Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.
Artale (1)
Principle of Compilers
Lecture IX: Principles of Code
Optimization
Alessandro Artale
Faculty of Computer Science – Free University of Bolzano
Room: 221
artale@inf.unibz.it
http://www.inf.unibz.it/ artale/
2003/2004 – Second Semester
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (2)
Summary of Lecture IX
Code Optimization
Basic Blocks and Flow Graphs
Sources of Optimization
1. Common Subexpression Elimination
2. Copy Propagation
3. Dead-Code Elimination
4. Constant Folding
5. Loop Optimization
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (3)
Code Optimization: Intro
Intermediate Code undergoes various transformations—called Optimiza-
tions—to make the resulting code running faster and taking less space.
Optimization never guarantees that the resulting code is the best possible.
We will consider only Machine-Independent Optimizations—i.e., they don’t
take into consideration any properties of the target machine.
The techniques used are a combination of Control-Flow and Data-Flow
analysis.
– Control-Flow Analysis. Identifies loops in the flow graph of a program
since such loops are usually good candidates for improvement.
– Data-Flow Analysis. Collects information about the way variables are
used in a program.
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (4)
Criteria for Code-Improving Transformations
The best transformations are those that yield the most benefit for the least
effort.
1. A transformation must preserve the meaning of a program. It’s better to
miss an opportunity to apply a transformation rather than risk changing
what the program does.
2. A transformation must, on the average, speed up a program by a
measurable amount.
3. Avoid code-optimization for programs that run occasionally or during
debugging.
4. Remember! Dramatic improvements are usually obtained by improving
the source code: The programmer is always responsible in finding the
best possible data structures and algorithms for solving a problem.
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (5)
Quicksort: An Example Program
We will use the sorting program Quicksort to illustrate the effects of the
various optimization techniques.
void quicksort(m,n)
int m,n;
int i,j,v,x;
if (n <= m) return;
i = m-1; j = n; v = a[n]; /* fragment begins here */
while (1)
do i = i+1; while (a[i]<v);
do j = j-1; while (a[j]>v);
if (i>=j) break;
x = a[i]; a[i] = a[j]; a[j] =x;
x = a[i]; a[i] = a[n]; a[n] =x; /* fragment ends here */
quicksort(m,j); quicksort(i+1,n);
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (6)
Quicksort: An Example Program (Cont.)
The following is the three-address code for a fragment of Quicksort.
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (7)
Summary
Code Optimization
Basic Blocks and Flow Graphs
Sources of Optimization
1. Common Subexpression Elimination
2. Copy Propagation
3. Dead-Code Elimination
4. Constant Folding
5. Loop Optimization
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (8)
Basic Blocks and Flow Graphs
The Machine-Independent Code-Optimization phase consists of control-flow
and data-flow analysis followed by the application of transformations.
During Control-Flow analysis, a program is represented as a Flow Graph
where:
– Nodes represent Basic Blocks: Sequence of consecutive statements in
which flow-of-control enters at the beginning and leaves at the end
without halt or branches;
– Edges represent the flow of control.
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (9)
Flow Graph:An Example
Flow graph for the three-address code fragment for quicksort. Each is a
basic block.
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (10)
Summary
Code Optimization
Basic Blocks and Flow Graphs
Sources of Optimization
1. Common Subexpression Elimination
2. Copy Propagation
3. Dead-Code Elimination
4. Constant Folding
5. Loop Optimization
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (11)
The Principal Sources of Optimization
We distinguish local transformations—involving only statements in a single
basic block—from global transformations.
A basic block computes a set of expressions: A number of transformations
can be applied to a basic block without changing the expressions computed
by the block.
1. Common Subexpressions elimination;
2. Copy Propagation;
3. Dead-Code elimination;
4. Constant Folding.
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (12)
Common Subexpressions Elimination
Frequently a program will include calculations of the same value.
An occurrence of an expression is called a common subexpression if
was previously computed, and the values of variables in have no changed
since the previous computation.
Example. Consider the basic block . The assignments to both and
have common subexpressions and can be eliminated.
After local common subexpression elimination, is transformed as:
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (13)
Common Subexpressions Elimination (Cont.)
Example (Cont.) After local elimination, still evaluates and
which are common subexpressions.
is evaluated in
by . Then, the statements
can be replaced by
Now, is also a common subexpression, computed in by . Then,
the statements
can be replaced by
.
Analogously, the value of is the same as the value assigned to in block
; while can be eliminated and replaced by .
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (14)
Common Subexpressions Elimination (Cont.)
Example. The following flow graph shows the result of eliminating both
local and global common subexpressions from basic blocks and .
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (15)
DAGs for Determining Common Subexpressions
To individuate common subexpressions we represent a basic block as a DAG
showing how expressions are re-used in a block.
A DAG for a Basic Block has the following labels and nodes:
1. Leaves contain unique identifiers, either variable names or constants.
2. Interior nodes contain an operator symbol.
3. Nodes can optionally be associated to a list of variables representing
those variables having the value computed at the node.
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (16)
DAGs for Blocks: An Example
The following shows both a three-address code of a basic block and its
associated DAG.
+ ,prod
prod
*
[] [] (1)
a + ,i
b * 20
4 i 1
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (17)
Summary
Code Optimization
Basic Blocks and Flow Graphs
Sources of Optimization
1. Common Subexpression Elimination
2. Copy Propagation
3. Dead-Code Elimination
4. Constant Folding
5. Loop Optimization
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (18)
Copy Propagation
Copy Propagation Rule: Given the copy statement x := y use y for x
whenever possible after the copy statement.
Copy Propagation applied to Block yields:
This transformation together with Dead-Code Elimination (see next slide)
will give us the opportunity to eliminate the assignment altogether.
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (19)
Dead-Code Elimination
A variable is live at a point in a program if its value can be used subsequently,
otherwise it is dead.
A piece of code is dead if data computed is never used elsewhere.
Dead-Code may appear as the result of previous transformation.
Dead-Code works well together with Copy Propagation.
Example. Considering the Block after Copy Propagation we can see that
x is never reused all over the code. Thus, x is a dead variable and we can
eliminate the assignment from .
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (20)
Constant Folding
Based on deducing at compile-time that the value of an expression (and in
particular of a variable) is a constant.
Constant Folding is the transformation that substitutes an expression with a
constant.
Constant Folding is useful to discover Dead-Code.
Example. Consider the conditional statement:
.
If, by Constant Folding, we discover that is always false we can eliminate
both the if-test and the jump to L.
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (21)
Summary
Code Optimization
Basic Blocks and Flow Graphs
Sources of Optimization
1. Common Subexpression Elimination
2. Copy Propagation
3. Dead-Code Elimination
4. Constant Folding
5. Loop Optimization
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (22)
Loop Optimization
The running time of a program can be improved if we decrease the amount
of instructions in an inner loop.
Three techniques are useful:
1. Code Motion
2. Reduction in Strength
3. Induction-Variable elimination
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (23)
Code Motion
If the computation of an expression is loop-invariant this transformation
places such computation before the loop.
Example. Consider the following while statement:
while (i limit - 2) do
The expression limit - 2 is loop invariant. Code motion transformation will
result in:
t := limit -2;
while (i t) do
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (24)
Reduction in Strength
Is based on the replacement of a computation with a less expensive one.
Example. Consider the assignment
in Block .
j is decremented by 1 each time, then
.
Thus, we may replace by
.
Problem: We need to initialize to before entering the Block .
– Result. The substitution of a multiplication by a subtraction will speed up
the resulting code.
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (25)
Induction Variables
A variable x is an Induction Variable of a loop if every time the variable x
changes values, it is incremented or decremented by some constant.
A common situation is one in which an induction variable, say i, indexes an
array, and some other induction variable, say t, is the actual offset to access
the array:
– Often we can get rid of i.
– In general, when there are two or more Induction Variables it is possible
to get rid of all but one.
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (26)
Induction Variables Elimination: An Example
Example. Consider the loop of Bock . The variables j and are Induction
Variables. The same applies for variables i and in Block .
After Reduction in Strength is applied to both and , the only use of i and
j is to determine the test in .
Since and the test is equivalent to .
After this replacement in the test, both i (in Block ) and j (in Block )
become dead-variables and can be eliminated! (see next slide for the new
optimized code).
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (27)
Induction Variables Elimination: An Example (Cont.)
Flow Graph after Reduction in Strength and Induction-Variables elimination.
Free University of Bolzano–Principles of Compilers. Lecture IX, 2003/2004 – A.Artale (28)
Summary of Lecture IX
Code Optimization
Basic Blocks and Flow Graphs
Sources of Optimization
1. Common Subexpression Elimination
2. Copy Propagation
3. Dead-Code Elimination
4. Constant Folding
5. Loop Optimization
