X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FRegAllocLinearScan.cpp;h=fb9745262b7c16435af227a1afd2def9525d24ff;hb=a6b677d209013d82b03fa8da937296b6faae7de0;hp=a26924be6279a3350bfecd10d38b3702bf2180dc;hpb=51cd9d6e073932fcb37f1857c66249d6c7d368ee;p=oota-llvm.git diff --git a/lib/CodeGen/RegAllocLinearScan.cpp b/lib/CodeGen/RegAllocLinearScan.cpp index a26924be627..fb9745262b7 100644 --- a/lib/CodeGen/RegAllocLinearScan.cpp +++ b/lib/CodeGen/RegAllocLinearScan.cpp @@ -12,8 +12,9 @@ //===----------------------------------------------------------------------===// #define DEBUG_TYPE "regalloc" -#include "PhysRegTracker.h" #include "VirtRegMap.h" +#include "VirtRegRewriter.h" +#include "Spiller.h" #include "llvm/Function.h" #include "llvm/CodeGen/LiveIntervalAnalysis.h" #include "llvm/CodeGen/LiveStackAnalysis.h" @@ -26,52 +27,81 @@ #include "llvm/CodeGen/RegisterCoalescer.h" #include "llvm/Target/TargetRegisterInfo.h" #include "llvm/Target/TargetMachine.h" +#include "llvm/Target/TargetOptions.h" #include "llvm/Target/TargetInstrInfo.h" #include "llvm/ADT/EquivalenceClasses.h" +#include "llvm/ADT/SmallSet.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/STLExtras.h" #include "llvm/Support/Debug.h" +#include "llvm/Support/ErrorHandling.h" #include "llvm/Support/Compiler.h" +#include "llvm/Support/ErrorHandling.h" +#include "llvm/Support/raw_ostream.h" #include #include #include #include #include + using namespace llvm; STATISTIC(NumIters , "Number of iterations performed"); STATISTIC(NumBacktracks, "Number of times we had to backtrack"); STATISTIC(NumCoalesce, "Number of copies coalesced"); +STATISTIC(NumDowngrade, "Number of registers downgraded"); static cl::opt NewHeuristic("new-spilling-heuristic", cl::desc("Use new spilling heuristic"), cl::init(false), cl::Hidden); +static cl::opt +PreSplitIntervals("pre-alloc-split", + cl::desc("Pre-register allocation live interval splitting"), + cl::init(false), cl::Hidden); + +static cl::opt +NewSpillFramework("new-spill-framework", + cl::desc("New spilling framework"), + cl::init(false), cl::Hidden); + static RegisterRegAlloc -linearscanRegAlloc("linearscan", " linear scan register allocator", +linearscanRegAlloc("linearscan", "linear scan register allocator", createLinearScanRegisterAllocator); namespace { struct VISIBILITY_HIDDEN RALinScan : public MachineFunctionPass { static char ID; - RALinScan() : MachineFunctionPass((intptr_t)&ID) {} + RALinScan() : MachineFunctionPass(&ID) {} typedef std::pair IntervalPtr; - typedef std::vector IntervalPtrs; + typedef SmallVector IntervalPtrs; private: /// RelatedRegClasses - This structure is built the first time a function is /// compiled, and keeps track of which register classes have registers that /// belong to multiple classes or have aliases that are in other classes. EquivalenceClasses RelatedRegClasses; - std::map OneClassForEachPhysReg; + DenseMap OneClassForEachPhysReg; + + // NextReloadMap - For each register in the map, it maps to the another + // register which is defined by a reload from the same stack slot and + // both reloads are in the same basic block. + DenseMap NextReloadMap; + + // DowngradedRegs - A set of registers which are being "downgraded", i.e. + // un-favored for allocation. + SmallSet DowngradedRegs; + + // DowngradeMap - A map from virtual registers to physical registers being + // downgraded for the virtual registers. + DenseMap DowngradeMap; MachineFunction* mf_; MachineRegisterInfo* mri_; const TargetMachine* tm_; const TargetRegisterInfo* tri_; const TargetInstrInfo* tii_; - MachineRegisterInfo *reginfo_; BitVector allocatableRegs_; LiveIntervals* li_; LiveStacks* ls_; @@ -94,11 +124,19 @@ namespace { IntervalPtrs inactive_; typedef std::priority_queue, + SmallVector, greater_ptr > IntervalHeap; IntervalHeap unhandled_; - std::auto_ptr prt_; - std::auto_ptr vrm_; + + /// regUse_ - Tracks register usage. + SmallVector regUse_; + SmallVector regUseBackUp_; + + /// vrm_ - Tracks register assignments. + VirtRegMap* vrm_; + + std::auto_ptr rewriter_; + std::auto_ptr spiller_; public: @@ -107,14 +145,21 @@ namespace { } virtual void getAnalysisUsage(AnalysisUsage &AU) const { + AU.setPreservesCFG(); AU.addRequired(); + if (StrongPHIElim) + AU.addRequiredID(StrongPHIEliminationID); // Make sure PassManager knows which analyses to make available // to coalescing and which analyses coalescing invalidates. AU.addRequiredTransitive(); + if (PreSplitIntervals) + AU.addRequiredID(PreAllocSplittingID); AU.addRequired(); AU.addPreserved(); AU.addRequired(); AU.addPreserved(); + AU.addRequired(); + AU.addPreserved(); AU.addPreservedID(MachineDominatorsID); MachineFunctionPass::getAnalysisUsage(AU); } @@ -138,10 +183,24 @@ namespace { /// ones to the active list. void processInactiveIntervals(unsigned CurPoint); + /// hasNextReloadInterval - Return the next liveinterval that's being + /// defined by a reload from the same SS as the specified one. + LiveInterval *hasNextReloadInterval(LiveInterval *cur); + + /// DowngradeRegister - Downgrade a register for allocation. + void DowngradeRegister(LiveInterval *li, unsigned Reg); + + /// UpgradeRegister - Upgrade a register for allocation. + void UpgradeRegister(unsigned Reg); + /// assignRegOrStackSlotAtInterval - assign a register if one /// is available, or spill. void assignRegOrStackSlotAtInterval(LiveInterval* cur); + void updateSpillWeights(std::vector &Weights, + unsigned reg, float weight, + const TargetRegisterClass *RC); + /// findIntervalsToSpill - Determine the intervals to spill for the /// specified interval. It's passed the physical registers whose spill /// weight is the lowest among all the registers whose live intervals @@ -161,12 +220,76 @@ namespace { unsigned attemptTrivialCoalescing(LiveInterval &cur, unsigned Reg); /// - /// register handling helpers + /// Register usage / availability tracking helpers. + /// + + void initRegUses() { + regUse_.resize(tri_->getNumRegs(), 0); + regUseBackUp_.resize(tri_->getNumRegs(), 0); + } + + void finalizeRegUses() { +#ifndef NDEBUG + // Verify all the registers are "freed". + bool Error = false; + for (unsigned i = 0, e = tri_->getNumRegs(); i != e; ++i) { + if (regUse_[i] != 0) { + cerr << tri_->getName(i) << " is still in use!\n"; + Error = true; + } + } + if (Error) + llvm_unreachable(0); +#endif + regUse_.clear(); + regUseBackUp_.clear(); + } + + void addRegUse(unsigned physReg) { + assert(TargetRegisterInfo::isPhysicalRegister(physReg) && + "should be physical register!"); + ++regUse_[physReg]; + for (const unsigned* as = tri_->getAliasSet(physReg); *as; ++as) + ++regUse_[*as]; + } + + void delRegUse(unsigned physReg) { + assert(TargetRegisterInfo::isPhysicalRegister(physReg) && + "should be physical register!"); + assert(regUse_[physReg] != 0); + --regUse_[physReg]; + for (const unsigned* as = tri_->getAliasSet(physReg); *as; ++as) { + assert(regUse_[*as] != 0); + --regUse_[*as]; + } + } + + bool isRegAvail(unsigned physReg) const { + assert(TargetRegisterInfo::isPhysicalRegister(physReg) && + "should be physical register!"); + return regUse_[physReg] == 0; + } + + void backUpRegUses() { + regUseBackUp_ = regUse_; + } + + void restoreRegUses() { + regUse_ = regUseBackUp_; + } + + /// + /// Register handling helpers. /// /// getFreePhysReg - return a free physical register for this virtual /// register interval if we have one, otherwise return 0. unsigned getFreePhysReg(LiveInterval* cur); + unsigned getFreePhysReg(LiveInterval* cur, + const TargetRegisterClass *RC, + unsigned MaxInactiveCount, + SmallVector &inactiveCounts, + bool SkipDGRegs); /// assignVirt2StackSlot - assigns this virtual register to a /// stack slot. returns the stack slot @@ -194,17 +317,15 @@ static RegisterPass X("linearscan-regalloc", "Linear Scan Register Allocator"); void RALinScan::ComputeRelatedRegClasses() { - const TargetRegisterInfo &TRI = *tri_; - // First pass, add all reg classes to the union, and determine at least one // reg class that each register is in. bool HasAliases = false; - for (TargetRegisterInfo::regclass_iterator RCI = TRI.regclass_begin(), - E = TRI.regclass_end(); RCI != E; ++RCI) { + for (TargetRegisterInfo::regclass_iterator RCI = tri_->regclass_begin(), + E = tri_->regclass_end(); RCI != E; ++RCI) { RelatedRegClasses.insert(*RCI); for (TargetRegisterClass::iterator I = (*RCI)->begin(), E = (*RCI)->end(); I != E; ++I) { - HasAliases = HasAliases || *TRI.getAliasSet(*I) != 0; + HasAliases = HasAliases || *tri_->getAliasSet(*I) != 0; const TargetRegisterClass *&PRC = OneClassForEachPhysReg[*I]; if (PRC) { @@ -221,10 +342,10 @@ void RALinScan::ComputeRelatedRegClasses() { // belongs to, add info about aliases. We don't need to do this for targets // without register aliases. if (HasAliases) - for (std::map::iterator + for (DenseMap::iterator I = OneClassForEachPhysReg.begin(), E = OneClassForEachPhysReg.end(); I != E; ++I) - for (const unsigned *AS = TRI.getAliasSet(I->first); *AS; ++AS) + for (const unsigned *AS = tri_->getAliasSet(I->first); *AS; ++AS) RelatedRegClasses.unionSets(I->second, OneClassForEachPhysReg[*AS]); } @@ -236,37 +357,54 @@ void RALinScan::ComputeRelatedRegClasses() { /// different register classes or because the coalescer was overly /// conservative. unsigned RALinScan::attemptTrivialCoalescing(LiveInterval &cur, unsigned Reg) { - if ((cur.preference && cur.preference == Reg) || !cur.containsOneValue()) + unsigned Preference = vrm_->getRegAllocPref(cur.reg); + if ((Preference && Preference == Reg) || !cur.containsOneValue()) return Reg; - VNInfo *vni = cur.getValNumInfo(0); - if (!vni->def || vni->def == ~1U || vni->def == ~0U) + VNInfo *vni = cur.begin()->valno; + if (!vni->def || vni->isUnused() || !vni->isDefAccurate()) return Reg; MachineInstr *CopyMI = li_->getInstructionFromIndex(vni->def); - unsigned SrcReg, DstReg; - if (!CopyMI || !tii_->isMoveInstr(*CopyMI, SrcReg, DstReg)) + unsigned SrcReg, DstReg, SrcSubReg, DstSubReg, PhysReg; + if (!CopyMI || + !tii_->isMoveInstr(*CopyMI, SrcReg, DstReg, SrcSubReg, DstSubReg)) return Reg; + PhysReg = SrcReg; if (TargetRegisterInfo::isVirtualRegister(SrcReg)) { if (!vrm_->isAssignedReg(SrcReg)) return Reg; - else - SrcReg = vrm_->getPhys(SrcReg); + PhysReg = vrm_->getPhys(SrcReg); } - if (Reg == SrcReg) + if (Reg == PhysReg) return Reg; - const TargetRegisterClass *RC = reginfo_->getRegClass(cur.reg); - if (!RC->contains(SrcReg)) + const TargetRegisterClass *RC = mri_->getRegClass(cur.reg); + if (!RC->contains(PhysReg)) return Reg; // Try to coalesce. - if (!li_->conflictsWithPhysRegDef(cur, *vrm_, SrcReg)) { - DOUT << "Coalescing: " << cur << " -> " << tri_->getName(SrcReg) + if (!li_->conflictsWithPhysRegDef(cur, *vrm_, PhysReg)) { + DOUT << "Coalescing: " << cur << " -> " << tri_->getName(PhysReg) << '\n'; vrm_->clearVirt(cur.reg); - vrm_->assignVirt2Phys(cur.reg, SrcReg); + vrm_->assignVirt2Phys(cur.reg, PhysReg); + + // Remove unnecessary kills since a copy does not clobber the register. + if (li_->hasInterval(SrcReg)) { + LiveInterval &SrcLI = li_->getInterval(SrcReg); + for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(cur.reg), + E = mri_->reg_end(); I != E; ++I) { + MachineOperand &O = I.getOperand(); + if (!O.isUse() || !O.isKill()) + continue; + MachineInstr *MI = &*I; + if (SrcLI.liveAt(li_->getDefIndex(li_->getInstructionIndex(MI)))) + O.setIsKill(false); + } + } + ++NumCoalesce; - return SrcReg; + return PhysReg; } return Reg; @@ -278,7 +416,6 @@ bool RALinScan::runOnMachineFunction(MachineFunction &fn) { tm_ = &fn.getTarget(); tri_ = tm_->getRegisterInfo(); tii_ = tm_->getInstrInfo(); - reginfo_ = &mf_->getRegInfo(); allocatableRegs_ = tri_->getAllocatableSet(fn); li_ = &getAnalysis(); ls_ = &getAnalysis(); @@ -292,24 +429,36 @@ bool RALinScan::runOnMachineFunction(MachineFunction &fn) { // If this is the first function compiled, compute the related reg classes. if (RelatedRegClasses.empty()) ComputeRelatedRegClasses(); - - if (!prt_.get()) prt_.reset(new PhysRegTracker(*tri_)); - vrm_.reset(new VirtRegMap(*mf_)); - if (!spiller_.get()) spiller_.reset(createSpiller()); + // Also resize register usage trackers. + initRegUses(); + + vrm_ = &getAnalysis(); + if (!rewriter_.get()) rewriter_.reset(createVirtRegRewriter()); + + if (NewSpillFramework) { + spiller_.reset(createSpiller(mf_, li_, ls_, vrm_)); + } + initIntervalSets(); linearScan(); // Rewrite spill code and update the PhysRegsUsed set. - spiller_->runOnMachineFunction(*mf_, *vrm_); - vrm_.reset(); // Free the VirtRegMap + rewriter_->runOnMachineFunction(*mf_, *vrm_, li_); assert(unhandled_.empty() && "Unhandled live intervals remain!"); + + finalizeRegUses(); + fixed_.clear(); active_.clear(); inactive_.clear(); handled_.clear(); + NextReloadMap.clear(); + DowngradedRegs.clear(); + DowngradeMap.clear(); + spiller_.reset(0); return true; } @@ -322,12 +471,14 @@ void RALinScan::initIntervalSets() active_.empty() && inactive_.empty() && "interval sets should be empty on initialization"); + handled_.reserve(li_->getNumIntervals()); + for (LiveIntervals::iterator i = li_->begin(), e = li_->end(); i != e; ++i) { - if (TargetRegisterInfo::isPhysicalRegister(i->second.reg)) { - reginfo_->setPhysRegUsed(i->second.reg); - fixed_.push_back(std::make_pair(&i->second, i->second.begin())); + if (TargetRegisterInfo::isPhysicalRegister(i->second->reg)) { + mri_->setPhysRegUsed(i->second->reg); + fixed_.push_back(std::make_pair(i->second, i->second->begin())); } else - unhandled_.push(&i->second); + unhandled_.push(i->second); } } @@ -335,7 +486,8 @@ void RALinScan::linearScan() { // linear scan algorithm DOUT << "********** LINEAR SCAN **********\n"; - DOUT << "********** Function: " << mf_->getFunction()->getName() << '\n'; + DEBUG(errs() << "********** Function: " + << mf_->getFunction()->getName() << '\n'); DEBUG(printIntervals("fixed", fixed_.begin(), fixed_.end())); @@ -363,7 +515,7 @@ void RALinScan::linearScan() DEBUG(printIntervals("inactive", inactive_.begin(), inactive_.end())); } - // expire any remaining active intervals + // Expire any remaining active intervals while (!active_.empty()) { IntervalPtr &IP = active_.back(); unsigned reg = IP.first->reg; @@ -371,11 +523,11 @@ void RALinScan::linearScan() assert(TargetRegisterInfo::isVirtualRegister(reg) && "Can only allocate virtual registers!"); reg = vrm_->getPhys(reg); - prt_->delRegUse(reg); + delRegUse(reg); active_.pop_back(); } - // expire any remaining inactive intervals + // Expire any remaining inactive intervals DEBUG(for (IntervalPtrs::reverse_iterator i = inactive_.rbegin(); i != inactive_.rend(); ++i) DOUT << "\tinterval " << *i->first << " expired\n"); @@ -385,11 +537,11 @@ void RALinScan::linearScan() MachineFunction::iterator EntryMBB = mf_->begin(); SmallVector LiveInMBBs; for (LiveIntervals::iterator i = li_->begin(), e = li_->end(); i != e; ++i) { - LiveInterval &cur = i->second; + LiveInterval &cur = *i->second; unsigned Reg = 0; bool isPhys = TargetRegisterInfo::isPhysicalRegister(cur.reg); if (isPhys) - Reg = i->second.reg; + Reg = cur.reg; else if (vrm_->isAssignedReg(cur.reg)) Reg = attemptTrivialCoalescing(cur, vrm_->getPhys(cur.reg)); if (!Reg) @@ -397,19 +549,30 @@ void RALinScan::linearScan() // Ignore splited live intervals. if (!isPhys && vrm_->getPreSplitReg(cur.reg)) continue; + for (LiveInterval::Ranges::const_iterator I = cur.begin(), E = cur.end(); I != E; ++I) { const LiveRange &LR = *I; - if (li_->findLiveInMBBs(LR, LiveInMBBs)) { + if (li_->findLiveInMBBs(LR.start, LR.end, LiveInMBBs)) { for (unsigned i = 0, e = LiveInMBBs.size(); i != e; ++i) - if (LiveInMBBs[i] != EntryMBB) + if (LiveInMBBs[i] != EntryMBB) { + assert(TargetRegisterInfo::isPhysicalRegister(Reg) && + "Adding a virtual register to livein set?"); LiveInMBBs[i]->addLiveIn(Reg); + } LiveInMBBs.clear(); } } } DOUT << *vrm_; + + // Look for physical registers that end up not being allocated even though + // register allocator had to spill other registers in its register class. + if (ls_->getNumIntervals() == 0) + return; + if (!vrm_->FindUnusedRegisters(li_)) + return; } /// processActiveIntervals - expire old intervals and move non-overlapping ones @@ -430,7 +593,7 @@ void RALinScan::processActiveIntervals(unsigned CurPoint) assert(TargetRegisterInfo::isVirtualRegister(reg) && "Can only allocate virtual registers!"); reg = vrm_->getPhys(reg); - prt_->delRegUse(reg); + delRegUse(reg); // Pop off the end of the list. active_[i] = active_.back(); @@ -443,7 +606,7 @@ void RALinScan::processActiveIntervals(unsigned CurPoint) assert(TargetRegisterInfo::isVirtualRegister(reg) && "Can only allocate virtual registers!"); reg = vrm_->getPhys(reg); - prt_->delRegUse(reg); + delRegUse(reg); // add to inactive. inactive_.push_back(std::make_pair(Interval, IntervalPos)); @@ -484,7 +647,7 @@ void RALinScan::processInactiveIntervals(unsigned CurPoint) assert(TargetRegisterInfo::isVirtualRegister(reg) && "Can only allocate virtual registers!"); reg = vrm_->getPhys(reg); - prt_->addRegUse(reg); + addRegUse(reg); // add to active active_.push_back(std::make_pair(Interval, IntervalPos)); @@ -501,12 +664,35 @@ void RALinScan::processInactiveIntervals(unsigned CurPoint) /// updateSpillWeights - updates the spill weights of the specifed physical /// register and its weight. -static void updateSpillWeights(std::vector &Weights, - unsigned reg, float weight, - const TargetRegisterInfo *TRI) { +void RALinScan::updateSpillWeights(std::vector &Weights, + unsigned reg, float weight, + const TargetRegisterClass *RC) { + SmallSet Processed; + SmallSet SuperAdded; + SmallVector Supers; Weights[reg] += weight; - for (const unsigned* as = TRI->getAliasSet(reg); *as; ++as) + Processed.insert(reg); + for (const unsigned* as = tri_->getAliasSet(reg); *as; ++as) { Weights[*as] += weight; + Processed.insert(*as); + if (tri_->isSubRegister(*as, reg) && + SuperAdded.insert(*as) && + RC->contains(*as)) { + Supers.push_back(*as); + } + } + + // If the alias is a super-register, and the super-register is in the + // register class we are trying to allocate. Then add the weight to all + // sub-registers of the super-register even if they are not aliases. + // e.g. allocating for GR32, bh is not used, updating bl spill weight. + // bl should get the same spill weight otherwise it will be choosen + // as a spill candidate since spilling bh doesn't make ebx available. + for (unsigned i = 0, e = Supers.size(); i != e; ++i) { + for (const unsigned *sr = tri_->getSubRegisters(Supers[i]); *sr; ++sr) + if (!Processed.count(*sr)) + Weights[*sr] += weight; + } } static @@ -531,19 +717,20 @@ static void RevertVectorIteratorsTo(RALinScan::IntervalPtrs &V, unsigned Point){ /// addStackInterval - Create a LiveInterval for stack if the specified live /// interval has been spilled. static void addStackInterval(LiveInterval *cur, LiveStacks *ls_, - LiveIntervals *li_, float &Weight, - VirtRegMap &vrm_) { + LiveIntervals *li_, + MachineRegisterInfo* mri_, VirtRegMap &vrm_) { int SS = vrm_.getStackSlot(cur->reg); if (SS == VirtRegMap::NO_STACK_SLOT) return; - LiveInterval &SI = ls_->getOrCreateInterval(SS); - SI.weight += Weight; + + const TargetRegisterClass *RC = mri_->getRegClass(cur->reg); + LiveInterval &SI = ls_->getOrCreateInterval(SS, RC); VNInfo *VNI; - if (SI.getNumValNums()) + if (SI.hasAtLeastOneValue()) VNI = SI.getValNumInfo(0); else - VNI = SI.getNextValue(~0U, 0, ls_->getVNInfoAllocator()); + VNI = SI.getNextValue(0, 0, false, ls_->getVNInfoAllocator()); LiveInterval &RI = li_->getInterval(cur->reg); // FIXME: This may be overly conservative. @@ -552,10 +739,10 @@ static void addStackInterval(LiveInterval *cur, LiveStacks *ls_, /// getConflictWeight - Return the number of conflicts between cur /// live interval and defs and uses of Reg weighted by loop depthes. -static float getConflictWeight(LiveInterval *cur, unsigned Reg, - LiveIntervals *li_, - MachineRegisterInfo *mri_, - const MachineLoopInfo *loopInfo) { +static +float getConflictWeight(LiveInterval *cur, unsigned Reg, LiveIntervals *li_, + MachineRegisterInfo *mri_, + const MachineLoopInfo *loopInfo) { float Conflicts = 0; for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(Reg), E = mri_->reg_end(); I != E; ++I) { @@ -651,6 +838,43 @@ static bool weightsAreClose(float w1, float w2) { return (diff / w2) <= 0.05f; // Within 5%. } +LiveInterval *RALinScan::hasNextReloadInterval(LiveInterval *cur) { + DenseMap::iterator I = NextReloadMap.find(cur->reg); + if (I == NextReloadMap.end()) + return 0; + return &li_->getInterval(I->second); +} + +void RALinScan::DowngradeRegister(LiveInterval *li, unsigned Reg) { + bool isNew = DowngradedRegs.insert(Reg); + isNew = isNew; // Silence compiler warning. + assert(isNew && "Multiple reloads holding the same register?"); + DowngradeMap.insert(std::make_pair(li->reg, Reg)); + for (const unsigned *AS = tri_->getAliasSet(Reg); *AS; ++AS) { + isNew = DowngradedRegs.insert(*AS); + isNew = isNew; // Silence compiler warning. + assert(isNew && "Multiple reloads holding the same register?"); + DowngradeMap.insert(std::make_pair(li->reg, *AS)); + } + ++NumDowngrade; +} + +void RALinScan::UpgradeRegister(unsigned Reg) { + if (Reg) { + DowngradedRegs.erase(Reg); + for (const unsigned *AS = tri_->getAliasSet(Reg); *AS; ++AS) + DowngradedRegs.erase(*AS); + } +} + +namespace { + struct LISorter { + bool operator()(LiveInterval* A, LiveInterval* B) { + return A->beginNumber() < B->beginNumber(); + } + }; +} + /// assignRegOrStackSlotAtInterval - assign a register if one is available, or /// spill. void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) @@ -658,9 +882,9 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) DOUT << "\tallocating current interval: "; // This is an implicitly defined live interval, just assign any register. - const TargetRegisterClass *RC = reginfo_->getRegClass(cur->reg); + const TargetRegisterClass *RC = mri_->getRegClass(cur->reg); if (cur->empty()) { - unsigned physReg = cur->preference; + unsigned physReg = vrm_->getRegAllocPref(cur->reg); if (!physReg) physReg = *RC->allocation_order_begin(*mf_); DOUT << tri_->getName(physReg) << '\n'; @@ -669,48 +893,55 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) return; } - PhysRegTracker backupPrt = *prt_; + backUpRegUses(); std::vector > SpillWeightsToAdd; unsigned StartPosition = cur->beginNumber(); const TargetRegisterClass *RCLeader = RelatedRegClasses.getLeaderValue(RC); - // If this live interval is defined by a move instruction and its source is - // assigned a physical register that is compatible with the target register - // class, then we should try to assign it the same register. + // If start of this live interval is defined by a move instruction and its + // source is assigned a physical register that is compatible with the target + // register class, then we should try to assign it the same register. // This can happen when the move is from a larger register class to a smaller // one, e.g. X86::mov32to32_. These move instructions are not coalescable. - if (!cur->preference && cur->containsOneValue()) { - VNInfo *vni = cur->getValNumInfo(0); - if (vni->def && vni->def != ~1U && vni->def != ~0U) { + if (!vrm_->getRegAllocPref(cur->reg) && cur->hasAtLeastOneValue()) { + VNInfo *vni = cur->begin()->valno; + if (vni->def && !vni->isUnused() && vni->isDefAccurate()) { MachineInstr *CopyMI = li_->getInstructionFromIndex(vni->def); - unsigned SrcReg, DstReg; - if (CopyMI && tii_->isMoveInstr(*CopyMI, SrcReg, DstReg)) { + unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; + if (CopyMI && + tii_->isMoveInstr(*CopyMI, SrcReg, DstReg, SrcSubReg, DstSubReg)) { unsigned Reg = 0; if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) Reg = SrcReg; else if (vrm_->isAssignedReg(SrcReg)) Reg = vrm_->getPhys(SrcReg); - if (Reg && allocatableRegs_[Reg] && RC->contains(Reg)) - cur->preference = Reg; + if (Reg) { + if (SrcSubReg) + Reg = tri_->getSubReg(Reg, SrcSubReg); + if (DstSubReg) + Reg = tri_->getMatchingSuperReg(Reg, DstSubReg, RC); + if (Reg && allocatableRegs_[Reg] && RC->contains(Reg)) + mri_->setRegAllocationHint(cur->reg, 0, Reg); + } } } } - // for every interval in inactive we overlap with, mark the + // For every interval in inactive we overlap with, mark the // register as not free and update spill weights. for (IntervalPtrs::const_iterator i = inactive_.begin(), e = inactive_.end(); i != e; ++i) { unsigned Reg = i->first->reg; assert(TargetRegisterInfo::isVirtualRegister(Reg) && "Can only allocate virtual registers!"); - const TargetRegisterClass *RegRC = reginfo_->getRegClass(Reg); + const TargetRegisterClass *RegRC = mri_->getRegClass(Reg); // If this is not in a related reg class to the register we're allocating, // don't check it. if (RelatedRegClasses.getLeaderValue(RegRC) == RCLeader && cur->overlapsFrom(*i->first, i->second-1)) { Reg = vrm_->getPhys(Reg); - prt_->addRegUse(Reg); + addRegUse(Reg); SpillWeightsToAdd.push_back(std::make_pair(Reg, i->first->weight)); } } @@ -752,7 +983,7 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) // Okay, the register picked by our speculative getFreePhysReg call turned // out to be in use. Actually add all of the conflicting fixed registers to - // prt so we can do an accurate query. + // regUse_ so we can do an accurate query. if (ConflictsWithFixed) { // For every interval in fixed we overlap with, mark the register as not // free and update spill weights. @@ -769,13 +1000,13 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) --II; if (cur->overlapsFrom(*I, II)) { unsigned reg = I->reg; - prt_->addRegUse(reg); + addRegUse(reg); SpillWeightsToAdd.push_back(std::make_pair(reg, I->weight)); } } } - // Using the newly updated prt_ object, which includes conflicts in the + // Using the newly updated regUse_ object, which includes conflicts in the // future, see if there are any registers available. physReg = getFreePhysReg(cur); } @@ -783,17 +1014,26 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) // Restore the physical register tracker, removing information about the // future. - *prt_ = backupPrt; + restoreRegUses(); - // if we find a free register, we are done: assign this virtual to + // If we find a free register, we are done: assign this virtual to // the free physical register and add this interval to the active // list. if (physReg) { DOUT << tri_->getName(physReg) << '\n'; vrm_->assignVirt2Phys(cur->reg, physReg); - prt_->addRegUse(physReg); + addRegUse(physReg); active_.push_back(std::make_pair(cur, cur->begin())); handled_.push_back(cur); + + // "Upgrade" the physical register since it has been allocated. + UpgradeRegister(physReg); + if (LiveInterval *NextReloadLI = hasNextReloadInterval(cur)) { + // "Downgrade" physReg to try to keep physReg from being allocated until + // the next reload from the same SS is allocated. + mri_->setRegAllocationHint(NextReloadLI->reg, 0, physReg); + DowngradeRegister(cur, physReg); + } return; } DOUT << "no free registers\n"; @@ -802,7 +1042,7 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) std::vector SpillWeights(tri_->getNumRegs(), 0.0f); for (std::vector >::iterator I = SpillWeightsToAdd.begin(), E = SpillWeightsToAdd.end(); I != E; ++I) - updateSpillWeights(SpillWeights, I->first, I->second, tri_); + updateSpillWeights(SpillWeights, I->first, I->second, RC); // for each interval in active, update spill weights. for (IntervalPtrs::const_iterator i = active_.begin(), e = active_.end(); @@ -811,14 +1051,14 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) assert(TargetRegisterInfo::isVirtualRegister(reg) && "Can only allocate virtual registers!"); reg = vrm_->getPhys(reg); - updateSpillWeights(SpillWeights, reg, i->first->weight, tri_); + updateSpillWeights(SpillWeights, reg, i->first->weight, RC); } DOUT << "\tassigning stack slot at interval "<< *cur << ":\n"; // Find a register to spill. float minWeight = HUGE_VALF; - unsigned minReg = 0; /*cur->preference*/; // Try the preferred register first. + unsigned minReg = 0; bool Found = false; std::vector > RegsWeights; @@ -851,9 +1091,26 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) if (minWeight == HUGE_VALF) { // All registers must have inf weight. Just grab one! minReg = BestPhysReg ? BestPhysReg : *RC->allocation_order_begin(*mf_); - if (cur->weight == HUGE_VALF || cur->getSize() == 1) + if (cur->weight == HUGE_VALF || + li_->getApproximateInstructionCount(*cur) == 0) { // Spill a physical register around defs and uses. - li_->spillPhysRegAroundRegDefsUses(*cur, minReg, *vrm_); + if (li_->spillPhysRegAroundRegDefsUses(*cur, minReg, *vrm_)) { + // spillPhysRegAroundRegDefsUses may have invalidated iterator stored + // in fixed_. Reset them. + for (unsigned i = 0, e = fixed_.size(); i != e; ++i) { + IntervalPtr &IP = fixed_[i]; + LiveInterval *I = IP.first; + if (I->reg == minReg || tri_->isSubRegister(minReg, I->reg)) + IP.second = I->advanceTo(I->begin(), StartPosition); + } + + DowngradedRegs.clear(); + assignRegOrStackSlotAtInterval(cur); + } else { + llvm_report_error("Ran out of registers during register allocation!"); + } + return; + } } // Find up to 3 registers to consider as spill candidates. @@ -869,29 +1126,60 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) DOUT << tri_->getName(RegsWeights[i].first) << " (" << RegsWeights[i].second << ")\n"); - // if the current has the minimum weight, we need to spill it and + // If the current has the minimum weight, we need to spill it and // add any added intervals back to unhandled, and restart // linearscan. if (cur->weight != HUGE_VALF && cur->weight <= minWeight) { DOUT << "\t\t\tspilling(c): " << *cur << '\n'; - float SSWeight; - std::vector added = - li_->addIntervalsForSpills(*cur, loopInfo, *vrm_, SSWeight); - addStackInterval(cur, ls_, li_, SSWeight, *vrm_); + SmallVector spillIs; + std::vector added; + + if (!NewSpillFramework) { + added = li_->addIntervalsForSpills(*cur, spillIs, loopInfo, *vrm_); + } else { + added = spiller_->spill(cur); + } + + std::sort(added.begin(), added.end(), LISorter()); + addStackInterval(cur, ls_, li_, mri_, *vrm_); if (added.empty()) return; // Early exit if all spills were folded. - // Merge added with unhandled. Note that we know that - // addIntervalsForSpills returns intervals sorted by their starting + // Merge added with unhandled. Note that we have already sorted + // intervals returned by addIntervalsForSpills by their starting // point. - for (unsigned i = 0, e = added.size(); i != e; ++i) - unhandled_.push(added[i]); + // This also update the NextReloadMap. That is, it adds mapping from a + // register defined by a reload from SS to the next reload from SS in the + // same basic block. + MachineBasicBlock *LastReloadMBB = 0; + LiveInterval *LastReload = 0; + int LastReloadSS = VirtRegMap::NO_STACK_SLOT; + for (unsigned i = 0, e = added.size(); i != e; ++i) { + LiveInterval *ReloadLi = added[i]; + if (ReloadLi->weight == HUGE_VALF && + li_->getApproximateInstructionCount(*ReloadLi) == 0) { + unsigned ReloadIdx = ReloadLi->beginNumber(); + MachineBasicBlock *ReloadMBB = li_->getMBBFromIndex(ReloadIdx); + int ReloadSS = vrm_->getStackSlot(ReloadLi->reg); + if (LastReloadMBB == ReloadMBB && LastReloadSS == ReloadSS) { + // Last reload of same SS is in the same MBB. We want to try to + // allocate both reloads the same register and make sure the reg + // isn't clobbered in between if at all possible. + assert(LastReload->beginNumber() < ReloadIdx); + NextReloadMap.insert(std::make_pair(LastReload->reg, ReloadLi->reg)); + } + LastReloadMBB = ReloadMBB; + LastReload = ReloadLi; + LastReloadSS = ReloadSS; + } + unhandled_.push(ReloadLi); + } return; } ++NumBacktracks; - // push the current interval back to unhandled since we are going + // Push the current interval back to unhandled since we are going // to re-run at least this iteration. Since we didn't modify it it // should go back right in the front of the list unhandled_.push(cur); @@ -912,7 +1200,8 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) // The earliest start of a Spilled interval indicates up to where // in handled we need to roll back - unsigned earliestStart = cur->beginNumber(); + + LiveInterval *earliestStartInterval = cur; // Spill live intervals of virtual regs mapped to the physical register we // want to clear (and its aliases). We only spill those that overlap with the @@ -924,15 +1213,23 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) LiveInterval *sli = spillIs.back(); spillIs.pop_back(); DOUT << "\t\t\tspilling(a): " << *sli << '\n'; - earliestStart = std::min(earliestStart, sli->beginNumber()); - float SSWeight; - std::vector newIs = - li_->addIntervalsForSpills(*sli, loopInfo, *vrm_, SSWeight); - addStackInterval(sli, ls_, li_, SSWeight, *vrm_); + earliestStartInterval = + (earliestStartInterval->beginNumber() < sli->beginNumber()) ? + earliestStartInterval : sli; + + std::vector newIs; + if (!NewSpillFramework) { + newIs = li_->addIntervalsForSpills(*sli, spillIs, loopInfo, *vrm_); + } else { + newIs = spiller_->spill(sli); + } + addStackInterval(sli, ls_, li_, mri_, *vrm_); std::copy(newIs.begin(), newIs.end(), std::back_inserter(added)); spilled.insert(sli->reg); } + unsigned earliestStart = earliestStartInterval->beginNumber(); + DOUT << "\t\trolling back to: " << earliestStart << '\n'; // Scan handled in reverse order up to the earliest start of a @@ -947,14 +1244,14 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) handled_.pop_back(); // When undoing a live interval allocation we must know if it is active or - // inactive to properly update the PhysRegTracker and the VirtRegMap. + // inactive to properly update regUse_ and the VirtRegMap. IntervalPtrs::iterator it; if ((it = FindIntervalInVector(active_, i)) != active_.end()) { active_.erase(it); assert(!TargetRegisterInfo::isPhysicalRegister(i->reg)); if (!spilled.count(i->reg)) unhandled_.push(i); - prt_->delRegUse(vrm_->getPhys(i->reg)); + delRegUse(vrm_->getPhys(i->reg)); vrm_->clearVirt(i->reg); } else if ((it = FindIntervalInVector(inactive_, i)) != inactive_.end()) { inactive_.erase(it); @@ -969,10 +1266,15 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) unhandled_.push(i); } - // It interval has a preference, it must be defined by a copy. Clear the - // preference now since the source interval allocation may have been undone - // as well. - i->preference = 0; + DenseMap::iterator ii = DowngradeMap.find(i->reg); + if (ii == DowngradeMap.end()) + // It interval has a preference, it must be defined by a copy. Clear the + // preference now since the source interval allocation may have been + // undone as well. + mri_->setRegAllocationHint(i->reg, 0, 0); + else { + UpgradeRegister(ii->second); + } } // Rewind the iterators in the active, inactive, and fixed lists back to the @@ -981,7 +1283,7 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) RevertVectorIteratorsTo(inactive_, earliestStart); RevertVectorIteratorsTo(fixed_, earliestStart); - // scan the rest and undo each interval that expired after t and + // Scan the rest and undo each interval that expired after t and // insert it in active (the next iteration of the algorithm will // put it in inactive if required) for (unsigned i = 0, e = handled_.size(); i != e; ++i) { @@ -991,13 +1293,99 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) DOUT << "\t\t\tundo changes for: " << *HI << '\n'; active_.push_back(std::make_pair(HI, HI->begin())); assert(!TargetRegisterInfo::isPhysicalRegister(HI->reg)); - prt_->addRegUse(vrm_->getPhys(HI->reg)); + addRegUse(vrm_->getPhys(HI->reg)); + } + } + + // Merge added with unhandled. + // This also update the NextReloadMap. That is, it adds mapping from a + // register defined by a reload from SS to the next reload from SS in the + // same basic block. + MachineBasicBlock *LastReloadMBB = 0; + LiveInterval *LastReload = 0; + int LastReloadSS = VirtRegMap::NO_STACK_SLOT; + std::sort(added.begin(), added.end(), LISorter()); + for (unsigned i = 0, e = added.size(); i != e; ++i) { + LiveInterval *ReloadLi = added[i]; + if (ReloadLi->weight == HUGE_VALF && + li_->getApproximateInstructionCount(*ReloadLi) == 0) { + unsigned ReloadIdx = ReloadLi->beginNumber(); + MachineBasicBlock *ReloadMBB = li_->getMBBFromIndex(ReloadIdx); + int ReloadSS = vrm_->getStackSlot(ReloadLi->reg); + if (LastReloadMBB == ReloadMBB && LastReloadSS == ReloadSS) { + // Last reload of same SS is in the same MBB. We want to try to + // allocate both reloads the same register and make sure the reg + // isn't clobbered in between if at all possible. + assert(LastReload->beginNumber() < ReloadIdx); + NextReloadMap.insert(std::make_pair(LastReload->reg, ReloadLi->reg)); + } + LastReloadMBB = ReloadMBB; + LastReload = ReloadLi; + LastReloadSS = ReloadSS; + } + unhandled_.push(ReloadLi); + } +} + +unsigned RALinScan::getFreePhysReg(LiveInterval* cur, + const TargetRegisterClass *RC, + unsigned MaxInactiveCount, + SmallVector &inactiveCounts, + bool SkipDGRegs) { + unsigned FreeReg = 0; + unsigned FreeRegInactiveCount = 0; + + std::pair Hint = mri_->getRegAllocationHint(cur->reg); + // Resolve second part of the hint (if possible) given the current allocation. + unsigned physReg = Hint.second; + if (physReg && + TargetRegisterInfo::isVirtualRegister(physReg) && vrm_->hasPhys(physReg)) + physReg = vrm_->getPhys(physReg); + + TargetRegisterClass::iterator I, E; + tie(I, E) = tri_->getAllocationOrder(RC, Hint.first, physReg, *mf_); + assert(I != E && "No allocatable register in this register class!"); + + // Scan for the first available register. + for (; I != E; ++I) { + unsigned Reg = *I; + // Ignore "downgraded" registers. + if (SkipDGRegs && DowngradedRegs.count(Reg)) + continue; + if (isRegAvail(Reg)) { + FreeReg = Reg; + if (FreeReg < inactiveCounts.size()) + FreeRegInactiveCount = inactiveCounts[FreeReg]; + else + FreeRegInactiveCount = 0; + break; + } + } + + // If there are no free regs, or if this reg has the max inactive count, + // return this register. + if (FreeReg == 0 || FreeRegInactiveCount == MaxInactiveCount) + return FreeReg; + + // Continue scanning the registers, looking for the one with the highest + // inactive count. Alkis found that this reduced register pressure very + // slightly on X86 (in rev 1.94 of this file), though this should probably be + // reevaluated now. + for (; I != E; ++I) { + unsigned Reg = *I; + // Ignore "downgraded" registers. + if (SkipDGRegs && DowngradedRegs.count(Reg)) + continue; + if (isRegAvail(Reg) && Reg < inactiveCounts.size() && + FreeRegInactiveCount < inactiveCounts[Reg]) { + FreeReg = Reg; + FreeRegInactiveCount = inactiveCounts[Reg]; + if (FreeRegInactiveCount == MaxInactiveCount) + break; // We found the one with the max inactive count. } } - // merge added with unhandled - for (unsigned i = 0, e = added.size(); i != e; ++i) - unhandled_.push(added[i]); + return FreeReg; } /// getFreePhysReg - return a free physical register for this virtual register @@ -1006,7 +1394,7 @@ unsigned RALinScan::getFreePhysReg(LiveInterval *cur) { SmallVector inactiveCounts; unsigned MaxInactiveCount = 0; - const TargetRegisterClass *RC = reginfo_->getRegClass(cur->reg); + const TargetRegisterClass *RC = mri_->getRegClass(cur->reg); const TargetRegisterClass *RCLeader = RelatedRegClasses.getLeaderValue(RC); for (IntervalPtrs::iterator i = inactive_.begin(), e = inactive_.end(); @@ -1017,7 +1405,7 @@ unsigned RALinScan::getFreePhysReg(LiveInterval *cur) { // If this is not in a related reg class to the register we're allocating, // don't check it. - const TargetRegisterClass *RegRC = reginfo_->getRegClass(reg); + const TargetRegisterClass *RegRC = mri_->getRegClass(reg); if (RelatedRegClasses.getLeaderValue(RegRC) == RCLeader) { reg = vrm_->getPhys(reg); if (inactiveCounts.size() <= reg) @@ -1027,55 +1415,23 @@ unsigned RALinScan::getFreePhysReg(LiveInterval *cur) { } } - unsigned FreeReg = 0; - unsigned FreeRegInactiveCount = 0; - // If copy coalescer has assigned a "preferred" register, check if it's // available first. - if (cur->preference) { - if (prt_->isRegAvail(cur->preference)) { - DOUT << "\t\tassigned the preferred register: " - << tri_->getName(cur->preference) << "\n"; - return cur->preference; - } else - DOUT << "\t\tunable to assign the preferred register: " - << tri_->getName(cur->preference) << "\n"; + unsigned Preference = vrm_->getRegAllocPref(cur->reg); + if (Preference) { + DOUT << "(preferred: " << tri_->getName(Preference) << ") "; + if (isRegAvail(Preference) && + RC->contains(Preference)) + return Preference; } - // Scan for the first available register. - TargetRegisterClass::iterator I = RC->allocation_order_begin(*mf_); - TargetRegisterClass::iterator E = RC->allocation_order_end(*mf_); - assert(I != E && "No allocatable register in this register class!"); - for (; I != E; ++I) - if (prt_->isRegAvail(*I)) { - FreeReg = *I; - if (FreeReg < inactiveCounts.size()) - FreeRegInactiveCount = inactiveCounts[FreeReg]; - else - FreeRegInactiveCount = 0; - break; - } - - // If there are no free regs, or if this reg has the max inactive count, - // return this register. - if (FreeReg == 0 || FreeRegInactiveCount == MaxInactiveCount) return FreeReg; - - // Continue scanning the registers, looking for the one with the highest - // inactive count. Alkis found that this reduced register pressure very - // slightly on X86 (in rev 1.94 of this file), though this should probably be - // reevaluated now. - for (; I != E; ++I) { - unsigned Reg = *I; - if (prt_->isRegAvail(Reg) && Reg < inactiveCounts.size() && - FreeRegInactiveCount < inactiveCounts[Reg]) { - FreeReg = Reg; - FreeRegInactiveCount = inactiveCounts[Reg]; - if (FreeRegInactiveCount == MaxInactiveCount) - break; // We found the one with the max inactive count. - } + if (!DowngradedRegs.empty()) { + unsigned FreeReg = getFreePhysReg(cur, RC, MaxInactiveCount, inactiveCounts, + true); + if (FreeReg) + return FreeReg; } - - return FreeReg; + return getFreePhysReg(cur, RC, MaxInactiveCount, inactiveCounts, false); } FunctionPass* llvm::createLinearScanRegisterAllocator() {