[llvm] [AMDGPU] Detect VALU-after-MFMA hazard in post-RA scheduler's getHazardType (PR #184084)

[Anshil Gandhi via llvm-commits]

================
@@ -234,6 +234,10 @@ GCNHazardRecognizer::getHazardType(SUnit *SU, int Stalls) {
       checkMAIVALUHazards(MI) > 0)
     return HazardType;
 
+  if (SIInstrInfo::isVALU(*MI) && !SIInstrInfo::isMFMA(*MI) &&
+      checkVALUImmediatelyAfterMFMAHazard(MI) > 0)
+    return HazardType;
----------------
gandhi56 wrote:

A VALU instruction cannot be coexecuted during the first pass of an 8-pass MFMA execution. on hardware. So the hazard recognizer should signal a NoopHazard because MFMA does not necessarily occupy HWVALU and the scheduler also needs to realize the coexec model.

So forcefully returning a NoopHazard would make more sense, right?

https://github.com/llvm/llvm-project/pull/184084