Profile guided optimization (pgo)

pgo is an optimization technique to optimize a program for its usual workload.

It is applied in two phases:

  1. Collect profiling data (best with representative benchmarks).
  2. Optimize program based on collected profiling data.

The following simple program is used as demonstrator.

#include <stdio.h>

#define NOINLINE __attribute__((noinline))

NOINLINE void foo() { puts("foo()"); }
NOINLINE void bar() { puts("bar()"); }

int main(int argc, char *argv[]) {
  if (argc == 2) {
    foo();
  } else {
    bar();
  }
}

clang

On the actual machine with clang 15.0.7, the following code is generated for the main() function.

# clang -o test test.c -O3

0000000000001160 <main>:
    1160:  50                   push   rax
    ; Jump if argc != 2.
    1161:  83 ff 02             cmp    edi,0x2
    1164:  75 09                jne    116f <main+0xf>
    ; foor() is on the hot path (fall-through).
    1166:  e8 d5 ff ff ff       call   1140 <_Z3foov>
    116b:  31 c0                xor    eax,eax
    116d:  59                   pop    rcx
    116e:  c3                   ret
    ; bar() is on the cold path (branch).
    116f:  e8 dc ff ff ff       call   1150 <_Z3barv>
    1174:  31 c0                xor    eax,eax
    1176:  59                   pop    rcx
    1177:  c3                   ret

The following shows how to compile with profiling instrumentation and how to optimize the final program with the collected profiling data (llvm pgo).

The arguments to ./test are chosen such that 9/10 runs call bar(), which is currently on the cold path.

# Compile test program with profiling instrumentation.
clang -o test test.cc -O3 -fprofile-instr-generate

# Collect profiling data from multiple runs.
for i in {0..10}; do
    LLVM_PROFILE_FILE="prof.clang/%p.profraw" ./test $(seq 0 $i)
done

# Merge raw profiling data into single profile data.
llvm-profdata merge -o pgo.profdata prof.clang/*.profraw

# Optimize test program with profiling data.
clang -o test test.cc -O3 -fprofile-use=pgo.profdata

NOTE: If LLVM_PROFILE_FILE is not given the profile data is written to default.profraw which is re-written on each run. If the LLVM_PROFILE_FILE contains a %m in the filename, a unique integer will be generated and consecutive runs will update the same generated profraw file, LLVM_PROFILE_FILE can specify a new file every time, however that requires more storage in general.

After optimizing the program with the profiling data, the main() function looks as follows.

0000000000001060 <main>:
    1060:  50                    push   rax
    ; Jump if argc == 2.
    1061:  83 ff 02              cmp    edi,0x2
    1064:  74 09                 je     106f <main+0xf>
    ; bar() is on the hot path (fall-through).
    1066:  e8 e5 ff ff ff        call   1050 <_Z3barv>
    106b:  31 c0                 xor    eax,eax
    106d:  59                    pop    rcx
    106e:  c3                    ret
    ; foo() is on the cold path (branch).
    106f:  e8 cc ff ff ff        call   1040 <_Z3foov>
    1074:  31 c0                 xor    eax,eax
    1076:  59                    pop    rcx
    1077:  c3                    ret

gcc

With gcc 13.2.1 on the current machine, the optimizer puts bar() on the hot path by default.

0000000000001040 <main>:
    1040:  48 83 ec 08          sub    rsp,0x8
    ; Jump if argc == 2.
    1044:  83 ff 02             cmp    edi,0x2
    1047:  74 0c                je     1055 <main+0x15>
    ; bar () is on the hot path (fall-through).
    1049:  e8 22 01 00 00       call   1170 <_Z3barv>
    104e:  31 c0                xor    eax,eax
    1050:  48 83 c4 08          add    rsp,0x8
    1054:  c3                   ret
    ; foo() is on the cold path (branch).
    1055:  e8 06 01 00 00       call   1160 <_Z3foov>
    105a:  eb f2                jmp    104e <main+0xe>
    105c:  0f 1f 40 00          nop    DWORD PTR [rax+0x0]

The following shows how to compile with profiling instrumentation and how to optimize the final program with the collected profiling data.

The arguments to ./test are chosen such that 2/3 runs call foo(), which is currently on the cold path.

gcc -o test test.cc -O3 -fprofile-generate
./test 1
./test 1
./test 2 2
gcc -o test test.cc -O3 -fprofile-use

NOTE: Consecutive runs update the generated test.gcda profile data file rather than re-write it.

After optimizing the program with the profiling data, the main() function

0000000000001040 <main.cold>:
    ; bar() is on the cold path (branch).
    1040:  e8 05 00 00 00       call   104a <_Z3barv>
    1045:  e9 25 00 00 00       jmp    106f <main+0xf>

0000000000001060 <main>:
    1060:  51                   push   rcx
    ; Jump if argc != 2.
    1061:  83 ff 02             cmp    edi,0x2
    1064:  0f 85 d6 ff ff ff    jne    1040 <main.cold>
    ; for() is on the hot path (fall-through).
    106a:  e8 11 01 00 00       call   1180 <_Z3foov>
    106f:  31 c0                xor    eax,eax
    1071:  5a                   pop    rdx
    1072:  c3                   ret