[llvm] [AMDGPU] Do not copy PHI operands whose class differs only in AV-ness (PR #226213)

[Pablo Reble via llvm-commits]

================
@@ -7765,6 +7765,17 @@ SIInstrInfo::legalizeOperands(MachineInstr &MI,
       if (!Op.isReg() || !Op.getReg().isVirtual())
         continue;
 
+      // No copy is needed if the classes differ only in AV-ness, e.g. an av_32
----------------
reble wrote:

Thanks for your feedback. I'll rework this and run some tests locally, making sure we don't introduce any performance regressions on MFMA-heavy kernels.

https://github.com/llvm/llvm-project/pull/226213