This is an archive of the discontinued LLVM Phabricator instance.

[GlobalISel] Add artifact combine for trunc(concat_vectors(a, ...) -> concat_vectors(trunc(a), ...)
AbandonedPublic

Authored by aemerson on Sep 22 2020, 4:23 PM.

Details

Summary

This makes the large G_CONCAT_VECTORS easier to handle.

Diff Detail

Event Timeline

aemerson created this revision.Sep 22 2020, 4:23 PM
aemerson requested review of this revision.Sep 22 2020, 4:23 PM

On further investigation I'm not sure this is the best way forward.

arsenm added inline comments.Sep 23 2020, 5:44 AM
llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-flat.mir
4651–4656

This is worse (although this test is broken since a <4 x s8> value should not be legal, and should be a verifier error to copy this to a physical register

aemerson abandoned this revision.Sep 23 2020, 7:15 AM
aemerson added inline comments.
llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-flat.mir
4651–4656

Yeah I’m going to drop this in favour of custom legalisation.