//===- TwoAddressInstructionPass.cpp - Two-Address instruction pass -------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This file implements the TwoAddress instruction pass which is used // by most register allocators. Two-Address instructions are rewritten // from: // // A = B op C // // to: // // A = B // A op= C // // Note that if a register allocator chooses to use this pass, that it // has to be capable of handling the non-SSA nature of these rewritten // virtual registers. // // It is also worth noting that the duplicate operand of the two // address instruction is removed. // //===----------------------------------------------------------------------===// #include "llvm/CodeGen/TwoAddressInstructionPass.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/iterator_range.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/CodeGen/LiveInterval.h" #include "llvm/CodeGen/LiveIntervals.h" #include "llvm/CodeGen/LiveVariables.h" #include "llvm/CodeGen/MachineBasicBlock.h" #include "llvm/CodeGen/MachineDominators.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineInstr.h" #include "llvm/CodeGen/MachineInstrBuilder.h" #include "llvm/CodeGen/MachineLoopInfo.h" #include "llvm/CodeGen/MachineOperand.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/Passes.h" #include "llvm/CodeGen/SlotIndexes.h" #include "llvm/CodeGen/TargetInstrInfo.h" #include "llvm/CodeGen/TargetOpcodes.h" #include "llvm/CodeGen/TargetRegisterInfo.h" #include "llvm/CodeGen/TargetSubtargetInfo.h" #include "llvm/MC/MCInstrDesc.h" #include "llvm/Pass.h" #include "llvm/Support/CodeGen.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Target/TargetMachine.h" #include #include #include using namespace llvm; #define DEBUG_TYPE "twoaddressinstruction" STATISTIC(NumTwoAddressInstrs, "Number of two-address instructions"); STATISTIC(NumCommuted , "Number of instructions commuted to coalesce"); STATISTIC(NumAggrCommuted , "Number of instructions aggressively commuted"); STATISTIC(NumConvertedTo3Addr, "Number of instructions promoted to 3-address"); STATISTIC(NumReSchedUps, "Number of instructions re-scheduled up"); STATISTIC(NumReSchedDowns, "Number of instructions re-scheduled down"); // Temporary flag to disable rescheduling. static cl::opt EnableRescheduling("twoaddr-reschedule", cl::desc("Coalesce copies by rescheduling (default=true)"), cl::init(true), cl::Hidden); // Limit the number of dataflow edges to traverse when evaluating the benefit // of commuting operands. static cl::opt MaxDataFlowEdge( "dataflow-edge-limit", cl::Hidden, cl::init(3), cl::desc("Maximum number of dataflow edges to traverse when evaluating " "the benefit of commuting operands")); namespace { class TwoAddressInstructionImpl { MachineFunction *MF = nullptr; const TargetInstrInfo *TII = nullptr; const TargetRegisterInfo *TRI = nullptr; const InstrItineraryData *InstrItins = nullptr; MachineRegisterInfo *MRI = nullptr; LiveVariables *LV = nullptr; LiveIntervals *LIS = nullptr; AliasAnalysis *AA = nullptr; CodeGenOptLevel OptLevel = CodeGenOptLevel::None; // The current basic block being processed. MachineBasicBlock *MBB = nullptr; // Keep track the distance of a MI from the start of the current basic block. DenseMap DistanceMap; // Set of already processed instructions in the current block. SmallPtrSet Processed; // A map from virtual registers to physical registers which are likely targets // to be coalesced to due to copies from physical registers to virtual // registers. e.g. v1024 = move r0. DenseMap SrcRegMap; // A map from virtual registers to physical registers which are likely targets // to be coalesced to due to copies to physical registers from virtual // registers. e.g. r1 = move v1024. DenseMap DstRegMap; MachineInstr *getSingleDef(Register Reg, MachineBasicBlock *BB) const; bool isRevCopyChain(Register FromReg, Register ToReg, int Maxlen); bool noUseAfterLastDef(Register Reg, unsigned Dist, unsigned &LastDef); bool isCopyToReg(MachineInstr &MI, Register &SrcReg, Register &DstReg, bool &IsSrcPhys, bool &IsDstPhys) const; bool isPlainlyKilled(const MachineInstr *MI, LiveRange &LR) const; bool isPlainlyKilled(const MachineInstr *MI, Register Reg) const; bool isPlainlyKilled(const MachineOperand &MO) const; bool isKilled(MachineInstr &MI, Register Reg, bool allowFalsePositives) const; MachineInstr *findOnlyInterestingUse(Register Reg, MachineBasicBlock *MBB, bool &IsCopy, Register &DstReg, bool &IsDstPhys) const; bool regsAreCompatible(Register RegA, Register RegB) const; void removeMapRegEntry(const MachineOperand &MO, DenseMap &RegMap) const; void removeClobberedSrcRegMap(MachineInstr *MI); bool regOverlapsSet(const SmallVectorImpl &Set, Register Reg) const; bool isProfitableToCommute(Register RegA, Register RegB, Register RegC, MachineInstr *MI, unsigned Dist); bool commuteInstruction(MachineInstr *MI, unsigned DstIdx, unsigned RegBIdx, unsigned RegCIdx, unsigned Dist); bool isProfitableToConv3Addr(Register RegA, Register RegB); bool convertInstTo3Addr(MachineBasicBlock::iterator &mi, MachineBasicBlock::iterator &nmi, Register RegA, Register RegB, unsigned &Dist); bool isDefTooClose(Register Reg, unsigned Dist, MachineInstr *MI); bool rescheduleMIBelowKill(MachineBasicBlock::iterator &mi, MachineBasicBlock::iterator &nmi, Register Reg); bool rescheduleKillAboveMI(MachineBasicBlock::iterator &mi, MachineBasicBlock::iterator &nmi, Register Reg); bool tryInstructionTransform(MachineBasicBlock::iterator &mi, MachineBasicBlock::iterator &nmi, unsigned SrcIdx, unsigned DstIdx, unsigned &Dist, bool shouldOnlyCommute); bool tryInstructionCommute(MachineInstr *MI, unsigned DstOpIdx, unsigned BaseOpIdx, bool BaseOpKilled, unsigned Dist); void scanUses(Register DstReg); void processCopy(MachineInstr *MI); using TiedPairList = SmallVector, 4>; using TiedOperandMap = SmallDenseMap; bool collectTiedOperands(MachineInstr *MI, TiedOperandMap&); void processTiedPairs(MachineInstr *MI, TiedPairList&, unsigned &Dist); void eliminateRegSequence(MachineBasicBlock::iterator&); bool processStatepoint(MachineInstr *MI, TiedOperandMap &TiedOperands); public: TwoAddressInstructionImpl(MachineFunction &MF, MachineFunctionPass *P); TwoAddressInstructionImpl(MachineFunction &MF, MachineFunctionAnalysisManager &MFAM); void setOptLevel(CodeGenOptLevel Level) { OptLevel = Level; } bool run(); }; class TwoAddressInstructionLegacyPass : public MachineFunctionPass { public: static char ID; // Pass identification, replacement for typeid TwoAddressInstructionLegacyPass() : MachineFunctionPass(ID) { initializeTwoAddressInstructionLegacyPassPass( *PassRegistry::getPassRegistry()); } /// Pass entry point. bool runOnMachineFunction(MachineFunction &MF) override { TwoAddressInstructionImpl Impl(MF, this); // Disable optimizations if requested. We cannot skip the whole pass as some // fixups are necessary for correctness. if (skipFunction(MF.getFunction())) Impl.setOptLevel(CodeGenOptLevel::None); return Impl.run(); } void getAnalysisUsage(AnalysisUsage &AU) const override { AU.setPreservesCFG(); AU.addUsedIfAvailable(); AU.addUsedIfAvailable(); AU.addPreserved(); AU.addPreserved(); AU.addPreserved(); AU.addPreservedID(MachineLoopInfoID); AU.addPreservedID(MachineDominatorsID); MachineFunctionPass::getAnalysisUsage(AU); } }; } // end anonymous namespace PreservedAnalyses TwoAddressInstructionPass::run(MachineFunction &MF, MachineFunctionAnalysisManager &MFAM) { // Disable optimizations if requested. We cannot skip the whole pass as some // fixups are necessary for correctness. TwoAddressInstructionImpl Impl(MF, MFAM); if (MF.getFunction().hasOptNone()) Impl.setOptLevel(CodeGenOptLevel::None); MFPropsModifier _(*this, MF); bool Changed = Impl.run(); if (!Changed) return PreservedAnalyses::all(); auto PA = getMachineFunctionPassPreservedAnalyses(); PA.preserve(); PA.preserve(); PA.preserve(); PA.preserve(); PA.preserve(); PA.preserveSet(); return PA; } char TwoAddressInstructionLegacyPass::ID = 0; char &llvm::TwoAddressInstructionPassID = TwoAddressInstructionLegacyPass::ID; INITIALIZE_PASS_BEGIN(TwoAddressInstructionLegacyPass, DEBUG_TYPE, "Two-Address instruction pass", false, false) INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) INITIALIZE_PASS_END(TwoAddressInstructionLegacyPass, DEBUG_TYPE, "Two-Address instruction pass", false, false) TwoAddressInstructionImpl::TwoAddressInstructionImpl( MachineFunction &Func, MachineFunctionAnalysisManager &MFAM) : MF(&Func), TII(Func.getSubtarget().getInstrInfo()), TRI(Func.getSubtarget().getRegisterInfo()), InstrItins(Func.getSubtarget().getInstrItineraryData()), MRI(&Func.getRegInfo()), LV(MFAM.getCachedResult(Func)), LIS(MFAM.getCachedResult(Func)), OptLevel(Func.getTarget().getOptLevel()) { auto &FAM = MFAM.getResult(Func) .getManager(); AA = FAM.getCachedResult(Func.getFunction()); } TwoAddressInstructionImpl::TwoAddressInstructionImpl(MachineFunction &Func, MachineFunctionPass *P) : MF(&Func), TII(Func.getSubtarget().getInstrInfo()), TRI(Func.getSubtarget().getRegisterInfo()), InstrItins(Func.getSubtarget().getInstrItineraryData()), MRI(&Func.getRegInfo()), OptLevel(Func.getTarget().getOptLevel()) { auto *LVWrapper = P->getAnalysisIfAvailable(); LV = LVWrapper ? &LVWrapper->getLV() : nullptr; auto *LISWrapper = P->getAnalysisIfAvailable(); LIS = LISWrapper ? &LISWrapper->getLIS() : nullptr; if (auto *AAPass = P->getAnalysisIfAvailable()) AA = &AAPass->getAAResults(); else AA = nullptr; } /// Return the MachineInstr* if it is the single def of the Reg in current BB. MachineInstr * TwoAddressInstructionImpl::getSingleDef(Register Reg, MachineBasicBlock *BB) const { MachineInstr *Ret = nullptr; for (MachineInstr &DefMI : MRI->def_instructions(Reg)) { if (DefMI.getParent() != BB || DefMI.isDebugValue()) continue; if (!Ret) Ret = &DefMI; else if (Ret != &DefMI) return nullptr; } return Ret; } /// Check if there is a reversed copy chain from FromReg to ToReg: /// %Tmp1 = copy %Tmp2; /// %FromReg = copy %Tmp1; /// %ToReg = add %FromReg ... /// %Tmp2 = copy %ToReg; /// MaxLen specifies the maximum length of the copy chain the func /// can walk through. bool TwoAddressInstructionImpl::isRevCopyChain(Register FromReg, Register ToReg, int Maxlen) { Register TmpReg = FromReg; for (int i = 0; i < Maxlen; i++) { MachineInstr *Def = getSingleDef(TmpReg, MBB); if (!Def || !Def->isCopy()) return false; TmpReg = Def->getOperand(1).getReg(); if (TmpReg == ToReg) return true; } return false; } /// Return true if there are no intervening uses between the last instruction /// in the MBB that defines the specified register and the two-address /// instruction which is being processed. It also returns the last def location /// by reference. bool TwoAddressInstructionImpl::noUseAfterLastDef(Register Reg, unsigned Dist, unsigned &LastDef) { LastDef = 0; unsigned LastUse = Dist; for (MachineOperand &MO : MRI->reg_operands(Reg)) { MachineInstr *MI = MO.getParent(); if (MI->getParent() != MBB || MI->isDebugValue()) continue; DenseMap::iterator DI = DistanceMap.find(MI); if (DI == DistanceMap.end()) continue; if (MO.isUse() && DI->second < LastUse) LastUse = DI->second; if (MO.isDef() && DI->second > LastDef) LastDef = DI->second; } return !(LastUse > LastDef && LastUse < Dist); } /// Return true if the specified MI is a copy instruction or an extract_subreg /// instruction. It also returns the source and destination registers and /// whether they are physical registers by reference. bool TwoAddressInstructionImpl::isCopyToReg(MachineInstr &MI, Register &SrcReg, Register &DstReg, bool &IsSrcPhys, bool &IsDstPhys) const { SrcReg = 0; DstReg = 0; if (MI.isCopy()) { DstReg = MI.getOperand(0).getReg(); SrcReg = MI.getOperand(1).getReg(); } else if (MI.isInsertSubreg() || MI.isSubregToReg()) { DstReg = MI.getOperand(0).getReg(); SrcReg = MI.getOperand(2).getReg(); } else { return false; } IsSrcPhys = SrcReg.isPhysical(); IsDstPhys = DstReg.isPhysical(); return true; } bool TwoAddressInstructionImpl::isPlainlyKilled(const MachineInstr *MI, LiveRange &LR) const { // This is to match the kill flag version where undefs don't have kill flags. if (!LR.hasAtLeastOneValue()) return false; SlotIndex useIdx = LIS->getInstructionIndex(*MI); LiveInterval::const_iterator I = LR.find(useIdx); assert(I != LR.end() && "Reg must be live-in to use."); return !I->end.isBlock() && SlotIndex::isSameInstr(I->end, useIdx); } /// Test if the given register value, which is used by the /// given instruction, is killed by the given instruction. bool TwoAddressInstructionImpl::isPlainlyKilled(const MachineInstr *MI, Register Reg) const { // FIXME: Sometimes tryInstructionTransform() will add instructions and // test whether they can be folded before keeping them. In this case it // sets a kill before recursively calling tryInstructionTransform() again. // If there is no interval available, we assume that this instruction is // one of those. A kill flag is manually inserted on the operand so the // check below will handle it. if (LIS && !LIS->isNotInMIMap(*MI)) { if (Reg.isVirtual()) return isPlainlyKilled(MI, LIS->getInterval(Reg)); // Reserved registers are considered always live. if (MRI->isReserved(Reg)) return false; return all_of(TRI->regunits(Reg), [&](MCRegUnit U) { return isPlainlyKilled(MI, LIS->getRegUnit(U)); }); } return MI->killsRegister(Reg, /*TRI=*/nullptr); } /// Test if the register used by the given operand is killed by the operand's /// instruction. bool TwoAddressInstructionImpl::isPlainlyKilled( const MachineOperand &MO) const { return MO.isKill() || isPlainlyKilled(MO.getParent(), MO.getReg()); } /// Test if the given register value, which is used by the given /// instruction, is killed by the given instruction. This looks through /// coalescable copies to see if the original value is potentially not killed. /// /// For example, in this code: /// /// %reg1034 = copy %reg1024 /// %reg1035 = copy killed %reg1025 /// %reg1036 = add killed %reg1034, killed %reg1035 /// /// %reg1034 is not considered to be killed, since it is copied from a /// register which is not killed. Treating it as not killed lets the /// normal heuristics commute the (two-address) add, which lets /// coalescing eliminate the extra copy. /// /// If allowFalsePositives is true then likely kills are treated as kills even /// if it can't be proven that they are kills. bool TwoAddressInstructionImpl::isKilled(MachineInstr &MI, Register Reg, bool allowFalsePositives) const { MachineInstr *DefMI = &MI; while (true) { // All uses of physical registers are likely to be kills. if (Reg.isPhysical() && (allowFalsePositives || MRI->hasOneUse(Reg))) return true; if (!isPlainlyKilled(DefMI, Reg)) return false; if (Reg.isPhysical()) return true; MachineRegisterInfo::def_iterator Begin = MRI->def_begin(Reg); // If there are multiple defs, we can't do a simple analysis, so just // go with what the kill flag says. if (std::next(Begin) != MRI->def_end()) return true; DefMI = Begin->getParent(); bool IsSrcPhys, IsDstPhys; Register SrcReg, DstReg; // If the def is something other than a copy, then it isn't going to // be coalesced, so follow the kill flag. if (!isCopyToReg(*DefMI, SrcReg, DstReg, IsSrcPhys, IsDstPhys)) return true; Reg = SrcReg; } } /// 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, Register Reg, Register &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; } /// Given a register, if all its uses are in the same basic block, return the /// last use instruction if it's a copy or a two-address use. MachineInstr *TwoAddressInstructionImpl::findOnlyInterestingUse( Register Reg, MachineBasicBlock *MBB, bool &IsCopy, Register &DstReg, bool &IsDstPhys) const { MachineOperand *UseOp = nullptr; for (MachineOperand &MO : MRI->use_nodbg_operands(Reg)) { MachineInstr *MI = MO.getParent(); if (MI->getParent() != MBB) return nullptr; if (isPlainlyKilled(MI, Reg)) UseOp = &MO; } if (!UseOp) return nullptr; MachineInstr &UseMI = *UseOp->getParent(); Register SrcReg; bool IsSrcPhys; if (isCopyToReg(UseMI, SrcReg, DstReg, IsSrcPhys, IsDstPhys)) { IsCopy = true; return &UseMI; } IsDstPhys = false; if (isTwoAddrUse(UseMI, Reg, DstReg)) { IsDstPhys = DstReg.isPhysical(); return &UseMI; } if (UseMI.isCommutable()) { unsigned Src1 = TargetInstrInfo::CommuteAnyOperandIndex; unsigned Src2 = UseOp->getOperandNo(); if (TII->findCommutedOpIndices(UseMI, Src1, Src2)) { MachineOperand &MO = UseMI.getOperand(Src1); if (MO.isReg() && MO.isUse() && isTwoAddrUse(UseMI, MO.getReg(), DstReg)) { IsDstPhys = DstReg.isPhysical(); return &UseMI; } } } return nullptr; } /// Return the physical register the specified virtual register might be mapped /// to. static MCRegister getMappedReg(Register Reg, DenseMap &RegMap) { while (Reg.isVirtual()) { DenseMap::iterator SI = RegMap.find(Reg); if (SI == RegMap.end()) return 0; Reg = SI->second; } if (Reg.isPhysical()) return Reg; return 0; } /// Return true if the two registers are equal or aliased. bool TwoAddressInstructionImpl::regsAreCompatible(Register RegA, Register RegB) const { if (RegA == RegB) return true; if (!RegA || !RegB) return false; return TRI->regsOverlap(RegA, RegB); } /// From RegMap remove entries mapped to a physical register which overlaps MO. void TwoAddressInstructionImpl::removeMapRegEntry( const MachineOperand &MO, DenseMap &RegMap) const { assert( (MO.isReg() || MO.isRegMask()) && "removeMapRegEntry must be called with a register or regmask operand."); SmallVector Srcs; for (auto SI : RegMap) { Register ToReg = SI.second; if (ToReg.isVirtual()) continue; if (MO.isReg()) { Register Reg = MO.getReg(); if (TRI->regsOverlap(ToReg, Reg)) Srcs.push_back(SI.first); } else if (MO.clobbersPhysReg(ToReg)) Srcs.push_back(SI.first); } for (auto SrcReg : Srcs) RegMap.erase(SrcReg); } /// If a physical register is clobbered, old entries mapped to it should be /// deleted. For example /// /// %2:gr64 = COPY killed $rdx /// MUL64r %3:gr64, implicit-def $rax, implicit-def $rdx /// /// After the MUL instruction, $rdx contains different value than in the COPY /// instruction. So %2 should not map to $rdx after MUL. void TwoAddressInstructionImpl::removeClobberedSrcRegMap(MachineInstr *MI) { if (MI->isCopy()) { // If a virtual register is copied to its mapped physical register, it // doesn't change the potential coalescing between them, so we don't remove // entries mapped to the physical register. For example // // %100 = COPY $r8 // ... // $r8 = COPY %100 // // The first copy constructs SrcRegMap[%100] = $r8, the second copy doesn't // destroy the content of $r8, and should not impact SrcRegMap. Register Dst = MI->getOperand(0).getReg(); if (!Dst || Dst.isVirtual()) return; Register Src = MI->getOperand(1).getReg(); if (regsAreCompatible(Dst, getMappedReg(Src, SrcRegMap))) return; } for (const MachineOperand &MO : MI->operands()) { if (MO.isRegMask()) { removeMapRegEntry(MO, SrcRegMap); continue; } if (!MO.isReg() || !MO.isDef()) continue; Register Reg = MO.getReg(); if (!Reg || Reg.isVirtual()) continue; removeMapRegEntry(MO, SrcRegMap); } } // Returns true if Reg is equal or aliased to at least one register in Set. bool TwoAddressInstructionImpl::regOverlapsSet( const SmallVectorImpl &Set, Register Reg) const { for (unsigned R : Set) if (TRI->regsOverlap(R, Reg)) return true; return false; } /// Return true if it's potentially profitable to commute the two-address /// instruction that's being processed. bool TwoAddressInstructionImpl::isProfitableToCommute(Register RegA, Register RegB, Register RegC, MachineInstr *MI, unsigned Dist) { if (OptLevel == CodeGenOptLevel::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 = EXTRACT_SUBREG killed %reg1027, 1 // %reg1029 = COPY %reg1028 // %reg1029 = SHR8ri %reg1029, 7, implicit dead %eflags // insert => %reg1030 = COPY %reg1028 // %reg1030 = ADD8rr killed %reg1028, killed %reg1029, implicit dead %eflags // In this case, it might not be possible to coalesce the second COPY // instruction if the first one is coalesced. So it would be profitable to // commute it: // %reg1028 = EXTRACT_SUBREG killed %reg1027, 1 // %reg1029 = COPY %reg1028 // %reg1029 = SHR8ri %reg1029, 7, implicit dead %eflags // insert => %reg1030 = COPY %reg1029 // %reg1030 = ADD8rr killed %reg1029, killed %reg1028, implicit dead %eflags if (!isPlainlyKilled(MI, RegC)) return false; // Ok, we have something like: // %reg1030 = ADD8rr killed %reg1028, killed %reg1029, implicit dead %eflags // let's see if it's worth commuting it. // Look for situations like this: // %reg1024 = MOV r1 // %reg1025 = MOV r0 // %reg1026 = ADD %reg1024, %reg1025 // r0 = MOV %reg1026 // Commute the ADD to hopefully eliminate an otherwise unavoidable copy. MCRegister ToRegA = getMappedReg(RegA, DstRegMap); if (ToRegA) { MCRegister FromRegB = getMappedReg(RegB, SrcRegMap); MCRegister FromRegC = getMappedReg(RegC, SrcRegMap); bool CompB = FromRegB && regsAreCompatible(FromRegB, ToRegA); bool CompC = FromRegC && regsAreCompatible(FromRegC, ToRegA); // 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, MaxDataFlowEdge)) return true; if (isRevCopyChain(RegB, RegA, MaxDataFlowEdge)) return false; // Look for other target specific commute preference. bool Commute; if (TII->hasCommutePreference(*MI, Commute)) return Commute; // 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; } /// Commute a two-address instruction and update the basic block, distance map, /// and live variables if needed. Return true if it is successful. bool TwoAddressInstructionImpl::commuteInstruction(MachineInstr *MI, unsigned DstIdx, unsigned RegBIdx, unsigned RegCIdx, unsigned Dist) { Register RegC = MI->getOperand(RegCIdx).getReg(); LLVM_DEBUG(dbgs() << "2addr: COMMUTING : " << *MI); MachineInstr *NewMI = TII->commuteInstruction(*MI, false, RegBIdx, RegCIdx); if (NewMI == nullptr) { LLVM_DEBUG(dbgs() << "2addr: COMMUTING FAILED!\n"); return false; } LLVM_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. MCRegister FromRegC = getMappedReg(RegC, SrcRegMap); if (FromRegC) { Register RegA = MI->getOperand(DstIdx).getReg(); SrcRegMap[RegA] = FromRegC; } return true; } /// Return true if it is profitable to convert the given 2-address instruction /// to a 3-address one. bool TwoAddressInstructionImpl::isProfitableToConv3Addr(Register RegA, Register RegB) { // Look for situations like this: // %reg1024 = MOV r1 // %reg1025 = MOV r0 // %reg1026 = ADD %reg1024, %reg1025 // r2 = MOV %reg1026 // Turn ADD into a 3-address instruction to avoid a copy. MCRegister FromRegB = getMappedReg(RegB, SrcRegMap); if (!FromRegB) return false; MCRegister ToRegA = getMappedReg(RegA, DstRegMap); return (ToRegA && !regsAreCompatible(FromRegB, ToRegA)); } /// Convert the specified two-address instruction into a three address one. /// Return true if this transformation was successful. bool TwoAddressInstructionImpl::convertInstTo3Addr( MachineBasicBlock::iterator &mi, MachineBasicBlock::iterator &nmi, Register RegA, Register RegB, unsigned &Dist) { MachineInstrSpan MIS(mi, MBB); MachineInstr *NewMI = TII->convertToThreeAddress(*mi, LV, LIS); if (!NewMI) return false; LLVM_DEBUG(dbgs() << "2addr: CONVERTING 2-ADDR: " << *mi); LLVM_DEBUG(dbgs() << "2addr: TO 3-ADDR: " << *NewMI); // If the old instruction is debug value tracked, an update is required. if (auto OldInstrNum = mi->peekDebugInstrNum()) { assert(mi->getNumExplicitDefs() == 1); assert(NewMI->getNumExplicitDefs() == 1); // Find the old and new def location. unsigned OldIdx = mi->defs().begin()->getOperandNo(); unsigned NewIdx = NewMI->defs().begin()->getOperandNo(); // Record that one def has been replaced by the other. unsigned NewInstrNum = NewMI->getDebugInstrNum(); MF->makeDebugValueSubstitution(std::make_pair(OldInstrNum, OldIdx), std::make_pair(NewInstrNum, NewIdx)); } MBB->erase(mi); // Nuke the old inst. for (MachineInstr &MI : MIS) DistanceMap.insert(std::make_pair(&MI, Dist++)); Dist--; mi = NewMI; nmi = std::next(mi); // Update source and destination register maps. SrcRegMap.erase(RegA); DstRegMap.erase(RegB); return true; } /// Scan forward recursively for only uses, update maps if the use is a copy or /// a two-address instruction. void TwoAddressInstructionImpl::scanUses(Register DstReg) { SmallVector VirtRegPairs; bool IsDstPhys; bool IsCopy = false; Register NewReg; Register Reg = DstReg; while (MachineInstr *UseMI = findOnlyInterestingUse(Reg, MBB, IsCopy, NewReg, IsDstPhys)) { if (IsCopy && !Processed.insert(UseMI).second) break; DenseMap::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; } SrcRegMap[NewReg] = Reg; VirtRegPairs.push_back(NewReg); Reg = NewReg; } if (!VirtRegPairs.empty()) { unsigned ToReg = VirtRegPairs.back(); VirtRegPairs.pop_back(); while (!VirtRegPairs.empty()) { unsigned FromReg = VirtRegPairs.pop_back_val(); 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!"); } } /// 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 TwoAddressInstructionImpl::processCopy(MachineInstr *MI) { if (Processed.count(MI)) return; bool IsSrcPhys, IsDstPhys; Register SrcReg, DstReg; if (!isCopyToReg(*MI, 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); } /// 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 TwoAddressInstructionImpl::rescheduleMIBelowKill( MachineBasicBlock::iterator &mi, MachineBasicBlock::iterator &nmi, Register 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::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; Register 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; SmallVector Uses; SmallVector Kills; SmallVector Defs; for (const MachineOperand &MO : MI->operands()) { if (!MO.isReg()) continue; Register MOReg = MO.getReg(); if (!MOReg) continue; if (MO.isDef()) Defs.push_back(MOReg); else { Uses.push_back(MOReg); if (MOReg != Reg && isPlainlyKilled(MO)) Kills.push_back(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 != MBB->end()) { End = skipDebugInstructionsForward(End, MBB->end()); if (End->isCopy() && regOverlapsSet(Defs, End->getOperand(1).getReg())) Defs.push_back(End->getOperand(0).getReg()); else break; ++End; } // Check if the reschedule will not break dependencies. unsigned NumVisited = 0; MachineBasicBlock::iterator KillPos = KillMI; ++KillPos; for (MachineInstr &OtherMI : make_range(End, KillPos)) { // Debug or pseudo instructions cannot be counted against the limit. if (OtherMI.isDebugOrPseudoInstr()) 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 (const MachineOperand &MO : OtherMI.operands()) { if (!MO.isReg()) continue; Register MOReg = MO.getReg(); if (!MOReg) continue; if (MO.isDef()) { if (regOverlapsSet(Uses, MOReg)) // Physical register use would be clobbered. return false; if (!MO.isDead() && regOverlapsSet(Defs, 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 (regOverlapsSet(Defs, MOReg)) return false; bool isKill = isPlainlyKilled(MO); if (MOReg != Reg && ((isKill && regOverlapsSet(Uses, MOReg)) || regOverlapsSet(Kills, 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)->isDebugInstr()) --Begin; nmi = End; MachineBasicBlock::iterator InsertPos = KillPos; if (LIS) { // We have to move the copies (and any interleaved debug instructions) // first so that the MBB is still well-formed when calling handleMove(). for (MachineBasicBlock::iterator MBBI = AfterMI; MBBI != End;) { auto CopyMI = MBBI++; MBB->splice(InsertPos, MBB, CopyMI); if (!CopyMI->isDebugOrPseudoInstr()) 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); } LLVM_DEBUG(dbgs() << "\trescheduled below kill: " << *KillMI); return true; } /// Return true if the re-scheduling will put the given instruction too close /// to the defs of its register dependencies. bool TwoAddressInstructionImpl::isDefTooClose(Register 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::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; } /// 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 TwoAddressInstructionImpl::rescheduleKillAboveMI( MachineBasicBlock::iterator &mi, MachineBasicBlock::iterator &nmi, Register 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::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; Register DstReg; if (isTwoAddrUse(*KillMI, Reg, DstReg)) return false; bool SeenStore = true; if (!KillMI->isSafeToMove(AA, SeenStore)) return false; SmallVector Uses; SmallVector Kills; SmallVector Defs; SmallVector LiveDefs; for (const MachineOperand &MO : KillMI->operands()) { if (!MO.isReg()) continue; Register MOReg = MO.getReg(); if (MO.isUse()) { if (!MOReg) continue; if (isDefTooClose(MOReg, DI->second, MI)) return false; bool isKill = isPlainlyKilled(MO); if (MOReg == Reg && !isKill) return false; Uses.push_back(MOReg); if (isKill && MOReg != Reg) Kills.push_back(MOReg); } else if (MOReg.isPhysical()) { Defs.push_back(MOReg); if (!MO.isDead()) LiveDefs.push_back(MOReg); } } // Check if the reschedule will not break depedencies. unsigned NumVisited = 0; for (MachineInstr &OtherMI : make_range(mi, MachineBasicBlock::iterator(KillMI))) { // Debug or pseudo instructions cannot be counted against the limit. if (OtherMI.isDebugOrPseudoInstr()) 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 OtherDefs; for (const MachineOperand &MO : OtherMI.operands()) { if (!MO.isReg()) continue; Register MOReg = MO.getReg(); if (!MOReg) continue; if (MO.isUse()) { if (regOverlapsSet(Defs, MOReg)) // Moving KillMI can clobber the physical register if the def has // not been seen. return false; if (regOverlapsSet(Kills, MOReg)) // Don't want to extend other live ranges and update kills. return false; if (&OtherMI != MI && MOReg == Reg && !isPlainlyKilled(MO)) // We can't schedule across a use of the register in question. return false; } else { OtherDefs.push_back(MOReg); } } for (Register MOReg : OtherDefs) { if (regOverlapsSet(Uses, MOReg)) return false; if (MOReg.isPhysical() && regOverlapsSet(LiveDefs, MOReg)) return false; // Physical register def is seen. llvm::erase(Defs, 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)->isDebugInstr()) --InsertPos; MachineBasicBlock::iterator From = KillMI; MachineBasicBlock::iterator To = std::next(From); while (std::prev(From)->isDebugInstr()) --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); } LLVM_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 TwoAddressInstructionImpl::tryInstructionCommute(MachineInstr *MI, unsigned DstOpIdx, unsigned BaseOpIdx, bool BaseOpKilled, unsigned Dist) { if (!MI->isCommutable()) return false; bool MadeChange = false; Register DstOpReg = MI->getOperand(DstOpIdx).getReg(); Register 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 search for // other commutable operands and does not change the values of passed // variables. if (OtherOpIdx == BaseOpIdx || !MI->getOperand(OtherOpIdx).isReg() || !TII->findCommutedOpIndices(*MI, BaseOpIdx, OtherOpIdx)) continue; Register 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 OtherOpKilled = isKilled(*MI, OtherOpReg, false); bool DoCommute = !BaseOpKilled && OtherOpKilled; 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, DstOpIdx, BaseOpIdx, OtherOpIdx, Dist)) { MadeChange = true; ++NumCommuted; if (AggressiveCommute) ++NumAggrCommuted; // There might be more than two commutable operands, update BaseOp and // continue scanning. // FIXME: This assumes that the new instruction's operands are in the // same positions and were simply swapped. BaseOpReg = OtherOpReg; BaseOpKilled = OtherOpKilled; // Resamples OpsNum in case the number of operands was reduced. This // happens with X86. OpsNum = MI->getDesc().getNumOperands(); } } return MadeChange; } /// 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 TwoAddressInstructionImpl::tryInstructionTransform( MachineBasicBlock::iterator &mi, MachineBasicBlock::iterator &nmi, unsigned SrcIdx, unsigned DstIdx, unsigned &Dist, bool shouldOnlyCommute) { if (OptLevel == CodeGenOptLevel::None) return false; MachineInstr &MI = *mi; Register regA = MI.getOperand(DstIdx).getReg(); Register regB = MI.getOperand(SrcIdx).getReg(); assert(regB.isVirtual() && "cannot make instruction into two-address form"); bool regBKilled = isKilled(MI, regB, true); if (regA.isVirtual()) 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 aggressively 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, 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. LLVM_DEBUG(dbgs() << "2addr: UNFOLDING: " << MI); const TargetRegisterClass *RC = TRI->getAllocatableClass( TII->getRegClass(UnfoldMCID, LoadRegIndex, TRI, *MF)); Register Reg = MRI->createVirtualRegister(RC); SmallVector NewMIs; if (!TII->unfoldMemoryOperand(*MF, MI, Reg, /*UnfoldLoad=*/true, /*UnfoldStore=*/false, NewMIs)) { LLVM_DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n"); return false; } assert(NewMIs.size() == 2 && "Unfolded a load into multiple instructions!"); // The load was previously folded, so this is the only use. NewMIs[1]->addRegisterKilled(Reg, TRI); // Tentatively insert the instructions into the block so that they // look "normal" to the transformation logic. MBB->insert(mi, NewMIs[0]); MBB->insert(mi, NewMIs[1]); DistanceMap.insert(std::make_pair(NewMIs[0], Dist++)); DistanceMap.insert(std::make_pair(NewMIs[1], Dist)); LLVM_DEBUG(dbgs() << "2addr: NEW LOAD: " << *NewMIs[0] << "2addr: NEW INST: " << *NewMIs[1]); // Transform the instruction, now that it no longer has a load. unsigned NewDstIdx = NewMIs[1]->findRegisterDefOperandIdx(regA, /*TRI=*/nullptr); unsigned NewSrcIdx = NewMIs[1]->findRegisterUseOperandIdx(regB, /*TRI=*/nullptr); MachineBasicBlock::iterator NewMI = NewMIs[1]; bool TransformResult = tryInstructionTransform(NewMI, mi, NewSrcIdx, NewDstIdx, Dist, true); (void)TransformResult; assert(!TransformResult && "tryInstructionTransform() should return false."); if (NewMIs[1]->getOperand(NewSrcIdx).isKill()) { // Success, or at least we made an improvement. Keep the unfolded // instructions and discard the original. if (LV) { for (const MachineOperand &MO : MI.operands()) { if (MO.isReg() && MO.getReg().isVirtual()) { if (MO.isUse()) { if (MO.isKill()) { if (NewMIs[0]->killsRegister(MO.getReg(), /*TRI=*/nullptr)) LV->replaceKillInstruction(MO.getReg(), MI, *NewMIs[0]); else { assert(NewMIs[1]->killsRegister(MO.getReg(), /*TRI=*/nullptr) && "Kill missing after load unfold!"); LV->replaceKillInstruction(MO.getReg(), MI, *NewMIs[1]); } } } else if (LV->removeVirtualRegisterDead(MO.getReg(), MI)) { if (NewMIs[1]->registerDefIsDead(MO.getReg(), /*TRI=*/nullptr)) LV->addVirtualRegisterDead(MO.getReg(), *NewMIs[1]); else { assert(NewMIs[0]->registerDefIsDead(MO.getReg(), /*TRI=*/nullptr) && "Dead flag missing after load unfold!"); LV->addVirtualRegisterDead(MO.getReg(), *NewMIs[0]); } } } } LV->addVirtualRegisterKilled(Reg, *NewMIs[1]); } SmallVector OrigRegs; if (LIS) { for (const MachineOperand &MO : MI.operands()) { if (MO.isReg()) OrigRegs.push_back(MO.getReg()); } LIS->RemoveMachineInstrFromMaps(MI); } MI.eraseFromParent(); DistanceMap.erase(&MI); // Update LiveIntervals. if (LIS) { MachineBasicBlock::iterator Begin(NewMIs[0]); MachineBasicBlock::iterator End(NewMIs[1]); LIS->repairIntervalsInRange(MBB, Begin, End, OrigRegs); } mi = NewMIs[1]; } else { // Transforming didn't eliminate the tie and didn't lead to an // improvement. Clean up the unfolded instructions and keep the // original. LLVM_DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n"); NewMIs[0]->eraseFromParent(); NewMIs[1]->eraseFromParent(); DistanceMap.erase(NewMIs[0]); DistanceMap.erase(NewMIs[1]); Dist--; } } } } return false; } // Collect tied operands of MI that need to be handled. // Rewrite trivial cases immediately. // Return true if any tied operands where found, including the trivial ones. bool TwoAddressInstructionImpl::collectTiedOperands( MachineInstr *MI, TiedOperandMap &TiedOperands) { bool AnyOps = false; unsigned NumOps = MI->getNumOperands(); for (unsigned SrcIdx = 0; SrcIdx < NumOps; ++SrcIdx) { unsigned DstIdx = 0; if (!MI->isRegTiedToDefOperand(SrcIdx, &DstIdx)) continue; AnyOps = true; MachineOperand &SrcMO = MI->getOperand(SrcIdx); MachineOperand &DstMO = MI->getOperand(DstIdx); Register SrcReg = SrcMO.getReg(); Register DstReg = DstMO.getReg(); // Tied constraint already satisfied? if (SrcReg == DstReg) continue; assert(SrcReg && SrcMO.isUse() && "two address instruction invalid"); // Deal with undef uses immediately - simply rewrite the src operand. if (SrcMO.isUndef() && !DstMO.getSubReg()) { // Constrain the DstReg register class if required. if (DstReg.isVirtual()) { const TargetRegisterClass *RC = MRI->getRegClass(SrcReg); MRI->constrainRegClass(DstReg, RC); } SrcMO.setReg(DstReg); SrcMO.setSubReg(0); LLVM_DEBUG(dbgs() << "\t\trewrite undef:\t" << *MI); continue; } TiedOperands[SrcReg].push_back(std::make_pair(SrcIdx, DstIdx)); } return AnyOps; } // Process a list of tied MI operands that all use the same source register. // The tied pairs are of the form (SrcIdx, DstIdx). void TwoAddressInstructionImpl::processTiedPairs(MachineInstr *MI, TiedPairList &TiedPairs, unsigned &Dist) { bool IsEarlyClobber = llvm::any_of(TiedPairs, [MI](auto const &TP) { return MI->getOperand(TP.second).isEarlyClobber(); }); bool RemovedKillFlag = false; bool AllUsesCopied = true; unsigned LastCopiedReg = 0; SlotIndex LastCopyIdx; Register RegB = 0; unsigned SubRegB = 0; for (auto &TP : TiedPairs) { unsigned SrcIdx = TP.first; unsigned DstIdx = TP.second; const MachineOperand &DstMO = MI->getOperand(DstIdx); Register RegA = DstMO.getReg(); // Grab RegB from the instruction because it may have changed if the // instruction was commuted. RegB = MI->getOperand(SrcIdx).getReg(); SubRegB = MI->getOperand(SrcIdx).getSubReg(); if (RegA == RegB) { // The register is tied to multiple destinations (or else we would // not have continued this far), but this use of the register // already matches the tied destination. Leave it. AllUsesCopied = false; continue; } LastCopiedReg = RegA; assert(RegB.isVirtual() && "cannot make instruction into two-address form"); #ifndef NDEBUG // First, verify that we don't have a use of "a" in the instruction // (a = b + a for example) because our transformation will not // work. This should never occur because we are in SSA form. for (unsigned i = 0; i != MI->getNumOperands(); ++i) assert(i == DstIdx || !MI->getOperand(i).isReg() || MI->getOperand(i).getReg() != RegA); #endif // Emit a copy. MachineInstrBuilder MIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), TII->get(TargetOpcode::COPY), RegA); // If this operand is folding a truncation, the truncation now moves to the // copy so that the register classes remain valid for the operands. MIB.addReg(RegB, 0, SubRegB); const TargetRegisterClass *RC = MRI->getRegClass(RegB); if (SubRegB) { if (RegA.isVirtual()) { assert(TRI->getMatchingSuperRegClass(RC, MRI->getRegClass(RegA), SubRegB) && "tied subregister must be a truncation"); // The superreg class will not be used to constrain the subreg class. RC = nullptr; } else { assert(TRI->getMatchingSuperReg(RegA, SubRegB, MRI->getRegClass(RegB)) && "tied subregister must be a truncation"); } } // Update DistanceMap. MachineBasicBlock::iterator PrevMI = MI; --PrevMI; DistanceMap.insert(std::make_pair(&*PrevMI, Dist)); DistanceMap[MI] = ++Dist; if (LIS) { LastCopyIdx = LIS->InsertMachineInstrInMaps(*PrevMI).getRegSlot(); SlotIndex endIdx = LIS->getInstructionIndex(*MI).getRegSlot(IsEarlyClobber); if (RegA.isVirtual()) { LiveInterval &LI = LIS->getInterval(RegA); VNInfo *VNI = LI.getNextValue(LastCopyIdx, LIS->getVNInfoAllocator()); LI.addSegment(LiveRange::Segment(LastCopyIdx, endIdx, VNI)); for (auto &S : LI.subranges()) { VNI = S.getNextValue(LastCopyIdx, LIS->getVNInfoAllocator()); S.addSegment(LiveRange::Segment(LastCopyIdx, endIdx, VNI)); } } else { for (MCRegUnit Unit : TRI->regunits(RegA)) { if (LiveRange *LR = LIS->getCachedRegUnit(Unit)) { VNInfo *VNI = LR->getNextValue(LastCopyIdx, LIS->getVNInfoAllocator()); LR->addSegment(LiveRange::Segment(LastCopyIdx, endIdx, VNI)); } } } } LLVM_DEBUG(dbgs() << "\t\tprepend:\t" << *MIB); MachineOperand &MO = MI->getOperand(SrcIdx); assert(MO.isReg() && MO.getReg() == RegB && MO.isUse() && "inconsistent operand info for 2-reg pass"); if (isPlainlyKilled(MO)) { MO.setIsKill(false); RemovedKillFlag = true; } // Make sure regA is a legal regclass for the SrcIdx operand. if (RegA.isVirtual() && RegB.isVirtual()) MRI->constrainRegClass(RegA, RC); MO.setReg(RegA); // The getMatchingSuper asserts guarantee that the register class projected // by SubRegB is compatible with RegA with no subregister. So regardless of // whether the dest oper writes a subreg, the source oper should not. MO.setSubReg(0); } if (AllUsesCopied) { LaneBitmask RemainingUses = LaneBitmask::getNone(); // Replace other (un-tied) uses of regB with LastCopiedReg. for (MachineOperand &MO : MI->all_uses()) { if (MO.getReg() == RegB) { if (MO.getSubReg() == SubRegB && !IsEarlyClobber) { if (isPlainlyKilled(MO)) { MO.setIsKill(false); RemovedKillFlag = true; } MO.setReg(LastCopiedReg); MO.setSubReg(0); } else { RemainingUses |= TRI->getSubRegIndexLaneMask(MO.getSubReg()); } } } // Update live variables for regB. if (RemovedKillFlag && RemainingUses.none() && LV && LV->getVarInfo(RegB).removeKill(*MI)) { MachineBasicBlock::iterator PrevMI = MI; --PrevMI; LV->addVirtualRegisterKilled(RegB, *PrevMI); } if (RemovedKillFlag && RemainingUses.none()) SrcRegMap[LastCopiedReg] = RegB; // Update LiveIntervals. if (LIS) { SlotIndex UseIdx = LIS->getInstructionIndex(*MI); auto Shrink = [=](LiveRange &LR, LaneBitmask LaneMask) { LiveRange::Segment *S = LR.getSegmentContaining(LastCopyIdx); if (!S) return true; if ((LaneMask & RemainingUses).any()) return false; if (S->end.getBaseIndex() != UseIdx) return false; S->end = LastCopyIdx; return true; }; LiveInterval &LI = LIS->getInterval(RegB); bool ShrinkLI = true; for (auto &S : LI.subranges()) ShrinkLI &= Shrink(S, S.LaneMask); if (ShrinkLI) Shrink(LI, LaneBitmask::getAll()); } } else if (RemovedKillFlag) { // Some tied uses of regB matched their destination registers, so // regB is still used in this instruction, but a kill flag was // removed from a different tied use of regB, so now we need to add // a kill flag to one of the remaining uses of regB. for (MachineOperand &MO : MI->all_uses()) { if (MO.getReg() == RegB) { MO.setIsKill(true); break; } } } } // For every tied operand pair this function transforms statepoint from // RegA = STATEPOINT ... RegB(tied-def N) // to // RegB = STATEPOINT ... RegB(tied-def N) // and replaces all uses of RegA with RegB. // No extra COPY instruction is necessary because tied use is killed at // STATEPOINT. bool TwoAddressInstructionImpl::processStatepoint( MachineInstr *MI, TiedOperandMap &TiedOperands) { bool NeedCopy = false; for (auto &TO : TiedOperands) { Register RegB = TO.first; if (TO.second.size() != 1) { NeedCopy = true; continue; } unsigned SrcIdx = TO.second[0].first; unsigned DstIdx = TO.second[0].second; MachineOperand &DstMO = MI->getOperand(DstIdx); Register RegA = DstMO.getReg(); assert(RegB == MI->getOperand(SrcIdx).getReg()); if (RegA == RegB) continue; // CodeGenPrepare can sink pointer compare past statepoint, which // breaks assumption that statepoint kills tied-use register when // in SSA form (see note in IR/SafepointIRVerifier.cpp). Fall back // to generic tied register handling to avoid assertion failures. // TODO: Recompute LIS/LV information for new range here. if (LIS) { const auto &UseLI = LIS->getInterval(RegB); const auto &DefLI = LIS->getInterval(RegA); if (DefLI.overlaps(UseLI)) { LLVM_DEBUG(dbgs() << "LIS: " << printReg(RegB, TRI, 0) << " UseLI overlaps with DefLI\n"); NeedCopy = true; continue; } } else if (LV && LV->getVarInfo(RegB).findKill(MI->getParent()) != MI) { // Note that MachineOperand::isKill does not work here, because it // is set only on first register use in instruction and for statepoint // tied-use register will usually be found in preceeding deopt bundle. LLVM_DEBUG(dbgs() << "LV: " << printReg(RegB, TRI, 0) << " not killed by statepoint\n"); NeedCopy = true; continue; } if (!MRI->constrainRegClass(RegB, MRI->getRegClass(RegA))) { LLVM_DEBUG(dbgs() << "MRI: couldn't constrain" << printReg(RegB, TRI, 0) << " to register class of " << printReg(RegA, TRI, 0) << '\n'); NeedCopy = true; continue; } MRI->replaceRegWith(RegA, RegB); if (LIS) { VNInfo::Allocator &A = LIS->getVNInfoAllocator(); LiveInterval &LI = LIS->getInterval(RegB); LiveInterval &Other = LIS->getInterval(RegA); SmallVector NewVNIs; for (const VNInfo *VNI : Other.valnos) { assert(VNI->id == NewVNIs.size() && "assumed"); NewVNIs.push_back(LI.createValueCopy(VNI, A)); } for (auto &S : Other) { VNInfo *VNI = NewVNIs[S.valno->id]; LiveRange::Segment NewSeg(S.start, S.end, VNI); LI.addSegment(NewSeg); } LIS->removeInterval(RegA); } if (LV) { if (MI->getOperand(SrcIdx).isKill()) LV->removeVirtualRegisterKilled(RegB, *MI); LiveVariables::VarInfo &SrcInfo = LV->getVarInfo(RegB); LiveVariables::VarInfo &DstInfo = LV->getVarInfo(RegA); SrcInfo.AliveBlocks |= DstInfo.AliveBlocks; DstInfo.AliveBlocks.clear(); for (auto *KillMI : DstInfo.Kills) LV->addVirtualRegisterKilled(RegB, *KillMI, false); } } return !NeedCopy; } /// Reduce two-address instructions to two operands. bool TwoAddressInstructionImpl::run() { bool MadeChange = false; LLVM_DEBUG(dbgs() << "********** REWRITING TWO-ADDR INSTRS **********\n"); LLVM_DEBUG(dbgs() << "********** Function: " << MF->getName() << '\n'); // This pass takes the function out of SSA form. MRI->leaveSSA(); // This pass will rewrite the tied-def to meet the RegConstraint. MF->getProperties() .set(MachineFunctionProperties::Property::TiedOpsRewritten); TiedOperandMap TiedOperands; for (MachineBasicBlock &MBBI : *MF) { MBB = &MBBI; unsigned Dist = 0; DistanceMap.clear(); SrcRegMap.clear(); DstRegMap.clear(); Processed.clear(); for (MachineBasicBlock::iterator mi = MBB->begin(), me = MBB->end(); mi != me; ) { MachineBasicBlock::iterator nmi = std::next(mi); // Skip debug instructions. if (mi->isDebugInstr()) { mi = nmi; continue; } // Expand REG_SEQUENCE instructions. This will position mi at the first // expanded instruction. if (mi->isRegSequence()) eliminateRegSequence(mi); DistanceMap.insert(std::make_pair(&*mi, ++Dist)); processCopy(&*mi); // First scan through all the tied register uses in this instruction // and record a list of pairs of tied operands for each register. if (!collectTiedOperands(&*mi, TiedOperands)) { removeClobberedSrcRegMap(&*mi); mi = nmi; continue; } ++NumTwoAddressInstrs; MadeChange = true; LLVM_DEBUG(dbgs() << '\t' << *mi); // If the instruction has a single pair of tied operands, try some // transformations that may either eliminate the tied operands or // improve the opportunities for coalescing away the register copy. if (TiedOperands.size() == 1) { SmallVectorImpl> &TiedPairs = TiedOperands.begin()->second; if (TiedPairs.size() == 1) { unsigned SrcIdx = TiedPairs[0].first; unsigned DstIdx = TiedPairs[0].second; Register SrcReg = mi->getOperand(SrcIdx).getReg(); Register DstReg = mi->getOperand(DstIdx).getReg(); if (SrcReg != DstReg && tryInstructionTransform(mi, nmi, SrcIdx, DstIdx, Dist, false)) { // The tied operands have been eliminated or shifted further down // the block to ease elimination. Continue processing with 'nmi'. TiedOperands.clear(); removeClobberedSrcRegMap(&*mi); mi = nmi; continue; } } } if (mi->getOpcode() == TargetOpcode::STATEPOINT && processStatepoint(&*mi, TiedOperands)) { TiedOperands.clear(); LLVM_DEBUG(dbgs() << "\t\trewrite to:\t" << *mi); mi = nmi; continue; } // Now iterate over the information collected above. for (auto &TO : TiedOperands) { processTiedPairs(&*mi, TO.second, Dist); LLVM_DEBUG(dbgs() << "\t\trewrite to:\t" << *mi); } // Rewrite INSERT_SUBREG as COPY now that we no longer need SSA form. if (mi->isInsertSubreg()) { // From %reg = INSERT_SUBREG %reg, %subreg, subidx // To %reg:subidx = COPY %subreg unsigned SubIdx = mi->getOperand(3).getImm(); mi->removeOperand(3); assert(mi->getOperand(0).getSubReg() == 0 && "Unexpected subreg idx"); mi->getOperand(0).setSubReg(SubIdx); mi->getOperand(0).setIsUndef(mi->getOperand(1).isUndef()); mi->removeOperand(1); mi->setDesc(TII->get(TargetOpcode::COPY)); LLVM_DEBUG(dbgs() << "\t\tconvert to:\t" << *mi); // Update LiveIntervals. if (LIS) { Register Reg = mi->getOperand(0).getReg(); LiveInterval &LI = LIS->getInterval(Reg); if (LI.hasSubRanges()) { // The COPY no longer defines subregs of %reg except for // %reg.subidx. LaneBitmask LaneMask = TRI->getSubRegIndexLaneMask(mi->getOperand(0).getSubReg()); SlotIndex Idx = LIS->getInstructionIndex(*mi).getRegSlot(); for (auto &S : LI.subranges()) { if ((S.LaneMask & LaneMask).none()) { LiveRange::iterator DefSeg = S.FindSegmentContaining(Idx); if (mi->getOperand(0).isUndef()) { S.removeValNo(DefSeg->valno); } else { LiveRange::iterator UseSeg = std::prev(DefSeg); S.MergeValueNumberInto(DefSeg->valno, UseSeg->valno); } } } // The COPY no longer has a use of %reg. LIS->shrinkToUses(&LI); } else { // The live interval for Reg did not have subranges but now it needs // them because we have introduced a subreg def. Recompute it. LIS->removeInterval(Reg); LIS->createAndComputeVirtRegInterval(Reg); } } } // Clear TiedOperands here instead of at the top of the loop // since most instructions do not have tied operands. TiedOperands.clear(); removeClobberedSrcRegMap(&*mi); mi = nmi; } } return MadeChange; } /// Eliminate a REG_SEQUENCE instruction as part of the de-ssa process. /// /// The instruction is turned into a sequence of sub-register copies: /// /// %dst = REG_SEQUENCE %v1, ssub0, %v2, ssub1 /// /// Becomes: /// /// undef %dst:ssub0 = COPY %v1 /// %dst:ssub1 = COPY %v2 void TwoAddressInstructionImpl::eliminateRegSequence( MachineBasicBlock::iterator &MBBI) { MachineInstr &MI = *MBBI; Register DstReg = MI.getOperand(0).getReg(); SmallVector OrigRegs; VNInfo *DefVN = nullptr; if (LIS) { OrigRegs.push_back(MI.getOperand(0).getReg()); for (unsigned i = 1, e = MI.getNumOperands(); i < e; i += 2) OrigRegs.push_back(MI.getOperand(i).getReg()); if (LIS->hasInterval(DstReg)) { DefVN = LIS->getInterval(DstReg) .Query(LIS->getInstructionIndex(MI)) .valueOut(); } } LaneBitmask UndefLanes = LaneBitmask::getNone(); bool DefEmitted = false; for (unsigned i = 1, e = MI.getNumOperands(); i < e; i += 2) { MachineOperand &UseMO = MI.getOperand(i); Register SrcReg = UseMO.getReg(); unsigned SubIdx = MI.getOperand(i+1).getImm(); // Nothing needs to be inserted for undef operands. if (UseMO.isUndef()) { UndefLanes |= TRI->getSubRegIndexLaneMask(SubIdx); continue; } // Defer any kill flag to the last operand using SrcReg. Otherwise, we // might insert a COPY that uses SrcReg after is was killed. bool isKill = UseMO.isKill(); if (isKill) for (unsigned j = i + 2; j < e; j += 2) if (MI.getOperand(j).getReg() == SrcReg) { MI.getOperand(j).setIsKill(); UseMO.setIsKill(false); isKill = false; break; } // Insert the sub-register copy. MachineInstr *CopyMI = BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), TII->get(TargetOpcode::COPY)) .addReg(DstReg, RegState::Define, SubIdx) .add(UseMO); // The first def needs an undef flag because there is no live register // before it. if (!DefEmitted) { CopyMI->getOperand(0).setIsUndef(true); // Return an iterator pointing to the first inserted instr. MBBI = CopyMI; } DefEmitted = true; // Update LiveVariables' kill info. if (LV && isKill && !SrcReg.isPhysical()) LV->replaceKillInstruction(SrcReg, MI, *CopyMI); LLVM_DEBUG(dbgs() << "Inserted: " << *CopyMI); } MachineBasicBlock::iterator EndMBBI = std::next(MachineBasicBlock::iterator(MI)); if (!DefEmitted) { LLVM_DEBUG(dbgs() << "Turned: " << MI << " into an IMPLICIT_DEF"); MI.setDesc(TII->get(TargetOpcode::IMPLICIT_DEF)); for (int j = MI.getNumOperands() - 1, ee = 0; j > ee; --j) MI.removeOperand(j); } else { if (LIS) { // Force live interval recomputation if we moved to a partial definition // of the register. Undef flags must be propagate to uses of undefined // subregister for accurate interval computation. if (UndefLanes.any() && DefVN && MRI->shouldTrackSubRegLiveness(DstReg)) { auto &LI = LIS->getInterval(DstReg); for (MachineOperand &UseOp : MRI->use_operands(DstReg)) { unsigned SubReg = UseOp.getSubReg(); if (UseOp.isUndef() || !SubReg) continue; auto *VN = LI.getVNInfoAt(LIS->getInstructionIndex(*UseOp.getParent())); if (DefVN != VN) continue; LaneBitmask LaneMask = TRI->getSubRegIndexLaneMask(SubReg); if ((UndefLanes & LaneMask).any()) UseOp.setIsUndef(true); } LIS->removeInterval(DstReg); } LIS->RemoveMachineInstrFromMaps(MI); } LLVM_DEBUG(dbgs() << "Eliminated: " << MI); MI.eraseFromParent(); } // Udpate LiveIntervals. if (LIS) LIS->repairIntervalsInRange(MBB, MBBI, EndMBBI, OrigRegs); }