ARMCC::CondCodes Pred, unsigned PredReg, unsigned Scratch,
DebugLoc dl, SmallVector<std::pair<unsigned, bool>, 8> &Regs);
void MergeOpsUpdate(MachineBasicBlock &MBB,
- MachineBasicBlock::iterator MBBI,
+ MemOpQueue &MemOps,
+ unsigned memOpsBegin,
+ unsigned memOpsEnd,
+ unsigned insertAfter,
int Offset,
unsigned Base,
bool BaseKill,
unsigned PredReg,
unsigned Scratch,
DebugLoc dl,
- MemOpQueue &MemOps,
- unsigned memOpsFrom,
- unsigned memOpsTo,
SmallVector<MachineBasicBlock::iterator, 4> &Merges);
void MergeLDR_STR(MachineBasicBlock &MBB, unsigned SIndex, unsigned Base,
int Opcode, unsigned Size,
// success.
void ARMLoadStoreOpt::
MergeOpsUpdate(MachineBasicBlock &MBB,
- MachineBasicBlock::iterator MBBI,
+ MemOpQueue &memOps,
+ unsigned memOpsBegin,
+ unsigned memOpsEnd,
+ unsigned insertAfter,
int Offset,
unsigned Base,
bool BaseKill,
unsigned PredReg,
unsigned Scratch,
DebugLoc dl,
- MemOpQueue &MemOps,
- unsigned memOpsFrom,
- unsigned memOpsTo,
SmallVector<MachineBasicBlock::iterator, 4> &Merges) {
// First calculate which of the registers should be killed by the merged
// instruction.
SmallVector<std::pair<unsigned, bool>, 8> Regs;
- for (unsigned i = memOpsFrom; i < memOpsTo; ++i) {
- const MachineOperand &MO = MemOps[i].MBBI->getOperand(0);
- Regs.push_back(std::make_pair(MO.getReg(), MO.isKill()));
+ const unsigned insertPos = memOps[insertAfter].Position;
+ for (unsigned i = memOpsBegin; i < memOpsEnd; ++i) {
+ const MachineOperand &MO = memOps[i].MBBI->getOperand(0);
+ unsigned Reg = MO.getReg();
+ bool isKill = MO.isKill();
+
+ // If we are inserting the merged operation after an unmerged operation that
+ // uses the same register, make sure to transfer any kill flag.
+ for (unsigned j = memOpsEnd, e = memOps.size(); !isKill && j != e; ++j)
+ if (memOps[j].Position<insertPos) {
+ const MachineOperand &MOJ = memOps[j].MBBI->getOperand(0);
+ if (MOJ.getReg() == Reg && MOJ.isKill())
+ isKill = true;
+ }
+
+ Regs.push_back(std::make_pair(Reg, isKill));
}
- if (!MergeOps(MBB, MBBI, Offset, Base, BaseKill, Opcode,
+ // Try to do the merge.
+ MachineBasicBlock::iterator Loc = memOps[insertAfter].MBBI;
+ Loc++;
+ if (!MergeOps(MBB, Loc, Offset, Base, BaseKill, Opcode,
Pred, PredReg, Scratch, dl, Regs))
return;
// Merge succeeded, update records.
- Merges.push_back(prior(MBBI));
- for (unsigned i = memOpsFrom; i < memOpsTo; ++i) {
- MBB.erase(MemOps[i].MBBI);
- MemOps[i].Merged = true;
+ Merges.push_back(prior(Loc));
+ for (unsigned i = memOpsBegin; i < memOpsEnd; ++i) {
+ // Remove kill flags from any unmerged memops that come before insertPos.
+ if (Regs[i-memOpsBegin].second)
+ for (unsigned j = memOpsEnd, e = memOps.size(); j != e; ++j)
+ if (memOps[j].Position<insertPos) {
+ MachineOperand &MOJ = memOps[j].MBBI->getOperand(0);
+ if (MOJ.getReg() == Regs[i-memOpsBegin].first && MOJ.isKill())
+ MOJ.setIsKill(false);
+ }
+ MBB.erase(memOps[i].MBBI);
+ memOps[i].Merged = true;
}
}
bool isAM4 = isi32Load(Opcode) || isi32Store(Opcode);
int Offset = MemOps[SIndex].Offset;
int SOffset = Offset;
- unsigned Pos = MemOps[SIndex].Position;
+ unsigned insertAfter = SIndex;
MachineBasicBlock::iterator Loc = MemOps[SIndex].MBBI;
DebugLoc dl = Loc->getDebugLoc();
- unsigned PReg = Loc->getOperand(0).getReg();
- unsigned PRegNum = ARMRegisterInfo::getRegisterNumbering(PReg);
+ const MachineOperand &PMO = Loc->getOperand(0);
+ unsigned PReg = PMO.getReg();
+ unsigned PRegNum = PMO.isUndef() ? UINT_MAX
+ : ARMRegisterInfo::getRegisterNumbering(PReg);
for (unsigned i = SIndex+1, e = MemOps.size(); i != e; ++i) {
int NewOffset = MemOps[i].Offset;
- unsigned Reg = MemOps[i].MBBI->getOperand(0).getReg();
- unsigned RegNum = ARMRegisterInfo::getRegisterNumbering(Reg);
+ const MachineOperand &MO = MemOps[i].MBBI->getOperand(0);
+ unsigned Reg = MO.getReg();
+ unsigned RegNum = MO.isUndef() ? UINT_MAX
+ : ARMRegisterInfo::getRegisterNumbering(Reg);
// AM4 - register numbers in ascending order.
// AM5 - consecutive register numbers in ascending order.
- if (NewOffset == Offset + (int)Size &&
+ if (Reg != ARM::SP &&
+ NewOffset == Offset + (int)Size &&
((isAM4 && RegNum > PRegNum) || RegNum == PRegNum+1)) {
Offset += Size;
PRegNum = RegNum;
} else {
// Can't merge this in. Try merge the earlier ones first.
- MergeOpsUpdate(MBB, ++Loc, SOffset, Base, false, Opcode, Pred, PredReg,
- Scratch, dl, MemOps, SIndex, i, Merges);
+ MergeOpsUpdate(MBB, MemOps, SIndex, i, insertAfter, SOffset,
+ Base, false, Opcode, Pred, PredReg, Scratch, dl, Merges);
MergeLDR_STR(MBB, i, Base, Opcode, Size, Pred, PredReg, Scratch,
MemOps, Merges);
return;
}
- if (MemOps[i].Position > Pos) {
- Pos = MemOps[i].Position;
- Loc = MemOps[i].MBBI;
- }
+ if (MemOps[i].Position > MemOps[insertAfter].Position)
+ insertAfter = i;
}
bool BaseKill = Loc->findRegisterUseOperandIdx(Base, true) != -1;
- MergeOpsUpdate(MBB, ++Loc, SOffset, Base, BaseKill, Opcode, Pred, PredReg,
- Scratch, dl, MemOps, SIndex, MemOps.size(), Merges);
+ MergeOpsUpdate(MBB, MemOps, SIndex, MemOps.size(), insertAfter, SOffset,
+ Base, BaseKill, Opcode, Pred, PredReg, Scratch, dl, Merges);
return;
}
/// isMemoryOp - Returns true if instruction is a memory operations (that this
/// pass is capable of operating on).
static bool isMemoryOp(const MachineInstr *MI) {
+ if (MI->hasOneMemOperand()) {
+ const MachineMemOperand *MMO = *MI->memoperands_begin();
+
+ // Don't touch volatile memory accesses - we may be changing their order.
+ if (MMO->isVolatile())
+ return false;
+
+ // Unaligned ldr/str is emulated by some kernels, but unaligned ldm/stm is
+ // not.
+ if (MMO->getAlignment() < 4)
+ return false;
+ }
+
+ // str <undef> could probably be eliminated entirely, but for now we just want
+ // to avoid making a mess of it.
+ // FIXME: Use str <undef> as a wildcard to enable better stm folding.
+ if (MI->getNumOperands() > 0 && MI->getOperand(0).isReg() &&
+ MI->getOperand(0).isUndef())
+ return false;
+
+ // Likewise don't mess with references to undefined addresses.
+ if (MI->getNumOperands() > 1 && MI->getOperand(1).isReg() &&
+ MI->getOperand(1).isUndef())
+ return false;
+
int Opcode = MI->getOpcode();
switch (Opcode) {
default: break;