Code Optimization

(最適化)

14th lecture, July 20, 2018

Language Theory and Compilers

http://www.sw.it.aoyama.ac.jp/2018/Compiler/lecture14.html

Martin J. Dürst

© 2005-18 Martin J. Dürst 青山学院大学

Today's Schedule

• Remaining schedule
• Summary, homework, and leftovers from last time
• Code optimization
  • Goals, requirements
  • Methods
  • Techniques
• 授業改善のための学生アンケート

Remaining Schedule

• July 20: Optimization
• July 24 (Friday lectures on Tuesday): Executing environment: virtual machines, garbage collection,...
• July 27, 11:10-12:35: Term final exam
  

Summary of Previous Lecture

• Code generation is machine dependent
• The resulting code can be represented in assembly language
• There are no machine instructions for conditionals, if/for/... statements
  ⇒ we have to use conditional jump instructions
• jump uses comparison with 0 as a condition
• In C, jump can be expressed with goto
• Restricted C is an intermediate representation between (full) C and assembly language

Homework from Two Weeks Ago

For the Turing machine given by the following state transition table:

1. Draw the state transition diagram for this machine
2. Show in detail how this machine processes the input ..._1101000_...
3. Guess and explain what kind of calculation this machine does if the tape contains only a single contiguous sequence of '0'es and '1'es (surrounded by blanks) and where the leftmost non-blank symbol is a '1'

(this Turing machine always starts on the rightmost non-blank symbol)

Homework 1: Code Generation for Logical And

C fragment:

if (c<4 && f>=12)
    a=d;

Use hint: Conversion to nested if statement:

if (c<4)
    if (f>=12)
        a = d;

Homework 1: Example Solution

Homework 2: Code Generation for for Statement

(bonus problem)

C fragment:

for (i=0; i<20; i++)
    x *= y;

Preparation (use hint): Conversion to while statement:

i = 0;
while (i<20) {
    x *= y;
    i++;
}

Homework 2: Example Solution

Goals of Optimization

• Make program execution faster
• Reduce code size
• Additional considerations:
  • Program functionality (do not change meaning)
  • Speed of compilation
  • Ease of debugging
  • Properties of target CPU/machine

Compilers provide options to choose different levels of optimization

Optimization Methods

• Peephole optimization: Local exchange of instructions,...
  
• Control flow analysis
  • Split program into basic blocks:
    • No jumps from outside
    • No jumps to other locations
    → linear control flow inside block
  • Create a graph where:
    • Basic blocks are nodes
    • Jumps from (the end of) one block to (the start of) another are directed edges
  • Example: Control flow analysis for Homework 2
• Data flow analysis
  Use control flow graph to analyse which variable/register assignments affect which variable/register usages
  

Optimization Techniques: Faster and Smaller

Repeatedly try to use different techniques to continuously optimize the code further
(good example for Ruby implementation (video))

• Constant folding
  24 * 60 * 60 ⇒ 86400
• Constant propagation
  a = 3; b = a * 4; ⇒ a = 3; b = 3 * 4;
• Move common code out of loops
  while (a<500) { a += b*c*d; }
  ⇒ e = b*c*d; while (a<500) { a += e; }
• Dead code elimination: Eliminate code that will never be executed
• Reuse of register values
  STORE a, R1
LOAD R1, a
  ⇒
  STORE a, R1

Optimization Techniques: Faster but Larger

Make execution faster even if the amout of code may increase

• Function inlining:

  double square(double x) { return x*x; }
  square(a);

  ⇒a*a
• Loop unrolling:
  • Unrolling a loop with a small number of repetitions
    for (i=0; i<5; i++) a+=b;

    ⇒ a = a+b+b+b+b+b;

  • Partial unrolling (e.g. 20 repetitions are split into 4 repetitions of 5 calculations)
    for (i=0; i<20; i++) a+=b;
    ⇒
    for (i=0; i<20; i+=5)
    a = a+b+b+b+b+b;

Optimization Techniques: Machine Dependent

Highly dependent on machine type

• Exchange of instructions (different execution time for different instructions)
  Example: x*2 ⇒ x+x or x<<1
  Example: x*5 ⇒ x<<2 + x
• Instruction reordering (example: instructions such as LOAD may take longer, but run in parallel)

Example of Instruction Reordering

Expression: 743 * a

Before optimization:

        CONST   R1, 743
        LOAD    R2, a
                           ; invisible waiting time
        MUL     R3, R1, R2

After optimization:

        LOAD    R2, a
        CONST   R1, 743    ; executed while LOAD still in progress
        MUL     R3, R1, R2

Dynamic Compilation

(also: Just-in-time compilation)

• Check execution count of functions and blocks
• Check use of frequent parameter values (e.g. 0, 1)
• Recompile to take advantage of this knowledge

Examples of Actual Optimization

source

without optimization (gcc -O0 -S code.c)

with optimization (gcc -O1 -S code.c)

more optimized (gcc -O2 -S code.c)

much more optimized (gcc -O3 -S code.c, with parallel processing instructions)

(assembly language for Intel PCs (CISC))

Optimization options for gcc: English, Japanese

授業改善のための学生アンケート

WEB アンケート

お願い: 自由記述に必ず良かった点、問題点を具体的に書きましょう

(悪い例: 発音が分かりにくい; 良い例: さ行が濁っているかどうか分かりにくい)

Homework

(no need to submit)

Prepare for term final exam. In particular, have a look at some optimization problems.

Glossary

peephole optimization

ピープホール最適化

control flow analysis

制御フロー解析

basic block

基本ブロック

data flow analysis

データフロー解析

constant folding

静式評価、定数たたみこみ

constant propagation

定数伝播

dead code elimination

無用命令の削除

function inlining

関数呼び出しの展開

loop unrolling

繰り返しの展開

dynamic compilation

動的コンパイル