Reduce the default number of CGUs for incremental.

[nnethercote]

Because it gives better performance.

r? @ghost

[nnethercote]

My local measurements suggested that 128 was clearly better than 256, and 64 was slightly better again. Let's see what CI says.

@bors try @rust-timer queue

[bors]

⌛ Trying commit 5407fe39ae91e062d1d31d69197af4ae351d1f5f with merge 3b2fafd65a7058220b97fb6097fec300c5482372...

[bors]

☀️ Try build successful - checks-actions
Build commit: 3b2fafd65a7058220b97fb6097fec300c5482372 (3b2fafd65a7058220b97fb6097fec300c5482372)

[rust-timer]

Finished benchmarking commit (3b2fafd65a7058220b97fb6097fec300c5482372): comparison URL.

Overall result: ❌✅ regressions and improvements - ACTION NEEDED

Benchmarking this pull request likely means that it is perf-sensitive, so we're automatically marking it as not fit for rolling up. While you can manually mark this PR as fit for rollup, we strongly recommend not doing so since this PR may lead to changes in compiler perf.

Next Steps: If you can justify the regressions found in this try perf run, please indicate this with @rustbot label: +perf-regression-triaged along with sufficient written justification. If you cannot justify the regressions please fix the regressions and do another perf run. If the next run shows neutral or positive results, the label will be automatically removed.

@bors rollup=never
@rustbot label: -S-waiting-on-perf +perf-regression

Instruction count

This is a highly reliable metric that was used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	35.5%	[0.5%, 226.1%]	7
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-0.9%	[-3.5%, -0.2%]	32
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	5.6%	[-3.5%, 226.1%]	39

Max RSS (memory usage)

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	20.1%	[20.1%, 20.1%]	1
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-2.2%	[-3.4%, -1.2%]	4
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	2.3%	[-3.4%, 20.1%]	5

Cycles

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	48.2%	[1.9%, 220.8%]	5
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-1.9%	[-3.7%, -1.1%]	13
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	12.0%	[-3.7%, 220.8%]	18

Binary size

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-1.7%	[-6.6%, -0.4%]	54
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	-1.7%	[-6.6%, -0.4%]	54

Bootstrap: 643.756s -> 643.951s (0.03%)

[nnethercote]

Instruction counts and cycles should be ignored here. Walltime, max-rss, and binary size are the important metrics.

That one terrible webrender results skews everything badly. It's probably a bit of freak random pathological case, and we can pretty much ignore it because opt builds are non-incremental by default. Also, the secondary benchmarks here probably aren't much use, because they're mostly very small and likely to have just one or two functions, hence one or two code units.

With that in mind, looking at just the primary check/debug/doc incremental cases:

• walltime is moderately improved
• max-rss is moderately improved
• binary size is greatly improved

So overall I think the results are pretty decent, better than they might appear at first glance, and this is worth considering. 256 codegen units (and thus 256 threads) is a lot!

[nnethercote]

Now let's try 64 CGUs.

@bors try @rust-timer queue

[bors]

⌛ Trying commit a0954f7 with merge 26269dfb9e9b933270bec418381c5dbf44934013...

[bors]

☀️ Try build successful - checks-actions
Build commit: 26269dfb9e9b933270bec418381c5dbf44934013 (26269dfb9e9b933270bec418381c5dbf44934013)

[rust-timer]

Finished benchmarking commit (26269dfb9e9b933270bec418381c5dbf44934013): comparison URL.

Overall result: ❌✅ regressions and improvements - ACTION NEEDED

Benchmarking this pull request likely means that it is perf-sensitive, so we're automatically marking it as not fit for rolling up. While you can manually mark this PR as fit for rollup, we strongly recommend not doing so since this PR may lead to changes in compiler perf.

Next Steps: If you can justify the regressions found in this try perf run, please indicate this with @rustbot label: +perf-regression-triaged along with sufficient written justification. If you cannot justify the regressions please fix the regressions and do another perf run. If the next run shows neutral or positive results, the label will be automatically removed.

@bors rollup=never
@rustbot label: -S-waiting-on-perf +perf-regression

Instruction count

This is a highly reliable metric that was used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	126.6%	[0.5%, 928.4%]	15
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-1.6%	[-5.4%, -0.2%]	50
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	28.0%	[-5.4%, 928.4%]	65

Max RSS (memory usage)

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	7.1%	[1.8%, 20.2%]	9
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-2.6%	[-5.0%, -1.2%]	5
Improvements ✅ (secondary)	-3.4%	[-3.4%, -3.4%]	2
All ❌✅ (primary)	3.6%	[-5.0%, 20.2%]	14

Cycles

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	159.4%	[2.3%, 831.4%]	11
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-3.0%	[-5.7%, -1.5%]	22
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	51.1%	[-5.7%, 831.4%]	33

Binary size

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-4.2%	[-12.7%, -0.4%]	69
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	-4.2%	[-12.7%, -0.4%]	69

Bootstrap: 642.109s -> 643.649s (0.24%)

[nnethercote]

Larger benchmarks have more CGUs, and reducing the max number of CGUs means they do more CGU merging, which increases the chance that a small change causes an outsized amount of codegen being redone. That explains the drastic regressions on the incr-patched results for the big benchmarks like cranelift-codegen, webrender, and cargo.

#112441 is an experiment in the other direction -- it disables all CGU merging in incremental builds, effectively setting the CGU limit to infinity.