[llvm] [AMDGPU][GFX11] Do not rewrite V_FMA/FMAC_* to V_FMAAK_F16_t16 on operand legalization. (PR #66202)

[Ivan Kosarev via llvm-commits]

================
@@ -3318,6 +3318,12 @@ bool SIInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
       if (pseudoToMCOpcode(NewOpc) == -1)
         return false;
 
+      // V_FMAAK_F16_t16 takes VGPR_32_Lo128 operands, so the rewrite
+      // would also require restricting their register classes. For now
----------------
kosarev wrote:

This looks irrelevant to me. The comment aims to explain why we bail out for fake True16 and `VGPR_32_Lo128`, not `VGPR_16_Lo128`. The instruction in the condition is supposed to be changed to `V_FMAAK_F16_fake16` when we have one, and then removed along with the rest of the fake16 support.

https://github.com/llvm/llvm-project/pull/66202