-
Notifications
You must be signed in to change notification settings - Fork 12.8k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Use BOLT in CI to optimize librustc_driver
#102487
Conversation
@bors try @rust-timer queue |
Awaiting bors try build completion. @rustbot label: +S-waiting-on-perf |
⌛ Trying commit 543c574ff7b86b3e8641a701d73f0b0c90b85a5d with merge 23a2a394654d7fe2fe4228351ed8485dca6aec98... |
💔 Test failed - checks-actions |
This comment has been minimized.
This comment has been minimized.
@bors try @rust-timer queue |
Awaiting bors try build completion. @rustbot label: +S-waiting-on-perf |
⌛ Trying commit 799dd5fc0c85929b600f677e5a1df07c7df3343c with merge 9f86f5f6789c609097d1af2b3c7fb1762d175dc7... |
☀️ Try build successful - checks-actions |
Queued 9f86f5f6789c609097d1af2b3c7fb1762d175dc7 with parent 877877a, future comparison URL. |
Finished benchmarking commit (9f86f5f6789c609097d1af2b3c7fb1762d175dc7): comparison URL. Overall result: ❌ regressions - ACTION NEEDEDBenchmarking 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 @bors rollup=never Instruction countThis is a highly reliable metric that was used to determine the overall result at the top of this comment.
Max RSS (memory usage)ResultsThis 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.
CyclesResultsThis 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.
Footnotes |
The cycle results are not bad, but nothing too impresive sadly. Maybe this isn't really worth it. I think that the Google team that tried this has similar experiences. I wonder if BOLT is less effective for Rust libraries vs C/C++ in general. Another reason might be that BOLT is more effective for binaries than for shared libraries. I'll try to statically link rustc and apply BOLT to it, to see what happens. |
I'd love to see results for statically linking rustc in general, to see how much benefit that provides. |
There are results of a few different variants involving static linking in #97154 |
@bors try @rust-timer queue |
Awaiting bors try build completion. @rustbot label: +S-waiting-on-perf |
Use BOLT in CI to optimize `librustc_driver` Based on rust-lang#94381. r? `@ghost`
☀️ Try build successful - checks-actions |
This comment has been minimized.
This comment has been minimized.
Finished benchmarking commit (8debfc8): comparison URL. Overall result: no relevant changes - no action neededBenchmarking 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 Instruction countThis benchmark run did not return any relevant results for this metric. Max RSS (memory usage)ResultsThis 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.
CyclesThis benchmark run did not return any relevant results for this metric. Binary sizeThis benchmark run did not return any relevant results for this metric. Bootstrap: 629.545s -> 630.771s (0.19%) |
@bors try @rust-timer queue |
This comment has been minimized.
This comment has been minimized.
Use BOLT in CI to optimize `librustc_driver` Based on rust-lang#94381. r? `@ghost`
☀️ Try build successful - checks-actions |
This comment has been minimized.
This comment has been minimized.
Finished benchmarking commit (9928903): comparison URL. Overall result: ❌ regressions - ACTION NEEDEDBenchmarking 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 @bors rollup=never Instruction countThis is a highly reliable metric that was used to determine the overall result at the top of this comment.
Max RSS (memory usage)ResultsThis 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.
CyclesResultsThis 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.
Binary sizeThis benchmark run did not return any relevant results for this metric. Bootstrap: 628.193s -> 626.601s (-0.25%) |
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#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.
Implemented in #116352. |
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.
Based on #94381.
r? @ghost