2023-01-03 Triage Log

Fairly busy week with some massive performance improvements at the expense of some significant albeit smaller regressions. The main wins came in a long-standing PR from @cjgillot to enable encoding spans in metadata relative to their enclosing item. This causes more work in full compilation which causes some regressions up to 5% but can lead to very large wins in incremental compilation scenarios (up to ~70%). For example, the clap crate compiles 68% faster after a small 1 line change than it did previously.

Triage done by @rylev. Revision range: b38a6d..b43596

Summary:

(instructions:u)	mean	range	count
Regressions ❌ (primary)	1.6%	[0.3%, 4.6%]	97
Regressions ❌ (secondary)	1.8%	[0.2%, 7.6%]	60
Improvements ✅ (primary)	-9.7%	[-68.7%, -0.2%]	53
Improvements ✅ (secondary)	-1.7%	[-15.3%, -0.1%]	62
All ❌✅ (primary)	-2.4%	[-68.7%, 4.6%]	150

1 Regressions, 1 Improvements, 4 Mixed; 1 of them in rollups 47 artifact comparisons made in total

Regressions

Rollup of 8 pull requests #106228 (Comparison Link)

(instructions:u)	mean	range	count
Regressions ❌ (primary)	0.4%	[0.2%, 0.6%]	7
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	0.4%	[0.2%, 0.6%]	7

Kicked off a perf run on the only likely culprit. The regression is small enough that I don‘t think it’s a big deal if we don't figure this out.

Allow trait method paths to satisfy const Fn bounds #106210 (Comparison Link)

(instructions:u)	mean	range	count
Regressions ❌ (primary)	0.3%	[0.2%, 0.5%]	16
Regressions ❌ (secondary)	0.4%	[0.2%, 0.7%]	16
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-1.4%	[-1.4%, -1.4%]	1
All ❌✅ (primary)	0.3%	[0.2%, 0.5%]	16

Massive amount of noise reversed in #106268. No idea where this noise comes from.

Improvements

fix comment for TokenCursor::desugar #106268 (Comparison Link)

(instructions:u)	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-0.3%	[-0.5%, -0.2%]	17
Improvements ✅ (secondary)	-0.3%	[-0.7%, -0.2%]	15
All ❌✅ (primary)	-0.3%	[-0.5%, -0.2%]	17

Massive amount of noise coming in #106210 reversed here. No idea where this noise comes from.

Abolish QueryVTable in favour of more assoc items on QueryConfig #106307 (Comparison Link)

(instructions:u)	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	0.6%	[0.4%, 1.0%]	4
Improvements ✅ (primary)	-0.5%	[-1.0%, -0.2%]	63
Improvements ✅ (secondary)	-0.6%	[-1.3%, -0.4%]	60
All ❌✅ (primary)	-0.5%	[-1.0%, -0.2%]	63

Mixed

Catch panics/unwinding in destruction of TLS values #105426 (Comparison Link)

(instructions:u)	mean	range	count
Regressions ❌ (primary)	0.7%	[0.3%, 1.9%]	11
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-0.3%	[-0.6%, -0.2%]	41
Improvements ✅ (secondary)	-1.8%	[-14.6%, -0.2%]	31
All ❌✅ (primary)	-0.1%	[-0.6%, 1.9%]	52

Very much unsure where the performance changes are coming from. Some show LLVM regressions but not all. LLVM also does not explain the improvements.
Ask in the PR for thoughts.

Use some more const_eval_select in pointer methods for compile times #106275 (Comparison Link)

(instructions:u)	mean	range	count
Regressions ❌ (primary)	0.4%	[0.4%, 0.4%]	1
Regressions ❌ (secondary)	0.7%	[0.6%, 0.9%]	2
Improvements ✅ (primary)	-0.5%	[-0.9%, -0.2%]	28
Improvements ✅ (secondary)	-0.8%	[-0.8%, -0.7%]	3
All ❌✅ (primary)	-0.4%	[-0.9%, 0.4%]	29

Author investigated why this might have caused a regression in regex and found nothing. The gains much outweigh the regressions so we should be fine keeping it.

Encode spans relative to the enclosing item -- enable on nightly #84762 (Comparison Link)

(instructions:u)	mean	range	count
Regressions ❌ (primary)	1.7%	[0.2%, 5.0%]	108
Regressions ❌ (secondary)	1.8%	[0.2%, 7.6%]	90
Improvements ✅ (primary)	-20.0%	[-68.9%, -0.6%]	25
Improvements ✅ (secondary)	-2.7%	[-4.8%, -1.5%]	11
All ❌✅ (primary)	-2.4%	[-68.9%, 5.0%]	133

There was an FCP to decide whether the trade off of some regressions for some huge incremental performance wins was worth it.
This FCP passed and so the PR was merged.

Reenable limited top-down MIR inlining #106364 (Comparison Link)

(instructions:u)	mean	range	count
Regressions ❌ (primary)	0.5%	[0.3%, 1.1%]	4
Regressions ❌ (secondary)	0.6%	[0.5%, 0.6%]	4
Improvements ✅ (primary)	-0.6%	[-1.2%, -0.2%]	9
Improvements ✅ (secondary)	-0.9%	[-2.4%, -0.2%]	18
All ❌✅ (primary)	-0.2%	[-1.2%, 1.1%]	13

As pointed out here this was effectively a performance wash with regressions and improvements canceling themselves out.