[opt-dist] Profile-guided optimization of LLVM regressed on Windows hosts #144318
Description
It has been found in #143898 (comment) that PGO build of LLVM is slower than general one a bit (~1%). Later in #144255 (comment) it has been proved that performance of MSVC dist compiler is better (the same 1%) (all in case of serde and ruff compilation). Zulip discussion has been with a question what to do with dist-msvc at the moment.
If you want to test it locally, here are the SHAs of above PR's jobs:
#144255: parent - 6781992; build - 978fa84
#143898: parent - cccf075; build - a094db8
for testing you may use rustup-toolchain-install-master tool (cargo install --git https://github.com/kennytm/rustup-toolchain-install-master), then install CI toolchains (rustup-toolchain-install-master --component=cargo <SHA> -i x86_64-pc-windows-msvc so you'll have cargo too) and then compile your favourite crate/program with hyperfine to get perf results (hyperfine "cargo +<SHA_PARENT> build --release" "cargo +<SHA_BUILD> build --release" --prepare "cargo clean" --runs 5, this will build code 5 times for each command to get a more reliable output)
please leave a cc for Windows folks and other who can help/do extra profiling, feel free to edit my message to add more info