[PATCH] D133489: [AMDGPU] Fix shrinking of F16 FMA on newer subtargets

Thu Sep 8 08:24:23 PDT 2022

foad added inline comments.

================
Comment at: llvm/test/CodeGen/AMDGPU/gfx10-shrink-mad-fma.mir:195
     %1:vgpr_32 = IMPLICIT_DEF
-    %2:vgpr_32 = V_FMA_F16_e64 0, 18688, 0, %0, 0, %1, 0, 0, implicit $mode, implicit $exec
     SI_RETURN implicit %2
----------------
Joe_Nash wrote:
> I guess the test case before was technically invalid, because V_FMA_F16_e64 should not exist on GFX10, only GFX8. But afaik we do not check whether an instruction is valid on the subtarget in SIInstrInfo::verifyInstruction.
Right. See https://reviews.llvm.org/D129506#inline-1246267.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D133489/new/

https://reviews.llvm.org/D133489