[AIEX] Add mechanism to release resources from Scoreboard

Author: krishnamtibrewala

Diff for: llvm/lib/Target/AIE/AIEHazardRecognizer.cpp

@@ -108,6 +108,15 @@ FuncUnitWrapper &FuncUnitWrapper::operator|=(const FuncUnitWrapper &Other) {
   return *this;
 }
 
+FuncUnitWrapper &FuncUnitWrapper::operator^=(const FuncUnitWrapper &Other) {
+  // XOR operation with the same FuncUnitWrapper will release resources.
+  Required ^= Other.Required;
+  Reserved ^= Other.Reserved;
+  Slots ^= Other.Slots;
+  MemoryBanks ^= Other.MemoryBanks;
+  return *this;
+}
+
 bool FuncUnitWrapper::conflict(const FuncUnitWrapper &Other) const {
   if ((Required & Other.Required) != 0 || (Slots & Other.Slots) != 0 ||
       (MemoryBanks & Other.MemoryBanks) != 0 ||
@@ -591,6 +600,18 @@ void AIEHazardRecognizer::emitInScoreboard(
                  TII->getMemoryCycles(SchedClass), DeltaCycles, FUDepthLimit);
 }
 
+void AIEHazardRecognizer::releaseFromScoreboard(
+    ResourceScoreboard<FuncUnitWrapper> &TheScoreboard, const MCInstrDesc &Desc,
+    MemoryBankBits MemoryBanks,
+    iterator_range<const MachineOperand *> MIOperands,
+    const MachineRegisterInfo &MRI, int DeltaCycles) const {
+  const unsigned SchedClass = TII->getSchedClass(Desc, MIOperands, MRI);
+  const SlotBits SlotSet =
+      getSlotSet(Desc, *TII->getFormatInterface(), IgnoreUnknownSlotSets);
+  releaseResources(TheScoreboard, ItinData, SchedClass, SlotSet, MemoryBanks,
+                   TII->getMemoryCycles(SchedClass), DeltaCycles, FUDepthLimit);
+}
+
 void AIEHazardRecognizer::enterResources(
     ResourceScoreboard<FuncUnitWrapper> &Scoreboard,
     const InstrItineraryData *ItinData, unsigned SchedClass, SlotBits SlotSet,
@@ -632,6 +653,47 @@ void AIEHazardRecognizer::enterResources(
   });
 }
 
+void AIEHazardRecognizer::releaseResources(
+    ResourceScoreboard<FuncUnitWrapper> &Scoreboard,
+    const InstrItineraryData *ItinData, unsigned SchedClass, SlotBits SlotSet,
+    MemoryBankBits MemoryBanks, SmallVector<int, 2> MemoryAccessCycles,
+    int DeltaCycles, std::optional<int> FUDepthLimit) {
+  assert(Scoreboard.isValidDelta(DeltaCycles));
+
+  // Remove slot usage
+  FuncUnitWrapper EmissionCycle(/*Req=*/0, /*Res=*/0, SlotSet);
+  Scoreboard[DeltaCycles] ^= EmissionCycle;
+
+  // Remove memory bank usage
+  if (!MemoryAccessCycles.empty()) {
+    FuncUnitWrapper MemoryBankAccessCycle(/*Req=*/0, /*Res=*/0, /*SlotSet=*/0,
+                                          MemoryBanks);
+    for (int Cycles : MemoryAccessCycles) {
+      Scoreboard[DeltaCycles + Cycles - 1] ^= MemoryBankAccessCycle;
+    }
+  }
+
+  int Cycle = DeltaCycles;
+  Scoreboard[Cycle].IssueCount--;
+  for (const InstrStage &IS : ItinData->getStages(SchedClass)) {
+    if (FUDepthLimit && (Cycle - DeltaCycles) >= *FUDepthLimit) {
+      break;
+    }
+    const FuncUnitWrapper ResourceToRelease(IS);
+    for (unsigned int C = 0; C < IS.getCycles(); ++C) {
+      Scoreboard[Cycle + C] ^= ResourceToRelease;
+    }
+
+    // Advance the cycle to the next stage.
+    Cycle += IS.getNextCycles();
+  }
+
+  LLVM_DEBUG({
+    dbgs() << "Scoreboard after release resources:\n";
+    Scoreboard.dump();
+  });
+}
+
 unsigned AIEHazardRecognizer::getPipelineDepth() const { return PipelineDepth; }
 
 unsigned AIEHazardRecognizer::getMaxLatency() const { return MaxLatency; }

Diff for: llvm/lib/Target/AIE/AIEHazardRecognizer.h

@@ -107,6 +107,7 @@ class FuncUnitWrapper {
   void dump() const;
 
   FuncUnitWrapper &operator|=(const FuncUnitWrapper &Other);
+  FuncUnitWrapper &operator^=(const FuncUnitWrapper &Other);
   bool conflict(const FuncUnitWrapper &Other) const;
 };
 
@@ -178,6 +179,13 @@ class AIEHazardRecognizer : public ScheduleHazardRecognizer {
                         iterator_range<const MachineOperand *> MIOperands,
                         const MachineRegisterInfo &MRI, int DeltaCycles);
 
+  void releaseFromScoreboard(ResourceScoreboard<FuncUnitWrapper> &Scoreboard,
+                             const MCInstrDesc &Desc,
+                             MemoryBankBits MemoryBanks,
+                             iterator_range<const MachineOperand *> MIOperands,
+                             const MachineRegisterInfo &MRI,
+                             int DeltaCycles) const;
+
   /// Block all scoreboard resources at DeltaCycles
   void blockCycleInScoreboard(int DeltaCycle);
 
@@ -254,6 +262,14 @@ class AIEHazardRecognizer : public ScheduleHazardRecognizer {
                              SmallVector<int, 2> MemoryAccessCycles,
                              int DeltaCycles, std::optional<int> FUDepthLimit);
 
+  static void releaseResources(ResourceScoreboard<FuncUnitWrapper> &Scoreboard,
+                               const InstrItineraryData *ItinData,
+                               unsigned SchedClass, SlotBits SlotSet,
+                               MemoryBankBits MemoryBanks,
+                               SmallVector<int, 2> MemoryAccessCycles,
+                               int DeltaCycles,
+                               std::optional<int> FUDepthLimit);
+
 private:
   ResourceScoreboard<FuncUnitWrapper> Scoreboard;
   const AIEBaseInstrInfo *TII;