Build `rustc` with a single CGU on x64 Linux #115554

Kobzol · 2023-09-05T07:05:34Z

This PR adds the rust.codegen-units=1 setting when compiling the 64-bit Linux rustc artifact (the one used for try builds and Linux rustup distribution). This had mixed results in the past, however after the bump to LLVM 17, the results now seem pretty incredible. Instruction counts, cycles, wall time, max RSS and even artifact sizes see large improvements.

The last try build with this setting took 1h 8m, which is basically the same duration for try builds that we have seen recently. So there shouldn't be any large hit to CI/build time.

I hope that this could potentially also reduce codegen noise of rustc a little bit, since small changes within a single rustc crate should no longer perturb optimizations because of CGU movement. We still do cross-crate LTO, so it won't eliminate it though.

r? @Mark-Simulacrum

Kobzol · 2023-09-05T07:06:57Z

@bors try @rust-timer queue

bors · 2023-09-05T07:07:05Z

⌛ Trying commit 1bffa88 with merge 2d8544092e7b47541698b5a6425f9f1aa131b994...

Zoxc · 2023-09-05T07:19:19Z

Why only on the Linux builder? This should be done on every builder.

Kobzol · 2023-09-05T07:22:29Z

Sure, but we don't have to do that in a single step. We can easily gauge the CI build time and compile performance effects of this on Linux, thanks to try builds and perf.RLO, but it's not that easy for other targets. It's possible that on Windows/macOS this might be a hit to CI times, so we should first investigate that (along with checking how it affects their performance, although that should hopefully be an improvement).

It's also possible that this might cause some issues for nightly users, so I would like to test it for a few weeks on nightly on Linux first, before we enable it globally across the board.

lqd · 2023-09-05T08:16:29Z

It's possible that on Windows/macOS this might be a hit to CI times, so we should first investigate that (along with checking how it affects their performance, although that should hopefully be an improvement)

Are we expecting LVM17 to be different from our previous measurements of CI times and performance?

Kobzol · 2023-09-05T08:18:33Z

Well, some things might have changed in the meantime. Apple CI is now much faster (so that's a good thing). And performance has changed a lot on Linux (but again, towards improvements, so hopefully this will be true on other platforms too). It looks like things are improving in general, but I'd still like to do this on Linux first to see if there are no unexpected issues with 1 CGU.

bors · 2023-09-05T08:19:58Z

☀️ Try build successful - checks-actions
Build commit: 2d8544092e7b47541698b5a6425f9f1aa131b994 (2d8544092e7b47541698b5a6425f9f1aa131b994)

lqd · 2023-09-05T08:21:01Z

I'm only talking about the comment about other builders, not this PR or the landing strategy: do we need to re-run the 1 CGU PRs on the other builders to gather up-to-date timings on LLVM17 and the new apple builder?

Kobzol · 2023-09-05T08:25:26Z

I think that it would be good, yes. It also depends how we want to enable it for other builders. We can allowlist it for Windows and macOS and test these two separately, or just enable it across the board for everything and then take a look at the total CI time.

rust-timer · 2023-09-05T12:21:30Z

Finished benchmarking commit (2d8544092e7b47541698b5a6425f9f1aa131b994): 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)	0.8%	[0.3%, 2.2%]	36
Regressions ❌ (secondary)	1.4%	[0.3%, 2.4%]	21
Improvements ✅ (primary)	-1.2%	[-4.1%, -0.3%]	89
Improvements ✅ (secondary)	-1.7%	[-3.6%, -0.4%]	114
All ❌✅ (primary)	-0.6%	[-4.1%, 2.2%]	125

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)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-2.1%	[-4.9%, -0.6%]	74
Improvements ✅ (secondary)	-3.0%	[-7.6%, -0.8%]	109
All ❌✅ (primary)	-2.1%	[-4.9%, -0.6%]	74

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)	-	-	0
Regressions ❌ (secondary)	2.7%	[2.4%, 3.1%]	4
Improvements ✅ (primary)	-2.6%	[-6.9%, -0.9%]	101
Improvements ✅ (secondary)	-4.1%	[-15.4%, -0.9%]	121
All ❌✅ (primary)	-2.6%	[-6.9%, -0.9%]	101

Binary size

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

Bootstrap: 627.535s -> 623.723s (-0.61%)
Artifact size: 316.30 MiB -> 271.45 MiB (-14.18%)

Mark-Simulacrum · 2023-09-10T15:51:47Z

@bors r rollup=never

bors · 2023-09-10T15:51:48Z

📌 Commit 1bffa88 has been approved by Mark-Simulacrum

It is now in the queue for this repository.

bors · 2023-09-10T18:18:57Z

⌛ Testing commit 1bffa88 with merge d4cecd0...

Build `rustc` with a single CGU on x64 Linux This PR adds the `rust.codegen-units=1` setting when compiling the 64-bit Linux `rustc` artifact (the one used for try builds and Linux rustup distribution). This had mixed results in the past, however after the bump to LLVM 17, the results now seem pretty [incredible](rust-lang#115554 (comment)). Instruction counts, cycles, wall time, max RSS and even artifact sizes see large improvements. The last [try build](https://github.com/rust-lang-ci/rust/actions/runs/6077686494/job/16487768049) with this setting took 1h 8m, which is basically the same duration for try builds that we have seen recently. So there shouldn't be any large hit to CI/build time. I hope that this could potentially also reduce codegen noise of `rustc` a little bit, since small changes within a single `rustc` crate should no longer perturn optimizations because of CGU movement. We still do cross-crate LTO, so it won't eliminate it though. r? `@Mark-Simulacrum`

klensy · 2023-09-10T18:23:33Z

This affects cargo too https://perf.rust-lang.org/compare.html?start=626a6ab93fafd01b37b1e26c96cb6eec0d39f3eb&end=2d8544092e7b47541698b5a6425f9f1aa131b994&stat=instructions:u&tab=artifact-size, so not only rustc? Probably other toolchain components, but they didn't measured here.

So this also set codegen-units-std = 1, probably the source of change for other components

rust/config.example.toml

Line 460 in 7418413

#codegen-units-std = codegen-units

klensy · 2023-09-10T19:46:19Z

But at the same time, libstd.so size didn't changed at all, strange.

bors · 2023-10-01T11:46:28Z

📌 Commit ca59652 has been approved by Mark-Simulacrum

It is now in the queue for this repository.

bors · 2023-10-01T12:49:32Z

⌛ Testing commit ca59652 with merge 871407a...

bors · 2023-10-01T14:37:45Z

☀️ Test successful - checks-actions
Approved by: Mark-Simulacrum
Pushing 871407a to master...

rust-timer · 2023-10-01T16:00:37Z

Finished benchmarking commit (871407a): comparison URL.

Overall result: ❌✅ regressions and improvements - ACTION NEEDED

Next Steps: If you can justify the regressions found in this perf run, please indicate this with @rustbot label: perf-regression-triaged along with sufficient written justification. If you cannot justify the regressions please open an issue or create a new PR that fixes the regressions, add a comment linking to the newly created issue or PR, and then add the perf-regression-triaged label to this PR.

@rustbot label: perf-regression
cc @rust-lang/wg-compiler-performance

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)	0.8%	[0.3%, 2.6%]	39
Regressions ❌ (secondary)	1.6%	[0.7%, 2.5%]	23
Improvements ✅ (primary)	-1.2%	[-3.9%, -0.2%]	89
Improvements ✅ (secondary)	-1.6%	[-3.5%, -0.4%]	124
All ❌✅ (primary)	-0.6%	[-3.9%, 2.6%]	128

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)	-	-	0
Regressions ❌ (secondary)	2.9%	[2.9%, 2.9%]	1
Improvements ✅ (primary)	-2.5%	[-4.6%, -0.5%]	90
Improvements ✅ (secondary)	-3.0%	[-6.7%, -0.6%]	118
All ❌✅ (primary)	-2.5%	[-4.6%, -0.5%]	90

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)	-	-	0
Regressions ❌ (secondary)	6.0%	[4.7%, 7.3%]	6
Improvements ✅ (primary)	-2.2%	[-5.9%, -0.6%]	134
Improvements ✅ (secondary)	-3.7%	[-13.9%, -1.0%]	140
All ❌✅ (primary)	-2.2%	[-5.9%, -0.6%]	134

Binary size

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

Bootstrap: 630.664s -> 627.128s (-0.56%)
Artifact size: 317.36 MiB -> 273.31 MiB (-13.88%)

Kobzol · 2023-10-01T17:09:54Z

Cycle count, Max RSS and artifact size is overwhelmingly positive.

@rustbot label: perf-regression-triaged

nnethercote · 2023-10-01T20:29:04Z

Summary of perf results:

icounts: -0.49% mean change across all measurements
cycles: -1.91% (ditto)
wall-time: -1.57% (ditto)
max-rss: -1.96% (ditto)
librustc_driver size: -31.6%
rustdoc artifact size: -20.7%
cargo artifact size: -19.9%

Fantastic results!

The bootstrap time also went down a tiny bit (-0.56%) but that is misleading. The bootstrap benchmark doesn't use the same configuration as the shipped compiler: no codegen-units=1 and no PGO/BOLT/LTO. @Kobzol told me that try build times have gone from something like 1h8m to 1h14m.

Build `rustc` with a single CGU on x64 Linux This PR adds the `rust.codegen-units=1` setting when compiling the 64-bit Linux `rustc` artifact (the one used for try builds and Linux rustup distribution). This had mixed results in the past, however after the bump to LLVM 17, the results now seem pretty [incredible](rust-lang/rust#115554 (comment)). Instruction counts, cycles, wall time, max RSS and even artifact sizes see large improvements. The last [try build](https://github.com/rust-lang-ci/rust/actions/runs/6077686494/job/16487768049) with this setting took 1h 8m, which is basically the same duration for try builds that we have seen recently. So there shouldn't be any large hit to CI/build time. I hope that this could potentially also reduce codegen noise of `rustc` a little bit, since small changes within a single `rustc` crate should no longer perturb optimizations because of CGU movement. We still do cross-crate LTO, so it won't eliminate it though. r? `@Mark-Simulacrum`

Optimize `librustc_driver.so` with BOLT This PR optimizes `librustc_driver.so` on 64-bit Linux CI with BOLT. ### Code One thing that's not clear yet to me how to resolve is how to best pass a linker flag that we need for BOLT (the second commit). It is currently passed unconditionally, which is not a good idea. We somehow have to: 1) Only pass it when we actually plan to use BOLT. How to best do that? `config.toml` entry? Environment variable? CLI flag for bootstrap? BOLT optimization is done by `opt-dist`, therefore bootstrap doesn't know about it by default. 2) Only pass it to `librustc_driver.so` (see performance below). Some discussion of this flag already happened on [Zulip](https://rust-lang.zulipchat.com/#narrow/stream/326414-t-infra.2Fbootstrap/topic/Adding.20a.20one-off.20linker.20flag). ### Performance Latest perf. results can be found [here](rust-lang#102487 (comment)). Note that instruction counts are not very interesting here, there are only regressions on hello world programs. Probably caused by a larger C libstd (?). Summary: - ✔️ `-1.8%` mean improvement in cycle counts across many primary benchmarks. - ✔️ `-1.8%` mean Max-RSS improvement. - ✖️ 34 MiB ( 48%) artifact size regression of `librustc_driver.so`. - This is caused by building `librustc_driver.so` with relocations (which are required for BOLT). Hopefully, it will be [fixed](https://discourse.llvm.org/t/bolt-rfc-a-new-mode-to-rewrite-entire-binary/68674) in the future with BOLT improvements, but now trying to reduce this size increase is [tricky](rust-lang#114649). - Note that the size of this file was recently reduced in rust-lang#115554 by pretty much the same amount (33 MiB). So the size after this PR is basically the same as it was for the last ~year. - ✖️ 1.4 MiB ( 53%) artifact size regression of `rustc`. - This is annoying and pretty much unnecessary. It is caused by the way relocations are currently applied in this PR, because they are applied both to `librustc_driver.so` (where they are needed) and for `rustc` (where they aren't needed), since both are built with a single cargo invocation. We might need e.g. some tricks in the bootstrap `rustc` shim to only apply the relocation flag for the shared library and not for `rustc`. ### CI time CI (try build) got slower by ~5 minutes, which is fine, IMO. It can be further reduced by running LLVM and `librustc_driver` BOLT profile gathering at the same time (now they are gathered separately for LLVM and `librustc_driver`). r? `@Mark-Simulacrum` Also CC `@onur-ozkan,` primarily for the bootstrap linker flag issue.

…ulacrum Optimize `librustc_driver.so` with BOLT This PR optimizes `librustc_driver.so` on 64-bit Linux CI with BOLT. ### Code One thing that's not clear yet to me how to resolve is how to best pass a linker flag that we need for BOLT (the second commit). It is currently passed unconditionally, which is not a good idea. We somehow have to: 1) Only pass it when we actually plan to use BOLT. How to best do that? `config.toml` entry? Environment variable? CLI flag for bootstrap? BOLT optimization is done by `opt-dist`, therefore bootstrap doesn't know about it by default. 2) Only pass it to `librustc_driver.so` (see performance below). Some discussion of this flag already happened on [Zulip](https://rust-lang.zulipchat.com/#narrow/stream/326414-t-infra.2Fbootstrap/topic/Adding.20a.20one-off.20linker.20flag). ### Performance Latest perf. results can be found [here](rust-lang#102487 (comment)). Note that instruction counts are not very interesting here, there are only regressions on hello world programs. Probably caused by a larger C libstd (?). Summary: - ✔️ `-1.8%` mean improvement in cycle counts across many primary benchmarks. - ✔️ `-1.8%` mean Max-RSS improvement. - ✖️ 34 MiB ( 48%) artifact size regression of `librustc_driver.so`. - This is caused by building `librustc_driver.so` with relocations (which are required for BOLT). Hopefully, it will be [fixed](https://discourse.llvm.org/t/bolt-rfc-a-new-mode-to-rewrite-entire-binary/68674) in the future with BOLT improvements, but now trying to reduce this size increase is [tricky](rust-lang#114649). - Note that the size of this file was recently reduced in rust-lang#115554 by pretty much the same amount (33 MiB). So the size after this PR is basically the same as it was for the last ~year. - ✖️ 1.4 MiB ( 53%) artifact size regression of `rustc`. - This is annoying and pretty much unnecessary. It is caused by the way relocations are currently applied in this PR, because they are applied both to `librustc_driver.so` (where they are needed) and for `rustc` (where they aren't needed), since both are built with a single cargo invocation. We might need e.g. some tricks in the bootstrap `rustc` shim to only apply the relocation flag for the shared library and not for `rustc`. ### CI time CI (try build) got slower by ~5 minutes, which is fine, IMO. It can be further reduced by running LLVM and `librustc_driver` BOLT profile gathering at the same time (now they are gathered separately for LLVM and `librustc_driver`). r? `@Mark-Simulacrum` Also CC `@onur-ozkan,` primarily for the bootstrap linker flag issue.

Optimize `librustc_driver.so` with BOLT This PR optimizes `librustc_driver.so` on 64-bit Linux CI with BOLT. ### Code One thing that's not clear yet to me how to resolve is how to best pass a linker flag that we need for BOLT (the second commit). It is currently passed unconditionally, which is not a good idea. We somehow have to: 1) Only pass it when we actually plan to use BOLT. How to best do that? `config.toml` entry? Environment variable? CLI flag for bootstrap? BOLT optimization is done by `opt-dist`, therefore bootstrap doesn't know about it by default. 2) Only pass it to `librustc_driver.so` (see performance below). Some discussion of this flag already happened on [Zulip](https://rust-lang.zulipchat.com/#narrow/stream/326414-t-infra.2Fbootstrap/topic/Adding.20a.20one-off.20linker.20flag). ### Performance Latest perf. results can be found [here](rust-lang/rust#102487 (comment)). Note that instruction counts are not very interesting here, there are only regressions on hello world programs. Probably caused by a larger C libstd (?). Summary: - ✔️ `-1.8%` mean improvement in cycle counts across many primary benchmarks. - ✔️ `-1.8%` mean Max-RSS improvement. - ✖️ 34 MiB ( 48%) artifact size regression of `librustc_driver.so`. - This is caused by building `librustc_driver.so` with relocations (which are required for BOLT). Hopefully, it will be [fixed](https://discourse.llvm.org/t/bolt-rfc-a-new-mode-to-rewrite-entire-binary/68674) in the future with BOLT improvements, but now trying to reduce this size increase is [tricky](rust-lang/rust#114649). - Note that the size of this file was recently reduced in rust-lang/rust#115554 by pretty much the same amount (33 MiB). So the size after this PR is basically the same as it was for the last ~year. - ✖️ 1.4 MiB ( 53%) artifact size regression of `rustc`. - This is annoying and pretty much unnecessary. It is caused by the way relocations are currently applied in this PR, because they are applied both to `librustc_driver.so` (where they are needed) and for `rustc` (where they aren't needed), since both are built with a single cargo invocation. We might need e.g. some tricks in the bootstrap `rustc` shim to only apply the relocation flag for the shared library and not for `rustc`. ### CI time CI (try build) got slower by ~5 minutes, which is fine, IMO. It can be further reduced by running LLVM and `librustc_driver` BOLT profile gathering at the same time (now they are gathered separately for LLVM and `librustc_driver`). r? `@Mark-Simulacrum` Also CC `@onur-ozkan,` primarily for the bootstrap linker flag issue.

Refactor `opt-dist` to simplify local building This PR refactors the `opt-dist` tool to make it easier to invoke it locally, outside of CI, and thus simplify building PGO/BOLT optimized `rustc` builds e.g. for distro maintainers. It should also make it easier to run the PGO/BOLT workflow locally e.g. to profile performance or debug issues (looking at you, rust-lang/rust#115554).

rustbot assigned Mark-Simulacrum Sep 5, 2023

rustbot added A-testsuite Area: The testsuite used to check the correctness of rustc S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. T-infra Relevant to the infrastructure team, which will review and decide on the PR/issue. labels Sep 5, 2023

This comment has been minimized.

Sign in to view

rustbot added the S-waiting-on-perf Status: Waiting on a perf run to be completed. label Sep 5, 2023

This comment has been minimized.

Sign in to view

rustbot added perf-regression Performance regression. and removed S-waiting-on-perf Status: Waiting on a perf run to be completed. labels Sep 5, 2023

Kobzol mentioned this pull request Sep 9, 2023

Use BOLT in CI to optimize librustc_driver #102487

Closed

bors added S-waiting-on-bors Status: Waiting on bors to run and complete tests. Bors will change the label on completion. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Sep 10, 2023

This comment has been minimized.

Sign in to view

bors added S-waiting-on-bors Status: Waiting on bors to run and complete tests. Bors will change the label on completion. and removed S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. labels Oct 1, 2023

bors added the merged-by-bors This PR was explicitly merged by bors. label Oct 1, 2023

bors merged commit 871407a into rust-lang:master Oct 1, 2023
18 of 23 checks passed

rustbot added this to the 1.75.0 milestone Oct 1, 2023

lqd mentioned this pull request Oct 1, 2023

[WIP] Build rustc with a single CGU on x64 Linux #107651

Merged

Kobzol deleted the single-cgu branch October 1, 2023 15:31

rustbot added the perf-regression-triaged The performance regression has been triaged. label Oct 1, 2023

matthiaskrgr mentioned this pull request Oct 1, 2023

ice: stack overflow #116321

Open

This was referenced Oct 1, 2023

Tracking issue for speeding up rustc via its build configuration #103595

Open

Optimize librustc_driver.so with BOLT #116352

Merged

pvdrz mentioned this pull request Oct 4, 2023

Fix divergence from upstream master ferrocene/ferrocene#23

Merged

erikdesjardins mentioned this pull request Mar 15, 2024

Build core, alloc, std with 1 CGU #122522

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Build `rustc` with a single CGU on x64 Linux #115554

Build `rustc` with a single CGU on x64 Linux #115554

Kobzol commented Sep 5, 2023 •

edited

Loading

Kobzol commented Sep 5, 2023

This comment has been minimized.

bors commented Sep 5, 2023

Zoxc commented Sep 5, 2023

Kobzol commented Sep 5, 2023

lqd commented Sep 5, 2023

Kobzol commented Sep 5, 2023

bors commented Sep 5, 2023

This comment has been minimized.

lqd commented Sep 5, 2023

Kobzol commented Sep 5, 2023

rust-timer commented Sep 5, 2023

Mark-Simulacrum commented Sep 10, 2023

bors commented Sep 10, 2023

bors commented Sep 10, 2023

klensy commented Sep 10, 2023 •

edited

Loading

klensy commented Sep 10, 2023

This comment has been minimized.

bors commented Oct 1, 2023

bors commented Oct 1, 2023

bors commented Oct 1, 2023

rust-timer commented Oct 1, 2023

Kobzol commented Oct 1, 2023

nnethercote commented Oct 1, 2023 •

edited

Loading

Build rustc with a single CGU on x64 Linux #115554

Build rustc with a single CGU on x64 Linux #115554

Conversation

Kobzol commented Sep 5, 2023 • edited Loading

Kobzol commented Sep 5, 2023

This comment has been minimized.

bors commented Sep 5, 2023

Zoxc commented Sep 5, 2023

Kobzol commented Sep 5, 2023

lqd commented Sep 5, 2023

Kobzol commented Sep 5, 2023

bors commented Sep 5, 2023

This comment has been minimized.

lqd commented Sep 5, 2023

Kobzol commented Sep 5, 2023

rust-timer commented Sep 5, 2023

Overall result: ❌✅ regressions and improvements - ACTION NEEDED

Mark-Simulacrum commented Sep 10, 2023

bors commented Sep 10, 2023

bors commented Sep 10, 2023

klensy commented Sep 10, 2023 • edited Loading

klensy commented Sep 10, 2023

This comment has been minimized.

bors commented Oct 1, 2023

bors commented Oct 1, 2023

bors commented Oct 1, 2023

rust-timer commented Oct 1, 2023

Overall result: ❌✅ regressions and improvements - ACTION NEEDED

Kobzol commented Oct 1, 2023

nnethercote commented Oct 1, 2023 • edited Loading

Build `rustc` with a single CGU on x64 Linux #115554

Build `rustc` with a single CGU on x64 Linux #115554

Kobzol commented Sep 5, 2023 •

edited

Loading

klensy commented Sep 10, 2023 •

edited

Loading

nnethercote commented Oct 1, 2023 •

edited

Loading