find_program(FIND find)

find_program(FILE file)

find_program(GREP grep)

if((FIND) AND (FILE) AND (GREP))

add_test(

NAME "just_ascii"

COMMAND sh -c "${FIND} include src windows tools singleheader tests examples benchmark -path benchmark/checkperf-reference -prune -name '*.h' -o -name '*.cpp' -type f -exec ${FILE} '{}' \; |${GREP} -v ASCII || exit 0 && exit 1"

WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}

)

endif()

- All source code files (.h, .cpp) must be ASCII.

The simdjson library relies on the lookup2 algorithm for UTF-8 validation on x64 platforms.

This algorithm validate the length of multibyte characters (that each multibyte character has the right number of continuation characters, and that all continuation characters are part of a multibyte character).

This algorithm compares *expected* continuation characters with *actual* continuation bytes, and emits an error anytime there is a mismatch.

For example, in the string "𝄞₿֏ab", which has a 4-, 3-, 2- and 1-byte

characters, the file will look like this:

| Character | 𝄞 | | | | ₿ | | | ֏ | | a | b |

|-----------------------|----|----|----|----|----|----|----|----|----|----|----|

| Character Length | 4 | | | | 3 | | | 2 | | 1 | 1 |

| Byte | F0 | 9D | 84 | 9E | E2 | 82 | BF | D6 | 8F | 61 | 62 |

| is_second_byte | | X | | | | X | | | X | | |

| is_third_byte | | | X | | | | X | | | | |

| is_fourth_byte | | | | X | | | | | | | |

| expected_continuation | | X | X | X | | X | X | | X | | |

| is_continuation | | X | X | X | | X | X | | X | | |

The errors here are basically (Second Byte OR Third Byte OR Fourth Byte == Continuation):

- **Extra Continuations:** Any continuation that is not a second, third or fourth byte is not

part of a valid 2-, 3- or 4-byte character and is thus an error. It could be that it's just

floating around extra outside of any character, or that there is an illegal 5-byte character,

or maybe it's at the beginning of the file before any characters have started; but it's an

error in all these cases.

- **Missing Continuations:** Any second, third or fourth byte that *isn't* a continuation is an error, because that means

we started a new character before we were finished with the current one.

Because we want to know if a byte is the *second* (or third, or fourth) byte of a multibyte

character, we need to "shift the bytes" to find that out. This is what they mean:

- `is_continuation`: if the current byte is a continuation.

- `is_second_byte`: if 1 byte back is the start of a 2-, 3- or 4-byte character.

- `is_third_byte`: if 2 bytes back is the start of a 3- or 4-byte character.

- `is_fourth_byte`: if 3 bytes back is the start of a 4-byte character.

We use shuffles to go n bytes back, selecting part of the current `input` and part of the

`prev_input` (search for `.prev<1>`, `.prev<2>`, etc.). These are passed in by the caller

function, because the 1-byte-back data is used by other checks as well.

Once we have the right bytes, we have to get the masks. To do this, we treat UTF-8 bytes as

numbers, using signed `<` and `>` operations to check if they are continuations or leads.

In fact, we treat the numbers as *signed*, partly because it helps us, and partly because

Intel's SIMD presently only offers signed `<` and `>` operations (not unsigned ones).

In UTF-8, bytes that start with the bits 110, 1110 and 11110 are 2-, 3- and 4-byte "leads,"

respectively, meaning they expect to have 1, 2 and 3 "continuation bytes" after them.

Continuation bytes start with 10, and ASCII (1-byte characters) starts with 0.

When treated as signed numbers, they look like this:

| Type | High Bits | Binary Range | Signed |

|--------------|------------|--------------|--------|

| ASCII | `0` | `01111111` | 127 |

| | | `00000000` | 0 |

| 4+-Byte Lead | `1111` | `11111111` | -1 |

| | | `11110000 | -16 |

| 3-Byte Lead | `1110` | `11101111` | -17 |

| | | `11100000 | -32 |

| 2-Byte Lead | `110` | `11011111` | -33 |

| | | `11000000 | -64 |

| Continuation | `10` | `10111111` | -65 |

| | | `10000000 | -128 |

This makes it pretty easy to get the continuation mask! It's just a single comparison:

We can do something similar for the others, but it takes two comparisons instead of one: "is

the start of a 4-byte character" is `< -32` and `> -65`, for example. And 2+ bytes is `< 0` and

`> -64`. Surely we can do better, they're right next to each other!

Notice *why* continuations were a single comparison. The actual *range* would require two

comparisons--`< -64` and `> -129`--but all characters are always greater than -128, so we get

that for free. In fact, if we had *unsigned* comparisons, 2+, 3+ and 4+ comparisons would be

just as easy: 4+ would be `> 239`, 3+ would be `> 223`, and 2+ would be `> 191`.

Instead, we add 128 to each byte, shifting the range up to make comparison easy. This wraps

ASCII down into the negative, and puts 4+-Byte Lead at the top:

| Type | High Bits | Binary Range | Signed |

|----------------------|------------|--------------|-------|

| 4+-Byte Lead (+ 127) | `0111` | `01111111` | 127 |

| | | `01110000 | 112 |

|----------------------|------------|--------------|-------|

| 3-Byte Lead (+ 127) | `0110` | `01101111` | 111 |

| | | `01100000 | 96 |

|----------------------|------------|--------------|-------|

| 2-Byte Lead (+ 127) | `010` | `01011111` | 95 |

| | | `01000000 | 64 |

|----------------------|------------|--------------|-------|

| Continuation (+ 127) | `00` | `00111111` | 63 |

| | | `00000000 | 0 |

|----------------------|------------|--------------|-------|

| ASCII (+ 127) | `1` | `11111111` | -1 |

| | | `10000000` | -128 |

|----------------------|------------|--------------|-------|

*Now* we can use signed `>` on all of them:

NOTE: we use `^ 0x80` instead of `+ 128` in the code, which accomplishes the same thing, and even takes the same number

of cycles as `+`, but on many Intel architectures can be parallelized better (you can do 3

`^`'s at a time on Haswell, but only 2 `+`'s).

That doesn't look like it saved us any instructions, did it? Well, because we're adding the

same number to all of them, we can save one of those `+ 128` operations by assembling

`prev2_flipped` out of prev 1 and prev 3 instead of assembling it from input and adding 128

to it. One more instruction saved!

At this point, we have `is_continuation`, `is_first_byte`, `is_second_byte` and `is_third_byte`.

All we have left to do is check if they match!

But wait--there's more. The above statement is only 3 operations, but they *cannot be done in

parallel*. You have to do 2 `|`'s and then 1 `&`. Haswell, at least, has 3 ports that can do

bitwise operations, and we're only using 1!

There is one big case the above code doesn't explicitly talk about--what if is_second_byte

and is_third_byte are BOTH true? That means there is a 3-byte and 2-byte character right next

to each other (or any combination), and the continuation could be part of either of them!

Our algorithm using `&` and `|` won't detect that the continuation byte is problematic.

Never fear, though. If that situation occurs, we'll already have detected that the second

leading byte was an error, because it was supposed to be a part of the preceding multibyte

character, but it *wasn't a continuation*.

We could stop here, but it turns out that we can fix it using `+` and `-` instead of `|` and

`&`, which is both interesting and possibly useful (even though we're not using it here). It

exploits the fact that in SIMD, a *true* value is -1, and a *false* value is 0. So those

comparisons were giving us numbers!

Given that, if you do `is_second_byte + is_third_byte + is_fourth_byte`, under normal

circumstances you will either get 0 (0 + 0 + 0) or -1 (-1 + 0 + 0, etc.). Thus,

`(is_second_byte + is_third_byte + is_fourth_byte) - is_continuation` will yield 0 only if

*both* or *neither* are 0 (0-0 or -1 - -1). You'll get 1 or -1 if they are different. Because

*any* nonzero value is treated as an error (not just -1), we're just fine here :)

Further, if *more than one* multibyte character overlaps,

`is_second_byte + is_third_byte + is_fourth_byte` will be -2 or -3! Subtracting `is_continuation`

from *that* is guaranteed to give you a nonzero value (-1, -2 or -3). So it'll always be

considered an error.

One reason you might want to do this is parallelism. ^ and | are not associative, so

(A | B | C) ^ D will always be three operations in a row: either you do A | B -> | C -> ^ D, or

you do B | C -> | A -> ^ D. But addition and subtraction *are* associative: (A + B + C) - D can

be written as `(A + B) + (C - D)`. This means you can do A + B and C - D at the same time, and

then adds the result together. Same number of operations, but if the processor can run

independent things in parallel (which most can), it runs faster.

This doesn't help us on Intel, but might help us elsewhere: on Haswell, at least, | and ^ have

a super nice advantage in that more of them can be run at the same time (they can run on 3

ports, while + and - can run on 2)! This means that we can do A | B while we're still doing C,

saving us the cycle we would have earned by using +. Even more, using an instruction with a

wider array of ports can help *other* code run ahead, too, since these instructions can "get

out of the way," running on a port other instructions can't.

There's one more relevant trick up our sleeve, it turns out: it turns out on Intel we can "pay

for" the (prev<1> + 128) instruction, because it can be used to save an instruction in

check_special_cases()--but we'll talk about that there :)

/* auto-generated on Thu 21 May 2020 14:01:15 EDT. Do not edit! */

// For example, in the string "ab", which has a 4-, 3-, 2- and 1-byte

// | Character | | | | | | | | | | a | b |

// can't fit in binary64. The maximal value is about 1.7976931348623157 x

// are about 10^80 atoms in the universe. The estimate for the total number

// can't fit in binary64. The maximal value is about 1.7976931348623157 x

// are about 10^80 atoms in the universe. The estimate for the total number

// For example, in the string "ab", which has a 4-, 3-, 2- and 1-byte

// | Character | | | | | | | | | | a | b |

// can't fit in binary64. The maximal value is about 1.7976931348623157 x

// are about 10^80 atoms in the universe. The estimate for the total number

// For example, in the string "ab", which has a 4-, 3-, 2- and 1-byte

// | Character | | | | | | | | | | a | b |

// can't fit in binary64. The maximal value is about 1.7976931348623157 x

// are about 10^80 atoms in the universe. The estimate for the total number