X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FMachineLICM.cpp;h=50091f39e5c2ee2467656e95503a614f24c9219d;hb=1fe75126274262c55648e802e47557ee0c7f9d90;hp=9b24a9a2391b549a0c367767c2b79f5e90347fcc;hpb=853d3fb8d24fab2258e9cd5dce3ec8ff4189eeda;p=oota-llvm.git diff --git a/lib/CodeGen/MachineLICM.cpp b/lib/CodeGen/MachineLICM.cpp index 9b24a9a2391..50091f39e5c 100644 --- a/lib/CodeGen/MachineLICM.cpp +++ b/lib/CodeGen/MachineLICM.cpp @@ -10,50 +10,72 @@ // This pass performs loop invariant code motion on machine instructions. We // attempt to remove as much code from the body of a loop as possible. // -// This pass does not attempt to throttle itself to limit register pressure. -// The register allocation phases are expected to perform rematerialization -// to recover when register pressure is high. -// // This pass is not intended to be a replacement or a complete alternative // for the LLVM-IR-level LICM pass. It is only designed to hoist simple // constructs that are not exposed before lowering and instruction selection. // //===----------------------------------------------------------------------===// -#define DEBUG_TYPE "machine-licm" #include "llvm/CodeGen/Passes.h" +#include "llvm/ADT/DenseMap.h" +#include "llvm/ADT/SmallSet.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/Analysis/AliasAnalysis.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/TargetRegisterInfo.h" -#include "llvm/Target/TargetInstrInfo.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/MC/MCInstrItineraries.h" +#include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" - +#include "llvm/Target/TargetInstrInfo.h" +#include "llvm/Target/TargetLowering.h" +#include "llvm/Target/TargetMachine.h" +#include "llvm/Target/TargetRegisterInfo.h" +#include "llvm/Target/TargetSubtargetInfo.h" using namespace llvm; -STATISTIC(NumHoisted, "Number of machine instructions hoisted out of loops"); -STATISTIC(NumCSEed, "Number of hoisted machine instructions CSEed"); +#define DEBUG_TYPE "machine-licm" + +static cl::opt +AvoidSpeculation("avoid-speculation", + cl::desc("MachineLICM should avoid speculation"), + cl::init(true), cl::Hidden); + +static cl::opt +HoistCheapInsts("hoist-cheap-insts", + cl::desc("MachineLICM should hoist even cheap instructions"), + cl::init(false), cl::Hidden); + +static cl::opt +SinkInstsToAvoidSpills("sink-insts-to-avoid-spills", + cl::desc("MachineLICM should sink instructions into " + "loops to avoid register spills"), + cl::init(false), cl::Hidden); + +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 MachineLICM : public MachineFunctionPass { - bool PreRegAlloc; - - const TargetMachine *TM; const TargetInstrInfo *TII; + const TargetLoweringBase *TLI; const TargetRegisterInfo *TRI; const MachineFrameInfo *MFI; - MachineRegisterInfo *RegInfo; + MachineRegisterInfo *MRI; + const InstrItineraryData *InstrItins; + bool PreRegAlloc; // Various analyses that we use... AliasAnalysis *AA; // Alias analysis info. @@ -66,25 +88,54 @@ namespace { MachineLoop *CurLoop; // The current loop we are working on. MachineBasicBlock *CurPreheader; // The preheader for CurLoop. - BitVector AllocatableSet; + // Exit blocks for CurLoop. + SmallVector ExitBlocks; + + bool isExitBlock(const MachineBasicBlock *MBB) const { + return std::find(ExitBlocks.begin(), ExitBlocks.end(), MBB) != + ExitBlocks.end(); + } + + // 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 potentail CSE instructions. + // For each opcode, keep a list of potential CSE instructions. DenseMap > CSEMap; + enum { + SpeculateFalse = 0, + SpeculateTrue = 1, + SpeculateUnknown = 2 + }; + + // If a MBB does not dominate loop exiting blocks then it may not safe + // to hoist loads from this block. + // Tri-state: 0 - false, 1 - true, 2 - unknown + unsigned SpeculationState; + public: static char ID; // Pass identification, replacement for typeid MachineLICM() : - MachineFunctionPass(&ID), PreRegAlloc(true) {} + MachineFunctionPass(ID), PreRegAlloc(true) { + initializeMachineLICMPass(*PassRegistry::getPassRegistry()); + } explicit MachineLICM(bool PreRA) : - MachineFunctionPass(&ID), PreRegAlloc(PreRA) {} - - virtual bool runOnMachineFunction(MachineFunction &MF); + MachineFunctionPass(ID), PreRegAlloc(PreRA) { + initializeMachineLICMPass(*PassRegistry::getPassRegistry()); + } - const char *getPassName() const { return "Machine Instruction LICM"; } + bool runOnMachineFunction(MachineFunction &MF) override; - virtual void getAnalysisUsage(AnalysisUsage &AU) const { - AU.setPreservesCFG(); + void getAnalysisUsage(AnalysisUsage &AU) const override { AU.addRequired(); AU.addRequired(); AU.addRequired(); @@ -93,7 +144,11 @@ namespace { MachineFunctionPass::getAnalysisUsage(AU); } - virtual void releaseMemory() { + void releaseMemory() override { + RegSeen.clear(); + RegPressure.clear(); + RegLimit.clear(); + BackTrace.clear(); CSEMap.clear(); } @@ -118,41 +173,99 @@ namespace { /// ProcessMI - Examine the instruction for potentai LICM candidate. Also /// gather register def and frame object update information. - void ProcessMI(MachineInstr *MI, unsigned *PhysRegDefs, + void ProcessMI(MachineInstr *MI, + BitVector &PhysRegDefs, + BitVector &PhysRegClobbers, SmallSet &StoredFIs, - SmallVector &Candidates); + SmallVectorImpl &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. + /// 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), /// and the instruction is hoistable. - /// + /// bool IsLoopInvariantInst(MachineInstr &I); + /// HasLoopPHIUse - Return true if the specified instruction is used by any + /// phi node in the current loop. + bool HasLoopPHIUse(const MachineInstr *MI) const; + + /// 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(const DenseMap &Cost, + bool Cheap); + + /// 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); + /// IsGuaranteedToExecute - Check if this mbb is guaranteed to execute. + /// If not then a load from this mbb may not be safe to hoist. + bool IsGuaranteedToExecute(MachineBasicBlock *BB); - /// isLoadFromConstantMemory - Return true if the given instruction is a - /// load from constant memory. - bool isLoadFromConstantMemory(MachineInstr *MI); + void EnterScope(MachineBasicBlock *MBB); + + void ExitScope(MachineBasicBlock *MBB); + + /// ExitScopeIfDone - Destroy scope for the MBB that corresponds to given + /// dominator tree node if its a leaf or all of its children are done. Walk + /// up the dominator tree to destroy ancestors which are now done. + void ExitScopeIfDone(MachineDomTreeNode *Node, + DenseMap &OpenChildren, + DenseMap &ParentMap); + + /// HoistOutOfLoop - Walk the specified loop in the CFG (defined by all + /// blocks dominated by the specified header 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 HoistOutOfLoop(MachineDomTreeNode *LoopHeaderNode); + void HoistRegion(MachineDomTreeNode *N, bool IsHeader); + + /// SinkIntoLoop - Sink instructions into loops if profitable. This + /// especially tries to prevent register spills caused by register pressure + /// if there is little to no overhead moving instructions into loops. + void SinkIntoLoop(); + + /// getRegisterClassIDAndCost - For a given MI, register, and the operand + /// index, return the ID and cost of its representative register class by + /// reference. + void getRegisterClassIDAndCost(const MachineInstr *MI, + unsigned Reg, unsigned OpIdx, + unsigned &RCId, unsigned &RCCost) const; + + /// 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 @@ -172,10 +285,14 @@ namespace { bool EliminateCSE(MachineInstr *MI, DenseMap >::iterator &CI); + /// MayCSE - Return true if the given instruction will be CSE'd if it's + /// hoisted out of the loop. + bool MayCSE(MachineInstr *MI); + /// 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 @@ -189,49 +306,79 @@ namespace { } // end anonymous namespace char MachineLICM::ID = 0; -static RegisterPass -X("machinelicm", "Machine Loop Invariant Code Motion"); - -FunctionPass *llvm::createMachineLICMPass(bool PreRegAlloc) { - return new MachineLICM(PreRegAlloc); -} +char &llvm::MachineLICMID = MachineLICM::ID; +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) /// 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->getLoopPredecessor()) return false; + // None of them did, so this is the outermost with a unique predecessor. return true; } bool MachineLICM::runOnMachineFunction(MachineFunction &MF) { - if (PreRegAlloc) - DEBUG(dbgs() << "******** Pre-regalloc Machine LICM ********\n"); - else - DEBUG(dbgs() << "******** Post-regalloc Machine LICM ********\n"); + if (skipOptnoneFunction(*MF.getFunction())) + return false; Changed = FirstInLoop = false; - TM = &MF.getTarget(); - TII = TM->getInstrInfo(); - TRI = TM->getRegisterInfo(); + TII = MF.getSubtarget().getInstrInfo(); + TLI = MF.getSubtarget().getTargetLowering(); + TRI = MF.getSubtarget().getRegisterInfo(); MFI = MF.getFrameInfo(); - RegInfo = &MF.getRegInfo(); - AllocatableSet = TRI->getAllocatableSet(MF); + MRI = &MF.getRegInfo(); + InstrItins = MF.getSubtarget().getInstrItineraryData(); + + PreRegAlloc = MRI->isSSA(); + + if (PreRegAlloc) + DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: "); + else + DEBUG(dbgs() << "******** Post-regalloc Machine LICM: "); + DEBUG(dbgs() << MF.getName() << " ********\n"); + + 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()] = TRI->getRegPressureLimit(*I, MF); + } // Get our Loop information... MLI = &getAnalysis(); DT = &getAnalysis(); AA = &getAnalysis(); - for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end(); I != E; ++I){ - CurLoop = *I; - CurPreheader = 0; + SmallVector Worklist(MLI->begin(), MLI->end()); + while (!Worklist.empty()) { + CurLoop = Worklist.pop_back_val(); + CurPreheader = nullptr; + ExitBlocks.clear(); // If this is done before regalloc, only visit outer-most preheader-sporting // loops. - if (PreRegAlloc && !LoopIsOuterMostWithPredecessor(CurLoop)) + if (PreRegAlloc && !LoopIsOuterMostWithPredecessor(CurLoop)) { + Worklist.append(CurLoop->begin(), CurLoop->end()); continue; + } + + CurLoop->getExitBlocks(ExitBlocks); if (!PreRegAlloc) HoistRegionPostRA(); @@ -240,8 +387,11 @@ bool MachineLICM::runOnMachineFunction(MachineFunction &MF) { // being hoisted. MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader()); FirstInLoop = true; - HoistRegion(N); + HoistOutOfLoop(N); CSEMap.clear(); + + if (SinkInstsToAvoidSpills) + SinkIntoLoop(); } } @@ -253,10 +403,10 @@ bool MachineLICM::runOnMachineFunction(MachineFunction &MF) { 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()) + if (!(*o)->isStore() || !(*o)->getPseudoValue()) continue; if (const FixedStackPseudoSourceValue *Value = - dyn_cast((*o)->getValue())) { + dyn_cast((*o)->getPseudoValue())) { if (Value->getFrameIndex() == FI) return true; } @@ -267,9 +417,10 @@ static bool InstructionStoresToFI(const MachineInstr *MI, int FI) { /// ProcessMI - Examine the instruction for potentai LICM candidate. Also /// gather register def and frame object update information. void MachineLICM::ProcessMI(MachineInstr *MI, - unsigned *PhysRegDefs, + BitVector &PhysRegDefs, + BitVector &PhysRegClobbers, SmallSet &StoredFIs, - SmallVector &Candidates) { + SmallVectorImpl &Candidates) { bool RuledOut = false; bool HasNonInvariantUse = false; unsigned Def = 0; @@ -286,6 +437,13 @@ void MachineLICM::ProcessMI(MachineInstr *MI, continue; } + // We can't hoist an instruction defining a physreg that is clobbered in + // the loop. + if (MO.isRegMask()) { + PhysRegClobbers.setBitsNotInMask(MO.getRegMask()); + continue; + } + if (!MO.isReg()) continue; unsigned Reg = MO.getReg(); @@ -295,7 +453,7 @@ void MachineLICM::ProcessMI(MachineInstr *MI, "Not expecting virtual register!"); if (!MO.isDef()) { - if (Reg && PhysRegDefs[Reg]) + if (Reg && (PhysRegDefs.test(Reg) || PhysRegClobbers.test(Reg))) // If it's using a non-loop-invariant register, then it's obviously not // safe to hoist. HasNonInvariantUse = true; @@ -303,9 +461,8 @@ void MachineLICM::ProcessMI(MachineInstr *MI, } if (MO.isImplicit()) { - ++PhysRegDefs[Reg]; - for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) - ++PhysRegDefs[*AS]; + for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) + PhysRegClobbers.set(*AI); if (!MO.isDead()) // Non-dead implicit def? This cannot be hoisted. RuledOut = true; @@ -322,14 +479,17 @@ void MachineLICM::ProcessMI(MachineInstr *MI, Def = Reg; // If we have already seen another instruction that defines the same - // register, then this is not safe. - if (++PhysRegDefs[Reg] > 1) + // register, then this is not safe. Two defs is indicated by setting a + // PhysRegClobbers bit. + for (MCRegAliasIterator AS(Reg, TRI, true); AS.isValid(); ++AS) { + if (PhysRegDefs.test(*AS)) + PhysRegClobbers.set(*AS); + PhysRegDefs.set(*AS); + } + if (PhysRegClobbers.test(Reg)) // 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; } // Only consider reloads for now and remats which do not have register @@ -345,33 +505,59 @@ void MachineLICM::ProcessMI(MachineInstr *MI, /// HoistRegionPostRA - Walk the specified region of the CFG and hoist loop /// invariants out to the preheader. void MachineLICM::HoistRegionPostRA() { + MachineBasicBlock *Preheader = getCurPreheader(); + if (!Preheader) + return; + unsigned NumRegs = TRI->getNumRegs(); - unsigned *PhysRegDefs = new unsigned[NumRegs]; - std::fill(PhysRegDefs, PhysRegDefs + NumRegs, 0); + BitVector PhysRegDefs(NumRegs); // Regs defined once in the loop. + BitVector PhysRegClobbers(NumRegs); // Regs defined more than once. 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(); + const std::vector &Blocks = CurLoop->getBlocks(); for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { MachineBasicBlock *BB = Blocks[i]; + + // If the header of the loop containing this basic block is a landing pad, + // then don't try to hoist instructions out of this loop. + const MachineLoop *ML = MLI->getLoopFor(BB); + if (ML && ML->getHeader()->isLandingPad()) continue; + // 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 (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) + PhysRegDefs.set(*AI); } + SpeculationState = SpeculateUnknown; for (MachineBasicBlock::iterator MII = BB->begin(), E = BB->end(); MII != E; ++MII) { MachineInstr *MI = &*MII; - ProcessMI(MI, PhysRegDefs, StoredFIs, Candidates); + ProcessMI(MI, PhysRegDefs, PhysRegClobbers, StoredFIs, Candidates); + } + } + + // Gather the registers read / clobbered by the terminator. + BitVector TermRegs(NumRegs); + MachineBasicBlock::iterator TI = Preheader->getFirstTerminator(); + if (TI != Preheader->end()) { + for (unsigned i = 0, e = TI->getNumOperands(); i != e; ++i) { + const MachineOperand &MO = TI->getOperand(i); + if (!MO.isReg()) + continue; + unsigned Reg = MO.getReg(); + if (!Reg) + continue; + for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) + TermRegs.set(*AI); } } @@ -380,19 +566,25 @@ void MachineLICM::HoistRegionPostRA() { // 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. + // 3. Make sure candidate def should not clobber + // registers read by the terminator. Similarly its def should not be + // clobbered by the terminator. 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) { + unsigned Def = Candidates[i].Def; + if (!PhysRegClobbers.test(Def) && !TermRegs.test(Def)) { 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()]) { + unsigned Reg = MO.getReg(); + if (PhysRegDefs.test(Reg) || + PhysRegClobbers.test(Reg)) { // If it's using a non-loop-invariant register, then it's obviously // not safe to hoist. Safe = false; @@ -403,14 +595,12 @@ void MachineLICM::HoistRegionPostRA() { 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(); + const std::vector &Blocks = CurLoop->getBlocks(); for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { MachineBasicBlock *BB = Blocks[i]; if (!BB->isLiveIn(Reg)) @@ -433,26 +623,17 @@ void MachineLICM::AddToLiveIns(unsigned Reg) { /// 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"; - }); + DEBUG(dbgs() << "Hoisting to BB#" << Preheader->getNumber() << " from BB#" + << MI->getParent()->getNumber() << ": " << *MI); // 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 + // 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); @@ -461,54 +642,330 @@ void MachineLICM::HoistPostRA(MachineInstr *MI, unsigned Def) { Changed = true; } -/// 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. +// IsGuaranteedToExecute - Check if this mbb is guaranteed to execute. +// If not then a load from this mbb may not be safe to hoist. +bool MachineLICM::IsGuaranteedToExecute(MachineBasicBlock *BB) { + if (SpeculationState != SpeculateUnknown) + return SpeculationState == SpeculateFalse; + + if (BB != CurLoop->getHeader()) { + // Check loop exiting blocks. + SmallVector CurrentLoopExitingBlocks; + CurLoop->getExitingBlocks(CurrentLoopExitingBlocks); + for (unsigned i = 0, e = CurrentLoopExitingBlocks.size(); i != e; ++i) + if (!DT->dominates(BB, CurrentLoopExitingBlocks[i])) { + SpeculationState = SpeculateTrue; + return false; + } + } + + SpeculationState = SpeculateFalse; + return true; +} + +void MachineLICM::EnterScope(MachineBasicBlock *MBB) { + DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n'); + + // Remember livein register pressure. + BackTrace.push_back(RegPressure); +} + +void MachineLICM::ExitScope(MachineBasicBlock *MBB) { + DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n'); + BackTrace.pop_back(); +} + +/// ExitScopeIfDone - Destroy scope for the MBB that corresponds to the given +/// dominator tree node if its a leaf or all of its children are done. Walk +/// up the dominator tree to destroy ancestors which are now done. +void MachineLICM::ExitScopeIfDone(MachineDomTreeNode *Node, + DenseMap &OpenChildren, + DenseMap &ParentMap) { + if (OpenChildren[Node]) + return; + + // Pop scope. + ExitScope(Node->getBlock()); + + // Now traverse upwards to pop ancestors whose offsprings are all done. + while (MachineDomTreeNode *Parent = ParentMap[Node]) { + unsigned Left = --OpenChildren[Parent]; + if (Left != 0) + break; + ExitScope(Parent->getBlock()); + Node = Parent; + } +} + +/// HoistOutOfLoop - Walk the specified loop in the CFG (defined by all +/// blocks dominated by the specified header 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 MachineLICM::HoistRegion(MachineDomTreeNode *N) { - assert(N != 0 && "Null dominator tree node?"); - MachineBasicBlock *BB = N->getBlock(); +void MachineLICM::HoistOutOfLoop(MachineDomTreeNode *HeaderN) { + MachineBasicBlock *Preheader = getCurPreheader(); + if (!Preheader) + return; + + SmallVector Scopes; + SmallVector WorkList; + DenseMap ParentMap; + DenseMap OpenChildren; + + // Perform a DFS walk to determine the order of visit. + WorkList.push_back(HeaderN); + while (!WorkList.empty()) { + MachineDomTreeNode *Node = WorkList.pop_back_val(); + assert(Node && "Null dominator tree node?"); + MachineBasicBlock *BB = Node->getBlock(); + + // If the header of the loop containing this basic block is a landing pad, + // then don't try to hoist instructions out of this loop. + const MachineLoop *ML = MLI->getLoopFor(BB); + if (ML && ML->getHeader()->isLandingPad()) + continue; + + // If this subregion is not in the top level loop at all, exit. + if (!CurLoop->contains(BB)) + continue; + + Scopes.push_back(Node); + const std::vector &Children = Node->getChildren(); + unsigned NumChildren = Children.size(); + + // 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) + NumChildren = 0; + + OpenChildren[Node] = NumChildren; + // Add children in reverse order as then the next popped worklist node is + // the first child of this node. This means we ultimately traverse the + // DOM tree in exactly the same order as if we'd recursed. + for (int i = (int)NumChildren-1; i >= 0; --i) { + MachineDomTreeNode *Child = Children[i]; + ParentMap[Child] = Node; + WorkList.push_back(Child); + } + } + + if (Scopes.size() == 0) + return; - // If this subregion is not in the top level loop at all, exit. - if (!CurLoop->contains(BB)) return; + // Compute registers which are livein into the loop headers. + RegSeen.clear(); + BackTrace.clear(); + InitRegPressure(Preheader); + + // Now perform LICM. + for (unsigned i = 0, e = Scopes.size(); i != e; ++i) { + MachineDomTreeNode *Node = Scopes[i]; + MachineBasicBlock *MBB = Node->getBlock(); + + EnterScope(MBB); + + // Process the block + SpeculationState = SpeculateUnknown; + for (MachineBasicBlock::iterator + MII = MBB->begin(), E = MBB->end(); MII != E; ) { + MachineBasicBlock::iterator NextMII = MII; ++NextMII; + MachineInstr *MI = &*MII; + if (!Hoist(MI, Preheader)) + UpdateRegPressure(MI); + MII = NextMII; + } + + // If it's a leaf node, it's done. Traverse upwards to pop ancestors. + ExitScopeIfDone(Node, OpenChildren, ParentMap); + } +} + +void MachineLICM::SinkIntoLoop() { + MachineBasicBlock *Preheader = getCurPreheader(); + if (!Preheader) + return; + + SmallVector Candidates; + for (MachineBasicBlock::instr_iterator I = Preheader->instr_begin(); + I != Preheader->instr_end(); ++I) { + // We need to ensure that we can safely move this instruction into the loop. + // As such, it must not have side-effects, e.g. such as a call has. + if (IsLoopInvariantInst(*I) && !HasLoopPHIUse(I)) + Candidates.push_back(I); + } - for (MachineBasicBlock::iterator - MII = BB->begin(), E = BB->end(); MII != E; ) { - MachineBasicBlock::iterator NextMII = MII; ++NextMII; - Hoist(&*MII); - MII = NextMII; + for (MachineInstr *I : Candidates) { + const MachineOperand &MO = I->getOperand(0); + if (!MO.isDef() || !MO.isReg() || !MO.getReg()) + continue; + if (!MRI->hasOneDef(MO.getReg())) + continue; + bool CanSink = true; + MachineBasicBlock *B = nullptr; + for (MachineInstr &MI : MRI->use_instructions(MO.getReg())) { + // FIXME: Come up with a proper cost model that estimates whether sinking + // the instruction (and thus possibly executing it on every loop + // iteration) is more expensive than a register. + // For now assumes that copies are cheap and thus almost always worth it. + if (!MI.isCopy()) { + CanSink = false; + break; + } + if (!B) { + B = MI.getParent(); + continue; + } + B = DT->findNearestCommonDominator(B, MI.getParent()); + if (!B) { + CanSink = false; + break; + } + } + if (!CanSink || !B || B == Preheader) + continue; + B->splice(B->getFirstNonPHI(), Preheader, I); } +} - const std::vector &Children = N->getChildren(); - 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()); +} + +/// getRegisterClassIDAndCost - For a given MI, register, and the operand +/// index, return the ID and cost of its representative register class. +void +MachineLICM::getRegisterClassIDAndCost(const MachineInstr *MI, + unsigned Reg, unsigned OpIdx, + unsigned &RCId, unsigned &RCCost) const { + const TargetRegisterClass *RC = MRI->getRegClass(Reg); + MVT VT = *RC->vt_begin(); + if (VT == MVT::Untyped) { + RCId = RC->getID(); + RCCost = 1; + } else { + RCId = TLI->getRepRegClassFor(VT)->getID(); + RCCost = TLI->getRepRegClassCostFor(VT); + } +} + +/// 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 = nullptr, *FBB = nullptr; + SmallVector Cond; + if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond, false) && Cond.empty()) + InitRegPressure(*BB->pred_begin()); + } + + 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).second; + unsigned RCId, RCCost; + getRegisterClassIDAndCost(MI, Reg, i, RCId, RCCost); + if (MO.isDef()) + RegPressure[RCId] += RCCost; + else { + bool isKill = isOperandKill(MO, MRI); + if (isNew && !isKill) + // Haven't seen this, it must be a livein. + RegPressure[RCId] += RCCost; + else if (!isNew && isKill) + RegPressure[RCId] -= RCCost; + } + } + } +} + +/// UpdateRegPressure - Update estimate of register pressure after the +/// specified instruction. +void MachineLICM::UpdateRegPressure(const MachineInstr *MI) { + if (MI->isImplicitDef()) + return; + + 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; + + bool isNew = RegSeen.insert(Reg).second; + if (MO.isDef()) + Defs.push_back(Reg); + else if (!isNew && isOperandKill(MO, MRI)) { + unsigned RCId, RCCost; + getRegisterClassIDAndCost(MI, Reg, i, RCId, RCCost); + if (RCCost > RegPressure[RCId]) + RegPressure[RCId] = 0; + else + RegPressure[RCId] -= RCCost; + } + } + + unsigned Idx = 0; + while (!Defs.empty()) { + unsigned Reg = Defs.pop_back_val(); + unsigned RCId, RCCost; + getRegisterClassIDAndCost(MI, Reg, Idx, RCId, RCCost); + RegPressure[RCId] += RCCost; + ++Idx; + } +} + +/// isLoadFromGOTOrConstantPool - Return true if this machine instruction +/// loads from global offset table or constant pool. +static bool isLoadFromGOTOrConstantPool(MachineInstr &MI) { + assert (MI.mayLoad() && "Expected MI that loads!"); + for (MachineInstr::mmo_iterator I = MI.memoperands_begin(), + E = MI.memoperands_end(); I != E; ++I) { + if (const PseudoSourceValue *PSV = (*I)->getPseudoValue()) { + if (PSV == PSV->getGOT() || PSV == PSV->getConstantPool()) + return true; + } + } + return false; } /// 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) { - if (I.isImplicitDef()) + // Check if it's safe to move the instruction. + bool DontMoveAcrossStore = true; + if (!I.isSafeToMove(TII, AA, DontMoveAcrossStore)) return false; - const TargetInstrDesc &TID = I.getDesc(); - - // Ignore stuff that we obviously can't hoist. - if (TID.mayStore() || TID.isCall() || TID.isTerminator() || - TID.hasUnmodeledSideEffects()) + // If it is load then check if it is guaranteed to execute by making sure that + // it dominates all exiting blocks. If it doesn't, then there is a path out of + // the loop which does not execute this load, so we can't hoist it. Loads + // from constant memory are not safe to speculate all the time, for example + // indexed load from a jump table. + // Stores and side effects are already checked by isSafeToMove. + if (I.mayLoad() && !isLoadFromGOTOrConstantPool(I) && + !IsGuaranteedToExecute(I.getParent())) 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 (!I.isInvariantLoad(AA)) - // FIXME: we should be able to hoist loads with no other side effects if - // there are no other instructions which can change memory in this loop. - // This is a trivial form of alias analysis. - return false; - } return true; } @@ -516,7 +973,7 @@ bool MachineLICM::IsLICMCandidate(MachineInstr &I) { /// 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; @@ -537,18 +994,8 @@ bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) { // 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 (!RegInfo->def_empty(Reg)) - return false; - if (AllocatableSet.test(Reg)) + if (!MRI->isConstantPhysReg(Reg, *I.getParent()->getParent())) 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 (!RegInfo->def_empty(AliasReg)) - return false; - if (AllocatableSet.test(AliasReg)) - return false; - } // Otherwise it's safe to move. continue; } else if (!MO.isDead()) { @@ -564,12 +1011,12 @@ bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) { 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))) + if (CurLoop->contains(MRI->getVRegDef(Reg))) return false; } @@ -578,71 +1025,281 @@ bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) { } -/// HasPHIUses - Return true if the specified register has any PHI use. -static bool HasPHIUses(unsigned Reg, MachineRegisterInfo *RegInfo) { - for (MachineRegisterInfo::use_iterator UI = RegInfo->use_begin(Reg), - UE = RegInfo->use_end(); UI != UE; ++UI) { - MachineInstr *UseMI = &*UI; - if (UseMI->isPHI()) +/// HasLoopPHIUse - Return true if the specified instruction is used by a +/// phi node and hoisting it could cause a copy to be inserted. +bool MachineLICM::HasLoopPHIUse(const MachineInstr *MI) const { + SmallVector Work(1, MI); + do { + MI = Work.pop_back_val(); + for (ConstMIOperands MO(MI); MO.isValid(); ++MO) { + if (!MO->isReg() || !MO->isDef()) + continue; + unsigned Reg = MO->getReg(); + if (!TargetRegisterInfo::isVirtualRegister(Reg)) + continue; + for (MachineInstr &UseMI : MRI->use_instructions(Reg)) { + // A PHI may cause a copy to be inserted. + if (UseMI.isPHI()) { + // A PHI inside the loop causes a copy because the live range of Reg is + // extended across the PHI. + if (CurLoop->contains(&UseMI)) + return true; + // A PHI in an exit block can cause a copy to be inserted if the PHI + // has multiple predecessors in the loop with different values. + // For now, approximate by rejecting all exit blocks. + if (isExitBlock(UseMI.getParent())) + return true; + continue; + } + // Look past copies as well. + if (UseMI.isCopy() && CurLoop->contains(&UseMI)) + Work.push_back(&UseMI); + } + } + } while (!Work.empty()); + 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 (MachineInstr &UseMI : MRI->use_nodbg_instructions(Reg)) { + 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 (TII->isAsCheapAsAMove(&MI) || 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(const DenseMap& Cost, + bool CheapInstr) { + for (const auto &ClassAndCost : Cost) { + if (ClassAndCost.second <= 0) + continue; + + unsigned Class = ClassAndCost.first; + int Limit = RegLimit[Class]; + + // Don't hoist cheap instructions if they would increase register pressure, + // even if we're under the limit. + if (CheapInstr && !HoistCheapInsts) return true; + + for (const auto &RP : BackTrace) + if (static_cast(RP[Class]) + ClassAndCost.second >= Limit) + return true; } + return false; } -/// isLoadFromConstantMemory - Return true if the given instruction is a -/// load from constant memory. Machine LICM will hoist these even if they are -/// not re-materializable. -bool MachineLICM::isLoadFromConstantMemory(MachineInstr *MI) { - if (!MI->getDesc().mayLoad()) return false; - if (!MI->hasOneMemOperand()) return false; - MachineMemOperand *MMO = *MI->memoperands_begin(); - if (MMO->isVolatile()) return false; - if (!MMO->getValue()) return false; - const PseudoSourceValue *PSV = dyn_cast(MMO->getValue()); - if (PSV) { - MachineFunction &MF = *MI->getParent()->getParent(); - return PSV->isConstant(MF.getFrameInfo()); - } else { - return AA->pointsToConstantMemory(MMO->getValue()); +/// 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; + + unsigned RCId, RCCost; + getRegisterClassIDAndCost(MI, Reg, i, RCId, RCCost); + 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) { + SmallVectorImpl &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) { - // FIXME: For now, only hoist re-materilizable instructions. LICM will - // increase register pressure. We want to make sure it doesn't increase - // spilling. - // 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 (!TII->isTriviallyReMaterializable(&MI, AA)) { - if (!isLoadFromConstantMemory(&MI)) - return false; + if (MI.isImplicitDef()) + return true; + + // Besides removing computation from the loop, hoisting an instruction has + // these effects: + // + // - The value defined by the instruction becomes live across the entire + // loop. This increases register pressure in the loop. + // + // - If the value is used by a PHI in the loop, a copy will be required for + // lowering the PHI after extending the live range. + // + // - When hoisting the last use of a value in the loop, that value no longer + // needs to be live in the loop. This lowers register pressure in the loop. + + bool CheapInstr = IsCheapInstruction(MI); + bool CreatesCopy = HasLoopPHIUse(&MI); + + // Don't hoist a cheap instruction if it would create a copy in the loop. + if (CheapInstr && CreatesCopy) { + DEBUG(dbgs() << "Won't hoist cheap instr with loop PHI use: " << MI); + 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) { + // Rematerializable instructions should always be hoisted since the register + // allocator can just pull them down again when needed. + if (TII->isTriviallyReMaterializable(&MI, AA)) + return true; + + // 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. + // Cheap instructions will only be hoisted if they don't increase register + // pressure at all. + // 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.isDef()) + if (!MO.isReg() || MO.isImplicit()) continue; - if (HasPHIUses(MO.getReg(), RegInfo)) - return false; + unsigned Reg = MO.getReg(); + if (!TargetRegisterInfo::isVirtualRegister(Reg)) + continue; + + unsigned RCId, RCCost; + getRegisterClassIDAndCost(&MI, Reg, i, RCId, RCCost); + if (MO.isDef()) { + if (HasHighOperandLatency(MI, i, Reg)) { + DEBUG(dbgs() << "Hoist High Latency: " << MI); + ++NumHighLatency; + return true; + } + Cost[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. + Cost[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, CheapInstr)) { + DEBUG(dbgs() << "Hoist non-reg-pressure: " << MI); + ++NumLowRP; + return true; + } + + // Don't risk increasing register pressure if it would create copies. + if (CreatesCopy) { + DEBUG(dbgs() << "Won't hoist instr with loop PHI use: " << MI); + return false; + } + + // Do not "speculate" in high register pressure situation. If an + // instruction is not guaranteed to be executed in the loop, it's best to be + // conservative. + if (AvoidSpeculation && + (!IsGuaranteedToExecute(MI.getParent()) && !MayCSE(&MI))) { + DEBUG(dbgs() << "Won't speculate: " << MI); + return false; + } + + // High register pressure situation, only hoist if the instruction is going + // to be remat'ed. + if (!TII->isTriviallyReMaterializable(&MI, AA) && + !MI.isInvariantLoad(AA)) { + DEBUG(dbgs() << "Can't remat / high reg-pressure: " << MI); + return false; } return true; } MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) { + // Don't unfold simple loads. + if (MI->canFoldAsLoad()) + return nullptr; + // 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 (!isLoadFromConstantMemory(MI)) - return 0; + if (!MI->isInvariantLoad(AA)) + return nullptr; // Next determine the register class for a temporary register. unsigned LoadRegIndex; @@ -651,14 +1308,14 @@ MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) { /*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); + if (NewOpc == 0) return nullptr; + const MCInstrDesc &MID = TII->get(NewOpc); + if (MID.getNumDefs() != 1) return nullptr; + MachineFunction &MF = *MI->getParent()->getParent(); + const TargetRegisterClass *RC = TII->getRegClass(MID, LoadRegIndex, TRI, MF); // Ok, we're unfolding. Create a temporary register and do the unfold. - unsigned Reg = RegInfo->createVirtualRegister(RC); + unsigned Reg = MRI->createVirtualRegister(RC); - MachineFunction &MF = *MI->getParent()->getParent(); SmallVector NewMIs; bool Success = TII->unfoldMemoryOperand(MF, MI, Reg, @@ -671,15 +1328,20 @@ MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) { assert(NewMIs.size() == 2 && "Unfolded a load into multiple instructions!"); MachineBasicBlock *MBB = MI->getParent(); - MBB->insert(MI, NewMIs[0]); - MBB->insert(MI, NewMIs[1]); + MachineBasicBlock::iterator Pos = MI; + MBB->insert(Pos, NewMIs[0]); + MBB->insert(Pos, 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; + return nullptr; } + + // 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]; @@ -688,21 +1350,8 @@ MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) { void MachineLICM::InitCSEMap(MachineBasicBlock *BB) { for (MachineBasicBlock::iterator I = BB->begin(),E = BB->end(); I != E; ++I) { const MachineInstr *MI = &*I; - // FIXME: For now, only hoist re-materilizable instructions. LICM will - // increase register pressure. We want to make sure it doesn't increase - // spilling. - if (TII->isTriviallyReMaterializable(MI, AA)) { - 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)); - } - } + unsigned Opcode = MI->getOpcode(); + CSEMap[Opcode].push_back(MI); } } @@ -711,15 +1360,17 @@ 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)) + if (TII->produceSameValue(MI, PrevMI, (PreRegAlloc ? MRI : nullptr))) return PrevMI; } - return 0; + return nullptr; } bool MachineLICM::EliminateCSE(MachineInstr *MI, DenseMap >::iterator &CI) { - if (CI == CSEMap.end()) + // 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)) { @@ -727,6 +1378,7 @@ bool MachineLICM::EliminateCSE(MachineInstr *MI, // Replace virtual registers defined by MI by their counterparts defined // by Dup. + SmallVector Defs; for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { const MachineOperand &MO = MI->getOperand(i); @@ -737,11 +1389,33 @@ bool MachineLICM::EliminateCSE(MachineInstr *MI, "Instructions with different phys regs are not identical!"); if (MO.isReg() && MO.isDef() && - !TargetRegisterInfo::isPhysicalRegister(MO.getReg())) { - RegInfo->replaceRegWith(MO.getReg(), Dup->getOperand(i).getReg()); - RegInfo->clearKillFlags(Dup->getOperand(i).getReg()); + !TargetRegisterInfo::isPhysicalRegister(MO.getReg())) + Defs.push_back(i); + } + + SmallVector OrigRCs; + for (unsigned i = 0, e = Defs.size(); i != e; ++i) { + unsigned Idx = Defs[i]; + unsigned Reg = MI->getOperand(Idx).getReg(); + unsigned DupReg = Dup->getOperand(Idx).getReg(); + OrigRCs.push_back(MRI->getRegClass(DupReg)); + + if (!MRI->constrainRegClass(DupReg, MRI->getRegClass(Reg))) { + // Restore old RCs if more than one defs. + for (unsigned j = 0; j != i; ++j) + MRI->setRegClass(Dup->getOperand(Defs[j]).getReg(), OrigRCs[j]); + return false; } } + + for (unsigned i = 0, e = Defs.size(); i != e; ++i) { + unsigned Idx = Defs[i]; + unsigned Reg = MI->getOperand(Idx).getReg(); + unsigned DupReg = Dup->getOperand(Idx).getReg(); + MRI->replaceRegWith(Reg, DupReg); + MRI->clearKillFlags(DupReg); + } + MI->eraseFromParent(); ++NumCSEed; return true; @@ -749,18 +1423,29 @@ bool MachineLICM::EliminateCSE(MachineInstr *MI, return false; } +/// MayCSE - Return true if the given instruction will be CSE'd if it's +/// hoisted out of the loop. +bool MachineLICM::MayCSE(MachineInstr *MI) { + unsigned Opcode = MI->getOpcode(); + DenseMap >::iterator + CI = CSEMap.find(Opcode); + // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate + // the undef property onto uses. + if (CI == CSEMap.end() || MI->isImplicitDef()) + return false; + + return LookForDuplicate(MI, CI->second) != nullptr; +} + /// 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) { - MachineBasicBlock *Preheader = getCurPreheader(); - if (!Preheader) 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; + if (!MI) return false; } // Now move the instructions to the predecessor, inserting it before any @@ -791,27 +1476,29 @@ void MachineLICM::Hoist(MachineInstr *MI) { // 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()) - RegInfo->clearKillFlags(MO.getReg()); + 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)); - } + else + CSEMap[Opcode].push_back(MI); } ++NumHoisted; Changed = true; + + return true; } MachineBasicBlock *MachineLICM::getCurPreheader() { @@ -820,7 +1507,7 @@ MachineBasicBlock *MachineLICM::getCurPreheader() { // If we've tried to get a preheader and failed, don't try again. if (CurPreheader == reinterpret_cast(-1)) - return 0; + return nullptr; if (!CurPreheader) { CurPreheader = CurLoop->getLoopPreheader(); @@ -828,13 +1515,13 @@ MachineBasicBlock *MachineLICM::getCurPreheader() { MachineBasicBlock *Pred = CurLoop->getLoopPredecessor(); if (!Pred) { CurPreheader = reinterpret_cast(-1); - return 0; + return nullptr; } CurPreheader = Pred->SplitCriticalEdge(CurLoop->getHeader(), this); if (!CurPreheader) { CurPreheader = reinterpret_cast(-1); - return 0; + return nullptr; } } }