx, DWORD PTR [ecx+ecx*8] ; edx = a*9
lea eax, DWORD PTR [edx+ecx*4] ; eax = a*9 + a*4 = a*13
ret
These two instructions performs faster than one IMUL.

MOVSB/MOVSW/MOVSD/MOVSQ copy byte/ 16-bit word/ 32-bit word/ 64-bit word from the address
which is in SI/ESI/RSI into the address which is in DI/EDI/RDI.
Together with the REP prefix, it is to be repeated in a loop, the count is to be stored in the CX/ECX/RCX
register: it works like memcpy() in C. If the block size is known to the compiler in the compile stage,
memcpy() is often inlined into a short code fragment using REP MOVSx, sometimes even as several
instructions.
The memcpy(EDI, ESI, 15) equivalent is:
; copy 15 bytes from ESI to EDI
CLD ; set direction to forward
MOV ECX, 3
REP MOVSD ; copy 12 bytes
MOVSW ; copy 2 more bytes
MOVSB ; copy remaining byte
( Supposedly, it works faster than copying 15 bytes using just one REP MOVSB).
MOVSX load with sign extension see also: ( 1.23.1 on page 200)
MOVZX load and clear all other bitsi see also: ( 1.23.1 on page 201)
MOV load value. this instruction name is misnomer, resulting in some confusion (data is not moved
but copied), in other architectures the same instructions is usually named “LOAD” and/or “STORE”
or something like that.
One important thing: if you set the low 16-bit part of a 32-bit register in 32-bit mode, the high 16
bits remains as they were. But if you modify the low 32-bit part of the register in 64-bit mode, the
high 32 bits of the register will be cleared.
Supposedly, it was done to simplify porting code to x86-64.
MUL unsigned multiply. IMUL often used instead of MUL, read more about it: 2.2.1 on page 450.
NEG negation: op = −op Same as NOT op / ADD op, 1.
NOP NOP. Its opcode is 0x90, it is in fact the XCHG EAX,EAX idle instruction. This implies that x86 does
not have a dedicated NOP instruction (as in many RISC). This book has at least one listing where
GDB shows NOP as 16-bit XCHG instruction: 1.11.1 on page 48.
More examples of such operations: ( .1.7 on page 1024).
NOP may be generated by the compiler for aligning labels on a 16-byte boundary. Another very
popular usage of NOP is to replace manually (patch) some instruction like a conditional jump to NOP
in order to disable its execution.
NOT op1: op1 = ¬op1. logical inversion Important feature—the instruction doesn’t change flags.
OR logical “or”
POP get a value from the stack: value=SS:[ESP]; ESP=ESP+4 (or 8)
PUSH push a value into the stack: ESP=ESP-4 (or 8); SS:[ESP]=value
RET return from subroutine: POP tmp; JMP tmp.
In fact, RET is an assembly language macro, in Windows and *NIX environment it is translated into
RETN (“return near”) or, in MS-DOS times, where the memory was addressed differently ( 11.6 on
page 986), into RETF (“return far”).
RET can have an operand. Then it works like this:
POP tmp; ADD ESP op1; JMP tmp. RET with an operand usually ends functions in the stdcall calling
convention, see also: 6.1.2 on page 725.
SAHF copy bits from AH to CPU flags:
1015
7 6 4 2 0
SF ZF AF PF CF
This instruction is often used in FPU-related code.

SBB (subtraction with borrow) subtract values, decrement the result if the CF flag is set. SBB is often
used for subtraction of large values, for example, to subtract two 64-bit values in 32-bit environment
using two SUB and SBB instructions. For example:
; work with 64-bit values: subtract val2 from val1.
; .lo means lowest 32 bits, .hi means highest.
SUB val1.lo, val2.lo
SBB val1.hi, val2.hi ; use CF that was set or cleared at the previous instruction
One more example: 1.34 on page 392.

SCASB/SCASW/SCASD/SCASQ (M) compare byte/ 16-bit word/ 32-bit word/ 64-bit word that’s stored in
AX/EAX/RAX with a variable whose address is in DI/EDI/RDI. Set flags as CMP does.
This instruction is often used with the REPNE prefix: continue to scan the buffer until a special value
stored in AX/EAX/RAX is found. Hence “NE” in REPNE: continue to scan while the compared values
are not equal and stop when equal.
It is often used like the strlen() C standard function, to determine an ASCIIZ string’s length:
Example:
lea edi, string
mov ecx, 0FFFFFFFFh ; scan 232 − 1 bytes, i.e., almost infinitely
xor eax, eax ; 0 is the terminator
repne scasb
add edi, 0FFFFFFFFh ; correct it
; now EDI points to the last character of the ASCIIZ string.
; lets determine string length

; current ECX = -1-strlen
not ecx
dec ecx
; now ECX contain string length
If we use a different AX/EAX/RAX value, the function acts like the memchr() standard C function, i.e.,
it finds a specific byte.
SHL shift value left
SHR shift value right:
7 6 5 4 3 2 1 0
CF 7 6 5 4 3 2 1 0 0
7 6 5 4 3 2 1 0
0 7 6 5 4 3 2 1 0 CF
These instructions are frequently used for multiplication and division by 2n . Another very frequent
application is processing bit fields: 1.28 on page 303.
SHRD op1, op2, op3: shift value in op2 right by op3 bits, taking bits from op1.
Example: 1.34 on page 392.
STOSB/STOSW/STOSD/STOSQ store byte/ 16-bit word/ 32-bit word/ 64-bit word from AX/EAX/RAX into
the address which is in DI/EDI/RDI.
1016
Together with the REP prefix, it is to be repeated in a loop, the counter is in the CX/ECX/RCX register:
it works like memset() in C. If the block size is known to the compiler on compile stage, memset() is
often inlined into a short code fragment using REP MOVSx, sometimes even as several instructions.
memset(EDI, 0xAA, 15) equivalent is:
; store 15 0xAA bytes to EDI
CLD ; set direction to forward
MOV EAX, 0AAAAAAAAh
MOV ECX, 3
REP STOSD ; write 12 bytes
STOSW ; write 2 more bytes
STOSB ; write remaining byte
( Supposedly, it works faster than storing 15 bytes using just one REP STOSB).
SUB subtract values. A frequently occurring pattern is SUB reg,reg, which implies zeroing of reg.
TEST same as AND but without saving the result, see also: 1.28 on page 303
XOR op1, op2: XOR6 values. op1 = op1⊕op2. A frequently occurring pattern is XOR reg,reg, which implies
zeroing of reg. See also: 2.6 on page 456.
Less frequently used instructions
BSF bit scan forward, see also: 1.36.2 on page 417

BSR bit scan reverse
BSWAP (byte swap), change value endianness.
BTC bit test and complement
BTR bit test and reset
BTS bit test and set
BT bit test
CBW/CWD/CWDE/CDQ/CDQE Sign-extend value:
CBW convert byte in AL to word in AX
CWD convert word in AX to doubleword in DX:AX
CWDE convert word in AX to doubleword in EAX
CDQ convert doubleword in EAX to quadword in EDX:EAX
CDQE (x64) convert doubleword in EAX to quadword in RAX
These instructions consider the value’s sign, extending it to high part of the newly constructed value.
See also: 1.34.5 on page 401.
Interestingly to know these instructions was initially named as SEX (Sign EXtend), as Stephen P. Morse
(one of Intel 8086 CPU designers) wrote in [Stephen P. Morse, The 8086 Primer, (1980)]7 :
The process of stretching numbers by extending the sign bit is called sign extension.
The 8086 provides instructions (Fig. 3.29) to facilitate the task of sign extension. These
instructions were initially named SEX (sign extend) but were later renamed to the more
conservative CBW (convert byte to word) and CWD (convert word to double word).
CLD clear DF flag.

CLI (M) clear IF flag.
CMC (M) toggle CF flag
CMOVcc conditional MOV: load if the condition is true. The condition codes are the same as in the Jcc
instructions ( .1.6 on page 1013).
6 eXclusive OR
7 Also available as https://archive.org/details/The8086Primer
1017
CMPSB/CMPSW/CMPSD/CMPSQ (M) compare byte/ 16-bit word/ 32-bit word/ 64-bit word from the ad-
dress which is in SI/ESI/RSI with the variable at the address stored in DI/EDI/RDI. Set flags as CMP
does.
Together with the REP prefix, it is to be repeated in a loop, the counter is stored in the CX/ECX/RCX
register, the process will run until the ZF flag is zero (e.g., until the compared values are equal to
each other, hence “E” in REPE).
It works like memcmp() in C.
Example from the Windows NT kernel (WRK v1.2):
Listing 3: base\ntos\rtl\i386\movemem.asm
; ULONG
; RtlCompareMemory (
; IN PVOID Source1,
; IN PVOID Source2,
; IN ULONG Length
; )
;
; Routine Description:
;
; This function compares two blocks of memory and returns the number
; of bytes that compared equal.
;
; Arguments:
;
; Source1 (esp+4) - Supplies a pointer to the first block of memory to
; compare.
;
; Source2 (esp+8) - Supplies a pointer to the second block of memory to
; compare.
;
; Length (esp+12) - Supplies the Length, in bytes, of the memory to be
; compared.
;
; Return Value:
;
; The number of bytes that compared equal is returned as the function
; value. If all bytes compared equal, then the length of the original
; block of memory is returned.
;
;--
RcmSource1 equ [esp+12]

RcmSource2 equ [esp+16]
RcmLength equ [esp+20]
CODE_ALIGNMENT
cPublicProc _RtlCompareMemory,3
cPublicFpo 3,0
push esi ; save registers

push edi ;
cld ; clear direction
mov esi,RcmSource1 ; (esi) -> first block to compare
mov edi,RcmSource2 ; (edi) -> second block to compare
;
; Compare dwords, if any.
;
rcm10: mov ecx,RcmLength ; (ecx) = length in bytes

shr ecx,2 ; (ecx) = length in dwords
jz rcm20 ; no dwords, try bytes
repe cmpsd ; compare dwords
jnz rcm40 ; mismatch, go find byte
;
; Compare residual bytes, if any.
;
1018
rcm20: mov ecx,RcmLength ; (ecx) = length in bytes
and ecx,3 ; (ecx) = length mod 4
jz rcm30 ; 0 odd bytes, go do dwords
repe cmpsb ; compare odd bytes
jnz rcm50 ; mismatch, go report how far we got
;
; All bytes in the block match.
;
rcm30: mov eax,RcmLength ; set number of matching bytes

pop edi ; restore registers
pop esi ;
stdRET _RtlCompareMemory
;
; When we come to rcm40, esi (and edi) points to the dword after the
; one which caused the mismatch. Back up 1 dword and find the byte.
; Since we know the dword didn't match, we can assume one byte won't.
;
rcm40: sub esi,4 ; back up

sub edi,4 ; back up
mov ecx,5 ; ensure that ecx doesn't count out
repe cmpsb ; find mismatch byte
;
; When we come to rcm50, esi points to the byte after the one that
; did not match, which is TWO after the last byte that did match.
;
rcm50: dec esi ; back up

sub esi,RcmSource1 ; compute bytes that matched
mov eax,esi ;
pop edi ; restore registers
pop esi ;
stdRET _RtlCompareMemory
stdENDP _RtlCompareMemory
N.B.: this function uses a 32-bit word comparison (CMPSD) if the block size is a multiple of 4, or
per-byte comparison (CMPSB) otherwise.
CPUID get information about the CPU’s features. see also: ( 1.30.6 on page 366).
DIV unsigned division
IDIV signed division
INT (M): INT x is analogous to PUSHF; CALL dword ptr [x*4] in 16-bit environment. It was widely
used in MS-DOS, functioning as a syscall vector. The registers AX/BX/CX/DX/SI/DI were filled with
the arguments and then the flow jumped to the address in the Interrupt Vector Table (located at the
beginning of the address space). It was popular because INT has a short opcode (2 bytes) and the
program which needs some MS-DOS services is not bother to determine the address of the service’s
entry point. The interrupt handler returns the control flow to caller using the IRET instruction.
The most busy MS-DOS interrupt number was 0x21, serving a huge part of its API. See also: [Ralf
Brown Ralf Brown’s Interrupt List], for the most comprehensive interrupt lists and other MS-DOS
information.
In the post-MS-DOS era, this instruction was still used as syscall both in Linux and Windows ( 6.3 on
page 738), but was later replaced by the SYSENTER or SYSCALL instructions.
INT 3 (M): this instruction is somewhat close to INT, it has its own 1-byte opcode (0xCC), and is actively
used while debugging. Often, the debuggers just write the 0xCC byte at the address of the breakpoint
to be set, and when an exception is raised, the original byte is restored and the original instruction
at this address is re-executed.
As of Windows NT, an EXCEPTION_BREAKPOINT exception is to be raised when the CPU executes this
instruction. This debugging event may be intercepted and handled by a host debugger, if one is
1019
loaded. If it is not loaded, Windows offers to run one of the registered system debuggers. If MSVS8 is
installed, its debugger may be loaded and connected to the process. In order to protect from reverse
engineering, a lot of anti-debugging methods check integrity of the loaded code.
MSVC has compiler intrinsic for the instruction: __debugbreak()9 .
There is also a win32 function in kernel32.dll named DebugBreak()10 , which also executes INT 3.
IN (M) input data from port. The instruction usually can be seen in OS drivers or in old MS-DOS code, for
example ( 8.8.3 on page 835).
IRET : was used in the MS-DOS environment for returning from an interrupt handler after it was called
by the INT instruction. Equivalent to POP tmp; POPF; JMP tmp.
LOOP (M) decrement CX/ECX/RCX, jump if it is still not zero.
LOOP instruction was often used in DOS-code which works with external devices. To add small delay,
this was done:
MOV CX, nnnn
LABEL: LOOP LABEL
Drawback is obvious: length of delay depends on CPU speed.

OUT (M) output data to port. The instruction usually can be seen in OS drivers or in old MS-DOS code, for
example ( 8.8.3 on page 835).
POPA (M) restores values of (R|E)DI, (R|E)SI, (R|E)BP, (R|E)BX, (R|E)DX, (R|E)CX, (R|E)AX registers from
the stack.
POPCNT population count. Counts the number of 1 bits in the value.
See: 2.7 on page 460.
POPF restore flags from the stack (AKA EFLAGS register)
PUSHA (M) pushes the values of the (R|E)AX, (R|E)CX, (R|E)DX, (R|E)BX, (R|E)BP, (R|E)SI, (R|E)DI registers
to the stack.
PUSHF push flags (AKA EFLAGS register)
RCL (M) rotate left via CF flag:
7 6 5 4 3 2 1 0 CF
CF 7 6 5 4 3 2 1 0
RCR (M) rotate right via CF flag:
CF 7 6 5 4 3 2 1 0
7 6 5 4 3 2 1 0 CF
ROL/ROR (M) cyclic shift

ROL: rotate left:
7 6 5 4 3 2 1 0
CF 7 6 5 4 3 2 1 0
ROR: rotate right:

8 Microsoft Visual Studio
9 MSDN
10 MSDN
1020
7 6 5 4 3 2 1 0
7 6 5 4 3 2 1 0 CF
Despite the fact that almost all CPUs have these instructions, there are no corresponding operations
in C/C++, so the compilers of these PLs usually do not generate these instructions.
For the programmer’s convenience, at least MSVC has the pseudofunctions (compiler intrinsics) _rotl()
and _rotr()11 , which are translated by the compiler directly to these instructions.
SAL Arithmetic shift left, synonymous to SHL
SAR Arithmetic shift right
7 6 5 4 3 2 1 0
7 6 5 4 3 2 1 0 CF
Hence, the sign bit always stays at the place of the MSB.
SETcc op: load 1 to operand (byte only) if the condition is true or zero otherwise. The condition codes
are the same as in the Jcc instructions ( .1.6 on page 1013).
STC (M) set CF flag
STD (M) set DF flag. This instruction is not generated by compilers and generally rare. For example, it
can be found in the ntoskrnl.exe Windows kernel file, in the hand-written memory copy routines.
STI (M) set IF flag
SYSCALL (AMD) call syscall ( 6.3 on page 738)
SYSENTER (Intel) call syscall ( 6.3 on page 738)
UD2 (M) undefined instruction, raises exception. Used for testing.
XCHG (M) exchange the values in the operands
This instruction is rare: compilers don’t generate it, because starting at Pentium, XCHG with address
in memory in operand executes as if it has LOCK prefix ([Michael Abrash, Graphics Programming
Black Book, 1997chapter 19]). Perhaps, Intel engineers did so for compatibility with synchronizing
primitives. Hence, XCHG starting at Pentium can be slow. On the other hand, XCHG was very popular
in assembly language programmers. So if you see XCHG in code, it can be a sign that this piece of
code is written manually. However, at least Borland Delphi compiler generates this instruction.
FPU instructions
-R suffix in the mnemonic usually implies that the operands are reversed, -P suffix implies that one element
is popped from the stack after the instruction’s execution, -PP suffix implies that two elements are popped.
-P instructions are often useful when we do not need the value in the FPU stack to be present anymore
after the operation.
FABS replace value in ST(0) by absolute value in ST(0)
FADD op: ST(0)=op+ST(0)
FADD ST(0), ST(i): ST(0)=ST(0)+ST(i)
FADDP ST(1)=ST(0)+ST(1); pop one element from the stack, i.e., the values in the stack are replaced by
their sum
FCHS ST(0)=-ST(0)
FCOM compare ST(0) with ST(1)
FCOM op: compare ST(0) with op
FCOMP compare ST(0) with ST(1); pop one element from the stack
11 MSDN
1021
FCOMPP compare ST(0) with ST(1); pop two elements from the stack
FDIVR op: ST(0)=op/ST(0)
FDIVR ST(i), ST(j): ST(i)=ST(j)/ST(i)
FDIVRP op: ST(0)=op/ST(0); pop one element from the stack
FDIVRP ST(i), ST(j): ST(i)=ST(j)/ST(i); pop one element from the stack
FDIV op: ST(0)=ST(0)/op
FDIV ST(i), ST(j): ST(i)=ST(i)/ST(j)
FDIVP ST(1)=ST(0)/ST(1); pop one element from the stack, i.e., the dividend and divisor values in the
stack are replaced by quotient
FILD op: convert integer and push it to the stack.
FIST op: convert ST(0) to integer op
FISTP op: convert ST(0) to integer op; pop one element from the stack
FLD1 push 1 to stack
FLDCW op: load FPU control word ( .1.3 on page 1009) from 16-bit op.
FLDZ push zero to stack
FLD op: push op to the stack.
FMUL op: ST(0)=ST(0)*op
FMUL ST(i), ST(j): ST(i)=ST(i)*ST(j)
FMULP op: ST(0)=ST(0)*op; pop one element from the stack
FMULP ST(i), ST(j): ST(i)=ST(i)*ST(j); pop one element from the stack
FSINCOS : tmp=ST(0); ST(1)=sin(tmp); ST(0)=cos(tmp)
√
FSQRT : ST (0) = ST (0)
FSTCW op: store FPU control word ( .1.3 on page 1009) into 16-bit op after checking for pending excep-
tions.
FNSTCW op: store FPU control word ( .1.3 on page 1009) into 16-bit op.
FSTSW op: store FPU status word ( .1.3 on page 1010) into 16-bit op after checking for pending excep-
tions.
FNSTSW op: store FPU status word ( .1.3 on page 1010) into 16-bit op.
FST op: copy ST(0) to op
FSTP op: copy ST(0) to op; pop one element from the stack
FSUBR op: ST(0)=op-ST(0)
FSUBR ST(0), ST(i): ST(0)=ST(i)-ST(0)
FSUBRP ST(1)=ST(0)-ST(1); pop one element from the stack, i.e., the value in the stack is replaced by
the difference
FSUB op: ST(0)=ST(0)-op
FSUB ST(0), ST(i): ST(0)=ST(0)-ST(i)
FSUBP ST(1)=ST(1)-ST(0); pop one element from the stack, i.e., the value in the stack is replaced by the
difference
FUCOM ST(i): compare ST(0) and ST(i)
FUCOM compare ST(0) and ST(1)
FUCOMP compare ST(0) and ST(1); pop one element from stack.
FUCOMPP compare ST(0) and ST(1); pop two elements from stack.
The instructions perform just like FCOM, but an exception is raised only if one of the operands is
SNaN, while QNaN numbers are processed smoothly.
FXCH ST(i) exchange values in ST(0) and ST(i)
1022
FXCH exchange values in ST(0) and ST(1)
Instructions having printable ASCII opcode
(In 32-bit mode).

These can be suitable for shellcode construction. See also: 8.14.1 on page 887.
ASCII character hexadecimal code x86 instruction

0 30 XOR
1 31 XOR
2 32 XOR
3 33 XOR
4 34 XOR
5 35 XOR
7 37 AAA
8 38 CMP
9 39 CMP
: 3a CMP
; 3b CMP
< 3c CMP
= 3d CMP
? 3f AAS
@ 40 INC
A 41 INC
B 42 INC
C 43 INC
D 44 INC
E 45 INC
F 46 INC
G 47 INC
H 48 DEC
I 49 DEC
J 4a DEC
K 4b DEC
L 4c DEC
M 4d DEC
N 4e DEC
O 4f DEC
P 50 PUSH
Q 51 PUSH
R 52 PUSH
S 53 PUSH
T 54 PUSH
U 55 PUSH
V 56 PUSH
W 57 PUSH
X 58 POP
Y 59 POP
Z 5a POP
[ 5b POP
\ 5c POP
] 5d POP
^ 5e POP
_ 5f POP
` 60 PUSHA
a 61 POPA
h 68 PUSH
i 69 IMUL
j 6a PUSH
k 6b IMUL
p 70 JO
q 71 JNO
r 72 JB
s 73 JAE
1023
t 74 JE
u 75 JNE
v 76 JBE
w 77 JA
x 78 JS
y 79 JNS
z 7a JP
Also:
ASCII character hexadecimal code x86 instruction

f 66 (in 32-bit mode) switch to
16-bit operand size
g 67 in 32-bit mode) switch to
16-bit address size
In summary: AAA, AAS, CMP, DEC, IMUL, INC, JA, JAE, JB, JBE, JE, JNE, JNO, JNS, JO, JP, JS, POP, POPA, PUSH,
PUSHA, XOR.
.1.7 npad
It is an assembly language macro for aligning labels on a specific boundary.
That’s often needed for the busy labels to where the control flow is often passed, e.g., loop body starts.
So the CPU can load the data or code from the memory effectively, through the memory bus, cache lines,
etc.
Taken from listing.inc (MSVC):
By the way, it is a curious example of the different NOP variations. All these instructions have no effects
whatsoever, but have a different size.
Having a single idle instruction instead of couple of NOP-s, is accepted to be better for CPU performance.
;; LISTING.INC
;;
;; This file contains assembler macros and is included by the files created
;; with the -FA compiler switch to be assembled by MASM (Microsoft Macro
;; Assembler).
;;
;; Copyright (c) 1993-2003, Microsoft Corporation. All rights reserved.
;; non destructive nops

npad macro size
if size eq 1
nop
else
if size eq 2
mov edi, edi
else
if size eq 3
; lea ecx, [ecx+00]
DB 8DH, 49H, 00H
else
if size eq 4
; lea esp, [esp+00]
DB 8DH, 64H, 24H, 00H
else
if size eq 5
add eax, DWORD PTR 0
else
if size eq 6
; lea ebx, [ebx+00000000]
DB 8DH, 9BH, 00H, 00H, 00H, 00H
else
1024
if size eq 7
; lea esp, [esp+00000000]
DB 8DH, 0A4H, 24H, 00H, 00H, 00H, 00H
else
if size eq 8
; jmp .+8; .npad 6
DB 0EBH, 06H, 8DH, 9BH, 00H, 00H, 00H, 00H
else
if size eq 9
; jmp .+9; .npad 7
DB 0EBH, 07H, 8DH, 0A4H, 24H, 00H, 00H, 00H, 00H
else
if size eq 10
; jmp .+A; .npad 7; .npad 1
DB 0EBH, 08H, 8DH, 0A4H, 24H, 00H, 00H, 00H, 00H, 90H
else
if size eq 11
; jmp .+B; .npad 7; .npad 2
DB 0EBH, 09H, 8DH, 0A4H, 24H, 00H, 00H, 00H, 00H, 8BH, 0FFH
else
if size eq 12
; jmp .+C; .npad 7; .npad 3
DB 0EBH, 0AH, 8DH, 0A4H, 24H, 00H, 00H, 00H, 00H, 8DH, 49H, 00H
else
if size eq 13
; jmp .+D; .npad 7; .npad 4
DB 0EBH, 0BH, 8DH, 0A4H, 24H, 00H, 00H, 00H, 00H, 8DH, 64H, 24H, 00H
else
if size eq 14
; jmp .+E; .npad 7; .npad 5
DB 0EBH, 0CH, 8DH, 0A4H, 24H, 00H, 00H, 00H, 00H, 05H, 00H, 00H, 00H, 00H
else
if size eq 15
; jmp .+F; .npad 7; .npad 6
DB 0EBH, 0DH, 8DH, 0A4H, 24H, 00H, 00H, 00H, 00H, 8DH, 9BH, 00H, 00H, 00H, 00H
else
%out error: unsupported npad size
.err
endif
endif
endif
endif
endif
endif
endif
endif
endif
endif
endif
endif
endif
endif
endif
endm
.2 ARM
.2.1 Terminology
ARM was initially developed as 32-bit CPU, so that’s why a word here, unlike x86, is 32-bit.
byte 8-bit. The DB assembly directive is used for defining variables and arrays of bytes.
halfword 16-bit. DCW assembly directive —”—.
word 32-bit. DCD assembly directive —”—.
doubleword 64-bit.
1025
quadword 128-bit.
.2.2 Versions
• ARMv4: Thumb mode introduced.
• ARMv6: used in iPhone 1st gen., iPhone 3G (Samsung 32-bit RISC ARM 1176JZ(F)-S that supports
Thumb-2)
• ARMv7: Thumb-2 was added (2003). was used in iPhone 3GS, iPhone 4, iPad 1st gen. (ARM Cortex-
A8), iPad 2 (Cortex-A9), iPad 3rd gen.
• ARMv7s: New instructions added. iPhone 5, iPhone 5c, iPad 4th gen. (Apple A6).
• ARMv8: 64-bit CPU, AKA ARM64 AKA AArch64. Was used in iPhone 5S, iPad Air (Apple A7). There is
no Thumb mode in 64-bit mode, only ARM (4-byte instructions).
.2.3 32-bit ARM (AArch32)

General purpose registers
• R0 — function result is usually returned using R0

• R1...R12 — GPRs
• R13 — AKA SP (stack pointer)
• R14 — AKA LR (link register)
• R15 — AKA PC (program counter)
R0-R3 are also called “scratch registers”: the function’s arguments are usually passed in them, and the
values in them are not required to be restored upon the function’s exit.
Current Program Status Register (CPSR)
Bit Description
0..4 M — processor mode
5 T — Thumb state
6 F — FIQ disable
7 I — IRQ disable
8 A — imprecise data abort disable
9 E — data endianness
10..15, 25, 26 IT — if-then state
16..19 GE — greater-than-or-equal-to
20..23 DNM — do not modify
24 J — Java state
27 Q — sticky overflow
28 V — overflow
29 C — carry/borrow/extend
30 Z — zero bit
31 N — negative/less than
VFP (floating point) and NEON registers
0..31bits 32..64 65..96 97..127

Q0128 bits
D064 bits D1
S032 bits S1 S2 S3
S-registers are 32-bit, used for the storage of single precision numbers.
D-registers are 64-bit ones, used for the storage of double precision numbers.
D- and S-registers share the same physical space in the CPU—it is possible to access a D-register via the
S-registers (it is senseless though).
Likewise, the NEON Q-registers are 128-bit ones and share the same physical space in the CPU with the
other floating point registers.
1026
In VFP 32 S-registers are present: S0..S31.
In VFPv2 there 16 D-registers are added, which in fact occupy the same space as S0..S31.
In VFPv3 (NEON or “Advanced SIMD”) there are 16 more D-registers, D0..D31, but the D16..D31 registers
are not sharing space with any other S-registers.
In NEON or “Advanced SIMD” another 16 128-bit Q-registers were added, which share the same space as
D0..D31.
.2.4 64-bit ARM (AArch64)

General purpose registers
The number of registers has been doubled since AArch32.

• X0 — function result is usually returned using X0
• X0...X7 — Function arguments are passed here.
• X8
• X9...X15 — are temporary registers, the callee function can use and not restore them.
• X16
• X17
• X18
• X19...X29 — callee function can use them, but must restore them upon exit.
• X29 — used as FP (at least GCC)
• X30 — “Procedure Link Register” AKA LR (link register).
• X31—register always contains zero AKA XZR or “Zero Register”. It’s 32-bit part is called WZR.
• SP, not a general purpose register anymore.
See also: [Procedure Call Standard for the ARM 64-bit Architecture (AArch64), (2013)]12 .
The 32-bit part of each X-register is also accessible via W-registers (W0, W1, etc.).
High 32-bit part low 32-bit part
X0
W0
.2.5 Instructions
There is a -S suffix for some instructions in ARM, indicating that the instruction sets the flags according
to the result. Instructions which lacks this suffix are not modify flags. For example ADD unlike ADDS will
add two numbers, but the flags will not be touched. Such instructions are convenient to use between
CMP where the flags are set and, e.g. conditional jumps, where the flags are used. They are also better in
terms of data dependency analysis (because less number of registers are modified during execution).
1027
Conditional codes table
Code Description Flags
EQ Equal Z == 1
NE Not equal Z == 0
CS AKA HS (Higher or Same) Carry set / Unsigned, Greater than, equal C == 1
CC AKA LO (LOwer) Carry clear / Unsigned, Less than C == 0
MI Minus, negative / Less than N == 1
PL Plus, positive or zero / Greater than, equal N == 0
VS Overflow V == 1
VC No overflow V == 0
HI Unsigned higher / Greater than C == 1 and
Z == 0
LS Unsigned lower or same / Less than or equal C == 0 or
Z == 1
GE Signed greater than or equal / Greater than or equal N == V
LT Signed less than / Less than N != V
GT Signed greater than / Greater than Z == 0 and
N == V
LE Signed less than or equal / Less than, equal Z == 1 or
N != V
None / AL Always Any
.3 MIPS
.3.1 Registers
( O32 calling convention )
General purpose registers GPR
Number Pseudoname Description

$0 $ZERO Always zero. Writing to this register is like NOP.
$1 $AT Used as a temporary register
for assembly macros and pseudo instructions.
$2 …$3 $V0 …$V1 Function result is returned here.
$4 …$7 $A0 …$A3 Function arguments.
$8 …$15 $T0 …$T7 Used for temporary data.
$16 …$23 $S0 …$S7 Used for temporary data∗ .
$24 …$25 $T8 …$T9 Used for temporary data.
$26 …$27 $K0 …$K1 Reserved for OS kernel.
$28 $GP Global Pointer∗∗ .
$29 $SP SP∗ .
$30 $FP FP∗ .
$31 $RA RA.
n/a PC PC.
n/a HI high 32 bit of multiplication or division remainder∗∗∗ .
n/a LO low 32 bit of multiplication and division remainder∗∗∗ .
Floating-point registers
Name Description
$F0..$F1 Function result returned here.
$F2..$F3 Not used.
$F4..$F11 Used for temporary data.
$F12..$F15 First two function arguments.
$F16..$F19 Used for temporary data.
$F20..$F31 Used for temporary data∗ .
∗
— Callee must preserve the value.
∗∗
— Callee must preserve the value ( except in PIC code).
∗∗∗
— accessible using the MFHI and MFLO instructions.
1028
.3.2 Instructions
There are 3 kinds of instructions:
• R-type: those which have 3 registers. R-instruction usually have the following form:
instruction destination, source1, source2
One important thing to keep in mind is that when the first and second register are the same, IDA may
show the instruction in its shorter form:
instruction destination/source1, source2
That somewhat reminds us of the Intel syntax for x86 assembly language.
• I-type: those which have 2 registers and a 16-bit immediate value.
• J-type: jump/branch instructions, have 26 bits for encoding the offset.
Jump instructions
What is the difference between B- instructions (BEQ, B, etc.) and J- ones (JAL, JALR, etc.)?
The B-instructions have an I-type, hence, the B-instructions’ offset is encoded as a 16-bit immediate. JR
and JALR are R-type and jump to an absolute address specified in a register. J and JAL are J-type, hence
the offset is encoded as a 26-bit immediate.
In short, B-instructions can encode a condition (B is in fact pseudo instruction for BEQ $ZERO, $ZERO,
LABEL), while J-instructions can’t.
.4 Some GCC library functions

name meaning
__divdi3 signed division
__moddi3 getting remainder (modulo) of signed division
__udivdi3 unsigned division
__umoddi3 getting remainder (modulo) of unsigned division
.5 Some MSVC library functions

ll in function name stands for “long long”, e.g., a 64-bit data type.
name meaning
__alldiv signed division
__allmul multiplication
__allrem remainder of signed division
__allshl shift left
__allshr signed shift right
__aulldiv unsigned division
__aullrem remainder of unsigned division
__aullshr unsigned shift right
Multiplication and shift left procedures are the same for both signed and unsigned numbers, hence there
is only one function for each operation here .
The source code of these function can be found in the installed MSVS, in VC/crt/src/intel/*.asm.
.6 Cheatsheets
.6.1 IDA
Hot-keys cheatsheet:
1029
key meaning
Space switch listing and graph view
C convert to code
D convert to data
A convert to string
* convert to array
U undefine
O make offset of operand
H make decimal number
R make char
B make binary number
Q make hexadecimal number
N rename identifier
? calculator
G jump to address
: add comment
Ctrl-X show references to the current function, label, variable
(incl. in local stack)
X show references to the function, label, variable, etc.
Alt-I search for constant
Ctrl-I search for the next occurrence of constant
Alt-B search for byte sequence
Ctrl-B search for the next occurrence of byte sequence
Alt-T search for text (including instructions, etc.)
Ctrl-T search for the next occurrence of text
Alt-P edit current function
Enter jump to function, variable, etc.
Esc get back
Num - fold function or selected area
Num + unhide function or area
Function/area folding may be useful for hiding function parts when you realize what they do. this is used
in myscript13 for hiding some often used patterns of inline code.
.6.2 OllyDbg
Hot-keys cheatsheet:
hot-key meaning
F7 trace into
F8 step over
F9 run
Ctrl-F2 restart
.6.3 MSVC
Some useful options which were used through this book. .
option meaning
/O1 minimize space
/Ob0 no inline expansion
/Ox maximum optimizations
/GS- disable security checks (buffer overflows)
/Fa(file) generate assembly listing
/Zi enable debugging information
/Zp(n) pack structs on n-byte boundary
/MD produced executable will use MSVCR*.DLL
Some information about MSVC versions: 5.1.1 on page 692.
.6.4 GCC
Some useful options which were used through this book.
13 GitHub
1030
option meaning
-Os code size optimization
-O3 maximum optimization
-regparm= how many arguments are to be passed in registers
-o file set name of output file
-g produce debugging information in resulting executable
-S generate assembly listing file
-masm=intel produce listing in Intel syntax
-fno-inline do not inline functions
.6.5 GDB
Some of commands we used in this book:
option meaning
break filename.c:number set a breakpoint on line number in source code
break function set a breakpoint on function
break *address set a breakpoint on address
b —”—
p variable print value of variable
run run
r —”—
cont continue execution
c —”—
bt print stack
set disassembly-flavor intel set Intel syntax
disas disassemble current function
disas function disassemble function
disas function,+50 disassemble portion
disas $eip,+0x10 —”—
disas/r disassemble with opcodes
info registers print all registers
info float print FPU-registers
info locals dump local variables (if known)
x/w ... dump memory as 32-bit word
x/w $rdi dump memory as 32-bit word
at address in RDI
x/10w ... dump 10 memory words
x/s ... dump memory as string
x/i ... dump memory as code
x/10c ... dump 10 characters
x/b ... dump bytes
x/h ... dump 16-bit halfwords
x/g ... dump giant (64-bit) words
finish execute till the end of function
next next instruction (don’t dive into functions)
step next instruction (dive into functions)
set step-mode on do not use line number information while stepping
frame n switch stack frame
info break list of breakpoints
del n delete breakpoint
set args ... set command-line arguments
1031
Acronyms Used
1032
OS Operating System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
OOP Object-Oriented Programming. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .542
PL Programming Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
PRNG Pseudorandom Number Generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
ROM Read-Only Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
ALU Arithmetic Logic Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
PID Program/process ID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 799
LF Line Feed (10 or ’\n’ in C/C++) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524
CR Carriage Return (13 or ’\r’ in C/C++) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524
LIFO Last In First Out . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
MSB Most Significant Bit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
LSB Least Significant Bit
NSA National Security Agency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
CFB Cipher Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845
CSPRNG Cryptographically Secure Pseudorandom Number Generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 846
ABI Application Binary Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
RA Return Address . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
PE Portable Executable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
SP stack pointer. SP/ESP/RSP in x86/x64. SP in ARM. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
DLL Dynamic-Link Library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 748
PC Program Counter. IP/EIP/RIP in x86/64. PC in ARM.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .19
LR Link Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
IDA Interactive Disassembler and Debugger developed by Hex-Rays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
IAT Import Address Table. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .748
INT Import Name Table. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .748
RVA Relative Virtual Address . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 748
1033
VA Virtual Address . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 748
OEP Original Entry Point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 738
MSVC Microsoft Visual C++
MSVS Microsoft Visual Studio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1020
ASLR Address Space Layout Randomization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607
MFC Microsoft Foundation Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 752
TLS Thread Local Storage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .280
AKA Also Known As . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .29
CRT C Runtime library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
CPU Central Processing Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
GPU Graphics Processing Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 855
FPU Floating-Point Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
CISC Complex Instruction Set Computing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .19
RISC Reduced Instruction Set Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
GUI Graphical User Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 744
RTTI Run-Time Type Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
BSS Block Started by Symbol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
SIMD Single Instruction, Multiple Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
BSOD Blue Screen of Death . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 739
DBMS Database Management Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
ISA Instruction Set Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
HPC High-Performance Computing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .517
SEH Structured Exception Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
ELF Executable File format widely used in *NIX systems including Linux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
TIB Thread Information Block. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .280
PIC Position Independent Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .537
1034
NAN Not a Number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1011
NOP No Operation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6
BEQ (PowerPC, ARM) Branch if Equal. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .94
BNE (PowerPC, ARM) Branch if Not Equal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
BLR (PowerPC) Branch to Link Register. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .819
XOR eXclusive OR. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1017
MCU Microcontroller Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495
RAM Random-Access Memory. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3
GCC GNU Compiler Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
EGA Enhanced Graphics Adapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 986
VGA Video Graphics Array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 986
API Application Programming Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 618
ASCII American Standard Code for Information Interchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
ASCIIZ ASCII Zero (null-terminated ASCII string ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
IA64 Intel Architecture 64 (Itanium) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
EPIC Explicitly Parallel Instruction Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 984
OOE Out-of-Order Execution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .462
MSDN Microsoft Developer Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610
STL (C++) Standard Template Library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564
PODT (C++) Plain Old Data Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575
HDD Hard Disk Drive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 587
VM Virtual Memory
WRK Windows Research Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 708
GPR General Purpose Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
SSDT System Service Dispatch Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 739
RE Reverse Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1001
1035
RAID Redundant Array of Independent Disks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
BCD Binary-Coded Decimal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
BOM Byte Order Mark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 699
GDB GNU Debugger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
FP Frame Pointer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
MBR Master Boot Record . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704
JPE Jump Parity Even (x86 instruction) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
CIDR Classless Inter-Domain Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
STMFD Store Multiple Full Descending (ARM instruction)
LDMFD Load Multiple Full Descending (ARM instruction)
STMED Store Multiple Empty Descending (ARM instruction). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30
LDMED Load Multiple Empty Descending (ARM instruction) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
STMFA Store Multiple Full Ascending (ARM instruction) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
LDMFA Load Multiple Full Ascending (ARM instruction) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
STMEA Store Multiple Empty Ascending (ARM instruction) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
LDMEA Load Multiple Empty Ascending (ARM instruction) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
APSR (ARM) Application Program Status Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
FPSCR (ARM) Floating-Point Status and Control Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
RFC Request for Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 703
TOS Top of Stack. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .655
LVA (Java) Local Variable Array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 661
JVM Java Virtual Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
JIT Just-In-Time compilation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
CDFS Compact Disc File System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715
CD Compact Disc
ADC Analog-to-Digital Converter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 711
1036
EOF End of File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
DIY Do It Yourself . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614
MMU Memory Management Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 605
DES Data Encryption Standard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
MIME Multipurpose Internet Mail Extensions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .444
DBI Dynamic Binary Instrumentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523
XML Extensible Markup Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623
JSON JavaScript Object Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623
URL Uniform Resource Locator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
IV Initialization Vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
RSA Rivest Shamir Adleman . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .939
CPRNG Cryptographically secure PseudoRandom Number Generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 940
GiB Gibibyte . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 954
CRC Cyclic redundancy check . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 974
AES Advanced Encryption Standard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 974
GC Garbage Collector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612
IDE Integrated development environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
BB Basic Block . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 996
1037
Glossary
anti-pattern Generally considered as bad practice. 31, 76, 461

arithmetic mean a sum of all values divided by their count. 518
atomic operation “ατ oµoς” stands for “indivisible” in Greek, so an atomic operation is guaranteed not
to be interrupted by other threads. 633, 778
basic block a group of instructions that do not have jump/branch instructions, and also don’t have jumps
inside the block from the outside. In IDA it looks just like as a list of instructions without empty lines.
683, 987, 988
callee A function being called by another. 32, 46, 66, 86, 97, 99, 101, 418, 462, 544, 644, 725–728, 730,
731, 1028
caller A function calling another. 6–8, 10, 28, 46, 86, 97, 98, 100, 108, 155, 418, 471, 544, 725, 727,
728, 731
compiler intrinsic A function specific to a compiler which is not an usual library function. The compiler
generates a specific machine code instead of a call to it. Often, it’s a pseudofunction for a specific
CPU instruction. Read more: ( 11.3 on page 981). 1019
CP/M Control Program for Microcomputers: a very basic disk OS used before MS-DOS. 888
decrement Decrease by 1. 18, 183, 184, 202, 437, 718, 1013, 1015, 1020
dongle Dongle is a small piece of hardware connected to LPT printer port (in past) or to USB. 818
endianness Byte order. 21, 78, 343, 1017
GiB Gibibyte: 230 or 1024 mebibytes or 1073741824 bytes. 15
heap usually, a big chunk of memory provided by the OS so that applications can divide it by themselves
as they wish. malloc()/free() work with the heap. 30, 344, 560, 562, 563, 575, 577, 592, 593, 623,
747, 748
increment Increase by 1. 16, 19, 184, 187, 202, 208, 324, 327, 437, 1013
integral data type usual numbers, but not a real ones. may be used for passing variables of boolean
data type and enumerations. 230
jump offset a part of the JMP or Jcc instruction’s opcode, to be added to the address of the next instruc-
tion, and this is how the new PC is calculated. May be negative as well. 93, 132, 1013
leaf function A function which does not call any other function. 27, 31
link register (RISC) A register where the return address is usually stored. This makes it possible to call
leaf functions without using the stack, i.e., faster. 31, 819, 1026, 1027
loop unwinding It is when a compiler, instead of generating loop code for n iterations, generates just n
copies of the loop body, in order to get rid of the instructions for loop maintenance. 186
name mangling used at least in C++, where the compiler needs to encode the name of class, method
and argument types in one string, which will become the internal name of the function. You can read
more about it here: 3.21.1 on page 542. 542, 693, 694
1038
NaN not a number: a special cases for floating point numbers, usually signaling about errors. 232, 254,
986
NEON AKA “Advanced SIMD”—SIMD from ARM. 1026
NOP “no operation”, idle instruction. 718
NTAPI API available only in the Windows NT line. Largely not documented by Microsoft. 786
padding Padding in English language means to stuff a pillow with something to give it a desired (bigger)
form. In computer science, padding means to add more bytes to a block so it will have desired size,
like 2n bytes.. 701
PDB (Win32) Debugging information file, usually just function names, but sometimes also function argu-
ments and local variables names. 692, 750, 786, 787, 794, 795, 871
POKE BASIC language instruction for writing a byte at a specific address. 718
product Multiplication result. 97, 223, 226, 405, 430, 450
quotient Division result. 216, 219, 221, 222, 226, 429, 497, 519
real number numbers which may contain a dot. this is float and double in C/C++. 216
register allocator The part of the compiler that assigns CPU registers to local variables. 201, 305, 418
reverse engineering act of understanding how the thing works, sometimes in order to clone it. iii, 1019
security cookie A random value, different at each execution. You can read more about it here: 1.26.3
on page 279. 769
stack frame A part of the stack that contains information specific to the current function: local variables,
function arguments, RA, etc.. 67, 97, 98, 478, 769
stack pointer A register pointing to a place in the stack. 9, 11, 19, 29, 34, 42, 54, 55, 73, 99, 544, 644,
725–728, 1007, 1014, 1026, 1033
stdout standard output. 20, 34, 155
tail call It is when the compiler (or interpreter) transforms the recursion (tail recursion) into an iteration
for efficiency. 481
thunk function Tiny function with a single role: call another function. 22, 41, 389, 819, 828
tracer My own simple debugging tool. You can read more about it here: 7.2.1 on page 781. 188–190,
541, 610, 696, 706, 709, 765, 773, 814, 815, 873, 879, 883, 885, 902, 980
user mode A restricted CPU mode in which it all application software code is executed. cf. kernel mode.
835
Windows NT Windows NT, 2000, XP, Vista, 7, 8, 10. 289, 416, 643, 699, 739, 749, 777, 890, 1019
word data type fitting in GPR. In the computers older than PCs, the memory size was often measured in
words rather than bytes.. 443–446, 451, 566, 624
xoring often used in the English language, which implying applying the XOR operation. 769, 831, 834
1039
Index
.NET, 754 FMOV, 439

0x0BADF00D, 75 FMRS, 319
0xCCCCCCCC, 75 IT, 150, 259, 282
LDMccFD, 135
Ada, 105 LDMEA, 30
AES, 844 LDMED, 30
Alpha AXP, 2 LDMFA, 30
AMD, 730 LDMFD, 19, 30, 135
Angband, 301 LDP, 24
Angry Birds, 260 LDR, 56, 73, 81, 268, 285, 437
Apollo Guidance Computer, 210 LDRB, 361
ARM, 208, 722, 819, 1025 LDRB.W, 208
Addressing modes, 437 LDRSB, 208
ARM mode, 2 LEA, 462
ARM1, 446 LSL, 331, 334
armel, 226 LSL.W, 331
armhf, 226 LSLR, 535
Condition codes, 135 LSLS, 269, 318, 535
D-registers, 225, 1026 LSR, 334
Data processing instructions, 499 LSRS, 318
DCB, 19 MADD, 103
hard float, 226 MLA, 102, 103
if-then block, 259 MOV, 8, 19, 20, 331, 498
Instructions MOVcc, 146, 150
ADC, 395 MOVK, 439
ADD, 20, 105, 135, 191, 319, 331, 499, 1027 MOVT, 20, 498
ADDAL, 135 MOVT.W, 21
ADDCC, 174 MOVW, 21
ADDS, 103, 395, 1027 MUL, 105
ADR, 18, 135 MULS, 103
ADRcc, 135, 163, 462 MVNS, 208
ADRP/ADD pair, 23, 54, 82, 286, 300, 439 NEG, 506
ANDcc, 535 ORR, 313
ASR, 334 POP, 18–20, 29, 31
ASRS, 313, 499 PUSH, 20, 29, 31
B, 54, 135, 136 RET, 24
Bcc, 95, 96, 146 RSB, 140, 295, 331, 506
BCS, 136, 262 SBC, 395
BEQ, 94, 163 SMMUL, 499
BGE, 136 STMEA, 30
BIC, 313, 318, 336 STMED, 30
BL, 19–23, 135, 440 STMFA, 30, 57
BLcc, 135 STMFD, 18, 30
BLE, 136 STMIA, 55
BLS, 136 STMIB, 57
BLT, 191 STP, 23, 54
BLX, 21 STR, 55, 268
BNE, 136 SUB, 55, 295, 331
BX, 102, 175 SUBcc, 535
CMP, 94, 95, 135, 163, 174, 191, 331, 1027 SUBEQ, 209
CSEL, 144, 148, 150, 331 SUBS, 395
EOR, 318 SXTB, 362
FCMPE, 262 SXTW, 300
FCSEL, 262
1040
TEST, 201 C language elements
TST, 307, 331 C99, 108
VADD, 226 bool, 303
VDIV, 226 restrict, 514
VLDR, 226 variable length arrays, 282
VMOV, 226, 259 Comma, 992
VMOVGT, 259 const, 9, 81, 468
VMRS, 259 for, 183, 483
VMUL, 226 if, 123, 154
XOR, 141, 319 Pointers, 66, 73, 109, 381, 418, 596
Leaf function, 31 Post-decrement, 437
Mode switching, 102, 175 Post-increment, 437
mode switching, 21 Pre-decrement, 437
Optional operators Pre-increment, 437
ASR, 331, 499 ptrdiff_t, 613
LSL, 268, 295, 331, 439 return, 10, 86, 107
LSR, 331, 499 Short-circuit, 525, 527, 993
ROR, 331 switch, 153, 154, 163
RRX, 331 while, 200
Pipeline, 174 C standard library
Registers alloca(), 34, 282, 462, 760
APSR, 259 assert(), 288, 703
FPSCR, 259 atexit(), 565
Link Register, 19, 31, 54, 176, 1026 atoi(), 500, 862
R0, 106, 1026 close(), 742
scratch registers, 208, 1026 exit(), 471
X0, 1026 fread(), 620
Z, 94, 1026 free(), 462, 592
S-registers, 225, 1026 fwrite(), 620
soft float, 226 getenv(), 863
Thumb mode, 2, 136, 175 localtime(), 652
Thumb-2 mode, 2, 175, 259, 261 localtime_r(), 352
ARM64 longjmp, 625
lo12, 55 longjmp(), 155
ASLR, 748 malloc(), 345, 462, 592
AT&T syntax, 12, 36 memchr(), 1016
AWK, 708 memcmp(), 448, 513, 704, 1017
memcpy(), 12, 66, 511, 624, 1015
Base address, 748 memmove(), 624
base32, 701 memset(), 264, 510, 883, 1016
Base64, 700 open(), 742
base64, 703, 841, 941 pow(), 228
base64scanner, 460, 701 puts(), 20
bash, 106 qsort(), 381, 470
BASIC rand(), 336, 695, 792, 794, 813, 839
POKE, 718 read(), 620, 742
BeagleBone, 851 realloc(), 462
binary grep, 706, 780 scanf(), 66
Binary Ninja, 780 setjmp, 625
Binary tree, 582 srand(), 813
BIND.EXE, 753 strcat(), 514
BinNavi, 780 strcmp(), 448, 471, 507, 742
binutils, 376 strcpy(), 12, 509, 840
Binwalk, 933 strlen(), 200, 414, 509, 526, 1016
Bitcoin, 630, 851 strstr(), 470
Boehm garbage collector, 612 strtok, 211
Boolector, 41 time(), 652, 813
Booth’s multiplication algorithm, 216 toupper(), 533
Borland C++, 603 va_arg, 518
Borland C++Builder, 694 va_list, 522
Borland Delphi, 14, 694, 697, 1021 vprintf, 522
BSoD, 738 write(), 620
BSS, 749 C++, 874
Buffer Overflow, 271, 278, 769 C++11, 575, 734
1041
exceptions, 760 Fused multiply–add, 102, 103
ostream, 557 Fuzzing, 506
References, 558
RTTI, 557 Garbage collector, 612, 677
STL, 692 GCC, 693, 1029, 1030
std::forward_list, 575 GDB, 27, 47, 51, 278, 389, 390, 781, 1030
std::list, 566 GeoIP, 930
std::map, 582 GHex, 780
std::set, 582 Glibc, 389, 624, 738
std::string, 559 Global variables, 76
std::vector, 575 GNU Scientific Library, 356
C11, 734 GnuPG, 940
Callbacks, 381 GraphViz, 611
Canary, 279 grep usage, 190, 260, 692, 706, 709, 872
cdecl, 42, 725
Chess, 459 Hash functions, 462
Cipher Feedback mode, 845 HASP, 705
clusterization, 938 Heartbleed, 623, 850
COFF, 826 Heisenbug, 630, 638
column-major order, 290 Hex-Rays, 107, 197, 296, 301, 613, 637, 988
Compiler intrinsic, 35, 450, 981 Hiew, 92, 132, 153, 697, 702, 750, 751, 754, 780,
Compiler’s anomalies, 146, 299, 313, 330, 493, 531, 980
981 Honeywell 6070, 444
Core dump, 606
ICQ, 718
Cray, 405, 446, 457, 460
IDA, 86, 153, 376, 514, 686, 699, 780, 781, 962,
CRC32, 462, 482
1029
CRT, 744, 765
var_?, 55, 73
CryptoMiniSat, 425
IEEE 754, 217, 315, 374, 425, 1005
CryptoPP, 724, 843
Inline code, 192, 312, 506, 548, 579
Cygwin, 693, 696, 754, 782
Integer overflow, 105
Data general Nova, 216 Intel
De Morgan’s laws, 994 8080, 208
DEC Alpha, 404 8086, 208, 312, 835
DES, 405, 418 Memory model, 650, 986
dlopen(), 742 8253, 889
dlsym(), 742 80286, 835, 987
dmalloc, 606 80386, 312, 987
Donald E. Knuth, 446 80486, 217
DOSBox, 890 FPU, 217
DosBox, 709 Intel 4004, 443
double, 217, 730 Intel C++, 9, 406, 981, 987, 1014
Doubly linked list, 458, 566 Intel syntax, 12, 18
dtruss, 781 iPod/iPhone/iPad, 18
Duff’s device, 494 Itanium, 404, 984
Dynamically loaded libraries, 21
JAD, 5
Edsger W. Dijkstra, 593 Java, 445, 654
EICAR, 887 John Carmack, 524
ELF, 79 JPEG, 938
Entropy, 911, 929 jumptable, 167, 175
Error messages, 702
Keil, 18
fastcall, 14, 33, 65, 305, 726 kernel panic, 738
fetchmail, 444 kernel space, 738
FidoNet, 701
LAPACK, 22
FILETIME, 402
LARGE_INTEGER, 402
FIXUP, 816
LD_PRELOAD, 742
float, 217, 730
Linker, 81, 542
Forth, 676
Linux, 306, 739, 874
FORTRAN, 22
libc.so.6, 305, 389
Fortran, 290, 514, 593, 693
LISP, 599
FreeBSD, 705
LLDB, 781
Function epilogue, 28, 54, 56, 135, 361, 708
LLVM, 18
Function prologue, 10, 28, 31, 55, 279, 708
1042
long double, 217 MOVE, 25, 83
Loop unwinding, 186 NEGU, 141
LZMA, 933 NOP, 27, 83
NOT, 210
Mac OS Classic, 818 Registers
Mac OS X, 782 FCCR, 263
Mathematica, 593, 804 HI, 499
MD5, 462, 704 LO, 499
memfrob(), 843 MS-DOS, 14, 33, 280, 603, 647, 704, 709, 718, 748,
Memoization, 805 835, 887, 888, 942, 986, 1005, 1015, 1019,
MFC, 751, 863 1020
Microsoft, 402 DOS extenders, 987
Microsoft Word, 623 MSVC, 1029, 1030
MIDI, 704 MSVCRT.DLL, 901
MinGW, 693, 901
minifloat, 439 Name mangling, 542
MIPS, 2, 712, 723, 749, 819, 937 Native API, 749
Branch delay slot, 8 Non-a-numbers (NaNs), 254
Global Pointer, 24, 296 Notepad, 934
Instructions NSA, 460
ADD, 105
ADDIU, 25, 84, 85 objdump, 376, 741, 754, 780
ADDU, 105 octet, 444
AND, 314 OEP, 748, 754
BC1F, 263 OllyDbg, 44, 69, 78, 98, 110, 126, 169, 187, 203,
BC1T, 263 220, 233, 244, 266, 273, 276, 290, 291, 322,
BEQ, 96, 137 343, 360, 361, 366, 369, 384, 751, 781, 1030
BLTZ, 141 OOP
BNE, 137 Polymorphism, 542
BNEZ, 177 opaque predicate, 538
BREAK, 499 OpenMP, 630, 695
C.LT.D, 263 OpenSSL, 623, 850
J, 6, 8, 25 OpenWatcom, 693, 727
JAL, 105 Oracle RDBMS, 9, 405, 702, 757, 874, 882, 883, 954,
JALR, 25, 105 964, 981, 987
JR, 166
LB, 197 Page (memory), 416
LBU, 197 Pascal, 697
LI, 441 PDP-11, 437
LUI, 25, 84, 85, 317, 441 PGP, 701
LW, 25, 74, 85, 166, 441 Phrack, 701
MFHI, 105, 500, 1028 Pin, 523
MFLO, 105, 499, 1028 PNG, 936
MTC1, 378 position-independent code, 18, 739
MULT, 105 PowerPC, 2, 24, 818
NOR, 210 Propagating Cipher Block Chaining, 855
OR, 27 Punched card, 264
ORI, 314, 441 puts() instead of printf(), 20, 71, 106, 133
SB, 197 Python, 523, 593
SLL, 177, 212, 333 ctypes, 733
SLLV, 333
Qt, 14
SLT, 137
Quake, 524
SLTIU, 177
Quake III Arena, 380
SLTU, 137, 139, 177
SRL, 217 Racket, 992
SUBU, 141 rada.re, 13
SW, 61 Radare, 781
Load delay slot, 166 radare2, 939
O32, 61, 65, 1027 rafind2, 780
Pseudoinstructions RAID4, 457
B, 194 RAM, 81
BEQZ, 139 Raspberry Pi, 18
LA, 27 ReactOS, 763
LI, 8 Recursion, 29, 30, 481
1043
Tail recursion, 481 xxd, 780, 917
Register allocation, 418 Unrolled loop, 192, 282, 494, 496, 510
Relocation, 21 uptime, 742
Resource Hacker, 785 UPX, 940
Reverse Polish notation, 264 USB, 820
RISC pipeline, 135 UseNet, 701
ROM, 81 user space, 738
ROT13, 843 user32.dll, 152
row-major order, 290 UTF-16, 445
RSA, 5 UTF-16LE, 698, 699
RVA, 748 UTF-8, 698, 941
Uuencode, 941
SAP, 692, 871 Uuencoding, 701
Scheme, 992
SCO OpenServer, 826 VA, 748
Scratch space, 729 Valgrind, 638
Security cookie, 279, 769 Variance, 841
Security through obscurity, 703
SHA1, 462 Watcom, 693
SHA512, 630 win32
Shadow space, 100, 101, 426 FindResource(), 599
Shellcode, 537, 739, 748, 888, 1022 GetOpenFileName, 210
Signed numbers, 124, 448 GetProcAddress(), 610
SIMD, 425, 513 HINSTANCE, 611
Software cracking, 14, 151, 609 HMODULE, 611
SQLite, 611 LoadLibrary(), 610
SSE, 425 MAKEINTRESOURCE(), 599
SSE2, 425 WinDbg, 781
Stack, 29, 97, 155 Windows, 777
Stack frame, 67 API, 1005
Stack overflow, 30 EnableMenuItem, 786
stdcall, 725, 980 IAT, 748
strace, 742, 781 INT, 748
strtoll(), 854 KERNEL32.DLL, 304
Stuxnet, 705 MSVCR80.DLL, 382
Syntactic Sugar, 154 NTAPI, 786
syscall, 305, 738, 781 ntoskrnl.exe, 874
Sysinternals, 702, 782 PDB, 692, 750, 786, 794, 871
Structured Exception Handling, 36, 755
Tabulation hashing, 459 TIB, 280, 755, 1008
Tagged pointers, 599 Win32, 303, 699, 742, 748, 987
TCP/IP, 461 GetProcAddress, 753
thiscall, 542, 544, 727 LoadLibrary, 753
Thumb-2 mode, 21 MulDiv(), 450, 803
thunk-functions, 22, 753, 819, 828 Ordinal, 751
TLS, 280, 733, 749, 754, 1008 RaiseException(), 755
Callbacks, 737, 754 SetUnhandledExceptionFilter(), 757
Tor, 701 Windows 2000, 749
tracer, 188, 386, 388, 696, 706, 709, 765, 773, 781, Windows 3.x, 643, 987
844, 873, 879, 883, 885, 980 Windows NT4, 749
Turbo C++, 603 Windows Vista, 748, 786
Windows XP, 749, 754, 794
uClibc, 624 Windows 2000, 403
UCS-2, 445 Windows 98, 152
UFS2, 705 Windows File Protection, 153
Unicode, 698 Windows Research Kernel, 404
UNIX Wine, 763
chmod, 4 Wolfram Mathematica, 911
diff, 718
fork, 625 x86
getopt, 854 AVX, 405
grep, 702, 980 Flags
mmap(), 603 CF, 33, 1013, 1015, 1017, 1020, 1021
od, 780 DF, 1017, 1021
strings, 701, 780 IF, 1017, 1021
1044
FPU, 1009 FNSTCW, 1022
Instructions FNSTSW, 231, 254, 1022
AAA, 1024 FSCALE, 379
AAS, 1024 FSINCOS, 1022
ADC, 394, 647, 1013 FSQRT, 1022
ADD, 9, 42, 97, 501, 647, 1013 FST, 1022
ADDSD, 426 FSTCW, 1022
ADDSS, 436 FSTP, 229, 1022
ADRcc, 143 FSTSW, 1022
AESDEC, 844 FSUB, 1022
AESENC, 844 FSUBP, 1022
AESKEYGENASSIST, 846 FSUBR, 1022
AND, 10, 304, 307, 320, 334, 368, 1013, 1017 FSUBRP, 1022
BSF, 417, 1017 FUCOM, 254, 1022
BSR, 1017 FUCOMI, 256
BSWAP, 461, 1017 FUCOMP, 1022
BT, 1017 FUCOMPP, 254, 1022
BTC, 316, 1017 FWAIT, 217
BTR, 316, 778, 1017 FXCH, 982, 1022
BTS, 316, 1017 IDIV, 449, 497, 1019
CALL, 9, 30, 720, 753, 856, 929, 1013 IMUL, 97, 299, 449, 450, 599, 1013, 1024
CBW, 449, 1017 IN, 720, 835, 889, 1019
CDQ, 401, 449, 1017 INC, 202, 980, 1013, 1024
CDQE, 449, 1017 INT, 33, 888, 1019
CLD, 1017 INT3, 696
CLI, 1017 IRET, 1019, 1020
CMC, 1017 JA, 124, 255, 449, 1013, 1024
CMOVcc, 135, 143, 144, 147, 150, 462, 1017 JAE, 124, 1013, 1024
CMP, 86, 470, 1013, 1024 JB, 124, 449, 1013, 1024
CMPSB, 704, 1017 JBE, 124, 1013, 1024
CMPSD, 1017 JC, 1013
CMPSQ, 1017 Jcc, 96, 146
CMPSW, 1017 JCXZ, 1013
COMISD, 434 JE, 155, 1013, 1024
COMISS, 436 JECXZ, 1013
CPUID, 366, 1019 JG, 124, 449, 1013
CWD, 449, 647, 899, 1017 JGE, 124, 1013
CWDE, 449, 1017 JL, 124, 449, 1013
DEC, 202, 1013, 1024 JLE, 124, 1013
DIV, 449, 1019 JMP, 30, 41, 54, 753, 980, 1013
DIVSD, 426, 707 JNA, 1013
FABS, 1021 JNAE, 1013
FADD, 1021 JNB, 1013
FADDP, 219, 225, 1021 JNBE, 255, 1013
FATRET, 329, 330 JNC, 1013
FCHS, 1021 JNE, 86, 124, 1013, 1024
FCMOVcc, 256 JNG, 1013
FCOM, 243, 254, 1021 JNGE, 1013
FCOMP, 231, 1021 JNL, 1013
FCOMPP, 1021 JNLE, 1013
FDIV, 218, 706, 1021 JNO, 1013, 1024
FDIVP, 219, 1021 JNS, 1013, 1024
FDIVR, 225, 1021 JNZ, 1013
FDIVRP, 1021 JO, 1013, 1024
FDUP, 676 JP, 232, 1013, 1024
FILD, 1021 JPO, 1013
FIST, 1022 JRCXZ, 1013
FISTP, 1022 JS, 1013, 1024
FLD, 229, 231, 1022 JZ, 94, 155, 981, 1013
FLD1, 1022 LAHF, 1014
FLDCW, 1022 LEA, 67, 99, 347, 472, 484, 501, 729, 790,
FLDZ, 1022 856, 1014
FMUL, 219, 1022 LEAVE, 11, 1014
FMULP, 1022 LES, 840, 898
1045
LOCK, 777 SYSCALL, 1019, 1021
LODSB, 889 SYSENTER, 739, 1019, 1021
LOOP, 183, 199, 708, 898, 1020 TEST, 200, 304, 307, 334, 1017
MAXSD, 434 UD2, 1021
MOV, 8, 10, 12, 510, 511, 720, 751, 856, 929, XADD, 778
980, 1015 XCHG, 1015, 1021
MOVDQA, 408 XOR, 10, 86, 206, 517, 707, 831, 980, 1017,
MOVDQU, 408 1024
MOVSB, 1015 MMX, 404
MOVSD, 433, 512, 1015 Prefixes
MOVSDX, 433 LOCK, 778, 1012
MOVSQ, 1015 REP, 1012, 1015, 1016
MOVSS, 436 REPE/REPNE, 1012
MOVSW, 1015 REPNE, 1016
MOVSX, 200, 208, 360–362, 449, 1015 Registers
MOVSXD, 283 AF, 444
MOVZX, 201, 345, 819, 1015 AH, 1014, 1015
MUL, 449, 450, 599, 1015 CS, 986
MULSD, 426 DF, 624
NEG, 504, 1015 DR6, 1011
NOP, 484, 980, 1015, 1024 DR7, 1011
NOT, 206, 208, 1015 DS, 986
OR, 307, 526, 1015 EAX, 86, 106
OUT, 720, 835, 1020 EBP, 67, 97
PADDD, 408 ECX, 542
PCMPEQB, 417 ES, 898, 986
PLMULHW, 405 ESP, 42, 67
PLMULLD, 405 Flags, 86, 126, 1008
PMOVMSKB, 417 FS, 735
POP, 9, 29, 30, 1015, 1024 GS, 280, 735, 738
POPA, 1020, 1024 JMP, 173
POPCNT, 1020 RIP, 741
POPF, 889, 1020 SS, 986
PUSH, 9, 11, 29, 30, 67, 720, 856, 929, 1015, ZF, 86, 304
1024 SSE, 405
PUSHA, 1020, 1024 SSE2, 405
PUSHF, 1020 x86-64, 14, 15, 50, 66, 72, 93, 99, 418, 425, 721,
PXOR, 417 727, 741, 1005, 1011
RCL, 708, 1020 Xcode, 18
RCR, 1020 XML, 701, 840
RET, 6, 7, 10, 30, 279, 544, 644, 980, 1015 XOR, 845
ROL, 330, 981, 1020
ROR, 981, 1020 Z80, 444
SAHF, 254, 1015 zlib, 625, 843
SAL, 636, 1020 Zobrist hashing, 459
SAR, 334, 449, 517, 636, 898, 1021 ZX Spectrum, 454
SBB, 394, 1015
SCASB, 890, 1016
SCASD, 1016
SCASQ, 1016
SCASW, 1016
SET, 467
SETcc, 137, 201, 255, 1021
SHL, 212, 265, 334, 636, 1016
SHR, 216, 334, 368, 636, 1016
SHRD, 400, 1016
STC, 1021
STD, 1021
STI, 1021
STOSB, 496, 1016
STOSD, 1016
STOSQ, 510, 1016
STOSW, 1016
SUB, 10, 86, 155, 470, 501, 1013, 1017
1046

Re4b en PDF

Uploaded by

Copyright:

Available Formats

Re4b en PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Re4b en PDF

Uploaded by

Copyright:

Available Formats

This is my own bulletin board

Can anyone help me with Low Fragmentation Heap in Windows?

Survey: how do you use DLL injection aside of API interception?

Why two titles? Read here: on page xiii.

Call for translators!

2 Important fundamentals 443

3 Slightly more advanced examples 464

5 Finding important/interesting stuff in the code 692

8 Case studies 784

9 Examples of reversing proprietary file formats 908

10 Dynamic binary instrumentation 972

11 Other things 980

12 Books/blogs worth reading 998

Acronyms Used 1033

2 Important fundamentals 443

3 Slightly more advanced examples 464

5 Finding important/interesting stuff in the code 692

8 Case studies 784

10 Dynamic binary instrumentation 972

11 Other things 980

12 Books/blogs worth reading 998

Acronyms Used 1033

About reverse engineering

Exercises and tasks

Praise for this book

About the Persian/Farsi translation

About the Chinese translation

1.1 The method

Optimization levels and debug information

A couple of words about different ISAs

Nowadays octal numbers seem to be used

Alan A. A. Donovan, Brian W. Kernighan —

1.2.3 Converting From One Radix To Another

hexadecimal binary decimal

decimal binary meaning

So each bit is mapped to a flag: read/write/execute.

Multi-Precision Arithmetic and Radix

How to Pronounce Non-Decimal Numbers

Numbers in a non-decimal base are usually pronounced by digit by digit: “one-zero-zero-one-one-...”.

Floating point numbers

1.3 An Empty Function

Listing 1.1: C/C++ Code

Let’s compile it!

Listing 1.2: Optimizing GCC/MSVC (assembly output)

Listing 1.4: Optimizing GCC 4.4.5 (assembly output)

…while IDA13 does it by pseudo name:

Listing 1.5: Optimizing GCC 4.4.5 (IDA)

A Note About MIPS Instructions and Register Names

1.3.4 Empty Functions in Practice

Listing 1.6: C/C++ code

dbg_print ("we did something\n");

Listing 1.7: C/C++ code

1.4 Returning Values

Listing 1.8: C/C++ Code

Let’s compile it.

Listing 1.9: Optimizing GCC/MSVC (assembly output)

Listing 1.11: Optimizing GCC 4.4.5 (assembly output)

…while IDA does it by their pseudo names:

Listing 1.12: Optimizing GCC 4.4.5 (IDA)