Optimize AArch64 code with IR-PGO

What IR-PGO is

Intermediate representation-level profile-guided optimization (IR-PGO) adds counters to LLVM IR before LLVM runs its optimization passes. The counters collect execution frequencies during the training run.

Compared with frontend PGO (FE-PGO), IR-PGO usually has lower instrumentation overhead and creates smaller raw profiles. IR-PGO is generally the better instrumentation-based choice for optimization.

Build the instrumented binary

Build an instrumented binary. The -fprofile-generate option tells LLVM to add IR-level counters and write raw profiles to the specified directory.

    

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

    

Run the instrumented binary:

    

        
        
./out/bsort.irpgo.instr

    

Confirm that the training run created at least one raw profile:

    

        
        
ls prof/ir/*.profraw

    

Merge and inspect the profile data

With -fprofile-generate=prof/ir, Clang writes a uniquely named raw profile in the specified directory. Multiple training runs or processes can create or update profiles there without clobbering unrelated profile data. For more information, see Clang profile filename patterns .

Before Clang can use the profile during an optimized build, merge the raw profiles with llvm-profdata:

    

        
        
llvm-profdata merge prof/ir -output=prof/ir.profdata

    

The merge step also converts the profiles to the .profdata format used during optimization.

Inspect the execution counts recorded for sort_array:

    

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

    

The output is similar to:

    

        
        Counters:
  _Z10sort_arrayPi:
    Hash: 0x08380d8f3e6d4c88
    Counters: 6
    Block counts: [9882, 49988097, 25224415, 10000, 1, 9882]
Instrumentation level: IR  entry_first = 0  instrument_loop_entries = 0
Functions shown: 1
Total functions: 13
Maximum function count: 9882
Maximum internal block count: 49988097
Total number of blocks: 38
Total count: 150915523

        
    

The profile contains six IR-level 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 IR-PGO and LTO

Build the optimized binary using the merged profile:

    

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

    

Run the optimized binary:

    

        
        
./out/bsort.irpgo.opt

    

What you’ve accomplished and what’s next

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

Next, you’ll extend the IR-PGO workflow with a context-sensitive profiling pass.