Evaluate Profile-Guided Optimization (PGO) and LLVM BOLT about pgvecto.rs HOT 6 OPEN

Absolutely, it could still be incredibly profitable for them if e.g. the team working on pgvecto.rs happened to have someone available to basically commit full time to making sure:

• PGO works
• BOLT works
• they don't introduce any problems
• maybe doing some other opportunistic performance opts
• and keeping an eye on that for a while after as their setup evolves. how long? ...I dunno!

Anyways, BOLT-related miscompiles:

• rust-lang/rust#109067
• rust-lang/rust#108978

Generic BOLT issues:

• rust-lang/rust#105783
• rust-lang/rust#112286
• rust-lang/rust#115344 the author of this seems familiar :^)

Relevant perf tracking issues for rustc:

• rust-lang/rust#103595
• rust-lang/rust#79442

Rust partially mitigated this by extending testing to the optimized artifacts as well:

• rust-lang/rust#109535

from pgvecto.rs.

Since I agree with your comments about BOLT and PGO stability issues, I still think PGO and BOLT combination is still a good thing to consider for the projects. Yes, combining PGO and BOLT can cause the issue but the decision on the tradeoff between "PGO+BOLT miscompilation possibilities" vs "gains from BOLT with PGO" should be considered for each project individually based on the projects' merits ("performance for all money", "stability first") and available resources like active contributors (as you said above).

Regarding BOLT + PGO issues - do you have a list of them anywhere near you? I am very interested in the story behind them and definitely will add to my PGO repository as the warnings for anyone who will test PGO with BOLT.

Thanks in advance!

from pgvecto.rs.

Thanks for the issue!

/cc @usamoi @VoVAllen

from pgvecto.rs.

This person has made many comments to this effect, however, I recommend not combining PGO and BOLT. It has caused many upstream issues for rust-lang, and Rust gets away with it mostly by having a lot of active contributors, such that when there's an all-hands-on-deck moment due to miscompilation, someone shows up to resolve it fairly quickly. It basically needs continuous monitoring to make sure nothing has gone wrong, because for whatever reason, rustc seems to be very good at generating code that surprises LLVM's tooling.

One or the other, independently, is probably safe to use, however. Rust had been using PGO for some time without BOLT, and it was only when used in tandem that they caused these issues.

from pgvecto.rs.

Thanks a lot for the links!

from pgvecto.rs.

It has caused many upstream issues for rust-lang

I'm not aware of any miscompilations caused by BOLT so far. We had miscompilations caused by LTO on Windows, but we don't use BOLT there.

Maybe you meant LTO + PGO? But that just sometimes fails to compile, I haven't seen any miscompilations related to that.

from pgvecto.rs.