Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Why ObjectTextSizeOz is bigger than ObjectTextSizeO0 ？ #824

Open
sparks-code opened this issue Dec 9, 2023 · 6 comments
Open

Why ObjectTextSizeOz is bigger than ObjectTextSizeO0 ？ #824

sparks-code opened this issue Dec 9, 2023 · 6 comments
Labels
Question Further information is requested

Comments

@sparks-code
Copy link

sparks-code commented Dec 9, 2023 •
edited
Loading

❓ Questions and Help

I tested the code size in relation to the platform and this is my code. My compilergym version is 0.25 and the benchmark i used is cbench.

cbench = env1.datasets["benchmark://cbench-v1"]
cbench_list = list(cbench.benchmark_uris())
for bench in cbench_list:
    with compiler_gym.make("llvm-autophase-codesize-v0") as env:  
        env.reset(benchmark = bench)
        print(env.benchmark)
        print(f'''ObjectTextSize = {env.observation.ObjectTextSizeBytes()},
        ObjectTextSizeO0 = {env.observation.ObjectTextSizeO0()},
        ObjectTextSizeO3= {env.observation.ObjectTextSizeO3()},
        ObjectTextSizeOz = {env.observation.ObjectTextSizeOz()},
        ''')

this is result
image
I want to know why ObjectTextSizeOz is bigger than ObjectTextSizeO0 ？
I optimize adpcm program with -O3 and -Oz in the command line

~/.local/share/compiler_gym/llvm-v0/bin/clang -O0 -c adpcm.bc -o adpcm_O0.o
~/.local/share/compiler_gym/llvm-v0/bin/clang -Oz -c adpcm.bc -o adpcm_Oz.o
~/.local/share/compiler_gym/llvm-v0/bin/clang -O3 -c adpcm.bc -o adpcm_O3.o

and the result is
image
why the result is different from compilergym？Is this normal?

Additional Context
@sparks-code sparks-code added the Question Further information is requested label Dec 9, 2023
@sparks-code
Copy link
Author 

benchmark://npb-v0/1
ObjectTextSize = 1833,
        ObjectTextSizeO0 = 1833,
        ObjectTextSizeO3= 1849,
        ObjectTextSizeOz = 1849,
        
benchmark://npb-v0/2
ObjectTextSize = 3204,
        ObjectTextSizeO0 = 3204,
        ObjectTextSizeO3= 4751,
        ObjectTextSizeOz = 3833,
        
benchmark://npb-v0/3
ObjectTextSize = 2644,
        ObjectTextSizeO0 = 2644,
        ObjectTextSizeO3= 3031,
        ObjectTextSizeOz = 2735,
        
benchmark://npb-v0/4
ObjectTextSize = 4531,
        ObjectTextSizeO0 = 4531,
        ObjectTextSizeO3= 4973,
        ObjectTextSizeOz = 4341,
        
benchmark://npb-v0/5
ObjectTextSize = 4279,
        ObjectTextSizeO0 = 4279,
        ObjectTextSizeO3= 6679,
        ObjectTextSizeOz = 6262,
        
benchmark://npb-v0/6
ObjectTextSize = 4796,
        ObjectTextSizeO0 = 4796,
        ObjectTextSizeO3= 6755,
        ObjectTextSizeOz = 6737,
        
benchmark://npb-v0/7
ObjectTextSize = 665,
        ObjectTextSizeO0 = 665,
        ObjectTextSizeO3= 1576,
        ObjectTextSizeOz = 1692,
        
benchmark://npb-v0/8
ObjectTextSize = 5237,
        ObjectTextSizeO0 = 5237,
        ObjectTextSizeO3= 4350,
        ObjectTextSizeOz = 3921,
        
benchmark://npb-v0/9
ObjectTextSize = 603,
        ObjectTextSizeO0 = 603,
        ObjectTextSizeO3= 1328,
        ObjectTextSizeOz = 658,
        
benchmark://npb-v0/10
ObjectTextSize = 170,
        ObjectTextSizeO0 = 170,
        ObjectTextSizeO3= 189,
        ObjectTextSizeOz = 189,
        
benchmark://npb-v0/11
ObjectTextSize = 2640,
        ObjectTextSizeO0 = 2640,
        ObjectTextSizeO3= 5382,
        ObjectTextSizeOz = 3384,
        
benchmark://npb-v0/12
ObjectTextSize = 16140,
        ObjectTextSizeO0 = 16140,
        ObjectTextSizeO3= 30611,
        ObjectTextSizeOz = 26386,
        
benchmark://npb-v0/13
ObjectTextSize = 4423,
        ObjectTextSizeO0 = 4423,
        ObjectTextSizeO3= 24981,
        ObjectTextSizeOz = 15896,
        
benchmark://npb-v0/14
ObjectTextSize = 1809,
        ObjectTextSizeO0 = 1809,
        ObjectTextSizeO3= 1738,
        ObjectTextSizeOz = 1738,
        
benchmark://npb-v0/15
ObjectTextSize = 398,
        ObjectTextSizeO0 = 398,
        ObjectTextSizeO3= 330,
        ObjectTextSizeOz = 378,
        
benchmark://npb-v0/16
ObjectTextSize = 97,
        ObjectTextSizeO0 = 97,
        ObjectTextSizeO3= 112,
        ObjectTextSizeOz = 112,
        
benchmark://npb-v0/17
ObjectTextSize = 4279,
        ObjectTextSizeO0 = 4279,
        ObjectTextSizeO3= 6679,
        ObjectTextSizeOz = 6262,
        
benchmark://npb-v0/18
ObjectTextSize = 1148,
        ObjectTextSizeO0 = 1148,
        ObjectTextSizeO3= 4132,
        ObjectTextSizeOz = 3051,
        
benchmark://npb-v0/19
ObjectTextSize = 13025,
        ObjectTextSizeO0 = 13025,
        ObjectTextSizeO3= 12016,
        ObjectTextSizeOz = 10906,
        
benchmark://npb-v0/20
ObjectTextSize = 170,
        ObjectTextSizeO0 = 170,
        ObjectTextSizeO3= 189,
        ObjectTextSizeOz = 189,
        
benchmark://npb-v0/21
ObjectTextSize = 398,
        ObjectTextSizeO0 = 398,
        ObjectTextSizeO3= 330,
        ObjectTextSizeOz = 378,
        
benchmark://npb-v0/22
ObjectTextSize = 8268,
        ObjectTextSizeO0 = 8268,
        ObjectTextSizeO3= 11602,
        ObjectTextSizeOz = 10180,
        
benchmark://npb-v0/23
ObjectTextSize = 911,
        ObjectTextSizeO0 = 911,
        ObjectTextSizeO3= 646,
        ObjectTextSizeOz = 644,
        
benchmark://npb-v0/24
ObjectTextSize = 170,
        ObjectTextSizeO0 = 170,
        ObjectTextSizeO3= 189,
        ObjectTextSizeOz = 189,
        
benchmark://npb-v0/25
ObjectTextSize = 1833,
        ObjectTextSizeO0 = 1833,
        ObjectTextSizeO3= 1849,
        ObjectTextSizeOz = 1849,
        
benchmark://npb-v0/26
ObjectTextSize = 1809,
        ObjectTextSizeO0 = 1809,
        ObjectTextSizeO3= 1738,
        ObjectTextSizeOz = 1738,
        
benchmark://npb-v0/27
ObjectTextSize = 398,
        ObjectTextSizeO0 = 398,
        ObjectTextSizeO3= 330,
        ObjectTextSizeOz = 378,
        
benchmark://npb-v0/28
ObjectTextSize = 3884,
        ObjectTextSizeO0 = 3884,
        ObjectTextSizeO3= 5680,
        ObjectTextSizeOz = 5133,
        
benchmark://npb-v0/29
ObjectTextSize = 11026,
        ObjectTextSizeO0 = 11026,
        ObjectTextSizeO3= 14282,
        ObjectTextSizeOz = 12264,
        
benchmark://npb-v0/30
ObjectTextSize = 17889,
        ObjectTextSizeO0 = 17889,
        ObjectTextSizeO3= 23847,
        ObjectTextSizeOz = 21185,
        
benchmark://npb-v0/31
ObjectTextSize = 170,
        ObjectTextSizeO0 = 170,
        ObjectTextSizeO3= 189,
        ObjectTextSizeOz = 189,
        
benchmark://npb-v0/32
ObjectTextSize = 398,
        ObjectTextSizeO0 = 398,
        ObjectTextSizeO3= 330,
        ObjectTextSizeOz = 378,
        
benchmark://npb-v0/33
ObjectTextSize = 915,
        ObjectTextSizeO0 = 915,
        ObjectTextSizeO3= 712,
        ObjectTextSizeOz = 664,
        
benchmark://npb-v0/34
ObjectTextSize = 13357,
        ObjectTextSizeO0 = 13357,
        ObjectTextSizeO3= 9921,
        ObjectTextSizeOz = 9208,
        
benchmark://npb-v0/35
ObjectTextSize = 3647,
        ObjectTextSizeO0 = 3647,
        ObjectTextSizeO3= 3256,
        ObjectTextSizeOz = 3256,
        
benchmark://npb-v0/36
ObjectTextSize = 1133,
        ObjectTextSizeO0 = 1133,
        ObjectTextSizeO3= 4096,
        ObjectTextSizeOz = 3036,
        
benchmark://npb-v0/37
ObjectTextSize = 561,
        ObjectTextSizeO0 = 561,
        ObjectTextSizeO3= 1136,
        ObjectTextSizeOz = 566,
        
benchmark://npb-v0/38
ObjectTextSize = 4134,
        ObjectTextSizeO0 = 4134,
        ObjectTextSizeO3= 8208,
        ObjectTextSizeOz = 7100,
        
benchmark://npb-v0/39
ObjectTextSize = 14460,
        ObjectTextSizeO0 = 14460,
        ObjectTextSizeO3= 24136,
        ObjectTextSizeOz = 22412,
        
benchmark://npb-v0/40
ObjectTextSize = 1355,
        ObjectTextSizeO0 = 1355,
        ObjectTextSizeO3= 1082,
        ObjectTextSizeOz = 1048,
        
benchmark://npb-v0/41
ObjectTextSize = 170,
        ObjectTextSizeO0 = 170,
        ObjectTextSizeO3= 189,
        ObjectTextSizeOz = 189,
        
benchmark://npb-v0/42
ObjectTextSize = 18330,
        ObjectTextSizeO0 = 18330,
        ObjectTextSizeO3= 11743,
        ObjectTextSizeOz = 11560,
        
benchmark://npb-v0/43
ObjectTextSize = 1803,
        ObjectTextSizeO0 = 1803,
        ObjectTextSizeO3= 1934,
        ObjectTextSizeOz = 1839,
        
benchmark://npb-v0/44
ObjectTextSize = 1683,
        ObjectTextSizeO0 = 1683,
        ObjectTextSizeO3= 1580,
        ObjectTextSizeOz = 1580,
        
benchmark://npb-v0/45
ObjectTextSize = 943,
        ObjectTextSizeO0 = 943,
        ObjectTextSizeO3= 3125,
        ObjectTextSizeOz = 2647,
        
benchmark://npb-v0/46
ObjectTextSize = 832,
        ObjectTextSizeO0 = 832,
        ObjectTextSizeO3= 798,
        ObjectTextSizeOz = 798,
        
benchmark://npb-v0/47
ObjectTextSize = 3749,
        ObjectTextSizeO0 = 3749,
        ObjectTextSizeO3= 6345,
        ObjectTextSizeOz = 6068,
        
benchmark://npb-v0/48
ObjectTextSize = 321,
        ObjectTextSizeO0 = 321,
        ObjectTextSizeO3= 260,
        ObjectTextSizeOz = 260,
        
benchmark://npb-v0/49
ObjectTextSize = 911,
        ObjectTextSizeO0 = 911,
        ObjectTextSizeO3= 646,
        ObjectTextSizeOz = 644,
        
benchmark://npb-v0/50
ObjectTextSize = 13047,
        ObjectTextSizeO0 = 13047,
        ObjectTextSizeO3= 12093,
        ObjectTextSizeOz = 10991

In addition to Cbench, a similar situation occurs with many benchmarks in NPB

@sparks-code
Copy link
Author

@Celebio @hughleat @ryanrussell @colesbury @ChrisCummins

@sparks-code
Copy link
Author

On the other hand, I use 2 ways to test ,
1.
~/.local/share/compiler_gym/llvm-v0/bin/opt adpcm.bc -O0 -o adpcm_opt_O0.bc
~/.local/share/compiler_gym/llvm-v0/bin/clang -c adpcm_opt_O0.bc -o adpcm_opt_O0.o
~/.local/share/compiler_gym/llvm-v0/bin/llvm-size adpcm_opt_O0.o
text data bss dec hex filename
1901 432 0 2333 91d adpcm_opt_O0.o

~/.local/share/compiler_gym/llvm-v0/bin/opt adpcm.bc -Oz -o adpcm_opt_Oz.bc
~/.local/share/compiler_gym/llvm-v0/bin/clang -c adpcm_opt_Oz.bc -o adpcm_opt_Oz.o
~/.local/share/compiler_gym/llvm-v0/bin/llvm-size adpcm_opt_Oz.o
text data bss dec hex filename
2386 0 0 2386 952 adpcm_opt_Oz.o

~/.local/share/compiler_gym/llvm-v0/bin/opt adpcm.bc -O3 -o adpcm_opt_O3.bc
~/.local/share/compiler_gym/llvm-v0/bin/clang -c adpcm_opt_O3.bc -o adpcm_opt_O3.o
~/.local/share/compiler_gym/llvm-v0/bin/llvm-size adpcm_opt_O3.o
text data bss dec hex filename
2725 0 0 2725 aa5 adpcm_opt_O3.o

@sparks-code
Copy link
Author

sparks-code commented Dec 9, 2023 •
edited
Loading

~/.local/share/compiler_gym/llvm-v0/bin/clang -c -O0 adpcm.bc -o adpcm_O0.o
~/.local/share/compiler_gym/llvm-v0/bin/llvm-size adpcm_O0.o
text data bss dec hex filename
1901 432 0 2333 91d adpcm_O0.o

~/.local/share/compiler_gym/llvm-v0/bin/clang -c -Oz adpcm.bc -o adpcm_Oz.o
~/.local/share/compiler_gym/llvm-v0/bin/llvm-size adpcm_Oz.o
text data bss dec hex filename
1629 0 0 1629 65d adpcm_Oz.o

~/.local/share/compiler_gym/llvm-v0/bin/clang -c -O3 adpcm.bc -o adpcm_O3.o
~/.local/share/compiler_gym/llvm-v0/bin/llvm-size adpcm_O3.o
text data bss dec hex filename
1648 0 0 1648 670 adpcm_O3.o
Why are the text section sizes of these two methods so different?

@yingluosanqian
Copy link

I optimize adpcm program with -O3 and -Oz in the command line

~/.local/share/compiler_gym/llvm-v0/bin/clang -O0 -c adpcm.bc -o adpcm_O0.o
~/.local/share/compiler_gym/llvm-v0/bin/clang -Oz -c adpcm.bc -o adpcm_Oz.o
~/.local/share/compiler_gym/llvm-v0/bin/clang -O3 -c adpcm.bc -o adpcm_O3.o

why the result is different from compilergym？Is this normal?

IIUC, compiler_gym behave like below, this may explain why these two results are different.
(See 179 line in this file CompilerGym/compiler_gym/envs/llvm/llvm_benchmark.py for details)

clang -O1 -Xclang -disable-llvm-passes -Xclang -disable-llvm-optzns -c adpcm.bc
opt -O0 -c adpcm.bc -o adpcm_O0.bc
opt -Oz -c adpcm.bc -o adpcm_Oz.bc
opt -O3 -c adpcm.bc -o adpcm_O3.bc

However, it also confused me that ObjectTextSizeOz > ObjectTextSizeO0. I don't know what happend...

@ChrisCummins
Copy link
Contributor

Yes @yingluosanqian you are correct, the LlvmEnv first uses clang to lower to a bitcode, then uses opt to do the actual optimization. Unfortunately, doing this does not get the same results as running a single clang invocation. This is kind of frustrating, but I don't think there is a simple workaround, since clang doesn't expose pass-ordering flags. It'd require a fairly substantial change to clang to expose the action space. Bottom line, these janky results are sadly unavoidable :(

Cheers,
Chris

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
Question Further information is requested
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
3 participants
@ChrisCummins @sparks-code @yingluosanqian