I noticed that we are currently passing -Zincremental-verify-ich unconditionally:
https://github.com/rust-lang/rustc-perf/blob/a513efc21cce31d0dc55abf16060334581853dc3/collector/src/bin/rustc-fake.rs#L78-L85
This at minimum slows down that compilation, and likely also pollutes the profiling data with incorrect measurements (for actual users - perf.rlo would obviously be happy about it).
I'm not sure how much extra noise we'd see if we didn't do this though; the default is to verify 1/32 of the hashes which might mean that random-ish code sometimes get better PGO and sometimes worse. I'm also not sure how much benefit there is to not doing this.
cc @Kobzol
I noticed that we are currently passing -Zincremental-verify-ich unconditionally:
https://github.com/rust-lang/rustc-perf/blob/a513efc21cce31d0dc55abf16060334581853dc3/collector/src/bin/rustc-fake.rs#L78-L85
This at minimum slows down that compilation, and likely also pollutes the profiling data with incorrect measurements (for actual users - perf.rlo would obviously be happy about it).
I'm not sure how much extra noise we'd see if we didn't do this though; the default is to verify 1/32 of the hashes which might mean that random-ish code sometimes get better PGO and sometimes worse. I'm also not sure how much benefit there is to not doing this.
cc @Kobzol