/// foldMemoryOperand - Try folding stack slot references in Ops into MI.
/// Return true on success, and MI will be erased.
bool InlineSpiller::foldMemoryOperand(MachineBasicBlock::iterator MI,
const SmallVectorImpl<unsigned> &Ops) {
// TargetInstrInfo::foldMemoryOperand only expects explicit, non-tied
// operands.
SmallVector<unsigned, 8> FoldOps;
for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
unsigned Idx = Ops[i];
MachineOperand &MO = MI->getOperand(Idx);
if (MO.isImplicit())
continue;
// FIXME: Teach targets to deal with subregs.
if (MO.getSubReg())
return false;
// Tied use operands should not be passed to foldMemoryOperand.
if (!MI->isRegTiedToDefOperand(Idx))
FoldOps.push_back(Idx);
}
MachineInstr *FoldMI = tii_.foldMemoryOperand(MI, FoldOps, stackSlot_);
if (!FoldMI)
return false;
lis_.ReplaceMachineInstrInMaps(MI, FoldMI);
vrm_.addSpillSlotUse(stackSlot_, FoldMI);
MI->eraseFromParent();
DEBUG(dbgs() << "\tfolded: " << *FoldMI);
return true;
}
/// runOnMachineFunction - Reduce two-address instructions to two operands.
///
bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
DEBUG(errs() << "Machine Function\n");
const TargetMachine &TM = MF.getTarget();
MRI = &MF.getRegInfo();
TII = TM.getInstrInfo();
TRI = TM.getRegisterInfo();
LV = getAnalysisIfAvailable<LiveVariables>();
AA = &getAnalysis<AliasAnalysis>();
bool MadeChange = false;
DEBUG(errs() << "********** REWRITING TWO-ADDR INSTRS **********\n");
DEBUG(errs() << "********** Function: "
<< MF.getFunction()->getName() << '\n');
// ReMatRegs - Keep track of the registers whose def's are remat'ed.
BitVector ReMatRegs;
ReMatRegs.resize(MRI->getLastVirtReg()+1);
typedef DenseMap<unsigned, SmallVector<std::pair<unsigned, unsigned>, 4> >
TiedOperandMap;
TiedOperandMap TiedOperands(4);
SmallPtrSet<MachineInstr*, 8> Processed;
for (MachineFunction::iterator mbbi = MF.begin(), mbbe = MF.end();
mbbi != mbbe; ++mbbi) {
unsigned Dist = 0;
DistanceMap.clear();
SrcRegMap.clear();
DstRegMap.clear();
Processed.clear();
for (MachineBasicBlock::iterator mi = mbbi->begin(), me = mbbi->end();
mi != me; ) {
MachineBasicBlock::iterator nmi = next(mi);
const TargetInstrDesc &TID = mi->getDesc();
bool FirstTied = true;
DistanceMap.insert(std::make_pair(mi, ++Dist));
ProcessCopy(&*mi, &*mbbi, Processed);
// First scan through all the tied register uses in this instruction
// and record a list of pairs of tied operands for each register.
unsigned NumOps = (mi->getOpcode() == TargetInstrInfo::INLINEASM)
? mi->getNumOperands() : TID.getNumOperands();
for (unsigned SrcIdx = 0; SrcIdx < NumOps; ++SrcIdx) {
unsigned DstIdx = 0;
if (!mi->isRegTiedToDefOperand(SrcIdx, &DstIdx))
continue;
if (FirstTied) {
FirstTied = false;
++NumTwoAddressInstrs;
DEBUG(errs() << '\t' << *mi);
}
assert(mi->getOperand(SrcIdx).isReg() &&
mi->getOperand(SrcIdx).getReg() &&
mi->getOperand(SrcIdx).isUse() &&
"two address instruction invalid");
unsigned regB = mi->getOperand(SrcIdx).getReg();
TiedOperandMap::iterator OI = TiedOperands.find(regB);
if (OI == TiedOperands.end()) {
SmallVector<std::pair<unsigned, unsigned>, 4> TiedPair;
OI = TiedOperands.insert(std::make_pair(regB, TiedPair)).first;
}
OI->second.push_back(std::make_pair(SrcIdx, DstIdx));
}
// Now iterate over the information collected above.
for (TiedOperandMap::iterator OI = TiedOperands.begin(),
OE = TiedOperands.end(); OI != OE; ++OI) {
SmallVector<std::pair<unsigned, unsigned>, 4> &TiedPairs = OI->second;
// If the instruction has a single pair of tied operands, try some
// transformations that may either eliminate the tied operands or
// improve the opportunities for coalescing away the register copy.
if (TiedOperands.size() == 1 && TiedPairs.size() == 1) {
unsigned SrcIdx = TiedPairs[0].first;
unsigned DstIdx = TiedPairs[0].second;
// If the registers are already equal, nothing needs to be done.
if (mi->getOperand(SrcIdx).getReg() ==
mi->getOperand(DstIdx).getReg())
break; // Done with this instruction.
if (TryInstructionTransform(mi, nmi, mbbi, SrcIdx, DstIdx, Dist))
break; // The tied operands have been eliminated.
}
bool RemovedKillFlag = false;
bool AllUsesCopied = true;
unsigned LastCopiedReg = 0;
unsigned regB = OI->first;
for (unsigned tpi = 0, tpe = TiedPairs.size(); tpi != tpe; ++tpi) {
unsigned SrcIdx = TiedPairs[tpi].first;
unsigned DstIdx = TiedPairs[tpi].second;
unsigned regA = mi->getOperand(DstIdx).getReg();
// Grab regB from the instruction because it may have changed if the
// instruction was commuted.
regB = mi->getOperand(SrcIdx).getReg();
if (regA == regB) {
// The register is tied to multiple destinations (or else we would
// not have continued this far), but this use of the register
// already matches the tied destination. Leave it.
AllUsesCopied = false;
continue;
}
LastCopiedReg = regA;
assert(TargetRegisterInfo::isVirtualRegister(regB) &&
"cannot make instruction into two-address form");
#ifndef NDEBUG
// First, verify that we don't have a use of "a" in the instruction
// (a = b + a for example) because our transformation will not
// work. This should never occur because we are in SSA form.
for (unsigned i = 0; i != mi->getNumOperands(); ++i)
assert(i == DstIdx ||
!mi->getOperand(i).isReg() ||
mi->getOperand(i).getReg() != regA);
#endif
// Emit a copy or rematerialize the definition.
const TargetRegisterClass *rc = MRI->getRegClass(regB);
MachineInstr *DefMI = MRI->getVRegDef(regB);
// If it's safe and profitable, remat the definition instead of
// copying it.
if (DefMI &&
DefMI->getDesc().isAsCheapAsAMove() &&
DefMI->isSafeToReMat(TII, regB, AA) &&
isProfitableToReMat(regB, rc, mi, DefMI, mbbi, Dist)){
DEBUG(errs() << "2addr: REMATTING : " << *DefMI << "\n");
unsigned regASubIdx = mi->getOperand(DstIdx).getSubReg();
TII->reMaterialize(*mbbi, mi, regA, regASubIdx, DefMI, TRI);
ReMatRegs.set(regB);
++NumReMats;
} else {
bool Emitted = TII->copyRegToReg(*mbbi, mi, regA, regB, rc, rc);
(void)Emitted;
assert(Emitted && "Unable to issue a copy instruction!\n");
}
MachineBasicBlock::iterator prevMI = prior(mi);
// Update DistanceMap.
DistanceMap.insert(std::make_pair(prevMI, Dist));
DistanceMap[mi] = ++Dist;
DEBUG(errs() << "\t\tprepend:\t" << *prevMI);
MachineOperand &MO = mi->getOperand(SrcIdx);
assert(MO.isReg() && MO.getReg() == regB && MO.isUse() &&
"inconsistent operand info for 2-reg pass");
if (MO.isKill()) {
MO.setIsKill(false);
RemovedKillFlag = true;
}
MO.setReg(regA);
}
if (AllUsesCopied) {
// Replace other (un-tied) uses of regB with LastCopiedReg.
for (unsigned i = 0, e = mi->getNumOperands(); i != e; ++i) {
MachineOperand &MO = mi->getOperand(i);
if (MO.isReg() && MO.getReg() == regB && MO.isUse()) {
if (MO.isKill()) {
MO.setIsKill(false);
RemovedKillFlag = true;
}
MO.setReg(LastCopiedReg);
}
}
// Update live variables for regB.
if (RemovedKillFlag && LV && LV->getVarInfo(regB).removeKill(mi))
LV->addVirtualRegisterKilled(regB, prior(mi));
} else if (RemovedKillFlag) {
// Some tied uses of regB matched their destination registers, so
// regB is still used in this instruction, but a kill flag was
// removed from a different tied use of regB, so now we need to add
// a kill flag to one of the remaining uses of regB.
for (unsigned i = 0, e = mi->getNumOperands(); i != e; ++i) {
MachineOperand &MO = mi->getOperand(i);
if (MO.isReg() && MO.getReg() == regB && MO.isUse()) {
MO.setIsKill(true);
break;
}
}
}
MadeChange = true;
DEBUG(errs() << "\t\trewrite to:\t" << *mi);
}
// Clear TiedOperands here instead of at the top of the loop
// since most instructions do not have tied operands.
TiedOperands.clear();
mi = nmi;
}
}
// Some remat'ed instructions are dead.
int VReg = ReMatRegs.find_first();
while (VReg != -1) {
if (MRI->use_empty(VReg)) {
MachineInstr *DefMI = MRI->getVRegDef(VReg);
DefMI->eraseFromParent();
}
VReg = ReMatRegs.find_next(VReg);
}
return MadeChange;
}
/// reMaterializeFor - Attempt to rematerialize li_->reg before MI instead of
/// reloading it.
bool InlineSpiller::reMaterializeFor(MachineBasicBlock::iterator MI) {
SlotIndex UseIdx = lis_.getInstructionIndex(MI).getUseIndex();
VNInfo *OrigVNI = li_->getVNInfoAt(UseIdx);
if (!OrigVNI) {
DEBUG(dbgs() << "\tadding <undef> flags: ");
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (MO.isReg() && MO.isUse() && MO.getReg() == li_->reg)
MO.setIsUndef();
}
DEBUG(dbgs() << UseIdx << '\t' << *MI);
return true;
}
if (!reMattable_.count(OrigVNI)) {
DEBUG(dbgs() << "\tusing non-remat valno " << OrigVNI->id << ": "
<< UseIdx << '\t' << *MI);
return false;
}
MachineInstr *OrigMI = lis_.getInstructionFromIndex(OrigVNI->def);
if (!allUsesAvailableAt(OrigMI, OrigVNI->def, UseIdx)) {
usedValues_.insert(OrigVNI);
DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << *MI);
return false;
}
// If the instruction also writes li_->reg, it had better not require the same
// register for uses and defs.
bool Reads, Writes;
SmallVector<unsigned, 8> Ops;
tie(Reads, Writes) = MI->readsWritesVirtualRegister(li_->reg, &Ops);
if (Writes) {
for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(Ops[i]);
if (MO.isUse() ? MI->isRegTiedToDefOperand(Ops[i]) : MO.getSubReg()) {
usedValues_.insert(OrigVNI);
DEBUG(dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << *MI);
return false;
}
}
}
// Alocate a new register for the remat.
unsigned NewVReg = mri_.createVirtualRegister(rc_);
vrm_.grow();
LiveInterval &NewLI = lis_.getOrCreateInterval(NewVReg);
NewLI.markNotSpillable();
newIntervals_->push_back(&NewLI);
// Finally we can rematerialize OrigMI before MI.
MachineBasicBlock &MBB = *MI->getParent();
tii_.reMaterialize(MBB, MI, NewLI.reg, 0, OrigMI, tri_);
MachineBasicBlock::iterator RematMI = MI;
SlotIndex DefIdx = lis_.InsertMachineInstrInMaps(--RematMI).getDefIndex();
DEBUG(dbgs() << "\tremat: " << DefIdx << '\t' << *RematMI);
// Replace operands
for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(Ops[i]);
if (MO.isReg() && MO.isUse() && MO.getReg() == li_->reg) {
MO.setReg(NewVReg);
MO.setIsKill();
}
}
DEBUG(dbgs() << "\t " << UseIdx << '\t' << *MI);
VNInfo *DefVNI = NewLI.getNextValue(DefIdx, 0, true,
lis_.getVNInfoAllocator());
NewLI.addRange(LiveRange(DefIdx, UseIdx.getDefIndex(), DefVNI));
DEBUG(dbgs() << "\tinterval: " << NewLI << '\n');
return true;
}