Evaluate using Profile-Guided Optimization (PGO) and Post-Link Time Optimization (PLO) for Zluda

[zamazan4ik]

Hi!

Recently I tested Profile-Guided Optimization (PGO) on a lot of projects - all my current results are available here. According to the tests, PGO helps in many real-life cases, including compilers, static analysis, parsing, etc. (e.g. Clang gets up to +20% compilation speed with enabled PGO). I think trying to speed up ZLUDA tools like compiler can be especially a good idea.

I can propose the following action points:

• Try to apply PGO to ZLUDA tools and perform all required benchmarks. For Rust projects, I can recommend starting with cargo-pgo - it makes a bit easier to perform PGO routines.
• If PGO works fine for ZLUDA - document PGO usage with ZLUDA somewhere (maybe in the README file), and integrate PGO usage with ZLUDA into the build scripts
• Apply Post-Link Optimization (PLO) optimizations with tools like LLVM BOLT. However, I recommend doing it only after PGO integration.

I would be happy to answer all your questions about PGO and PLO.