OSDN Git Service

(root) / android-x86 / external-llvm-project.git / commit
? search:
summary | shortlog | log | commit | commitdiff | tree
(parent: cf9fa2a) | patch
[AMDGPU] Don't combine memory intrs to v3i16
authorSebastian Neubauer <sebastian.neubauer@amd.com>
Tue, 21 Jul 2020 08:28:12 +0000 (10:28 +0200)
committerSebastian Neubauer <sebastian.neubauer@amd.com>
Wed, 22 Jul 2020 10:44:01 +0000 (12:44 +0200)
commit2c659082bda6319732118e746fe025d8d5f9bfac
treee1941b4130f3da1521f018f8e99da92cd098b863tree | snapshot
parentcf9fa2aa018105313efc77c7fb6c5d9dce700297commit | diff
[AMDGPU] Don't combine memory intrs to v3i16

v3i16 and v3f16 currently cannot be legalized and lowered so they should
not be emitted by inst combining.

Moved the check down to still allow extracting 1 or 2 elements via the dmask.

Fixes image intrinsics being combined to return v3x16.

Differential Revision: https://reviews.llvm.org/D84223
llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp diff | blob | history
llvm/test/Transforms/InstCombine/AMDGPU/amdgcn-demanded-vector-elts.ll diff | blob | history
external/llvm-project
About OSDN
Find Software
Develop Software
Help