[Mlir-commits] [llvm] [mlir] [AMDGPU] Carry MFMA C/D policy into MIR (PR #236909)

[Matt Arsenault]

================
@@ -3772,3 +3769,17 @@ raw_ostream &operator<<(raw_ostream &OS, const AMDGPU::TargetIDSetting S) {
 }
 
 } // namespace llvm
+
+std::optional<llvm::AMDGPU::MFMACDRegClass>
+llvm::AMDGPU::getMFMACDRegClass(const CallBase &Call) {
+  Attribute A = Call.getAttributes().getFnAttr(MFMACDRegClassAttr);
+  if (!A.isValid())
+    return MFMACDRegClass::Auto;
+  if (!Call.getIntrinsicID() || !Intrinsic::getBaseName(Call.getIntrinsicID())
+                                     .starts_with("llvm.amdgcn.mfma."))
----------------
arsenm wrote:

This is not a reasonable way to implement passing information into the backend 

https://github.com/llvm/llvm-project/pull/236909