//===-- llvm/CodeGen/Rewriter.cpp - Rewriter -----------------------------===// // // The LLVM Compiler Infrastructure // // This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// #define DEBUG_TYPE "virtregrewriter" #include "VirtRegRewriter.h" #include "VirtRegMap.h" #include "llvm/Function.h" #include "llvm/CodeGen/LiveIntervalAnalysis.h" #include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineInstrBuilder.h" #include "llvm/CodeGen/MachineRegisterInfo.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/TargetInstrInfo.h" #include "llvm/Target/TargetLowering.h" #include "llvm/ADT/DepthFirstIterator.h" #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/Statistic.h" using namespace llvm; STATISTIC(NumDSE , "Number of dead stores elided"); STATISTIC(NumDSS , "Number of dead spill slots removed"); STATISTIC(NumCommutes, "Number of instructions commuted"); STATISTIC(NumDRM , "Number of re-materializable defs elided"); STATISTIC(NumStores , "Number of stores added"); STATISTIC(NumPSpills , "Number of physical register spills"); STATISTIC(NumOmitted , "Number of reloads omitted"); STATISTIC(NumAvoided , "Number of reloads deemed unnecessary"); STATISTIC(NumCopified, "Number of available reloads turned into copies"); STATISTIC(NumReMats , "Number of re-materialization"); STATISTIC(NumLoads , "Number of loads added"); STATISTIC(NumReused , "Number of values reused"); STATISTIC(NumDCE , "Number of copies elided"); STATISTIC(NumSUnfold , "Number of stores unfolded"); STATISTIC(NumModRefUnfold, "Number of modref unfolded"); namespace { enum RewriterName { local, trivial }; } static cl::opt RewriterOpt("rewriter", cl::desc("Rewriter to use (default=local)"), cl::Prefix, cl::values(clEnumVal(local, "local rewriter"), clEnumVal(trivial, "trivial rewriter"), clEnumValEnd), cl::init(local)); static cl::opt ScheduleSpills("schedule-spills", cl::desc("Schedule spill code"), cl::init(false)); VirtRegRewriter::~VirtRegRewriter() {} /// substitutePhysReg - Replace virtual register in MachineOperand with a /// physical register. Do the right thing with the sub-register index. /// Note that operands may be added, so the MO reference is no longer valid. static void substitutePhysReg(MachineOperand &MO, unsigned Reg, const TargetRegisterInfo &TRI) { if (MO.getSubReg()) { MO.substPhysReg(Reg, TRI); // Any kill flags apply to the full virtual register, so they also apply to // the full physical register. // We assume that partial defs have already been decorated with a super-reg // operand by LiveIntervals. MachineInstr &MI = *MO.getParent(); if (MO.isUse() && !MO.isUndef() && (MO.isKill() || MI.isRegTiedToDefOperand(&MO-&MI.getOperand(0)))) MI.addRegisterKilled(Reg, &TRI, /*AddIfNotFound=*/ true); } else { MO.setReg(Reg); } } namespace { /// This class is intended for use with the new spilling framework only. It /// rewrites vreg def/uses to use the assigned preg, but does not insert any /// spill code. struct TrivialRewriter : public VirtRegRewriter { bool runOnMachineFunction(MachineFunction &MF, VirtRegMap &VRM, LiveIntervals* LIs) { DEBUG(dbgs() << "********** REWRITE MACHINE CODE **********\n"); DEBUG(dbgs() << "********** Function: " << MF.getFunction()->getName() << '\n'); DEBUG(dbgs() << "**** Machine Instrs" << "(NOTE! Does not include spills and reloads!) ****\n"); DEBUG(MF.dump()); MachineRegisterInfo *mri = &MF.getRegInfo(); const TargetRegisterInfo *tri = MF.getTarget().getRegisterInfo(); bool changed = false; for (LiveIntervals::iterator liItr = LIs->begin(), liEnd = LIs->end(); liItr != liEnd; ++liItr) { const LiveInterval *li = liItr->second; unsigned reg = li->reg; if (TargetRegisterInfo::isPhysicalRegister(reg)) { if (!li->empty()) mri->setPhysRegUsed(reg); } else { if (!VRM.hasPhys(reg)) continue; unsigned pReg = VRM.getPhys(reg); mri->setPhysRegUsed(pReg); // Copy the register use-list before traversing it. SmallVector, 32> reglist; for (MachineRegisterInfo::reg_iterator I = mri->reg_begin(reg), E = mri->reg_end(); I != E; ++I) reglist.push_back(std::make_pair(&*I, I.getOperandNo())); for (unsigned N=0; N != reglist.size(); ++N) substitutePhysReg(reglist[N].first->getOperand(reglist[N].second), pReg, *tri); changed |= !reglist.empty(); } } DEBUG(dbgs() << "**** Post Machine Instrs ****\n"); DEBUG(MF.dump()); return changed; } }; } // ************************************************************************ // namespace { /// AvailableSpills - As the local rewriter is scanning and rewriting an MBB /// from top down, keep track of which spill slots or remat are available in /// each register. /// /// Note that not all physregs are created equal here. In particular, some /// physregs are reloads that we are allowed to clobber or ignore at any time. /// Other physregs are values that the register allocated program is using /// that we cannot CHANGE, but we can read if we like. We keep track of this /// on a per-stack-slot / remat id basis as the low bit in the value of the /// SpillSlotsAvailable entries. The predicate 'canClobberPhysReg()' checks /// this bit and addAvailable sets it if. class AvailableSpills { const TargetRegisterInfo *TRI; const TargetInstrInfo *TII; // SpillSlotsOrReMatsAvailable - This map keeps track of all of the spilled // or remat'ed virtual register values that are still available, due to // being loaded or stored to, but not invalidated yet. std::map SpillSlotsOrReMatsAvailable; // PhysRegsAvailable - This is the inverse of SpillSlotsOrReMatsAvailable, // indicating which stack slot values are currently held by a physreg. This // is used to invalidate entries in SpillSlotsOrReMatsAvailable when a // physreg is modified. std::multimap PhysRegsAvailable; void disallowClobberPhysRegOnly(unsigned PhysReg); void ClobberPhysRegOnly(unsigned PhysReg); public: AvailableSpills(const TargetRegisterInfo *tri, const TargetInstrInfo *tii) : TRI(tri), TII(tii) { } /// clear - Reset the state. void clear() { SpillSlotsOrReMatsAvailable.clear(); PhysRegsAvailable.clear(); } const TargetRegisterInfo *getRegInfo() const { return TRI; } /// getSpillSlotOrReMatPhysReg - If the specified stack slot or remat is /// available in a physical register, return that PhysReg, otherwise /// return 0. unsigned getSpillSlotOrReMatPhysReg(int Slot) const { std::map::const_iterator I = SpillSlotsOrReMatsAvailable.find(Slot); if (I != SpillSlotsOrReMatsAvailable.end()) { return I->second >> 1; // Remove the CanClobber bit. } return 0; } /// addAvailable - Mark that the specified stack slot / remat is available /// in the specified physreg. If CanClobber is true, the physreg can be /// modified at any time without changing the semantics of the program. void addAvailable(int SlotOrReMat, unsigned Reg, bool CanClobber = true) { // If this stack slot is thought to be available in some other physreg, // remove its record. ModifyStackSlotOrReMat(SlotOrReMat); PhysRegsAvailable.insert(std::make_pair(Reg, SlotOrReMat)); SpillSlotsOrReMatsAvailable[SlotOrReMat]= (Reg << 1) | (unsigned)CanClobber; if (SlotOrReMat > VirtRegMap::MAX_STACK_SLOT) DEBUG(dbgs() << "Remembering RM#" << SlotOrReMat-VirtRegMap::MAX_STACK_SLOT-1); else DEBUG(dbgs() << "Remembering SS#" << SlotOrReMat); DEBUG(dbgs() << " in physreg " << TRI->getName(Reg) << (CanClobber ? " canclobber" : "") << "\n"); } /// canClobberPhysRegForSS - Return true if the spiller is allowed to change /// the value of the specified stackslot register if it desires. The /// specified stack slot must be available in a physreg for this query to /// make sense. bool canClobberPhysRegForSS(int SlotOrReMat) const { assert(SpillSlotsOrReMatsAvailable.count(SlotOrReMat) && "Value not available!"); return SpillSlotsOrReMatsAvailable.find(SlotOrReMat)->second & 1; } /// canClobberPhysReg - Return true if the spiller is allowed to clobber the /// physical register where values for some stack slot(s) might be /// available. bool canClobberPhysReg(unsigned PhysReg) const { std::multimap::const_iterator I = PhysRegsAvailable.lower_bound(PhysReg); while (I != PhysRegsAvailable.end() && I->first == PhysReg) { int SlotOrReMat = I->second; I++; if (!canClobberPhysRegForSS(SlotOrReMat)) return false; } return true; } /// disallowClobberPhysReg - Unset the CanClobber bit of the specified /// stackslot register. The register is still available but is no longer /// allowed to be modifed. void disallowClobberPhysReg(unsigned PhysReg); /// ClobberPhysReg - This is called when the specified physreg changes /// value. We use this to invalidate any info about stuff that lives in /// it and any of its aliases. void ClobberPhysReg(unsigned PhysReg); /// ModifyStackSlotOrReMat - This method is called when the value in a stack /// slot changes. This removes information about which register the /// previous value for this slot lives in (as the previous value is dead /// now). void ModifyStackSlotOrReMat(int SlotOrReMat); /// ClobberSharingStackSlots - When a register mapped to a stack slot changes, /// other stack slots sharing the same register are no longer valid. void ClobberSharingStackSlots(int StackSlot); /// AddAvailableRegsToLiveIn - Availability information is being kept coming /// into the specified MBB. Add available physical registers as potential /// live-in's. If they are reused in the MBB, they will be added to the /// live-in set to make register scavenger and post-allocation scheduler. void AddAvailableRegsToLiveIn(MachineBasicBlock &MBB, BitVector &RegKills, std::vector &KillOps); }; } // ************************************************************************ // // Given a location where a reload of a spilled register or a remat of // a constant is to be inserted, attempt to find a safe location to // insert the load at an earlier point in the basic-block, to hide // latency of the load and to avoid address-generation interlock // issues. static MachineBasicBlock::iterator ComputeReloadLoc(MachineBasicBlock::iterator const InsertLoc, MachineBasicBlock::iterator const Begin, unsigned PhysReg, const TargetRegisterInfo *TRI, bool DoReMat, int SSorRMId, const TargetInstrInfo *TII, const MachineFunction &MF) { if (!ScheduleSpills) return InsertLoc; // Spill backscheduling is of primary interest to addresses, so // don't do anything if the register isn't in the register class // used for pointers. const TargetLowering *TL = MF.getTarget().getTargetLowering(); if (!TL->isTypeLegal(TL->getPointerTy())) // Believe it or not, this is true on 16-bit targets like PIC16. return InsertLoc; const TargetRegisterClass *ptrRegClass = TL->getRegClassFor(TL->getPointerTy()); if (!ptrRegClass->contains(PhysReg)) return InsertLoc; // Scan upwards through the preceding instructions. If an instruction doesn't // reference the stack slot or the register we're loading, we can // backschedule the reload up past it. MachineBasicBlock::iterator NewInsertLoc = InsertLoc; while (NewInsertLoc != Begin) { MachineBasicBlock::iterator Prev = prior(NewInsertLoc); for (unsigned i = 0; i < Prev->getNumOperands(); ++i) { MachineOperand &Op = Prev->getOperand(i); if (!DoReMat && Op.isFI() && Op.getIndex() == SSorRMId) goto stop; } if (Prev->findRegisterUseOperandIdx(PhysReg) != -1 || Prev->findRegisterDefOperand(PhysReg)) goto stop; for (const unsigned *Alias = TRI->getAliasSet(PhysReg); *Alias; ++Alias) if (Prev->findRegisterUseOperandIdx(*Alias) != -1 || Prev->findRegisterDefOperand(*Alias)) goto stop; NewInsertLoc = Prev; } stop:; // If we made it to the beginning of the block, turn around and move back // down just past any existing reloads. They're likely to be reloads/remats // for instructions earlier than what our current reload/remat is for, so // they should be scheduled earlier. if (NewInsertLoc == Begin) { int FrameIdx; while (InsertLoc != NewInsertLoc && (TII->isLoadFromStackSlot(NewInsertLoc, FrameIdx) || TII->isTriviallyReMaterializable(NewInsertLoc))) ++NewInsertLoc; } return NewInsertLoc; } namespace { // ReusedOp - For each reused operand, we keep track of a bit of information, // in case we need to rollback upon processing a new operand. See comments // below. struct ReusedOp { // The MachineInstr operand that reused an available value. unsigned Operand; // StackSlotOrReMat - The spill slot or remat id of the value being reused. unsigned StackSlotOrReMat; // PhysRegReused - The physical register the value was available in. unsigned PhysRegReused; // AssignedPhysReg - The physreg that was assigned for use by the reload. unsigned AssignedPhysReg; // VirtReg - The virtual register itself. unsigned VirtReg; ReusedOp(unsigned o, unsigned ss, unsigned prr, unsigned apr, unsigned vreg) : Operand(o), StackSlotOrReMat(ss), PhysRegReused(prr), AssignedPhysReg(apr), VirtReg(vreg) {} }; /// ReuseInfo - This maintains a collection of ReuseOp's for each operand that /// is reused instead of reloaded. class ReuseInfo { MachineInstr &MI; std::vector Reuses; BitVector PhysRegsClobbered; public: ReuseInfo(MachineInstr &mi, const TargetRegisterInfo *tri) : MI(mi) { PhysRegsClobbered.resize(tri->getNumRegs()); } bool hasReuses() const { return !Reuses.empty(); } /// addReuse - If we choose to reuse a virtual register that is already /// available instead of reloading it, remember that we did so. void addReuse(unsigned OpNo, unsigned StackSlotOrReMat, unsigned PhysRegReused, unsigned AssignedPhysReg, unsigned VirtReg) { // If the reload is to the assigned register anyway, no undo will be // required. if (PhysRegReused == AssignedPhysReg) return; // Otherwise, remember this. Reuses.push_back(ReusedOp(OpNo, StackSlotOrReMat, PhysRegReused, AssignedPhysReg, VirtReg)); } void markClobbered(unsigned PhysReg) { PhysRegsClobbered.set(PhysReg); } bool isClobbered(unsigned PhysReg) const { return PhysRegsClobbered.test(PhysReg); } /// GetRegForReload - We are about to emit a reload into PhysReg. If there /// is some other operand that is using the specified register, either pick /// a new register to use, or evict the previous reload and use this reg. unsigned GetRegForReload(const TargetRegisterClass *RC, unsigned PhysReg, MachineFunction &MF, MachineInstr *MI, AvailableSpills &Spills, std::vector &MaybeDeadStores, SmallSet &Rejected, BitVector &RegKills, std::vector &KillOps, VirtRegMap &VRM); /// GetRegForReload - Helper for the above GetRegForReload(). Add a /// 'Rejected' set to remember which registers have been considered and /// rejected for the reload. This avoids infinite looping in case like /// this: /// t1 := op t2, t3 /// t2 <- assigned r0 for use by the reload but ended up reuse r1 /// t3 <- assigned r1 for use by the reload but ended up reuse r0 /// t1 <- desires r1 /// sees r1 is taken by t2, tries t2's reload register r0 /// sees r0 is taken by t3, tries t3's reload register r1 /// sees r1 is taken by t2, tries t2's reload register r0 ... unsigned GetRegForReload(unsigned VirtReg, unsigned PhysReg, MachineInstr *MI, AvailableSpills &Spills, std::vector &MaybeDeadStores, BitVector &RegKills, std::vector &KillOps, VirtRegMap &VRM) { SmallSet Rejected; MachineFunction &MF = *MI->getParent()->getParent(); const TargetRegisterClass* RC = MF.getRegInfo().getRegClass(VirtReg); return GetRegForReload(RC, PhysReg, MF, MI, Spills, MaybeDeadStores, Rejected, RegKills, KillOps, VRM); } }; } // ****************** // // Utility Functions // // ****************** // /// findSinglePredSuccessor - Return via reference a vector of machine basic /// blocks each of which is a successor of the specified BB and has no other /// predecessor. static void findSinglePredSuccessor(MachineBasicBlock *MBB, SmallVectorImpl &Succs){ for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), SE = MBB->succ_end(); SI != SE; ++SI) { MachineBasicBlock *SuccMBB = *SI; if (SuccMBB->pred_size() == 1) Succs.push_back(SuccMBB); } } /// ResurrectConfirmedKill - Helper for ResurrectKill. This register is killed /// but not re-defined and it's being reused. Remove the kill flag for the /// register and unset the kill's marker and last kill operand. static void ResurrectConfirmedKill(unsigned Reg, const TargetRegisterInfo* TRI, BitVector &RegKills, std::vector &KillOps) { DEBUG(dbgs() << "Resurrect " << TRI->getName(Reg) << "\n"); MachineOperand *KillOp = KillOps[Reg]; KillOp->setIsKill(false); // KillOps[Reg] might be a def of a super-register. unsigned KReg = KillOp->getReg(); if (!RegKills[KReg]) return; assert(KillOps[KReg]->getParent() == KillOp->getParent() && "invalid superreg kill flags"); KillOps[KReg] = NULL; RegKills.reset(KReg); // If it's a def of a super-register. Its other sub-regsters are no // longer killed as well. for (const unsigned *SR = TRI->getSubRegisters(KReg); *SR; ++SR) { DEBUG(dbgs() << " Resurrect subreg " << TRI->getName(*SR) << "\n"); assert(KillOps[*SR]->getParent() == KillOp->getParent() && "invalid subreg kill flags"); KillOps[*SR] = NULL; RegKills.reset(*SR); } } /// ResurrectKill - Invalidate kill info associated with a previous MI. An /// optimization may have decided that it's safe to reuse a previously killed /// register. If we fail to erase the invalid kill flags, then the register /// scavenger may later clobber the register used by this MI. Note that this /// must be done even if this MI is being deleted! Consider: /// /// USE $r1 (vreg1) /// ... /// $r1(vreg3) = COPY $r1 (vreg2) /// /// RegAlloc has smartly assigned all three vregs to the same physreg. Initially /// vreg1's only use is a kill. The rewriter doesn't know it should be live /// until it rewrites vreg2. At that points it sees that the copy is dead and /// deletes it. However, deleting the copy implicitly forwards liveness of $r1 /// (it's copy coalescing). We must resurrect $r1 by removing the kill flag at /// vreg1 before deleting the copy. static void ResurrectKill(MachineInstr &MI, unsigned Reg, const TargetRegisterInfo* TRI, BitVector &RegKills, std::vector &KillOps) { if (RegKills[Reg] && KillOps[Reg]->getParent() != &MI) { ResurrectConfirmedKill(Reg, TRI, RegKills, KillOps); return; } // No previous kill for this reg. Check for subreg kills as well. // d4 = // store d4, fi#0 // ... // = s8 // ... // = d4 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) { unsigned SReg = *SR; if (RegKills[SReg] && KillOps[SReg]->getParent() != &MI) ResurrectConfirmedKill(SReg, TRI, RegKills, KillOps); } } /// InvalidateKills - MI is going to be deleted. If any of its operands are /// marked kill, then invalidate the information. static void InvalidateKills(MachineInstr &MI, const TargetRegisterInfo* TRI, BitVector &RegKills, std::vector &KillOps, SmallVector *KillRegs = NULL) { for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { MachineOperand &MO = MI.getOperand(i); if (!MO.isReg() || !MO.isUse() || !MO.isKill() || MO.isUndef()) continue; unsigned Reg = MO.getReg(); if (TargetRegisterInfo::isVirtualRegister(Reg)) continue; if (KillRegs) KillRegs->push_back(Reg); assert(Reg < KillOps.size()); if (KillOps[Reg] == &MO) { // This operand was the kill, now no longer. KillOps[Reg] = NULL; RegKills.reset(Reg); for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) { if (RegKills[*SR]) { assert(KillOps[*SR] == &MO && "bad subreg kill flags"); KillOps[*SR] = NULL; RegKills.reset(*SR); } } } else { // This operand may have reused a previously killed reg. Keep it live in // case it continues to be used after erasing this instruction. ResurrectKill(MI, Reg, TRI, RegKills, KillOps); } } } /// InvalidateRegDef - If the def operand of the specified def MI is now dead /// (since its spill instruction is removed), mark it isDead. Also checks if /// the def MI has other definition operands that are not dead. Returns it by /// reference. static bool InvalidateRegDef(MachineBasicBlock::iterator I, MachineInstr &NewDef, unsigned Reg, bool &HasLiveDef, const TargetRegisterInfo *TRI) { // Due to remat, it's possible this reg isn't being reused. That is, // the def of this reg (by prev MI) is now dead. MachineInstr *DefMI = I; MachineOperand *DefOp = NULL; for (unsigned i = 0, e = DefMI->getNumOperands(); i != e; ++i) { MachineOperand &MO = DefMI->getOperand(i); if (!MO.isReg() || !MO.isDef() || !MO.isKill() || MO.isUndef()) continue; if (MO.getReg() == Reg) DefOp = &MO; else if (!MO.isDead()) HasLiveDef = true; } if (!DefOp) return false; bool FoundUse = false, Done = false; MachineBasicBlock::iterator E = &NewDef; ++I; ++E; for (; !Done && I != E; ++I) { MachineInstr *NMI = I; for (unsigned j = 0, ee = NMI->getNumOperands(); j != ee; ++j) { MachineOperand &MO = NMI->getOperand(j); if (!MO.isReg() || MO.getReg() == 0 || (MO.getReg() != Reg && !TRI->isSubRegister(Reg, MO.getReg()))) continue; if (MO.isUse()) FoundUse = true; Done = true; // Stop after scanning all the operands of this MI. } } if (!FoundUse) { // Def is dead! DefOp->setIsDead(); return true; } return false; } /// UpdateKills - Track and update kill info. If a MI reads a register that is /// marked kill, then it must be due to register reuse. Transfer the kill info /// over. static void UpdateKills(MachineInstr &MI, const TargetRegisterInfo* TRI, BitVector &RegKills, std::vector &KillOps) { // These do not affect kill info at all. if (MI.isDebugValue()) return; for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { MachineOperand &MO = MI.getOperand(i); if (!MO.isReg() || !MO.isUse() || MO.isUndef()) continue; unsigned Reg = MO.getReg(); if (Reg == 0) continue; // This operand may have reused a previously killed reg. Keep it live. ResurrectKill(MI, Reg, TRI, RegKills, KillOps); if (MO.isKill()) { RegKills.set(Reg); KillOps[Reg] = &MO; for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) { RegKills.set(*SR); KillOps[*SR] = &MO; } } } for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { const MachineOperand &MO = MI.getOperand(i); if (!MO.isReg() || !MO.getReg() || !MO.isDef()) continue; unsigned Reg = MO.getReg(); RegKills.reset(Reg); KillOps[Reg] = NULL; // It also defines (or partially define) aliases. for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) { RegKills.reset(*SR); KillOps[*SR] = NULL; } for (const unsigned *SR = TRI->getSuperRegisters(Reg); *SR; ++SR) { RegKills.reset(*SR); KillOps[*SR] = NULL; } } } /// ReMaterialize - Re-materialize definition for Reg targeting DestReg. /// static void ReMaterialize(MachineBasicBlock &MBB, MachineBasicBlock::iterator &MII, unsigned DestReg, unsigned Reg, const TargetInstrInfo *TII, const TargetRegisterInfo *TRI, VirtRegMap &VRM) { MachineInstr *ReMatDefMI = VRM.getReMaterializedMI(Reg); #ifndef NDEBUG const MCInstrDesc &MCID = ReMatDefMI->getDesc(); assert(MCID.getNumDefs() == 1 && "Don't know how to remat instructions that define > 1 values!"); #endif TII->reMaterialize(MBB, MII, DestReg, 0, ReMatDefMI, *TRI); MachineInstr *NewMI = prior(MII); for (unsigned i = 0, e = NewMI->getNumOperands(); i != e; ++i) { MachineOperand &MO = NewMI->getOperand(i); if (!MO.isReg() || MO.getReg() == 0) continue; unsigned VirtReg = MO.getReg(); if (TargetRegisterInfo::isPhysicalRegister(VirtReg)) continue; assert(MO.isUse()); unsigned Phys = VRM.getPhys(VirtReg); assert(Phys && "Virtual register is not assigned a register?"); substitutePhysReg(MO, Phys, *TRI); } ++NumReMats; } /// findSuperReg - Find the SubReg's super-register of given register class /// where its SubIdx sub-register is SubReg. static unsigned findSuperReg(const TargetRegisterClass *RC, unsigned SubReg, unsigned SubIdx, const TargetRegisterInfo *TRI) { for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end(); I != E; ++I) { unsigned Reg = *I; if (TRI->getSubReg(Reg, SubIdx) == SubReg) return Reg; } return 0; } // ******************************** // // Available Spills Implementation // // ******************************** // /// disallowClobberPhysRegOnly - Unset the CanClobber bit of the specified /// stackslot register. The register is still available but is no longer /// allowed to be modifed. void AvailableSpills::disallowClobberPhysRegOnly(unsigned PhysReg) { std::multimap::iterator I = PhysRegsAvailable.lower_bound(PhysReg); while (I != PhysRegsAvailable.end() && I->first == PhysReg) { int SlotOrReMat = I->second; I++; assert((SpillSlotsOrReMatsAvailable[SlotOrReMat] >> 1) == PhysReg && "Bidirectional map mismatch!"); SpillSlotsOrReMatsAvailable[SlotOrReMat] &= ~1; DEBUG(dbgs() << "PhysReg " << TRI->getName(PhysReg) << " copied, it is available for use but can no longer be modified\n"); } } /// disallowClobberPhysReg - Unset the CanClobber bit of the specified /// stackslot register and its aliases. The register and its aliases may /// still available but is no longer allowed to be modifed. void AvailableSpills::disallowClobberPhysReg(unsigned PhysReg) { for (const unsigned *AS = TRI->getAliasSet(PhysReg); *AS; ++AS) disallowClobberPhysRegOnly(*AS); disallowClobberPhysRegOnly(PhysReg); } /// ClobberPhysRegOnly - This is called when the specified physreg changes /// value. We use this to invalidate any info about stuff we thing lives in it. void AvailableSpills::ClobberPhysRegOnly(unsigned PhysReg) { std::multimap::iterator I = PhysRegsAvailable.lower_bound(PhysReg); while (I != PhysRegsAvailable.end() && I->first == PhysReg) { int SlotOrReMat = I->second; PhysRegsAvailable.erase(I++); assert((SpillSlotsOrReMatsAvailable[SlotOrReMat] >> 1) == PhysReg && "Bidirectional map mismatch!"); SpillSlotsOrReMatsAvailable.erase(SlotOrReMat); DEBUG(dbgs() << "PhysReg " << TRI->getName(PhysReg) << " clobbered, invalidating "); if (SlotOrReMat > VirtRegMap::MAX_STACK_SLOT) DEBUG(dbgs() << "RM#" << SlotOrReMat-VirtRegMap::MAX_STACK_SLOT-1 <<"\n"); else DEBUG(dbgs() << "SS#" << SlotOrReMat << "\n"); } } /// ClobberPhysReg - This is called when the specified physreg changes /// value. We use this to invalidate any info about stuff we thing lives in /// it and any of its aliases. void AvailableSpills::ClobberPhysReg(unsigned PhysReg) { for (const unsigned *AS = TRI->getAliasSet(PhysReg); *AS; ++AS) ClobberPhysRegOnly(*AS); ClobberPhysRegOnly(PhysReg); } /// AddAvailableRegsToLiveIn - Availability information is being kept coming /// into the specified MBB. Add available physical registers as potential /// live-in's. If they are reused in the MBB, they will be added to the /// live-in set to make register scavenger and post-allocation scheduler. void AvailableSpills::AddAvailableRegsToLiveIn(MachineBasicBlock &MBB, BitVector &RegKills, std::vector &KillOps) { std::set NotAvailable; for (std::multimap::iterator I = PhysRegsAvailable.begin(), E = PhysRegsAvailable.end(); I != E; ++I) { unsigned Reg = I->first; const TargetRegisterClass* RC = TRI->getMinimalPhysRegClass(Reg); // FIXME: A temporary workaround. We can't reuse available value if it's // not safe to move the def of the virtual register's class. e.g. // X86::RFP* register classes. Do not add it as a live-in. if (!TII->isSafeToMoveRegClassDefs(RC)) // This is no longer available. NotAvailable.insert(Reg); else { MBB.addLiveIn(Reg); if (RegKills[Reg]) ResurrectConfirmedKill(Reg, TRI, RegKills, KillOps); } // Skip over the same register. std::multimap::iterator NI = llvm::next(I); while (NI != E && NI->first == Reg) { ++I; ++NI; } } for (std::set::iterator I = NotAvailable.begin(), E = NotAvailable.end(); I != E; ++I) { ClobberPhysReg(*I); for (const unsigned *SubRegs = TRI->getSubRegisters(*I); *SubRegs; ++SubRegs) ClobberPhysReg(*SubRegs); } } /// ModifyStackSlotOrReMat - This method is called when the value in a stack /// slot changes. This removes information about which register the previous /// value for this slot lives in (as the previous value is dead now). void AvailableSpills::ModifyStackSlotOrReMat(int SlotOrReMat) { std::map::iterator It = SpillSlotsOrReMatsAvailable.find(SlotOrReMat); if (It == SpillSlotsOrReMatsAvailable.end()) return; unsigned Reg = It->second >> 1; SpillSlotsOrReMatsAvailable.erase(It); // This register may hold the value of multiple stack slots, only remove this // stack slot from the set of values the register contains. std::multimap::iterator I = PhysRegsAvailable.lower_bound(Reg); for (; ; ++I) { assert(I != PhysRegsAvailable.end() && I->first == Reg && "Map inverse broken!"); if (I->second == SlotOrReMat) break; } PhysRegsAvailable.erase(I); } void AvailableSpills::ClobberSharingStackSlots(int StackSlot) { std::map::iterator It = SpillSlotsOrReMatsAvailable.find(StackSlot); if (It == SpillSlotsOrReMatsAvailable.end()) return; unsigned Reg = It->second >> 1; // Erase entries in PhysRegsAvailable for other stack slots. std::multimap::iterator I = PhysRegsAvailable.lower_bound(Reg); while (I != PhysRegsAvailable.end() && I->first == Reg) { std::multimap::iterator NextI = llvm::next(I); if (I->second != StackSlot) { DEBUG(dbgs() << "Clobbered sharing SS#" << I->second << " in " << PrintReg(Reg, TRI) << '\n'); SpillSlotsOrReMatsAvailable.erase(I->second); PhysRegsAvailable.erase(I); } I = NextI; } } // ************************** // // Reuse Info Implementation // // ************************** // /// GetRegForReload - We are about to emit a reload into PhysReg. If there /// is some other operand that is using the specified register, either pick /// a new register to use, or evict the previous reload and use this reg. unsigned ReuseInfo::GetRegForReload(const TargetRegisterClass *RC, unsigned PhysReg, MachineFunction &MF, MachineInstr *MI, AvailableSpills &Spills, std::vector &MaybeDeadStores, SmallSet &Rejected, BitVector &RegKills, std::vector &KillOps, VirtRegMap &VRM) { const TargetInstrInfo* TII = MF.getTarget().getInstrInfo(); const TargetRegisterInfo *TRI = Spills.getRegInfo(); if (Reuses.empty()) return PhysReg; // This is most often empty. for (unsigned ro = 0, e = Reuses.size(); ro != e; ++ro) { ReusedOp &Op = Reuses[ro]; // If we find some other reuse that was supposed to use this register // exactly for its reload, we can change this reload to use ITS reload // register. That is, unless its reload register has already been // considered and subsequently rejected because it has also been reused // by another operand. if (Op.PhysRegReused == PhysReg && Rejected.count(Op.AssignedPhysReg) == 0 && RC->contains(Op.AssignedPhysReg)) { // Yup, use the reload register that we didn't use before. unsigned NewReg = Op.AssignedPhysReg; Rejected.insert(PhysReg); return GetRegForReload(RC, NewReg, MF, MI, Spills, MaybeDeadStores, Rejected, RegKills, KillOps, VRM); } else { // Otherwise, we might also have a problem if a previously reused // value aliases the new register. If so, codegen the previous reload // and use this one. unsigned PRRU = Op.PhysRegReused; if (TRI->regsOverlap(PRRU, PhysReg)) { // Okay, we found out that an alias of a reused register // was used. This isn't good because it means we have // to undo a previous reuse. MachineBasicBlock *MBB = MI->getParent(); const TargetRegisterClass *AliasRC = MBB->getParent()->getRegInfo().getRegClass(Op.VirtReg); // Copy Op out of the vector and remove it, we're going to insert an // explicit load for it. ReusedOp NewOp = Op; Reuses.erase(Reuses.begin()+ro); // MI may be using only a sub-register of PhysRegUsed. unsigned RealPhysRegUsed = MI->getOperand(NewOp.Operand).getReg(); unsigned SubIdx = 0; assert(TargetRegisterInfo::isPhysicalRegister(RealPhysRegUsed) && "A reuse cannot be a virtual register"); if (PRRU != RealPhysRegUsed) { // What was the sub-register index? SubIdx = TRI->getSubRegIndex(PRRU, RealPhysRegUsed); assert(SubIdx && "Operand physreg is not a sub-register of PhysRegUsed"); } // Ok, we're going to try to reload the assigned physreg into the // slot that we were supposed to in the first place. However, that // register could hold a reuse. Check to see if it conflicts or // would prefer us to use a different register. unsigned NewPhysReg = GetRegForReload(RC, NewOp.AssignedPhysReg, MF, MI, Spills, MaybeDeadStores, Rejected, RegKills, KillOps, VRM); bool DoReMat = NewOp.StackSlotOrReMat > VirtRegMap::MAX_STACK_SLOT; int SSorRMId = DoReMat ? VRM.getReMatId(NewOp.VirtReg) : (int) NewOp.StackSlotOrReMat; // Back-schedule reloads and remats. MachineBasicBlock::iterator InsertLoc = ComputeReloadLoc(MI, MBB->begin(), PhysReg, TRI, DoReMat, SSorRMId, TII, MF); if (DoReMat) { ReMaterialize(*MBB, InsertLoc, NewPhysReg, NewOp.VirtReg, TII, TRI, VRM); } else { TII->loadRegFromStackSlot(*MBB, InsertLoc, NewPhysReg, NewOp.StackSlotOrReMat, AliasRC, TRI); MachineInstr *LoadMI = prior(InsertLoc); VRM.addSpillSlotUse(NewOp.StackSlotOrReMat, LoadMI); // Any stores to this stack slot are not dead anymore. MaybeDeadStores[NewOp.StackSlotOrReMat] = NULL; ++NumLoads; } Spills.ClobberPhysReg(NewPhysReg); Spills.ClobberPhysReg(NewOp.PhysRegReused); unsigned RReg = SubIdx ? TRI->getSubReg(NewPhysReg, SubIdx) :NewPhysReg; MI->getOperand(NewOp.Operand).setReg(RReg); MI->getOperand(NewOp.Operand).setSubReg(0); Spills.addAvailable(NewOp.StackSlotOrReMat, NewPhysReg); UpdateKills(*prior(InsertLoc), TRI, RegKills, KillOps); DEBUG(dbgs() << '\t' << *prior(InsertLoc)); DEBUG(dbgs() << "Reuse undone!\n"); --NumReused; // Finally, PhysReg is now available, go ahead and use it. return PhysReg; } } } return PhysReg; } // ************************************************************************ // /// FoldsStackSlotModRef - Return true if the specified MI folds the specified /// stack slot mod/ref. It also checks if it's possible to unfold the /// instruction by having it define a specified physical register instead. static bool FoldsStackSlotModRef(MachineInstr &MI, int SS, unsigned PhysReg, const TargetInstrInfo *TII, const TargetRegisterInfo *TRI, VirtRegMap &VRM) { if (VRM.hasEmergencySpills(&MI) || VRM.isSpillPt(&MI)) return false; bool Found = false; VirtRegMap::MI2VirtMapTy::const_iterator I, End; for (tie(I, End) = VRM.getFoldedVirts(&MI); I != End; ++I) { unsigned VirtReg = I->second.first; VirtRegMap::ModRef MR = I->second.second; if (MR & VirtRegMap::isModRef) if (VRM.getStackSlot(VirtReg) == SS) { Found= TII->getOpcodeAfterMemoryUnfold(MI.getOpcode(), true, true) != 0; break; } } if (!Found) return false; // Does the instruction uses a register that overlaps the scratch register? for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { MachineOperand &MO = MI.getOperand(i); if (!MO.isReg() || MO.getReg() == 0) continue; unsigned Reg = MO.getReg(); if (TargetRegisterInfo::isVirtualRegister(Reg)) { if (!VRM.hasPhys(Reg)) continue; Reg = VRM.getPhys(Reg); } if (TRI->regsOverlap(PhysReg, Reg)) return false; } return true; } /// FindFreeRegister - Find a free register of a given register class by looking /// at (at most) the last two machine instructions. static unsigned FindFreeRegister(MachineBasicBlock::iterator MII, MachineBasicBlock &MBB, const TargetRegisterClass *RC, const TargetRegisterInfo *TRI, BitVector &AllocatableRegs) { BitVector Defs(TRI->getNumRegs()); BitVector Uses(TRI->getNumRegs()); SmallVector LocalUses; SmallVector Kills; // Take a look at 2 instructions at most. unsigned Count = 0; while (Count < 2) { if (MII == MBB.begin()) break; MachineInstr *PrevMI = prior(MII); MII = PrevMI; if (PrevMI->isDebugValue()) continue; // Skip over dbg_value instructions. ++Count; for (unsigned i = 0, e = PrevMI->getNumOperands(); i != e; ++i) { MachineOperand &MO = PrevMI->getOperand(i); if (!MO.isReg() || MO.getReg() == 0) continue; unsigned Reg = MO.getReg(); if (MO.isDef()) { Defs.set(Reg); for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) Defs.set(*AS); } else { LocalUses.push_back(Reg); if (MO.isKill() && AllocatableRegs[Reg]) Kills.push_back(Reg); } } for (unsigned i = 0, e = Kills.size(); i != e; ++i) { unsigned Kill = Kills[i]; if (!Defs[Kill] && !Uses[Kill] && RC->contains(Kill)) return Kill; } for (unsigned i = 0, e = LocalUses.size(); i != e; ++i) { unsigned Reg = LocalUses[i]; Uses.set(Reg); for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) Uses.set(*AS); } } return 0; } static void AssignPhysToVirtReg(MachineInstr *MI, unsigned VirtReg, unsigned PhysReg, const TargetRegisterInfo &TRI) { for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { MachineOperand &MO = MI->getOperand(i); if (MO.isReg() && MO.getReg() == VirtReg) substitutePhysReg(MO, PhysReg, TRI); } } namespace { struct RefSorter { bool operator()(const std::pair &A, const std::pair &B) { return A.second < B.second; } }; // ***************************** // // Local Spiller Implementation // // ***************************** // class LocalRewriter : public VirtRegRewriter { MachineRegisterInfo *MRI; const TargetRegisterInfo *TRI; const TargetInstrInfo *TII; VirtRegMap *VRM; LiveIntervals *LIs; BitVector AllocatableRegs; DenseMap DistanceMap; DenseMap > Slot2DbgValues; MachineBasicBlock *MBB; // Basic block currently being processed. public: bool runOnMachineFunction(MachineFunction &MF, VirtRegMap &VRM, LiveIntervals* LIs); private: void EraseInstr(MachineInstr *MI) { VRM->RemoveMachineInstrFromMaps(MI); LIs->RemoveMachineInstrFromMaps(MI); MI->eraseFromParent(); } bool OptimizeByUnfold2(unsigned VirtReg, int SS, MachineBasicBlock::iterator &MII, std::vector &MaybeDeadStores, AvailableSpills &Spills, BitVector &RegKills, std::vector &KillOps); bool OptimizeByUnfold(MachineBasicBlock::iterator &MII, std::vector &MaybeDeadStores, AvailableSpills &Spills, BitVector &RegKills, std::vector &KillOps); bool CommuteToFoldReload(MachineBasicBlock::iterator &MII, unsigned VirtReg, unsigned SrcReg, int SS, AvailableSpills &Spills, BitVector &RegKills, std::vector &KillOps, const TargetRegisterInfo *TRI); void SpillRegToStackSlot(MachineBasicBlock::iterator &MII, int Idx, unsigned PhysReg, int StackSlot, const TargetRegisterClass *RC, bool isAvailable, MachineInstr *&LastStore, AvailableSpills &Spills, SmallSet &ReMatDefs, BitVector &RegKills, std::vector &KillOps); void TransferDeadness(unsigned Reg, BitVector &RegKills, std::vector &KillOps); bool InsertEmergencySpills(MachineInstr *MI); bool InsertRestores(MachineInstr *MI, AvailableSpills &Spills, BitVector &RegKills, std::vector &KillOps); bool InsertSpills(MachineInstr *MI); void ProcessUses(MachineInstr &MI, AvailableSpills &Spills, std::vector &MaybeDeadStores, BitVector &RegKills, ReuseInfo &ReusedOperands, std::vector &KillOps); void RewriteMBB(LiveIntervals *LIs, AvailableSpills &Spills, BitVector &RegKills, std::vector &KillOps); }; } bool LocalRewriter::runOnMachineFunction(MachineFunction &MF, VirtRegMap &vrm, LiveIntervals* lis) { MRI = &MF.getRegInfo(); TRI = MF.getTarget().getRegisterInfo(); TII = MF.getTarget().getInstrInfo(); VRM = &vrm; LIs = lis; AllocatableRegs = TRI->getAllocatableSet(MF); DEBUG(dbgs() << "\n**** Local spiller rewriting function '" << MF.getFunction()->getName() << "':\n"); DEBUG(dbgs() << "**** Machine Instrs (NOTE! Does not include spills and" " reloads!) ****\n"); DEBUG(MF.print(dbgs(), LIs->getSlotIndexes())); // Spills - Keep track of which spilled values are available in physregs // so that we can choose to reuse the physregs instead of emitting // reloads. This is usually refreshed per basic block. AvailableSpills Spills(TRI, TII); // Keep track of kill information. BitVector RegKills(TRI->getNumRegs()); std::vector KillOps; KillOps.resize(TRI->getNumRegs(), NULL); // SingleEntrySuccs - Successor blocks which have a single predecessor. SmallVector SinglePredSuccs; SmallPtrSet EarlyVisited; // Traverse the basic blocks depth first. MachineBasicBlock *Entry = MF.begin(); SmallPtrSet Visited; for (df_ext_iterator > DFI = df_ext_begin(Entry, Visited), E = df_ext_end(Entry, Visited); DFI != E; ++DFI) { MBB = *DFI; if (!EarlyVisited.count(MBB)) RewriteMBB(LIs, Spills, RegKills, KillOps); // If this MBB is the only predecessor of a successor. Keep the // availability information and visit it next. do { // Keep visiting single predecessor successor as long as possible. SinglePredSuccs.clear(); findSinglePredSuccessor(MBB, SinglePredSuccs); if (SinglePredSuccs.empty()) MBB = 0; else { // FIXME: More than one successors, each of which has MBB has // the only predecessor. MBB = SinglePredSuccs[0]; if (!Visited.count(MBB) && EarlyVisited.insert(MBB)) { Spills.AddAvailableRegsToLiveIn(*MBB, RegKills, KillOps); RewriteMBB(LIs, Spills, RegKills, KillOps); } } } while (MBB); // Clear the availability info. Spills.clear(); } DEBUG(dbgs() << "**** Post Machine Instrs ****\n"); DEBUG(MF.print(dbgs(), LIs->getSlotIndexes())); // Mark unused spill slots. MachineFrameInfo *MFI = MF.getFrameInfo(); int SS = VRM->getLowSpillSlot(); if (SS != VirtRegMap::NO_STACK_SLOT) { for (int e = VRM->getHighSpillSlot(); SS <= e; ++SS) { SmallVector &DbgValues = Slot2DbgValues[SS]; if (!VRM->isSpillSlotUsed(SS)) { MFI->RemoveStackObject(SS); for (unsigned j = 0, ee = DbgValues.size(); j != ee; ++j) { MachineInstr *DVMI = DbgValues[j]; DEBUG(dbgs() << "Removing debug info referencing FI#" << SS << '\n'); EraseInstr(DVMI); } ++NumDSS; } DbgValues.clear(); } } Slot2DbgValues.clear(); return true; } /// OptimizeByUnfold2 - Unfold a series of load / store folding instructions if /// a scratch register is available. /// xorq %r12, %r13 /// addq %rax, -184(%rbp) /// addq %r13, -184(%rbp) /// ==> /// xorq %r12, %r13 /// movq -184(%rbp), %r12 /// addq %rax, %r12 /// addq %r13, %r12 /// movq %r12, -184(%rbp) bool LocalRewriter:: OptimizeByUnfold2(unsigned VirtReg, int SS, MachineBasicBlock::iterator &MII, std::vector &MaybeDeadStores, AvailableSpills &Spills, BitVector &RegKills, std::vector &KillOps) { MachineBasicBlock::iterator NextMII = llvm::next(MII); // Skip over dbg_value instructions. while (NextMII != MBB->end() && NextMII->isDebugValue()) NextMII = llvm::next(NextMII); if (NextMII == MBB->end()) return false; if (TII->getOpcodeAfterMemoryUnfold(MII->getOpcode(), true, true) == 0) return false; // Now let's see if the last couple of instructions happens to have freed up // a register. const TargetRegisterClass* RC = MRI->getRegClass(VirtReg); unsigned PhysReg = FindFreeRegister(MII, *MBB, RC, TRI, AllocatableRegs); if (!PhysReg) return false; MachineFunction &MF = *MBB->getParent(); TRI = MF.getTarget().getRegisterInfo(); MachineInstr &MI = *MII; if (!FoldsStackSlotModRef(MI, SS, PhysReg, TII, TRI, *VRM)) return false; // If the next instruction also folds the same SS modref and can be unfoled, // then it's worthwhile to issue a load from SS into the free register and // then unfold these instructions. if (!FoldsStackSlotModRef(*NextMII, SS, PhysReg, TII, TRI, *VRM)) return false; // Back-schedule reloads and remats. ComputeReloadLoc(MII, MBB->begin(), PhysReg, TRI, false, SS, TII, MF); // Load from SS to the spare physical register. TII->loadRegFromStackSlot(*MBB, MII, PhysReg, SS, RC, TRI); // This invalidates Phys. Spills.ClobberPhysReg(PhysReg); // Remember it's available. Spills.addAvailable(SS, PhysReg); MaybeDeadStores[SS] = NULL; // Unfold current MI. SmallVector NewMIs; if (!TII->unfoldMemoryOperand(MF, &MI, VirtReg, false, false, NewMIs)) llvm_unreachable("Unable unfold the load / store folding instruction!"); assert(NewMIs.size() == 1); AssignPhysToVirtReg(NewMIs[0], VirtReg, PhysReg, *TRI); VRM->transferRestorePts(&MI, NewMIs[0]); MII = MBB->insert(MII, NewMIs[0]); InvalidateKills(MI, TRI, RegKills, KillOps); EraseInstr(&MI); ++NumModRefUnfold; // Unfold next instructions that fold the same SS. do { MachineInstr &NextMI = *NextMII; NextMII = llvm::next(NextMII); NewMIs.clear(); if (!TII->unfoldMemoryOperand(MF, &NextMI, VirtReg, false, false, NewMIs)) llvm_unreachable("Unable unfold the load / store folding instruction!"); assert(NewMIs.size() == 1); AssignPhysToVirtReg(NewMIs[0], VirtReg, PhysReg, *TRI); VRM->transferRestorePts(&NextMI, NewMIs[0]); MBB->insert(NextMII, NewMIs[0]); InvalidateKills(NextMI, TRI, RegKills, KillOps); EraseInstr(&NextMI); ++NumModRefUnfold; // Skip over dbg_value instructions. while (NextMII != MBB->end() && NextMII->isDebugValue()) NextMII = llvm::next(NextMII); if (NextMII == MBB->end()) break; } while (FoldsStackSlotModRef(*NextMII, SS, PhysReg, TII, TRI, *VRM)); // Store the value back into SS. TII->storeRegToStackSlot(*MBB, NextMII, PhysReg, true, SS, RC, TRI); MachineInstr *StoreMI = prior(NextMII); VRM->addSpillSlotUse(SS, StoreMI); VRM->virtFolded(VirtReg, StoreMI, VirtRegMap::isMod); return true; } /// OptimizeByUnfold - Turn a store folding instruction into a load folding /// instruction. e.g. /// xorl %edi, %eax /// movl %eax, -32(%ebp) /// movl -36(%ebp), %eax /// orl %eax, -32(%ebp) /// ==> /// xorl %edi, %eax /// orl -36(%ebp), %eax /// mov %eax, -32(%ebp) /// This enables unfolding optimization for a subsequent instruction which will /// also eliminate the newly introduced store instruction. bool LocalRewriter:: OptimizeByUnfold(MachineBasicBlock::iterator &MII, std::vector &MaybeDeadStores, AvailableSpills &Spills, BitVector &RegKills, std::vector &KillOps) { MachineFunction &MF = *MBB->getParent(); MachineInstr &MI = *MII; unsigned UnfoldedOpc = 0; unsigned UnfoldPR = 0; unsigned UnfoldVR = 0; int FoldedSS = VirtRegMap::NO_STACK_SLOT; VirtRegMap::MI2VirtMapTy::const_iterator I, End; for (tie(I, End) = VRM->getFoldedVirts(&MI); I != End; ) { // Only transform a MI that folds a single register. if (UnfoldedOpc) return false; UnfoldVR = I->second.first; VirtRegMap::ModRef MR = I->second.second; // MI2VirtMap be can updated which invalidate the iterator. // Increment the iterator first. ++I; if (VRM->isAssignedReg(UnfoldVR)) continue; // If this reference is not a use, any previous store is now dead. // Otherwise, the store to this stack slot is not dead anymore. FoldedSS = VRM->getStackSlot(UnfoldVR); MachineInstr* DeadStore = MaybeDeadStores[FoldedSS]; if (DeadStore && (MR & VirtRegMap::isModRef)) { unsigned PhysReg = Spills.getSpillSlotOrReMatPhysReg(FoldedSS); if (!PhysReg || !DeadStore->readsRegister(PhysReg)) continue; UnfoldPR = PhysReg; UnfoldedOpc = TII->getOpcodeAfterMemoryUnfold(MI.getOpcode(), false, true); } } if (!UnfoldedOpc) { if (!UnfoldVR) return false; // Look for other unfolding opportunities. return OptimizeByUnfold2(UnfoldVR, FoldedSS, MII, MaybeDeadStores, Spills, RegKills, KillOps); } for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { MachineOperand &MO = MI.getOperand(i); if (!MO.isReg() || MO.getReg() == 0 || !MO.isUse()) continue; unsigned VirtReg = MO.getReg(); if (TargetRegisterInfo::isPhysicalRegister(VirtReg) || MO.getSubReg()) continue; if (VRM->isAssignedReg(VirtReg)) { unsigned PhysReg = VRM->getPhys(VirtReg); if (PhysReg && TRI->regsOverlap(PhysReg, UnfoldPR)) return false; } else if (VRM->isReMaterialized(VirtReg)) continue; int SS = VRM->getStackSlot(VirtReg); unsigned PhysReg = Spills.getSpillSlotOrReMatPhysReg(SS); if (PhysReg) { if (TRI->regsOverlap(PhysReg, UnfoldPR)) return false; continue; } if (VRM->hasPhys(VirtReg)) { PhysReg = VRM->getPhys(VirtReg); if (!TRI->regsOverlap(PhysReg, UnfoldPR)) continue; } // Ok, we'll need to reload the value into a register which makes // it impossible to perform the store unfolding optimization later. // Let's see if it is possible to fold the load if the store is // unfolded. This allows us to perform the store unfolding // optimization. SmallVector NewMIs; if (TII->unfoldMemoryOperand(MF, &MI, UnfoldVR, false, false, NewMIs)) { assert(NewMIs.size() == 1); MachineInstr *NewMI = NewMIs.back(); MBB->insert(MII, NewMI); NewMIs.clear(); int Idx = NewMI->findRegisterUseOperandIdx(VirtReg, false); assert(Idx != -1); SmallVector Ops; Ops.push_back(Idx); MachineInstr *FoldedMI = TII->foldMemoryOperand(NewMI, Ops, SS); NewMI->eraseFromParent(); if (FoldedMI) { VRM->addSpillSlotUse(SS, FoldedMI); if (!VRM->hasPhys(UnfoldVR)) VRM->assignVirt2Phys(UnfoldVR, UnfoldPR); VRM->virtFolded(VirtReg, FoldedMI, VirtRegMap::isRef); MII = FoldedMI; InvalidateKills(MI, TRI, RegKills, KillOps); EraseInstr(&MI); return true; } } } return false; } /// CommuteChangesDestination - We are looking for r0 = op r1, r2 and /// where SrcReg is r1 and it is tied to r0. Return true if after /// commuting this instruction it will be r0 = op r2, r1. static bool CommuteChangesDestination(MachineInstr *DefMI, const MCInstrDesc &MCID, unsigned SrcReg, const TargetInstrInfo *TII, unsigned &DstIdx) { if (MCID.getNumDefs() != 1 && MCID.getNumOperands() != 3) return false; if (!DefMI->getOperand(1).isReg() || DefMI->getOperand(1).getReg() != SrcReg) return false; unsigned DefIdx; if (!DefMI->isRegTiedToDefOperand(1, &DefIdx) || DefIdx != 0) return false; unsigned SrcIdx1, SrcIdx2; if (!TII->findCommutedOpIndices(DefMI, SrcIdx1, SrcIdx2)) return false; if (SrcIdx1 == 1 && SrcIdx2 == 2) { DstIdx = 2; return true; } return false; } /// CommuteToFoldReload - /// Look for /// r1 = load fi#1 /// r1 = op r1, r2 /// store r1, fi#1 /// /// If op is commutable and r2 is killed, then we can xform these to /// r2 = op r2, fi#1 /// store r2, fi#1 bool LocalRewriter:: CommuteToFoldReload(MachineBasicBlock::iterator &MII, unsigned VirtReg, unsigned SrcReg, int SS, AvailableSpills &Spills, BitVector &RegKills, std::vector &KillOps, const TargetRegisterInfo *TRI) { if (MII == MBB->begin() || !MII->killsRegister(SrcReg)) return false; MachineInstr &MI = *MII; MachineBasicBlock::iterator DefMII = prior(MII); MachineInstr *DefMI = DefMII; const MCInstrDesc &MCID = DefMI->getDesc(); unsigned NewDstIdx; if (DefMII != MBB->begin() && MCID.isCommutable() && CommuteChangesDestination(DefMI, MCID, SrcReg, TII, NewDstIdx)) { MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx); unsigned NewReg = NewDstMO.getReg(); if (!NewDstMO.isKill() || TRI->regsOverlap(NewReg, SrcReg)) return false; MachineInstr *ReloadMI = prior(DefMII); int FrameIdx; unsigned DestReg = TII->isLoadFromStackSlot(ReloadMI, FrameIdx); if (DestReg != SrcReg || FrameIdx != SS) return false; int UseIdx = DefMI->findRegisterUseOperandIdx(DestReg, false); if (UseIdx == -1) return false; unsigned DefIdx; if (!MI.isRegTiedToDefOperand(UseIdx, &DefIdx)) return false; assert(DefMI->getOperand(DefIdx).isReg() && DefMI->getOperand(DefIdx).getReg() == SrcReg); // Now commute def instruction. MachineInstr *CommutedMI = TII->commuteInstruction(DefMI, true); if (!CommutedMI) return false; MBB->insert(MII, CommutedMI); SmallVector Ops; Ops.push_back(NewDstIdx); MachineInstr *FoldedMI = TII->foldMemoryOperand(CommutedMI, Ops, SS); // Not needed since foldMemoryOperand returns new MI. CommutedMI->eraseFromParent(); if (!FoldedMI) return false; VRM->addSpillSlotUse(SS, FoldedMI); VRM->virtFolded(VirtReg, FoldedMI, VirtRegMap::isRef); // Insert new def MI and spill MI. const TargetRegisterClass* RC = MRI->getRegClass(VirtReg); TII->storeRegToStackSlot(*MBB, &MI, NewReg, true, SS, RC, TRI); MII = prior(MII); MachineInstr *StoreMI = MII; VRM->addSpillSlotUse(SS, StoreMI); VRM->virtFolded(VirtReg, StoreMI, VirtRegMap::isMod); MII = FoldedMI; // Update MII to backtrack. // Delete all 3 old instructions. InvalidateKills(*ReloadMI, TRI, RegKills, KillOps); EraseInstr(ReloadMI); InvalidateKills(*DefMI, TRI, RegKills, KillOps); EraseInstr(DefMI); InvalidateKills(MI, TRI, RegKills, KillOps); EraseInstr(&MI); // If NewReg was previously holding value of some SS, it's now clobbered. // This has to be done now because it's a physical register. When this // instruction is re-visited, it's ignored. Spills.ClobberPhysReg(NewReg); ++NumCommutes; return true; } return false; } /// SpillRegToStackSlot - Spill a register to a specified stack slot. Check if /// the last store to the same slot is now dead. If so, remove the last store. void LocalRewriter:: SpillRegToStackSlot(MachineBasicBlock::iterator &MII, int Idx, unsigned PhysReg, int StackSlot, const TargetRegisterClass *RC, bool isAvailable, MachineInstr *&LastStore, AvailableSpills &Spills, SmallSet &ReMatDefs, BitVector &RegKills, std::vector &KillOps) { MachineBasicBlock::iterator oldNextMII = llvm::next(MII); TII->storeRegToStackSlot(*MBB, llvm::next(MII), PhysReg, true, StackSlot, RC, TRI); MachineInstr *StoreMI = prior(oldNextMII); VRM->addSpillSlotUse(StackSlot, StoreMI); DEBUG(dbgs() << "Store:\t" << *StoreMI); // If there is a dead store to this stack slot, nuke it now. if (LastStore) { DEBUG(dbgs() << "Removed dead store:\t" << *LastStore); ++NumDSE; SmallVector KillRegs; InvalidateKills(*LastStore, TRI, RegKills, KillOps, &KillRegs); MachineBasicBlock::iterator PrevMII = LastStore; bool CheckDef = PrevMII != MBB->begin(); if (CheckDef) --PrevMII; EraseInstr(LastStore); if (CheckDef) { // Look at defs of killed registers on the store. Mark the defs // as dead since the store has been deleted and they aren't // being reused. for (unsigned j = 0, ee = KillRegs.size(); j != ee; ++j) { bool HasOtherDef = false; if (InvalidateRegDef(PrevMII, *MII, KillRegs[j], HasOtherDef, TRI)) { MachineInstr *DeadDef = PrevMII; if (ReMatDefs.count(DeadDef) && !HasOtherDef) { // FIXME: This assumes a remat def does not have side effects. EraseInstr(DeadDef); ++NumDRM; } } } } } // Allow for multi-instruction spill sequences, as on PPC Altivec. Presume // the last of multiple instructions is the actual store. LastStore = prior(oldNextMII); // If the stack slot value was previously available in some other // register, change it now. Otherwise, make the register available, // in PhysReg. Spills.ModifyStackSlotOrReMat(StackSlot); Spills.ClobberPhysReg(PhysReg); Spills.addAvailable(StackSlot, PhysReg, isAvailable); ++NumStores; } /// isSafeToDelete - Return true if this instruction doesn't produce any side /// effect and all of its defs are dead. static bool isSafeToDelete(MachineInstr &MI) { const MCInstrDesc &MCID = MI.getDesc(); if (MCID.mayLoad() || MCID.mayStore() || MCID.isTerminator() || MCID.isCall() || MCID.isBarrier() || MCID.isReturn() || MI.isLabel() || MI.isDebugValue() || MI.hasUnmodeledSideEffects()) return false; // Technically speaking inline asm without side effects and no defs can still // be deleted. But there is so much bad inline asm code out there, we should // let them be. if (MI.isInlineAsm()) return false; for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { MachineOperand &MO = MI.getOperand(i); if (!MO.isReg() || !MO.getReg()) continue; if (MO.isDef() && !MO.isDead()) return false; if (MO.isUse() && MO.isKill()) // FIXME: We can't remove kill markers or else the scavenger will assert. // An alternative is to add a ADD pseudo instruction to replace kill // markers. return false; } return true; } /// TransferDeadness - A identity copy definition is dead and it's being /// removed. Find the last def or use and mark it as dead / kill. void LocalRewriter:: TransferDeadness(unsigned Reg, BitVector &RegKills, std::vector &KillOps) { SmallPtrSet Seens; SmallVector,8> Refs; for (MachineRegisterInfo::reg_iterator RI = MRI->reg_begin(Reg), RE = MRI->reg_end(); RI != RE; ++RI) { MachineInstr *UDMI = &*RI; if (UDMI->isDebugValue() || UDMI->getParent() != MBB) continue; DenseMap::iterator DI = DistanceMap.find(UDMI); if (DI == DistanceMap.end()) continue; if (Seens.insert(UDMI)) Refs.push_back(std::make_pair(UDMI, DI->second)); } if (Refs.empty()) return; std::sort(Refs.begin(), Refs.end(), RefSorter()); while (!Refs.empty()) { MachineInstr *LastUDMI = Refs.back().first; Refs.pop_back(); MachineOperand *LastUD = NULL; for (unsigned i = 0, e = LastUDMI->getNumOperands(); i != e; ++i) { MachineOperand &MO = LastUDMI->getOperand(i); if (!MO.isReg() || MO.getReg() != Reg) continue; if (!LastUD || (LastUD->isUse() && MO.isDef())) LastUD = &MO; if (LastUDMI->isRegTiedToDefOperand(i)) break; } if (LastUD->isDef()) { // If the instruction has no side effect, delete it and propagate // backward further. Otherwise, mark is dead and we are done. if (!isSafeToDelete(*LastUDMI)) { LastUD->setIsDead(); break; } EraseInstr(LastUDMI); } else { LastUD->setIsKill(); RegKills.set(Reg); KillOps[Reg] = LastUD; break; } } } /// InsertEmergencySpills - Insert emergency spills before MI if requested by /// VRM. Return true if spills were inserted. bool LocalRewriter::InsertEmergencySpills(MachineInstr *MI) { if (!VRM->hasEmergencySpills(MI)) return false; MachineBasicBlock::iterator MII = MI; SmallSet UsedSS; std::vector &EmSpills = VRM->getEmergencySpills(MI); for (unsigned i = 0, e = EmSpills.size(); i != e; ++i) { unsigned PhysReg = EmSpills[i]; const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(PhysReg); assert(RC && "Unable to determine register class!"); int SS = VRM->getEmergencySpillSlot(RC); if (UsedSS.count(SS)) llvm_unreachable("Need to spill more than one physical registers!"); UsedSS.insert(SS); TII->storeRegToStackSlot(*MBB, MII, PhysReg, true, SS, RC, TRI); MachineInstr *StoreMI = prior(MII); VRM->addSpillSlotUse(SS, StoreMI); // Back-schedule reloads and remats. MachineBasicBlock::iterator InsertLoc = ComputeReloadLoc(llvm::next(MII), MBB->begin(), PhysReg, TRI, false, SS, TII, *MBB->getParent()); TII->loadRegFromStackSlot(*MBB, InsertLoc, PhysReg, SS, RC, TRI); MachineInstr *LoadMI = prior(InsertLoc); VRM->addSpillSlotUse(SS, LoadMI); ++NumPSpills; DistanceMap.insert(std::make_pair(LoadMI, DistanceMap.size())); } return true; } /// InsertRestores - Restore registers before MI is requested by VRM. Return /// true is any instructions were inserted. bool LocalRewriter::InsertRestores(MachineInstr *MI, AvailableSpills &Spills, BitVector &RegKills, std::vector &KillOps) { if (!VRM->isRestorePt(MI)) return false; MachineBasicBlock::iterator MII = MI; std::vector &RestoreRegs = VRM->getRestorePtRestores(MI); for (unsigned i = 0, e = RestoreRegs.size(); i != e; ++i) { unsigned VirtReg = RestoreRegs[e-i-1]; // Reverse order. if (!VRM->getPreSplitReg(VirtReg)) continue; // Split interval spilled again. unsigned Phys = VRM->getPhys(VirtReg); MRI->setPhysRegUsed(Phys); // Check if the value being restored if available. If so, it must be // from a predecessor BB that fallthrough into this BB. We do not // expect: // BB1: // r1 = load fi#1 // ... // = r1 // ... # r1 not clobbered // ... // = load fi#1 bool DoReMat = VRM->isReMaterialized(VirtReg); int SSorRMId = DoReMat ? VRM->getReMatId(VirtReg) : VRM->getStackSlot(VirtReg); unsigned InReg = Spills.getSpillSlotOrReMatPhysReg(SSorRMId); if (InReg == Phys) { // If the value is already available in the expected register, save // a reload / remat. if (SSorRMId) DEBUG(dbgs() << "Reusing RM#" << SSorRMId-VirtRegMap::MAX_STACK_SLOT-1); else DEBUG(dbgs() << "Reusing SS#" << SSorRMId); DEBUG(dbgs() << " from physreg " << TRI->getName(InReg) << " for " << PrintReg(VirtReg) <<" instead of reloading into physreg " << TRI->getName(Phys) << '\n'); // Reusing a physreg may resurrect it. But we expect ProcessUses to update // the kill flags for the current instruction after processing it. ++NumOmitted; continue; } else if (InReg && InReg != Phys) { if (SSorRMId) DEBUG(dbgs() << "Reusing RM#" << SSorRMId-VirtRegMap::MAX_STACK_SLOT-1); else DEBUG(dbgs() << "Reusing SS#" << SSorRMId); DEBUG(dbgs() << " from physreg " << TRI->getName(InReg) << " for " << PrintReg(VirtReg) <<" by copying it into physreg " << TRI->getName(Phys) << '\n'); // If the reloaded / remat value is available in another register, // copy it to the desired register. // Back-schedule reloads and remats. MachineBasicBlock::iterator InsertLoc = ComputeReloadLoc(MII, MBB->begin(), Phys, TRI, DoReMat, SSorRMId, TII, *MBB->getParent()); MachineInstr *CopyMI = BuildMI(*MBB, InsertLoc, MI->getDebugLoc(), TII->get(TargetOpcode::COPY), Phys) .addReg(InReg, RegState::Kill); // This invalidates Phys. Spills.ClobberPhysReg(Phys); // Remember it's available. Spills.addAvailable(SSorRMId, Phys); CopyMI->setAsmPrinterFlag(MachineInstr::ReloadReuse); UpdateKills(*CopyMI, TRI, RegKills, KillOps); DEBUG(dbgs() << '\t' << *CopyMI); ++NumCopified; continue; } // Back-schedule reloads and remats. MachineBasicBlock::iterator InsertLoc = ComputeReloadLoc(MII, MBB->begin(), Phys, TRI, DoReMat, SSorRMId, TII, *MBB->getParent()); if (VRM->isReMaterialized(VirtReg)) { ReMaterialize(*MBB, InsertLoc, Phys, VirtReg, TII, TRI, *VRM); } else { const TargetRegisterClass* RC = MRI->getRegClass(VirtReg); TII->loadRegFromStackSlot(*MBB, InsertLoc, Phys, SSorRMId, RC, TRI); MachineInstr *LoadMI = prior(InsertLoc); VRM->addSpillSlotUse(SSorRMId, LoadMI); ++NumLoads; DistanceMap.insert(std::make_pair(LoadMI, DistanceMap.size())); } // This invalidates Phys. Spills.ClobberPhysReg(Phys); // Remember it's available. Spills.addAvailable(SSorRMId, Phys); UpdateKills(*prior(InsertLoc), TRI, RegKills, KillOps); DEBUG(dbgs() << '\t' << *prior(MII)); } return true; } /// InsertSpills - Insert spills after MI if requested by VRM. Return /// true if spills were inserted. bool LocalRewriter::InsertSpills(MachineInstr *MI) { if (!VRM->isSpillPt(MI)) return false; MachineBasicBlock::iterator MII = MI; std::vector > &SpillRegs = VRM->getSpillPtSpills(MI); for (unsigned i = 0, e = SpillRegs.size(); i != e; ++i) { unsigned VirtReg = SpillRegs[i].first; bool isKill = SpillRegs[i].second; if (!VRM->getPreSplitReg(VirtReg)) continue; // Split interval spilled again. const TargetRegisterClass *RC = MRI->getRegClass(VirtReg); unsigned Phys = VRM->getPhys(VirtReg); int StackSlot = VRM->getStackSlot(VirtReg); MachineBasicBlock::iterator oldNextMII = llvm::next(MII); TII->storeRegToStackSlot(*MBB, llvm::next(MII), Phys, isKill, StackSlot, RC, TRI); MachineInstr *StoreMI = prior(oldNextMII); VRM->addSpillSlotUse(StackSlot, StoreMI); DEBUG(dbgs() << "Store:\t" << *StoreMI); VRM->virtFolded(VirtReg, StoreMI, VirtRegMap::isMod); } return true; } /// ProcessUses - Process all of MI's spilled operands and all available /// operands. void LocalRewriter::ProcessUses(MachineInstr &MI, AvailableSpills &Spills, std::vector &MaybeDeadStores, BitVector &RegKills, ReuseInfo &ReusedOperands, std::vector &KillOps) { // Clear kill info. SmallSet KilledMIRegs; SmallVector VirtUseOps; for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { MachineOperand &MO = MI.getOperand(i); if (!MO.isReg() || MO.getReg() == 0) continue; // Ignore non-register operands. unsigned VirtReg = MO.getReg(); if (TargetRegisterInfo::isPhysicalRegister(VirtReg)) { // Ignore physregs for spilling, but remember that it is used by this // function. MRI->setPhysRegUsed(VirtReg); continue; } // We want to process implicit virtual register uses first. if (MO.isImplicit()) // If the virtual register is implicitly defined, emit a implicit_def // before so scavenger knows it's "defined". // FIXME: This is a horrible hack done the by register allocator to // remat a definition with virtual register operand. VirtUseOps.insert(VirtUseOps.begin(), i); else VirtUseOps.push_back(i); // A partial def causes problems because the same operand both reads and // writes the register. This rewriter is designed to rewrite uses and defs // separately, so a partial def would already have been rewritten to a // physreg by the time we get to processing defs. // Add an implicit use operand to model the partial def. if (MO.isDef() && MO.getSubReg() && MI.readsVirtualRegister(VirtReg) && MI.findRegisterUseOperandIdx(VirtReg) == -1) { VirtUseOps.insert(VirtUseOps.begin(), MI.getNumOperands()); MI.addOperand(MachineOperand::CreateReg(VirtReg, false, // isDef true)); // isImplicit DEBUG(dbgs() << "Partial redef: " << MI); } } // Process all of the spilled uses and all non spilled reg references. SmallVector PotentialDeadStoreSlots; KilledMIRegs.clear(); for (unsigned j = 0, e = VirtUseOps.size(); j != e; ++j) { unsigned i = VirtUseOps[j]; unsigned VirtReg = MI.getOperand(i).getReg(); assert(TargetRegisterInfo::isVirtualRegister(VirtReg) && "Not a virtual register?"); unsigned SubIdx = MI.getOperand(i).getSubReg(); if (VRM->isAssignedReg(VirtReg)) { // This virtual register was assigned a physreg! unsigned Phys = VRM->getPhys(VirtReg); MRI->setPhysRegUsed(Phys); if (MI.getOperand(i).isDef()) ReusedOperands.markClobbered(Phys); substitutePhysReg(MI.getOperand(i), Phys, *TRI); if (VRM->isImplicitlyDefined(VirtReg)) // FIXME: Is this needed? BuildMI(*MBB, &MI, MI.getDebugLoc(), TII->get(TargetOpcode::IMPLICIT_DEF), Phys); continue; } // This virtual register is now known to be a spilled value. if (!MI.getOperand(i).isUse()) continue; // Handle defs in the loop below (handle use&def here though) bool AvoidReload = MI.getOperand(i).isUndef(); // Check if it is defined by an implicit def. It should not be spilled. // Note, this is for correctness reason. e.g. // 8 %reg1024 = IMPLICIT_DEF // 12 %reg1024 = INSERT_SUBREG %reg1024, %reg1025, 2 // The live range [12, 14) are not part of the r1024 live interval since // it's defined by an implicit def. It will not conflicts with live // interval of r1025. Now suppose both registers are spilled, you can // easily see a situation where both registers are reloaded before // the INSERT_SUBREG and both target registers that would overlap. bool DoReMat = VRM->isReMaterialized(VirtReg); int SSorRMId = DoReMat ? VRM->getReMatId(VirtReg) : VRM->getStackSlot(VirtReg); int ReuseSlot = SSorRMId; // Check to see if this stack slot is available. unsigned PhysReg = Spills.getSpillSlotOrReMatPhysReg(SSorRMId); // If this is a sub-register use, make sure the reuse register is in the // right register class. For example, for x86 not all of the 32-bit // registers have accessible sub-registers. // Similarly so for EXTRACT_SUBREG. Consider this: // EDI = op // MOV32_mr fi#1, EDI // ... // = EXTRACT_SUBREG fi#1 // fi#1 is available in EDI, but it cannot be reused because it's not in // the right register file. if (PhysReg && !AvoidReload && SubIdx) { const TargetRegisterClass* RC = MRI->getRegClass(VirtReg); if (!RC->contains(PhysReg)) PhysReg = 0; } if (PhysReg && !AvoidReload) { // This spilled operand might be part of a two-address operand. If this // is the case, then changing it will necessarily require changing the // def part of the instruction as well. However, in some cases, we // aren't allowed to modify the reused register. If none of these cases // apply, reuse it. bool CanReuse = true; bool isTied = MI.isRegTiedToDefOperand(i); if (isTied) { // Okay, we have a two address operand. We can reuse this physreg as // long as we are allowed to clobber the value and there isn't an // earlier def that has already clobbered the physreg. CanReuse = !ReusedOperands.isClobbered(PhysReg) && Spills.canClobberPhysReg(PhysReg); } // If this is an asm, and a PhysReg alias is used elsewhere as an // earlyclobber operand, we can't also use it as an input. if (MI.isInlineAsm()) { for (unsigned k = 0, e = MI.getNumOperands(); k != e; ++k) { MachineOperand &MOk = MI.getOperand(k); if (MOk.isReg() && MOk.isEarlyClobber() && TRI->regsOverlap(MOk.getReg(), PhysReg)) { CanReuse = false; DEBUG(dbgs() << "Not reusing physreg " << TRI->getName(PhysReg) << " for " << PrintReg(VirtReg) << ": " << MOk << '\n'); break; } } } if (CanReuse) { // If this stack slot value is already available, reuse it! if (ReuseSlot > VirtRegMap::MAX_STACK_SLOT) DEBUG(dbgs() << "Reusing RM#" << ReuseSlot-VirtRegMap::MAX_STACK_SLOT-1); else DEBUG(dbgs() << "Reusing SS#" << ReuseSlot); DEBUG(dbgs() << " from physreg " << TRI->getName(PhysReg) << " for " << PrintReg(VirtReg) << " instead of reloading into " << PrintReg(VRM->getPhys(VirtReg), TRI) << '\n'); unsigned RReg = SubIdx ? TRI->getSubReg(PhysReg, SubIdx) : PhysReg; MI.getOperand(i).setReg(RReg); MI.getOperand(i).setSubReg(0); // Reusing a physreg may resurrect it. But we expect ProcessUses to // update the kill flags for the current instr after processing it. // The only technical detail we have is that we don't know that // PhysReg won't be clobbered by a reloaded stack slot that occurs // later in the instruction. In particular, consider 'op V1, V2'. // If V1 is available in physreg R0, we would choose to reuse it // here, instead of reloading it into the register the allocator // indicated (say R1). However, V2 might have to be reloaded // later, and it might indicate that it needs to live in R0. When // this occurs, we need to have information available that // indicates it is safe to use R1 for the reload instead of R0. // // To further complicate matters, we might conflict with an alias, // or R0 and R1 might not be compatible with each other. In this // case, we actually insert a reload for V1 in R1, ensuring that // we can get at R0 or its alias. ReusedOperands.addReuse(i, ReuseSlot, PhysReg, VRM->getPhys(VirtReg), VirtReg); if (isTied) // Only mark it clobbered if this is a use&def operand. ReusedOperands.markClobbered(PhysReg); ++NumReused; if (MI.getOperand(i).isKill() && ReuseSlot <= VirtRegMap::MAX_STACK_SLOT) { // The store of this spilled value is potentially dead, but we // won't know for certain until we've confirmed that the re-use // above is valid, which means waiting until the other operands // are processed. For now we just track the spill slot, we'll // remove it after the other operands are processed if valid. PotentialDeadStoreSlots.push_back(ReuseSlot); } // Mark is isKill if it's there no other uses of the same virtual // register and it's not a two-address operand. IsKill will be // unset if reg is reused. if (!isTied && KilledMIRegs.count(VirtReg) == 0) { MI.getOperand(i).setIsKill(); KilledMIRegs.insert(VirtReg); } continue; } // CanReuse // Otherwise we have a situation where we have a two-address instruction // whose mod/ref operand needs to be reloaded. This reload is already // available in some register "PhysReg", but if we used PhysReg as the // operand to our 2-addr instruction, the instruction would modify // PhysReg. This isn't cool if something later uses PhysReg and expects // to get its initial value. // // To avoid this problem, and to avoid doing a load right after a store, // we emit a copy from PhysReg into the designated register for this // operand. // // This case also applies to an earlyclobber'd PhysReg. unsigned DesignatedReg = VRM->getPhys(VirtReg); assert(DesignatedReg && "Must map virtreg to physreg!"); // Note that, if we reused a register for a previous operand, the // register we want to reload into might not actually be // available. If this occurs, use the register indicated by the // reuser. if (ReusedOperands.hasReuses()) DesignatedReg = ReusedOperands. GetRegForReload(VirtReg, DesignatedReg, &MI, Spills, MaybeDeadStores, RegKills, KillOps, *VRM); // If the mapped designated register is actually the physreg we have // incoming, we don't need to inserted a dead copy. if (DesignatedReg == PhysReg) { // If this stack slot value is already available, reuse it! if (ReuseSlot > VirtRegMap::MAX_STACK_SLOT) DEBUG(dbgs() << "Reusing RM#" << ReuseSlot-VirtRegMap::MAX_STACK_SLOT-1); else DEBUG(dbgs() << "Reusing SS#" << ReuseSlot); DEBUG(dbgs() << " from physreg " << TRI->getName(PhysReg) << " for " << PrintReg(VirtReg) << " instead of reloading into same physreg.\n"); unsigned RReg = SubIdx ? TRI->getSubReg(PhysReg, SubIdx) : PhysReg; MI.getOperand(i).setReg(RReg); MI.getOperand(i).setSubReg(0); ReusedOperands.markClobbered(RReg); ++NumReused; continue; } MRI->setPhysRegUsed(DesignatedReg); ReusedOperands.markClobbered(DesignatedReg); // Back-schedule reloads and remats. MachineBasicBlock::iterator InsertLoc = ComputeReloadLoc(&MI, MBB->begin(), PhysReg, TRI, DoReMat, SSorRMId, TII, *MBB->getParent()); MachineInstr *CopyMI = BuildMI(*MBB, InsertLoc, MI.getDebugLoc(), TII->get(TargetOpcode::COPY), DesignatedReg).addReg(PhysReg); CopyMI->setAsmPrinterFlag(MachineInstr::ReloadReuse); UpdateKills(*CopyMI, TRI, RegKills, KillOps); // This invalidates DesignatedReg. Spills.ClobberPhysReg(DesignatedReg); Spills.addAvailable(ReuseSlot, DesignatedReg); unsigned RReg = SubIdx ? TRI->getSubReg(DesignatedReg, SubIdx) : DesignatedReg; MI.getOperand(i).setReg(RReg); MI.getOperand(i).setSubReg(0); DEBUG(dbgs() << '\t' << *prior(InsertLoc)); ++NumReused; continue; } // if (PhysReg) // Otherwise, reload it and remember that we have it. PhysReg = VRM->getPhys(VirtReg); assert(PhysReg && "Must map virtreg to physreg!"); // Note that, if we reused a register for a previous operand, the // register we want to reload into might not actually be // available. If this occurs, use the register indicated by the // reuser. if (ReusedOperands.hasReuses()) PhysReg = ReusedOperands.GetRegForReload(VirtReg, PhysReg, &MI, Spills, MaybeDeadStores, RegKills, KillOps, *VRM); MRI->setPhysRegUsed(PhysReg); ReusedOperands.markClobbered(PhysReg); if (AvoidReload) ++NumAvoided; else { // Back-schedule reloads and remats. MachineBasicBlock::iterator InsertLoc = ComputeReloadLoc(MI, MBB->begin(), PhysReg, TRI, DoReMat, SSorRMId, TII, *MBB->getParent()); if (DoReMat) { ReMaterialize(*MBB, InsertLoc, PhysReg, VirtReg, TII, TRI, *VRM); } else { const TargetRegisterClass* RC = MRI->getRegClass(VirtReg); TII->loadRegFromStackSlot(*MBB, InsertLoc, PhysReg, SSorRMId, RC,TRI); MachineInstr *LoadMI = prior(InsertLoc); VRM->addSpillSlotUse(SSorRMId, LoadMI); ++NumLoads; DistanceMap.insert(std::make_pair(LoadMI, DistanceMap.size())); } // This invalidates PhysReg. Spills.ClobberPhysReg(PhysReg); // Any stores to this stack slot are not dead anymore. if (!DoReMat) MaybeDeadStores[SSorRMId] = NULL; Spills.addAvailable(SSorRMId, PhysReg); // Assumes this is the last use. IsKill will be unset if reg is reused // unless it's a two-address operand. if (!MI.isRegTiedToDefOperand(i) && KilledMIRegs.count(VirtReg) == 0) { MI.getOperand(i).setIsKill(); KilledMIRegs.insert(VirtReg); } UpdateKills(*prior(InsertLoc), TRI, RegKills, KillOps); DEBUG(dbgs() << '\t' << *prior(InsertLoc)); } unsigned RReg = SubIdx ? TRI->getSubReg(PhysReg, SubIdx) : PhysReg; MI.getOperand(i).setReg(RReg); MI.getOperand(i).setSubReg(0); } // Ok - now we can remove stores that have been confirmed dead. for (unsigned j = 0, e = PotentialDeadStoreSlots.size(); j != e; ++j) { // This was the last use and the spilled value is still available // for reuse. That means the spill was unnecessary! int PDSSlot = PotentialDeadStoreSlots[j]; MachineInstr* DeadStore = MaybeDeadStores[PDSSlot]; if (DeadStore) { DEBUG(dbgs() << "Removed dead store:\t" << *DeadStore); InvalidateKills(*DeadStore, TRI, RegKills, KillOps); EraseInstr(DeadStore); MaybeDeadStores[PDSSlot] = NULL; ++NumDSE; } } } /// rewriteMBB - Keep track of which spills are available even after the /// register allocator is done with them. If possible, avoid reloading vregs. void LocalRewriter::RewriteMBB(LiveIntervals *LIs, AvailableSpills &Spills, BitVector &RegKills, std::vector &KillOps) { DEBUG(dbgs() << "\n**** Local spiller rewriting MBB '" << MBB->getName() << "':\n"); MachineFunction &MF = *MBB->getParent(); // MaybeDeadStores - When we need to write a value back into a stack slot, // keep track of the inserted store. If the stack slot value is never read // (because the value was used from some available register, for example), and // subsequently stored to, the original store is dead. This map keeps track // of inserted stores that are not used. If we see a subsequent store to the // same stack slot, the original store is deleted. std::vector MaybeDeadStores; MaybeDeadStores.resize(MF.getFrameInfo()->getObjectIndexEnd(), NULL); // ReMatDefs - These are rematerializable def MIs which are not deleted. SmallSet ReMatDefs; // Keep track of the registers we have already spilled in case there are // multiple defs of the same register in MI. SmallSet SpilledMIRegs; RegKills.reset(); KillOps.clear(); KillOps.resize(TRI->getNumRegs(), NULL); DistanceMap.clear(); for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end(); MII != E; ) { MachineBasicBlock::iterator NextMII = llvm::next(MII); if (OptimizeByUnfold(MII, MaybeDeadStores, Spills, RegKills, KillOps)) NextMII = llvm::next(MII); if (InsertEmergencySpills(MII)) NextMII = llvm::next(MII); InsertRestores(MII, Spills, RegKills, KillOps); if (InsertSpills(MII)) NextMII = llvm::next(MII); bool Erased = false; bool BackTracked = false; MachineInstr &MI = *MII; // Remember DbgValue's which reference stack slots. if (MI.isDebugValue() && MI.getOperand(0).isFI()) Slot2DbgValues[MI.getOperand(0).getIndex()].push_back(&MI); /// ReusedOperands - Keep track of operand reuse in case we need to undo /// reuse. ReuseInfo ReusedOperands(MI, TRI); ProcessUses(MI, Spills, MaybeDeadStores, RegKills, ReusedOperands, KillOps); DEBUG(dbgs() << '\t' << MI); // If we have folded references to memory operands, make sure we clear all // physical registers that may contain the value of the spilled virtual // register // Copy the folded virts to a small vector, we may change MI2VirtMap. SmallVector, 4> FoldedVirts; // C++0x FTW! for (std::pair FVRange = VRM->getFoldedVirts(&MI); FVRange.first != FVRange.second; ++FVRange.first) FoldedVirts.push_back(FVRange.first->second); SmallSet FoldedSS; for (unsigned FVI = 0, FVE = FoldedVirts.size(); FVI != FVE; ++FVI) { unsigned VirtReg = FoldedVirts[FVI].first; VirtRegMap::ModRef MR = FoldedVirts[FVI].second; DEBUG(dbgs() << "Folded " << PrintReg(VirtReg) << " MR: " << MR); int SS = VRM->getStackSlot(VirtReg); if (SS == VirtRegMap::NO_STACK_SLOT) continue; FoldedSS.insert(SS); DEBUG(dbgs() << " - StackSlot: " << SS << "\n"); // If this folded instruction is just a use, check to see if it's a // straight load from the virt reg slot. if ((MR & VirtRegMap::isRef) && !(MR & VirtRegMap::isMod)) { int FrameIdx; unsigned DestReg = TII->isLoadFromStackSlot(&MI, FrameIdx); if (DestReg && FrameIdx == SS) { // If this spill slot is available, turn it into a copy (or nothing) // instead of leaving it as a load! if (unsigned InReg = Spills.getSpillSlotOrReMatPhysReg(SS)) { DEBUG(dbgs() << "Promoted Load To Copy: " << MI); if (DestReg != InReg) { MachineOperand *DefMO = MI.findRegisterDefOperand(DestReg); MachineInstr *CopyMI = BuildMI(*MBB, &MI, MI.getDebugLoc(), TII->get(TargetOpcode::COPY)) .addReg(DestReg, RegState::Define, DefMO->getSubReg()) .addReg(InReg, RegState::Kill); // Revisit the copy so we make sure to notice the effects of the // operation on the destreg (either needing to RA it if it's // virtual or needing to clobber any values if it's physical). NextMII = CopyMI; NextMII->setAsmPrinterFlag(MachineInstr::ReloadReuse); BackTracked = true; } else { DEBUG(dbgs() << "Removing now-noop copy: " << MI); // InvalidateKills resurrects any prior kill of the copy's source // allowing the source reg to be reused in place of the copy. Spills.disallowClobberPhysReg(InReg); } InvalidateKills(MI, TRI, RegKills, KillOps); EraseInstr(&MI); Erased = true; goto ProcessNextInst; } } else { unsigned PhysReg = Spills.getSpillSlotOrReMatPhysReg(SS); SmallVector NewMIs; if (PhysReg && TII->unfoldMemoryOperand(MF, &MI, PhysReg, false, false, NewMIs)){ MBB->insert(MII, NewMIs[0]); InvalidateKills(MI, TRI, RegKills, KillOps); EraseInstr(&MI); Erased = true; --NextMII; // backtrack to the unfolded instruction. BackTracked = true; goto ProcessNextInst; } } } // If this reference is not a use, any previous store is now dead. // Otherwise, the store to this stack slot is not dead anymore. MachineInstr* DeadStore = MaybeDeadStores[SS]; if (DeadStore) { bool isDead = !(MR & VirtRegMap::isRef); MachineInstr *NewStore = NULL; if (MR & VirtRegMap::isModRef) { unsigned PhysReg = Spills.getSpillSlotOrReMatPhysReg(SS); SmallVector NewMIs; // We can reuse this physreg as long as we are allowed to clobber // the value and there isn't an earlier def that has already clobbered // the physreg. if (PhysReg && !ReusedOperands.isClobbered(PhysReg) && Spills.canClobberPhysReg(PhysReg) && !TII->isStoreToStackSlot(&MI, SS)) { // Not profitable! MachineOperand *KillOpnd = DeadStore->findRegisterUseOperand(PhysReg, true); // Note, if the store is storing a sub-register, it's possible the // super-register is needed below. if (KillOpnd && !KillOpnd->getSubReg() && TII->unfoldMemoryOperand(MF, &MI, PhysReg, false, true,NewMIs)){ MBB->insert(MII, NewMIs[0]); NewStore = NewMIs[1]; MBB->insert(MII, NewStore); VRM->addSpillSlotUse(SS, NewStore); InvalidateKills(MI, TRI, RegKills, KillOps); EraseInstr(&MI); Erased = true; --NextMII; --NextMII; // backtrack to the unfolded instruction. BackTracked = true; isDead = true; ++NumSUnfold; } } } if (isDead) { // Previous store is dead. // If we get here, the store is dead, nuke it now. DEBUG(dbgs() << "Removed dead store:\t" << *DeadStore); InvalidateKills(*DeadStore, TRI, RegKills, KillOps); EraseInstr(DeadStore); if (!NewStore) ++NumDSE; } MaybeDeadStores[SS] = NULL; if (NewStore) { // Treat this store as a spill merged into a copy. That makes the // stack slot value available. VRM->virtFolded(VirtReg, NewStore, VirtRegMap::isMod); goto ProcessNextInst; } } // If the spill slot value is available, and this is a new definition of // the value, the value is not available anymore. if (MR & VirtRegMap::isMod) { // Notice that the value in this stack slot has been modified. Spills.ModifyStackSlotOrReMat(SS); // If this is *just* a mod of the value, check to see if this is just a // store to the spill slot (i.e. the spill got merged into the copy). If // so, realize that the vreg is available now, and add the store to the // MaybeDeadStore info. int StackSlot; if (!(MR & VirtRegMap::isRef)) { if (unsigned SrcReg = TII->isStoreToStackSlot(&MI, StackSlot)) { assert(TargetRegisterInfo::isPhysicalRegister(SrcReg) && "Src hasn't been allocated yet?"); if (CommuteToFoldReload(MII, VirtReg, SrcReg, StackSlot, Spills, RegKills, KillOps, TRI)) { NextMII = llvm::next(MII); BackTracked = true; goto ProcessNextInst; } // Okay, this is certainly a store of SrcReg to [StackSlot]. Mark // this as a potentially dead store in case there is a subsequent // store into the stack slot without a read from it. MaybeDeadStores[StackSlot] = &MI; // If the stack slot value was previously available in some other // register, change it now. Otherwise, make the register // available in PhysReg. Spills.addAvailable(StackSlot, SrcReg, MI.killsRegister(SrcReg)); } } } } // Process all of the spilled defs. SpilledMIRegs.clear(); for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { MachineOperand &MO = MI.getOperand(i); if (!(MO.isReg() && MO.getReg() && MO.isDef())) continue; unsigned VirtReg = MO.getReg(); if (!TargetRegisterInfo::isVirtualRegister(VirtReg)) { // Check to see if this is a noop copy. If so, eliminate the // instruction before considering the dest reg to be changed. // Also check if it's copying from an "undef", if so, we can't // eliminate this or else the undef marker is lost and it will // confuses the scavenger. This is extremely rare. if (MI.isIdentityCopy() && !MI.getOperand(1).isUndef() && MI.getNumOperands() == 2) { ++NumDCE; DEBUG(dbgs() << "Removing now-noop copy: " << MI); SmallVector KillRegs; InvalidateKills(MI, TRI, RegKills, KillOps, &KillRegs); if (MO.isDead() && !KillRegs.empty()) { // Source register or an implicit super/sub-register use is killed. assert(TRI->regsOverlap(KillRegs[0], MI.getOperand(0).getReg())); // Last def is now dead. TransferDeadness(MI.getOperand(1).getReg(), RegKills, KillOps); } EraseInstr(&MI); Erased = true; Spills.disallowClobberPhysReg(VirtReg); goto ProcessNextInst; } // If it's not a no-op copy, it clobbers the value in the destreg. Spills.ClobberPhysReg(VirtReg); ReusedOperands.markClobbered(VirtReg); // Check to see if this instruction is a load from a stack slot into // a register. If so, this provides the stack slot value in the reg. int FrameIdx; if (unsigned DestReg = TII->isLoadFromStackSlot(&MI, FrameIdx)) { assert(DestReg == VirtReg && "Unknown load situation!"); // If it is a folded reference, then it's not safe to clobber. bool Folded = FoldedSS.count(FrameIdx); // Otherwise, if it wasn't available, remember that it is now! Spills.addAvailable(FrameIdx, DestReg, !Folded); goto ProcessNextInst; } continue; } unsigned SubIdx = MO.getSubReg(); bool DoReMat = VRM->isReMaterialized(VirtReg); if (DoReMat) ReMatDefs.insert(&MI); // The only vregs left are stack slot definitions. int StackSlot = VRM->getStackSlot(VirtReg); const TargetRegisterClass *RC = MRI->getRegClass(VirtReg); // If this def is part of a two-address operand, make sure to execute // the store from the correct physical register. unsigned PhysReg; unsigned TiedOp; if (MI.isRegTiedToUseOperand(i, &TiedOp)) { PhysReg = MI.getOperand(TiedOp).getReg(); if (SubIdx) { unsigned SuperReg = findSuperReg(RC, PhysReg, SubIdx, TRI); assert(SuperReg && TRI->getSubReg(SuperReg, SubIdx) == PhysReg && "Can't find corresponding super-register!"); PhysReg = SuperReg; } } else { PhysReg = VRM->getPhys(VirtReg); if (ReusedOperands.isClobbered(PhysReg)) { // Another def has taken the assigned physreg. It must have been a // use&def which got it due to reuse. Undo the reuse! PhysReg = ReusedOperands.GetRegForReload(VirtReg, PhysReg, &MI, Spills, MaybeDeadStores, RegKills, KillOps, *VRM); } } // If StackSlot is available in a register that also holds other stack // slots, clobber those stack slots now. Spills.ClobberSharingStackSlots(StackSlot); assert(PhysReg && "VR not assigned a physical register?"); MRI->setPhysRegUsed(PhysReg); unsigned RReg = SubIdx ? TRI->getSubReg(PhysReg, SubIdx) : PhysReg; ReusedOperands.markClobbered(RReg); MI.getOperand(i).setReg(RReg); MI.getOperand(i).setSubReg(0); if (!MO.isDead() && SpilledMIRegs.insert(VirtReg)) { MachineInstr *&LastStore = MaybeDeadStores[StackSlot]; SpillRegToStackSlot(MII, -1, PhysReg, StackSlot, RC, true, LastStore, Spills, ReMatDefs, RegKills, KillOps); NextMII = llvm::next(MII); // Check to see if this is a noop copy. If so, eliminate the // instruction before considering the dest reg to be changed. if (MI.isIdentityCopy()) { ++NumDCE; DEBUG(dbgs() << "Removing now-noop copy: " << MI); InvalidateKills(MI, TRI, RegKills, KillOps); EraseInstr(&MI); Erased = true; UpdateKills(*LastStore, TRI, RegKills, KillOps); goto ProcessNextInst; } } } ProcessNextInst: // Delete dead instructions without side effects. if (!Erased && !BackTracked && isSafeToDelete(MI)) { InvalidateKills(MI, TRI, RegKills, KillOps); EraseInstr(&MI); Erased = true; } if (!Erased) DistanceMap.insert(std::make_pair(&MI, DistanceMap.size())); if (!Erased && !BackTracked) { for (MachineBasicBlock::iterator II = &MI; II != NextMII; ++II) UpdateKills(*II, TRI, RegKills, KillOps); } MII = NextMII; } } llvm::VirtRegRewriter* llvm::createVirtRegRewriter() { switch (RewriterOpt) { default: llvm_unreachable("Unreachable!"); case local: return new LocalRewriter(); case trivial: return new TrivialRewriter(); } }