//===-- LiveIntervalAnalysis.cpp - Live Interval Analysis -----------------===// // // The LLVM Compiler Infrastructure // // This file was developed by the LLVM research group and 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/LoopInfo.h" #include "llvm/CodeGen/LiveVariables.h" #include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineInstr.h" #include "llvm/CodeGen/Passes.h" #include "llvm/CodeGen/SSARegMap.h" #include "llvm/Target/MRegisterInfo.h" #include "llvm/Target/TargetInstrInfo.h" #include "llvm/Target/TargetMachine.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/STLExtras.h" #include #include using namespace llvm; namespace { // Hidden options for help debugging. cl::opt DisableReMat("disable-rematerialization", cl::init(false), cl::Hidden); cl::opt SplitAtBB("split-intervals-at-bb", cl::init(false), cl::Hidden); cl::opt SplitLimit("split-limit", cl::init(-1), cl::Hidden); } STATISTIC(numIntervals, "Number of original intervals"); STATISTIC(numIntervalsAfter, "Number of intervals after coalescing"); STATISTIC(numFolds , "Number of loads/stores folded into instructions"); STATISTIC(numSplits , "Number of intervals split"); char LiveIntervals::ID = 0; namespace { RegisterPass X("liveintervals", "Live Interval Analysis"); } void LiveIntervals::getAnalysisUsage(AnalysisUsage &AU) const { AU.addPreserved(); AU.addRequired(); AU.addPreservedID(PHIEliminationID); AU.addRequiredID(PHIEliminationID); AU.addRequiredID(TwoAddressInstructionPassID); MachineFunctionPass::getAnalysisUsage(AU); } void LiveIntervals::releaseMemory() { Idx2MBBMap.clear(); mi2iMap_.clear(); i2miMap_.clear(); r2iMap_.clear(); // Release VNInfo memroy regions after all VNInfo objects are dtor'd. VNInfoAllocator.Reset(); for (unsigned i = 0, e = ClonedMIs.size(); i != e; ++i) delete ClonedMIs[i]; } namespace llvm { inline bool operator<(unsigned V, const IdxMBBPair &IM) { return V < IM.first; } inline bool operator<(const IdxMBBPair &IM, unsigned V) { return IM.first < V; } struct Idx2MBBCompare { bool operator()(const IdxMBBPair &LHS, const IdxMBBPair &RHS) const { return LHS.first < RHS.first; } }; } /// runOnMachineFunction - Register allocate the whole function /// bool LiveIntervals::runOnMachineFunction(MachineFunction &fn) { mf_ = &fn; tm_ = &fn.getTarget(); mri_ = tm_->getRegisterInfo(); tii_ = tm_->getInstrInfo(); lv_ = &getAnalysis(); allocatableRegs_ = mri_->getAllocatableSet(fn); // Number MachineInstrs and MachineBasicBlocks. // Initialize MBB indexes to a sentinal. MBB2IdxMap.resize(mf_->getNumBlockIDs(), std::make_pair(~0U,~0U)); unsigned MIIndex = 0; for (MachineFunction::iterator MBB = mf_->begin(), E = mf_->end(); MBB != E; ++MBB) { unsigned StartIdx = MIIndex; for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; ++I) { bool inserted = mi2iMap_.insert(std::make_pair(I, MIIndex)).second; assert(inserted && "multiple MachineInstr -> index mappings"); i2miMap_.push_back(I); MIIndex += InstrSlots::NUM; } // Set the MBB2IdxMap entry for this MBB. MBB2IdxMap[MBB->getNumber()] = std::make_pair(StartIdx, MIIndex - 1); Idx2MBBMap.push_back(std::make_pair(StartIdx, MBB)); } std::sort(Idx2MBBMap.begin(), Idx2MBBMap.end(), Idx2MBBCompare()); computeIntervals(); numIntervals += getNumIntervals(); DOUT << "********** INTERVALS **********\n"; for (iterator I = begin(), E = end(); I != E; ++I) { I->second.print(DOUT, mri_); DOUT << "\n"; } numIntervalsAfter += getNumIntervals(); DEBUG(dump()); return true; } /// print - Implement the dump method. void LiveIntervals::print(std::ostream &O, const Module* ) const { O << "********** INTERVALS **********\n"; for (const_iterator I = begin(), E = end(); I != E; ++I) { I->second.print(DOUT, mri_); DOUT << "\n"; } O << "********** MACHINEINSTRS **********\n"; for (MachineFunction::iterator mbbi = mf_->begin(), mbbe = mf_->end(); mbbi != mbbe; ++mbbi) { O << ((Value*)mbbi->getBasicBlock())->getName() << ":\n"; for (MachineBasicBlock::iterator mii = mbbi->begin(), mie = mbbi->end(); mii != mie; ++mii) { O << getInstructionIndex(mii) << '\t' << *mii; } } } /// 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 (unsigned index = getBaseIndex(I->start), end = getBaseIndex(I->end-1) + InstrSlots::NUM; index != end; index += InstrSlots::NUM) { // skip deleted instructions while (index != end && !getInstructionFromIndex(index)) index += InstrSlots::NUM; if (index == end) break; MachineInstr *MI = getInstructionFromIndex(index); unsigned SrcReg, DstReg; if (tii_->isMoveInstr(*MI, SrcReg, DstReg)) if (SrcReg == li.reg || DstReg == li.reg) continue; for (unsigned i = 0; i != MI->getNumOperands(); ++i) { MachineOperand& mop = MI->getOperand(i); if (!mop.isRegister()) continue; unsigned PhysReg = mop.getReg(); if (PhysReg == 0 || PhysReg == li.reg) continue; if (MRegisterInfo::isVirtualRegister(PhysReg)) { if (!vrm.hasPhys(PhysReg)) continue; PhysReg = vrm.getPhys(PhysReg); } if (PhysReg && mri_->regsOverlap(PhysReg, reg)) return true; } } } return false; } void LiveIntervals::printRegName(unsigned reg) const { if (MRegisterInfo::isPhysicalRegister(reg)) cerr << mri_->getName(reg); else cerr << "%reg" << reg; } void LiveIntervals::handleVirtualRegisterDef(MachineBasicBlock *mbb, MachineBasicBlock::iterator mi, unsigned MIIdx, LiveInterval &interval) { DOUT << "\t\tregister: "; DEBUG(printRegName(interval.reg)); LiveVariables::VarInfo& vi = lv_->getVarInfo(interval.reg); // 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. if (interval.empty()) { // Get the Idx of the defining instructions. unsigned defIndex = getDefIndex(MIIdx); VNInfo *ValNo; unsigned SrcReg, DstReg; if (tii_->isMoveInstr(*mi, SrcReg, DstReg)) ValNo = interval.getNextValue(defIndex, SrcReg, VNInfoAllocator); else if (mi->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG) ValNo = interval.getNextValue(defIndex, mi->getOperand(1).getReg(), VNInfoAllocator); else ValNo = interval.getNextValue(defIndex, 0, 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? unsigned killIdx; if (vi.Kills[0] != mi) killIdx = getUseIndex(getInstructionIndex(vi.Kills[0]))+1; else killIdx = defIndex+1; // If the kill happens after the definition, we have an intra-block // live range. if (killIdx > defIndex) { assert(vi.AliveBlocks.none() && "Shouldn't be alive across any blocks!"); LiveRange LR(defIndex, killIdx, ValNo); interval.addRange(LR); DOUT << " +" << LR << "\n"; interval.addKill(ValNo, 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, getInstructionIndex(&mbb->back()) + InstrSlots::NUM, ValNo); DOUT << " +" << 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 (unsigned i = 0, e = vi.AliveBlocks.size(); i != e; ++i) { if (vi.AliveBlocks[i]) { MachineBasicBlock *MBB = mf_->getBlockNumbered(i); if (!MBB->empty()) { LiveRange LR(getMBBStartIdx(i), getInstructionIndex(&MBB->back()) + InstrSlots::NUM, ValNo); interval.addRange(LR); DOUT << " +" << 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]; unsigned killIdx = getUseIndex(getInstructionIndex(Kill))+1; LiveRange LR(getMBBStartIdx(Kill->getParent()), killIdx, ValNo); interval.addRange(LR); interval.addKill(ValNo, killIdx); DOUT << " +" << 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->isRegReDefinedByTwoAddr(interval.reg)) { // 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. unsigned DefIndex = getDefIndex(getInstructionIndex(vi.DefInst)); unsigned RedefIndex = getDefIndex(MIIdx); const LiveRange *OldLR = interval.getLiveRangeContaining(RedefIndex-1); VNInfo *OldValNo = OldLR->valno; unsigned OldEnd = OldLR->end; // 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(0, 0, VNInfoAllocator); interval.copyValNumInfo(ValNo, OldValNo); // Value#0 is now defined by the 2-addr instruction. OldValNo->def = RedefIndex; OldValNo->reg = 0; // Add the new live interval which replaces the range for the input copy. LiveRange LR(DefIndex, RedefIndex, ValNo); DOUT << " replace range with " << LR; interval.addRange(LR); interval.addKill(ValNo, RedefIndex); interval.removeKills(ValNo, RedefIndex, OldEnd); // If this redefinition is dead, we need to add a dummy unit live // range covering the def slot. if (lv_->RegisterDefIsDead(mi, interval.reg)) interval.addRange(LiveRange(RedefIndex, RedefIndex+1, OldValNo)); DOUT << " RESULT: "; interval.print(DOUT, mri_); } 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()) { assert(vi.Kills.size() == 1 && "PHI elimination vreg should have one kill, the PHI itself!"); // Remove the old range that we now know has an incorrect number. VNInfo *VNI = interval.getValNumInfo(0); MachineInstr *Killer = vi.Kills[0]; unsigned Start = getMBBStartIdx(Killer->getParent()); unsigned End = getUseIndex(getInstructionIndex(Killer))+1; DOUT << " Removing [" << Start << "," << End << "] from: "; interval.print(DOUT, mri_); DOUT << "\n"; interval.removeRange(Start, End); interval.addKill(VNI, Start); VNI->hasPHIKill = true; DOUT << " RESULT: "; interval.print(DOUT, mri_); // 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(~0, 0, VNInfoAllocator)); DOUT << " replace range with " << LR; interval.addRange(LR); interval.addKill(LR.valno, End); DOUT << " RESULT: "; interval.print(DOUT, mri_); } // 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. unsigned defIndex = getDefIndex(MIIdx); VNInfo *ValNo; unsigned SrcReg, DstReg; if (tii_->isMoveInstr(*mi, SrcReg, DstReg)) ValNo = interval.getNextValue(defIndex, SrcReg, VNInfoAllocator); else if (mi->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG) ValNo = interval.getNextValue(defIndex, mi->getOperand(1).getReg(), VNInfoAllocator); else ValNo = interval.getNextValue(defIndex, 0, VNInfoAllocator); unsigned killIndex = getInstructionIndex(&mbb->back()) + InstrSlots::NUM; LiveRange LR(defIndex, killIndex, ValNo); interval.addRange(LR); interval.addKill(ValNo, killIndex); ValNo->hasPHIKill = true; DOUT << " +" << LR; } } DOUT << '\n'; } void LiveIntervals::handlePhysicalRegisterDef(MachineBasicBlock *MBB, MachineBasicBlock::iterator mi, unsigned MIIdx, LiveInterval &interval, unsigned SrcReg) { // A physical register cannot be live across basic block, so its // lifetime must end somewhere in its defining basic block. DOUT << "\t\tregister: "; DEBUG(printRegName(interval.reg)); unsigned baseIndex = MIIdx; unsigned start = getDefIndex(baseIndex); unsigned 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) if (lv_->RegisterDefIsDead(mi, interval.reg)) { DOUT << " dead"; end = getDefIndex(start) + 1; 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) while (++mi != MBB->end()) { baseIndex += InstrSlots::NUM; if (lv_->KillsRegister(mi, interval.reg)) { DOUT << " killed"; end = getUseIndex(baseIndex) + 1; goto exit; } else if (lv_->ModifiesRegister(mi, interval.reg)) { // 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) DOUT << " dead"; end = getDefIndex(start) + 1; goto exit; } } // 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. assert(!SrcReg && "physreg was not killed in defining block!"); end = getDefIndex(start) + 1; // It's dead. exit: assert(start < end && "did not find end of interval?"); // Already exists? Extend old live interval. LiveInterval::iterator OldLR = interval.FindLiveRangeContaining(start); VNInfo *ValNo = (OldLR != interval.end()) ? OldLR->valno : interval.getNextValue(start, SrcReg, VNInfoAllocator); LiveRange LR(start, end, ValNo); interval.addRange(LR); interval.addKill(LR.valno, end); DOUT << " +" << LR << '\n'; } void LiveIntervals::handleRegisterDef(MachineBasicBlock *MBB, MachineBasicBlock::iterator MI, unsigned MIIdx, unsigned reg) { if (MRegisterInfo::isVirtualRegister(reg)) handleVirtualRegisterDef(MBB, MI, MIIdx, getOrCreateInterval(reg)); else if (allocatableRegs_[reg]) { unsigned SrcReg, DstReg; if (MI->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG) SrcReg = MI->getOperand(1).getReg(); else if (!tii_->isMoveInstr(*MI, SrcReg, DstReg)) SrcReg = 0; handlePhysicalRegisterDef(MBB, MI, MIIdx, getOrCreateInterval(reg), SrcReg); // Def of a register also defines its sub-registers. for (const unsigned* AS = mri_->getSubRegisters(reg); *AS; ++AS) // Avoid processing some defs more than once. if (!MI->findRegisterDefOperand(*AS)) handlePhysicalRegisterDef(MBB, MI, MIIdx, getOrCreateInterval(*AS), 0); } } void LiveIntervals::handleLiveInRegister(MachineBasicBlock *MBB, unsigned MIIdx, LiveInterval &interval, bool isAlias) { DOUT << "\t\tlivein register: "; DEBUG(printRegName(interval.reg)); // 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(); unsigned baseIndex = MIIdx; unsigned start = baseIndex; unsigned end = start; while (mi != MBB->end()) { if (lv_->KillsRegister(mi, interval.reg)) { DOUT << " killed"; end = getUseIndex(baseIndex) + 1; goto exit; } else if (lv_->ModifiesRegister(mi, interval.reg)) { // 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) DOUT << " dead"; end = getDefIndex(start) + 1; goto exit; } baseIndex += InstrSlots::NUM; ++mi; } exit: // Live-in register might not be used at all. if (end == MIIdx) { if (isAlias) { DOUT << " dead"; end = getDefIndex(MIIdx) + 1; } else { DOUT << " live through"; end = baseIndex; } } LiveRange LR(start, end, interval.getNextValue(start, 0, VNInfoAllocator)); interval.addRange(LR); interval.addKill(LR.valno, end); DOUT << " +" << LR << '\n'; } /// 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() { DOUT << "********** COMPUTING LIVE INTERVALS **********\n" << "********** Function: " << ((Value*)mf_->getFunction())->getName() << '\n'; // Track the index of the current machine instr. unsigned MIIndex = 0; for (MachineFunction::iterator MBBI = mf_->begin(), E = mf_->end(); MBBI != E; ++MBBI) { MachineBasicBlock *MBB = MBBI; DOUT << ((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 = mri_->getSubRegisters(*LI); *AS; ++AS) if (!hasInterval(*AS)) handleLiveInRegister(MBB, MIIndex, getOrCreateInterval(*AS), true); } for (; MI != miEnd; ++MI) { DOUT << MIIndex << "\t" << *MI; // Handle defs. for (int i = MI->getNumOperands() - 1; i >= 0; --i) { MachineOperand &MO = MI->getOperand(i); // handle register defs - build intervals if (MO.isRegister() && MO.getReg() && MO.isDef()) handleRegisterDef(MBB, MI, MIIndex, MO.getReg()); } MIIndex += InstrSlots::NUM; } } } bool LiveIntervals::findLiveInMBBs(const LiveRange &LR, SmallVectorImpl &MBBs) const { std::vector::const_iterator I = std::lower_bound(Idx2MBBMap.begin(), Idx2MBBMap.end(), LR.start); bool ResVal = false; while (I != Idx2MBBMap.end()) { if (LR.end <= I->first) break; MBBs.push_back(I->second); ResVal = true; ++I; } return ResVal; } LiveInterval LiveIntervals::createInterval(unsigned reg) { float Weight = MRegisterInfo::isPhysicalRegister(reg) ? HUGE_VALF : 0.0F; return LiveInterval(reg, Weight); } //===----------------------------------------------------------------------===// // Register allocator hooks. // /// 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) { if (DisableReMat) return false; if (tii_->isTriviallyReMaterializable(MI)) return true; int FrameIdx = 0; if (!tii_->isLoadFromStackSlot(MI, FrameIdx) || !mf_->getFrameInfo()->isFixedObjectIndex(FrameIdx)) return false; // This is a load from fixed stack slot. It can be rematerialized unless it's // re-defined by a two-address instruction. for (LiveInterval::const_vni_iterator i = li.vni_begin(), e = li.vni_end(); i != e; ++i) { const VNInfo *VNI = *i; if (VNI == ValNo) continue; unsigned DefIdx = VNI->def; if (DefIdx == ~1U) continue; // Dead val#. MachineInstr *DefMI = (DefIdx == ~0u) ? NULL : getInstructionFromIndex(DefIdx); if (DefMI && DefMI->isRegReDefinedByTwoAddr(li.reg)) return false; } return true; } /// 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, unsigned InstrIdx, SmallVector &Ops, bool isSS, int Slot, unsigned Reg) { unsigned MRInfo = 0; const TargetInstrDescriptor *TID = MI->getInstrDescriptor(); SmallVector FoldOps; for (unsigned i = 0, e = Ops.size(); i != e; ++i) { unsigned OpIdx = Ops[i]; // FIXME: fold subreg use. if (MI->getOperand(OpIdx).getSubReg()) return false; if (MI->getOperand(OpIdx).isDef()) MRInfo |= (unsigned)VirtRegMap::isMod; else { // Filter out two-address use operand(s). if (TID->getOperandConstraint(OpIdx, TOI::TIED_TO) != -1) { MRInfo = VirtRegMap::isModRef; continue; } MRInfo |= (unsigned)VirtRegMap::isRef; } FoldOps.push_back(OpIdx); } MachineInstr *fmi = isSS ? mri_->foldMemoryOperand(MI, FoldOps, Slot) : mri_->foldMemoryOperand(MI, FoldOps, DefMI); if (fmi) { // Attempt to fold the memory reference into the instruction. If // we can do this, we don't need to insert spill code. if (lv_) lv_->instructionChanged(MI, fmi); else LiveVariables::transferKillDeadInfo(MI, fmi, mri_); MachineBasicBlock &MBB = *MI->getParent(); if (isSS && !mf_->getFrameInfo()->isFixedObjectIndex(Slot)) vrm.virtFolded(Reg, MI, fmi, (VirtRegMap::ModRef)MRInfo); vrm.transferSpillPts(MI, fmi); vrm.transferRestorePts(MI, fmi); mi2iMap_.erase(MI); i2miMap_[InstrIdx /InstrSlots::NUM] = fmi; mi2iMap_[fmi] = InstrIdx; MI = MBB.insert(MBB.erase(MI), fmi); ++numFolds; return true; } return false; } 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; } /// rewriteInstructionForSpills, rewriteInstructionsForSpills - Helper functions /// for addIntervalsForSpills to rewrite uses / defs for the given live range. void LiveIntervals:: rewriteInstructionForSpills(const LiveInterval &li, bool TrySplit, unsigned id, unsigned index, unsigned end, MachineInstr *MI, MachineInstr *ReMatOrigDefMI, MachineInstr *ReMatDefMI, unsigned Slot, int LdSlot, bool isLoad, bool isLoadSS, bool DefIsReMat, bool CanDelete, VirtRegMap &vrm, SSARegMap *RegMap, const TargetRegisterClass* rc, SmallVector &ReMatIds, unsigned &NewVReg, bool &HasDef, bool &HasUse, const LoopInfo *loopInfo, std::map &MBBVRegsMap, std::vector &NewLIs) { RestartInstruction: for (unsigned i = 0; i != MI->getNumOperands(); ++i) { MachineOperand& mop = MI->getOperand(i); if (!mop.isRegister()) continue; unsigned Reg = mop.getReg(); unsigned RegI = Reg; if (Reg == 0 || MRegisterInfo::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) { DOUT << "\t\t\t\tErasing re-materlizable def: "; DOUT << 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; } } // Do not fold load / store here if we are splitting. We'll find an // optimal point to insert a load / store later. if (TryFold) TryFold = !TrySplit && NewVReg == 0; // 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.isRegister()) continue; unsigned RegJ = MOj.getReg(); if (RegJ == 0 || MRegisterInfo::isPhysicalRegister(RegJ)) continue; if (RegJ == RegI) { Ops.push_back(j); HasUse |= MOj.isUse(); HasDef |= MOj.isDef(); } } if (TryFold && tryFoldMemoryOperand(MI, vrm, ReMatDefMI, index, Ops, FoldSS, FoldSlot, Reg)) { // Folding the load/store can completely change the instruction in // unpredictable ways, rescan it from the beginning. HasUse = false; HasDef = false; goto RestartInstruction; } // Create a new virtual register for the spill interval. bool CreatedNewVReg = false; if (NewVReg == 0) { NewVReg = RegMap->createVirtualRegister(rc); vrm.grow(); CreatedNewVReg = true; } mop.setReg(NewVReg); // Reuse NewVReg for other reads. for (unsigned j = 0, e = Ops.size(); j != e; ++j) MI->getOperand(Ops[j]).setReg(NewVReg); if (CreatedNewVReg) { if (DefIsReMat) { vrm.setVirtIsReMaterialized(NewVReg, ReMatDefMI/*, CanDelete*/); if (ReMatIds[id] == VirtRegMap::MAX_STACK_SLOT) { // Each valnum may have its own remat id. ReMatIds[id] = vrm.assignVirtReMatId(NewVReg); } else { vrm.assignVirtReMatId(NewVReg, ReMatIds[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); } // 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), getUseIndex(index)+1, nI.getNextValue(~0U, 0, VNInfoAllocator)); DOUT << " +" << LR; nI.addRange(LR); } else { // Extend the split live interval to this def / use. unsigned End = getUseIndex(index)+1; LiveRange LR(nI.ranges[nI.ranges.size()-1].end, End, nI.getValNumInfo(nI.getNumValNums()-1)); DOUT << " +" << LR; nI.addRange(LR); } } if (HasDef) { LiveRange LR(getDefIndex(index), getStoreIndex(index), nI.getNextValue(~0U, 0, VNInfoAllocator)); DOUT << " +" << LR; nI.addRange(LR); } DOUT << "\t\t\t\tAdded new interval: "; nI.print(DOUT, mri_); DOUT << '\n'; } } bool LiveIntervals::anyKillInMBBAfterIdx(const LiveInterval &li, const VNInfo *VNI, MachineBasicBlock *MBB, unsigned Idx) const { unsigned End = getMBBEndIdx(MBB); for (unsigned j = 0, ee = VNI->kills.size(); j != ee; ++j) { unsigned KillIdx = VNI->kills[j]; if (KillIdx > Idx && KillIdx < End) return true; } return false; } static const VNInfo *findDefinedVNInfo(const LiveInterval &li, unsigned DefIdx) { const VNInfo *VNI = NULL; for (LiveInterval::const_vni_iterator i = li.vni_begin(), e = li.vni_end(); i != e; ++i) if ((*i)->def == DefIdx) { VNI = *i; break; } return VNI; } 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, SSARegMap *RegMap, const TargetRegisterClass* rc, SmallVector &ReMatIds, const LoopInfo *loopInfo, BitVector &SpillMBBs, std::map > &SpillIdxes, BitVector &RestoreMBBs, std::map > &RestoreIdxes, std::map &MBBVRegsMap, std::vector &NewLIs) { unsigned NewVReg = 0; unsigned index = getBaseIndex(I->start); unsigned end = getBaseIndex(I->end-1) + InstrSlots::NUM; bool TrySplitMI = TrySplit && vrm.getPreSplitReg(li.reg) == 0; for (; index != end; index += InstrSlots::NUM) { // skip deleted instructions while (index != end && !getInstructionFromIndex(index)) index += InstrSlots::NUM; if (index == end) break; MachineInstr *MI = getInstructionFromIndex(index); MachineBasicBlock *MBB = MI->getParent(); NewVReg = 0; if (TrySplitMI) { std::map::const_iterator NVI = MBBVRegsMap.find(MBB->getNumber()); if (NVI != MBBVRegsMap.end()) { NewVReg = NVI->second; // One common case: // x = use // ... // ... // def = ... // = use // It's better to start a new interval to avoid artifically // extend the new interval. // FIXME: Too slow? Can we fix it after rewriteInstructionsForSpills? bool MIHasUse = false; bool MIHasDef = false; for (unsigned i = 0; i != MI->getNumOperands(); ++i) { MachineOperand& mop = MI->getOperand(i); if (!mop.isRegister() || mop.getReg() != li.reg) continue; if (mop.isUse()) MIHasUse = true; else MIHasDef = true; } if (MIHasDef && !MIHasUse) { MBBVRegsMap.erase(MBB->getNumber()); NewVReg = 0; } } } bool IsNew = NewVReg == 0; bool HasDef = false; bool HasUse = false; rewriteInstructionForSpills(li, TrySplitMI, I->valno->id, index, end, MI, ReMatOrigDefMI, ReMatDefMI, Slot, LdSlot, isLoad, isLoadSS, DefIsReMat, CanDelete, vrm, RegMap, rc, ReMatIds, NewVReg, HasDef, HasUse, loopInfo, MBBVRegsMap, NewLIs); if (!HasDef && !HasUse) continue; // Update weight of spill interval. LiveInterval &nI = getOrCreateInterval(NewVReg); if (!TrySplitMI) { // 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. unsigned MBBId = MBB->getNumber(); 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 = findDefinedVNInfo(li, getDefIndex(index)); if (VNI) HasKill = anyKillInMBBAfterIdx(li, VNI, MBB, getDefIndex(index)); } std::map >::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 ((int)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 && (int)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) { std::map >::iterator SII = SpillIdxes.find(MBBId); if (SII != SpillIdxes.end() && SII->second.back().vreg == NewVReg && (int)index > SII->second.back().index) // Use(s) following the last def, it's not safe to fold the spill. SII->second.back().canFold = false; std::map >::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->getBasicBlock()); nI.weight += getSpillWeight(HasDef, HasUse, loopDepth); } } bool LiveIntervals::alsoFoldARestore(int Id, int index, unsigned vr, BitVector &RestoreMBBs, std::map > &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, int index, unsigned vr, BitVector &RestoreMBBs, std::map > &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 = -1; } std::vector LiveIntervals:: addIntervalsForSpills(const LiveInterval &li, const LoopInfo *loopInfo, VirtRegMap &vrm) { // Since this is called after the analysis is done we don't know if // LiveVariables is available lv_ = getAnalysisToUpdate(); assert(li.weight != HUGE_VALF && "attempt to spill already spilled interval!"); DOUT << "\t\t\t\tadding intervals for spills for interval: "; li.print(DOUT, mri_); DOUT << '\n'; // Each bit specify whether it a spill is required in the MBB. BitVector SpillMBBs(mf_->getNumBlockIDs()); std::map > SpillIdxes; BitVector RestoreMBBs(mf_->getNumBlockIDs()); std::map > RestoreIdxes; std::map MBBVRegsMap; std::vector NewLIs; SSARegMap *RegMap = mf_->getSSARegMap(); const TargetRegisterClass* rc = RegMap->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); 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->getInstrDescriptor()->Flags & M_LOAD_FLAG)); 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, RegMap, rc, ReMatIds, loopInfo, SpillMBBs, SpillIdxes, RestoreMBBs, RestoreIdxes, MBBVRegsMap, NewLIs); } else { rewriteInstructionsForSpills(li, false, I, NULL, 0, Slot, 0, false, false, false, false, vrm, RegMap, rc, ReMatIds, loopInfo, SpillMBBs, SpillIdxes, RestoreMBBs, RestoreIdxes, MBBVRegsMap, NewLIs); } IsFirstRange = false; } return NewLIs; } bool TrySplit = SplitAtBB && !intervalIsInOneMBB(li); if (SplitLimit != -1 && (int)numSplits >= SplitLimit) TrySplit = false; 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; unsigned DefIdx = VNI->def; if (DefIdx == ~1U) continue; // Dead val#. // Is the def for the val# rematerializable? MachineInstr *ReMatDefMI = (DefIdx == ~0u) ? 0 : getInstructionFromIndex(DefIdx); if (ReMatDefMI && isReMaterializable(li, VNI, ReMatDefMI)) { // 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. vrm must // delete these! ReMatDefs[VN] = ReMatDefMI = ReMatDefMI->clone(); vrm.setVirtIsReMaterialized(li.reg, ReMatDefMI); 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) Slot = vrm.assignVirt2StackSlot(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->getInstrDescriptor()->Flags & M_LOAD_FLAG)); rewriteInstructionsForSpills(li, TrySplit, I, ReMatOrigDefMI, ReMatDefMI, Slot, LdSlot, isLoad, isLoadSS, DefIsReMat, CanDelete, vrm, RegMap, rc, ReMatIds, loopInfo, SpillMBBs, SpillIdxes, RestoreMBBs, RestoreIdxes, MBBVRegsMap, NewLIs); } // Insert spills / restores if we are splitting. if (!TrySplit) return NewLIs; 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) { int index = spills[i].index; unsigned VReg = spills[i].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.isRegister() || 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 > 0) // Also folded uses, do not issue a load. eraseRestoreInfo(Id, index, VReg, RestoreMBBs, RestoreIdxes); } } // Else tell the spiller to issue a spill. if (!Folded) vrm.addSpillPoint(VReg, MI); } 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) { int index = restores[i].index; if (index == -1) continue; unsigned VReg = restores[i].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.isRegister() || 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 (!vrm.isReMaterialized(VReg)) 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->getInstrDescriptor()->Flags & M_LOAD_FLAG)) Folded = tryFoldMemoryOperand(MI, vrm, ReMatDefMI, index, Ops, isLoadSS, LdSlot, VReg); } } // If folding is not possible / failed, then tell the spiller to issue a // load / rematerialization for us. if (!Folded) vrm.addRestorePoint(VReg, MI); } Id = RestoreMBBs.find_next(Id); } // Finalize spill weights. for (unsigned i = 0, e = NewLIs.size(); i != e; ++i) NewLIs[i]->weight /= NewLIs[i]->getSize(); return NewLIs; }