#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;
STATISTIC(NumHoisted,
const char *getPassName() const { return "Machine Instruction LICM"; }
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.setPreservesCFG();
AU.addRequired<MachineLoopInfo>();
AU.addRequired<MachineDominatorTree>();
AU.addRequired<AliasAnalysis>();
///
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;
- /// IncreaseHighRegPressure - Visit BBs from preheader to current BB, check
- /// if hoisting an instruction of the given cost matrix can cause high
+ /// CanCauseHighRegPressure - Visit BBs from header to current BB,
+ /// check if hoisting an instruction of the given cost matrix can cause high
/// register pressure.
- bool IncreaseHighRegPressure(DenseMap<unsigned, int> &Cost);
+ 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.
/// this does not count live through (livein but not used) registers.
void InitRegPressure(MachineBasicBlock *BB);
- /// UpdateRegPressureBefore / UpdateRegPressureAfter - Update estimate of
- /// register pressure before and after executing a specifi instruction.
- void UpdateRegPressureBefore(const MachineInstr *MI,
- SmallVector<unsigned, 4> &Defs);
- void UpdateRegPressureAfter(SmallVector<unsigned, 4> &Defs);
-
- /// isLoadFromConstantMemory - Return true if the given instruction is a
- /// load from constant memory.
- bool isLoadFromConstantMemory(MachineInstr *MI);
+ /// 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
/// 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, MachineBasicBlock *Preheader);
+ /// 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
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...
return;
if (IsHeader) {
- // Compute registers which are liveout of preheader.
+ // Compute registers which are livein into the loop headers.
RegSeen.clear();
BackTrace.clear();
InitRegPressure(Preheader);
// Remember livein register pressure.
BackTrace.push_back(RegPressure);
- SmallVector<unsigned, 4> Defs;
for (MachineBasicBlock::iterator
MII = BB->begin(), E = BB->end(); MII != E; ) {
MachineBasicBlock::iterator NextMII = MII; ++NextMII;
MachineInstr *MI = &*MII;
-
- assert(Defs.empty());
- UpdateRegPressureBefore(MI, Defs);
- Hoist(MI, Preheader);
- UpdateRegPressureAfter(Defs);
-
+ if (!Hoist(MI, Preheader))
+ UpdateRegPressure(MI);
MII = NextMII;
}
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;
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())
RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
else {
- if (isNew && !MO.isKill())
+ 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 && MO.isKill())
+ else if (!isNew && isKill)
RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
}
}
}
}
-/// UpdateRegPressureBefore / UpdateRegPressureAfter - Update estimate of
-/// register pressure before and after executing a specifi instruction.
-void MachineLICM::UpdateRegPressureBefore(const MachineInstr *MI,
- SmallVector<unsigned, 4> &Defs) {
- bool NoImpact = MI->isImplicitDef() || MI->isPHI();
+/// 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))
+ if (!TargetRegisterInfo::isVirtualRegister(Reg))
continue;
bool isNew = RegSeen.insert(Reg);
- if (NoImpact)
- continue;
-
if (MO.isDef())
Defs.push_back(Reg);
- else {
- if (!isNew && MO.isKill()) {
- const TargetRegisterClass *RC = MRI->getRegClass(Reg);
- EVT VT = *RC->vt_begin();
- unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
- unsigned RCCost = TLI->getRepRegClassCostFor(VT);
+ 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);
- assert(RCCost <= RegPressure[RCId]);
+ if (RCCost > RegPressure[RCId])
+ RegPressure[RCId] = 0;
+ else
RegPressure[RCId] -= RCCost;
- }
}
}
-}
-void MachineLICM::UpdateRegPressureAfter(SmallVector<unsigned, 4> &Defs) {
while (!Defs.empty()) {
unsigned Reg = Defs.pop_back_val();
- RegSeen.insert(Reg);
const TargetRegisterClass *RC = MRI->getRegClass(Reg);
EVT VT = *RC->vt_begin();
unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
}
-/// 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<PseudoSourceValue>(MMO->getValue());
- if (PSV) {
- MachineFunction &MF = *MI->getParent()->getParent();
- return PSV->isConstant(MF.getFrameInfo());
- } else {
- return AA->pointsToConstantMemory(MMO->getValue());
- }
-}
-
/// 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) {
return false;
}
-/// IncreaseHighRegPressure - Visit BBs from preheader to current BB, check
+/// 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::IncreaseHighRegPressure(DenseMap<unsigned, int> &Cost) {
- for (unsigned i = BackTrace.size(); i != 0; --i) {
- bool AnyIncrease = false;
- SmallVector<unsigned, 8> &RP = BackTrace[i-1];
- for (DenseMap<unsigned, int>::iterator CI = Cost.begin(), CE = Cost.end();
- CI != CE; ++CI) {
- if (CI->second <= 0)
- continue;
- AnyIncrease = true;
- unsigned RCId = CI->first;
+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;
}
-
- if (!AnyIncrease)
- // Hoisting the instruction doesn't increase register pressure.
- return false;
}
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 (!TargetRegisterInfo::isVirtualRegister(Reg))
+ continue;
+
+ const TargetRegisterClass *RC = MRI->getRegClass(Reg);
+ EVT VT = *RC->vt_begin();
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ unsigned RCCost = TLI->getRepRegClassCostFor(VT);
+ if (MO.isDef()) {
+ DenseMap<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) {
// 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 {
// 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))
+ if (!TargetRegisterInfo::isVirtualRegister(Reg))
continue;
if (MO.isDef()) {
if (HasHighOperandLatency(MI, i, Reg)) {
unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
unsigned RCCost = TLI->getRepRegClassCostFor(VT);
DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
- // If the instruction is not register pressure neutrail (or better),
- // check if hoisting it will cause high register pressure in BB's
- // leading up to this point.
if (CI != Cost.end())
CI->second += RCCost;
else
Cost.insert(std::make_pair(RCId, RCCost));
- } else if (MO.isKill()) {
+ } 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
+ // neutral.
const TargetRegisterClass *RC = MRI->getRegClass(Reg);
EVT VT = *RC->vt_begin();
unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
}
}
- // Visit BBs from preheader to current BB, if hoisting this doesn't cause
+ // Visit BBs from header to current BB, if hoisting this doesn't cause
// high register pressure, then it's safe to proceed.
- if (!IncreaseHighRegPressure(Cost)) {
+ 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))
+ !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;
}
// 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.
NewMIs[1]->eraseFromParent();
return 0;
}
+
+ // Update register pressure for the unfolded instruction.
+ UpdateRegPressure(NewMIs[1]);
+
// Otherwise we successfully unfolded a load that we can hoist.
MI->eraseFromParent();
return NewMIs[0];
void MachineLICM::InitCSEMap(MachineBasicBlock *BB) {
for (MachineBasicBlock::iterator I = BB->begin(),E = BB->end(); I != E; ++I) {
const MachineInstr *MI = &*I;
- // 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<unsigned, std::vector<const MachineInstr*> >::iterator
- CI = CSEMap.find(Opcode);
- if (CI != CSEMap.end())
- CI->second.push_back(MI);
- else {
- std::vector<const MachineInstr*> CSEMIs;
- CSEMIs.push_back(MI);
- CSEMap.insert(std::make_pair(Opcode, CSEMIs));
- }
+ unsigned Opcode = MI->getOpcode();
+ DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator
+ CI = CSEMap.find(Opcode);
+ if (CI != CSEMap.end())
+ CI->second.push_back(MI);
+ else {
+ std::vector<const MachineInstr*> CSEMIs;
+ CSEMIs.push_back(MI);
+ CSEMap.insert(std::make_pair(Opcode, CSEMIs));
}
}
}
std::vector<const MachineInstr*> &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;
/// 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) {
+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
// 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.
++NumHoisted;
Changed = true;
+
+ return true;
}
MachineBasicBlock *MachineLICM::getCurPreheader() {