Evaluate Profile-Guided Optimization (PGO) and LLVM BOLT

[zamazan4ik]

Evaluate Profile-Guided Optimization (PGO) and LLVM BOLT

[kulakowski]

Thanks for the issue! And I found https://github.com/zamazan4ik/awesome-pgo via your github profile, it's quite a helpful resource.

LTO, PGO, and BOLT are definitely things we've discussed internally, as well as the challenges you mentioned like collecting a good enough profile, and integrating them into the build and release processes. And in fact we are using ThinLTO in release builds now.

As the project matures we'll get a better idea of when we will invest more in this!

[marvin-hansen]

Actually, LTO & PGO is relatively easy in practice.

I made a code example a few month ago with a step-by-step quick guide:

https://github.com/rust-academy/rust-lto-pgo/blob/main/docs/results-health.txt

cargo-pgo is your best friend. It's super simple and fast to use:
https://github.com/Kobzol/cargo-pgo

The biggest gain for me was massively reduced tail end latency meaning everything past the 75 percentile saw
significant reduced latency and I would say users generally appreciate less loading time when systems are busy.