#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
+#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Target/TargetInstrItineraries.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/ADT/DenseMap.h"
using namespace llvm;
-STATISTIC(NumHoisted, "Number of machine instructions hoisted out of loops");
-STATISTIC(NumCSEed, "Number of hoisted machine instructions CSEed");
+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");
const TargetMachine *TM;
const TargetInstrInfo *TII;
+ const TargetLowering *TLI;
const TargetRegisterInfo *TRI;
const MachineFrameInfo *MFI;
- MachineRegisterInfo *RegInfo;
+ MachineRegisterInfo *MRI;
+ const InstrItineraryData *InstrItins;
// Various analyses that we use...
AliasAnalysis *AA; // Alias analysis info.
// State that is updated as we process loops
bool Changed; // True if a loop is changed.
+ bool FirstInLoop; // True if it's the first LICM in the loop.
MachineLoop *CurLoop; // The current loop we are working on.
MachineBasicBlock *CurPreheader; // The preheader for CurLoop.
BitVector AllocatableSet;
- // For each opcode, keep a list of potentail CSE instructions.
+ // Track 'estimated' register pressure.
+ SmallSet<unsigned, 32> RegSeen;
+ SmallVector<unsigned, 8> RegPressure;
+
+ // Register pressure "limit" per register class. If the pressure
+ // is higher than the limit, then it's considered high.
+ SmallVector<unsigned, 8> RegLimit;
+
+ // Register pressure on path leading from loop preheader to current BB.
+ SmallVector<SmallVector<unsigned, 8>, 16> BackTrace;
+
+ // For each opcode, keep a list of potential CSE instructions.
DenseMap<unsigned, std::vector<const MachineInstr*> > CSEMap;
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) {}
+ MachineFunctionPass(ID), PreRegAlloc(PreRA) {
+ initializeMachineLICMPass(*PassRegistry::getPassRegistry());
+ }
virtual bool runOnMachineFunction(MachineFunction &MF);
const char *getPassName() const { return "Machine Instruction LICM"; }
- // FIXME: Loop preheaders?
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.setPreservesCFG();
AU.addRequired<MachineLoopInfo>();
AU.addRequired<MachineDominatorTree>();
AU.addRequired<AliasAnalysis>();
}
virtual void releaseMemory() {
+ RegSeen.clear();
+ RegPressure.clear();
+ RegLimit.clear();
+ BackTrace.clear();
+ for (DenseMap<unsigned,std::vector<const MachineInstr*> >::iterator
+ CI = CSEMap.begin(), CE = CSEMap.end(); CI != CE; ++CI)
+ CI->second.clear();
CSEMap.clear();
}
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
///
bool IsLoopInvariantInst(MachineInstr &I);
+ /// HasHighOperandLatency - Compute operand latency between a def of 'Reg'
+ /// and an use in the current loop, return true if the target considered
+ /// it 'high'.
+ bool HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx,
+ unsigned Reg) const;
+
+ bool IsCheapInstruction(MachineInstr &MI) const;
+
+ /// CanCauseHighRegPressure - Visit BBs from header to current BB,
+ /// check if hoisting an instruction of the given cost matrix can cause high
+ /// register pressure.
+ bool CanCauseHighRegPressure(DenseMap<unsigned, int> &Cost);
+
+ /// UpdateBackTraceRegPressure - Traverse the back trace from header to
+ /// the current block and update their register pressures to reflect the
+ /// effect of hoisting MI from the current block to the preheader.
+ void UpdateBackTraceRegPressure(const MachineInstr *MI);
+
/// IsProfitableToHoist - Return true if it is potentially profitable to
/// hoist the given loop invariant.
bool IsProfitableToHoist(MachineInstr &MI);
/// visit definitions before uses, allowing us to hoist a loop body in one
/// pass without iteration.
///
- void HoistRegion(MachineDomTreeNode *N);
+ void HoistRegion(MachineDomTreeNode *N, bool IsHeader = false);
+
+ /// InitRegPressure - Find all virtual register references that are liveout
+ /// of the preheader to initialize the starting "register pressure". Note
+ /// this does not count live through (livein but not used) registers.
+ void InitRegPressure(MachineBasicBlock *BB);
+
+ /// UpdateRegPressure - Update estimate of register pressure after the
+ /// specified instruction.
+ void UpdateRegPressure(const MachineInstr *MI);
/// isLoadFromConstantMemory - Return true if the given instruction is a
/// load from constant memory.
/// Hoist - When an instruction is found to only use loop invariant operands
/// that is safe to hoist, this instruction is called to do the dirty work.
- ///
- void Hoist(MachineInstr *MI);
+ /// It returns true if the instruction is hoisted.
+ bool Hoist(MachineInstr *MI, MachineBasicBlock *Preheader);
/// InitCSEMap - Initialize the CSE map with instructions that are in the
/// current loop preheader that may become duplicates of instructions that
/// are hoisted out of the loop.
void InitCSEMap(MachineBasicBlock *BB);
+
+ /// getCurPreheader - Get the preheader for the current loop, splitting
+ /// a critical edge if needed.
+ MachineBasicBlock *getCurPreheader();
};
} // end anonymous namespace
char MachineLICM::ID = 0;
-static RegisterPass<MachineLICM>
-X("machinelicm", "Machine Loop Invariant Code Motion");
+INITIALIZE_PASS_BEGIN(MachineLICM, "machinelicm",
+ "Machine Loop Invariant Code Motion", false, false)
+INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
+INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
+INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
+INITIALIZE_PASS_END(MachineLICM, "machinelicm",
+ "Machine Loop Invariant Code Motion", false, false)
FunctionPass *llvm::createMachineLICMPass(bool PreRegAlloc) {
return new MachineLICM(PreRegAlloc);
}
-/// LoopIsOuterMostWithPreheader - Test if the given loop is the outer-most
-/// loop that has a preheader.
-static bool LoopIsOuterMostWithPreheader(MachineLoop *CurLoop) {
+/// LoopIsOuterMostWithPredecessor - Test if the given loop is the outer-most
+/// loop that has a unique predecessor.
+static bool LoopIsOuterMostWithPredecessor(MachineLoop *CurLoop) {
+ // Check whether this loop even has a unique predecessor.
+ if (!CurLoop->getLoopPredecessor())
+ return false;
+ // Ok, now check to see if any of its outer loops do.
for (MachineLoop *L = CurLoop->getParentLoop(); L; L = L->getParentLoop())
- if (L->getLoopPreheader())
+ if (L->getLoopPredecessor())
return false;
+ // None of them did, so this is the outermost with a unique predecessor.
return true;
}
bool MachineLICM::runOnMachineFunction(MachineFunction &MF) {
if (PreRegAlloc)
- DEBUG(dbgs() << "******** Pre-regalloc Machine LICM ********\n");
+ DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: ");
else
- DEBUG(dbgs() << "******** Post-regalloc Machine LICM ********\n");
+ DEBUG(dbgs() << "******** Post-regalloc Machine LICM: ");
+ DEBUG(dbgs() << MF.getFunction()->getName() << " ********\n");
- Changed = false;
+ Changed = FirstInLoop = false;
TM = &MF.getTarget();
TII = TM->getInstrInfo();
+ TLI = TM->getTargetLowering();
TRI = TM->getRegisterInfo();
MFI = MF.getFrameInfo();
- RegInfo = &MF.getRegInfo();
+ MRI = &MF.getRegInfo();
+ InstrItins = TM->getInstrItineraryData();
AllocatableSet = TRI->getAllocatableSet(MF);
+ if (PreRegAlloc) {
+ // Estimate register pressure during pre-regalloc pass.
+ unsigned NumRC = TRI->getNumRegClasses();
+ RegPressure.resize(NumRC);
+ std::fill(RegPressure.begin(), RegPressure.end(), 0);
+ RegLimit.resize(NumRC);
+ for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
+ E = TRI->regclass_end(); I != E; ++I)
+ RegLimit[(*I)->getID()] = TLI->getRegPressureLimit(*I, MF);
+ }
+
// Get our Loop information...
MLI = &getAnalysis<MachineLoopInfo>();
DT = &getAnalysis<MachineDominatorTree>();
AA = &getAnalysis<AliasAnalysis>();
- for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end(); I != E; ++I){
- CurLoop = *I;
+ SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end());
+ while (!Worklist.empty()) {
+ CurLoop = Worklist.pop_back_val();
+ CurPreheader = 0;
// If this is done before regalloc, only visit outer-most preheader-sporting
// loops.
- if (PreRegAlloc && !LoopIsOuterMostWithPreheader(CurLoop))
- continue;
-
- // Determine the block to which to hoist instructions. If we can't find a
- // suitable loop preheader, we can't do any hoisting.
- //
- // FIXME: We are only hoisting if the basic block coming into this loop
- // has only one successor. This isn't the case in general because we haven't
- // broken critical edges or added preheaders.
- CurPreheader = CurLoop->getLoopPreheader();
- if (!CurPreheader)
+ if (PreRegAlloc && !LoopIsOuterMostWithPredecessor(CurLoop)) {
+ Worklist.append(CurLoop->begin(), CurLoop->end());
continue;
+ }
if (!PreRegAlloc)
HoistRegionPostRA();
// CSEMap is initialized for loop header when the first instruction is
// being hoisted.
MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader());
- HoistRegion(N);
+ FirstInLoop = true;
+ HoistRegion(N, true);
CSEMap.clear();
}
}
delete[] PhysRegDefs;
}
-/// AddToLiveIns - Add register 'Reg' to the livein sets of BBs in the
-/// current loop.
+/// 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<MachineBasicBlock*> Blocks = CurLoop->getBlocks();
- for (unsigned i = 0, e = Blocks.size(); i != e; ++i)
- Blocks[i]->addLiveIn(Reg);
+ for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
+ MachineBasicBlock *BB = Blocks[i];
+ if (!BB->isLiveIn(Reg))
+ BB->addLiveIn(Reg);
+ for (MachineBasicBlock::iterator
+ MII = BB->begin(), E = BB->end(); MII != E; ++MII) {
+ MachineInstr *MI = &*MII;
+ for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+ MachineOperand &MO = MI->getOperand(i);
+ if (!MO.isReg() || !MO.getReg() || MO.isDef()) continue;
+ if (MO.getReg() == Reg || TRI->isSuperRegister(Reg, MO.getReg()))
+ MO.setIsKill(false);
+ }
+ }
+ }
}
/// HoistPostRA - When an instruction is found to only use loop invariant
/// operands that is safe to hoist, this instruction is called to do the
/// dirty work.
void MachineLICM::HoistPostRA(MachineInstr *MI, unsigned Def) {
+ MachineBasicBlock *Preheader = getCurPreheader();
+ if (!Preheader) return;
+
// Now move the instructions to the predecessor, inserting it before any
// terminator instructions.
DEBUG({
dbgs() << "Hoisting " << *MI;
- if (CurPreheader->getBasicBlock())
+ if (Preheader->getBasicBlock())
dbgs() << " to MachineBasicBlock "
- << CurPreheader->getName();
+ << Preheader->getName();
if (MI->getParent()->getBasicBlock())
dbgs() << " from MachineBasicBlock "
<< MI->getParent()->getName();
// Splice the instruction to the preheader.
MachineBasicBlock *MBB = MI->getParent();
- CurPreheader->splice(CurPreheader->getFirstTerminator(), MBB, MI);
+ Preheader->splice(Preheader->getFirstTerminator(), MBB, MI);
// Add register to livein list to all the BBs in the current loop since a
// loop invariant must be kept live throughout the whole loop. This is
/// first order w.r.t the DominatorTree. This allows us to visit definitions
/// before uses, allowing us to hoist a loop body in one pass without iteration.
///
-void MachineLICM::HoistRegion(MachineDomTreeNode *N) {
+void MachineLICM::HoistRegion(MachineDomTreeNode *N, bool IsHeader) {
assert(N != 0 && "Null dominator tree node?");
MachineBasicBlock *BB = N->getBlock();
// If this subregion is not in the top level loop at all, exit.
if (!CurLoop->contains(BB)) return;
+ MachineBasicBlock *Preheader = getCurPreheader();
+ if (!Preheader)
+ return;
+
+ if (IsHeader) {
+ // Compute registers which are livein into the loop headers.
+ RegSeen.clear();
+ BackTrace.clear();
+ InitRegPressure(Preheader);
+ }
+
+ // Remember livein register pressure.
+ BackTrace.push_back(RegPressure);
+
for (MachineBasicBlock::iterator
MII = BB->begin(), E = BB->end(); MII != E; ) {
MachineBasicBlock::iterator NextMII = MII; ++NextMII;
- Hoist(&*MII);
+ MachineInstr *MI = &*MII;
+ if (!Hoist(MI, Preheader))
+ UpdateRegPressure(MI);
MII = NextMII;
}
- const std::vector<MachineDomTreeNode*> &Children = N->getChildren();
- for (unsigned I = 0, E = Children.size(); I != E; ++I)
- HoistRegion(Children[I]);
+ // 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<MachineDomTreeNode*> &Children = N->getChildren();
+ for (unsigned I = 0, E = Children.size(); I != E; ++I)
+ HoistRegion(Children[I]);
+ }
+
+ BackTrace.pop_back();
+}
+
+static bool isOperandKill(const MachineOperand &MO, MachineRegisterInfo *MRI) {
+ return MO.isKill() || MRI->hasOneNonDBGUse(MO.getReg());
+}
+
+/// InitRegPressure - Find all virtual register references that are liveout of
+/// the preheader to initialize the starting "register pressure". Note this
+/// does not count live through (livein but not used) registers.
+void MachineLICM::InitRegPressure(MachineBasicBlock *BB) {
+ std::fill(RegPressure.begin(), RegPressure.end(), 0);
+
+ // If the preheader has only a single predecessor and it ends with a
+ // fallthrough or an unconditional branch, then scan its predecessor for live
+ // defs as well. This happens whenever the preheader is created by splitting
+ // the critical edge from the loop predecessor to the loop header.
+ if (BB->pred_size() == 1) {
+ MachineBasicBlock *TBB = 0, *FBB = 0;
+ SmallVector<MachineOperand, 4> 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 (!Reg || TargetRegisterInfo::isPhysicalRegister(Reg))
+ continue;
+
+ bool isNew = RegSeen.insert(Reg);
+ const TargetRegisterClass *RC = MRI->getRegClass(Reg);
+ EVT VT = *RC->vt_begin();
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ if (MO.isDef())
+ RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
+ else {
+ bool isKill = isOperandKill(MO, MRI);
+ if (isNew && !isKill)
+ // Haven't seen this, it must be a livein.
+ RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
+ else if (!isNew && isKill)
+ RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
+ }
+ }
+ }
+}
+
+/// UpdateRegPressure - Update estimate of register pressure after the
+/// specified instruction.
+void MachineLICM::UpdateRegPressure(const MachineInstr *MI) {
+ if (MI->isImplicitDef())
+ return;
+
+ SmallVector<unsigned, 4> 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 (!Reg || TargetRegisterInfo::isPhysicalRegister(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);
+
+ if (RCCost > RegPressure[RCId])
+ RegPressure[RCId] = 0;
+ else
+ RegPressure[RCId] -= RCCost;
+ }
+ }
+
+ while (!Defs.empty()) {
+ unsigned Reg = Defs.pop_back_val();
+ const TargetRegisterClass *RC = MRI->getRegClass(Reg);
+ EVT VT = *RC->vt_begin();
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ unsigned RCCost = TLI->getRepRegClassCostFor(VT);
+ RegPressure[RCId] += RCCost;
+ }
}
/// IsLICMCandidate - Returns true if the instruction may be a suitable
/// candidate for LICM. e.g. If the instruction is a call, then it's obviously
/// not safe to hoist it.
bool MachineLICM::IsLICMCandidate(MachineInstr &I) {
- 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())
- 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;
}
// 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))
+ if (!MRI->def_empty(Reg))
return false;
if (AllocatableSet.test(Reg))
return false;
// Check for a def among the register's aliases too.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
- if (!RegInfo->def_empty(AliasReg))
+ if (!MRI->def_empty(AliasReg))
return false;
if (AllocatableSet.test(AliasReg))
return false;
if (!MO.isUse())
continue;
- assert(RegInfo->getVRegDef(Reg) &&
+ assert(MRI->getVRegDef(Reg) &&
"Machine instr not mapped for this vreg?!");
// If the loop contains the definition of an operand, then the instruction
// isn't loop invariant.
- if (CurLoop->contains(RegInfo->getVRegDef(Reg)))
+ if (CurLoop->contains(MRI->getVRegDef(Reg)))
return false;
}
/// 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) {
+static bool HasPHIUses(unsigned Reg, MachineRegisterInfo *MRI) {
+ for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg),
+ UE = MRI->use_end(); UI != UE; ++UI) {
MachineInstr *UseMI = &*UI;
if (UseMI->isPHI())
return true;
MachineFunction &MF = *MI->getParent()->getParent();
return PSV->isConstant(MF.getFrameInfo());
} else {
- return AA->pointsToConstantMemory(MMO->getValue());
+ 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) const {
+ if (!InstrItins || InstrItins->isEmpty() || MRI->use_nodbg_empty(Reg))
+ return false;
+
+ for (MachineRegisterInfo::use_nodbg_iterator I = MRI->use_nodbg_begin(Reg),
+ E = MRI->use_nodbg_end(); I != E; ++I) {
+ MachineInstr *UseMI = &*I;
+ if (UseMI->isCopyLike())
+ continue;
+ if (!CurLoop->contains(UseMI->getParent()))
+ continue;
+ for (unsigned i = 0, e = UseMI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = UseMI->getOperand(i);
+ if (!MO.isReg() || !MO.isUse())
+ continue;
+ unsigned MOReg = MO.getReg();
+ if (MOReg != Reg)
+ continue;
+
+ if (TII->hasHighOperandLatency(InstrItins, MRI, &MI, DefIdx, UseMI, i))
+ return true;
+ }
+
+ // Only look at the first in loop use.
+ break;
+ }
+
+ return false;
+}
+
+/// IsCheapInstruction - Return true if the instruction is marked "cheap" or
+/// the operand latency between its def and a use is one or less.
+bool MachineLICM::IsCheapInstruction(MachineInstr &MI) const {
+ if (MI.getDesc().isAsCheapAsAMove() || MI.isCopyLike())
+ return true;
+ if (!InstrItins || InstrItins->isEmpty())
+ return false;
+
+ bool isCheap = false;
+ unsigned NumDefs = MI.getDesc().getNumDefs();
+ for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) {
+ MachineOperand &DefMO = MI.getOperand(i);
+ if (!DefMO.isReg() || !DefMO.isDef())
+ continue;
+ --NumDefs;
+ unsigned Reg = DefMO.getReg();
+ if (TargetRegisterInfo::isPhysicalRegister(Reg))
+ continue;
+
+ if (!TII->hasLowDefLatency(InstrItins, &MI, i))
+ return false;
+ isCheap = true;
+ }
+
+ return isCheap;
+}
+
+/// CanCauseHighRegPressure - Visit BBs from header to current BB, check
+/// if hoisting an instruction of the given cost matrix can cause high
+/// register pressure.
+bool MachineLICM::CanCauseHighRegPressure(DenseMap<unsigned, int> &Cost) {
+ for (DenseMap<unsigned, int>::iterator CI = Cost.begin(), CE = Cost.end();
+ CI != CE; ++CI) {
+ if (CI->second <= 0)
+ continue;
+
+ unsigned RCId = CI->first;
+ for (unsigned i = BackTrace.size(); i != 0; --i) {
+ SmallVector<unsigned, 8> &RP = BackTrace[i-1];
+ if (RP[RCId] + CI->second >= RegLimit[RCId])
+ return true;
+ }
+ }
+
+ return false;
+}
+
+/// UpdateBackTraceRegPressure - Traverse the back trace from header to the
+/// current block and update their register pressures to reflect the effect
+/// of hoisting MI from the current block to the preheader.
+void MachineLICM::UpdateBackTraceRegPressure(const MachineInstr *MI) {
+ if (MI->isImplicitDef())
+ return;
+
+ // First compute the 'cost' of the instruction, i.e. its contribution
+ // to register pressure.
+ DenseMap<unsigned, int> 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))
+ continue;
+
+ const TargetRegisterClass *RC = MRI->getRegClass(Reg);
+ EVT VT = *RC->vt_begin();
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ unsigned RCCost = TLI->getRepRegClassCostFor(VT);
+ if (MO.isDef()) {
+ DenseMap<unsigned, int>::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<unsigned, int>::iterator CI = Cost.find(RCId);
+ if (CI != Cost.end())
+ CI->second -= RCCost;
+ else
+ Cost.insert(std::make_pair(RCId, -RCCost));
+ }
+ }
+
+ // Update register pressure of blocks from loop header to current block.
+ for (unsigned i = 0, e = BackTrace.size(); i != e; ++i) {
+ SmallVector<unsigned, 8> &RP = BackTrace[i];
+ for (DenseMap<unsigned, int>::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.
+ if (MI.isImplicitDef())
+ return true;
+
+ // If the instruction is cheap, only hoist if it is re-materilizable. LICM
+ // will increase register pressure. It's probably not worth it if the
+ // instruction is cheap.
// Also hoist loads from constant memory, e.g. load from stubs, GOT. Hoisting
// these tend to help performance in low register pressure situation. The
// trade off is it may cause spill in high pressure situation. It will end up
// adding a store in the loop preheader. But the reload is no more expensive.
// The side benefit is these loads are frequently CSE'ed.
- if (!TII->isTriviallyReMaterializable(&MI, AA)) {
- if (!isLoadFromConstantMemory(&MI))
+ if (IsCheapInstruction(MI)) {
+ if (!TII->isTriviallyReMaterializable(&MI, AA))
+ return false;
+ } else {
+ // Estimate register pressure to determine whether to LICM the instruction.
+ // In low register pressure situation, we can be more aggressive about
+ // hoisting. Also, favors hoisting long latency instructions even in
+ // moderately high pressure situation.
+ // FIXME: If there are long latency loop-invariant instructions inside the
+ // loop at this point, why didn't the optimizer's LICM hoist them?
+ DenseMap<unsigned, int> 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))
+ continue;
+ if (MO.isDef()) {
+ if (HasHighOperandLatency(MI, i, Reg)) {
+ ++NumHighLatency;
+ return true;
+ }
+
+ const TargetRegisterClass *RC = MRI->getRegClass(Reg);
+ EVT VT = *RC->vt_begin();
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ unsigned RCCost = TLI->getRepRegClassCostFor(VT);
+ DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
+ if (CI != Cost.end())
+ CI->second += RCCost;
+ else
+ Cost.insert(std::make_pair(RCId, RCCost));
+ } else if (isOperandKill(MO, MRI)) {
+ // Is a virtual register use is a kill, hoisting it out of the loop
+ // may actually reduce register pressure or be register pressure
+ // neutral.
+ const TargetRegisterClass *RC = MRI->getRegClass(Reg);
+ EVT VT = *RC->vt_begin();
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ unsigned RCCost = TLI->getRepRegClassCostFor(VT);
+ DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
+ if (CI != Cost.end())
+ CI->second -= RCCost;
+ else
+ Cost.insert(std::make_pair(RCId, -RCCost));
+ }
+ }
+
+ // Visit BBs from header to current BB, if hoisting this doesn't cause
+ // high register pressure, then it's safe to proceed.
+ if (!CanCauseHighRegPressure(Cost)) {
+ ++NumLowRP;
+ return true;
+ }
+
+ // High register pressure situation, only hoist if the instruction is going to
+ // be remat'ed.
+ if (!TII->isTriviallyReMaterializable(&MI, AA) &&
+ !isLoadFromConstantMemory(&MI))
return false;
}
const MachineOperand &MO = MI.getOperand(i);
if (!MO.isReg() || !MO.isDef())
continue;
- if (HasPHIUses(MO.getReg(), RegInfo))
+ if (HasPHIUses(MO.getReg(), MRI))
return false;
}
}
MachineInstr *MachineLICM::ExtractHoistableLoad(MachineInstr *MI) {
+ // Don't unfold simple loads.
+ if (MI->getDesc().canFoldAsLoad())
+ return 0;
+
// If not, we may be able to unfold a load and hoist that.
// First test whether the instruction is loading from an amenable
// memory location.
if (TID.getNumDefs() != 1) return 0;
const TargetRegisterClass *RC = TID.OpInfo[LoadRegIndex].getRegClass(TRI);
// Ok, we're unfolding. Create a temporary register and do the unfold.
- unsigned Reg = RegInfo->createVirtualRegister(RC);
+ unsigned Reg = MRI->createVirtualRegister(RC);
MachineFunction &MF = *MI->getParent()->getParent();
SmallVector<MachineInstr *, 2> NewMIs;
NewMIs[1]->eraseFromParent();
return 0;
}
+
+ // Update register pressure for the unfolded instruction.
+ UpdateRegPressure(NewMIs[1]);
+
// Otherwise we successfully unfolded a load that we can hoist.
MI->eraseFromParent();
return NewMIs[0];
bool MachineLICM::EliminateCSE(MachineInstr *MI,
DenseMap<unsigned, std::vector<const MachineInstr*> >::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)) {
"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());
+ !TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
+ MRI->replaceRegWith(MO.getReg(), Dup->getOperand(i).getReg());
+ MRI->clearKillFlags(Dup->getOperand(i).getReg());
+ }
}
MI->eraseFromParent();
++NumCSEed;
/// 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) {
+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
// terminator instructions.
DEBUG({
dbgs() << "Hoisting " << *MI;
- if (CurPreheader->getBasicBlock())
+ if (Preheader->getBasicBlock())
dbgs() << " to MachineBasicBlock "
- << CurPreheader->getName();
+ << Preheader->getName();
if (MI->getParent()->getBasicBlock())
dbgs() << " from MachineBasicBlock "
<< MI->getParent()->getName();
// If this is the first instruction being hoisted to the preheader,
// initialize the CSE map with potential common expressions.
- InitCSEMap(CurPreheader);
+ if (FirstInLoop) {
+ InitCSEMap(Preheader);
+ FirstInLoop = false;
+ }
// Look for opportunity to CSE the hoisted instruction.
unsigned Opcode = MI->getOpcode();
CI = CSEMap.find(Opcode);
if (!EliminateCSE(MI, CI)) {
// Otherwise, splice the instruction to the preheader.
- CurPreheader->splice(CurPreheader->getFirstTerminator(),MI->getParent(),MI);
+ Preheader->splice(Preheader->getFirstTerminator(),MI->getParent(),MI);
+
+ // Update register pressure for BBs from header to this block.
+ UpdateBackTraceRegPressure(MI);
+
+ // Clear the kill flags of any register this instruction defines,
+ // since they may need to be live throughout the entire loop
+ // rather than just live for part of it.
+ for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+ MachineOperand &MO = MI->getOperand(i);
+ if (MO.isReg() && MO.isDef() && !MO.isDead())
+ MRI->clearKillFlags(MO.getReg());
+ }
// Add to the CSE map.
if (CI != CSEMap.end())
++NumHoisted;
Changed = true;
+
+ return true;
+}
+
+MachineBasicBlock *MachineLICM::getCurPreheader() {
+ // Determine the block to which to hoist instructions. If we can't find a
+ // suitable loop predecessor, we can't do any hoisting.
+
+ // If we've tried to get a preheader and failed, don't try again.
+ if (CurPreheader == reinterpret_cast<MachineBasicBlock *>(-1))
+ return 0;
+
+ if (!CurPreheader) {
+ CurPreheader = CurLoop->getLoopPreheader();
+ if (!CurPreheader) {
+ MachineBasicBlock *Pred = CurLoop->getLoopPredecessor();
+ if (!Pred) {
+ CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
+ return 0;
+ }
+
+ CurPreheader = Pred->SplitCriticalEdge(CurLoop->getHeader(), this);
+ if (!CurPreheader) {
+ CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1);
+ return 0;
+ }
+ }
+ }
+ return CurPreheader;
}