[llvm] [AMDGPU] Correctly insert s_nops for implicit read of SDWA (PR #100276)

[Jay Foad via llvm-commits]

================
@@ -935,6 +935,30 @@ int GCNHazardRecognizer::checkVALUHazards(MachineInstr *VALU) {
           if (Use.isReg() && TRI->regsOverlap(Def, Use.getReg()))
----------------
jayfoad wrote:

> I assume we don't do this / plan on doing this for MI300 (at least for VALUs)?

I can't promise that, but at least scanning all implicit operands would be correct by default. If necessary you could ignore certain implicit operands if you are sure it is still safe. What you are doing in this patch is the opposite of that.

https://github.com/llvm/llvm-project/pull/100276