[PATCH] D32036: AMDGPU: SimplifyDemandedElts for image intrinsics

[Marek Olšák via Phabricator via llvm-commits]

mareko added inline comments.


================
Comment at: lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp:1762
+
+          NewCall->setArgOperand(3, ConstantInt::get(DMask->getType(), NewDMask));
+        }
----------------
Gather4 opcodes always return 4 VGPRs and DMASK has a different meaning. Specifically, Gather4 reads 4 texels from memory and DMASK selects which color component is returned for the texels (i.e. 4x red channel, or 4x green channel, etc.) So DMASK shouldn't be changed by the compiler for gather4 opcodes.


https://reviews.llvm.org/D32036