LLVM Now Using PGO For Building x86_64 Windows Release Binaries: ~22% Faster Builds
Originally was a pull request to make use of both PGO and ThinLTO for the Windows release packaging. But they ended up dialing that back to just PGO for now. But with PGO + ThinLTO they were finding the installer binary around 5% smaller while the toolchain build around 20% faster.
With this commit to LLVM that was merged earlier this week just PGO is enabled by default but great results for toolchain performance:
Use PGO for x86_64 windows release packaging (#71067)
Applying this to 17.0.4 makes the toolchain 22% faster (as measured by building clang).
For those long experimenting with LTO and PGO, the results are likely not too surprising. PGO can yield great performance uplift assuming you have generated accurate profiles for the software use. Though due to complexities involved and the need to have accurate profiles, Clang PGO was shot down from the Linux kernel as one example. In any event PGO is one of the exciting compiler optimization techniques for helping to achieve maximum performance.