If there were no free VGPRs we would need two emergency spill slots for register
scavenging during PEI/frame index elimination. Reuse 'ResultReg' for scale
calculation so that only one spill is needed.
Details
- Reviewers
arsenm - Commits
- rG2cbb8c946a6e: [AMDGPU] Reuse register during frame index elimination
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp | ||
---|---|---|
1144–1145 | Typo Resue | |
llvm/test/CodeGen/AMDGPU/pei-scavenge-vgpr-spill.mir | ||
3 | Add other subtarget run lines to catch the different add handling? | |
24 | I assume these live ins are just to force all registers to be live? I think this will break when things are finally inverted to scavenge from the end of the block. I think there should also be an implicit use on the S_ENDPGM (I think you can use the all VGPR reg mask to shorten it) | |
38 | We ended up with an add of 0 which doesn't seem ideal |
llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp | ||
---|---|---|
1144–1145 | Same typo exists in the review title too |
Typo Resue