[AMDGPU] Mark SMEM cache invalidations as not reading memory
authorJay Foad <jay.foad@amd.com>
Thu, 19 May 2022 16:46:54 +0000 (17:46 +0100)
committerJay Foad <jay.foad@amd.com>
Fri, 20 May 2022 16:18:03 +0000 (17:18 +0100)
commit9af56c676e40efa551e899675e902cbb3f0db0b6
tree05a65393832d4e52dd8bbd5fe93a3d4d0bf23191
parent30628b0eccf89c36dbd176e925c7fc5c3bfa519d
[AMDGPU] Mark SMEM cache invalidations as not reading memory

This brings the MachineInstrs in line with the corresponding intrinsics
which have side effects but do not access memory. It also matches how
BUF cache invalidation instructions are defined.

The lit test changes are just because the machine scheduler previously
treated them like loads, and added an artificial scheduling edge from
them to the exit SU, which caused them to be scheduled earlier.

Differential Revision: https://reviews.llvm.org/D126074
llvm/lib/Target/AMDGPU/SMInstructions.td
llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.dcache.inv.vol.ll
llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.dcache.wb.ll
llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.dcache.wb.vol.ll