2020-08-03 Triage Log

See the README for details on how this log was gathered.

Triage done by simulacrum. Revision range: efc02b03d18b0cbaa55b1e421d792f70a39230b2..19cefa68640843956eedd86227ddc1d35dbc6754

8 regressions, 2 improvements, 1 of them on rollups. 1 outstanding nag from last week.

Regressions

• convert higher ranked Predicates to PredicateKind::ForAll #73503 (instructions) Up to 2.5% losses on stress tests, 1-2% on some other real-world crates. This was expected.
• mv std libs to library/ #73265 (instructions) Up to 3% losses on tiny crates, and stress tests show 1-3% regressions.
• Update cargo #74923 (instructions) This was up to a 28% loss on some incremental benchmarks, as a result of regressions in proc macro performance (expansion is slower).
• Replace all uses of log::log_enabled with Debug printers #74876 (instructions) A 1.5% regression on the ctfe stress test.
• std: Switch from libbacktrace to gimli (take 2) #74682 (instructions) An expected loss of up to 20% on smaller benchmarks. This was also a max-rss regression, though how big is difficult to determine as this statistics is noisy. See discussion on Zulip. Optimistically, there is a follow-up PR, “metadata: track the simplified Self type for every trait impl #75008”, which may eliminate a good portion of these losses.
• Normalize all opaque types when converting ParamEnv to Reveal::All #65989 (instructions) A slight performance regression, of up to 4%. Expected.
• Move ‘probably equal’ methods to librustc_parse #75004 (instructions) A 1% regression on some smaller crates, rustdoc-only.
• Rollup of 10 pull requests #75060 (instructions) Up to 1% regression; unclear cause -- possibly BTreeMap changes.

Improvements

• Cache non-exhaustive separately from attributes #74887 (instructions) Up to 2% wins on tiny crates. This arose as an attempt to mitigate some of the effects of re-landing gimli.
• Move from log to tracing #74726 (instructions) Up to 1.5% wins, scattered across the smaller benchmarks. Expected.

Nags:

• Disable polymorphisation #74633 From last week, this was a performance neutral PR even though the landing of polymorphisation was a regression. Investigation on why needs to be conducted.