+ // Since there are no intervening uses for both registers, then commute
+ // if the def of regC is closer. Its live interval is shorter.
+ return LastDefB && LastDefC && LastDefC > LastDefB;
+}
+
+/// commuteInstruction - Commute a two-address instruction and update the basic
+/// block, distance map, and live variables if needed. Return true if it is
+/// successful.
+bool TwoAddressInstructionPass::commuteInstruction(MachineInstr *MI,
+ unsigned RegBIdx,
+ unsigned RegCIdx,
+ unsigned Dist) {
+ unsigned RegC = MI->getOperand(RegCIdx).getReg();
+ DEBUG(dbgs() << "2addr: COMMUTING : " << *MI);
+ MachineInstr *NewMI = TII->commuteInstruction(MI, false, RegBIdx, RegCIdx);
+
+ if (NewMI == nullptr) {
+ DEBUG(dbgs() << "2addr: COMMUTING FAILED!\n");
+ return false;
+ }
+
+ DEBUG(dbgs() << "2addr: COMMUTED TO: " << *NewMI);
+ assert(NewMI == MI &&
+ "TargetInstrInfo::commuteInstruction() should not return a new "
+ "instruction unless it was requested.");
+
+ // Update source register map.
+ unsigned FromRegC = getMappedReg(RegC, SrcRegMap);
+ if (FromRegC) {
+ unsigned RegA = MI->getOperand(0).getReg();
+ SrcRegMap[RegA] = FromRegC;
+ }
+
+ return true;
+}
+
+/// isProfitableToConv3Addr - Return true if it is profitable to convert the
+/// given 2-address instruction to a 3-address one.
+bool
+TwoAddressInstructionPass::isProfitableToConv3Addr(unsigned RegA,unsigned RegB){
+ // Look for situations like this:
+ // %reg1024<def> = MOV r1
+ // %reg1025<def> = MOV r0
+ // %reg1026<def> = ADD %reg1024, %reg1025
+ // r2 = MOV %reg1026
+ // Turn ADD into a 3-address instruction to avoid a copy.
+ unsigned FromRegB = getMappedReg(RegB, SrcRegMap);
+ if (!FromRegB)
+ return false;
+ unsigned ToRegA = getMappedReg(RegA, DstRegMap);
+ return (ToRegA && !regsAreCompatible(FromRegB, ToRegA, TRI));
+}
+
+/// convertInstTo3Addr - Convert the specified two-address instruction into a
+/// three address one. Return true if this transformation was successful.
+bool
+TwoAddressInstructionPass::convertInstTo3Addr(MachineBasicBlock::iterator &mi,
+ MachineBasicBlock::iterator &nmi,
+ unsigned RegA, unsigned RegB,
+ unsigned Dist) {
+ // FIXME: Why does convertToThreeAddress() need an iterator reference?
+ MachineFunction::iterator MFI = MBB->getIterator();
+ MachineInstr *NewMI = TII->convertToThreeAddress(MFI, mi, LV);
+ assert(MBB->getIterator() == MFI &&
+ "convertToThreeAddress changed iterator reference");
+ if (!NewMI)
+ return false;
+
+ DEBUG(dbgs() << "2addr: CONVERTING 2-ADDR: " << *mi);
+ DEBUG(dbgs() << "2addr: TO 3-ADDR: " << *NewMI);
+ bool Sunk = false;
+
+ if (LIS)
+ LIS->ReplaceMachineInstrInMaps(mi, NewMI);
+
+ if (NewMI->findRegisterUseOperand(RegB, false, TRI))
+ // FIXME: Temporary workaround. If the new instruction doesn't
+ // uses RegB, convertToThreeAddress must have created more
+ // then one instruction.
+ Sunk = sink3AddrInstruction(NewMI, RegB, mi);
+
+ MBB->erase(mi); // Nuke the old inst.
+
+ if (!Sunk) {
+ DistanceMap.insert(std::make_pair(NewMI, Dist));
+ mi = NewMI;
+ nmi = std::next(mi);
+ }
+
+ // Update source and destination register maps.
+ SrcRegMap.erase(RegA);
+ DstRegMap.erase(RegB);
+ return true;
+}
+
+/// scanUses - Scan forward recursively for only uses, update maps if the use
+/// is a copy or a two-address instruction.
+void
+TwoAddressInstructionPass::scanUses(unsigned DstReg) {
+ SmallVector<unsigned, 4> VirtRegPairs;
+ bool IsDstPhys;
+ bool IsCopy = false;
+ unsigned NewReg = 0;
+ unsigned Reg = DstReg;
+ while (MachineInstr *UseMI = findOnlyInterestingUse(Reg, MBB, MRI, TII,IsCopy,
+ NewReg, IsDstPhys)) {
+ if (IsCopy && !Processed.insert(UseMI).second)
+ break;
+
+ DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(UseMI);
+ if (DI != DistanceMap.end())
+ // Earlier in the same MBB.Reached via a back edge.
+ break;
+
+ if (IsDstPhys) {
+ VirtRegPairs.push_back(NewReg);
+ break;
+ }
+ bool isNew = SrcRegMap.insert(std::make_pair(NewReg, Reg)).second;
+ if (!isNew)
+ assert(SrcRegMap[NewReg] == Reg && "Can't map to two src registers!");
+ VirtRegPairs.push_back(NewReg);
+ Reg = NewReg;
+ }
+
+ if (!VirtRegPairs.empty()) {
+ unsigned ToReg = VirtRegPairs.back();
+ VirtRegPairs.pop_back();
+ while (!VirtRegPairs.empty()) {
+ unsigned FromReg = VirtRegPairs.back();
+ VirtRegPairs.pop_back();
+ bool isNew = DstRegMap.insert(std::make_pair(FromReg, ToReg)).second;
+ if (!isNew)
+ assert(DstRegMap[FromReg] == ToReg &&"Can't map to two dst registers!");
+ ToReg = FromReg;
+ }
+ bool isNew = DstRegMap.insert(std::make_pair(DstReg, ToReg)).second;
+ if (!isNew)
+ assert(DstRegMap[DstReg] == ToReg && "Can't map to two dst registers!");
+ }
+}
+
+/// processCopy - If the specified instruction is not yet processed, process it
+/// if it's a copy. For a copy instruction, we find the physical registers the
+/// source and destination registers might be mapped to. These are kept in
+/// point-to maps used to determine future optimizations. e.g.
+/// v1024 = mov r0
+/// v1025 = mov r1
+/// v1026 = add v1024, v1025
+/// r1 = mov r1026
+/// If 'add' is a two-address instruction, v1024, v1026 are both potentially
+/// coalesced to r0 (from the input side). v1025 is mapped to r1. v1026 is
+/// potentially joined with r1 on the output side. It's worthwhile to commute
+/// 'add' to eliminate a copy.
+void TwoAddressInstructionPass::processCopy(MachineInstr *MI) {
+ if (Processed.count(MI))
+ return;
+
+ bool IsSrcPhys, IsDstPhys;
+ unsigned SrcReg, DstReg;
+ if (!isCopyToReg(*MI, TII, SrcReg, DstReg, IsSrcPhys, IsDstPhys))
+ return;
+
+ if (IsDstPhys && !IsSrcPhys)
+ DstRegMap.insert(std::make_pair(SrcReg, DstReg));
+ else if (!IsDstPhys && IsSrcPhys) {
+ bool isNew = SrcRegMap.insert(std::make_pair(DstReg, SrcReg)).second;
+ if (!isNew)
+ assert(SrcRegMap[DstReg] == SrcReg &&
+ "Can't map to two src physical registers!");
+
+ scanUses(DstReg);
+ }
+
+ Processed.insert(MI);
+ return;
+}
+
+/// rescheduleMIBelowKill - If there is one more local instruction that reads
+/// 'Reg' and it kills 'Reg, consider moving the instruction below the kill
+/// instruction in order to eliminate the need for the copy.
+bool TwoAddressInstructionPass::
+rescheduleMIBelowKill(MachineBasicBlock::iterator &mi,
+ MachineBasicBlock::iterator &nmi,
+ unsigned Reg) {
+ // Bail immediately if we don't have LV or LIS available. We use them to find
+ // kills efficiently.
+ if (!LV && !LIS)
+ return false;
+
+ MachineInstr *MI = &*mi;
+ DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
+ if (DI == DistanceMap.end())
+ // Must be created from unfolded load. Don't waste time trying this.
+ return false;
+
+ MachineInstr *KillMI = nullptr;
+ if (LIS) {
+ LiveInterval &LI = LIS->getInterval(Reg);
+ assert(LI.end() != LI.begin() &&
+ "Reg should not have empty live interval.");
+
+ SlotIndex MBBEndIdx = LIS->getMBBEndIdx(MBB).getPrevSlot();
+ LiveInterval::const_iterator I = LI.find(MBBEndIdx);
+ if (I != LI.end() && I->start < MBBEndIdx)
+ return false;
+
+ --I;
+ KillMI = LIS->getInstructionFromIndex(I->end);
+ } else {
+ KillMI = LV->getVarInfo(Reg).findKill(MBB);
+ }
+ if (!KillMI || MI == KillMI || KillMI->isCopy() || KillMI->isCopyLike())
+ // Don't mess with copies, they may be coalesced later.
+ return false;
+
+ if (KillMI->hasUnmodeledSideEffects() || KillMI->isCall() ||
+ KillMI->isBranch() || KillMI->isTerminator())
+ // Don't move pass calls, etc.
+ return false;
+
+ unsigned DstReg;
+ if (isTwoAddrUse(*KillMI, Reg, DstReg))
+ return false;
+
+ bool SeenStore = true;
+ if (!MI->isSafeToMove(AA, SeenStore))
+ return false;
+
+ if (TII->getInstrLatency(InstrItins, MI) > 1)
+ // FIXME: Needs more sophisticated heuristics.
+ return false;
+
+ SmallSet<unsigned, 2> Uses;
+ SmallSet<unsigned, 2> Kills;
+ SmallSet<unsigned, 2> Defs;
+ for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = MI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned MOReg = MO.getReg();
+ if (!MOReg)
+ continue;
+ if (MO.isDef())
+ Defs.insert(MOReg);
+ else {
+ Uses.insert(MOReg);
+ if (MOReg != Reg && (MO.isKill() ||
+ (LIS && isPlainlyKilled(MI, MOReg, LIS))))
+ Kills.insert(MOReg);
+ }
+ }
+
+ // Move the copies connected to MI down as well.
+ MachineBasicBlock::iterator Begin = MI;
+ MachineBasicBlock::iterator AfterMI = std::next(Begin);
+
+ MachineBasicBlock::iterator End = AfterMI;
+ while (End->isCopy() && Defs.count(End->getOperand(1).getReg())) {
+ Defs.insert(End->getOperand(0).getReg());
+ ++End;
+ }
+
+ // Check if the reschedule will not break depedencies.
+ unsigned NumVisited = 0;
+ MachineBasicBlock::iterator KillPos = KillMI;
+ ++KillPos;
+ for (MachineBasicBlock::iterator I = End; I != KillPos; ++I) {
+ MachineInstr *OtherMI = I;
+ // DBG_VALUE cannot be counted against the limit.
+ if (OtherMI->isDebugValue())
+ continue;
+ if (NumVisited > 10) // FIXME: Arbitrary limit to reduce compile time cost.
+ return false;
+ ++NumVisited;
+ if (OtherMI->hasUnmodeledSideEffects() || OtherMI->isCall() ||
+ OtherMI->isBranch() || OtherMI->isTerminator())
+ // Don't move pass calls, etc.
+ return false;
+ for (unsigned i = 0, e = OtherMI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = OtherMI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned MOReg = MO.getReg();
+ if (!MOReg)
+ continue;
+ if (MO.isDef()) {
+ if (Uses.count(MOReg))
+ // Physical register use would be clobbered.
+ return false;
+ if (!MO.isDead() && Defs.count(MOReg))
+ // May clobber a physical register def.
+ // FIXME: This may be too conservative. It's ok if the instruction
+ // is sunken completely below the use.
+ return false;
+ } else {
+ if (Defs.count(MOReg))
+ return false;
+ bool isKill = MO.isKill() ||
+ (LIS && isPlainlyKilled(OtherMI, MOReg, LIS));
+ if (MOReg != Reg &&
+ ((isKill && Uses.count(MOReg)) || Kills.count(MOReg)))
+ // Don't want to extend other live ranges and update kills.
+ return false;
+ if (MOReg == Reg && !isKill)
+ // We can't schedule across a use of the register in question.
+ return false;
+ // Ensure that if this is register in question, its the kill we expect.
+ assert((MOReg != Reg || OtherMI == KillMI) &&
+ "Found multiple kills of a register in a basic block");
+ }
+ }
+ }
+
+ // Move debug info as well.
+ while (Begin != MBB->begin() && std::prev(Begin)->isDebugValue())
+ --Begin;
+
+ nmi = End;
+ MachineBasicBlock::iterator InsertPos = KillPos;
+ if (LIS) {
+ // We have to move the copies first so that the MBB is still well-formed
+ // when calling handleMove().
+ for (MachineBasicBlock::iterator MBBI = AfterMI; MBBI != End;) {
+ MachineInstr *CopyMI = MBBI;
+ ++MBBI;
+ MBB->splice(InsertPos, MBB, CopyMI);
+ LIS->handleMove(CopyMI);
+ InsertPos = CopyMI;
+ }
+ End = std::next(MachineBasicBlock::iterator(MI));
+ }
+
+ // Copies following MI may have been moved as well.
+ MBB->splice(InsertPos, MBB, Begin, End);
+ DistanceMap.erase(DI);
+
+ // Update live variables
+ if (LIS) {
+ LIS->handleMove(MI);
+ } else {
+ LV->removeVirtualRegisterKilled(Reg, KillMI);
+ LV->addVirtualRegisterKilled(Reg, MI);
+ }
+
+ DEBUG(dbgs() << "\trescheduled below kill: " << *KillMI);
+ return true;
+}
+
+/// isDefTooClose - Return true if the re-scheduling will put the given
+/// instruction too close to the defs of its register dependencies.
+bool TwoAddressInstructionPass::isDefTooClose(unsigned Reg, unsigned Dist,
+ MachineInstr *MI) {
+ for (MachineInstr &DefMI : MRI->def_instructions(Reg)) {
+ if (DefMI.getParent() != MBB || DefMI.isCopy() || DefMI.isCopyLike())
+ continue;
+ if (&DefMI == MI)
+ return true; // MI is defining something KillMI uses
+ DenseMap<MachineInstr*, unsigned>::iterator DDI = DistanceMap.find(&DefMI);
+ if (DDI == DistanceMap.end())
+ return true; // Below MI
+ unsigned DefDist = DDI->second;
+ assert(Dist > DefDist && "Visited def already?");
+ if (TII->getInstrLatency(InstrItins, &DefMI) > (Dist - DefDist))
+ return true;
+ }
+ return false;
+}
+
+/// rescheduleKillAboveMI - If there is one more local instruction that reads
+/// 'Reg' and it kills 'Reg, consider moving the kill instruction above the
+/// current two-address instruction in order to eliminate the need for the
+/// copy.
+bool TwoAddressInstructionPass::
+rescheduleKillAboveMI(MachineBasicBlock::iterator &mi,
+ MachineBasicBlock::iterator &nmi,
+ unsigned Reg) {
+ // Bail immediately if we don't have LV or LIS available. We use them to find
+ // kills efficiently.
+ if (!LV && !LIS)
+ return false;
+
+ MachineInstr *MI = &*mi;
+ DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
+ if (DI == DistanceMap.end())
+ // Must be created from unfolded load. Don't waste time trying this.
+ return false;
+
+ MachineInstr *KillMI = nullptr;
+ if (LIS) {
+ LiveInterval &LI = LIS->getInterval(Reg);
+ assert(LI.end() != LI.begin() &&
+ "Reg should not have empty live interval.");
+
+ SlotIndex MBBEndIdx = LIS->getMBBEndIdx(MBB).getPrevSlot();
+ LiveInterval::const_iterator I = LI.find(MBBEndIdx);
+ if (I != LI.end() && I->start < MBBEndIdx)
+ return false;
+
+ --I;
+ KillMI = LIS->getInstructionFromIndex(I->end);
+ } else {
+ KillMI = LV->getVarInfo(Reg).findKill(MBB);
+ }
+ if (!KillMI || MI == KillMI || KillMI->isCopy() || KillMI->isCopyLike())
+ // Don't mess with copies, they may be coalesced later.
+ return false;
+
+ unsigned DstReg;
+ if (isTwoAddrUse(*KillMI, Reg, DstReg))
+ return false;
+
+ bool SeenStore = true;
+ if (!KillMI->isSafeToMove(AA, SeenStore))
+ return false;
+
+ SmallSet<unsigned, 2> Uses;
+ SmallSet<unsigned, 2> Kills;
+ SmallSet<unsigned, 2> Defs;
+ SmallSet<unsigned, 2> LiveDefs;
+ for (unsigned i = 0, e = KillMI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = KillMI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned MOReg = MO.getReg();
+ if (MO.isUse()) {
+ if (!MOReg)
+ continue;
+ if (isDefTooClose(MOReg, DI->second, MI))
+ return false;
+ bool isKill = MO.isKill() || (LIS && isPlainlyKilled(KillMI, MOReg, LIS));
+ if (MOReg == Reg && !isKill)
+ return false;
+ Uses.insert(MOReg);
+ if (isKill && MOReg != Reg)
+ Kills.insert(MOReg);
+ } else if (TargetRegisterInfo::isPhysicalRegister(MOReg)) {
+ Defs.insert(MOReg);
+ if (!MO.isDead())
+ LiveDefs.insert(MOReg);
+ }
+ }
+
+ // Check if the reschedule will not break depedencies.
+ unsigned NumVisited = 0;
+ MachineBasicBlock::iterator KillPos = KillMI;
+ for (MachineBasicBlock::iterator I = mi; I != KillPos; ++I) {
+ MachineInstr *OtherMI = I;
+ // DBG_VALUE cannot be counted against the limit.
+ if (OtherMI->isDebugValue())
+ continue;
+ if (NumVisited > 10) // FIXME: Arbitrary limit to reduce compile time cost.
+ return false;
+ ++NumVisited;
+ if (OtherMI->hasUnmodeledSideEffects() || OtherMI->isCall() ||
+ OtherMI->isBranch() || OtherMI->isTerminator())
+ // Don't move pass calls, etc.
+ return false;
+ SmallVector<unsigned, 2> OtherDefs;
+ for (unsigned i = 0, e = OtherMI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = OtherMI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned MOReg = MO.getReg();
+ if (!MOReg)
+ continue;
+ if (MO.isUse()) {
+ if (Defs.count(MOReg))
+ // Moving KillMI can clobber the physical register if the def has
+ // not been seen.
+ return false;
+ if (Kills.count(MOReg))
+ // Don't want to extend other live ranges and update kills.
+ return false;
+ if (OtherMI != MI && MOReg == Reg &&
+ !(MO.isKill() || (LIS && isPlainlyKilled(OtherMI, MOReg, LIS))))
+ // We can't schedule across a use of the register in question.
+ return false;
+ } else {
+ OtherDefs.push_back(MOReg);
+ }
+ }
+
+ for (unsigned i = 0, e = OtherDefs.size(); i != e; ++i) {
+ unsigned MOReg = OtherDefs[i];
+ if (Uses.count(MOReg))
+ return false;
+ if (TargetRegisterInfo::isPhysicalRegister(MOReg) &&
+ LiveDefs.count(MOReg))
+ return false;
+ // Physical register def is seen.
+ Defs.erase(MOReg);
+ }
+ }
+
+ // Move the old kill above MI, don't forget to move debug info as well.
+ MachineBasicBlock::iterator InsertPos = mi;
+ while (InsertPos != MBB->begin() && std::prev(InsertPos)->isDebugValue())
+ --InsertPos;
+ MachineBasicBlock::iterator From = KillMI;
+ MachineBasicBlock::iterator To = std::next(From);
+ while (std::prev(From)->isDebugValue())
+ --From;
+ MBB->splice(InsertPos, MBB, From, To);
+
+ nmi = std::prev(InsertPos); // Backtrack so we process the moved instr.
+ DistanceMap.erase(DI);
+
+ // Update live variables
+ if (LIS) {
+ LIS->handleMove(KillMI);
+ } else {
+ LV->removeVirtualRegisterKilled(Reg, KillMI);
+ LV->addVirtualRegisterKilled(Reg, MI);
+ }
+
+ DEBUG(dbgs() << "\trescheduled kill: " << *KillMI);
+ return true;
+}
+
+/// Tries to commute the operand 'BaseOpIdx' and some other operand in the
+/// given machine instruction to improve opportunities for coalescing and
+/// elimination of a register to register copy.
+///
+/// 'DstOpIdx' specifies the index of MI def operand.
+/// 'BaseOpKilled' specifies if the register associated with 'BaseOpIdx'
+/// operand is killed by the given instruction.
+/// The 'Dist' arguments provides the distance of MI from the start of the
+/// current basic block and it is used to determine if it is profitable
+/// to commute operands in the instruction.
+///
+/// Returns true if the transformation happened. Otherwise, returns false.
+bool TwoAddressInstructionPass::tryInstructionCommute(MachineInstr *MI,
+ unsigned DstOpIdx,
+ unsigned BaseOpIdx,
+ bool BaseOpKilled,
+ unsigned Dist) {
+ unsigned DstOpReg = MI->getOperand(DstOpIdx).getReg();
+ unsigned BaseOpReg = MI->getOperand(BaseOpIdx).getReg();
+ unsigned OpsNum = MI->getDesc().getNumOperands();
+ unsigned OtherOpIdx = MI->getDesc().getNumDefs();
+ for (; OtherOpIdx < OpsNum; OtherOpIdx++) {
+ // The call of findCommutedOpIndices below only checks if BaseOpIdx
+ // and OtherOpIdx are commutable, it does not really searches for
+ // other commutable operands and does not change the values of passed
+ // variables.
+ if (OtherOpIdx == BaseOpIdx ||
+ !TII->findCommutedOpIndices(MI, BaseOpIdx, OtherOpIdx))
+ continue;
+
+ unsigned OtherOpReg = MI->getOperand(OtherOpIdx).getReg();
+ bool AggressiveCommute = false;
+
+ // If OtherOp dies but BaseOp does not, swap the OtherOp and BaseOp
+ // operands. This makes the live ranges of DstOp and OtherOp joinable.
+ bool DoCommute =
+ !BaseOpKilled && isKilled(*MI, OtherOpReg, MRI, TII, LIS, false);
+
+ if (!DoCommute &&
+ isProfitableToCommute(DstOpReg, BaseOpReg, OtherOpReg, MI, Dist)) {
+ DoCommute = true;
+ AggressiveCommute = true;
+ }
+
+ // If it's profitable to commute, try to do so.
+ if (DoCommute && commuteInstruction(MI, BaseOpIdx, OtherOpIdx, Dist)) {
+ ++NumCommuted;
+ if (AggressiveCommute)
+ ++NumAggrCommuted;
+ return true;
+ }
+ }
+ return false;
+}
+
+/// tryInstructionTransform - For the case where an instruction has a single
+/// pair of tied register operands, attempt some transformations that may
+/// either eliminate the tied operands or improve the opportunities for
+/// coalescing away the register copy. Returns true if no copy needs to be
+/// inserted to untie mi's operands (either because they were untied, or
+/// because mi was rescheduled, and will be visited again later). If the
+/// shouldOnlyCommute flag is true, only instruction commutation is attempted.
+bool TwoAddressInstructionPass::
+tryInstructionTransform(MachineBasicBlock::iterator &mi,
+ MachineBasicBlock::iterator &nmi,
+ unsigned SrcIdx, unsigned DstIdx,
+ unsigned Dist, bool shouldOnlyCommute) {
+ if (OptLevel == CodeGenOpt::None)
+ return false;
+
+ MachineInstr &MI = *mi;
+ unsigned regA = MI.getOperand(DstIdx).getReg();
+ unsigned regB = MI.getOperand(SrcIdx).getReg();
+
+ assert(TargetRegisterInfo::isVirtualRegister(regB) &&
+ "cannot make instruction into two-address form");
+ bool regBKilled = isKilled(MI, regB, MRI, TII, LIS, true);
+
+ if (TargetRegisterInfo::isVirtualRegister(regA))
+ scanUses(regA);
+
+ bool Commuted = tryInstructionCommute(&MI, DstIdx, SrcIdx, regBKilled, Dist);
+
+ // If the instruction is convertible to 3 Addr, instead
+ // of returning try 3 Addr transformation aggresively and
+ // use this variable to check later. Because it might be better.
+ // For example, we can just use `leal (%rsi,%rdi), %eax` and `ret`
+ // instead of the following code.
+ // addl %esi, %edi
+ // movl %edi, %eax
+ // ret
+ if (Commuted && !MI.isConvertibleTo3Addr())
+ return false;
+
+ if (shouldOnlyCommute)
+ return false;
+
+ // If there is one more use of regB later in the same MBB, consider
+ // re-schedule this MI below it.
+ if (!Commuted && EnableRescheduling && rescheduleMIBelowKill(mi, nmi, regB)) {
+ ++NumReSchedDowns;
+ return true;
+ }
+
+ // If we commuted, regB may have changed so we should re-sample it to avoid
+ // confusing the three address conversion below.
+ if (Commuted) {
+ regB = MI.getOperand(SrcIdx).getReg();
+ regBKilled = isKilled(MI, regB, MRI, TII, LIS, true);
+ }
+
+ if (MI.isConvertibleTo3Addr()) {
+ // This instruction is potentially convertible to a true
+ // three-address instruction. Check if it is profitable.
+ if (!regBKilled || isProfitableToConv3Addr(regA, regB)) {
+ // Try to convert it.
+ if (convertInstTo3Addr(mi, nmi, regA, regB, Dist)) {
+ ++NumConvertedTo3Addr;
+ return true; // Done with this instruction.
+ }
+ }
+ }
+
+ // Return if it is commuted but 3 addr conversion is failed.
+ if (Commuted)
+ return false;
+
+ // If there is one more use of regB later in the same MBB, consider
+ // re-schedule it before this MI if it's legal.
+ if (EnableRescheduling && rescheduleKillAboveMI(mi, nmi, regB)) {
+ ++NumReSchedUps;
+ return true;
+ }
+
+ // If this is an instruction with a load folded into it, try unfolding
+ // the load, e.g. avoid this:
+ // movq %rdx, %rcx
+ // addq (%rax), %rcx
+ // in favor of this:
+ // movq (%rax), %rcx
+ // addq %rdx, %rcx
+ // because it's preferable to schedule a load than a register copy.
+ if (MI.mayLoad() && !regBKilled) {
+ // Determine if a load can be unfolded.
+ unsigned LoadRegIndex;
+ unsigned NewOpc =
+ TII->getOpcodeAfterMemoryUnfold(MI.getOpcode(),
+ /*UnfoldLoad=*/true,
+ /*UnfoldStore=*/false,
+ &LoadRegIndex);
+ if (NewOpc != 0) {
+ const MCInstrDesc &UnfoldMCID = TII->get(NewOpc);
+ if (UnfoldMCID.getNumDefs() == 1) {
+ // Unfold the load.
+ DEBUG(dbgs() << "2addr: UNFOLDING: " << MI);
+ const TargetRegisterClass *RC =
+ TRI->getAllocatableClass(
+ TII->getRegClass(UnfoldMCID, LoadRegIndex, TRI, *MF));
+ unsigned Reg = MRI->createVirtualRegister(RC);
+ SmallVector<MachineInstr *, 2> NewMIs;
+ if (!TII->unfoldMemoryOperand(*MF, &MI, Reg,
+ /*UnfoldLoad=*/true,/*UnfoldStore=*/false,
+ NewMIs)) {
+ DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n");
+ return false;