+/// isTwoAddrUse - Return true if the specified MI uses the specified register
+/// as a two-address use. If so, return the destination register by reference.
+static bool isTwoAddrUse(MachineInstr &MI, unsigned Reg, unsigned &DstReg) {
+ for (unsigned i = 0, NumOps = MI.getNumOperands(); i != NumOps; ++i) {
+ const MachineOperand &MO = MI.getOperand(i);
+ if (!MO.isReg() || !MO.isUse() || MO.getReg() != Reg)
+ continue;
+ unsigned ti;
+ if (MI.isRegTiedToDefOperand(i, &ti)) {
+ DstReg = MI.getOperand(ti).getReg();
+ return true;
+ }
+ }
+ return false;
+}
+
+/// findOnlyInterestingUse - Given a register, if has a single in-basic block
+/// use, return the use instruction if it's a copy or a two-address use.
+static
+MachineInstr *findOnlyInterestingUse(unsigned Reg, MachineBasicBlock *MBB,
+ MachineRegisterInfo *MRI,
+ const TargetInstrInfo *TII,
+ bool &IsCopy,
+ unsigned &DstReg, bool &IsDstPhys) {
+ if (!MRI->hasOneNonDBGUse(Reg))
+ // None or more than one use.
+ return nullptr;
+ MachineInstr &UseMI = *MRI->use_instr_nodbg_begin(Reg);
+ if (UseMI.getParent() != MBB)
+ return nullptr;
+ unsigned SrcReg;
+ bool IsSrcPhys;
+ if (isCopyToReg(UseMI, TII, SrcReg, DstReg, IsSrcPhys, IsDstPhys)) {
+ IsCopy = true;
+ return &UseMI;
+ }
+ IsDstPhys = false;
+ if (isTwoAddrUse(UseMI, Reg, DstReg)) {
+ IsDstPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
+ return &UseMI;
+ }
+ return nullptr;
+}
+
+/// getMappedReg - Return the physical register the specified virtual register
+/// might be mapped to.
+static unsigned
+getMappedReg(unsigned Reg, DenseMap<unsigned, unsigned> &RegMap) {
+ while (TargetRegisterInfo::isVirtualRegister(Reg)) {
+ DenseMap<unsigned, unsigned>::iterator SI = RegMap.find(Reg);
+ if (SI == RegMap.end())
+ return 0;
+ Reg = SI->second;
+ }
+ if (TargetRegisterInfo::isPhysicalRegister(Reg))
+ return Reg;
+ return 0;
+}
+
+/// regsAreCompatible - Return true if the two registers are equal or aliased.
+///
+static bool
+regsAreCompatible(unsigned RegA, unsigned RegB, const TargetRegisterInfo *TRI) {
+ if (RegA == RegB)
+ return true;
+ if (!RegA || !RegB)
+ return false;
+ return TRI->regsOverlap(RegA, RegB);
+}
+
+
+/// isProfitableToCommute - Return true if it's potentially profitable to commute
+/// the two-address instruction that's being processed.
+bool
+TwoAddressInstructionPass::
+isProfitableToCommute(unsigned regA, unsigned regB, unsigned regC,
+ MachineInstr *MI, unsigned Dist) {
+ if (OptLevel == CodeGenOpt::None)
+ return false;
+
+ // Determine if it's profitable to commute this two address instruction. In
+ // general, we want no uses between this instruction and the definition of
+ // the two-address register.
+ // e.g.
+ // %reg1028<def> = EXTRACT_SUBREG %reg1027<kill>, 1
+ // %reg1029<def> = MOV8rr %reg1028
+ // %reg1029<def> = SHR8ri %reg1029, 7, %EFLAGS<imp-def,dead>
+ // insert => %reg1030<def> = MOV8rr %reg1028
+ // %reg1030<def> = ADD8rr %reg1028<kill>, %reg1029<kill>, %EFLAGS<imp-def,dead>
+ // In this case, it might not be possible to coalesce the second MOV8rr
+ // instruction if the first one is coalesced. So it would be profitable to
+ // commute it:
+ // %reg1028<def> = EXTRACT_SUBREG %reg1027<kill>, 1
+ // %reg1029<def> = MOV8rr %reg1028
+ // %reg1029<def> = SHR8ri %reg1029, 7, %EFLAGS<imp-def,dead>
+ // insert => %reg1030<def> = MOV8rr %reg1029
+ // %reg1030<def> = ADD8rr %reg1029<kill>, %reg1028<kill>, %EFLAGS<imp-def,dead>
+
+ if (!isPlainlyKilled(MI, regC, LIS))
+ return false;
+
+ // Ok, we have something like:
+ // %reg1030<def> = ADD8rr %reg1028<kill>, %reg1029<kill>, %EFLAGS<imp-def,dead>
+ // let's see if it's worth commuting it.
+
+ // Look for situations like this:
+ // %reg1024<def> = MOV r1
+ // %reg1025<def> = MOV r0
+ // %reg1026<def> = ADD %reg1024, %reg1025
+ // r0 = MOV %reg1026
+ // Commute the ADD to hopefully eliminate an otherwise unavoidable copy.
+ unsigned ToRegA = getMappedReg(regA, DstRegMap);
+ if (ToRegA) {
+ unsigned FromRegB = getMappedReg(regB, SrcRegMap);
+ unsigned FromRegC = getMappedReg(regC, SrcRegMap);
+ bool CompB = FromRegB && regsAreCompatible(FromRegB, ToRegA, TRI);
+ bool CompC = FromRegC && regsAreCompatible(FromRegC, ToRegA, TRI);
+
+ // Compute if any of the following are true:
+ // -RegB is not tied to a register and RegC is compatible with RegA.
+ // -RegB is tied to the wrong physical register, but RegC is.
+ // -RegB is tied to the wrong physical register, and RegC isn't tied.
+ if ((!FromRegB && CompC) || (FromRegB && !CompB && (!FromRegC || CompC)))
+ return true;
+ // Don't compute if any of the following are true:
+ // -RegC is not tied to a register and RegB is compatible with RegA.
+ // -RegC is tied to the wrong physical register, but RegB is.
+ // -RegC is tied to the wrong physical register, and RegB isn't tied.
+ if ((!FromRegC && CompB) || (FromRegC && !CompC && (!FromRegB || CompB)))
+ return false;
+ }
+
+ // If there is a use of regC between its last def (could be livein) and this
+ // instruction, then bail.
+ unsigned LastDefC = 0;
+ if (!noUseAfterLastDef(regC, Dist, LastDefC))
+ return false;
+
+ // If there is a use of regB between its last def (could be livein) and this
+ // instruction, then go ahead and make this transformation.
+ unsigned LastDefB = 0;
+ if (!noUseAfterLastDef(regB, Dist, LastDefB))
+ return true;
+
+ // Look for situation like this:
+ // %reg101 = MOV %reg100
+ // %reg102 = ...
+ // %reg103 = ADD %reg102, %reg101
+ // ... = %reg103 ...
+ // %reg100 = MOV %reg103
+ // If there is a reversed copy chain from reg101 to reg103, commute the ADD
+ // to eliminate an otherwise unavoidable copy.
+ // FIXME:
+ // We can extend the logic further: If an pair of operands in an insn has
+ // been merged, the insn could be regarded as a virtual copy, and the virtual
+ // copy could also be used to construct a copy chain.
+ // To more generally minimize register copies, ideally the logic of two addr
+ // instruction pass should be integrated with register allocation pass where
+ // interference graph is available.
+ if (isRevCopyChain(regC, regA, 3))
+ return true;
+
+ if (isRevCopyChain(regB, regA, 3))
+ return false;
+
+ // 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");
+ }
+ }
+ }