Optimize AArch64 code with LLVM LTO and PGO
Introduction
Understand PGO and LTO for AArch64 code
Prepare your AArch64 environment and verify LLVM tool availability
Build AArch64 code with LTO
Optimize AArch64 code with S-PGO
Optimize AArch64 code with FE-PGO
Optimize AArch64 code with IR-PGO
Optimize AArch64 code with CSIR-PGO
Next Steps
Optimize AArch64 code with LLVM LTO and PGO
What CSIR-PGO is
After an initial intermediate representation-level profile-guided optimization (IR-PGO) build, context-sensitive IR-PGO (CSIR-PGO) adds a second, context-sensitive profiling pass.
The first pass is the standard IR-PGO instrumentation pass. The second pass instruments the program after inlining, enabling LLVM to distinguish execution counts from different calling contexts.
The additional context can improve optimization when a function’s behavior depends on its call site. It doesn’t guarantee better performance for every program.
Use CSIR-PGO when you want to provide LLVM with more detailed profile information and can afford an extra build and training run.
Build the context-sensitive instrumented binary
First, generate prof/ir.profdata by completing the
IR-PGO workflow
. Use that profile to guide Clang’s usual PGO-driven optimization decisions while it builds a second instrumented binary:
clang++ -O3 -flto -fuse-ld=lld \
-fprofile-use=prof/ir.profdata \
-fcs-profile-generate=prof/csir \
bsort.cpp -o out/bsort.csirpgo.instr
The -fcs-profile-generate option then adds context-sensitive counters after inlining.
Run the context-sensitive instrumented binary:
./out/bsort.csirpgo.instr
Confirm that the training run created at least one context-sensitive raw profile:
ls prof/csir/*.profraw
Merge and inspect the profiles
Merge the context-sensitive raw profiles with the existing prof/ir.profdata profile:
Don’t merge only the .profraw files: the final CSIR-PGO profile needs counts from both instrumentation passes.
llvm-profdata merge prof/ir.profdata prof/csir -output=prof/csir.profdata
Inspect the context-sensitive execution counts recorded for sort_array. For this example, sort_array should have six context-sensitive counters:
llvm-profdata show --showcs --counts --function=sort_array prof/csir.profdata
The output is similar to:
Counters:
ld-temp.o;_Z10sort_arrayPi:
Hash: 0x18c2aba34f0cfff9
Counters: 6
Block counts: [24763682, 25224415, 9999, 9882, 1, 9881]
Instrumentation level: IR entry_first = 0 instrument_loop_entries = 0
Functions shown: 1
Total functions: 12
Maximum function count: 24763682
Maximum internal block count: 25224415
Total number of blocks: 32
Total count: 75242276
The --showcs option selects context-sensitive records from the merged profile. Exact hashes and individual counter values can vary with the LLVM version and training workload.
Build with CSIR-PGO and LTO
Build the optimized binary using the merged CSIR-PGO profile:
clang++ -O3 -flto -fuse-ld=lld \
-fprofile-use=prof/csir.profdata \
bsort.cpp -o out/bsort.csirpgo.opt
Run the optimized binary:
./out/bsort.csirpgo.opt
What you’ve accomplished
You’ve added a second context-sensitive profiling pass on top of IR-PGO and built a CSIR-PGO optimized binary with LTO.
You can now apply the workflow that best matches your profiling environment to a representative workload from your own application.