Optimize AArch64 code with FE-PGO

What FE-PGO is

Frontend profile-guided optimization (FE-PGO) adds counters before Clang lowers the source code to LLVM intermediate representation (IR). The counters collect execution frequencies while the instrumented program runs. The resulting profile maps closely to source-level constructs.

Use FE-PGO when source-level profile information is important. For performance-focused PGO, IR-PGO and context-sensitive IR-PGO (CSIR-PGO) are usually better defaults.

Build the instrumented binary

Build an instrumented binary:

    

        
        
clang++ -O3 -flto -fuse-ld=lld \
    -fprofile-instr-generate=prof/fe.profraw \
    bsort.cpp -o out/bsort.fepgo.instr

    

The -fprofile-instr-generate option tells Clang to add frontend counters. The =prof/fe.profraw value configures the instrumented program to write its raw profile to that path when it exits.

Run the instrumented binary:

    

        
        
./out/bsort.fepgo.instr

    

Let the training run complete and exit normally so the profiling runtime can finish writing the raw profile. Then, confirm that it created prof/fe.profraw:

    

        
        
ls prof/fe.profraw

    

Convert and inspect the profile data

Before Clang can use the profile during an optimized build, convert it to the .profdata format:

    

        
        
llvm-profdata merge prof/fe.profraw -output=prof/fe.profdata

    

Inspect the execution counts recorded for sort_array:

    

        
        
llvm-profdata show --counts --function=sort_array prof/fe.profdata

    

The output is similar to:

    

        
        Counters:
  _Z10sort_arrayPi:
    Hash: 0x00000000000046d1
    Counters: 2
    Function count: 1
    Block counts: [10000]
Instrumentation level: Front-end
Functions shown: 1
Total functions: 16
Maximum function count: 5044883
Maximum internal block count: 49988097
Total number of blocks: 39
Total count: 100476579

        
    

The profile contains two frontend counters for sort_array. Exact hashes and counts can vary with the LLVM version and training workload. Use --all-functions to inspect the complete profile, but expect much more output for a large application.

Build with FE-PGO and LTO

Build the optimized binary using the converted profile:

    

        
        
clang++ -O3 -flto -fuse-ld=lld \
    -fprofile-instr-use=prof/fe.profdata \
    bsort.cpp -o out/bsort.fepgo.opt

    

Run the optimized binary:

    

        
        
./out/bsort.fepgo.opt

    

What you’ve accomplished and what’s next

You’ve collected an FE-PGO profile, converted it with llvm-profdata, and used it with LTO to build an optimized binary.

Next, you’ll build the same example with IR-PGO and LTO.