//===-- LiveIntervalAnalysis.cpp - Live Interval Analysis -----------------===// // // The LLVM Compiler Infrastructure // // This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // // This file implements the LiveInterval analysis pass which is used // by the Linear Scan Register allocator. This pass linearizes the // basic blocks of the function in DFS order and uses the // LiveVariables pass to conservatively compute live intervals for // each virtual and physical register. // //===----------------------------------------------------------------------===// #define DEBUG_TYPE "liveintervals" #include "llvm/CodeGen/LiveIntervalAnalysis.h" #include "VirtRegMap.h" #include "llvm/Value.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/CodeGen/LiveVariables.h" #include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineInstr.h" #include "llvm/CodeGen/MachineInstrBuilder.h" #include "llvm/CodeGen/MachineLoopInfo.h" #include "llvm/CodeGen/MachineMemOperand.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/Passes.h" #include "llvm/CodeGen/PseudoSourceValue.h" #include "llvm/Target/TargetRegisterInfo.h" #include "llvm/Target/TargetInstrInfo.h" #include "llvm/Target/TargetMachine.h" #include "llvm/Target/TargetOptions.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/raw_ostream.h" #include "llvm/ADT/DepthFirstIterator.h" #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/STLExtras.h" #include #include #include using namespace llvm; // Hidden options for help debugging. static cl::opt DisableReMat("disable-rematerialization", cl::init(false), cl::Hidden); static cl::opt EnableFastSpilling("fast-spill", cl::init(false), cl::Hidden); static cl::opt EarlyCoalescing("early-coalescing", cl::init(false)); static cl::opt CoalescingLimit("early-coalescing-limit", cl::init(-1), cl::Hidden); STATISTIC(numIntervals , "Number of original intervals"); STATISTIC(numFolds , "Number of loads/stores folded into instructions"); STATISTIC(numSplits , "Number of intervals split"); STATISTIC(numCoalescing, "Number of early coalescing performed"); char LiveIntervals::ID = 0; static RegisterPass X("liveintervals", "Live Interval Analysis"); void LiveIntervals::getAnalysisUsage(AnalysisUsage &AU) const { AU.setPreservesCFG(); AU.addRequired(); AU.addPreserved(); AU.addPreserved(); AU.addRequired(); AU.addPreservedID(MachineLoopInfoID); AU.addPreservedID(MachineDominatorsID); if (!StrongPHIElim) { AU.addPreservedID(PHIEliminationID); AU.addRequiredID(PHIEliminationID); } AU.addRequiredID(TwoAddressInstructionPassID); MachineFunctionPass::getAnalysisUsage(AU); } void LiveIntervals::releaseMemory() { // Free the live intervals themselves. for (DenseMap::iterator I = r2iMap_.begin(), E = r2iMap_.end(); I != E; ++I) delete I->second; MBB2IdxMap.clear(); Idx2MBBMap.clear(); mi2iMap_.clear(); i2miMap_.clear(); r2iMap_.clear(); terminatorGaps.clear(); phiJoinCopies.clear(); // Release VNInfo memroy regions after all VNInfo objects are dtor'd. VNInfoAllocator.Reset(); while (!CloneMIs.empty()) { MachineInstr *MI = CloneMIs.back(); CloneMIs.pop_back(); mf_->DeleteMachineInstr(MI); } } static bool CanTurnIntoImplicitDef(MachineInstr *MI, unsigned Reg, unsigned OpIdx, const TargetInstrInfo *tii_){ unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; if (tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubReg, DstSubReg) && Reg == SrcReg) return true; if (OpIdx == 2 && MI->getOpcode() == TargetInstrInfo::SUBREG_TO_REG) return true; if (OpIdx == 1 && MI->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG) return true; return false; } /// processImplicitDefs - Process IMPLICIT_DEF instructions and make sure /// there is one implicit_def for each use. Add isUndef marker to /// implicit_def defs and their uses. void LiveIntervals::processImplicitDefs() { SmallSet ImpDefRegs; SmallVector ImpDefMIs; 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) { MachineBasicBlock *MBB = *DFI; for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; ) { MachineInstr *MI = &*I; ++I; if (MI->getOpcode() == TargetInstrInfo::IMPLICIT_DEF) { unsigned Reg = MI->getOperand(0).getReg(); ImpDefRegs.insert(Reg); if (TargetRegisterInfo::isPhysicalRegister(Reg)) { for (const unsigned *SS = tri_->getSubRegisters(Reg); *SS; ++SS) ImpDefRegs.insert(*SS); } ImpDefMIs.push_back(MI); continue; } if (MI->getOpcode() == TargetInstrInfo::INSERT_SUBREG) { MachineOperand &MO = MI->getOperand(2); if (ImpDefRegs.count(MO.getReg())) { // %reg1032 = INSERT_SUBREG %reg1032, undef, 2 // This is an identity copy, eliminate it now. if (MO.isKill()) { LiveVariables::VarInfo& vi = lv_->getVarInfo(MO.getReg()); vi.removeKill(MI); } MI->eraseFromParent(); continue; } } bool ChangedToImpDef = false; 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) continue; if (!ImpDefRegs.count(Reg)) continue; // Use is a copy, just turn it into an implicit_def. if (CanTurnIntoImplicitDef(MI, Reg, i, tii_)) { bool isKill = MO.isKill(); MI->setDesc(tii_->get(TargetInstrInfo::IMPLICIT_DEF)); for (int j = MI->getNumOperands() - 1, ee = 0; j > ee; --j) MI->RemoveOperand(j); if (isKill) { ImpDefRegs.erase(Reg); LiveVariables::VarInfo& vi = lv_->getVarInfo(Reg); vi.removeKill(MI); } ChangedToImpDef = true; break; } MO.setIsUndef(); if (MO.isKill() || MI->isRegTiedToDefOperand(i)) { // Make sure other uses of for (unsigned j = i+1; j != e; ++j) { MachineOperand &MOJ = MI->getOperand(j); if (MOJ.isReg() && MOJ.isUse() && MOJ.getReg() == Reg) MOJ.setIsUndef(); } ImpDefRegs.erase(Reg); } } if (ChangedToImpDef) { // Backtrack to process this new implicit_def. --I; } else { for (unsigned i = 0; i != MI->getNumOperands(); ++i) { MachineOperand& MO = MI->getOperand(i); if (!MO.isReg() || !MO.isDef()) continue; ImpDefRegs.erase(MO.getReg()); } } } // Any outstanding liveout implicit_def's? for (unsigned i = 0, e = ImpDefMIs.size(); i != e; ++i) { MachineInstr *MI = ImpDefMIs[i]; unsigned Reg = MI->getOperand(0).getReg(); if (TargetRegisterInfo::isPhysicalRegister(Reg) || !ImpDefRegs.count(Reg)) { // Delete all "local" implicit_def's. That include those which define // physical registers since they cannot be liveout. MI->eraseFromParent(); continue; } // If there are multiple defs of the same register and at least one // is not an implicit_def, do not insert implicit_def's before the // uses. bool Skip = false; for (MachineRegisterInfo::def_iterator DI = mri_->def_begin(Reg), DE = mri_->def_end(); DI != DE; ++DI) { if (DI->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) { Skip = true; break; } } if (Skip) continue; // The only implicit_def which we want to keep are those that are live // out of its block. MI->eraseFromParent(); for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(Reg), UE = mri_->use_end(); UI != UE; ) { MachineOperand &RMO = UI.getOperand(); MachineInstr *RMI = &*UI; ++UI; MachineBasicBlock *RMBB = RMI->getParent(); if (RMBB == MBB) continue; // Turn a copy use into an implicit_def. unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; if (tii_->isMoveInstr(*RMI, SrcReg, DstReg, SrcSubReg, DstSubReg) && Reg == SrcReg) { RMI->setDesc(tii_->get(TargetInstrInfo::IMPLICIT_DEF)); for (int j = RMI->getNumOperands() - 1, ee = 0; j > ee; --j) RMI->RemoveOperand(j); continue; } const TargetRegisterClass* RC = mri_->getRegClass(Reg); unsigned NewVReg = mri_->createVirtualRegister(RC); RMO.setReg(NewVReg); RMO.setIsUndef(); RMO.setIsKill(); } } ImpDefRegs.clear(); ImpDefMIs.clear(); } } void LiveIntervals::computeNumbering() { Index2MiMap OldI2MI = i2miMap_; std::vector OldI2MBB = Idx2MBBMap; Idx2MBBMap.clear(); MBB2IdxMap.clear(); mi2iMap_.clear(); i2miMap_.clear(); terminatorGaps.clear(); phiJoinCopies.clear(); FunctionSize = 0; // Number MachineInstrs and MachineBasicBlocks. // Initialize MBB indexes to a sentinal. MBB2IdxMap.resize(mf_->getNumBlockIDs(), std::make_pair(LiveIndex(),LiveIndex())); LiveIndex MIIndex; for (MachineFunction::iterator MBB = mf_->begin(), E = mf_->end(); MBB != E; ++MBB) { LiveIndex StartIdx = MIIndex; // Insert an empty slot at the beginning of each block. MIIndex = getNextIndex(MIIndex); i2miMap_.push_back(0); for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; ++I) { if (I == MBB->getFirstTerminator()) { // Leave a gap for before terminators, this is where we will point // PHI kills. LiveIndex tGap(true, MIIndex); bool inserted = terminatorGaps.insert(std::make_pair(&*MBB, tGap)).second; assert(inserted && "Multiple 'first' terminators encountered during numbering."); inserted = inserted; // Avoid compiler warning if assertions turned off. i2miMap_.push_back(0); MIIndex = getNextIndex(MIIndex); } bool inserted = mi2iMap_.insert(std::make_pair(I, MIIndex)).second; assert(inserted && "multiple MachineInstr -> index mappings"); inserted = true; i2miMap_.push_back(I); MIIndex = getNextIndex(MIIndex); FunctionSize++; // Insert max(1, numdefs) empty slots after every instruction. unsigned Slots = I->getDesc().getNumDefs(); if (Slots == 0) Slots = 1; while (Slots--) { MIIndex = getNextIndex(MIIndex); i2miMap_.push_back(0); } } if (MBB->getFirstTerminator() == MBB->end()) { // Leave a gap for before terminators, this is where we will point // PHI kills. LiveIndex tGap(true, MIIndex); bool inserted = terminatorGaps.insert(std::make_pair(&*MBB, tGap)).second; assert(inserted && "Multiple 'first' terminators encountered during numbering."); inserted = inserted; // Avoid compiler warning if assertions turned off. i2miMap_.push_back(0); MIIndex = getNextIndex(MIIndex); } // Set the MBB2IdxMap entry for this MBB. MBB2IdxMap[MBB->getNumber()] = std::make_pair(StartIdx, getPrevSlot(MIIndex)); Idx2MBBMap.push_back(std::make_pair(StartIdx, MBB)); } std::sort(Idx2MBBMap.begin(), Idx2MBBMap.end(), Idx2MBBCompare()); if (!OldI2MI.empty()) for (iterator OI = begin(), OE = end(); OI != OE; ++OI) { for (LiveInterval::iterator LI = OI->second->begin(), LE = OI->second->end(); LI != LE; ++LI) { // Remap the start index of the live range to the corresponding new // number, or our best guess at what it _should_ correspond to if the // original instruction has been erased. This is either the following // instruction or its predecessor. unsigned index = LI->start.getVecIndex(); LiveIndex::Slot offset = LI->start.getSlot(); if (LI->start.isLoad()) { std::vector::const_iterator I = std::lower_bound(OldI2MBB.begin(), OldI2MBB.end(), LI->start); // Take the pair containing the index std::vector::const_iterator J = (I == OldI2MBB.end() && OldI2MBB.size()>0) ? (I-1): I; LI->start = getMBBStartIdx(J->second); } else { LI->start = LiveIndex( LiveIndex(mi2iMap_[OldI2MI[index]]), (LiveIndex::Slot)offset); } // Remap the ending index in the same way that we remapped the start, // except for the final step where we always map to the immediately // following instruction. index = (getPrevSlot(LI->end)).getVecIndex(); offset = LI->end.getSlot(); if (LI->end.isLoad()) { // VReg dies at end of block. std::vector::const_iterator I = std::lower_bound(OldI2MBB.begin(), OldI2MBB.end(), LI->end); --I; LI->end = getNextSlot(getMBBEndIdx(I->second)); } else { unsigned idx = index; while (index < OldI2MI.size() && !OldI2MI[index]) ++index; if (index != OldI2MI.size()) LI->end = LiveIndex(mi2iMap_[OldI2MI[index]], (idx == index ? offset : LiveIndex::LOAD)); else LI->end = LiveIndex(LiveIndex::NUM * i2miMap_.size()); } } for (LiveInterval::vni_iterator VNI = OI->second->vni_begin(), VNE = OI->second->vni_end(); VNI != VNE; ++VNI) { VNInfo* vni = *VNI; // Remap the VNInfo def index, which works the same as the // start indices above. VN's with special sentinel defs // don't need to be remapped. if (vni->isDefAccurate() && !vni->isUnused()) { unsigned index = vni->def.getVecIndex(); LiveIndex::Slot offset = vni->def.getSlot(); if (vni->def.isLoad()) { std::vector::const_iterator I = std::lower_bound(OldI2MBB.begin(), OldI2MBB.end(), vni->def); // Take the pair containing the index std::vector::const_iterator J = (I == OldI2MBB.end() && OldI2MBB.size()>0) ? (I-1): I; vni->def = getMBBStartIdx(J->second); } else { vni->def = LiveIndex(mi2iMap_[OldI2MI[index]], offset); } } // Remap the VNInfo kill indices, which works the same as // the end indices above. for (size_t i = 0; i < vni->kills.size(); ++i) { unsigned index = getPrevSlot(vni->kills[i]).getVecIndex(); LiveIndex::Slot offset = vni->kills[i].getSlot(); if (vni->kills[i].isLoad()) { assert("Value killed at a load slot."); /*std::vector::const_iterator I = std::lower_bound(OldI2MBB.begin(), OldI2MBB.end(), vni->kills[i]); --I; vni->kills[i] = getMBBEndIdx(I->second);*/ } else { if (vni->kills[i].isPHIIndex()) { std::vector::const_iterator I = std::lower_bound(OldI2MBB.begin(), OldI2MBB.end(), vni->kills[i]); --I; vni->kills[i] = terminatorGaps[I->second]; } else { assert(OldI2MI[index] != 0 && "Kill refers to instruction not present in index maps."); vni->kills[i] = LiveIndex(mi2iMap_[OldI2MI[index]], offset); } /* unsigned idx = index; while (index < OldI2MI.size() && !OldI2MI[index]) ++index; if (index != OldI2MI.size()) vni->kills[i] = mi2iMap_[OldI2MI[index]] + (idx == index ? offset : 0); else vni->kills[i] = InstrSlots::NUM * i2miMap_.size(); */ } } } } } void LiveIntervals::scaleNumbering(int factor) { // Need to // * scale MBB begin and end points // * scale all ranges. // * Update VNI structures. // * Scale instruction numberings // Scale the MBB indices. Idx2MBBMap.clear(); for (MachineFunction::iterator MBB = mf_->begin(), MBBE = mf_->end(); MBB != MBBE; ++MBB) { std::pair &mbbIndices = MBB2IdxMap[MBB->getNumber()]; mbbIndices.first = mbbIndices.first.scale(factor); mbbIndices.second = mbbIndices.second.scale(factor); Idx2MBBMap.push_back(std::make_pair(mbbIndices.first, MBB)); } std::sort(Idx2MBBMap.begin(), Idx2MBBMap.end(), Idx2MBBCompare()); // Scale terminator gaps. for (DenseMap::iterator TGI = terminatorGaps.begin(), TGE = terminatorGaps.end(); TGI != TGE; ++TGI) { terminatorGaps[TGI->first] = TGI->second.scale(factor); } // Scale the intervals. for (iterator LI = begin(), LE = end(); LI != LE; ++LI) { LI->second->scaleNumbering(factor); } // Scale MachineInstrs. Mi2IndexMap oldmi2iMap = mi2iMap_; LiveIndex highestSlot; for (Mi2IndexMap::iterator MI = oldmi2iMap.begin(), ME = oldmi2iMap.end(); MI != ME; ++MI) { LiveIndex newSlot = MI->second.scale(factor); mi2iMap_[MI->first] = newSlot; highestSlot = std::max(highestSlot, newSlot); } unsigned highestVIndex = highestSlot.getVecIndex(); i2miMap_.clear(); i2miMap_.resize(highestVIndex + 1); for (Mi2IndexMap::iterator MI = mi2iMap_.begin(), ME = mi2iMap_.end(); MI != ME; ++MI) { i2miMap_[MI->second.getVecIndex()] = const_cast(MI->first); } } /// runOnMachineFunction - Register allocate the whole function /// bool LiveIntervals::runOnMachineFunction(MachineFunction &fn) { mf_ = &fn; mri_ = &mf_->getRegInfo(); tm_ = &fn.getTarget(); tri_ = tm_->getRegisterInfo(); tii_ = tm_->getInstrInfo(); aa_ = &getAnalysis(); lv_ = &getAnalysis(); allocatableRegs_ = tri_->getAllocatableSet(fn); processImplicitDefs(); computeNumbering(); computeIntervals(); performEarlyCoalescing(); numIntervals += getNumIntervals(); DEBUG(dump()); return true; } /// print - Implement the dump method. void LiveIntervals::print(raw_ostream &OS, const Module* ) const { OS << "********** INTERVALS **********\n"; for (const_iterator I = begin(), E = end(); I != E; ++I) { I->second->print(OS, tri_); OS << "\n"; } printInstrs(OS); } void LiveIntervals::printInstrs(raw_ostream &OS) const { OS << "********** MACHINEINSTRS **********\n"; for (MachineFunction::iterator mbbi = mf_->begin(), mbbe = mf_->end(); mbbi != mbbe; ++mbbi) { OS << ((Value*)mbbi->getBasicBlock())->getName() << ":\n"; for (MachineBasicBlock::iterator mii = mbbi->begin(), mie = mbbi->end(); mii != mie; ++mii) { OS << getInstructionIndex(mii) << '\t' << *mii; } } } void LiveIntervals::dumpInstrs() const { printInstrs(errs()); } /// conflictsWithPhysRegDef - Returns true if the specified register /// is defined during the duration of the specified interval. bool LiveIntervals::conflictsWithPhysRegDef(const LiveInterval &li, VirtRegMap &vrm, unsigned reg) { for (LiveInterval::Ranges::const_iterator I = li.ranges.begin(), E = li.ranges.end(); I != E; ++I) { for (LiveIndex index = getBaseIndex(I->start), end = getNextIndex(getBaseIndex(getPrevSlot(I->end))); index != end; index = getNextIndex(index)) { // skip deleted instructions while (index != end && !getInstructionFromIndex(index)) index = getNextIndex(index); if (index == end) break; MachineInstr *MI = getInstructionFromIndex(index); unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; if (tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubReg, DstSubReg)) if (SrcReg == li.reg || DstReg == li.reg) continue; for (unsigned i = 0; i != MI->getNumOperands(); ++i) { MachineOperand& mop = MI->getOperand(i); if (!mop.isReg()) continue; unsigned PhysReg = mop.getReg(); if (PhysReg == 0 || PhysReg == li.reg) continue; if (TargetRegisterInfo::isVirtualRegister(PhysReg)) { if (!vrm.hasPhys(PhysReg)) continue; PhysReg = vrm.getPhys(PhysReg); } if (PhysReg && tri_->regsOverlap(PhysReg, reg)) return true; } } } return false; } /// conflictsWithPhysRegRef - Similar to conflictsWithPhysRegRef except /// it can check use as well. bool LiveIntervals::conflictsWithPhysRegRef(LiveInterval &li, unsigned Reg, bool CheckUse, SmallPtrSet &JoinedCopies) { for (LiveInterval::Ranges::const_iterator I = li.ranges.begin(), E = li.ranges.end(); I != E; ++I) { for (LiveIndex index = getBaseIndex(I->start), end = getNextIndex(getBaseIndex(getPrevSlot(I->end))); index != end; index = getNextIndex(index)) { // Skip deleted instructions. MachineInstr *MI = 0; while (index != end) { MI = getInstructionFromIndex(index); if (MI) break; index = getNextIndex(index); } if (index == end) break; if (JoinedCopies.count(MI)) continue; for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { MachineOperand& MO = MI->getOperand(i); if (!MO.isReg()) continue; if (MO.isUse() && !CheckUse) continue; unsigned PhysReg = MO.getReg(); if (PhysReg == 0 || TargetRegisterInfo::isVirtualRegister(PhysReg)) continue; if (tri_->isSubRegister(Reg, PhysReg)) return true; } } } return false; } #ifndef NDEBUG static void printRegName(unsigned reg, const TargetRegisterInfo* tri_) { if (TargetRegisterInfo::isPhysicalRegister(reg)) errs() << tri_->getName(reg); else errs() << "%reg" << reg; } #endif void LiveIntervals::handleVirtualRegisterDef(MachineBasicBlock *mbb, MachineBasicBlock::iterator mi, LiveIndex MIIdx, MachineOperand& MO, unsigned MOIdx, LiveInterval &interval) { DEBUG({ errs() << "\t\tregister: "; printRegName(interval.reg, tri_); }); // Virtual registers may be defined multiple times (due to phi // elimination and 2-addr elimination). Much of what we do only has to be // done once for the vreg. We use an empty interval to detect the first // time we see a vreg. LiveVariables::VarInfo& vi = lv_->getVarInfo(interval.reg); if (interval.empty()) { // Get the Idx of the defining instructions. LiveIndex defIndex = getDefIndex(MIIdx); // Earlyclobbers move back one, so that they overlap the live range // of inputs. if (MO.isEarlyClobber()) defIndex = getUseIndex(MIIdx); VNInfo *ValNo; MachineInstr *CopyMI = NULL; unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; if (mi->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG || mi->getOpcode() == TargetInstrInfo::INSERT_SUBREG || mi->getOpcode() == TargetInstrInfo::SUBREG_TO_REG || tii_->isMoveInstr(*mi, SrcReg, DstReg, SrcSubReg, DstSubReg)) CopyMI = mi; // Earlyclobbers move back one. ValNo = interval.getNextValue(defIndex, CopyMI, true, VNInfoAllocator); assert(ValNo->id == 0 && "First value in interval is not 0?"); // Loop over all of the blocks that the vreg is defined in. There are // two cases we have to handle here. The most common case is a vreg // whose lifetime is contained within a basic block. In this case there // will be a single kill, in MBB, which comes after the definition. if (vi.Kills.size() == 1 && vi.Kills[0]->getParent() == mbb) { // FIXME: what about dead vars? LiveIndex killIdx; if (vi.Kills[0] != mi) killIdx = getNextSlot(getUseIndex(getInstructionIndex(vi.Kills[0]))); else if (MO.isEarlyClobber()) // Earlyclobbers that die in this instruction move up one extra, to // compensate for having the starting point moved back one. This // gets them to overlap the live range of other outputs. killIdx = getNextSlot(getNextSlot(defIndex)); else killIdx = getNextSlot(defIndex); // If the kill happens after the definition, we have an intra-block // live range. if (killIdx > defIndex) { assert(vi.AliveBlocks.empty() && "Shouldn't be alive across any blocks!"); LiveRange LR(defIndex, killIdx, ValNo); interval.addRange(LR); DEBUG(errs() << " +" << LR << "\n"); ValNo->addKill(killIdx); return; } } // The other case we handle is when a virtual register lives to the end // of the defining block, potentially live across some blocks, then is // live into some number of blocks, but gets killed. Start by adding a // range that goes from this definition to the end of the defining block. LiveRange NewLR(defIndex, getNextSlot(getMBBEndIdx(mbb)), ValNo); DEBUG(errs() << " +" << NewLR); interval.addRange(NewLR); // Iterate over all of the blocks that the variable is completely // live in, adding [insrtIndex(begin), instrIndex(end)+4) to the // live interval. for (SparseBitVector<>::iterator I = vi.AliveBlocks.begin(), E = vi.AliveBlocks.end(); I != E; ++I) { LiveRange LR(getMBBStartIdx(*I), getNextSlot(getMBBEndIdx(*I)), // MBB ends at -1. ValNo); interval.addRange(LR); DEBUG(errs() << " +" << LR); } // Finally, this virtual register is live from the start of any killing // block to the 'use' slot of the killing instruction. for (unsigned i = 0, e = vi.Kills.size(); i != e; ++i) { MachineInstr *Kill = vi.Kills[i]; LiveIndex killIdx = getNextSlot(getUseIndex(getInstructionIndex(Kill))); LiveRange LR(getMBBStartIdx(Kill->getParent()), killIdx, ValNo); interval.addRange(LR); ValNo->addKill(killIdx); DEBUG(errs() << " +" << LR); } } else { // If this is the second time we see a virtual register definition, it // must be due to phi elimination or two addr elimination. If this is // the result of two address elimination, then the vreg is one of the // def-and-use register operand. if (mi->isRegTiedToUseOperand(MOIdx)) { // If this is a two-address definition, then we have already processed // the live range. The only problem is that we didn't realize there // are actually two values in the live interval. Because of this we // need to take the LiveRegion that defines this register and split it // into two values. assert(interval.containsOneValue()); LiveIndex DefIndex = getDefIndex(interval.getValNumInfo(0)->def); LiveIndex RedefIndex = getDefIndex(MIIdx); if (MO.isEarlyClobber()) RedefIndex = getUseIndex(MIIdx); const LiveRange *OldLR = interval.getLiveRangeContaining(getPrevSlot(RedefIndex)); VNInfo *OldValNo = OldLR->valno; // Delete the initial value, which should be short and continuous, // because the 2-addr copy must be in the same MBB as the redef. interval.removeRange(DefIndex, RedefIndex); // Two-address vregs should always only be redefined once. This means // that at this point, there should be exactly one value number in it. assert(interval.containsOneValue() && "Unexpected 2-addr liveint!"); // The new value number (#1) is defined by the instruction we claimed // defined value #0. VNInfo *ValNo = interval.getNextValue(OldValNo->def, OldValNo->getCopy(), false, // update at * VNInfoAllocator); ValNo->setFlags(OldValNo->getFlags()); // * <- updating here // Value#0 is now defined by the 2-addr instruction. OldValNo->def = RedefIndex; OldValNo->setCopy(0); if (MO.isEarlyClobber()) OldValNo->setHasRedefByEC(true); // Add the new live interval which replaces the range for the input copy. LiveRange LR(DefIndex, RedefIndex, ValNo); DEBUG(errs() << " replace range with " << LR); interval.addRange(LR); ValNo->addKill(RedefIndex); // If this redefinition is dead, we need to add a dummy unit live // range covering the def slot. if (MO.isDead()) interval.addRange( LiveRange(RedefIndex, MO.isEarlyClobber() ? getNextSlot(getNextSlot(RedefIndex)) : getNextSlot(RedefIndex), OldValNo)); DEBUG({ errs() << " RESULT: "; interval.print(errs(), tri_); }); } else { // Otherwise, this must be because of phi elimination. If this is the // first redefinition of the vreg that we have seen, go back and change // the live range in the PHI block to be a different value number. if (interval.containsOneValue()) { // Remove the old range that we now know has an incorrect number. VNInfo *VNI = interval.getValNumInfo(0); MachineInstr *Killer = vi.Kills[0]; phiJoinCopies.push_back(Killer); LiveIndex Start = getMBBStartIdx(Killer->getParent()); LiveIndex End = getNextSlot(getUseIndex(getInstructionIndex(Killer))); DEBUG({ errs() << " Removing [" << Start << "," << End << "] from: "; interval.print(errs(), tri_); errs() << "\n"; }); interval.removeRange(Start, End); assert(interval.ranges.size() == 1 && "Newly discovered PHI interval has >1 ranges."); MachineBasicBlock *killMBB = getMBBFromIndex(interval.endIndex()); VNI->addKill(terminatorGaps[killMBB]); VNI->setHasPHIKill(true); DEBUG({ errs() << " RESULT: "; interval.print(errs(), tri_); }); // Replace the interval with one of a NEW value number. Note that this // value number isn't actually defined by an instruction, weird huh? :) LiveRange LR(Start, End, interval.getNextValue(LiveIndex(mbb->getNumber()), 0, false, VNInfoAllocator)); LR.valno->setIsPHIDef(true); DEBUG(errs() << " replace range with " << LR); interval.addRange(LR); LR.valno->addKill(End); DEBUG({ errs() << " RESULT: "; interval.print(errs(), tri_); }); } // In the case of PHI elimination, each variable definition is only // live until the end of the block. We've already taken care of the // rest of the live range. LiveIndex defIndex = getDefIndex(MIIdx); if (MO.isEarlyClobber()) defIndex = getUseIndex(MIIdx); VNInfo *ValNo; MachineInstr *CopyMI = NULL; unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; if (mi->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG || mi->getOpcode() == TargetInstrInfo::INSERT_SUBREG || mi->getOpcode() == TargetInstrInfo::SUBREG_TO_REG || tii_->isMoveInstr(*mi, SrcReg, DstReg, SrcSubReg, DstSubReg)) CopyMI = mi; ValNo = interval.getNextValue(defIndex, CopyMI, true, VNInfoAllocator); LiveIndex killIndex = getNextSlot(getMBBEndIdx(mbb)); LiveRange LR(defIndex, killIndex, ValNo); interval.addRange(LR); ValNo->addKill(terminatorGaps[mbb]); ValNo->setHasPHIKill(true); DEBUG(errs() << " +" << LR); } } DEBUG(errs() << '\n'); } void LiveIntervals::handlePhysicalRegisterDef(MachineBasicBlock *MBB, MachineBasicBlock::iterator mi, LiveIndex MIIdx, MachineOperand& MO, LiveInterval &interval, MachineInstr *CopyMI) { // A physical register cannot be live across basic block, so its // lifetime must end somewhere in its defining basic block. DEBUG({ errs() << "\t\tregister: "; printRegName(interval.reg, tri_); }); LiveIndex baseIndex = MIIdx; LiveIndex start = getDefIndex(baseIndex); // Earlyclobbers move back one. if (MO.isEarlyClobber()) start = getUseIndex(MIIdx); LiveIndex end = start; // If it is not used after definition, it is considered dead at // the instruction defining it. Hence its interval is: // [defSlot(def), defSlot(def)+1) // For earlyclobbers, the defSlot was pushed back one; the extra // advance below compensates. if (MO.isDead()) { DEBUG(errs() << " dead"); if (MO.isEarlyClobber()) end = getNextSlot(getNextSlot(start)); else end = getNextSlot(start); goto exit; } // If it is not dead on definition, it must be killed by a // subsequent instruction. Hence its interval is: // [defSlot(def), useSlot(kill)+1) baseIndex = getNextIndex(baseIndex); while (++mi != MBB->end()) { while (baseIndex.getVecIndex() < i2miMap_.size() && getInstructionFromIndex(baseIndex) == 0) baseIndex = getNextIndex(baseIndex); if (mi->killsRegister(interval.reg, tri_)) { DEBUG(errs() << " killed"); end = getNextSlot(getUseIndex(baseIndex)); goto exit; } else { int DefIdx = mi->findRegisterDefOperandIdx(interval.reg, false, tri_); if (DefIdx != -1) { if (mi->isRegTiedToUseOperand(DefIdx)) { // Two-address instruction. end = getDefIndex(baseIndex); if (mi->getOperand(DefIdx).isEarlyClobber()) end = getUseIndex(baseIndex); } else { // Another instruction redefines the register before it is ever read. // Then the register is essentially dead at the instruction that defines // it. Hence its interval is: // [defSlot(def), defSlot(def)+1) DEBUG(errs() << " dead"); end = getNextSlot(start); } goto exit; } } baseIndex = getNextIndex(baseIndex); } // The only case we should have a dead physreg here without a killing or // instruction where we know it's dead is if it is live-in to the function // and never used. Another possible case is the implicit use of the // physical register has been deleted by two-address pass. end = getNextSlot(start); exit: assert(start < end && "did not find end of interval?"); // Already exists? Extend old live interval. LiveInterval::iterator OldLR = interval.FindLiveRangeContaining(start); bool Extend = OldLR != interval.end(); VNInfo *ValNo = Extend ? OldLR->valno : interval.getNextValue(start, CopyMI, true, VNInfoAllocator); if (MO.isEarlyClobber() && Extend) ValNo->setHasRedefByEC(true); LiveRange LR(start, end, ValNo); interval.addRange(LR); LR.valno->addKill(end); DEBUG(errs() << " +" << LR << '\n'); } void LiveIntervals::handleRegisterDef(MachineBasicBlock *MBB, MachineBasicBlock::iterator MI, LiveIndex MIIdx, MachineOperand& MO, unsigned MOIdx) { if (TargetRegisterInfo::isVirtualRegister(MO.getReg())) handleVirtualRegisterDef(MBB, MI, MIIdx, MO, MOIdx, getOrCreateInterval(MO.getReg())); else if (allocatableRegs_[MO.getReg()]) { MachineInstr *CopyMI = NULL; unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; if (MI->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG || MI->getOpcode() == TargetInstrInfo::INSERT_SUBREG || MI->getOpcode() == TargetInstrInfo::SUBREG_TO_REG || tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubReg, DstSubReg)) CopyMI = MI; handlePhysicalRegisterDef(MBB, MI, MIIdx, MO, getOrCreateInterval(MO.getReg()), CopyMI); // Def of a register also defines its sub-registers. for (const unsigned* AS = tri_->getSubRegisters(MO.getReg()); *AS; ++AS) // If MI also modifies the sub-register explicitly, avoid processing it // more than once. Do not pass in TRI here so it checks for exact match. if (!MI->modifiesRegister(*AS)) handlePhysicalRegisterDef(MBB, MI, MIIdx, MO, getOrCreateInterval(*AS), 0); } } void LiveIntervals::handleLiveInRegister(MachineBasicBlock *MBB, LiveIndex MIIdx, LiveInterval &interval, bool isAlias) { DEBUG({ errs() << "\t\tlivein register: "; printRegName(interval.reg, tri_); }); // Look for kills, if it reaches a def before it's killed, then it shouldn't // be considered a livein. MachineBasicBlock::iterator mi = MBB->begin(); LiveIndex baseIndex = MIIdx; LiveIndex start = baseIndex; while (baseIndex.getVecIndex() < i2miMap_.size() && getInstructionFromIndex(baseIndex) == 0) baseIndex = getNextIndex(baseIndex); LiveIndex end = baseIndex; bool SeenDefUse = false; while (mi != MBB->end()) { if (mi->killsRegister(interval.reg, tri_)) { DEBUG(errs() << " killed"); end = getNextSlot(getUseIndex(baseIndex)); SeenDefUse = true; break; } else if (mi->modifiesRegister(interval.reg, tri_)) { // Another instruction redefines the register before it is ever read. // Then the register is essentially dead at the instruction that defines // it. Hence its interval is: // [defSlot(def), defSlot(def)+1) DEBUG(errs() << " dead"); end = getNextSlot(getDefIndex(start)); SeenDefUse = true; break; } baseIndex = getNextIndex(baseIndex); ++mi; if (mi != MBB->end()) { while (baseIndex.getVecIndex() < i2miMap_.size() && getInstructionFromIndex(baseIndex) == 0) baseIndex = getNextIndex(baseIndex); } } // Live-in register might not be used at all. if (!SeenDefUse) { if (isAlias) { DEBUG(errs() << " dead"); end = getNextSlot(getDefIndex(MIIdx)); } else { DEBUG(errs() << " live through"); end = baseIndex; } } VNInfo *vni = interval.getNextValue(LiveIndex(MBB->getNumber()), 0, false, VNInfoAllocator); vni->setIsPHIDef(true); LiveRange LR(start, end, vni); interval.addRange(LR); LR.valno->addKill(end); DEBUG(errs() << " +" << LR << '\n'); } bool LiveIntervals::isProfitableToCoalesce(LiveInterval &DstInt, LiveInterval &SrcInt, SmallVector &IdentCopies, SmallVector &OtherCopies) { bool HaveConflict = false; unsigned NumIdent = 0; for (MachineRegisterInfo::def_iterator ri = mri_->def_begin(SrcInt.reg), re = mri_->def_end(); ri != re; ++ri) { MachineInstr *MI = &*ri; unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; if (!tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubReg, DstSubReg)) return false; if (SrcReg != DstInt.reg) { OtherCopies.push_back(MI); HaveConflict |= DstInt.liveAt(getInstructionIndex(MI)); } else { IdentCopies.push_back(MI); ++NumIdent; } } if (!HaveConflict) return false; // Let coalescer handle it return IdentCopies.size() > OtherCopies.size(); } void LiveIntervals::performEarlyCoalescing() { if (!EarlyCoalescing) return; /// Perform early coalescing: eliminate copies which feed into phi joins /// and whose sources are defined by the phi joins. for (unsigned i = 0, e = phiJoinCopies.size(); i != e; ++i) { MachineInstr *Join = phiJoinCopies[i]; if (CoalescingLimit != -1 && (int)numCoalescing == CoalescingLimit) break; unsigned PHISrc, PHIDst, SrcSubReg, DstSubReg; bool isMove= tii_->isMoveInstr(*Join, PHISrc, PHIDst, SrcSubReg, DstSubReg); #ifndef NDEBUG assert(isMove && "PHI join instruction must be a move!"); #else isMove = isMove; #endif LiveInterval &DstInt = getInterval(PHIDst); LiveInterval &SrcInt = getInterval(PHISrc); SmallVector IdentCopies; SmallVector OtherCopies; if (!isProfitableToCoalesce(DstInt, SrcInt, IdentCopies, OtherCopies)) continue; DEBUG(errs() << "PHI Join: " << *Join); assert(DstInt.containsOneValue() && "PHI join should have just one val#!"); VNInfo *VNI = DstInt.getValNumInfo(0); // Change the non-identity copies to directly target the phi destination. for (unsigned i = 0, e = OtherCopies.size(); i != e; ++i) { MachineInstr *PHICopy = OtherCopies[i]; DEBUG(errs() << "Moving: " << *PHICopy); LiveIndex MIIndex = getInstructionIndex(PHICopy); LiveIndex DefIndex = getDefIndex(MIIndex); LiveRange *SLR = SrcInt.getLiveRangeContaining(DefIndex); LiveIndex StartIndex = SLR->start; LiveIndex EndIndex = SLR->end; // Delete val# defined by the now identity copy and add the range from // beginning of the mbb to the end of the range. SrcInt.removeValNo(SLR->valno); DEBUG(errs() << " added range [" << StartIndex << ',' << EndIndex << "] to reg" << DstInt.reg << '\n'); if (DstInt.liveAt(StartIndex)) DstInt.removeRange(StartIndex, EndIndex); VNInfo *NewVNI = DstInt.getNextValue(DefIndex, PHICopy, true, VNInfoAllocator); NewVNI->setHasPHIKill(true); DstInt.addRange(LiveRange(StartIndex, EndIndex, NewVNI)); for (unsigned j = 0, ee = PHICopy->getNumOperands(); j != ee; ++j) { MachineOperand &MO = PHICopy->getOperand(j); if (!MO.isReg() || MO.getReg() != PHISrc) continue; MO.setReg(PHIDst); } } // Now let's eliminate all the would-be identity copies. for (unsigned i = 0, e = IdentCopies.size(); i != e; ++i) { MachineInstr *PHICopy = IdentCopies[i]; DEBUG(errs() << "Coalescing: " << *PHICopy); LiveIndex MIIndex = getInstructionIndex(PHICopy); LiveIndex DefIndex = getDefIndex(MIIndex); LiveRange *SLR = SrcInt.getLiveRangeContaining(DefIndex); LiveIndex StartIndex = SLR->start; LiveIndex EndIndex = SLR->end; // Delete val# defined by the now identity copy and add the range from // beginning of the mbb to the end of the range. SrcInt.removeValNo(SLR->valno); RemoveMachineInstrFromMaps(PHICopy); PHICopy->eraseFromParent(); DEBUG(errs() << " added range [" << StartIndex << ',' << EndIndex << "] to reg" << DstInt.reg << '\n'); DstInt.addRange(LiveRange(StartIndex, EndIndex, VNI)); } // Remove the phi join and update the phi block liveness. LiveIndex MIIndex = getInstructionIndex(Join); LiveIndex UseIndex = getUseIndex(MIIndex); LiveIndex DefIndex = getDefIndex(MIIndex); LiveRange *SLR = SrcInt.getLiveRangeContaining(UseIndex); LiveRange *DLR = DstInt.getLiveRangeContaining(DefIndex); DLR->valno->setCopy(0); DLR->valno->setIsDefAccurate(false); DstInt.addRange(LiveRange(SLR->start, SLR->end, DLR->valno)); SrcInt.removeRange(SLR->start, SLR->end); assert(SrcInt.empty()); removeInterval(PHISrc); RemoveMachineInstrFromMaps(Join); Join->eraseFromParent(); ++numCoalescing; } } /// computeIntervals - computes the live intervals for virtual /// registers. for some ordering of the machine instructions [1,N] a /// live interval is an interval [i, j) where 1 <= i <= j < N for /// which a variable is live void LiveIntervals::computeIntervals() { DEBUG(errs() << "********** COMPUTING LIVE INTERVALS **********\n" << "********** Function: " << ((Value*)mf_->getFunction())->getName() << '\n'); SmallVector UndefUses; for (MachineFunction::iterator MBBI = mf_->begin(), E = mf_->end(); MBBI != E; ++MBBI) { MachineBasicBlock *MBB = MBBI; // Track the index of the current machine instr. LiveIndex MIIndex = getMBBStartIdx(MBB); DEBUG(errs() << ((Value*)MBB->getBasicBlock())->getName() << ":\n"); MachineBasicBlock::iterator MI = MBB->begin(), miEnd = MBB->end(); // Create intervals for live-ins to this BB first. for (MachineBasicBlock::const_livein_iterator LI = MBB->livein_begin(), LE = MBB->livein_end(); LI != LE; ++LI) { handleLiveInRegister(MBB, MIIndex, getOrCreateInterval(*LI)); // Multiple live-ins can alias the same register. for (const unsigned* AS = tri_->getSubRegisters(*LI); *AS; ++AS) if (!hasInterval(*AS)) handleLiveInRegister(MBB, MIIndex, getOrCreateInterval(*AS), true); } // Skip over empty initial indices. while (MIIndex.getVecIndex() < i2miMap_.size() && getInstructionFromIndex(MIIndex) == 0) MIIndex = getNextIndex(MIIndex); for (; MI != miEnd; ++MI) { DEBUG(errs() << MIIndex << "\t" << *MI); // Handle defs. for (int i = MI->getNumOperands() - 1; i >= 0; --i) { MachineOperand &MO = MI->getOperand(i); if (!MO.isReg() || !MO.getReg()) continue; // handle register defs - build intervals if (MO.isDef()) handleRegisterDef(MBB, MI, MIIndex, MO, i); else if (MO.isUndef()) UndefUses.push_back(MO.getReg()); } // Skip over the empty slots after each instruction. unsigned Slots = MI->getDesc().getNumDefs(); if (Slots == 0) Slots = 1; while (Slots--) MIIndex = getNextIndex(MIIndex); // Skip over empty indices. while (MIIndex.getVecIndex() < i2miMap_.size() && getInstructionFromIndex(MIIndex) == 0) MIIndex = getNextIndex(MIIndex); } } // Create empty intervals for registers defined by implicit_def's (except // for those implicit_def that define values which are liveout of their // blocks. for (unsigned i = 0, e = UndefUses.size(); i != e; ++i) { unsigned UndefReg = UndefUses[i]; (void)getOrCreateInterval(UndefReg); } } bool LiveIntervals::findLiveInMBBs( LiveIndex Start, LiveIndex End, SmallVectorImpl &MBBs) const { std::vector::const_iterator I = std::lower_bound(Idx2MBBMap.begin(), Idx2MBBMap.end(), Start); bool ResVal = false; while (I != Idx2MBBMap.end()) { if (I->first >= End) break; MBBs.push_back(I->second); ResVal = true; ++I; } return ResVal; } bool LiveIntervals::findReachableMBBs( LiveIndex Start, LiveIndex End, SmallVectorImpl &MBBs) const { std::vector::const_iterator I = std::lower_bound(Idx2MBBMap.begin(), Idx2MBBMap.end(), Start); bool ResVal = false; while (I != Idx2MBBMap.end()) { if (I->first > End) break; MachineBasicBlock *MBB = I->second; if (getMBBEndIdx(MBB) > End) break; for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), SE = MBB->succ_end(); SI != SE; ++SI) MBBs.push_back(*SI); ResVal = true; ++I; } return ResVal; } LiveInterval* LiveIntervals::createInterval(unsigned reg) { float Weight = TargetRegisterInfo::isPhysicalRegister(reg) ? HUGE_VALF : 0.0F; return new LiveInterval(reg, Weight); } /// dupInterval - Duplicate a live interval. The caller is responsible for /// managing the allocated memory. LiveInterval* LiveIntervals::dupInterval(LiveInterval *li) { LiveInterval *NewLI = createInterval(li->reg); NewLI->Copy(*li, mri_, getVNInfoAllocator()); return NewLI; } /// getVNInfoSourceReg - Helper function that parses the specified VNInfo /// copy field and returns the source register that defines it. unsigned LiveIntervals::getVNInfoSourceReg(const VNInfo *VNI) const { if (!VNI->getCopy()) return 0; if (VNI->getCopy()->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG) { // If it's extracting out of a physical register, return the sub-register. unsigned Reg = VNI->getCopy()->getOperand(1).getReg(); if (TargetRegisterInfo::isPhysicalRegister(Reg)) Reg = tri_->getSubReg(Reg, VNI->getCopy()->getOperand(2).getImm()); return Reg; } else if (VNI->getCopy()->getOpcode() == TargetInstrInfo::INSERT_SUBREG || VNI->getCopy()->getOpcode() == TargetInstrInfo::SUBREG_TO_REG) return VNI->getCopy()->getOperand(2).getReg(); unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; if (tii_->isMoveInstr(*VNI->getCopy(), SrcReg, DstReg, SrcSubReg, DstSubReg)) return SrcReg; llvm_unreachable("Unrecognized copy instruction!"); return 0; } //===----------------------------------------------------------------------===// // Register allocator hooks. // /// getReMatImplicitUse - If the remat definition MI has one (for now, we only /// allow one) virtual register operand, then its uses are implicitly using /// the register. Returns the virtual register. unsigned LiveIntervals::getReMatImplicitUse(const LiveInterval &li, MachineInstr *MI) const { unsigned RegOp = 0; for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { MachineOperand &MO = MI->getOperand(i); if (!MO.isReg() || !MO.isUse()) continue; unsigned Reg = MO.getReg(); if (Reg == 0 || Reg == li.reg) continue; if (TargetRegisterInfo::isPhysicalRegister(Reg) && !allocatableRegs_[Reg]) continue; // FIXME: For now, only remat MI with at most one register operand. assert(!RegOp && "Can't rematerialize instruction with multiple register operand!"); RegOp = MO.getReg(); #ifndef NDEBUG break; #endif } return RegOp; } /// isValNoAvailableAt - Return true if the val# of the specified interval /// which reaches the given instruction also reaches the specified use index. bool LiveIntervals::isValNoAvailableAt(const LiveInterval &li, MachineInstr *MI, LiveIndex UseIdx) const { LiveIndex Index = getInstructionIndex(MI); VNInfo *ValNo = li.FindLiveRangeContaining(Index)->valno; LiveInterval::const_iterator UI = li.FindLiveRangeContaining(UseIdx); return UI != li.end() && UI->valno == ValNo; } /// isReMaterializable - Returns true if the definition MI of the specified /// val# of the specified interval is re-materializable. bool LiveIntervals::isReMaterializable(const LiveInterval &li, const VNInfo *ValNo, MachineInstr *MI, SmallVectorImpl &SpillIs, bool &isLoad) { if (DisableReMat) return false; if (!tii_->isTriviallyReMaterializable(MI, aa_)) return false; // Target-specific code can mark an instruction as being rematerializable // if it has one virtual reg use, though it had better be something like // a PIC base register which is likely to be live everywhere. unsigned ImpUse = getReMatImplicitUse(li, MI); if (ImpUse) { const LiveInterval &ImpLi = getInterval(ImpUse); for (MachineRegisterInfo::use_iterator ri = mri_->use_begin(li.reg), re = mri_->use_end(); ri != re; ++ri) { MachineInstr *UseMI = &*ri; LiveIndex UseIdx = getInstructionIndex(UseMI); if (li.FindLiveRangeContaining(UseIdx)->valno != ValNo) continue; if (!isValNoAvailableAt(ImpLi, MI, UseIdx)) return false; } // If a register operand of the re-materialized instruction is going to // be spilled next, then it's not legal to re-materialize this instruction. for (unsigned i = 0, e = SpillIs.size(); i != e; ++i) if (ImpUse == SpillIs[i]->reg) return false; } return true; } /// isReMaterializable - Returns true if the definition MI of the specified /// val# of the specified interval is re-materializable. bool LiveIntervals::isReMaterializable(const LiveInterval &li, const VNInfo *ValNo, MachineInstr *MI) { SmallVector Dummy1; bool Dummy2; return isReMaterializable(li, ValNo, MI, Dummy1, Dummy2); } /// isReMaterializable - Returns true if every definition of MI of every /// val# of the specified interval is re-materializable. bool LiveIntervals::isReMaterializable(const LiveInterval &li, SmallVectorImpl &SpillIs, bool &isLoad) { isLoad = false; for (LiveInterval::const_vni_iterator i = li.vni_begin(), e = li.vni_end(); i != e; ++i) { const VNInfo *VNI = *i; if (VNI->isUnused()) continue; // Dead val#. // Is the def for the val# rematerializable? if (!VNI->isDefAccurate()) return false; MachineInstr *ReMatDefMI = getInstructionFromIndex(VNI->def); bool DefIsLoad = false; if (!ReMatDefMI || !isReMaterializable(li, VNI, ReMatDefMI, SpillIs, DefIsLoad)) return false; isLoad |= DefIsLoad; } return true; } /// FilterFoldedOps - Filter out two-address use operands. Return /// true if it finds any issue with the operands that ought to prevent /// folding. static bool FilterFoldedOps(MachineInstr *MI, SmallVector &Ops, unsigned &MRInfo, SmallVector &FoldOps) { MRInfo = 0; for (unsigned i = 0, e = Ops.size(); i != e; ++i) { unsigned OpIdx = Ops[i]; MachineOperand &MO = MI->getOperand(OpIdx); // FIXME: fold subreg use. if (MO.getSubReg()) return true; if (MO.isDef()) MRInfo |= (unsigned)VirtRegMap::isMod; else { // Filter out two-address use operand(s). if (MI->isRegTiedToDefOperand(OpIdx)) { MRInfo = VirtRegMap::isModRef; continue; } MRInfo |= (unsigned)VirtRegMap::isRef; } FoldOps.push_back(OpIdx); } return false; } /// tryFoldMemoryOperand - Attempts to fold either a spill / restore from /// slot / to reg or any rematerialized load into ith operand of specified /// MI. If it is successul, MI is updated with the newly created MI and /// returns true. bool LiveIntervals::tryFoldMemoryOperand(MachineInstr* &MI, VirtRegMap &vrm, MachineInstr *DefMI, LiveIndex InstrIdx, SmallVector &Ops, bool isSS, int Slot, unsigned Reg) { // If it is an implicit def instruction, just delete it. if (MI->getOpcode() == TargetInstrInfo::IMPLICIT_DEF) { RemoveMachineInstrFromMaps(MI); vrm.RemoveMachineInstrFromMaps(MI); MI->eraseFromParent(); ++numFolds; return true; } // Filter the list of operand indexes that are to be folded. Abort if // any operand will prevent folding. unsigned MRInfo = 0; SmallVector FoldOps; if (FilterFoldedOps(MI, Ops, MRInfo, FoldOps)) return false; // The only time it's safe to fold into a two address instruction is when // it's folding reload and spill from / into a spill stack slot. if (DefMI && (MRInfo & VirtRegMap::isMod)) return false; MachineInstr *fmi = isSS ? tii_->foldMemoryOperand(*mf_, MI, FoldOps, Slot) : tii_->foldMemoryOperand(*mf_, MI, FoldOps, DefMI); if (fmi) { // Remember this instruction uses the spill slot. if (isSS) vrm.addSpillSlotUse(Slot, fmi); // Attempt to fold the memory reference into the instruction. If // we can do this, we don't need to insert spill code. MachineBasicBlock &MBB = *MI->getParent(); if (isSS && !mf_->getFrameInfo()->isImmutableObjectIndex(Slot)) vrm.virtFolded(Reg, MI, fmi, (VirtRegMap::ModRef)MRInfo); vrm.transferSpillPts(MI, fmi); vrm.transferRestorePts(MI, fmi); vrm.transferEmergencySpills(MI, fmi); mi2iMap_.erase(MI); i2miMap_[InstrIdx.getVecIndex()] = fmi; mi2iMap_[fmi] = InstrIdx; MI = MBB.insert(MBB.erase(MI), fmi); ++numFolds; return true; } return false; } /// canFoldMemoryOperand - Returns true if the specified load / store /// folding is possible. bool LiveIntervals::canFoldMemoryOperand(MachineInstr *MI, SmallVector &Ops, bool ReMat) const { // Filter the list of operand indexes that are to be folded. Abort if // any operand will prevent folding. unsigned MRInfo = 0; SmallVector FoldOps; if (FilterFoldedOps(MI, Ops, MRInfo, FoldOps)) return false; // It's only legal to remat for a use, not a def. if (ReMat && (MRInfo & VirtRegMap::isMod)) return false; return tii_->canFoldMemoryOperand(MI, FoldOps); } bool LiveIntervals::intervalIsInOneMBB(const LiveInterval &li) const { SmallPtrSet MBBs; for (LiveInterval::Ranges::const_iterator I = li.ranges.begin(), E = li.ranges.end(); I != E; ++I) { std::vector::const_iterator II = std::lower_bound(Idx2MBBMap.begin(), Idx2MBBMap.end(), I->start); if (II == Idx2MBBMap.end()) continue; if (I->end > II->first) // crossing a MBB. return false; MBBs.insert(II->second); if (MBBs.size() > 1) return false; } return true; } /// rewriteImplicitOps - Rewrite implicit use operands of MI (i.e. uses of /// interval on to-be re-materialized operands of MI) with new register. void LiveIntervals::rewriteImplicitOps(const LiveInterval &li, MachineInstr *MI, unsigned NewVReg, VirtRegMap &vrm) { // There is an implicit use. That means one of the other operand is // being remat'ed and the remat'ed instruction has li.reg as an // use operand. Make sure we rewrite that as well. for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { MachineOperand &MO = MI->getOperand(i); if (!MO.isReg()) continue; unsigned Reg = MO.getReg(); if (Reg == 0 || TargetRegisterInfo::isPhysicalRegister(Reg)) continue; if (!vrm.isReMaterialized(Reg)) continue; MachineInstr *ReMatMI = vrm.getReMaterializedMI(Reg); MachineOperand *UseMO = ReMatMI->findRegisterUseOperand(li.reg); if (UseMO) UseMO->setReg(NewVReg); } } /// rewriteInstructionForSpills, rewriteInstructionsForSpills - Helper functions /// for addIntervalsForSpills to rewrite uses / defs for the given live range. bool LiveIntervals:: rewriteInstructionForSpills(const LiveInterval &li, const VNInfo *VNI, bool TrySplit, LiveIndex index, LiveIndex end, MachineInstr *MI, MachineInstr *ReMatOrigDefMI, MachineInstr *ReMatDefMI, unsigned Slot, int LdSlot, bool isLoad, bool isLoadSS, bool DefIsReMat, bool CanDelete, VirtRegMap &vrm, const TargetRegisterClass* rc, SmallVector &ReMatIds, const MachineLoopInfo *loopInfo, unsigned &NewVReg, unsigned ImpUse, bool &HasDef, bool &HasUse, DenseMap &MBBVRegsMap, std::vector &NewLIs) { bool CanFold = false; RestartInstruction: for (unsigned i = 0; i != MI->getNumOperands(); ++i) { MachineOperand& mop = MI->getOperand(i); if (!mop.isReg()) continue; unsigned Reg = mop.getReg(); unsigned RegI = Reg; if (Reg == 0 || TargetRegisterInfo::isPhysicalRegister(Reg)) continue; if (Reg != li.reg) continue; bool TryFold = !DefIsReMat; bool FoldSS = true; // Default behavior unless it's a remat. int FoldSlot = Slot; if (DefIsReMat) { // If this is the rematerializable definition MI itself and // all of its uses are rematerialized, simply delete it. if (MI == ReMatOrigDefMI && CanDelete) { DEBUG(errs() << "\t\t\t\tErasing re-materlizable def: " << MI << '\n'); RemoveMachineInstrFromMaps(MI); vrm.RemoveMachineInstrFromMaps(MI); MI->eraseFromParent(); break; } // If def for this use can't be rematerialized, then try folding. // If def is rematerializable and it's a load, also try folding. TryFold = !ReMatDefMI || (ReMatDefMI && (MI == ReMatOrigDefMI || isLoad)); if (isLoad) { // Try fold loads (from stack slot, constant pool, etc.) into uses. FoldSS = isLoadSS; FoldSlot = LdSlot; } } // Scan all of the operands of this instruction rewriting operands // to use NewVReg instead of li.reg as appropriate. We do this for // two reasons: // // 1. If the instr reads the same spilled vreg multiple times, we // want to reuse the NewVReg. // 2. If the instr is a two-addr instruction, we are required to // keep the src/dst regs pinned. // // Keep track of whether we replace a use and/or def so that we can // create the spill interval with the appropriate range. HasUse = mop.isUse(); HasDef = mop.isDef(); SmallVector Ops; Ops.push_back(i); for (unsigned j = i+1, e = MI->getNumOperands(); j != e; ++j) { const MachineOperand &MOj = MI->getOperand(j); if (!MOj.isReg()) continue; unsigned RegJ = MOj.getReg(); if (RegJ == 0 || TargetRegisterInfo::isPhysicalRegister(RegJ)) continue; if (RegJ == RegI) { Ops.push_back(j); if (!MOj.isUndef()) { HasUse |= MOj.isUse(); HasDef |= MOj.isDef(); } } } // Create a new virtual register for the spill interval. // Create the new register now so we can map the fold instruction // to the new register so when it is unfolded we get the correct // answer. bool CreatedNewVReg = false; if (NewVReg == 0) { NewVReg = mri_->createVirtualRegister(rc); vrm.grow(); CreatedNewVReg = true; } if (!TryFold) CanFold = false; else { // Do not fold load / store here if we are splitting. We'll find an // optimal point to insert a load / store later. if (!TrySplit) { if (tryFoldMemoryOperand(MI, vrm, ReMatDefMI, index, Ops, FoldSS, FoldSlot, NewVReg)) { // Folding the load/store can completely change the instruction in // unpredictable ways, rescan it from the beginning. if (FoldSS) { // We need to give the new vreg the same stack slot as the // spilled interval. vrm.assignVirt2StackSlot(NewVReg, FoldSlot); } HasUse = false; HasDef = false; CanFold = false; if (isNotInMIMap(MI)) break; goto RestartInstruction; } } else { // We'll try to fold it later if it's profitable. CanFold = canFoldMemoryOperand(MI, Ops, DefIsReMat); } } mop.setReg(NewVReg); if (mop.isImplicit()) rewriteImplicitOps(li, MI, NewVReg, vrm); // Reuse NewVReg for other reads. for (unsigned j = 0, e = Ops.size(); j != e; ++j) { MachineOperand &mopj = MI->getOperand(Ops[j]); mopj.setReg(NewVReg); if (mopj.isImplicit()) rewriteImplicitOps(li, MI, NewVReg, vrm); } if (CreatedNewVReg) { if (DefIsReMat) { vrm.setVirtIsReMaterialized(NewVReg, ReMatDefMI); if (ReMatIds[VNI->id] == VirtRegMap::MAX_STACK_SLOT) { // Each valnum may have its own remat id. ReMatIds[VNI->id] = vrm.assignVirtReMatId(NewVReg); } else { vrm.assignVirtReMatId(NewVReg, ReMatIds[VNI->id]); } if (!CanDelete || (HasUse && HasDef)) { // If this is a two-addr instruction then its use operands are // rematerializable but its def is not. It should be assigned a // stack slot. vrm.assignVirt2StackSlot(NewVReg, Slot); } } else { vrm.assignVirt2StackSlot(NewVReg, Slot); } } else if (HasUse && HasDef && vrm.getStackSlot(NewVReg) == VirtRegMap::NO_STACK_SLOT) { // If this interval hasn't been assigned a stack slot (because earlier // def is a deleted remat def), do it now. assert(Slot != VirtRegMap::NO_STACK_SLOT); vrm.assignVirt2StackSlot(NewVReg, Slot); } // Re-matting an instruction with virtual register use. Add the // register as an implicit use on the use MI. if (DefIsReMat && ImpUse) MI->addOperand(MachineOperand::CreateReg(ImpUse, false, true)); // Create a new register interval for this spill / remat. LiveInterval &nI = getOrCreateInterval(NewVReg); if (CreatedNewVReg) { NewLIs.push_back(&nI); MBBVRegsMap.insert(std::make_pair(MI->getParent()->getNumber(), NewVReg)); if (TrySplit) vrm.setIsSplitFromReg(NewVReg, li.reg); } if (HasUse) { if (CreatedNewVReg) { LiveRange LR(getLoadIndex(index), getNextSlot(getUseIndex(index)), nI.getNextValue(LiveIndex(), 0, false, VNInfoAllocator)); DEBUG(errs() << " +" << LR); nI.addRange(LR); } else { // Extend the split live interval to this def / use. LiveIndex End = getNextSlot(getUseIndex(index)); LiveRange LR(nI.ranges[nI.ranges.size()-1].end, End, nI.getValNumInfo(nI.getNumValNums()-1)); DEBUG(errs() << " +" << LR); nI.addRange(LR); } } if (HasDef) { LiveRange LR(getDefIndex(index), getStoreIndex(index), nI.getNextValue(LiveIndex(), 0, false, VNInfoAllocator)); DEBUG(errs() << " +" << LR); nI.addRange(LR); } DEBUG({ errs() << "\t\t\t\tAdded new interval: "; nI.print(errs(), tri_); errs() << '\n'; }); } return CanFold; } bool LiveIntervals::anyKillInMBBAfterIdx(const LiveInterval &li, const VNInfo *VNI, MachineBasicBlock *MBB, LiveIndex Idx) const { LiveIndex End = getMBBEndIdx(MBB); for (unsigned j = 0, ee = VNI->kills.size(); j != ee; ++j) { if (VNI->kills[j].isPHIIndex()) continue; LiveIndex KillIdx = VNI->kills[j]; if (KillIdx > Idx && KillIdx < End) return true; } return false; } /// RewriteInfo - Keep track of machine instrs that will be rewritten /// during spilling. namespace { struct RewriteInfo { LiveIndex Index; MachineInstr *MI; bool HasUse; bool HasDef; RewriteInfo(LiveIndex i, MachineInstr *mi, bool u, bool d) : Index(i), MI(mi), HasUse(u), HasDef(d) {} }; struct RewriteInfoCompare { bool operator()(const RewriteInfo &LHS, const RewriteInfo &RHS) const { return LHS.Index < RHS.Index; } }; } void LiveIntervals:: rewriteInstructionsForSpills(const LiveInterval &li, bool TrySplit, LiveInterval::Ranges::const_iterator &I, MachineInstr *ReMatOrigDefMI, MachineInstr *ReMatDefMI, unsigned Slot, int LdSlot, bool isLoad, bool isLoadSS, bool DefIsReMat, bool CanDelete, VirtRegMap &vrm, const TargetRegisterClass* rc, SmallVector &ReMatIds, const MachineLoopInfo *loopInfo, BitVector &SpillMBBs, DenseMap > &SpillIdxes, BitVector &RestoreMBBs, DenseMap > &RestoreIdxes, DenseMap &MBBVRegsMap, std::vector &NewLIs) { bool AllCanFold = true; unsigned NewVReg = 0; LiveIndex start = getBaseIndex(I->start); LiveIndex end = getNextIndex(getBaseIndex(getPrevSlot(I->end))); // First collect all the def / use in this live range that will be rewritten. // Make sure they are sorted according to instruction index. std::vector RewriteMIs; for (MachineRegisterInfo::reg_iterator ri = mri_->reg_begin(li.reg), re = mri_->reg_end(); ri != re; ) { MachineInstr *MI = &*ri; MachineOperand &O = ri.getOperand(); ++ri; assert(!O.isImplicit() && "Spilling register that's used as implicit use?"); LiveIndex index = getInstructionIndex(MI); if (index < start || index >= end) continue; if (O.isUndef()) // Must be 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. continue; RewriteMIs.push_back(RewriteInfo(index, MI, O.isUse(), O.isDef())); } std::sort(RewriteMIs.begin(), RewriteMIs.end(), RewriteInfoCompare()); unsigned ImpUse = DefIsReMat ? getReMatImplicitUse(li, ReMatDefMI) : 0; // Now rewrite the defs and uses. for (unsigned i = 0, e = RewriteMIs.size(); i != e; ) { RewriteInfo &rwi = RewriteMIs[i]; ++i; LiveIndex index = rwi.Index; bool MIHasUse = rwi.HasUse; bool MIHasDef = rwi.HasDef; MachineInstr *MI = rwi.MI; // If MI def and/or use the same register multiple times, then there // are multiple entries. unsigned NumUses = MIHasUse; while (i != e && RewriteMIs[i].MI == MI) { assert(RewriteMIs[i].Index == index); bool isUse = RewriteMIs[i].HasUse; if (isUse) ++NumUses; MIHasUse |= isUse; MIHasDef |= RewriteMIs[i].HasDef; ++i; } MachineBasicBlock *MBB = MI->getParent(); if (ImpUse && MI != ReMatDefMI) { // Re-matting an instruction with virtual register use. Update the // register interval's spill weight to HUGE_VALF to prevent it from // being spilled. LiveInterval &ImpLi = getInterval(ImpUse); ImpLi.weight = HUGE_VALF; } unsigned MBBId = MBB->getNumber(); unsigned ThisVReg = 0; if (TrySplit) { DenseMap::iterator NVI = MBBVRegsMap.find(MBBId); if (NVI != MBBVRegsMap.end()) { ThisVReg = NVI->second; // One common case: // x = use // ... // ... // def = ... // = use // It's better to start a new interval to avoid artifically // extend the new interval. if (MIHasDef && !MIHasUse) { MBBVRegsMap.erase(MBB->getNumber()); ThisVReg = 0; } } } bool IsNew = ThisVReg == 0; if (IsNew) { // This ends the previous live interval. If all of its def / use // can be folded, give it a low spill weight. if (NewVReg && TrySplit && AllCanFold) { LiveInterval &nI = getOrCreateInterval(NewVReg); nI.weight /= 10.0F; } AllCanFold = true; } NewVReg = ThisVReg; bool HasDef = false; bool HasUse = false; bool CanFold = rewriteInstructionForSpills(li, I->valno, TrySplit, index, end, MI, ReMatOrigDefMI, ReMatDefMI, Slot, LdSlot, isLoad, isLoadSS, DefIsReMat, CanDelete, vrm, rc, ReMatIds, loopInfo, NewVReg, ImpUse, HasDef, HasUse, MBBVRegsMap, NewLIs); if (!HasDef && !HasUse) continue; AllCanFold &= CanFold; // Update weight of spill interval. LiveInterval &nI = getOrCreateInterval(NewVReg); if (!TrySplit) { // The spill weight is now infinity as it cannot be spilled again. nI.weight = HUGE_VALF; continue; } // Keep track of the last def and first use in each MBB. if (HasDef) { if (MI != ReMatOrigDefMI || !CanDelete) { bool HasKill = false; if (!HasUse) HasKill = anyKillInMBBAfterIdx(li, I->valno, MBB, getDefIndex(index)); else { // If this is a two-address code, then this index starts a new VNInfo. const VNInfo *VNI = li.findDefinedVNInfoForRegInt(getDefIndex(index)); if (VNI) HasKill = anyKillInMBBAfterIdx(li, VNI, MBB, getDefIndex(index)); } DenseMap >::iterator SII = SpillIdxes.find(MBBId); if (!HasKill) { if (SII == SpillIdxes.end()) { std::vector S; S.push_back(SRInfo(index, NewVReg, true)); SpillIdxes.insert(std::make_pair(MBBId, S)); } else if (SII->second.back().vreg != NewVReg) { SII->second.push_back(SRInfo(index, NewVReg, true)); } else if (index > SII->second.back().index) { // If there is an earlier def and this is a two-address // instruction, then it's not possible to fold the store (which // would also fold the load). SRInfo &Info = SII->second.back(); Info.index = index; Info.canFold = !HasUse; } SpillMBBs.set(MBBId); } else if (SII != SpillIdxes.end() && SII->second.back().vreg == NewVReg && index > SII->second.back().index) { // There is an earlier def that's not killed (must be two-address). // The spill is no longer needed. SII->second.pop_back(); if (SII->second.empty()) { SpillIdxes.erase(MBBId); SpillMBBs.reset(MBBId); } } } } if (HasUse) { DenseMap >::iterator SII = SpillIdxes.find(MBBId); if (SII != SpillIdxes.end() && SII->second.back().vreg == NewVReg && index > SII->second.back().index) // Use(s) following the last def, it's not safe to fold the spill. SII->second.back().canFold = false; DenseMap >::iterator RII = RestoreIdxes.find(MBBId); if (RII != RestoreIdxes.end() && RII->second.back().vreg == NewVReg) // If we are splitting live intervals, only fold if it's the first // use and there isn't another use later in the MBB. RII->second.back().canFold = false; else if (IsNew) { // Only need a reload if there isn't an earlier def / use. if (RII == RestoreIdxes.end()) { std::vector Infos; Infos.push_back(SRInfo(index, NewVReg, true)); RestoreIdxes.insert(std::make_pair(MBBId, Infos)); } else { RII->second.push_back(SRInfo(index, NewVReg, true)); } RestoreMBBs.set(MBBId); } } // Update spill weight. unsigned loopDepth = loopInfo->getLoopDepth(MBB); nI.weight += getSpillWeight(HasDef, HasUse, loopDepth); } if (NewVReg && TrySplit && AllCanFold) { // If all of its def / use can be folded, give it a low spill weight. LiveInterval &nI = getOrCreateInterval(NewVReg); nI.weight /= 10.0F; } } bool LiveIntervals::alsoFoldARestore(int Id, LiveIndex index, unsigned vr, BitVector &RestoreMBBs, DenseMap > &RestoreIdxes) { if (!RestoreMBBs[Id]) return false; std::vector &Restores = RestoreIdxes[Id]; for (unsigned i = 0, e = Restores.size(); i != e; ++i) if (Restores[i].index == index && Restores[i].vreg == vr && Restores[i].canFold) return true; return false; } void LiveIntervals::eraseRestoreInfo(int Id, LiveIndex index, unsigned vr, BitVector &RestoreMBBs, DenseMap > &RestoreIdxes) { if (!RestoreMBBs[Id]) return; std::vector &Restores = RestoreIdxes[Id]; for (unsigned i = 0, e = Restores.size(); i != e; ++i) if (Restores[i].index == index && Restores[i].vreg) Restores[i].index = LiveIndex(); } /// handleSpilledImpDefs - Remove IMPLICIT_DEF instructions which are being /// spilled and create empty intervals for their uses. void LiveIntervals::handleSpilledImpDefs(const LiveInterval &li, VirtRegMap &vrm, const TargetRegisterClass* rc, std::vector &NewLIs) { for (MachineRegisterInfo::reg_iterator ri = mri_->reg_begin(li.reg), re = mri_->reg_end(); ri != re; ) { MachineOperand &O = ri.getOperand(); MachineInstr *MI = &*ri; ++ri; if (O.isDef()) { assert(MI->getOpcode() == TargetInstrInfo::IMPLICIT_DEF && "Register def was not rewritten?"); RemoveMachineInstrFromMaps(MI); vrm.RemoveMachineInstrFromMaps(MI); MI->eraseFromParent(); } else { // This must be an use of an implicit_def so it's not part of the live // interval. Create a new empty live interval for it. // FIXME: Can we simply erase some of the instructions? e.g. Stores? unsigned NewVReg = mri_->createVirtualRegister(rc); vrm.grow(); vrm.setIsImplicitlyDefined(NewVReg); NewLIs.push_back(&getOrCreateInterval(NewVReg)); for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { MachineOperand &MO = MI->getOperand(i); if (MO.isReg() && MO.getReg() == li.reg) { MO.setReg(NewVReg); MO.setIsUndef(); } } } } } std::vector LiveIntervals:: addIntervalsForSpillsFast(const LiveInterval &li, const MachineLoopInfo *loopInfo, VirtRegMap &vrm) { unsigned slot = vrm.assignVirt2StackSlot(li.reg); std::vector added; assert(li.weight != HUGE_VALF && "attempt to spill already spilled interval!"); DEBUG({ errs() << "\t\t\t\tadding intervals for spills for interval: "; li.dump(); errs() << '\n'; }); const TargetRegisterClass* rc = mri_->getRegClass(li.reg); MachineRegisterInfo::reg_iterator RI = mri_->reg_begin(li.reg); while (RI != mri_->reg_end()) { MachineInstr* MI = &*RI; SmallVector Indices; bool HasUse = false; bool HasDef = false; for (unsigned i = 0; i != MI->getNumOperands(); ++i) { MachineOperand& mop = MI->getOperand(i); if (!mop.isReg() || mop.getReg() != li.reg) continue; HasUse |= MI->getOperand(i).isUse(); HasDef |= MI->getOperand(i).isDef(); Indices.push_back(i); } if (!tryFoldMemoryOperand(MI, vrm, NULL, getInstructionIndex(MI), Indices, true, slot, li.reg)) { unsigned NewVReg = mri_->createVirtualRegister(rc); vrm.grow(); vrm.assignVirt2StackSlot(NewVReg, slot); // create a new register for this spill LiveInterval &nI = getOrCreateInterval(NewVReg); // the spill weight is now infinity as it // cannot be spilled again nI.weight = HUGE_VALF; // Rewrite register operands to use the new vreg. for (SmallVectorImpl::iterator I = Indices.begin(), E = Indices.end(); I != E; ++I) { MI->getOperand(*I).setReg(NewVReg); if (MI->getOperand(*I).isUse()) MI->getOperand(*I).setIsKill(true); } // Fill in the new live interval. LiveIndex index = getInstructionIndex(MI); if (HasUse) { LiveRange LR(getLoadIndex(index), getUseIndex(index), nI.getNextValue(LiveIndex(), 0, false, getVNInfoAllocator())); DEBUG(errs() << " +" << LR); nI.addRange(LR); vrm.addRestorePoint(NewVReg, MI); } if (HasDef) { LiveRange LR(getDefIndex(index), getStoreIndex(index), nI.getNextValue(LiveIndex(), 0, false, getVNInfoAllocator())); DEBUG(errs() << " +" << LR); nI.addRange(LR); vrm.addSpillPoint(NewVReg, true, MI); } added.push_back(&nI); DEBUG({ errs() << "\t\t\t\tadded new interval: "; nI.dump(); errs() << '\n'; }); } RI = mri_->reg_begin(li.reg); } return added; } std::vector LiveIntervals:: addIntervalsForSpills(const LiveInterval &li, SmallVectorImpl &SpillIs, const MachineLoopInfo *loopInfo, VirtRegMap &vrm) { if (EnableFastSpilling) return addIntervalsForSpillsFast(li, loopInfo, vrm); assert(li.weight != HUGE_VALF && "attempt to spill already spilled interval!"); DEBUG({ errs() << "\t\t\t\tadding intervals for spills for interval: "; li.print(errs(), tri_); errs() << '\n'; }); // Each bit specify whether a spill is required in the MBB. BitVector SpillMBBs(mf_->getNumBlockIDs()); DenseMap > SpillIdxes; BitVector RestoreMBBs(mf_->getNumBlockIDs()); DenseMap > RestoreIdxes; DenseMap MBBVRegsMap; std::vector NewLIs; const TargetRegisterClass* rc = mri_->getRegClass(li.reg); unsigned NumValNums = li.getNumValNums(); SmallVector ReMatDefs; ReMatDefs.resize(NumValNums, NULL); SmallVector ReMatOrigDefs; ReMatOrigDefs.resize(NumValNums, NULL); SmallVector ReMatIds; ReMatIds.resize(NumValNums, VirtRegMap::MAX_STACK_SLOT); BitVector ReMatDelete(NumValNums); unsigned Slot = VirtRegMap::MAX_STACK_SLOT; // Spilling a split live interval. It cannot be split any further. Also, // it's also guaranteed to be a single val# / range interval. if (vrm.getPreSplitReg(li.reg)) { vrm.setIsSplitFromReg(li.reg, 0); // Unset the split kill marker on the last use. LiveIndex KillIdx = vrm.getKillPoint(li.reg); if (KillIdx != LiveIndex()) { MachineInstr *KillMI = getInstructionFromIndex(KillIdx); assert(KillMI && "Last use disappeared?"); int KillOp = KillMI->findRegisterUseOperandIdx(li.reg, true); assert(KillOp != -1 && "Last use disappeared?"); KillMI->getOperand(KillOp).setIsKill(false); } vrm.removeKillPoint(li.reg); bool DefIsReMat = vrm.isReMaterialized(li.reg); Slot = vrm.getStackSlot(li.reg); assert(Slot != VirtRegMap::MAX_STACK_SLOT); MachineInstr *ReMatDefMI = DefIsReMat ? vrm.getReMaterializedMI(li.reg) : NULL; int LdSlot = 0; bool isLoadSS = DefIsReMat && tii_->isLoadFromStackSlot(ReMatDefMI, LdSlot); bool isLoad = isLoadSS || (DefIsReMat && (ReMatDefMI->getDesc().canFoldAsLoad())); bool IsFirstRange = true; for (LiveInterval::Ranges::const_iterator I = li.ranges.begin(), E = li.ranges.end(); I != E; ++I) { // If this is a split live interval with multiple ranges, it means there // are two-address instructions that re-defined the value. Only the // first def can be rematerialized! if (IsFirstRange) { // Note ReMatOrigDefMI has already been deleted. rewriteInstructionsForSpills(li, false, I, NULL, ReMatDefMI, Slot, LdSlot, isLoad, isLoadSS, DefIsReMat, false, vrm, rc, ReMatIds, loopInfo, SpillMBBs, SpillIdxes, RestoreMBBs, RestoreIdxes, MBBVRegsMap, NewLIs); } else { rewriteInstructionsForSpills(li, false, I, NULL, 0, Slot, 0, false, false, false, false, vrm, rc, ReMatIds, loopInfo, SpillMBBs, SpillIdxes, RestoreMBBs, RestoreIdxes, MBBVRegsMap, NewLIs); } IsFirstRange = false; } handleSpilledImpDefs(li, vrm, rc, NewLIs); return NewLIs; } bool TrySplit = !intervalIsInOneMBB(li); if (TrySplit) ++numSplits; bool NeedStackSlot = false; for (LiveInterval::const_vni_iterator i = li.vni_begin(), e = li.vni_end(); i != e; ++i) { const VNInfo *VNI = *i; unsigned VN = VNI->id; if (VNI->isUnused()) continue; // Dead val#. // Is the def for the val# rematerializable? MachineInstr *ReMatDefMI = VNI->isDefAccurate() ? getInstructionFromIndex(VNI->def) : 0; bool dummy; if (ReMatDefMI && isReMaterializable(li, VNI, ReMatDefMI, SpillIs, dummy)) { // Remember how to remat the def of this val#. ReMatOrigDefs[VN] = ReMatDefMI; // Original def may be modified so we have to make a copy here. MachineInstr *Clone = mf_->CloneMachineInstr(ReMatDefMI); CloneMIs.push_back(Clone); ReMatDefs[VN] = Clone; bool CanDelete = true; if (VNI->hasPHIKill()) { // A kill is a phi node, not all of its uses can be rematerialized. // It must not be deleted. CanDelete = false; // Need a stack slot if there is any live range where uses cannot be // rematerialized. NeedStackSlot = true; } if (CanDelete) ReMatDelete.set(VN); } else { // Need a stack slot if there is any live range where uses cannot be // rematerialized. NeedStackSlot = true; } } // One stack slot per live interval. if (NeedStackSlot && vrm.getPreSplitReg(li.reg) == 0) { if (vrm.getStackSlot(li.reg) == VirtRegMap::NO_STACK_SLOT) Slot = vrm.assignVirt2StackSlot(li.reg); // This case only occurs when the prealloc splitter has already assigned // a stack slot to this vreg. else Slot = vrm.getStackSlot(li.reg); } // Create new intervals and rewrite defs and uses. for (LiveInterval::Ranges::const_iterator I = li.ranges.begin(), E = li.ranges.end(); I != E; ++I) { MachineInstr *ReMatDefMI = ReMatDefs[I->valno->id]; MachineInstr *ReMatOrigDefMI = ReMatOrigDefs[I->valno->id]; bool DefIsReMat = ReMatDefMI != NULL; bool CanDelete = ReMatDelete[I->valno->id]; int LdSlot = 0; bool isLoadSS = DefIsReMat && tii_->isLoadFromStackSlot(ReMatDefMI, LdSlot); bool isLoad = isLoadSS || (DefIsReMat && ReMatDefMI->getDesc().canFoldAsLoad()); rewriteInstructionsForSpills(li, TrySplit, I, ReMatOrigDefMI, ReMatDefMI, Slot, LdSlot, isLoad, isLoadSS, DefIsReMat, CanDelete, vrm, rc, ReMatIds, loopInfo, SpillMBBs, SpillIdxes, RestoreMBBs, RestoreIdxes, MBBVRegsMap, NewLIs); } // Insert spills / restores if we are splitting. if (!TrySplit) { handleSpilledImpDefs(li, vrm, rc, NewLIs); return NewLIs; } SmallPtrSet AddedKill; SmallVector Ops; if (NeedStackSlot) { int Id = SpillMBBs.find_first(); while (Id != -1) { std::vector &spills = SpillIdxes[Id]; for (unsigned i = 0, e = spills.size(); i != e; ++i) { LiveIndex index = spills[i].index; unsigned VReg = spills[i].vreg; LiveInterval &nI = getOrCreateInterval(VReg); bool isReMat = vrm.isReMaterialized(VReg); MachineInstr *MI = getInstructionFromIndex(index); bool CanFold = false; bool FoundUse = false; Ops.clear(); if (spills[i].canFold) { CanFold = true; for (unsigned j = 0, ee = MI->getNumOperands(); j != ee; ++j) { MachineOperand &MO = MI->getOperand(j); if (!MO.isReg() || MO.getReg() != VReg) continue; Ops.push_back(j); if (MO.isDef()) continue; if (isReMat || (!FoundUse && !alsoFoldARestore(Id, index, VReg, RestoreMBBs, RestoreIdxes))) { // MI has two-address uses of the same register. If the use // isn't the first and only use in the BB, then we can't fold // it. FIXME: Move this to rewriteInstructionsForSpills. CanFold = false; break; } FoundUse = true; } } // Fold the store into the def if possible. bool Folded = false; if (CanFold && !Ops.empty()) { if (tryFoldMemoryOperand(MI, vrm, NULL, index, Ops, true, Slot,VReg)){ Folded = true; if (FoundUse) { // Also folded uses, do not issue a load. eraseRestoreInfo(Id, index, VReg, RestoreMBBs, RestoreIdxes); nI.removeRange(getLoadIndex(index), getNextSlot(getUseIndex(index))); } nI.removeRange(getDefIndex(index), getStoreIndex(index)); } } // Otherwise tell the spiller to issue a spill. if (!Folded) { LiveRange *LR = &nI.ranges[nI.ranges.size()-1]; bool isKill = LR->end == getStoreIndex(index); if (!MI->registerDefIsDead(nI.reg)) // No need to spill a dead def. vrm.addSpillPoint(VReg, isKill, MI); if (isKill) AddedKill.insert(&nI); } } Id = SpillMBBs.find_next(Id); } } int Id = RestoreMBBs.find_first(); while (Id != -1) { std::vector &restores = RestoreIdxes[Id]; for (unsigned i = 0, e = restores.size(); i != e; ++i) { LiveIndex index = restores[i].index; if (index == LiveIndex()) continue; unsigned VReg = restores[i].vreg; LiveInterval &nI = getOrCreateInterval(VReg); bool isReMat = vrm.isReMaterialized(VReg); MachineInstr *MI = getInstructionFromIndex(index); bool CanFold = false; Ops.clear(); if (restores[i].canFold) { CanFold = true; for (unsigned j = 0, ee = MI->getNumOperands(); j != ee; ++j) { MachineOperand &MO = MI->getOperand(j); if (!MO.isReg() || MO.getReg() != VReg) continue; if (MO.isDef()) { // If this restore were to be folded, it would have been folded // already. CanFold = false; break; } Ops.push_back(j); } } // Fold the load into the use if possible. bool Folded = false; if (CanFold && !Ops.empty()) { if (!isReMat) Folded = tryFoldMemoryOperand(MI, vrm, NULL,index,Ops,true,Slot,VReg); else { MachineInstr *ReMatDefMI = vrm.getReMaterializedMI(VReg); int LdSlot = 0; bool isLoadSS = tii_->isLoadFromStackSlot(ReMatDefMI, LdSlot); // If the rematerializable def is a load, also try to fold it. if (isLoadSS || ReMatDefMI->getDesc().canFoldAsLoad()) Folded = tryFoldMemoryOperand(MI, vrm, ReMatDefMI, index, Ops, isLoadSS, LdSlot, VReg); if (!Folded) { unsigned ImpUse = getReMatImplicitUse(li, ReMatDefMI); if (ImpUse) { // Re-matting an instruction with virtual register use. Add the // register as an implicit use on the use MI and update the register // interval's spill weight to HUGE_VALF to prevent it from being // spilled. LiveInterval &ImpLi = getInterval(ImpUse); ImpLi.weight = HUGE_VALF; MI->addOperand(MachineOperand::CreateReg(ImpUse, false, true)); } } } } // If folding is not possible / failed, then tell the spiller to issue a // load / rematerialization for us. if (Folded) nI.removeRange(getLoadIndex(index), getNextSlot(getUseIndex(index))); else vrm.addRestorePoint(VReg, MI); } Id = RestoreMBBs.find_next(Id); } // Finalize intervals: add kills, finalize spill weights, and filter out // dead intervals. std::vector RetNewLIs; for (unsigned i = 0, e = NewLIs.size(); i != e; ++i) { LiveInterval *LI = NewLIs[i]; if (!LI->empty()) { LI->weight /= InstrSlots::NUM * getApproximateInstructionCount(*LI); if (!AddedKill.count(LI)) { LiveRange *LR = &LI->ranges[LI->ranges.size()-1]; LiveIndex LastUseIdx = getBaseIndex(LR->end); MachineInstr *LastUse = getInstructionFromIndex(LastUseIdx); int UseIdx = LastUse->findRegisterUseOperandIdx(LI->reg, false); assert(UseIdx != -1); if (!LastUse->isRegTiedToDefOperand(UseIdx)) { LastUse->getOperand(UseIdx).setIsKill(); vrm.addKillPoint(LI->reg, LastUseIdx); } } RetNewLIs.push_back(LI); } } handleSpilledImpDefs(li, vrm, rc, RetNewLIs); return RetNewLIs; } /// hasAllocatableSuperReg - Return true if the specified physical register has /// any super register that's allocatable. bool LiveIntervals::hasAllocatableSuperReg(unsigned Reg) const { for (const unsigned* AS = tri_->getSuperRegisters(Reg); *AS; ++AS) if (allocatableRegs_[*AS] && hasInterval(*AS)) return true; return false; } /// getRepresentativeReg - Find the largest super register of the specified /// physical register. unsigned LiveIntervals::getRepresentativeReg(unsigned Reg) const { // Find the largest super-register that is allocatable. unsigned BestReg = Reg; for (const unsigned* AS = tri_->getSuperRegisters(Reg); *AS; ++AS) { unsigned SuperReg = *AS; if (!hasAllocatableSuperReg(SuperReg) && hasInterval(SuperReg)) { BestReg = SuperReg; break; } } return BestReg; } /// getNumConflictsWithPhysReg - Return the number of uses and defs of the /// specified interval that conflicts with the specified physical register. unsigned LiveIntervals::getNumConflictsWithPhysReg(const LiveInterval &li, unsigned PhysReg) const { unsigned NumConflicts = 0; const LiveInterval &pli = getInterval(getRepresentativeReg(PhysReg)); for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(li.reg), E = mri_->reg_end(); I != E; ++I) { MachineOperand &O = I.getOperand(); MachineInstr *MI = O.getParent(); LiveIndex Index = getInstructionIndex(MI); if (pli.liveAt(Index)) ++NumConflicts; } return NumConflicts; } /// spillPhysRegAroundRegDefsUses - Spill the specified physical register /// around all defs and uses of the specified interval. Return true if it /// was able to cut its interval. bool LiveIntervals::spillPhysRegAroundRegDefsUses(const LiveInterval &li, unsigned PhysReg, VirtRegMap &vrm) { unsigned SpillReg = getRepresentativeReg(PhysReg); for (const unsigned *AS = tri_->getAliasSet(PhysReg); *AS; ++AS) // If there are registers which alias PhysReg, but which are not a // sub-register of the chosen representative super register. Assert // since we can't handle it yet. assert(*AS == SpillReg || !allocatableRegs_[*AS] || !hasInterval(*AS) || tri_->isSuperRegister(*AS, SpillReg)); bool Cut = false; LiveInterval &pli = getInterval(SpillReg); SmallPtrSet SeenMIs; for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(li.reg), E = mri_->reg_end(); I != E; ++I) { MachineOperand &O = I.getOperand(); MachineInstr *MI = O.getParent(); if (SeenMIs.count(MI)) continue; SeenMIs.insert(MI); LiveIndex Index = getInstructionIndex(MI); if (pli.liveAt(Index)) { vrm.addEmergencySpill(SpillReg, MI); LiveIndex StartIdx = getLoadIndex(Index); LiveIndex EndIdx = getNextSlot(getStoreIndex(Index)); if (pli.isInOneLiveRange(StartIdx, EndIdx)) { pli.removeRange(StartIdx, EndIdx); Cut = true; } else { std::string msg; raw_string_ostream Msg(msg); Msg << "Ran out of registers during register allocation!"; if (MI->getOpcode() == TargetInstrInfo::INLINEASM) { Msg << "\nPlease check your inline asm statement for invalid " << "constraints:\n"; MI->print(Msg, tm_); } llvm_report_error(Msg.str()); } for (const unsigned* AS = tri_->getSubRegisters(SpillReg); *AS; ++AS) { if (!hasInterval(*AS)) continue; LiveInterval &spli = getInterval(*AS); if (spli.liveAt(Index)) spli.removeRange(getLoadIndex(Index), getNextSlot(getStoreIndex(Index))); } } } return Cut; } LiveRange LiveIntervals::addLiveRangeToEndOfBlock(unsigned reg, MachineInstr* startInst) { LiveInterval& Interval = getOrCreateInterval(reg); VNInfo* VN = Interval.getNextValue( LiveIndex(getInstructionIndex(startInst), LiveIndex::DEF), startInst, true, getVNInfoAllocator()); VN->setHasPHIKill(true); VN->kills.push_back(terminatorGaps[startInst->getParent()]); LiveRange LR( LiveIndex(getInstructionIndex(startInst), LiveIndex::DEF), getNextSlot(getMBBEndIdx(startInst->getParent())), VN); Interval.addRange(LR); return LR; }