X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FMachineLICM.cpp;h=443fc2d97bdf591464d26f1918250b9e75a7d464;hb=b43034d700004e1fec3ddf177e21ac89478bcc6c;hp=fbf21139ac4c7ffc6568221ffbd16ab97820269b;hpb=c475c3608a5f0fc0c6bd43da04ae786649690070;p=oota-llvm.git diff --git a/lib/CodeGen/MachineLICM.cpp b/lib/CodeGen/MachineLICM.cpp index fbf21139ac4..443fc2d97bd 100644 --- a/lib/CodeGen/MachineLICM.cpp +++ b/lib/CodeGen/MachineLICM.cpp @@ -23,52 +23,145 @@ #define DEBUG_TYPE "machine-licm" #include "llvm/CodeGen/Passes.h" #include "llvm/CodeGen/MachineDominators.h" +#include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineLoopInfo.h" +#include "llvm/CodeGen/MachineMemOperand.h" #include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/CodeGen/PseudoSourceValue.h" +#include "llvm/Target/TargetLowering.h" #include "llvm/Target/TargetRegisterInfo.h" #include "llvm/Target/TargetInstrInfo.h" +#include "llvm/Target/TargetInstrItineraries.h" #include "llvm/Target/TargetMachine.h" +#include "llvm/Analysis/AliasAnalysis.h" +#include "llvm/ADT/DenseMap.h" +#include "llvm/ADT/SmallSet.h" #include "llvm/ADT/Statistic.h" -#include "llvm/Support/CommandLine.h" -#include "llvm/Support/Compiler.h" #include "llvm/Support/Debug.h" +#include "llvm/Support/raw_ostream.h" using namespace llvm; -STATISTIC(NumHoisted, "Number of machine instructions hoisted out of loops"); +STATISTIC(NumHoisted, + "Number of machine instructions hoisted out of loops"); +STATISTIC(NumLowRP, + "Number of instructions hoisted in low reg pressure situation"); +STATISTIC(NumHighLatency, + "Number of high latency instructions hoisted"); +STATISTIC(NumCSEed, + "Number of hoisted machine instructions CSEed"); +STATISTIC(NumPostRAHoisted, + "Number of machine instructions hoisted out of loops post regalloc"); namespace { - class VISIBILITY_HIDDEN MachineLICM : public MachineFunctionPass { + class MachineLICM : public MachineFunctionPass { + bool PreRegAlloc; + const TargetMachine *TM; const TargetInstrInfo *TII; + const TargetLowering *TLI; + const TargetRegisterInfo *TRI; + const MachineFrameInfo *MFI; + MachineRegisterInfo *MRI; + const InstrItineraryData *InstrItins; // Various analyses that we use... - MachineLoopInfo *LI; // Current MachineLoopInfo + AliasAnalysis *AA; // Alias analysis info. + MachineLoopInfo *MLI; // Current MachineLoopInfo MachineDominatorTree *DT; // Machine dominator tree for the cur loop - MachineRegisterInfo *RegInfo; // Machine register information // State that is updated as we process loops bool Changed; // True if a loop is changed. + bool FirstInLoop; // True if it's the first LICM in the loop. MachineLoop *CurLoop; // The current loop we are working on. MachineBasicBlock *CurPreheader; // The preheader for CurLoop. + + BitVector AllocatableSet; + + // Track 'estimated' register pressure. + SmallSet RegSeen; + SmallVector RegPressure; + + // Register pressure "limit" per register class. If the pressure + // is higher than the limit, then it's considered high. + SmallVector RegLimit; + + // Register pressure on path leading from loop preheader to current BB. + SmallVector, 16> BackTrace; + + // For each opcode, keep a list of potential CSE instructions. + DenseMap > CSEMap; + public: static char ID; // Pass identification, replacement for typeid - MachineLICM() : MachineFunctionPass(&ID) {} + MachineLICM() : + MachineFunctionPass(ID), PreRegAlloc(true) { + initializeMachineLICMPass(*PassRegistry::getPassRegistry()); + } + + explicit MachineLICM(bool PreRA) : + MachineFunctionPass(ID), PreRegAlloc(PreRA) { + initializeMachineLICMPass(*PassRegistry::getPassRegistry()); + } virtual bool runOnMachineFunction(MachineFunction &MF); const char *getPassName() const { return "Machine Instruction LICM"; } - // FIXME: Loop preheaders? virtual void getAnalysisUsage(AnalysisUsage &AU) const { - AU.setPreservesCFG(); AU.addRequired(); AU.addRequired(); + AU.addRequired(); AU.addPreserved(); AU.addPreserved(); MachineFunctionPass::getAnalysisUsage(AU); } + + virtual void releaseMemory() { + RegSeen.clear(); + RegPressure.clear(); + RegLimit.clear(); + BackTrace.clear(); + for (DenseMap >::iterator + CI = CSEMap.begin(), CE = CSEMap.end(); CI != CE; ++CI) + CI->second.clear(); + CSEMap.clear(); + } + private: + /// CandidateInfo - Keep track of information about hoisting candidates. + struct CandidateInfo { + MachineInstr *MI; + unsigned Def; + int FI; + CandidateInfo(MachineInstr *mi, unsigned def, int fi) + : MI(mi), Def(def), FI(fi) {} + }; + + /// HoistRegionPostRA - Walk the specified region of the CFG and hoist loop + /// invariants out to the preheader. + void HoistRegionPostRA(); + + /// HoistPostRA - When an instruction is found to only use loop invariant + /// operands that is safe to hoist, this instruction is called to do the + /// dirty work. + void HoistPostRA(MachineInstr *MI, unsigned Def); + + /// ProcessMI - Examine the instruction for potentai LICM candidate. Also + /// gather register def and frame object update information. + void ProcessMI(MachineInstr *MI, unsigned *PhysRegDefs, + SmallSet &StoredFIs, + SmallVector &Candidates); + + /// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the + /// current loop. + void AddToLiveIns(unsigned Reg); + + /// IsLICMCandidate - Returns true if the instruction may be a suitable + /// candidate for LICM. e.g. If the instruction is a call, then it's + /// obviously not safe to hoist it. + bool IsLICMCandidate(MachineInstr &I); + /// IsLoopInvariantInst - Returns true if the instruction is loop /// invariant. I.e., all virtual register operands are defined outside of /// the loop, physical registers aren't accessed (explicitly or implicitly), @@ -76,74 +169,377 @@ namespace { /// bool IsLoopInvariantInst(MachineInstr &I); + /// HasHighOperandLatency - Compute operand latency between a def of 'Reg' + /// and an use in the current loop, return true if the target considered + /// it 'high'. + bool HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx, + unsigned Reg) const; + + bool IsCheapInstruction(MachineInstr &MI) const; + + /// CanCauseHighRegPressure - Visit BBs from header to current BB, + /// check if hoisting an instruction of the given cost matrix can cause high + /// register pressure. + bool CanCauseHighRegPressure(DenseMap &Cost); + + /// UpdateBackTraceRegPressure - Traverse the back trace from header to + /// the current block and update their register pressures to reflect the + /// effect of hoisting MI from the current block to the preheader. + void UpdateBackTraceRegPressure(const MachineInstr *MI); + + /// IsProfitableToHoist - Return true if it is potentially profitable to + /// hoist the given loop invariant. + bool IsProfitableToHoist(MachineInstr &MI); + /// HoistRegion - Walk the specified region of the CFG (defined by all /// blocks dominated by the specified block, and that are in the current /// loop) in depth first order w.r.t the DominatorTree. This allows us to /// visit definitions before uses, allowing us to hoist a loop body in one /// pass without iteration. /// - void HoistRegion(MachineDomTreeNode *N); + void HoistRegion(MachineDomTreeNode *N, bool IsHeader = false); + + /// InitRegPressure - Find all virtual register references that are liveout + /// of the preheader to initialize the starting "register pressure". Note + /// this does not count live through (livein but not used) registers. + void InitRegPressure(MachineBasicBlock *BB); + + /// UpdateRegPressure - Update estimate of register pressure after the + /// specified instruction. + void UpdateRegPressure(const MachineInstr *MI); + + /// ExtractHoistableLoad - Unfold a load from the given machineinstr if + /// the load itself could be hoisted. Return the unfolded and hoistable + /// load, or null if the load couldn't be unfolded or if it wouldn't + /// be hoistable. + MachineInstr *ExtractHoistableLoad(MachineInstr *MI); + + /// LookForDuplicate - Find an instruction amount PrevMIs that is a + /// duplicate of MI. Return this instruction if it's found. + const MachineInstr *LookForDuplicate(const MachineInstr *MI, + std::vector &PrevMIs); + + /// EliminateCSE - Given a LICM'ed instruction, look for an instruction on + /// the preheader that compute the same value. If it's found, do a RAU on + /// with the definition of the existing instruction rather than hoisting + /// the instruction to the preheader. + bool EliminateCSE(MachineInstr *MI, + DenseMap >::iterator &CI); /// Hoist - When an instruction is found to only use loop invariant operands /// that is safe to hoist, this instruction is called to do the dirty work. - /// - void Hoist(MachineInstr &MI); + /// It returns true if the instruction is hoisted. + bool Hoist(MachineInstr *MI, MachineBasicBlock *Preheader); + + /// InitCSEMap - Initialize the CSE map with instructions that are in the + /// current loop preheader that may become duplicates of instructions that + /// are hoisted out of the loop. + void InitCSEMap(MachineBasicBlock *BB); + + /// getCurPreheader - Get the preheader for the current loop, splitting + /// a critical edge if needed. + MachineBasicBlock *getCurPreheader(); }; } // end anonymous namespace char MachineLICM::ID = 0; -static RegisterPass -X("machinelicm", "Machine Loop Invariant Code Motion"); +INITIALIZE_PASS_BEGIN(MachineLICM, "machinelicm", + "Machine Loop Invariant Code Motion", false, false) +INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) +INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) +INITIALIZE_AG_DEPENDENCY(AliasAnalysis) +INITIALIZE_PASS_END(MachineLICM, "machinelicm", + "Machine Loop Invariant Code Motion", false, false) -FunctionPass *llvm::createMachineLICMPass() { return new MachineLICM(); } +FunctionPass *llvm::createMachineLICMPass(bool PreRegAlloc) { + return new MachineLICM(PreRegAlloc); +} -/// LoopIsOuterMostWithPreheader - Test if the given loop is the outer-most -/// loop that has a preheader. -static bool LoopIsOuterMostWithPreheader(MachineLoop *CurLoop) { +/// LoopIsOuterMostWithPredecessor - Test if the given loop is the outer-most +/// loop that has a unique predecessor. +static bool LoopIsOuterMostWithPredecessor(MachineLoop *CurLoop) { + // Check whether this loop even has a unique predecessor. + if (!CurLoop->getLoopPredecessor()) + return false; + // Ok, now check to see if any of its outer loops do. for (MachineLoop *L = CurLoop->getParentLoop(); L; L = L->getParentLoop()) - if (L->getLoopPreheader()) + if (L->getLoopPredecessor()) return false; + // None of them did, so this is the outermost with a unique predecessor. return true; } -/// Hoist expressions out of the specified loop. Note, alias info for inner loop -/// is not preserved so it is not a good idea to run LICM multiple times on one -/// loop. -/// bool MachineLICM::runOnMachineFunction(MachineFunction &MF) { - DOUT << "******** Machine LICM ********\n"; + if (PreRegAlloc) + DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: "); + else + DEBUG(dbgs() << "******** Post-regalloc Machine LICM: "); + DEBUG(dbgs() << MF.getFunction()->getName() << " ********\n"); - Changed = false; + Changed = FirstInLoop = false; TM = &MF.getTarget(); TII = TM->getInstrInfo(); - RegInfo = &MF.getRegInfo(); + TLI = TM->getTargetLowering(); + TRI = TM->getRegisterInfo(); + MFI = MF.getFrameInfo(); + MRI = &MF.getRegInfo(); + InstrItins = TM->getInstrItineraryData(); + AllocatableSet = TRI->getAllocatableSet(MF); + + if (PreRegAlloc) { + // Estimate register pressure during pre-regalloc pass. + unsigned NumRC = TRI->getNumRegClasses(); + RegPressure.resize(NumRC); + std::fill(RegPressure.begin(), RegPressure.end(), 0); + RegLimit.resize(NumRC); + for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(), + E = TRI->regclass_end(); I != E; ++I) + RegLimit[(*I)->getID()] = TLI->getRegPressureLimit(*I, MF); + } // Get our Loop information... - LI = &getAnalysis(); - DT = &getAnalysis(); + MLI = &getAnalysis(); + DT = &getAnalysis(); + AA = &getAnalysis(); - for (MachineLoopInfo::iterator - I = LI->begin(), E = LI->end(); I != E; ++I) { - CurLoop = *I; + SmallVector Worklist(MLI->begin(), MLI->end()); + while (!Worklist.empty()) { + CurLoop = Worklist.pop_back_val(); + CurPreheader = 0; - // Only visit outer-most preheader-sporting loops. - if (!LoopIsOuterMostWithPreheader(CurLoop)) + // If this is done before regalloc, only visit outer-most preheader-sporting + // loops. + if (PreRegAlloc && !LoopIsOuterMostWithPredecessor(CurLoop)) { + Worklist.append(CurLoop->begin(), CurLoop->end()); continue; + } - // Determine the block to which to hoist instructions. If we can't find a - // suitable loop preheader, we can't do any hoisting. - // - // FIXME: We are only hoisting if the basic block coming into this loop - // has only one successor. This isn't the case in general because we haven't - // broken critical edges or added preheaders. - CurPreheader = CurLoop->getLoopPreheader(); - if (!CurPreheader) + if (!PreRegAlloc) + HoistRegionPostRA(); + else { + // CSEMap is initialized for loop header when the first instruction is + // being hoisted. + MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader()); + FirstInLoop = true; + HoistRegion(N, true); + CSEMap.clear(); + } + } + + return Changed; +} + +/// InstructionStoresToFI - Return true if instruction stores to the +/// specified frame. +static bool InstructionStoresToFI(const MachineInstr *MI, int FI) { + for (MachineInstr::mmo_iterator o = MI->memoperands_begin(), + oe = MI->memoperands_end(); o != oe; ++o) { + if (!(*o)->isStore() || !(*o)->getValue()) + continue; + if (const FixedStackPseudoSourceValue *Value = + dyn_cast((*o)->getValue())) { + if (Value->getFrameIndex() == FI) + return true; + } + } + return false; +} + +/// ProcessMI - Examine the instruction for potentai LICM candidate. Also +/// gather register def and frame object update information. +void MachineLICM::ProcessMI(MachineInstr *MI, + unsigned *PhysRegDefs, + SmallSet &StoredFIs, + SmallVector &Candidates) { + bool RuledOut = false; + bool HasNonInvariantUse = false; + unsigned Def = 0; + for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { + const MachineOperand &MO = MI->getOperand(i); + if (MO.isFI()) { + // Remember if the instruction stores to the frame index. + int FI = MO.getIndex(); + if (!StoredFIs.count(FI) && + MFI->isSpillSlotObjectIndex(FI) && + InstructionStoresToFI(MI, FI)) + StoredFIs.insert(FI); + HasNonInvariantUse = true; + continue; + } + + if (!MO.isReg()) + continue; + unsigned Reg = MO.getReg(); + if (!Reg) continue; + assert(TargetRegisterInfo::isPhysicalRegister(Reg) && + "Not expecting virtual register!"); - HoistRegion(DT->getNode(CurLoop->getHeader())); + if (!MO.isDef()) { + if (Reg && PhysRegDefs[Reg]) + // If it's using a non-loop-invariant register, then it's obviously not + // safe to hoist. + HasNonInvariantUse = true; + continue; + } + + if (MO.isImplicit()) { + ++PhysRegDefs[Reg]; + for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) + ++PhysRegDefs[*AS]; + if (!MO.isDead()) + // Non-dead implicit def? This cannot be hoisted. + RuledOut = true; + // No need to check if a dead implicit def is also defined by + // another instruction. + continue; + } + + // FIXME: For now, avoid instructions with multiple defs, unless + // it's a dead implicit def. + if (Def) + RuledOut = true; + else + Def = Reg; + + // If we have already seen another instruction that defines the same + // register, then this is not safe. + if (++PhysRegDefs[Reg] > 1) + // MI defined register is seen defined by another instruction in + // the loop, it cannot be a LICM candidate. + RuledOut = true; + for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) + if (++PhysRegDefs[*AS] > 1) + RuledOut = true; } - return Changed; + // Only consider reloads for now and remats which do not have register + // operands. FIXME: Consider unfold load folding instructions. + if (Def && !RuledOut) { + int FI = INT_MIN; + if ((!HasNonInvariantUse && IsLICMCandidate(*MI)) || + (TII->isLoadFromStackSlot(MI, FI) && MFI->isSpillSlotObjectIndex(FI))) + Candidates.push_back(CandidateInfo(MI, Def, FI)); + } +} + +/// HoistRegionPostRA - Walk the specified region of the CFG and hoist loop +/// invariants out to the preheader. +void MachineLICM::HoistRegionPostRA() { + unsigned NumRegs = TRI->getNumRegs(); + unsigned *PhysRegDefs = new unsigned[NumRegs]; + std::fill(PhysRegDefs, PhysRegDefs + NumRegs, 0); + + SmallVector Candidates; + SmallSet StoredFIs; + + // Walk the entire region, count number of defs for each register, and + // collect potential LICM candidates. + const std::vector Blocks = CurLoop->getBlocks(); + for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { + MachineBasicBlock *BB = Blocks[i]; + // Conservatively treat live-in's as an external def. + // FIXME: That means a reload that're reused in successor block(s) will not + // be LICM'ed. + for (MachineBasicBlock::livein_iterator I = BB->livein_begin(), + E = BB->livein_end(); I != E; ++I) { + unsigned Reg = *I; + ++PhysRegDefs[Reg]; + for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) + ++PhysRegDefs[*AS]; + } + + for (MachineBasicBlock::iterator + MII = BB->begin(), E = BB->end(); MII != E; ++MII) { + MachineInstr *MI = &*MII; + ProcessMI(MI, PhysRegDefs, StoredFIs, Candidates); + } + } + + // Now evaluate whether the potential candidates qualify. + // 1. Check if the candidate defined register is defined by another + // instruction in the loop. + // 2. If the candidate is a load from stack slot (always true for now), + // check if the slot is stored anywhere in the loop. + for (unsigned i = 0, e = Candidates.size(); i != e; ++i) { + if (Candidates[i].FI != INT_MIN && + StoredFIs.count(Candidates[i].FI)) + continue; + + if (PhysRegDefs[Candidates[i].Def] == 1) { + bool Safe = true; + MachineInstr *MI = Candidates[i].MI; + for (unsigned j = 0, ee = MI->getNumOperands(); j != ee; ++j) { + const MachineOperand &MO = MI->getOperand(j); + if (!MO.isReg() || MO.isDef() || !MO.getReg()) + continue; + if (PhysRegDefs[MO.getReg()]) { + // If it's using a non-loop-invariant register, then it's obviously + // not safe to hoist. + Safe = false; + break; + } + } + if (Safe) + HoistPostRA(MI, Candidates[i].Def); + } + } + + delete[] PhysRegDefs; +} + +/// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the current +/// loop, and make sure it is not killed by any instructions in the loop. +void MachineLICM::AddToLiveIns(unsigned Reg) { + const std::vector Blocks = CurLoop->getBlocks(); + for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { + MachineBasicBlock *BB = Blocks[i]; + if (!BB->isLiveIn(Reg)) + BB->addLiveIn(Reg); + for (MachineBasicBlock::iterator + MII = BB->begin(), E = BB->end(); MII != E; ++MII) { + MachineInstr *MI = &*MII; + for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { + MachineOperand &MO = MI->getOperand(i); + if (!MO.isReg() || !MO.getReg() || MO.isDef()) continue; + if (MO.getReg() == Reg || TRI->isSuperRegister(Reg, MO.getReg())) + MO.setIsKill(false); + } + } + } +} + +/// HoistPostRA - When an instruction is found to only use loop invariant +/// operands that is safe to hoist, this instruction is called to do the +/// dirty work. +void MachineLICM::HoistPostRA(MachineInstr *MI, unsigned Def) { + MachineBasicBlock *Preheader = getCurPreheader(); + if (!Preheader) return; + + // Now move the instructions to the predecessor, inserting it before any + // terminator instructions. + DEBUG({ + dbgs() << "Hoisting " << *MI; + if (Preheader->getBasicBlock()) + dbgs() << " to MachineBasicBlock " + << Preheader->getName(); + if (MI->getParent()->getBasicBlock()) + dbgs() << " from MachineBasicBlock " + << MI->getParent()->getName(); + dbgs() << "\n"; + }); + + // Splice the instruction to the preheader. + MachineBasicBlock *MBB = MI->getParent(); + Preheader->splice(Preheader->getFirstTerminator(), MBB, MI); + + // Add register to livein list to all the BBs in the current loop since a + // loop invariant must be kept live throughout the whole loop. This is + // important to ensure later passes do not scavenge the def register. + AddToLiveIns(Def); + + ++NumPostRAHoisted; + Changed = true; } /// HoistRegion - Walk the specified region of the CFG (defined by all blocks @@ -151,78 +547,159 @@ bool MachineLICM::runOnMachineFunction(MachineFunction &MF) { /// first order w.r.t the DominatorTree. This allows us to visit definitions /// before uses, allowing us to hoist a loop body in one pass without iteration. /// -void MachineLICM::HoistRegion(MachineDomTreeNode *N) { +void MachineLICM::HoistRegion(MachineDomTreeNode *N, bool IsHeader) { assert(N != 0 && "Null dominator tree node?"); MachineBasicBlock *BB = N->getBlock(); // If this subregion is not in the top level loop at all, exit. if (!CurLoop->contains(BB)) return; + MachineBasicBlock *Preheader = getCurPreheader(); + if (!Preheader) + return; + + if (IsHeader) { + // Compute registers which are livein into the loop headers. + RegSeen.clear(); + BackTrace.clear(); + InitRegPressure(Preheader); + } + + // Remember livein register pressure. + BackTrace.push_back(RegPressure); + for (MachineBasicBlock::iterator - I = BB->begin(), E = BB->end(); I != E; ) { - MachineInstr &MI = *I++; + MII = BB->begin(), E = BB->end(); MII != E; ) { + MachineBasicBlock::iterator NextMII = MII; ++NextMII; + MachineInstr *MI = &*MII; + if (!Hoist(MI, Preheader)) + UpdateRegPressure(MI); + MII = NextMII; + } - // Try hoisting the instruction out of the loop. We can only do this if - // all of the operands of the instruction are loop invariant and if it is - // safe to hoist the instruction. - Hoist(MI); + // Don't hoist things out of a large switch statement. This often causes + // code to be hoisted that wasn't going to be executed, and increases + // register pressure in a situation where it's likely to matter. + if (BB->succ_size() < 25) { + const std::vector &Children = N->getChildren(); + for (unsigned I = 0, E = Children.size(); I != E; ++I) + HoistRegion(Children[I]); } - const std::vector &Children = N->getChildren(); + BackTrace.pop_back(); +} - for (unsigned I = 0, E = Children.size(); I != E; ++I) - HoistRegion(Children[I]); +static bool isOperandKill(const MachineOperand &MO, MachineRegisterInfo *MRI) { + return MO.isKill() || MRI->hasOneNonDBGUse(MO.getReg()); } -/// IsLoopInvariantInst - Returns true if the instruction is loop -/// invariant. I.e., all virtual register operands are defined outside of the -/// loop, physical registers aren't accessed explicitly, and there are no side -/// effects that aren't captured by the operands or other flags. -/// -bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) { - const TargetInstrDesc &TID = I.getDesc(); - - // Ignore stuff that we obviously can't hoist. - if (TID.mayStore() || TID.isCall() || TID.isTerminator() || - TID.hasUnmodeledSideEffects()) - return false; - - if (TID.mayLoad()) { - // Okay, this instruction does a load. As a refinement, we allow the target - // to decide whether the loaded value is actually a constant. If so, we can - // actually use it as a load. - if (!TII->isInvariantLoad(&I)) - // FIXME: we should be able to sink loads with no other side effects if - // there is nothing that can change memory from here until the end of - // block. This is a trivial form of alias analysis. - return false; +/// InitRegPressure - Find all virtual register references that are liveout of +/// the preheader to initialize the starting "register pressure". Note this +/// does not count live through (livein but not used) registers. +void MachineLICM::InitRegPressure(MachineBasicBlock *BB) { + std::fill(RegPressure.begin(), RegPressure.end(), 0); + + // If the preheader has only a single predecessor and it ends with a + // fallthrough or an unconditional branch, then scan its predecessor for live + // defs as well. This happens whenever the preheader is created by splitting + // the critical edge from the loop predecessor to the loop header. + if (BB->pred_size() == 1) { + MachineBasicBlock *TBB = 0, *FBB = 0; + SmallVector Cond; + if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond, false) && Cond.empty()) + InitRegPressure(*BB->pred_begin()); } - DEBUG({ - DOUT << "--- Checking if we can hoist " << I; - if (I.getDesc().getImplicitUses()) { - DOUT << " * Instruction has implicit uses:\n"; - - const TargetRegisterInfo *TRI = TM->getRegisterInfo(); - for (const unsigned *ImpUses = I.getDesc().getImplicitUses(); - *ImpUses; ++ImpUses) - DOUT << " -> " << TRI->getName(*ImpUses) << "\n"; + for (MachineBasicBlock::iterator MII = BB->begin(), E = BB->end(); + MII != E; ++MII) { + MachineInstr *MI = &*MII; + for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) { + const MachineOperand &MO = MI->getOperand(i); + if (!MO.isReg() || MO.isImplicit()) + continue; + unsigned Reg = MO.getReg(); + if (!TargetRegisterInfo::isVirtualRegister(Reg)) + continue; + + bool isNew = RegSeen.insert(Reg); + const TargetRegisterClass *RC = MRI->getRegClass(Reg); + EVT VT = *RC->vt_begin(); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + if (MO.isDef()) + RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); + else { + bool isKill = isOperandKill(MO, MRI); + if (isNew && !isKill) + // Haven't seen this, it must be a livein. + RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); + else if (!isNew && isKill) + RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT); } + } + } +} - if (I.getDesc().getImplicitDefs()) { - DOUT << " * Instruction has implicit defines:\n"; +/// UpdateRegPressure - Update estimate of register pressure after the +/// specified instruction. +void MachineLICM::UpdateRegPressure(const MachineInstr *MI) { + if (MI->isImplicitDef()) + return; - const TargetRegisterInfo *TRI = TM->getRegisterInfo(); - for (const unsigned *ImpDefs = I.getDesc().getImplicitDefs(); - *ImpDefs; ++ImpDefs) - DOUT << " -> " << TRI->getName(*ImpDefs) << "\n"; - } - }); + SmallVector Defs; + for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) { + const MachineOperand &MO = MI->getOperand(i); + if (!MO.isReg() || MO.isImplicit()) + continue; + unsigned Reg = MO.getReg(); + if (!TargetRegisterInfo::isVirtualRegister(Reg)) + continue; - if (I.getDesc().getImplicitDefs() || I.getDesc().getImplicitUses()) { - DOUT << "Cannot hoist with implicit defines or uses\n"; - return false; + bool isNew = RegSeen.insert(Reg); + if (MO.isDef()) + Defs.push_back(Reg); + else if (!isNew && isOperandKill(MO, MRI)) { + const TargetRegisterClass *RC = MRI->getRegClass(Reg); + EVT VT = *RC->vt_begin(); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + unsigned RCCost = TLI->getRepRegClassCostFor(VT); + + if (RCCost > RegPressure[RCId]) + RegPressure[RCId] = 0; + else + RegPressure[RCId] -= RCCost; + } + } + + while (!Defs.empty()) { + unsigned Reg = Defs.pop_back_val(); + const TargetRegisterClass *RC = MRI->getRegClass(Reg); + EVT VT = *RC->vt_begin(); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + unsigned RCCost = TLI->getRepRegClassCostFor(VT); + RegPressure[RCId] += RCCost; } +} + +/// IsLICMCandidate - Returns true if the instruction may be a suitable +/// candidate for LICM. e.g. If the instruction is a call, then it's obviously +/// not safe to hoist it. +bool MachineLICM::IsLICMCandidate(MachineInstr &I) { + // Check if it's safe to move the instruction. + bool DontMoveAcrossStore = true; + if (!I.isSafeToMove(TII, AA, DontMoveAcrossStore)) + return false; + + return true; +} + +/// IsLoopInvariantInst - Returns true if the instruction is loop +/// invariant. I.e., all virtual register operands are defined outside of the +/// loop, physical registers aren't accessed explicitly, and there are no side +/// effects that aren't captured by the operands or other flags. +/// +bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) { + if (!IsLICMCandidate(I)) + return false; // The instruction is loop invariant if all of its operands are. for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { @@ -235,18 +712,44 @@ bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) { if (Reg == 0) continue; // Don't hoist an instruction that uses or defines a physical register. - if (TargetRegisterInfo::isPhysicalRegister(Reg)) - return false; + if (TargetRegisterInfo::isPhysicalRegister(Reg)) { + if (MO.isUse()) { + // If the physreg has no defs anywhere, it's just an ambient register + // and we can freely move its uses. Alternatively, if it's allocatable, + // it could get allocated to something with a def during allocation. + if (!MRI->def_empty(Reg)) + return false; + if (AllocatableSet.test(Reg)) + return false; + // Check for a def among the register's aliases too. + for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) { + unsigned AliasReg = *Alias; + if (!MRI->def_empty(AliasReg)) + return false; + if (AllocatableSet.test(AliasReg)) + return false; + } + // Otherwise it's safe to move. + continue; + } else if (!MO.isDead()) { + // A def that isn't dead. We can't move it. + return false; + } else if (CurLoop->getHeader()->isLiveIn(Reg)) { + // If the reg is live into the loop, we can't hoist an instruction + // which would clobber it. + return false; + } + } if (!MO.isUse()) continue; - assert(RegInfo->getVRegDef(Reg) && + assert(MRI->getVRegDef(Reg) && "Machine instr not mapped for this vreg?!"); // If the loop contains the definition of an operand, then the instruction // isn't loop invariant. - if (CurLoop->contains(RegInfo->getVRegDef(Reg)->getParent())) + if (CurLoop->contains(MRI->getVRegDef(Reg))) return false; } @@ -254,27 +757,446 @@ bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) { return true; } + +/// HasPHIUses - Return true if the specified register has any PHI use. +static bool HasPHIUses(unsigned Reg, MachineRegisterInfo *MRI) { + for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg), + UE = MRI->use_end(); UI != UE; ++UI) { + MachineInstr *UseMI = &*UI; + if (UseMI->isPHI()) + return true; + } + return false; +} + + +/// HasHighOperandLatency - Compute operand latency between a def of 'Reg' +/// and an use in the current loop, return true if the target considered +/// it 'high'. +bool MachineLICM::HasHighOperandLatency(MachineInstr &MI, + unsigned DefIdx, unsigned Reg) const { + if (!InstrItins || InstrItins->isEmpty() || MRI->use_nodbg_empty(Reg)) + return false; + + for (MachineRegisterInfo::use_nodbg_iterator I = MRI->use_nodbg_begin(Reg), + E = MRI->use_nodbg_end(); I != E; ++I) { + MachineInstr *UseMI = &*I; + if (UseMI->isCopyLike()) + continue; + if (!CurLoop->contains(UseMI->getParent())) + continue; + for (unsigned i = 0, e = UseMI->getNumOperands(); i != e; ++i) { + const MachineOperand &MO = UseMI->getOperand(i); + if (!MO.isReg() || !MO.isUse()) + continue; + unsigned MOReg = MO.getReg(); + if (MOReg != Reg) + continue; + + if (TII->hasHighOperandLatency(InstrItins, MRI, &MI, DefIdx, UseMI, i)) + return true; + } + + // Only look at the first in loop use. + break; + } + + return false; +} + +/// IsCheapInstruction - Return true if the instruction is marked "cheap" or +/// the operand latency between its def and a use is one or less. +bool MachineLICM::IsCheapInstruction(MachineInstr &MI) const { + if (MI.getDesc().isAsCheapAsAMove() || MI.isCopyLike()) + return true; + if (!InstrItins || InstrItins->isEmpty()) + return false; + + bool isCheap = false; + unsigned NumDefs = MI.getDesc().getNumDefs(); + for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) { + MachineOperand &DefMO = MI.getOperand(i); + if (!DefMO.isReg() || !DefMO.isDef()) + continue; + --NumDefs; + unsigned Reg = DefMO.getReg(); + if (TargetRegisterInfo::isPhysicalRegister(Reg)) + continue; + + if (!TII->hasLowDefLatency(InstrItins, &MI, i)) + return false; + isCheap = true; + } + + return isCheap; +} + +/// CanCauseHighRegPressure - Visit BBs from header to current BB, check +/// if hoisting an instruction of the given cost matrix can cause high +/// register pressure. +bool MachineLICM::CanCauseHighRegPressure(DenseMap &Cost) { + for (DenseMap::iterator CI = Cost.begin(), CE = Cost.end(); + CI != CE; ++CI) { + if (CI->second <= 0) + continue; + + unsigned RCId = CI->first; + for (unsigned i = BackTrace.size(); i != 0; --i) { + SmallVector &RP = BackTrace[i-1]; + if (RP[RCId] + CI->second >= RegLimit[RCId]) + return true; + } + } + + return false; +} + +/// UpdateBackTraceRegPressure - Traverse the back trace from header to the +/// current block and update their register pressures to reflect the effect +/// of hoisting MI from the current block to the preheader. +void MachineLICM::UpdateBackTraceRegPressure(const MachineInstr *MI) { + if (MI->isImplicitDef()) + return; + + // First compute the 'cost' of the instruction, i.e. its contribution + // to register pressure. + DenseMap Cost; + for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) { + const MachineOperand &MO = MI->getOperand(i); + if (!MO.isReg() || MO.isImplicit()) + continue; + unsigned Reg = MO.getReg(); + if (!TargetRegisterInfo::isVirtualRegister(Reg)) + continue; + + const TargetRegisterClass *RC = MRI->getRegClass(Reg); + EVT VT = *RC->vt_begin(); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + unsigned RCCost = TLI->getRepRegClassCostFor(VT); + if (MO.isDef()) { + DenseMap::iterator CI = Cost.find(RCId); + if (CI != Cost.end()) + CI->second += RCCost; + else + Cost.insert(std::make_pair(RCId, RCCost)); + } else if (isOperandKill(MO, MRI)) { + DenseMap::iterator CI = Cost.find(RCId); + if (CI != Cost.end()) + CI->second -= RCCost; + else + Cost.insert(std::make_pair(RCId, -RCCost)); + } + } + + // Update register pressure of blocks from loop header to current block. + for (unsigned i = 0, e = BackTrace.size(); i != e; ++i) { + SmallVector &RP = BackTrace[i]; + for (DenseMap::iterator CI = Cost.begin(), CE = Cost.end(); + CI != CE; ++CI) { + unsigned RCId = CI->first; + RP[RCId] += CI->second; + } + } +} + +/// IsProfitableToHoist - Return true if it is potentially profitable to hoist +/// the given loop invariant. +bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) { + if (MI.isImplicitDef()) + return true; + + // If the instruction is cheap, only hoist if it is re-materilizable. LICM + // will increase register pressure. It's probably not worth it if the + // instruction is cheap. + // Also hoist loads from constant memory, e.g. load from stubs, GOT. Hoisting + // these tend to help performance in low register pressure situation. The + // trade off is it may cause spill in high pressure situation. It will end up + // adding a store in the loop preheader. But the reload is no more expensive. + // The side benefit is these loads are frequently CSE'ed. + if (IsCheapInstruction(MI)) { + if (!TII->isTriviallyReMaterializable(&MI, AA)) + return false; + } else { + // Estimate register pressure to determine whether to LICM the instruction. + // In low register pressure situation, we can be more aggressive about + // hoisting. Also, favors hoisting long latency instructions even in + // moderately high pressure situation. + // FIXME: If there are long latency loop-invariant instructions inside the + // loop at this point, why didn't the optimizer's LICM hoist them? + DenseMap Cost; + for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) { + const MachineOperand &MO = MI.getOperand(i); + if (!MO.isReg() || MO.isImplicit()) + continue; + unsigned Reg = MO.getReg(); + if (!TargetRegisterInfo::isVirtualRegister(Reg)) + continue; + if (MO.isDef()) { + if (HasHighOperandLatency(MI, i, Reg)) { + ++NumHighLatency; + return true; + } + + const TargetRegisterClass *RC = MRI->getRegClass(Reg); + EVT VT = *RC->vt_begin(); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + unsigned RCCost = TLI->getRepRegClassCostFor(VT); + DenseMap::iterator CI = Cost.find(RCId); + if (CI != Cost.end()) + CI->second += RCCost; + else + Cost.insert(std::make_pair(RCId, RCCost)); + } else if (isOperandKill(MO, MRI)) { + // Is a virtual register use is a kill, hoisting it out of the loop + // may actually reduce register pressure or be register pressure + // neutral. + const TargetRegisterClass *RC = MRI->getRegClass(Reg); + EVT VT = *RC->vt_begin(); + unsigned RCId = TLI->getRepRegClassFor(VT)->getID(); + unsigned RCCost = TLI->getRepRegClassCostFor(VT); + DenseMap::iterator CI = Cost.find(RCId); + if (CI != Cost.end()) + CI->second -= RCCost; + else + Cost.insert(std::make_pair(RCId, -RCCost)); + } + } + + // Visit BBs from header to current BB, if hoisting this doesn't cause + // high register pressure, then it's safe to proceed. + if (!CanCauseHighRegPressure(Cost)) { + ++NumLowRP; + return true; + } + + // High register pressure situation, only hoist if the instruction is going to + // be remat'ed. + if (!TII->isTriviallyReMaterializable(&MI, AA) && + !MI.isInvariantLoad(AA)) + return false; + } + + // If result(s) of this instruction is used by PHIs, then don't hoist it. + // The presence of joins makes it difficult for current register allocator + // implementation to perform remat. + for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { + const MachineOperand &MO = MI.getOperand(i); + if (!MO.isReg() || !MO.isDef()) + continue; + if (HasPHIUses(MO.getReg(), MRI)) + return false; + } + + return true; +} + +MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) { + // Don't unfold simple loads. + if (MI->getDesc().canFoldAsLoad()) + return 0; + + // If not, we may be able to unfold a load and hoist that. + // First test whether the instruction is loading from an amenable + // memory location. + if (!MI->isInvariantLoad(AA)) + return 0; + + // Next determine the register class for a temporary register. + unsigned LoadRegIndex; + unsigned NewOpc = + TII->getOpcodeAfterMemoryUnfold(MI->getOpcode(), + /*UnfoldLoad=*/true, + /*UnfoldStore=*/false, + &LoadRegIndex); + if (NewOpc == 0) return 0; + const TargetInstrDesc &TID = TII->get(NewOpc); + if (TID.getNumDefs() != 1) return 0; + const TargetRegisterClass *RC = TID.OpInfo[LoadRegIndex].getRegClass(TRI); + // Ok, we're unfolding. Create a temporary register and do the unfold. + unsigned Reg = MRI->createVirtualRegister(RC); + + MachineFunction &MF = *MI->getParent()->getParent(); + SmallVector NewMIs; + bool Success = + TII->unfoldMemoryOperand(MF, MI, Reg, + /*UnfoldLoad=*/true, /*UnfoldStore=*/false, + NewMIs); + (void)Success; + assert(Success && + "unfoldMemoryOperand failed when getOpcodeAfterMemoryUnfold " + "succeeded!"); + assert(NewMIs.size() == 2 && + "Unfolded a load into multiple instructions!"); + MachineBasicBlock *MBB = MI->getParent(); + MBB->insert(MI, NewMIs[0]); + MBB->insert(MI, NewMIs[1]); + // If unfolding produced a load that wasn't loop-invariant or profitable to + // hoist, discard the new instructions and bail. + if (!IsLoopInvariantInst(*NewMIs[0]) || !IsProfitableToHoist(*NewMIs[0])) { + NewMIs[0]->eraseFromParent(); + NewMIs[1]->eraseFromParent(); + return 0; + } + + // Update register pressure for the unfolded instruction. + UpdateRegPressure(NewMIs[1]); + + // Otherwise we successfully unfolded a load that we can hoist. + MI->eraseFromParent(); + return NewMIs[0]; +} + +void MachineLICM::InitCSEMap(MachineBasicBlock *BB) { + for (MachineBasicBlock::iterator I = BB->begin(),E = BB->end(); I != E; ++I) { + const MachineInstr *MI = &*I; + unsigned Opcode = MI->getOpcode(); + DenseMap >::iterator + CI = CSEMap.find(Opcode); + if (CI != CSEMap.end()) + CI->second.push_back(MI); + else { + std::vector CSEMIs; + CSEMIs.push_back(MI); + CSEMap.insert(std::make_pair(Opcode, CSEMIs)); + } + } +} + +const MachineInstr* +MachineLICM::LookForDuplicate(const MachineInstr *MI, + std::vector &PrevMIs) { + for (unsigned i = 0, e = PrevMIs.size(); i != e; ++i) { + const MachineInstr *PrevMI = PrevMIs[i]; + if (TII->produceSameValue(MI, PrevMI, (PreRegAlloc ? MRI : 0))) + return PrevMI; + } + return 0; +} + +bool MachineLICM::EliminateCSE(MachineInstr *MI, + DenseMap >::iterator &CI) { + // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate + // the undef property onto uses. + if (CI == CSEMap.end() || MI->isImplicitDef()) + return false; + + if (const MachineInstr *Dup = LookForDuplicate(MI, CI->second)) { + DEBUG(dbgs() << "CSEing " << *MI << " with " << *Dup); + + // Replace virtual registers defined by MI by their counterparts defined + // by Dup. + for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { + const MachineOperand &MO = MI->getOperand(i); + + // Physical registers may not differ here. + assert((!MO.isReg() || MO.getReg() == 0 || + !TargetRegisterInfo::isPhysicalRegister(MO.getReg()) || + MO.getReg() == Dup->getOperand(i).getReg()) && + "Instructions with different phys regs are not identical!"); + + if (MO.isReg() && MO.isDef() && + !TargetRegisterInfo::isPhysicalRegister(MO.getReg())) { + MRI->replaceRegWith(MO.getReg(), Dup->getOperand(i).getReg()); + MRI->clearKillFlags(Dup->getOperand(i).getReg()); + } + } + MI->eraseFromParent(); + ++NumCSEed; + return true; + } + return false; +} + /// Hoist - When an instruction is found to use only loop invariant operands /// that are safe to hoist, this instruction is called to do the dirty work. /// -void MachineLICM::Hoist(MachineInstr &MI) { - if (!IsLoopInvariantInst(MI)) return; +bool MachineLICM::Hoist(MachineInstr *MI, MachineBasicBlock *Preheader) { + // First check whether we should hoist this instruction. + if (!IsLoopInvariantInst(*MI) || !IsProfitableToHoist(*MI)) { + // If not, try unfolding a hoistable load. + MI = ExtractHoistableLoad(MI); + if (!MI) return false; + } // Now move the instructions to the predecessor, inserting it before any // terminator instructions. DEBUG({ - DOUT << "Hoisting " << MI; - if (CurPreheader->getBasicBlock()) - DOUT << " to MachineBasicBlock " - << CurPreheader->getBasicBlock()->getName(); - if (MI.getParent()->getBasicBlock()) - DOUT << " from MachineBasicBlock " - << MI.getParent()->getBasicBlock()->getName(); - DOUT << "\n"; + dbgs() << "Hoisting " << *MI; + if (Preheader->getBasicBlock()) + dbgs() << " to MachineBasicBlock " + << Preheader->getName(); + if (MI->getParent()->getBasicBlock()) + dbgs() << " from MachineBasicBlock " + << MI->getParent()->getName(); + dbgs() << "\n"; }); - CurPreheader->splice(CurPreheader->getFirstTerminator(), MI.getParent(), &MI); + // If this is the first instruction being hoisted to the preheader, + // initialize the CSE map with potential common expressions. + if (FirstInLoop) { + InitCSEMap(Preheader); + FirstInLoop = false; + } + + // Look for opportunity to CSE the hoisted instruction. + unsigned Opcode = MI->getOpcode(); + DenseMap >::iterator + CI = CSEMap.find(Opcode); + if (!EliminateCSE(MI, CI)) { + // Otherwise, splice the instruction to the preheader. + Preheader->splice(Preheader->getFirstTerminator(),MI->getParent(),MI); + + // Update register pressure for BBs from header to this block. + UpdateBackTraceRegPressure(MI); + + // Clear the kill flags of any register this instruction defines, + // since they may need to be live throughout the entire loop + // rather than just live for part of it. + for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { + MachineOperand &MO = MI->getOperand(i); + if (MO.isReg() && MO.isDef() && !MO.isDead()) + MRI->clearKillFlags(MO.getReg()); + } + + // Add to the CSE map. + if (CI != CSEMap.end()) + CI->second.push_back(MI); + else { + std::vector CSEMIs; + CSEMIs.push_back(MI); + CSEMap.insert(std::make_pair(Opcode, CSEMIs)); + } + } ++NumHoisted; Changed = true; + + return true; +} + +MachineBasicBlock *MachineLICM::getCurPreheader() { + // Determine the block to which to hoist instructions. If we can't find a + // suitable loop predecessor, we can't do any hoisting. + + // If we've tried to get a preheader and failed, don't try again. + if (CurPreheader == reinterpret_cast(-1)) + return 0; + + if (!CurPreheader) { + CurPreheader = CurLoop->getLoopPreheader(); + if (!CurPreheader) { + MachineBasicBlock *Pred = CurLoop->getLoopPredecessor(); + if (!Pred) { + CurPreheader = reinterpret_cast(-1); + return 0; + } + + CurPreheader = Pred->SplitCriticalEdge(CurLoop->getHeader(), this); + if (!CurPreheader) { + CurPreheader = reinterpret_cast(-1); + return 0; + } + } + } + return CurPreheader; }