[llvm] AMDGPU: Handle rewriting VGPR MFMA fed from AGPR copy (PR #153022)

[Jeffrey Byrnes via llvm-commits]

================
@@ -60,6 +64,32 @@ class AMDGPURewriteAGPRCopyMFMAImpl {
     return TII.isMAI(MI) && AMDGPU::getMFMASrcCVDstAGPROp(MI.getOpcode()) != -1;
   }
 
+  /// Find AV_* registers assigned to AGPRs (or virtual registers which were
+  /// already required to be AGPR).
+  ///
+  /// \return the assigned physical register that \p VReg is assigned to if it
+  /// is an AGPR, otherwise MCRegister().
+  MCRegister getAssignedAGPR(Register VReg) const {
+    MCRegister PhysReg = VRM.getPhys(VReg);
+    if (!PhysReg)
+      return MCRegister();
+
+    const TargetRegisterClass *VirtRegRC = MRI.getRegClass(VReg);
+    if (!TRI.hasAGPRs(VirtRegRC))
+      return MCRegister();
+
+    if (!TRI.hasVGPRs(VirtRegRC))
+      return PhysReg;
+
+    // If this is an AV register, we have to check if the actual assignment is
+    // to an AGPR
+    const TargetRegisterClass *AssignedRC = TRI.getPhysRegBaseClass(PhysReg);
+    return TRI.isAGPRClass(AssignedRC) ? PhysReg : MCRegister();
----------------
jrbyrnes wrote:

To echo perlfu: it's unclear why we don't just use this instead of doing the VirtRegRC checks -- comment would be helpful.

https://github.com/llvm/llvm-project/pull/153022