X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FMachineLICM.cpp;h=428a9d9bde5260a7c8a519f827c370d24d243294;hb=c415af225d9546a66ac9f7368a973e0be25b438d;hp=3f060ccc4ea1dc7187da507ca3461f8c57dc8a07;hpb=134982daa9bcd87f79c357e3a2686804b9baddd9;p=oota-llvm.git diff --git a/lib/CodeGen/MachineLICM.cpp b/lib/CodeGen/MachineLICM.cpp index 3f060ccc4ea..428a9d9bde5 100644 --- a/lib/CodeGen/MachineLICM.cpp +++ b/lib/CodeGen/MachineLICM.cpp @@ -28,20 +28,25 @@ #include "llvm/CodeGen/MachineMemOperand.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/PseudoSourceValue.h" +#include "llvm/MC/MCInstrItineraries.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/Debug.h" #include "llvm/Support/raw_ostream.h" - using namespace llvm; +static cl::opt +AvoidSpeculation("avoid-speculation", + cl::desc("MachineLICM should avoid speculation"), + cl::init(true), cl::Hidden); + STATISTIC(NumHoisted, "Number of machine instructions hoisted out of loops"); STATISTIC(NumLowRP, @@ -55,8 +60,6 @@ STATISTIC(NumPostRAHoisted, namespace { class MachineLICM : public MachineFunctionPass { - bool PreRegAlloc; - const TargetMachine *TM; const TargetInstrInfo *TII; const TargetLowering *TLI; @@ -64,6 +67,7 @@ namespace { const MachineFrameInfo *MFI; MachineRegisterInfo *MRI; const InstrItineraryData *InstrItins; + bool PreRegAlloc; // Various analyses that we use... AliasAnalysis *AA; // Alias analysis info. @@ -76,8 +80,6 @@ namespace { 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; @@ -92,6 +94,17 @@ namespace { // 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() : @@ -106,10 +119,7 @@ namespace { virtual bool runOnMachineFunction(MachineFunction &MF); - const char *getPassName() const { return "Machine Instruction LICM"; } - virtual void getAnalysisUsage(AnalysisUsage &AU) const { - AU.setPreservesCFG(); AU.addRequired(); AU.addRequired(); AU.addRequired(); @@ -150,7 +160,9 @@ 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); @@ -167,13 +179,20 @@ namespace { /// 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); + /// HasAnyPHIUse - Return true if the specified register is used by any + /// phi node. + bool HasAnyPHIUse(unsigned Reg) 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); + 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 @@ -189,13 +208,36 @@ namespace { /// 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. + /// 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); + + 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 HoistRegion(MachineDomTreeNode *N, bool IsHeader = false); + void HoistOutOfLoop(MachineDomTreeNode *LoopHeaderNode); + void HoistRegion(MachineDomTreeNode *N, bool IsHeader); + + /// 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 @@ -206,10 +248,6 @@ namespace { /// specified instruction. void UpdateRegPressure(const MachineInstr *MI); - /// isLoadFromConstantMemory - Return true if the given instruction is a - /// load from constant memory. - bool isLoadFromConstantMemory(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 @@ -228,6 +266,10 @@ 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. /// It returns true if the instruction is hoisted. @@ -245,6 +287,7 @@ namespace { } // end anonymous namespace char MachineLICM::ID = 0; +char &llvm::MachineLICMID = MachineLICM::ID; INITIALIZE_PASS_BEGIN(MachineLICM, "machinelicm", "Machine Loop Invariant Code Motion", false, false) INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) @@ -253,10 +296,6 @@ INITIALIZE_AG_DEPENDENCY(AliasAnalysis) INITIALIZE_PASS_END(MachineLICM, "machinelicm", "Machine Loop Invariant Code Motion", false, false) -FunctionPass *llvm::createMachineLICMPass(bool PreRegAlloc) { - return new MachineLICM(PreRegAlloc); -} - /// LoopIsOuterMostWithPredecessor - Test if the given loop is the outer-most /// loop that has a unique predecessor. static bool LoopIsOuterMostWithPredecessor(MachineLoop *CurLoop) { @@ -272,12 +311,6 @@ static bool LoopIsOuterMostWithPredecessor(MachineLoop *CurLoop) { } bool MachineLICM::runOnMachineFunction(MachineFunction &MF) { - if (PreRegAlloc) - DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: "); - else - DEBUG(dbgs() << "******** Post-regalloc Machine LICM: "); - DEBUG(dbgs() << MF.getFunction()->getName() << " ********\n"); - Changed = FirstInLoop = false; TM = &MF.getTarget(); TII = TM->getInstrInfo(); @@ -286,7 +319,14 @@ bool MachineLICM::runOnMachineFunction(MachineFunction &MF) { MFI = MF.getFrameInfo(); MRI = &MF.getRegInfo(); InstrItins = TM->getInstrItineraryData(); - AllocatableSet = TRI->getAllocatableSet(MF); + + PreRegAlloc = MRI->isSSA(); + + if (PreRegAlloc) + DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: "); + else + DEBUG(dbgs() << "******** Post-regalloc Machine LICM: "); + DEBUG(dbgs() << MF.getFunction()->getName() << " ********\n"); if (PreRegAlloc) { // Estimate register pressure during pre-regalloc pass. @@ -296,7 +336,7 @@ bool MachineLICM::runOnMachineFunction(MachineFunction &MF) { 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); + RegLimit[(*I)->getID()] = TRI->getRegPressureLimit(*I, MF); } // Get our Loop information... @@ -323,7 +363,7 @@ bool MachineLICM::runOnMachineFunction(MachineFunction &MF) { // being hoisted. MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader()); FirstInLoop = true; - HoistRegion(N, true); + HoistOutOfLoop(N); CSEMap.clear(); } } @@ -350,7 +390,8 @@ 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) { bool RuledOut = false; @@ -369,6 +410,13 @@ 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(); @@ -378,7 +426,7 @@ 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; @@ -386,9 +434,8 @@ void MachineLICM::ProcessMI(MachineInstr *MI, } if (MO.isImplicit()) { - ++PhysRegDefs[Reg]; - for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) - ++PhysRegDefs[*AS]; + for (const uint16_t *AS = TRI->getOverlaps(Reg); *AS; ++AS) + PhysRegClobbers.set(*AS); if (!MO.isDead()) // Non-dead implicit def? This cannot be hoisted. RuledOut = true; @@ -405,14 +452,17 @@ 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) - // 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) + // register, then this is not safe. Two defs is indicated by setting a + // PhysRegClobbers bit. + for (const uint16_t *AS = TRI->getOverlaps(Reg); *AS; ++AS) { + if (PhysRegDefs.test(*AS)) + PhysRegClobbers.set(*AS); + if (PhysRegClobbers.test(*AS)) + // MI defined register is seen defined by another instruction in + // the loop, it cannot be a LICM candidate. RuledOut = true; + PhysRegDefs.set(*AS); + } } // Only consider reloads for now and remats which do not have register @@ -429,8 +479,8 @@ void MachineLICM::ProcessMI(MachineInstr *MI, /// invariants out to the preheader. void MachineLICM::HoistRegionPostRA() { 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; @@ -440,21 +490,27 @@ void MachineLICM::HoistRegionPostRA() { 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 (const uint16_t *AS = TRI->getOverlaps(Reg); *AS; ++AS) + PhysRegDefs.set(*AS); } + 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); } } @@ -468,14 +524,15 @@ void MachineLICM::HoistRegionPostRA() { StoredFIs.count(Candidates[i].FI)) continue; - if (PhysRegDefs[Candidates[i].Def] == 1) { + if (!PhysRegClobbers.test(Candidates[i].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()]) { + if (PhysRegDefs.test(MO.getReg()) || + PhysRegClobbers.test(MO.getReg())) { // If it's using a non-loop-invariant register, then it's obviously // not safe to hoist. Safe = false; @@ -486,8 +543,6 @@ void MachineLICM::HoistRegionPostRA() { HoistPostRA(MI, Candidates[i].Def); } } - - delete[] PhysRegDefs; } /// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the current @@ -520,22 +575,14 @@ void MachineLICM::HoistPostRA(MachineInstr *MI, unsigned Def) { // 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); @@ -544,57 +591,170 @@ 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. -/// -void MachineLICM::HoistRegion(MachineDomTreeNode *N, bool IsHeader) { - assert(N != 0 && "Null dominator tree node?"); - MachineBasicBlock *BB = N->getBlock(); +// 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; +} - // If this subregion is not in the top level loop at all, exit. - if (!CurLoop->contains(BB)) return; +void MachineLICM::EnterScope(MachineBasicBlock *MBB) { + DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n'); - MachineBasicBlock *Preheader = getCurPreheader(); - if (!Preheader) + // 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; - if (IsHeader) { + // 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::HoistOutOfLoop(MachineDomTreeNode *HeaderN) { + SmallVector Scopes; + SmallVector WorkList; + DenseMap ParentMap; + DenseMap OpenChildren; + + // Perform a DFS walk to determine the order of visit. + WorkList.push_back(HeaderN); + do { + MachineDomTreeNode *Node = WorkList.pop_back_val(); + assert(Node != 0 && "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); + } + } while (!WorkList.empty()); + + if (Scopes.size() != 0) { + MachineBasicBlock *Preheader = getCurPreheader(); + if (!Preheader) + return; + // Compute registers which are livein into the loop headers. RegSeen.clear(); BackTrace.clear(); InitRegPressure(Preheader); } - // Remember livein register pressure. - BackTrace.push_back(RegPressure); + // Now perform LICM. + for (unsigned i = 0, e = Scopes.size(); i != e; ++i) { + MachineDomTreeNode *Node = Scopes[i]; + MachineBasicBlock *MBB = Node->getBlock(); - for (MachineBasicBlock::iterator - MII = BB->begin(), E = BB->end(); MII != E; ) { - MachineBasicBlock::iterator NextMII = MII; ++NextMII; - MachineInstr *MI = &*MII; - if (!Hoist(MI, Preheader)) - UpdateRegPressure(MI); - MII = NextMII; - } + MachineBasicBlock *Preheader = getCurPreheader(); + if (!Preheader) + continue; - // 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]); - } + EnterScope(MBB); - BackTrace.pop_back(); + // 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); + } } 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); + EVT 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. @@ -620,22 +780,21 @@ void MachineLICM::InitRegPressure(MachineBasicBlock *BB) { if (!MO.isReg() || MO.isImplicit()) continue; unsigned Reg = MO.getReg(); - if (!Reg || TargetRegisterInfo::isPhysicalRegister(Reg)) + 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(); + unsigned RCId, RCCost; + getRegisterClassIDAndCost(MI, Reg, i, RCId, RCCost); if (MO.isDef()) - RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); + RegPressure[RCId] += RCCost; else { bool isKill = isOperandKill(MO, MRI); if (isNew && !isKill) // Haven't seen this, it must be a livein. - RegPressure[RCId] += TLI->getRepRegClassCostFor(VT); + RegPressure[RCId] += RCCost; else if (!isNew && isKill) - RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT); + RegPressure[RCId] -= RCCost; } } } @@ -653,18 +812,15 @@ void MachineLICM::UpdateRegPressure(const MachineInstr *MI) { if (!MO.isReg() || MO.isImplicit()) continue; unsigned Reg = MO.getReg(); - if (!Reg || TargetRegisterInfo::isPhysicalRegister(Reg)) + if (!TargetRegisterInfo::isVirtualRegister(Reg)) continue; 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); - + unsigned RCId, RCCost; + getRegisterClassIDAndCost(MI, Reg, i, RCId, RCCost); if (RCCost > RegPressure[RCId]) RegPressure[RCId] = 0; else @@ -672,16 +828,31 @@ void MachineLICM::UpdateRegPressure(const MachineInstr *MI) { } } + unsigned Idx = 0; 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); + 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 Value *V = (*I)->getValue()) { + if (const PseudoSourceValue *PSV = dyn_cast(V)) + 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. @@ -690,7 +861,17 @@ bool MachineLICM::IsLICMCandidate(MachineInstr &I) { bool DontMoveAcrossStore = true; if (!I.isSafeToMove(TII, AA, DontMoveAcrossStore)) return false; - + + // 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; + return true; } @@ -698,7 +879,7 @@ 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; @@ -719,18 +900,8 @@ 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 (!MRI->def_empty(Reg)) + if (!MRI->isConstantPhysReg(Reg, *I.getParent()->getParent())) 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()) { @@ -760,48 +931,38 @@ bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) { } -/// HasPHIUses - Return true if the specified register has any PHI use. -static bool HasPHIUses(unsigned Reg, MachineRegisterInfo *MRI) { +/// HasAnyPHIUse - Return true if the specified register is used by any +/// phi node. +bool MachineLICM::HasAnyPHIUse(unsigned Reg) const { for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg), UE = MRI->use_end(); UI != UE; ++UI) { MachineInstr *UseMI = &*UI; if (UseMI->isPHI()) return true; + // Look pass copies as well. + if (UseMI->isCopy()) { + unsigned Def = UseMI->getOperand(0).getReg(); + if (TargetRegisterInfo::isVirtualRegister(Def) && + HasAnyPHIUse(Def)) + 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(AliasAnalysis::Location(MMO->getValue(), - MMO->getSize(), - MMO->getTBAAInfo())); - } -} - /// 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) { - if (MRI->use_nodbg_empty(Reg)) + 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) { @@ -823,19 +984,48 @@ bool MachineLICM::HasHighOperandLatency(MachineInstr &MI, 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.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) + if (CI->second <= 0) continue; unsigned RCId = CI->first; + unsigned Limit = RegLimit[RCId]; + int Cost = CI->second; for (unsigned i = BackTrace.size(); i != 0; --i) { SmallVector &RP = BackTrace[i-1]; - if (RP[RCId] + CI->second >= RegLimit[RCId]) + if (RP[RCId] + Cost >= Limit) return true; } } @@ -858,13 +1048,11 @@ void MachineLICM::UpdateBackTraceRegPressure(const MachineInstr *MI) { if (!MO.isReg() || MO.isImplicit()) continue; unsigned Reg = MO.getReg(); - if (!Reg || TargetRegisterInfo::isPhysicalRegister(Reg)) + 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); + unsigned RCId, RCCost; + getRegisterClassIDAndCost(MI, Reg, i, RCId, RCCost); if (MO.isDef()) { DenseMap::iterator CI = Cost.find(RCId); if (CI != Cost.end()) @@ -905,32 +1093,33 @@ bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) { // 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 (MI.getDesc().isAsCheapAsAMove()) { + 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 + // 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 (!Reg || TargetRegisterInfo::isPhysicalRegister(Reg)) + if (!TargetRegisterInfo::isVirtualRegister(Reg)) continue; + + unsigned RCId, RCCost; + getRegisterClassIDAndCost(&MI, Reg, i, RCId, RCCost); 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; @@ -940,10 +1129,6 @@ bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) { // 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; @@ -959,21 +1144,27 @@ bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) { return true; } - // High register pressure situation, only hoist if the instruction is going to - // be remat'ed. + // 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))) + return false; + + // High register pressure situation, only hoist if the instruction is going + // to be remat'ed. if (!TII->isTriviallyReMaterializable(&MI, AA) && - !isLoadFromConstantMemory(&MI)) + !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. + // If result(s) of this instruction is used by PHIs outside of the loop, then + // don't hoist it if the instruction because it will introduce an extra copy. 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)) + if (HasAnyPHIUse(MO.getReg())) return false; } @@ -982,13 +1173,13 @@ bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) { MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) { // Don't unfold simple loads. - if (MI->getDesc().canFoldAsLoad()) + if (MI->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 (!isLoadFromConstantMemory(MI)) + if (!MI->isInvariantLoad(AA)) return 0; // Next determine the register class for a temporary register. @@ -999,9 +1190,9 @@ MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) { /*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); + const MCInstrDesc &MID = TII->get(NewOpc); + if (MID.getNumDefs() != 1) return 0; + const TargetRegisterClass *RC = TII->getRegClass(MID, LoadRegIndex, TRI); // Ok, we're unfolding. Create a temporary register and do the unfold. unsigned Reg = MRI->createVirtualRegister(RC); @@ -1018,8 +1209,9 @@ 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])) { @@ -1039,20 +1231,15 @@ 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(); + 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)); } } } @@ -1062,7 +1249,7 @@ 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 : 0))) return PrevMI; } return 0; @@ -1080,6 +1267,7 @@ 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); @@ -1090,11 +1278,33 @@ bool MachineLICM::EliminateCSE(MachineInstr *MI, "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()); + !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; @@ -1102,6 +1312,20 @@ 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) != 0; +} + /// 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. ///