PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and...

34
PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences Tokyo Institute of Technology

description

PLC '06 Contribution Systematic test whether the optimizations are correct Not the proof of optimizers Confirm that the optimizers are correct in fine granularity and with relatively high reliability Compare the programs before and after optimization Values of all the left-hand sides of assignment statements, etc. are output to trace file Compare the trace file In case of error, error messages are given together with the intermediate code shown in C-style, which makes finding the cause of errors easily We found four bugs through experiments, including two unknown bugs Time for testing may be long, but we regard it as not important

Transcript of PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and...

Page 1: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Experience in Testing Compiler Optimizers Using Comparison Checking

Masataka Sassa and Daijiro SudoDept. of Mathematical and

Computing SciencesTokyo Institute of Technology

Page 2: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Background

Compiler Optimization Improves time efficiency of object code Optimization is a program transformation

Should not change the meaning of programs Algorithm error Implementation error

Testing Optimizers Incorrect optimizations cause different output

Difficult to identify the place of incorrect optimization Even incorrect optimizations may not change the

output (seems correct at first glance) It will cause incorrect result in other untested programs

Page 3: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Contribution Systematic test whether the optimizations are correct

Not the proof of optimizers Confirm that the optimizers are correct in fine granularity and

with relatively high reliability Compare the programs before and after optimization

Values of all the left-hand sides of assignment statements, etc. are output to trace file

Compare the trace file In case of error, error messages are given together with the

intermediate code shown in C-style, which makes finding the cause of errors easily

We found four bugs through experiments, including two unknown bugs

Time for testing may be long, but we regard it as not important

Page 4: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Testing Optimizers through Comparison Checking

First proposed by Jaramillo [2002], but no enough experience

Our paper focuses on its experience in a real development of an optimizing compiler COINS (http://coins-project.org, around 230 kloc) to test its SSA optimizers

Page 5: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Outline of the method 1. Duplicate intermediate code (no. of optimization + 1)

source

front-end

intermediate code

copy interm code

interm code

loop transformation

to SSA

loop analysis

insert trace output

SSA back-trans

no opt

runvar info

comparison check

display error

object code

interm code

loop transformation

to SSA

loop analysis

insert trace output

SSA back-trans

optimize1

runvar info

object code

interm code

loop transformation

to SSA

loop analysis

insert trace output

SSA back-trans

optimize1/2

runvar info

object code

interm code

loop transformation

to SSA

loop analysis

insert trace output

SSA back-trans

optimize1/2/3

runvar info

object code

comparison check

display error

comparison check

display error

comparison check

display error

Page 6: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Static Single Assignment ( SSA ) Form

Definitions of variables are unique in the program text.Each variable definition is distinguished by attaching a suffix.Introduce the -function at the join point where different definitions of a variable reach.

a:=a:=

:=a

a2:=a1:=

a3:=φ(a1,a2):=a3

a:=x+ya:=a+3b:=x+y

normal form normal formSSA form SSA form

a1:=x0+y0a2:=a1+3b1:=x0+y0

Page 7: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

2. Loop transformation on natural loops

a>5L1

b>10L2

c=a+bc=3

d=a+ba=0

L3

a=1L4

printf(c)L5

a=10b=20c=0d=0

L0

a>5

c=a+bc=3

d=a+ba=0

a=1

printf(c)

b>10

L1

L2

L3L4

L5

a=10b=20c=0d=0

L0

Lt

Page 8: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

3. SSA translation

a>5

c=a+bc=3

d=a+ba=0

a=1

printf(c)

b>10

L1

L2

L3L4

L5

a=10b=20c=0d=0

L0

Lt

a_1=(a_0,a_4)c_1=(c_0,c_4)

d_1=(d_0,d_3)a_1>5

c_2=a_1+b_0c_3=3

d_2=a_1+b_0a_2=0

a_3=1

printf(c_1)

b_0>10

L1

L2

L3L4

L5

a_0=10b_0=20c_0=0d_0=0

L0

a_4= (a_2,a_3)c_4= (c_3,c_1)d_3= (d_2,d_1)

Lt

Page 9: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

4. SSA form optimization (1) Loop invariant code motion

a_1= (a_0,a_4)c_1=(c_0,c_4)d_1= (d_0,d_3)

a_1>5

c_2=a_1+b_0d_2=a_1+b_0

printf(c_1)

b_0>10

L1

L2

L3 L4

L5

a_0=10b_0=20c_0=0d_0=0c_3=3a_2=0a_3=1

L0

a_4= (a_2,a_3)c_4= (c_3,c_1)d_3= (d_2,d_1)

Lt

a_1= (a_0,a_4)c_1= (c_0,c_4)

d_1= (d_0,d_3)a_1>5

c_2=a_1+b_0c_3=3

d_2=a_1+b_0a_2=0

a_3=1

printf(c_1)

b_0>10

L1

L2

L3L4

L5

a_0=10b_0=20c_0=0d_0=0

L0

a_4= (a_2,a_3)c_4= (c_3,c_1)d_3= (d_2,d_1)

Lt

Page 10: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

4. SSA form optimization(2) Common subexpression elimination

a_1= (a_0,a_4)c_1= (c_0,c_4)d_1= (d_0,d_3)

a_1>5

c_2=a_1+b_0d_2=a_1+b_0

printf(c_1)

b_0>10

L1

L2

L3 L4

L5

a_0=10b_0=20c_0=0d_0=0c_3=3a_2=0a_3=1

L0

a_4= (a_2,a_3)c_4= (c_3,c_1)d_3= (d_2,d_1)

Lt

a_1= (a_0,a_4)c_1= (c_0,c_4)d_1= (d_0,d_3)

a_1>5

c_2=a_1+b_0d_2=c_2

printf(c_1)

b_0>10

L1

L2

L3 L4

L5

a_0=10b_0=20c_0=0d_0=0c_3=3a_2=0a_3=1

L0

a_4= (a_2,a_3)c_4= (c_3,c_1)d_3= (d_2,d_1)

Lt

Page 11: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

4. SSA form optimization(3) Dead code elimination

a_1= (a_0,a_4)c_1= (c_0,c_4)

a_1>5

printf(c_1)

b_0>10

L1

L2

L3 L4

L5

a_0=10b_0=20c_0=0

c_3=3a_2=0a_3=1

L0

a_4= (a_2,a_3)c_4= (c_3,c_1)

Lt

a_1= (a_0,a_4)c_1= (c_0,c_4)d_1= (d_0,d_3)

a_1>5

c_2=a_1+b_0d_2=c_2

printf(c_1)

b_0>10

L1

L2

L3 L4

L5

a_0=10b_0=20c_0=0d_0=0c_3=3a_2=0a_3=1

L0

a_4= (a_2,a_3)c_4= (c_3,c_1)d_3= (d_2,d_1)

Lt

Page 12: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Outline of the method (cont.)

5. Loop analysis6. Insertion of statements for outputting

trace The value of the left-hand side variable

after each assignment statement The values of both sides of relational

operator before each conditional expression

The values of parameters before and after each function call

Page 13: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Outline of the method (cont.)

7. SSA back-translation and C-style program output◆ Output intermediate code (in SSA form)

in a C-style program

8. Execution Give each duplicated compiler the same

input and execute

Page 14: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

8. Executionsource

front-end

intermediate code

copy interm code

interm code

loop transformation

to SSA

loop analysis

insert trace output

SSA back-trans

no opt

runvar info

comparison check

display error

object code

interm code

loop transformation

to SSA

loop analysis

insert trace output

SSA back-trans

optimize1

runvar info

object code

interm code

loop transformation

to SSA

loop analysis

insert trace output

SSA back-trans

optimize1/2

runvar info

object code

interm code

loop transformation

to SSA

loop analysis

insert trace output

SSA back-trans

optimize1/2/3

runvar info

object code

comparison check

display error

comparison check

display error

comparison check

display error

Page 15: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Example of trace file

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

(i) basic block number(ii) instruction number in the basic block(iii) variable name, parameters etc.(iv) value (of variable name etc.).

Page 16: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Output of intermediate code in C-style (part)

Page 17: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

9. Comparison Checking Result of trace (Error messages are output if the values do not match)

(a) no optimization (d) dead code elimination(b)loop invariant code motion

(c) common subexpresssionelimination

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 0_5:c_3:3 0_6:a_2:0 0_7:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:d_2:30 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.1

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:c_3:3 3_3:d_2:30 3_4:a_2:0 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.0

Loop Lv.1

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 0_5:c_3:3 0_6:a_2:0 0_7:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:d_2:30 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.0

Loop Lv.1

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:c_3:3 0_5:a_2:0 0_6:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 t_1:a_4:0 t_2:c_4:3 1_1:a_1:0 1_2:c_1:3 1_3:JUMP_l:0 1_3:JUMP_r:5 </L0>

Loop Lv.0

Loop Lv.1

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 0_5:c_3:3 0_6:a_2:0 0_7:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:d_2:30 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.0

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 0_5:c_3:3 0_6:a_2:0 0_7:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:d_2:30 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.0

Page 18: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

How to compare instructions moved by code motion optimization

In loop invariant code motion and partial redundancy elimination, code is moved.

The system compares instructions moved out of the loop after optimization with all instructions in the corresponding loop before optimization, by looking at the tags made by the loop analysis

Ex: in comparing figures (a) and (b), c_3 and a_2 are moved out of the loop

Page 19: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

9. Comparison Checking Result of trace (Error messages are output if the values do not match)

(a) no optimization (d) dead code elimination(b)loop invariant code motion

(c) common subexpresssionelimination

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 0_5:c_3:3 0_6:a_2:0 0_7:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:d_2:30 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.1

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:c_3:3 3_3:d_2:30 3_4:a_2:0 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.0

Loop Lv.1

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 0_5:c_3:3 0_6:a_2:0 0_7:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:d_2:30 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.0

Loop Lv.1

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:c_3:3 0_5:a_2:0 0_6:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 t_1:a_4:0 t_2:c_4:3 1_1:a_1:0 1_2:c_1:3 1_3:JUMP_l:0 1_3:JUMP_r:5 </L0>

Loop Lv.0

Loop Lv.1

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 0_5:c_3:3 0_6:a_2:0 0_7:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:d_2:30 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.0

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 0_5:c_3:3 0_6:a_2:0 0_7:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:d_2:30 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.0

Page 20: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

How to compare instructions deleted or added by optimization

We do not compare variables deleted or added by optimization.

We compare only those variables that retain correspondence.

Ex: in comparing figures (c) and (d), d_0, d_1, d_2, d_3, c_2 are deleted by dead code elimination, and are not compared.

Page 21: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

9. Comparison Checking Result of trace (Error messages are output if the values do not match)

(a) no optimization (d) dead code elimination(b)loop invariant code motion

(c) common subexpresssionelimination

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 0_5:c_3:3 0_6:a_2:0 0_7:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:d_2:30 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.1

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:c_3:3 3_3:d_2:30 3_4:a_2:0 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.0

Loop Lv.1

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 0_5:c_3:3 0_6:a_2:0 0_7:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:d_2:30 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.0

Loop Lv.1

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:c_3:3 0_5:a_2:0 0_6:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 t_1:a_4:0 t_2:c_4:3 1_1:a_1:0 1_2:c_1:3 1_3:JUMP_l:0 1_3:JUMP_r:5 </L0>

Loop Lv.0

Loop Lv.1

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 0_5:c_3:3 0_6:a_2:0 0_7:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:d_2:30 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.0

0_1:a_0:10 0_2:b_0:20 0_3:c_0:0 0_4:d_0:0 0_5:c_3:3 0_6:a_2:0 0_7:a_3:1 5_1:CALL_printf_1_1_bef:3 5_1:CALL_printf_1_1_aft:3

<L0> 1_1:a_1:10 1_2:c_1:0 1_3:d_1:0 1_4:JUMP_l:10 1_4:JUMP_r:5 2_1:JUMP_l:20 2_1:JUMP_r:10 3_1:c_2:30 3_2:d_2:30 t_1:a_4:0 t_2:c_4:3 t_3:d_3:30 1_1:a_1:0 1_2:c_1:3 1_3:d_1:30 1_4:JUMP_l:0 1_4:JUMP_r:5 </L0>

Loop Lv.0

Page 22: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Example trace output with errors

0_1:a:10_2:t_0:10_3:b:30_4:t_1:30_5:t_2:10_6:t_3:00_7:t_4:00_8:c:30_9:t_5:10_10:t_6:30_11:t_7:30_12:t_8:60_13:t_9:10_14:t_10:60_15:x:7

0_1:a:10_2:t_0:10_3:b:30_4:t_1:30_5:t_2:10_6:t_3:00_7:t_4:00_8:c:30_9:t_5:10_10:t_6:30_11:t_7:00_12:t_8:30_13:t_9:10_14:t_10:30_15:x:4

Errors in t_7, t_8, t_10, x

before optimization after optimization (CSEQP)

Page 23: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Example of error message

-----NG!-----opt::LoopLevel=0,BlkID=0,InstrNumber=11,t_7:0unopt::LoopLevel=0,BlkID=0,InstrNumber=11,t_7:3------------------NG!-----opt::LoopLevel=0,BlkID=0,InstrNumber=12,t_8:3unopt::LoopLevel=0,BlkID=0,InstrNumber=12,t_8:6------------------NG!-----opt::LoopLevel=0,BlkID=0,InstrNumber=14,t_10:3unopt::LoopLevel=0,BlkID=0,InstrNumber=14,t_10:6------------------NG!-----opt::LoopLevel=0,BlkID=0,InstrNumber=15,x:4unopt::LoopLevel=0,BlkID=0,InstrNumber=15, x:7-------------

Page 24: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Experiments

We used the COINS C compiler. We implemented our system in Java. Approximately 5000 lines were added,

of which 2000 lines are for comparison checker.

Page 25: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Tested optimizations

Tested optimizations are: copy propagation, constant propagation, common

subexpression elimination, CSEQP, partial redundancy elimination, loop invariant code motion, strength reduction and test replacement of induction variables, dead code elimination, removal of redundant -functions, and removal of empty basic blocks

The test programs used are: 700 small test programs 181.mcf from SPEC CINT2000.

Page 26: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Results

We found four bugs in the COINS' SSA optimizers: Two for PRE (partial redundancy elimination)

(unknown bugs) One for CSEQP (common subexpression

elimination based on question propagation) One for operator strength reduction and test

replacement of induction variable

Page 27: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

A bug in partial redundancy elimination

Page 28: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

False alarm

False alarm is a situation where comparison checking displays errors although the optimization is correct.

Major cause: Values representing addresses that are dynamically allocated at runtime. They may change if optimizations are applied.

Errors concerning source level integers usually have a small number of digits.

We provided an option to let the user specify the limiting number of digits for checking integer values.

By specifying this limit as seven digits, the ratio of false alarm becomes approximately 5%.

In practice, values of addresses that are a false alarm can be easily recognized.

Page 29: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Conclusions

Experience in testing compiler optimizers using comparison checking Almost all program values before and after optimizations are

output as a trace and checked Applied to the SSA optimizers of a real project COINS

Could find bugs in optimizers Four bugs of optimizers including two unknown bugs The false alarm can be made as low as approx. 5%

Can find erroneous optimization point easily Error message indicates the critical point Intermediate code in C-style helps locate the point Quite beneficial in practice

Page 30: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Page 31: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Difference with Jaramillo’s work

1. Alternate execution is not used and no breakpoints are used

2. Trace files are generated3. The intermediate code is displayed in C-style

program in the case of error4. False alarms by pointers are well managed5. SSA form is used 6. Other implementation issues, such as handling of

natural loops, checking conditional expressions, and handling of each type of optimization are described

Page 32: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Related work Method by Necula et al.

Validation using symbolic execution Found bugs of gcc2.7.2.2

loop expansion, register allocation Method by Zuck, Pnueli et al.

Make correspondence to computation model Confirm using validation tool

Method by Okuma and Minamide Prove definition of each pass using a theorem

prover Generate compiler programs automatically

Page 33: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

Trace file

A trace of the values of variables etc. called “Variable information” is taken at execution time before and after optimization

(i) basic block number (ii) instruction number in the basic block (iii) variable name etc. (iv) value (of variable name etc.).

Values of left-hand side of assignment statements Values of both sides of the comparison of conditional

expressions Values of parameters of function calls

Page 34: PLC '06 Experience in Testing Compiler Optimizers Using Comparison Checking Masataka Sassa and Daijiro Sudo Dept. of Mathematical and Computing Sciences.

PLC '06

_L2:((int)(*(((int *)(unsigned char *)&(a))))) = ((int)( 1));divexI32_0 = ((int)(((int)(*(((int *)(unsigned char *)&(a)))))));((int)(*(((int *)(unsigned char *)&(b))))) = ((int)(((int)(divexI32_0 + 2))));divexI32_1 = ((int)(((int)(*(((int *)(unsigned char *)&(b)))))));divexI32_2 = ((int)(((int)(*(((int *)(unsigned char *)&(a)))))));divexI32_3 = ((int)(((int)(*(((int *)(unsigned char *)&(c)))))));divexI32_4 = ((int)(((int)(divexI32_2 * divexI32_3))));((int)(*(((int *)(unsigned char *)&(c))))) = ((int)(((int)(divexI32_1 + divexI32_4))));divexI32_5 = ((int)(((int)(*(((int *)(unsigned char *)&(a)))))));divexI32_6 = ((int)(((int)(*(((int *)(unsigned char *)&(b)))))));divexI32_7 = ((int)(((int)(*(((int *)(unsigned char *)&(c)))))));divexI32_8 = ((int)(((int)(divexI32_6 + divexI32_7))));divexI32_9 = ((int)(((int)(*(((int *)(unsigned char *)&(a)))))));divexI32_10 = ((int)(((int)(divexI32_8 * divexI32_9))));((int)(*(((int *)(unsigned char *)&(x))))) = ((int)(((int)(divexI32_5 + divexI32_10))));

_L2:((int)(*(((int *)(unsigned char *)&(a))))) = ((int)( 1));divexI32_0 = ((int)(((int)(*(((int *)(unsigned char *)&(a)))))));((int)(*(((int *)(unsigned char *)&(b))))) = ((int)(((int)(divexI32_0 + 2))));divexI32_1 = ((int)(((int)(*(((int *)(unsigned char *)&(b)))))));divexI32_2 = ((int)(divexI32_0));divexI32_3 = ((int)(((int)(*(((int *)(unsigned char *)&(c)))))));divexI32_4 = ((int)(((int)(divexI32_2 * divexI32_3))));((int)(*(((int *)(unsigned char *)&(c))))) = ((int)(((int)(divexI32_1 + divexI32_4))));divexI32_5 = ((int)(divexI32_0));divexI32_6 = ((int)(divexI32_1));divexI32_7 = ((int)(divexI32_3));divexI32_8 = ((int)(((int)(divexI32_6 + divexI32_7))));divexI32_9 = ((int)(divexI32_0));divexI32_10 = ((int)(((int)(divexI32_8 * divexI32_9))));((int)(*(((int *)(unsigned char *)&(x))))) = ((int)(((int)(divexI32_5 + divexI32_10))));