[llvm-branch-commits] [llvm] [AMDGPU] Create V_FMAAK_F16/V_FMAMK_F16 in true16 with imm folding (PR #173317)

[Matt Arsenault via llvm-branch-commits]

================
@@ -3880,6 +3878,19 @@ bool SIInstrInfo::foldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
       // These come before src2.
       removeModOperands(UseMI);
       UseMI.setDesc(get(NewOpc));
+
+      if (NewOpc == AMDGPU::V_FMAAK_F16_t16 ||
+          NewOpc == AMDGPU::V_FMAAK_F16_fake16) {
+        auto Tmp = MRI->createVirtualRegister(
+            NewOpc == AMDGPU::V_FMAAK_F16_t16 ? &AMDGPU::VGPR_16_Lo128RegClass
+                                              : &AMDGPU::VGPR_32_Lo128RegClass);
----------------
arsenm wrote:

Avoid hardcoding the class like above 

https://github.com/llvm/llvm-project/pull/173317