projects / platform / upstream / llvm.git / commit
? search:
summary | shortlog | log | commit | commitdiff | tree
(parent: b4b7c98) | patch
[AMDGPU] Propagate AGPR RC from PHI to its PHI operands
authorStanislav Mekhanoshin <Stanislav.Mekhanoshin@amd.com>
Thu, 2 Apr 2020 23:06:45 +0000 (16:06 -0700)
committerStanislav Mekhanoshin <Stanislav.Mekhanoshin@amd.com>
Fri, 3 Apr 2020 18:23:02 +0000 (11:23 -0700)
commit0462795095e57d9cac3c68ff1f28bc4bef8c4865
tree5328f7268dd6abeba143651b3765ec75007dbbb2tree | snapshot
parentb4b7c989d6f1f255b49c34caf38748062d5402cccommit | diff
[AMDGPU] Propagate AGPR RC from PHI to its PHI operands

We can fix register class of PHI based on its all AGPR uses.
That leaves behind all PHIs which were already processed
earlier. Propagate RC back to PHI operands of a PHI.

Differential Revision: https://reviews.llvm.org/D77344
llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp diff | blob | history
llvm/test/CodeGen/AMDGPU/mfma-loop.ll diff | blob | history
Domain: System / Toolchain;