This is an archive of the discontinued LLVM Phabricator instance.

AMDGPU: Don't rematerialize mov with implicit operands
ClosedPublic

Authored by arsenm on Feb 4 2019, 1:02 PM.

Details

Summary

This was pulling the mov used for register indexing on gfx9 out of the loop.

Diff Detail

Event Timeline

arsenm created this revision.Feb 4 2019, 1:02 PM
rampitec added inline comments.Feb 4 2019, 1:08 PM
lib/Target/AMDGPU/SIInstrInfo.cpp
144

The all have implicit exec. Is that reflected in MI.getDesc().getNumOperands()?

arsenm marked an inline comment as done.Feb 4 2019, 1:09 PM
arsenm added inline comments.
lib/Target/AMDGPU/SIInstrInfo.cpp
144

Yes, this is the second meaning of implicit operands (not present in the static instruction definition). I originally hardcoded this to 3

rampitec accepted this revision.Feb 4 2019, 1:10 PM

LGTM

This revision is now accepted and ready to land.Feb 4 2019, 1:10 PM
arsenm closed this revision.Feb 4 2019, 2:27 PM

r353101