[perf] Disable inline_in_all_cgus when LTO is enabled #65052
Conversation
|
r? @estebank (rust_highfive has picked a reviewer for you, use r? to override) |
rust-highfive
added
the
S-waiting-on-review
|
@bors try @rust-timer queue A perf run can't hurt, I guess. |
|
Awaiting bors try build completion |
[perf] Disable inline_in_all_cgus when LTO is enabled I tried to run lolbench with this flag altered; however lolbench either errored or got stuck, so I end up running only 3 of the benchmarks manually. One showed no changes, another one showed 20% perf improvement somehow, and another showed slight (1%) regression. So I'm thinking of benchmarking it with perf instead; if we still have ThinLTO on CI then we can measure stage2 compiler performance as a metric of how much it affects generated code.
|
☀️ Try build successful - checks-azure |
|
Queued 8b81ee4 with parent 0221e26, future comparison URL. |
|
Finished benchmarking try commit 8b81ee4, comparison URL. |
|
Everything within the margin of noise, is ThinLTO used for benchmarks? |
|
It should be, yes, for release builds by default I believe. |
|
@ishitatsuyuki |
|
@bjorn3 I am aware of that, and what I expected is that this change will indirectly affect stage2 compiler, which in turn can be benchmarked for performance difference. So far what I have heard is that ThinLTO is enabled for both compiler builds and perf, so everything being identical sounds really strange. Something more likely is that the code/flag being changed has no effect at all, so maybe I should ask who had implemented this originally - |
| @@ -65,7 +65,7 @@ pub trait MonoItemExt<'tcx>: fmt::Debug { | |||
| fn instantiation_mode(&self, tcx: TyCtxt<'tcx>) -> InstantiationMode { | |||
| let inline_in_all_cgus = | |||
| tcx.sess.opts.debugging_opts.inline_in_all_cgus.unwrap_or_else(|| { | |||
| tcx.sess.opts.optimize != OptLevel::No | |||
| tcx.sess.opts.optimize != OptLevel::No && if tcx.sess.lto() == Lto::No | |||
There was a problem hiding this comment.
There is an if after the &&. Why does this code compile?
There was a problem hiding this comment.
Oh good catch, and yeah I have no idea why this happened to compile. I'll push a fix later.
There was a problem hiding this comment.
Strange. I can't reproduce on play.rust-lang.org.
|
Oh shit this file is dead code. The code is inside |
ishitatsuyuki
force-pushed
the
no-inline-cgu
branch
from
5da9200
to
3e0a02a
Compare
|
@Mark-Simulacrum Sorry, can I get another perf run? |
|
@bors try @rust-timer queue |
|
Awaiting bors try build completion |
|
@bors try |
1 similar comment
|
@bors try |
[perf] Disable inline_in_all_cgus when LTO is enabled I tried to run lolbench with this flag altered; however lolbench either errored or got stuck, so I end up running only 3 of the benchmarks manually. One showed no changes, another one showed 20% perf improvement somehow, and another showed slight (1%) regression. So I'm thinking of benchmarking it with perf instead; if we still have ThinLTO on CI then we can measure stage2 compiler performance as a metric of how much it affects generated code.
|
☀️ Try build successful - checks-azure |
|
Queued c5643b3 with parent 9e35a28, future comparison URL. |
|
Finished benchmarking try commit c5643b3, comparison URL. |
|
Checked this and it seems it’s mostly a big lose. Closing. |
I tried to run lolbench with this flag altered; however lolbench either errored or got stuck, so I end up running only 3 of the benchmarks manually. One showed no changes, another one showed 20% perf improvement somehow, and another showed slight (1%) regression.
So I'm thinking of benchmarking it with perf instead; if we still have ThinLTO on CI then we can measure stage2 compiler performance as a metric of how much it affects generated code.