MachineLoop *CurLoop; // The current loop we are working on.
MachineBasicBlock *CurPreheader; // The preheader for CurLoop.
+ // Exit blocks for CurLoop.
+ SmallVector<MachineBasicBlock*, 8> ExitBlocks;
+
+ bool isExitBlock(const MachineBasicBlock *MBB) const {
+ return std::find(ExitBlocks.begin(), ExitBlocks.end(), MBB) !=
+ ExitBlocks.end();
+ }
+
// Track 'estimated' register pressure.
SmallSet<unsigned, 32> RegSeen;
SmallVector<unsigned, 8> RegPressure;
///
bool IsLoopInvariantInst(MachineInstr &I);
- /// HasAnyPHIUse - Return true if the specified register is used by any
- /// phi node.
- bool HasAnyPHIUse(unsigned Reg) const;
+ /// 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
/// 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);
+ bool CanCauseHighRegPressure(DenseMap<unsigned, int> &Cost, bool Cheap);
/// UpdateBackTraceRegPressure - Traverse the back trace from header to
/// the current block and update their register pressures to reflect the
while (!Worklist.empty()) {
CurLoop = Worklist.pop_back_val();
CurPreheader = 0;
+ ExitBlocks.clear();
// If this is done before regalloc, only visit outer-most preheader-sporting
// loops.
continue;
}
+ CurLoop->getExitBlocks(ExitBlocks);
+
if (!PreRegAlloc)
HoistRegionPostRA();
else {
}
if (MO.isImplicit()) {
- for (const unsigned *AS = TRI->getOverlaps(Reg); *AS; ++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.
// If we have already seen another instruction that defines the same
// register, then this is not safe. Two defs is indicated by setting a
// PhysRegClobbers bit.
- for (const unsigned *AS = TRI->getOverlaps(Reg); *AS; ++AS) {
+ for (const uint16_t *AS = TRI->getOverlaps(Reg); *AS; ++AS) {
if (PhysRegDefs.test(*AS))
PhysRegClobbers.set(*AS);
if (PhysRegClobbers.test(*AS))
/// 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();
BitVector PhysRegDefs(NumRegs); // Regs defined once in the loop.
BitVector PhysRegClobbers(NumRegs); // Regs defined more than once.
for (MachineBasicBlock::livein_iterator I = BB->livein_begin(),
E = BB->livein_end(); I != E; ++I) {
unsigned Reg = *I;
- for (const unsigned *AS = TRI->getOverlaps(Reg); *AS; ++AS)
+ for (const uint16_t *AS = TRI->getOverlaps(Reg); *AS; ++AS)
PhysRegDefs.set(*AS);
}
}
}
+ // 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 (const uint16_t *AS = TRI->getOverlaps(Reg); *AS; ++AS)
+ TermRegs.set(*AS);
+ }
+ }
+
// Now evaluate whether the potential candidates qualify.
// 1. Check if the candidate defined register is defined by another
// instruction in the loop.
// 2. If the candidate is a load from stack slot (always true for now),
// check if the slot is stored anywhere in the loop.
+ // 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 (!PhysRegClobbers.test(Candidates[i].Def)) {
+ 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.test(MO.getReg()) ||
- PhysRegClobbers.test(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;
/// 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.
}
-/// 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;
+/// 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<const MachineInstr*, 8> 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 (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg),
+ UE = MRI->use_end(); UI != UE; ++UI) {
+ MachineInstr *UseMI = &*UI;
+ // 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;
}
/// 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) {
+bool MachineLICM::CanCauseHighRegPressure(DenseMap<unsigned, int> &Cost,
+ bool CheapInstr) {
for (DenseMap<unsigned, int>::iterator CI = Cost.begin(), CE = Cost.end();
CI != CE; ++CI) {
if (CI->second <= 0)
unsigned RCId = CI->first;
unsigned Limit = RegLimit[RCId];
int Cost = CI->second;
+
+ // Don't hoist cheap instructions if they would increase register pressure,
+ // even if we're under the limit.
+ if (CheapInstr)
+ return true;
+
for (unsigned i = BackTrace.size(); i != 0; --i) {
SmallVector<unsigned, 8> &RP = BackTrace[i-1];
if (RP[RCId] + Cost >= Limit)
if (MI.isImplicitDef())
return true;
- // If the instruction is cheap, only hoist if it is re-materilizable. LICM
- // will increase register pressure. It's probably not worth it if the
- // instruction is cheap.
- // Also hoist loads from constant memory, e.g. load from stubs, GOT. Hoisting
- // these tend to help performance in low register pressure situation. The
- // trade off is it may cause spill in high pressure situation. It will end up
- // adding a store in the loop preheader. But the reload is no more expensive.
- // The side benefit is these loads are frequently CSE'ed.
- if (IsCheapInstruction(MI)) {
- if (!TII->isTriviallyReMaterializable(&MI, AA))
- return false;
- } else {
- // Estimate register pressure to determine whether to LICM the instruction.
- // In low register pressure situation, we can be more aggressive about
- // hoisting. Also, favors hoisting long latency instructions even in
- // moderately high pressure situation.
- // FIXME: If there are long latency loop-invariant instructions inside the
- // loop at this point, why didn't the optimizer's LICM hoist them?
- DenseMap<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 (!TargetRegisterInfo::isVirtualRegister(Reg))
- continue;
+ // 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;
+ }
- unsigned RCId, RCCost;
- getRegisterClassIDAndCost(&MI, Reg, i, RCId, RCCost);
- if (MO.isDef()) {
- if (HasHighOperandLatency(MI, i, Reg)) {
- ++NumHighLatency;
- return true;
- }
+ // 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<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 (!TargetRegisterInfo::isVirtualRegister(Reg))
+ continue;
- 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.
- DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
- if (CI != Cost.end())
- CI->second -= RCCost;
- else
- Cost.insert(std::make_pair(RCId, -RCCost));
+ 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)) {
- ++NumLowRP;
- return true;
- }
+ // 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;
+ }
- // 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;
+ // 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;
+ }
- // High register pressure situation, only hoist if the instruction is going to
- // be remat'ed.
- if (!TII->isTriviallyReMaterializable(&MI, AA) &&
- !MI.isInvariantLoad(AA))
- return false;
+ // 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;
}
- // 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 (HasAnyPHIUse(MO.getReg()))
- 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;