2021-09-28 Triage Log

The largest story for the week are the massive improvements that come from enabling the new pass manager in LLVM which leads to consistent 5% to 30% improvements across almost all test cases. The regressions were mostly minor with clear paths for addressing the ones that were not made with some specific trade off in mind.

Triage done by @rylev. Revision range: 7743c9fadd64886d537966ba224b9c20e6014a59..83f147b3baf21acfc367a6da1045d212cd3957e4

4 Regressions, 4 Improvements, 3 Mixed; 0 of them in rollups 43 comparisons made in total

Regressions

Revise never type fallback algorithm #88804

Introduce Rvalue::ShallowInitBox #89030

  • Moderate regression in instruction counts (up to 1.9% on incr-patched: println builds of syn)
  • Perf regression is happening in real-world optimized builds which we would expect if we're making LLVM do more work.
  • The author has an idea for how to reduce the pressure on LLVM and perhaps win back some of the perf, but was unsure if the regression was large enough to warrant that investigation.
  • Left a comment asking the author to think about prioritizing that investigation.

Fix spacing of links in inline code. #88343

  • Large regression in instruction counts (up to 2.1% on incr-unchanged builds of webrender-wrench)
  • This change is only in doc comments for the standard library so large regressions are quite surprising.
  • This has impacted the incr-unchanged scenario of only one benchmark. Perhaps the docs led to a change in how incremental cache was being stored which could have an impact?

Suggest both of immutable and mutable trait implementations #89263

  • Moderate regression in instruction counts (up to 2.0% on full builds of diesel)
  • The regression occurs in evaluate_obligation which seems like it would be effected by this change, but the only impacted benchmark is diesel doc which doesn't trigger this diagnostic
  • The change is large enough though that it being noise would be quite surprising.
  • Left a comment

Improvements

Migrate in-tree crates to 2021 #89103

  • Moderate improvement in instruction counts (up to -2.9% on incr-full builds of ctfe-stress-4)

Disable visible path calculation for PrettyPrinter in Ok path of compiler #89120

  • Very large improvement in instruction counts (up to -23.1% on incr-full builds of await-call-tree)

Make Duration respect width when formatting using Debug #88999

  • Moderate improvement in instruction counts (up to -1.0% on full builds of deeply-nested-async)

Enable new pass manager with LLVM 13 #88243

  • Very large improvement in instruction counts (up to -31.7% on incr-patched: println builds of inflate)

Mixed

Use ZST for fmt unsafety #89139

  • Moderate improvement in instruction counts (up to -1.6% on full builds of cranelift-codegen)
  • Moderate regression in instruction counts (up to 0.5% on full builds of deeply-nested-async)
  • The regressions were already triaged here.

Support #[track_caller] on closures and generators #87064

  • Large improvement in instruction counts (up to -1.8% on incr-unchanged builds of webrender-wrench)
  • Small regression in instruction counts (up to 0.5% on incr-unchanged builds of helloworld)
  • The large majority of performance regressions were quite small and many were historically noisy benchmarks.
  • Looking at the non-noisy benchmarks nothing seemed to stand out as a clear culprit. -
  • There were some test cases that indicated large jumps in incr_comp_persist_dep_graph but this did not hold true across all incremental test cases, and nothing in the changes seemed to jump out at me as potentially causing this issue.
  • Added a comment in the PR asking for additional opinions.

Don't normalize opaque types with escaping late-bound regions #89285

  • Very large improvement in instruction counts (up to -86.7% on full builds of issue-88862)
  • Moderate regression in instruction counts (up to 1.3% on full builds of deeply-nested-async)
  • This has already been triaged.
  • Essentially it's a targeted fix for a severe perf regression that can happen in limited circumstances. The fix will be backported to the latest beta.

Untriaged Pull Requests