[llvm] [AArch64] Fold scalar-to-vector shuffles into DUP/FMOV (PR #166962)

[Amina Chabane via llvm-commits]

================
@@ -696,6 +696,15 @@ static bool is64bitDefwithZeroHigh64bit(MachineInstr *MI,
   const TargetRegisterClass *RC = MRI->getRegClass(MI->getOperand(0).getReg());
   if (RC != &AArch64::FPR64RegClass)
     return false;
+  if (MI->getOpcode() == TargetOpcode::COPY) {
----------------
Amichaxx wrote:

Perhaps I am misunderstanding, but I believe cross-class copies are not elided due to this block in `MachineCopyPropagation::isForwardableRegClassCopy`.
```
  Register UseDstReg = UseICopyOperands->Destination->getReg();
  bool Found = false;
  bool IsCrossClass = false;
  for (const TargetRegisterClass *RC : TRI->regclasses()) {
    if (RC->contains(CopySrcReg) && RC->contains(UseDstReg)) {
      Found = true;
      if (TRI->getCrossCopyRegClass(RC) != RC) {
        IsCrossClass = true;
        break;
      }
    }
  }
  if (!Found)
    return false;
  if (!IsCrossClass)
    return true;
```
Found remains false because they share no register class and the COPY is not forwarded. After which, I believe a COPY between GPR to FPR would always become an fmov.

https://github.com/llvm/llvm-project/pull/166962