[BOLT] CMOVConversion pass
authorAmir Ayupov <aaupov@fb.com>
Tue, 8 Feb 2022 04:16:13 +0000 (20:16 -0800)
committerAmir Ayupov <aaupov@fb.com>
Tue, 8 Mar 2022 18:44:31 +0000 (10:44 -0800)
commit687e4af1c05ae36af88900d41150e260d8f273c0
tree3e742ec4ac36a1daf0dfe7d19f43b2391331052d
parent151f809c558d3d7e67e4d4b7efe84218c3b8cfa7
[BOLT] CMOVConversion pass

Convert simple hammocks into cmov based on misprediction rate.

Test Plan:
- Assembly test: `cmov-conversion.s`
- Testing on a binary:
  # Bootstrap clang with `-x86-cmov-converter-force-all` and `-Wl,--emit-relocs`
  (Release build)
  # Collect perf.data:

    - `clang++ <opts> bolt/lib/Core/BinaryFunction.cpp -E > bf.cpp`
    - `perf record -e cycles:u -j any,u -- clang-15 bf.cpp -O2 -std=c++14 -c -o bf.o`
  # Optimize clang-15 with and w/o -cmov-conversion:
    - `llvm-bolt clang-15 -p perf.data -o clang-15.bolt`
    - `llvm-bolt clang-15 -p perf.data -cmov-conversion -o clang-15.bolt.cmovconv`
  # Run perf experiment:
    - test: `clang-15.bolt.cmovconv`,
    - control: `clang-15.bolt`,
    - workload (clang options): `bf.cpp -O2 -std=c++14 -c -o bf.o`
Results:
```
  task-clock [delta: -360.21 ± 356.75, delta(%): -1.7760 ± 1.7589, p-value: 0.047951, balance: -6]
  instructions  [delta: 44061118 ± 13246382, delta(%): 0.0690 ± 0.0207, p-value: 0.000001, balance: 50]
  icache-misses [delta: -5534468 ± 2779620, delta(%): -0.4331 ± 0.2175, p-value: 0.028014, balance: -28]
  branch-misses [delta: -1624270 ± 1113244, delta(%): -0.3456 ± 0.2368, p-value: 0.030300, balance: -22]
```

Reviewed By: rafauler

Differential Revision: https://reviews.llvm.org/D120177
bolt/include/bolt/Core/MCPlusBuilder.h
bolt/include/bolt/Passes/CMOVConversion.h [new file with mode: 0644]
bolt/lib/Passes/CMOVConversion.cpp [new file with mode: 0644]
bolt/lib/Passes/CMakeLists.txt
bolt/lib/Rewrite/BinaryPassManager.cpp
bolt/lib/Target/X86/X86MCPlusBuilder.cpp
bolt/test/X86/cmov-conversion.s [new file with mode: 0644]