[PATCH] D59829: AMDGPU: An extension to promote constant offset to the immediate

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: lib/Target/AMDGPU/SILoadStoreOptimizer.cpp:1230-1231
+  if (!BaseLoDef ||  !BaseHiDef ||
+      (BaseLoDef->getOpcode() != AMDGPU::V_ADD_I32_e32 &&
+       BaseLoDef->getOpcode() != AMDGPU::V_ADD_I32_e64)   ||
+      (BaseHiDef->getOpcode() != AMDGPU::V_ADDC_U32_e32 &&
----------------
cfang wrote:
> arsenm wrote:
> > Where are the _e32 versions coming from? I don't think you should be seeing these at this point.
> > 
> > You would need to verify the carry out is dead here. You should add a testcase where the vcc def of the add is used
> These _e32 version was generated in "SI Fold Operands". If you think should should not happen, we may need to fix that first.
This one is complicated. This still needs verification that the vcc use isn't needed


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D59829/new/

https://reviews.llvm.org/D59829