This is an archive of the discontinued LLVM Phabricator instance.

[clang][NewPM] Remove exception handling before loading pgo sample profile data
ClosedPublic

Authored by leonardchan on Jun 20 2019, 2:47 PM.

Diff Detail

Repository
rL LLVM

Event Timeline

leonardchan created this revision.Jun 20 2019, 2:47 PM
Herald added a project: Restricted Project. · View Herald TranscriptJun 20 2019, 2:47 PM

See inline comment, but I think we should just drop the testing of the function attribute bit here rather than adjusting the pipeline.

llvm/lib/Passes/PassBuilder.cpp
665–668 ↗(On Diff #205904)

Does the this need to go before the EarlyFPM as well as before the sample profile loader?

Also, this is ... a pretty expensive thing to do... Do we really need this? We've been using ThinLTO and sample PGO with the new PM for a long time and haven't seen any real issue. I think we can probably just leave this difference between the two pass managers and remove the test for this pass running rather than adding it.

CC-ing some others just to double check, but I'm not aware of any issues we've seen with PGO that were because of this discrepancy.

Adding Wei who works on SamplePGO on our end currently.

llvm/lib/Passes/PassBuilder.cpp
665–668 ↗(On Diff #205904)

We don't tend to use exception handling so might not be hitting the issue described in the headline.

If this is fixing an issue with SamplePGO and EH cleanup interactions, can you add a test that tests for that (i.e. is there a code optimization issue currently?).

Is there a bug reference somewhere?

leonardchan marked an inline comment as done.

See inline comment, but I think we should just drop the testing of the function attribute bit here rather than adjusting the pipeline.

Ok. Removed the pipeline change for now.

Is there a bug reference somewhere?

No bug has been filed for this, but I'm not sure if this is a bug if there aren't really any issues with PGO under the new PM.

llvm/lib/Passes/PassBuilder.cpp
665–668 ↗(On Diff #205904)

I'm not entirely familiar with the expected codegen for SamplePGO or PruneEH and don't really know if not adding these 2 passes breaks anything.

Under the legacy PM, exception handling was removed via PruneEH before creating the sample profile loader, so this change was meant to match the behavior since function-attrs and function(simplify-cfg) (the new PM equivalent for -prune-eh) did not initially run before SampleProfileLoaderPass was added to the pipeline.

The original patch that added CodeGen/pgo-sample.c (D21197) doesn't seem to have any codegen tests either and only checks that PruneEH runs before sample profile loader creation, so I'm also unsure of what code optimizations are meant to happen.

FWIW, I think we can wait for a bug to be filed or report come in with some functional test case before we change any behavior here.

I'm just suggesting how to make the test better below.

llvm/lib/Passes/PassBuilder.cpp
665–668 ↗(On Diff #205904)

Going back to this original issue, I think I understand more what is going on...

Specifically, I can imagine that having invoke still in place caused some issues.

So I would just keep the testing that (in the new PM) SimplifyCFG is run before the profile loading pass, and ignore function attrs.

leonardchan marked 2 inline comments as done.
leonardchan edited the summary of this revision. (Show Details)
leonardchan added inline comments.
llvm/lib/Passes/PassBuilder.cpp
665–668 ↗(On Diff #205904)

Done.

This revision is now accepted and ready to land.Jun 21 2019, 6:23 PM
This revision was automatically updated to reflect the committed changes.