[llvm] [AMDGPU] Do not run GCNNSAReassign pass for GFX12 (PR #78185)

[Jay Foad via llvm-commits]

================
@@ -237,7 +237,7 @@ GCNNSAReassign::CheckNSA(const MachineInstr &MI, bool Fast) const {
 
 bool GCNNSAReassign::runOnMachineFunction(MachineFunction &MF) {
   ST = &MF.getSubtarget<GCNSubtarget>();
-  if (!ST->hasNSAEncoding())
+  if (!ST->hasNSAEncoding() || ST->getGeneration() > GCNSubtarget::GFX11)
----------------
jayfoad wrote:

It _only_ has NSA encoding, so there is no point in running a pass that tries to optimize it back to a non-NSA encoded form (I guess we could add a `hasNonNSAEncoding` hook for that).

https://github.com/llvm/llvm-project/pull/78185