This is an archive of the discontinued LLVM Phabricator instance.

AMDGPU: Quick fix for extreme slowness in spill-scavenge-offset.ll test
ClosedPublic

Authored by nhaehnle on Feb 11 2016, 12:18 PM.

Diff Detail

Event Timeline

nhaehnle updated this revision to Diff 47700.Feb 11 2016, 12:18 PM
nhaehnle retitled this revision from to AMDGPU: Quick fix for extreme slowness in spill-scavenge-offset.ll test.
nhaehnle updated this object.
nhaehnle added a reviewer: arsenm.
nhaehnle added a subscriber: llvm-commits.
arsenm accepted this revision.Feb 11 2016, 12:27 PM
arsenm edited edge metadata.

LGTM

test/CodeGen/AMDGPU/spill-scavenge-offset.ll
14

While you're cleaning this up, can you replace this with the llvm.amdgcn.mbcnt expansion?

This revision is now accepted and ready to land.Feb 11 2016, 12:27 PM
This revision was automatically updated to reflect the committed changes.
nhaehnle marked an inline comment as done.Feb 11 2016, 4:12 PM

I have committed a variant of this now. FYI, I've played around with this a little bit more, and noticed that only the compile for Tonga is slow.

On Verde, the compile only takes 1.5s in debug mode with the default register allocator - still long, but basically reasonable given the size of the test and nothing compared to the craziness of Tonga.

I have no clue what causes this, maybe somebody has an idea.