Change TaskDeps to start preallocated with 128 capacity

[FractalFir]

This is a tiny change that makes TaskDeps::read_set start preallocated with capacity for 128 elements.

From local profiling, it looks like TaskDeps::read_set is one of the most-often resized hash-sets in rustc.

r? @Noratrieb

rustbot has assigned @Noratrieb.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

@bors try @rust-timer queue

⌛ Trying commit 02aa397 with merge 0fd98fe...

☀️ Try build successful - checks-actions
Build commit: 0fd98fe (0fd98fe325e1591d05ebd43e902c9273501ab6d3)

Finished benchmarking commit (0fd98fe): comparison URL.

Overall result: ❌✅ regressions and improvements - please read the text below

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 the most reliable metric that we have; it was used to determine the overall result at the top of this comment. However, even this metric can sometimes exhibit noise.

	mean	range	count
Regressions ❌ (primary)	0.4%	[0.2%, 0.6%]	11
Regressions ❌ (secondary)	0.4%	[0.3%, 0.4%]	6
Improvements ✅ (primary)	-3.8%	[-5.5%, -0.4%]	8
Improvements ✅ (secondary)	-0.5%	[-0.5%, -0.5%]	1
All ❌✅ (primary)	-1.4%	[-5.5%, 0.6%]	19

Max RSS (memory usage)

Results (primary 0.0%, secondary -0.1%)

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)	1.0%	[0.4%, 2.8%]	25
Regressions ❌ (secondary)	1.0%	[0.4%, 1.9%]	15
Improvements ✅ (primary)	-0.9%	[-3.4%, -0.4%]	28
Improvements ✅ (secondary)	-1.0%	[-2.5%, -0.4%]	19
All ❌✅ (primary)	0.0%	[-3.4%, 2.8%]	53

Cycles

Results (primary -0.3%, secondary 0.3%)

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)	0.9%	[0.4%, 3.3%]	46
Regressions ❌ (secondary)	1.3%	[0.4%, 6.3%]	69
Improvements ✅ (primary)	-1.5%	[-4.9%, -0.4%]	45
Improvements ✅ (secondary)	-0.9%	[-3.6%, -0.4%]	59
All ❌✅ (primary)	-0.3%	[-4.9%, 3.3%]	91

Binary size

Results (primary -0.7%)

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)	-0.7%	[-0.7%, -0.7%]	1
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	-0.7%	[-0.7%, -0.7%]	1

Bootstrap: 770.447s -> 768.842s (-0.21%)
Artifact size: 359.63 MiB -> 359.69 MiB (0.02%)

This seems like a perf improvement to me - since the average is down, and the regressions are much smaller than the improvements.

I'll have to look a bit more into the regressions, since, at least at the first glance, some of them seem a bit odd:

Eg. unicode-normalization-0.1.19 opt incr-patched: println sees a regression in the backend, which I don't think this change touches.

r? @nnethercote Since this is a micro-optimization, I think you are the right person to review this.

What is the general policy for PRs that have both regressions and improvements? Does each regression need to be separately justifed, or do they just need to be outweight by the perf gains?

What is the distribution of sizes of the set? You can probably add a Drop impl and just eprintln! the len() of the set and then pass that through counts or similar to get that.

IOW, is 1024 the right point? Or perhaps 2048, or some other point on the curve -- I could also imagine wanting to jump up (e.g., most sets are 10 elements but then immediately grow to 2048 elements).

I agree with @Mark-Simulacrum, some counts results would be useful. I don't have any sense how many TaskDeps values are created and how big they get.

Suggested change

	read_set: FxHashSet::with_capacity_and_hasher(1024, FxBuildHasher::default()),
	read_set: FxHashSet::with_capacity_and_hasher(1024, Default::default()),

would work too I think.

It works - thanks for the suggestion.

I did not know about counts - it will probably help a lot with further profiling!

This is the result of counts:

(  1)   577194 (82.8%, 82.8%): read_set.len():0
(  2)    11768 ( 1.7%, 84.5%): read_set.len():9
(  3)    10303 ( 1.5%, 86.0%): read_set.len():8
(  4)     8670 ( 1.2%, 87.2%): read_set.len():11
(  5)     8655 ( 1.2%, 88.5%): read_set.len():10
(  6)     7165 ( 1.0%, 89.5%): read_set.len():12
(  7)     5010 ( 0.7%, 90.2%): read_set.len():16
(  8)     4740 ( 0.7%, 90.9%): read_set.len():14
(  9)     4511 ( 0.6%, 91.5%): read_set.len():13
( 10)     3949 ( 0.6%, 92.1%): read_set.len():17
( 11)     3493 ( 0.5%, 92.6%): read_set.len():18
( 12)     3347 ( 0.5%, 93.1%): read_set.len():15
( 13)     3281 ( 0.5%, 93.6%): read_set.len():19
( 14)     2563 ( 0.4%, 93.9%): read_set.len():20
( 15)     2138 ( 0.3%, 94.2%): read_set.len():23
( 16)     1932 ( 0.3%, 94.5%): read_set.len():21
( 17)     1767 ( 0.3%, 94.8%): read_set.len():22
( 18)     1736 ( 0.2%, 95.0%): read_set.len():24
( 19)     1480 ( 0.2%, 95.2%): read_set.len():27
( 20)     1445 ( 0.2%, 95.4%): read_set.len():25
( 21)     1412 ( 0.2%, 95.6%): read_set.len():32
( 22)     1272 ( 0.2%, 95.8%): read_set.len():26
( 23)     1233 ( 0.2%, 96.0%): read_set.len():33
( 24)     1198 ( 0.2%, 96.2%): read_set.len():29
( 25)     1121 ( 0.2%, 96.3%): read_set.len():28
( 26)     1088 ( 0.2%, 96.5%): read_set.len():39
( 27)      936 ( 0.1%, 96.6%): read_set.len():31
( 28)      823 ( 0.1%, 96.7%): read_set.len():30
( 29)      814 ( 0.1%, 96.8%): read_set.len():34
( 30)      772 ( 0.1%, 97.0%): read_set.len():35
( 31)      737 ( 0.1%, 97.1%): read_set.len():42
( 32)      680 ( 0.1%, 97.2%): read_set.len():45
( 33)      662 ( 0.1%, 97.3%): read_set.len():37
( 34)      630 ( 0.1%, 97.3%): read_set.len():40
( 35)      606 ( 0.1%, 97.4%): read_set.len():38
( 36)      568 ( 0.1%, 97.5%): read_set.len():47
( 37)      565 ( 0.1%, 97.6%): read_set.len():48
( 38)      560 ( 0.1%, 97.7%): read_set.len():41
( 39)      551 ( 0.1%, 97.8%): read_set.len():44
( 40)      545 ( 0.1%, 97.8%): read_set.len():36
( 41)      472 ( 0.1%, 97.9%): read_set.len():49
( 42)      430 ( 0.1%, 98.0%): read_set.len():46
( 43)      423 ( 0.1%, 98.0%): read_set.len():43
( 44)      421 ( 0.1%, 98.1%): read_set.len():50
( 45)      417 ( 0.1%, 98.1%): read_set.len():51
( 46)      390 ( 0.1%, 98.2%): read_set.len():60
( 47)      325 ( 0.0%, 98.2%): read_set.len():52
( 48)      308 ( 0.0%, 98.3%): read_set.len():56
( 49)      286 ( 0.0%, 98.3%): read_set.len():62
( 50)      278 ( 0.0%, 98.4%): read_set.len():55
( 51)      274 ( 0.0%, 98.4%): read_set.len():59
( 52)      268 ( 0.0%, 98.4%): read_set.len():53
( 53)      268 ( 0.0%, 98.5%): read_set.len():54
( 54)      248 ( 0.0%, 98.5%): read_set.len():57
( 55)      231 ( 0.0%, 98.6%): read_set.len():61
( 56)      220 ( 0.0%, 98.6%): read_set.len():78
( 57)      219 ( 0.0%, 98.6%): read_set.len():64
( 58)      208 ( 0.0%, 98.6%): read_set.len():68
( 59)      191 ( 0.0%, 98.7%): read_set.len():58
( 60)      176 ( 0.0%, 98.7%): read_set.len():65
( 61)      173 ( 0.0%, 98.7%): read_set.len():69
( 62)      168 ( 0.0%, 98.7%): read_set.len():116
( 63)      168 ( 0.0%, 98.8%): read_set.len():74
( 64)      162 ( 0.0%, 98.8%): read_set.len():80
( 65)      159 ( 0.0%, 98.8%): read_set.len():67
( 66)      142 ( 0.0%, 98.8%): read_set.len():63
( 67)      141 ( 0.0%, 98.9%): read_set.len():66
( 68)      141 ( 0.0%, 98.9%): read_set.len():70
( 69)      141 ( 0.0%, 98.9%): read_set.len():72
( 70)      120 ( 0.0%, 98.9%): read_set.len():71
( 71)      117 ( 0.0%, 98.9%): read_set.len():76
( 72)      116 ( 0.0%, 99.0%): read_set.len():75
( 73)      111 ( 0.0%, 99.0%): read_set.len():77
( 74)      107 ( 0.0%, 99.0%): read_set.len():79
( 75)      107 ( 0.0%, 99.0%): read_set.len():81
( 76)      105 ( 0.0%, 99.0%): read_set.len():86
( 77)       99 ( 0.0%, 99.0%): read_set.len():85
( 78)       93 ( 0.0%, 99.0%): read_set.len():112
( 79)       92 ( 0.0%, 99.1%): read_set.len():91

So, it seems like this HashSet is usually either empty, or it has < 128 elements in it. I'll see if 128 elements is better locally.

I have done some more testing, and 128 seems to be a better cap for performance.

Additonally, I have discovered(by printing the caller on each resize) that ~30% of all resizes of HashSets occur within this file.

(  1)   118543 (28.3%, 28.3%): Resizing insert from Location { file: "/home/michal/rust/compiler/rustc_query_system/src/dep_graph/graph.rs", line: 511, col: 40 }
(  2)    84689 (20.2%, 48.6%): Resizing insert from Location { file: "/home/michal/rust/compiler/rustc_data_structures/src/obligation_forest/mod.rs", line: 678, col: 37 }
(  3)    32348 ( 7.7%, 56.3%): Resizing insert from Location { file: "/home/michal/rust/compiler/rustc_data_structures/src/transitive_relation.rs", line: 104, col: 20 }
(  4)    28910 ( 6.9%, 63.2%): Resizing insert from Location { file: "/home/michal/rust/compiler/rustc_type_ir/src/elaborate.rs", line: 97, col: 30 }
(  5)    26681 ( 6.4%, 69.6%): Resizing insert from Location { file: "compiler/rustc_hir_typeck/src/method/probe.rs", line: 852, col: 28 }
(  6)    18466 ( 4.4%, 74.0%): Resizing insert from Location { file: "/home/michal/rust/compiler/rustc_data_structures/src/graph/scc/mod.rs", line: 689, col: 57 }
(  7)    16360 ( 3.9%, 77.9%): Resizing insert from Location { file: "compiler/rustc_resolve/src/late.rs", line: 3918, col: 44 }

counts is invaluable for this kind of profiling, I use it all the time.

Do you want to do another perf run with the limit set to 128 or even 64?

@bors try @rust-timer queue

⌛ Trying commit e100fd4 with merge a21e95e...

☀️ Try build successful - checks-actions
Build commit: a21e95e (a21e95e41580e2394a6be7b196d7fff5cd833ee1)

@rust-timer build a21e95e

Finished benchmarking commit (a21e95e): comparison URL.

Overall result: ✅ improvements - no 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.

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

Instruction count

This is the most reliable metric that we have; it was used to determine the overall result at the top of this comment. However, even this metric can sometimes exhibit noise.

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-0.6%	[-0.8%, -0.5%]	6
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	-0.6%	[-0.8%, -0.5%]	6

Max RSS (memory usage)

Results (secondary 1.1%)

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)	2.9%	[2.8%, 3.0%]	2
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-2.5%	[-2.5%, -2.5%]	1
All ❌✅ (primary)	-	-	0

Cycles

Results (primary -2.1%, secondary 7.7%)

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)	7.7%	[7.6%, 7.8%]	2
Improvements ✅ (primary)	-2.1%	[-2.1%, -2.1%]	1
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	-2.1%	[-2.1%, -2.1%]	1

Binary size

This benchmark run did not return any relevant results for this metric.

Bootstrap: 770.531s -> 770.268s (-0.03%)
Artifact size: 361.97 MiB -> 361.95 MiB (-0.01%)

I think this is ready to merge - since it is a perf imporvement, and a 1-line change.

@bors r+

📌 Commit e100fd4 has been approved by nnethercote

It is now in the queue for this repository.

🌲 The tree is currently closed for pull requests below priority 100. This pull request will be tested once the tree is reopened.

⌛ Testing commit e100fd4 with merge 4f52199...

☀️ Test successful - checks-actions
Approved by: nnethercote
Pushing 4f52199 to master...

A job failed! Check out the build log: (web) (plain)

Click to see the possible cause of the failure (guessed by this bot)

gh pr comment ${HEAD_PR} -F output.log
shell: /usr/bin/bash -e {0}
##[endgroup]
fatal: ambiguous argument 'HEAD^1': unknown revision or path not in the working tree.
Use '--' to separate paths from revisions, like this:
'git <command> [<revision>...] -- [<file>...]'
##[error]Process completed with exit code 128.

Finished benchmarking commit (4f52199): comparison URL.

Overall result: ✅ improvements - no action needed

@rustbot label: -perf-regression

Instruction count

This is the most reliable metric that we have; it was used to determine the overall result at the top of this comment. However, even this metric can sometimes exhibit noise.

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-0.6%	[-0.8%, -0.5%]	6
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	-0.6%	[-0.8%, -0.5%]	6

Max RSS (memory usage)

Results (primary 2.7%, secondary -1.6%)

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)	2.7%	[2.7%, 2.7%]	1
Regressions ❌ (secondary)	2.4%	[2.4%, 2.4%]	1
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-2.6%	[-3.9%, -0.9%]	4
All ❌✅ (primary)	2.7%	[2.7%, 2.7%]	1

Cycles

This benchmark run did not return any relevant results for this metric.

Binary size

This benchmark run did not return any relevant results for this metric.

Bootstrap: 765.189s -> 765.417s (0.03%)
Artifact size: 362.04 MiB -> 362.05 MiB (0.00%)