This is an archive of the discontinued LLVM Phabricator instance.

AMDGPU/GlobalISel: Remove manual store select code
ClosedPublic

Authored by arsenm on Jul 22 2019, 6:10 AM.

Details

Reviewers
tstellar
nhaehnle
Summary

This regresses the weird types that are newly treated as legal load
types, but fixes incorrectly using flat instrucions on SI.

Diff Detail

Event Timeline

arsenm created this revision.Jul 22 2019, 6:10 AM

Which types are regressed?

Which types are regressed?

Types like s128, or v4s16. I'm considering adding dedicated VT-sized types to TableGen to ignore the type and only match the size to avoid needing to list all of these types

Which types are regressed?

Types like s128, or v4s16. I'm considering adding dedicated VT-sized types to TableGen to ignore the type and only match the size to avoid needing to list all of these types

Actually, this fixes v4s16. It will break other types we still let through, like v16s8. I think eventually we'll need to bitcast some loads in a step after D64899 (or at least for stores)

tstellar accepted this revision.Jul 22 2019, 7:46 AM

Ok, this seems fine. Those types aren't really use much.

This revision is now accepted and ready to land.Jul 22 2019, 7:46 AM
arsenm closed this revision.Jul 31 2019, 8:52 PM

r367512