2022-07-19 Triage Log

A rather rough week for compiler performance with regressions outweighing improvements by a considerable margin, in particular in real world crates. To add insult to injury, the biggest regressions came in rollups which make it difficult to trace the cause.

Triage done by @rylev. Revision range: b3f4c3119957aa0a250cab08ab586b7a9a680ef1..8bd12e8cca3f28f302b9cc0f1f47bb64bd1f98fd

Summary:

	mean	max	count
Regressions 😿 (primary)	1.5%	4.0%	176
Regressions 😿 (secondary)	1.8%	6.4%	147
Improvements 🎉 (primary)	N/A	N/A	0
Improvements 🎉 (secondary)	-1.6%	-4.1%	9
All 😿🎉 (primary)	1.5%	4.0%	176

7 Regressions, 5 Improvements, 3 Mixed; 4 of them in rollups 48 artifact comparisons made in total

Regressions

Rollup of 6 pull requests #99177 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	N/A	N/A	0
Regressions 😿 (secondary)	0.5%	0.6%	9
Improvements 🎉 (primary)	N/A	N/A	0
Improvements 🎉 (secondary)	N/A	N/A	0
All 😿🎉 (primary)	N/A	N/A	0

Relatively small regression in secondary benchmarks. Nothing stands out as an obvious culprit, so we can mark as triaged.

Pull Derefer before ElaborateDrops #98145 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	0.5%	0.8%	13
Regressions 😿 (secondary)	0.7%	1.1%	20
Improvements 🎉 (primary)	N/A	N/A	0
Improvements 🎉 (secondary)	N/A	N/A	0
All 😿🎉 (primary)	0.5%	0.8%	13

Slowdowns are happening in MIR borrow checking which this PR directly changes, so regressions are not surprising.
An issue should be opened to look into whether this can be improved.
A comment was left asking for a follow up.

Rollup of 5 pull requests #99210 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	2.6%	4.5%	7
Regressions 😿 (secondary)	1.6%	4.1%	55
Improvements 🎉 (primary)	N/A	N/A	0
Improvements 🎉 (secondary)	N/A	N/A	0
All 😿🎉 (primary)	2.6%	4.5%	7

Seems the issue comes from calling an LLVM API more often due to needing to have unstable feature information kept around longer for use with the asm! macro.
Ran a cachegrind diff and found that the likely culprit is #99155 calling LLVMRustHasFeature more often.
Left a comment asking for a follow up.

Rollup of 5 pull requests #99231 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	1.2%	3.2%	117
Regressions 😿 (secondary)	1.7%	6.4%	35
Improvements 🎉 (primary)	N/A	N/A	0
Improvements 🎉 (secondary)	-0.3%	-0.3%	1
All 😿🎉 (primary)	1.2%	3.2%	117

It's very hard to tell which PR might be responsible. Luckily, there are only 5 PRs in total so it should be possible to track the issue down.
Left a comment suggesting as much.

Support unstable moves via stable in unstable items #95956 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	0.5%	0.9%	13
Regressions 😿 (secondary)	0.6%	1.2%	10
Improvements 🎉 (primary)	N/A	N/A	0
Improvements 🎉 (secondary)	N/A	N/A	0
All 😿🎉 (primary)	0.5%	0.9%	13

The change seems to be localized to incremental compilation scenarios.
It seems queries like incr_comp_encode_dep_graph are being hit more often. I‘m unsure whether this is simply because of the additional pass that’s happening or something else.
Left a comment asking for more insight.

Rollup of 10 pull requests #99346 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	0.6%	2.6%	136
Regressions 😿 (secondary)	1.0%	5.8%	100
Improvements 🎉 (primary)	N/A	N/A	0
Improvements 🎉 (secondary)	N/A	N/A	0
All 😿🎉 (primary)	0.6%	2.6%	136

Regression was found in #98582 and will be reverted in #99368 or a follow up.

Use ICF (identical code folding) for building rustc #99062 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	0.3%	0.4%	2
Regressions 😿 (secondary)	2.6%	2.6%	3
Improvements 🎉 (primary)	N/A	N/A	0
Improvements 🎉 (secondary)	N/A	N/A	0
All 😿🎉 (primary)	0.3%	0.4%	2

Small performance changes and some wins for binary size.
Asked in the PR whether this can be marked as triaged.

Improvements

gather body owners #98203 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	N/A	N/A	0
Regressions 😿 (secondary)	0.6%	0.6%	1
Improvements 🎉 (primary)	-0.4%	-0.7%	30
Improvements 🎉 (secondary)	-0.5%	-0.8%	18
All 😿🎉 (primary)	-0.4%	-0.7%	30

interpret: get rid of MemPlaceMeta::Poison #99013 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	N/A	N/A	0
Regressions 😿 (secondary)	0.5%	0.5%	1
Improvements 🎉 (primary)	N/A	N/A	0
Improvements 🎉 (secondary)	-1.6%	-2.5%	12
All 😿🎉 (primary)	N/A	N/A	0

stop using FnCtxt outside of hir typeck #99015 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	N/A	N/A	0
Regressions 😿 (secondary)	N/A	N/A	0
Improvements 🎉 (primary)	-0.6%	-0.9%	11
Improvements 🎉 (secondary)	-0.4%	-0.6%	11
All 😿🎉 (primary)	-0.6%	-0.9%	11

Use constant eval to do strict mem::uninit/zeroed validity checks #99033 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	N/A	N/A	0
Regressions 😿 (secondary)	N/A	N/A	0
Improvements 🎉 (primary)	N/A	N/A	0
Improvements 🎉 (secondary)	-2.2%	-2.9%	6
All 😿🎉 (primary)	N/A	N/A	0

replace_bound_vars fast path: check predicates, don't check consts #99232 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	N/A	N/A	0
Regressions 😿 (secondary)	N/A	N/A	0
Improvements 🎉 (primary)	-0.6%	-0.7%	6
Improvements 🎉 (secondary)	N/A	N/A	0
All 😿🎉 (primary)	-0.6%	-0.7%	6

Mixed

Stop keeping metadata in memory before writing it to disk #96544 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	N/A	N/A	0
Regressions 😿 (secondary)	0.3%	0.4%	3
Improvements 🎉 (primary)	-0.3%	-0.3%	1
Improvements 🎉 (secondary)	-4.1%	-4.1%	2
All 😿🎉 (primary)	-0.3%	-0.3%	1

The perf here is a wash.

Final derive output improvements #99046 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	0.5%	0.6%	8
Regressions 😿 (secondary)	N/A	N/A	0
Improvements 🎉 (primary)	-0.6%	-0.9%	19
Improvements 🎉 (secondary)	-2.8%	-3.7%	6
All 😿🎉 (primary)	-0.2%	-0.9%	27

Improvements much outweigh the regressions as noted here.

proc_macro/bridge: stop using a remote object handle for proc_macro Ident and Literal #98189 (Comparison Link)

	mean	max	count
Regressions 😿 (primary)	1.0%	2.2%	4
Regressions 😿 (secondary)	N/A	N/A	0
Improvements 🎉 (primary)	-0.7%	-1.7%	19
Improvements 🎉 (secondary)	-3.0%	-3.1%	3
All 😿🎉 (primary)	-0.4%	2.2%	23

Still more improvements than regressions, and this is an active area of performance work as noted here.