Use instruction itinerary to determine what instructions are 'cheap'.

llvm-svn: 117348

include/llvm/Target/TargetInstrInfo.h

@@ -639,6 +639,12 @@ public:
                              const MachineInstr *UseMI, unsigned UseIdx) const {
     return false;
   }
+
+  /// hasLowDefLatency - Compute operand latency of a def of 'Reg', return true
+  /// if the target considered it 'low'.
+  virtual
+  bool hasLowDefLatency(const InstrItineraryData *ItinData,
+                        const MachineInstr *DefMI, unsigned DefIdx) const;
 };
 
 /// TargetInstrInfoImpl - This is the default implementation of

lib/CodeGen/MachineLICM.cpp

@@ -173,7 +173,10 @@ namespace {
     /// HasHighOperandLatency - Compute operand latency between a def of 'Reg'
     /// and an use in the current loop, return true if the target considered
     /// it 'high'.
-    bool HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx, unsigned Reg);
+    bool HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx,
+                               unsigned Reg) const;
+
+    bool IsCheapInstruction(MachineInstr &MI) const;
 
     /// CanCauseHighRegPressure - Visit BBs from header to current BB,
     /// check if hoisting an instruction of the given cost matrix can cause high
@@ -795,13 +798,15 @@ bool MachineLICM::isLoadFromConstantMemory(MachineInstr *MI) {
 /// and an use in the current loop, return true if the target considered
 /// it 'high'.
 bool MachineLICM::HasHighOperandLatency(MachineInstr &MI,
-                                        unsigned DefIdx, unsigned Reg) {
-  if (MRI->use_nodbg_empty(Reg))
+                                        unsigned DefIdx, unsigned Reg) const {
+  if (!InstrItins || InstrItins->isEmpty() || MRI->use_nodbg_empty(Reg))
     return false;
 
   for (MachineRegisterInfo::use_nodbg_iterator I = MRI->use_nodbg_begin(Reg),
          E = MRI->use_nodbg_end(); I != E; ++I) {
     MachineInstr *UseMI = &*I;
+    if (UseMI->isCopyLike())
+      continue;
     if (!CurLoop->contains(UseMI->getParent()))
       continue;
     for (unsigned i = 0, e = UseMI->getNumOperands(); i != e; ++i) {
@@ -823,6 +828,33 @@ bool MachineLICM::HasHighOperandLatency(MachineInstr &MI,
   return false;
 }
 
+/// IsCheapInstruction - Return true if the instruction is marked "cheap" or
+/// the operand latency between its def and a use is one or less.
+bool MachineLICM::IsCheapInstruction(MachineInstr &MI) const {
+  if (MI.getDesc().isAsCheapAsAMove() || MI.isCopyLike())
+    return true;
+  if (!InstrItins || InstrItins->isEmpty())
+    return false;
+
+  bool isCheap = false;
+  unsigned NumDefs = MI.getDesc().getNumDefs();
+  for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) {
+    MachineOperand &DefMO = MI.getOperand(i);
+    if (!DefMO.isReg() || !DefMO.isDef())
+      continue;
+    --NumDefs;
+    unsigned Reg = DefMO.getReg();
+    if (TargetRegisterInfo::isPhysicalRegister(Reg))
+      continue;
+
+    if (!TII->hasLowDefLatency(InstrItins, &MI, i))
+      return false;
+    isCheap = true;
+  }
+
+  return isCheap;
+}
+
 /// CanCauseHighRegPressure - Visit BBs from header to current BB, check
 /// if hoisting an instruction of the given cost matrix can cause high
 /// register pressure.
@@ -905,7 +937,7 @@ bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) {
   // trade off is it may cause spill in high pressure situation. It will end up
   // adding a store in the loop preheader. But the reload is no more expensive.
   // The side benefit is these loads are frequently CSE'ed.
-  if (MI.getDesc().isAsCheapAsAMove()) {
+  if (IsCheapInstruction(MI)) {
     if (!TII->isTriviallyReMaterializable(&MI, AA))
       return false;
   } else {

lib/Target/ARM/ARMBaseInstrInfo.cpp

@@ -1951,3 +1951,18 @@ hasHighOperandLatency(const InstrItineraryData *ItinData,
   return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON ||
          UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON;
 }
+
+bool ARMBaseInstrInfo::
+hasLowDefLatency(const InstrItineraryData *ItinData,
+                 const MachineInstr *DefMI, unsigned DefIdx) const {
+  if (!ItinData || ItinData->isEmpty())
+    return false;
+
+  unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask;
+  if (DDomain == ARMII::DomainGeneral) {
+    unsigned DefClass = DefMI->getDesc().getSchedClass();
+    int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
+    return (DefCycle != -1 && DefCycle <= 2);
+  }
+  return false;
+}

lib/Target/ARM/ARMBaseInstrInfo.h

@@ -382,6 +382,8 @@ private:
                              const MachineRegisterInfo *MRI,
                              const MachineInstr *DefMI, unsigned DefIdx,
                              const MachineInstr *UseMI, unsigned UseIdx) const;
+  bool hasLowDefLatency(const InstrItineraryData *ItinData,
+                        const MachineInstr *DefMI, unsigned DefIdx) const;
 };
 
 static inline

lib/Target/TargetInstrInfo.cpp

@@ -94,6 +94,16 @@ TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
   return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
 }
 
+bool TargetInstrInfo::hasLowDefLatency(const InstrItineraryData *ItinData,
+                                       const MachineInstr *DefMI,
+                                       unsigned DefIdx) const {
+  if (!ItinData || ItinData->isEmpty())
+    return false;
+
+  unsigned DefClass = DefMI->getDesc().getSchedClass();
+  int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
+  return (DefCycle != -1 && DefCycle <= 1);
+}
 
 /// insertNoop - Insert a noop into the instruction stream at the specified
 /// point.