src/tools/rustc-perf/triage/2021-05-11.md

2021-05-11 Triage Log

Not much change overall - both regressions and improvements were all minor, apart from the 2x compile-time improvement for libcore from PR #83278.

We did see a pretty clear 2% improvement in memory usage (max-rss), and that was again due to PR #83278.

2 Regressions, 3 Improvements, 0 Mixed 0 of them in rollups

Integrate attributes as part of the crate hash #83901

Moderate regression in instruction counts (up to 1.6% on full builds of externs-check)
There was a regression, seems roughly consistent with the results that we got from the perf run on the PR, so this regression was expected outcome.

Bump stdarch submodule #83278

Moderate regression in instruction counts (up to 3.3% on full builds of keccak-debug)
There are regressions, but as noted in PR comments, they were not as bad as anticipated. And also, we did continue to see the 2x win on compile-time for libcore.

Retain data in vectorized registers for longer #84915

Moderate improvement in instruction counts (up to -1.3% on full builds of keccak-debug)
(Slight improvement, as expected from perf run on the PR itself.)

lazify backtrace formatting for delayed diagnostics #84965

Moderate improvement in instruction counts (up to -2.2% on incr-unchanged builds of cranelift-codegen-check)
(Slight improvement, though PR itself is focused on a narrower case of interest, namely the time to run the ui test suite with debuginfo turned on. In that context, its a huge win, going from 42s to 34s according to the PR author.)

Optimize BufWriter #79930

Moderate improvement in instruction counts (up to -1.8% on incr-unchanged builds of encoding-check)
(Slight improvement, with some very minor regressions/noise mixed in. But we would not expect BufWriter activity to be on the hot-path of rustc itself, so its not surprising that the optimization here would have little effect on the perf.rlo benchmark suite.)

No new nags this week, and none leftover from last week.