[llvm] [AMDGPU][True16][MC] true16 for v_fma_f16 (PR #119477)

[Matt Arsenault via llvm-commits]

================
@@ -4033,14 +4033,29 @@ MachineInstr *SIInstrInfo::convertToThreeAddress(MachineInstr &MI,
   if (Src0Literal && !ST.hasVOP3Literal())
     return nullptr;
 
-  unsigned NewOpc = IsFMA ? IsF16 ? AMDGPU::V_FMA_F16_gfx9_e64
-                                  : IsF64 ? AMDGPU::V_FMA_F64_e64
-                                          : IsLegacy
-                                                ? AMDGPU::V_FMA_LEGACY_F32_e64
-                                                : AMDGPU::V_FMA_F32_e64
-                          : IsF16 ? AMDGPU::V_MAD_F16_e64
-                                  : IsLegacy ? AMDGPU::V_MAD_LEGACY_F32_e64
-                                             : AMDGPU::V_MAD_F32_e64;
+  auto getNewFMAInst = [&]() {
+    if (IsFMA) {
+      if (IsF16)
+        return ST.hasTrue16BitInsts() ? AMDGPU::V_FMA_F16_gfx9_fake16_e64
+                                      : AMDGPU::V_FMA_F16_gfx9_e64;
+      else if (IsF64)
+        return AMDGPU::V_FMA_F64_e64;
----------------
arsenm wrote:

No else after return 

https://github.com/llvm/llvm-project/pull/119477