[llvm] [AMDGPU][CoExecSched] Add free-proximity scheduling heuristic (PR #225334)

[Kazu Hirata via llvm-commits]

================
@@ -463,32 +476,163 @@ SUnit *HardwareUnitInfo::getNextTargetSU(bool LookDeep) const {
   return TargetSU;
 }
 
-void HardwareUnitInfo::insert(SUnit *SU, unsigned BlockingCycles) {
-  if (!AllSUs.insert(SU))
-    llvm_unreachable("HardwareUnit already contains SU!");
+int HardwareUnitInfo::compareDepth(SUnit *Candidate, SUnit *Existing) const {
+  const unsigned CurDepth = Existing->getDepth();
+  const unsigned CandDepth = Candidate->getDepth();
 
-  TotalCycles += BlockingCycles;
+  if (CandDepth > CurDepth)
+    return -1;
+  if (CandDepth == CurDepth)
+    return 0;
+  return 1;
+}
+
+int HardwareUnitInfo::compareRegFreeProximity(SUnit *Candidate,
+                                              SUnit *Existing) const {
+  const MachineInstr *CandMI = Candidate->getInstr();
+  const bool IsMemOp = SIInstrInfo::isDS(*CandMI) ||
+                       SIInstrInfo::isFLAT(*CandMI) ||
+                       SIInstrInfo::isVMEM(*CandMI);
+  if (IsMemOp)
+    return compareDepth(Candidate, Existing);
 
+  auto getRegStats = [](const SUnit *SU) {
+    SmallDenseMap<Register, unsigned, 8> RegMaxUnsched;
+
+    for (const SDep &Pred : SU->Preds) {
+      if (Pred.getKind() != SDep::Data)
+        continue;
+      Register Reg = Pred.getReg();
+      if (!Reg)
+        continue;
+      const SUnit *PredSU = Pred.getSUnit();
+      unsigned Unscheduled = 0;
+      for (const SDep &Succ : PredSU->Succs)
+        if (Succ.getKind() == SDep::Data && !Succ.getSUnit()->isScheduled)
+          ++Unscheduled;
+
+      LLVM_DEBUG({
+        dbgs() << "        pred SU(" << PredSU->NodeNum
+               << ") reg=" << printReg(Reg) << " unschedSuccs=" << Unscheduled
+               << " ";
+        if (PredSU->getInstr())
+          PredSU->getInstr()->print(dbgs(), /*IsStandalone=*/true,
+                                    /*SkipOpers=*/false, /*SkipDebugLoc=*/true);
+        else
+          dbgs() << "<no instr>";
+        dbgs() << "\n";
+      });
+      RegMaxUnsched[Reg] = std::max(RegMaxUnsched[Reg], Unscheduled);
----------------
kazutakahirata wrote:

Could we avoid duplicate hash lookups on `RegMaxUnsched` by binding a reference?

```suggestion
      auto &MaxUnsched = RegMaxUnsched[Reg];
      MaxUnsched = std::max(MaxUnsched, Unscheduled);
```

https://github.com/llvm/llvm-project/pull/225334