namespace {
struct ARMLoadStoreOpt : public MachineFunctionPass {
static char ID;
- ARMLoadStoreOpt() : MachineFunctionPass(&ID) {}
+ ARMLoadStoreOpt() : MachineFunctionPass(ID) {}
const TargetInstrInfo *TII;
const TargetRegisterInfo *TRI;
private:
struct MemOpQueueEntry {
int Offset;
+ unsigned Reg;
+ bool isKill;
unsigned Position;
MachineBasicBlock::iterator MBBI;
bool Merged;
- MemOpQueueEntry(int o, int p, MachineBasicBlock::iterator i)
- : Offset(o), Position(p), MBBI(i), Merged(false) {}
+ MemOpQueueEntry(int o, unsigned r, bool k, unsigned p,
+ MachineBasicBlock::iterator i)
+ : Offset(o), Reg(r), isKill(k), Position(p), MBBI(i), Merged(false) {}
};
typedef SmallVector<MemOpQueueEntry,8> MemOpQueue;
typedef MemOpQueue::iterator MemOpQueueIter;
static int getLoadStoreMultipleOpcode(int Opcode) {
switch (Opcode) {
- case ARM::LDR:
- NumLDMGened++;
+ case ARM::LDRi12:
+ ++NumLDMGened;
return ARM::LDM;
- case ARM::STR:
- NumSTMGened++;
+ case ARM::STRi12:
+ ++NumSTMGened;
return ARM::STM;
case ARM::t2LDRi8:
case ARM::t2LDRi12:
- NumLDMGened++;
+ ++NumLDMGened;
return ARM::t2LDM;
case ARM::t2STRi8:
case ARM::t2STRi12:
- NumSTMGened++;
+ ++NumSTMGened;
return ARM::t2STM;
case ARM::VLDRS:
- NumVLDMGened++;
+ ++NumVLDMGened;
return ARM::VLDMS;
case ARM::VSTRS:
- NumVSTMGened++;
+ ++NumVSTMGened;
return ARM::VSTMS;
case ARM::VLDRD:
- NumVLDMGened++;
+ ++NumVLDMGened;
return ARM::VLDMD;
case ARM::VSTRD:
- NumVSTMGened++;
+ ++NumVSTMGened;
return ARM::VSTMD;
default: llvm_unreachable("Unhandled opcode!");
}
}
static bool isi32Load(unsigned Opc) {
- return Opc == ARM::LDR || isT2i32Load(Opc);
+ return Opc == ARM::LDRi12 || isT2i32Load(Opc);
}
static bool isT2i32Store(unsigned Opc) {
}
static bool isi32Store(unsigned Opc) {
- return Opc == ARM::STR || isT2i32Store(Opc);
+ return Opc == ARM::STRi12 || isT2i32Store(Opc);
}
/// MergeOps - Create and insert a LDM or STM with Base as base register and
return false;
ARM_AM::AMSubMode Mode = ARM_AM::ia;
- bool isAM4 = isi32Load(Opcode) || isi32Store(Opcode);
- if (isAM4 && Offset == 4) {
- if (isThumb2)
- // Thumb2 does not support ldmib / stmib.
- return false;
+ // VFP and Thumb2 do not support IB or DA modes.
+ bool isNotVFP = isi32Load(Opcode) || isi32Store(Opcode);
+ bool haveIBAndDA = isNotVFP && !isThumb2;
+ if (Offset == 4 && haveIBAndDA)
Mode = ARM_AM::ib;
- } else if (isAM4 && Offset == -4 * (int)NumRegs + 4) {
- if (isThumb2)
- // Thumb2 does not support ldmda / stmda.
- return false;
+ else if (Offset == -4 * (int)NumRegs + 4 && haveIBAndDA)
Mode = ARM_AM::da;
- } else if (isAM4 && Offset == -4 * (int)NumRegs) {
+ else if (Offset == -4 * (int)NumRegs && isNotVFP)
+ // VLDM/VSTM do not support DB mode without also updating the base reg.
Mode = ARM_AM::db;
- } else if (Offset != 0) {
+ else if (Offset != 0) {
// If starting offset isn't zero, insert a MI to materialize a new base.
// But only do so if it is cost effective, i.e. merging more than two
// loads / stores.
BaseKill = true; // New base is always killed right its use.
}
- bool isDPR = (Opcode == ARM::VLDRD || Opcode == ARM::VSTRD);
bool isDef = (isi32Load(Opcode) || Opcode == ARM::VLDRS ||
Opcode == ARM::VLDRD);
Opcode = getLoadStoreMultipleOpcode(Opcode);
- MachineInstrBuilder MIB = (isAM4)
- ? BuildMI(MBB, MBBI, dl, TII->get(Opcode))
- .addReg(Base, getKillRegState(BaseKill))
- .addImm(ARM_AM::getAM4ModeImm(Mode)).addImm(Pred).addReg(PredReg)
- : BuildMI(MBB, MBBI, dl, TII->get(Opcode))
- .addReg(Base, getKillRegState(BaseKill))
- .addImm(ARM_AM::getAM5Opc(Mode, isDPR ? NumRegs<<1 : NumRegs))
- .addImm(Pred).addReg(PredReg);
+ MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, TII->get(Opcode))
+ .addReg(Base, getKillRegState(BaseKill))
+ .addImm(ARM_AM::getAM4ModeImm(Mode)).addImm(Pred).addReg(PredReg);
for (unsigned i = 0; i != NumRegs; ++i)
MIB = MIB.addReg(Regs[i].first, getDefRegState(isDef)
| getKillRegState(Regs[i].second));
// MergeOpsUpdate - call MergeOps and update MemOps and merges accordingly on
// success.
-void ARMLoadStoreOpt::
-MergeOpsUpdate(MachineBasicBlock &MBB,
- MemOpQueue &memOps,
- unsigned memOpsBegin,
- unsigned memOpsEnd,
- unsigned insertAfter,
- int Offset,
- unsigned Base,
- bool BaseKill,
- int Opcode,
- ARMCC::CondCodes Pred,
- unsigned PredReg,
- unsigned Scratch,
- DebugLoc dl,
- SmallVector<MachineBasicBlock::iterator, 4> &Merges) {
+void ARMLoadStoreOpt::MergeOpsUpdate(MachineBasicBlock &MBB,
+ MemOpQueue &memOps,
+ unsigned memOpsBegin, unsigned memOpsEnd,
+ unsigned insertAfter, int Offset,
+ unsigned Base, bool BaseKill,
+ int Opcode,
+ ARMCC::CondCodes Pred, unsigned PredReg,
+ unsigned Scratch,
+ DebugLoc dl,
+ 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;
const unsigned insertPos = memOps[insertAfter].Position;
+
+ SmallSet<unsigned, 4> UnavailRegs;
+ SmallSet<unsigned, 4> KilledRegs;
+ DenseMap<unsigned, unsigned> Killer;
+ for (unsigned i = 0; i < memOpsBegin; ++i) {
+ if (memOps[i].Position < insertPos && memOps[i].isKill) {
+ unsigned Reg = memOps[i].Reg;
+ if (memOps[i].Merged)
+ UnavailRegs.insert(Reg);
+ else {
+ KilledRegs.insert(Reg);
+ Killer[Reg] = i;
+ }
+ }
+ }
+ for (unsigned i = memOpsEnd, e = memOps.size(); i != e; ++i) {
+ if (memOps[i].Position < insertPos && memOps[i].isKill) {
+ unsigned Reg = memOps[i].Reg;
+ KilledRegs.insert(Reg);
+ Killer[Reg] = i;
+ }
+ }
+
+ SmallVector<std::pair<unsigned, bool>, 8> Regs;
for (unsigned i = memOpsBegin; i < memOpsEnd; ++i) {
- const MachineOperand &MO = memOps[i].MBBI->getOperand(0);
- unsigned Reg = MO.getReg();
- bool isKill = MO.isKill();
+ unsigned Reg = memOps[i].Reg;
+ if (UnavailRegs.count(Reg))
+ // Register is killed before and it's not easy / possible to update the
+ // kill marker on already merged instructions. Abort.
+ return;
// 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;
- }
-
+ bool isKill = memOps[i].isKill || KilledRegs.count(Reg);
Regs.push_back(std::make_pair(Reg, isKill));
}
// Try to do the merge.
MachineBasicBlock::iterator Loc = memOps[insertAfter].MBBI;
- Loc++;
+ ++Loc;
if (!MergeOps(MBB, Loc, Offset, Base, BaseKill, Opcode,
Pred, PredReg, Scratch, dl, Regs))
return;
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);
- }
+ if (Regs[i-memOpsBegin].second) {
+ unsigned Reg = Regs[i-memOpsBegin].first;
+ if (KilledRegs.count(Reg)) {
+ unsigned j = Killer[Reg];
+ memOps[j].MBBI->getOperand(0).setIsKill(false);
+ memOps[j].isKill = false;
+ }
+ }
MBB.erase(memOps[i].MBBI);
memOps[i].Merged = true;
}
ARMCC::CondCodes Pred, unsigned PredReg,
unsigned Scratch, MemOpQueue &MemOps,
SmallVector<MachineBasicBlock::iterator, 4> &Merges) {
- bool isAM4 = isi32Load(Opcode) || isi32Store(Opcode);
+ bool isNotVFP = isi32Load(Opcode) || isi32Store(Opcode);
int Offset = MemOps[SIndex].Offset;
int SOffset = Offset;
unsigned insertAfter = SIndex;
const MachineOperand &PMO = Loc->getOperand(0);
unsigned PReg = PMO.getReg();
unsigned PRegNum = PMO.isUndef() ? UINT_MAX
- : ARMRegisterInfo::getRegisterNumbering(PReg);
+ : getARMRegisterNumbering(PReg);
unsigned Count = 1;
for (unsigned i = SIndex+1, e = MemOps.size(); i != e; ++i) {
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.
- // Can only do up to 16 double-word registers per insn.
+ : getARMRegisterNumbering(Reg);
+ // Register numbers must be in ascending order. For VFP, the registers
+ // must also be consecutive and there is a limit of 16 double-word
+ // registers per instruction.
if (Reg != ARM::SP &&
NewOffset == Offset + (int)Size &&
- ((isAM4 && RegNum > PRegNum)
+ ((isNotVFP && RegNum > PRegNum)
|| ((Size < 8 || Count < 16) && RegNum == PRegNum+1))) {
Offset += Size;
PRegNum = RegNum;
return false;
// Make sure the offset fits in 8 bits.
- if (Bytes <= 0 || (Limit && Bytes >= Limit))
+ if (Bytes == 0 || (Limit && Bytes >= Limit))
return false;
unsigned Scale = (MI->getOpcode() == ARM::tSUBspi) ? 4 : 1; // FIXME
MI->getOpcode() != ARM::ADDri)
return false;
- if (Bytes <= 0 || (Limit && Bytes >= Limit))
+ if (Bytes == 0 || (Limit && Bytes >= Limit))
// Make sure the offset fits in 8 bits.
return false;
static inline unsigned getLSMultipleTransferSize(MachineInstr *MI) {
switch (MI->getOpcode()) {
default: return 0;
- case ARM::LDR:
- case ARM::STR:
+ case ARM::LDRi12:
+ case ARM::STRi12:
case ARM::t2LDRi8:
case ARM::t2LDRi12:
case ARM::t2STRi8:
case ARM::STM:
case ARM::t2LDM:
case ARM::t2STM:
- return (MI->getNumOperands() - 4) * 4;
case ARM::VLDMS:
case ARM::VSTMS:
+ return (MI->getNumOperands() - MI->getDesc().getNumOperands() + 1) * 4;
case ARM::VLDMD:
case ARM::VSTMD:
- return ARM_AM::getAM5Offset(MI->getOperand(1).getImm()) * 4;
+ return (MI->getNumOperands() - MI->getDesc().getNumOperands() + 1) * 8;
}
}
ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
int Opcode = MI->getOpcode();
DebugLoc dl = MI->getDebugLoc();
- bool isAM4 = (Opcode == ARM::LDM || Opcode == ARM::t2LDM ||
- Opcode == ARM::STM || Opcode == ARM::t2STM);
bool DoMerge = false;
ARM_AM::AMSubMode Mode = ARM_AM::ia;
- unsigned Offset = 0;
- if (isAM4) {
- // Can't use an updating ld/st if the base register is also a dest
- // register. e.g. ldmdb r0!, {r0, r1, r2}. The behavior is undefined.
- for (unsigned i = 3, e = MI->getNumOperands(); i != e; ++i) {
- if (MI->getOperand(i).getReg() == Base)
- return false;
- }
- Mode = ARM_AM::getAM4SubMode(MI->getOperand(1).getImm());
- } else {
- // VLDM{D|S}, VSTM{D|S} addressing mode 5 ops.
- Mode = ARM_AM::getAM5SubMode(MI->getOperand(1).getImm());
- Offset = ARM_AM::getAM5Offset(MI->getOperand(1).getImm());
+ // Can't use an updating ld/st if the base register is also a dest
+ // register. e.g. ldmdb r0!, {r0, r1, r2}. The behavior is undefined.
+ for (unsigned i = 3, e = MI->getNumOperands(); i != e; ++i) {
+ if (MI->getOperand(i).getReg() == Base)
+ return false;
}
+ Mode = ARM_AM::getAM4SubMode(MI->getOperand(1).getImm());
// Try merging with the previous instruction.
MachineBasicBlock::iterator BeginMBBI = MBB.begin();
MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
while (PrevMBBI != BeginMBBI && PrevMBBI->isDebugValue())
--PrevMBBI;
- if (isAM4) {
- if (Mode == ARM_AM::ia &&
- isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
- DoMerge = true;
- Mode = ARM_AM::db;
- } else if (isAM4 && Mode == ARM_AM::ib &&
- isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
- DoMerge = true;
- Mode = ARM_AM::da;
- }
- } else {
- if (Mode == ARM_AM::ia &&
- isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
- Mode = ARM_AM::db;
- DoMerge = true;
- }
+ if (Mode == ARM_AM::ia &&
+ isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
+ Mode = ARM_AM::db;
+ DoMerge = true;
+ } else if (Mode == ARM_AM::ib &&
+ isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
+ Mode = ARM_AM::da;
+ DoMerge = true;
}
if (DoMerge)
MBB.erase(PrevMBBI);
MachineBasicBlock::iterator NextMBBI = llvm::next(MBBI);
while (NextMBBI != EndMBBI && NextMBBI->isDebugValue())
++NextMBBI;
- if (isAM4) {
- if ((Mode == ARM_AM::ia || Mode == ARM_AM::ib) &&
- isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
- DoMerge = true;
- } else if ((Mode == ARM_AM::da || Mode == ARM_AM::db) &&
- isMatchingDecrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
- DoMerge = true;
- }
- } else {
- if (Mode == ARM_AM::ia &&
- isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
- DoMerge = true;
- }
+ if ((Mode == ARM_AM::ia || Mode == ARM_AM::ib) &&
+ isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
+ DoMerge = true;
+ } else if ((Mode == ARM_AM::da || Mode == ARM_AM::db) &&
+ isMatchingDecrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
+ DoMerge = true;
}
if (DoMerge) {
if (NextMBBI == I) {
unsigned NewOpc = getUpdatingLSMultipleOpcode(Opcode);
MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, TII->get(NewOpc))
.addReg(Base, getDefRegState(true)) // WB base register
- .addReg(Base, getKillRegState(BaseKill));
- if (isAM4) {
- // [t2]LDM_UPD, [t2]STM_UPD
- MIB.addImm(ARM_AM::getAM4ModeImm(Mode))
- .addImm(Pred).addReg(PredReg);
- } else {
- // VLDM[SD}_UPD, VSTM[SD]_UPD
- MIB.addImm(ARM_AM::getAM5Opc(Mode, Offset))
- .addImm(Pred).addReg(PredReg);
- }
+ .addReg(Base, getKillRegState(BaseKill))
+ .addImm(ARM_AM::getAM4ModeImm(Mode))
+ .addImm(Pred).addReg(PredReg);
// Transfer the rest of operands.
for (unsigned OpNum = 4, e = MI->getNumOperands(); OpNum != e; ++OpNum)
MIB.addOperand(MI->getOperand(OpNum));
static unsigned getPreIndexedLoadStoreOpcode(unsigned Opc) {
switch (Opc) {
- case ARM::LDR: return ARM::LDR_PRE;
- case ARM::STR: return ARM::STR_PRE;
+ case ARM::LDRi12: return ARM::LDR_PRE;
+ case ARM::STRi12: return ARM::STR_PRE;
case ARM::VLDRS: return ARM::VLDMS_UPD;
case ARM::VLDRD: return ARM::VLDMD_UPD;
case ARM::VSTRS: return ARM::VSTMS_UPD;
static unsigned getPostIndexedLoadStoreOpcode(unsigned Opc) {
switch (Opc) {
- case ARM::LDR: return ARM::LDR_POST;
- case ARM::STR: return ARM::STR_POST;
+ case ARM::LDRi12: return ARM::LDR_POST;
+ case ARM::STRi12: return ARM::STR_POST;
case ARM::VLDRS: return ARM::VLDMS_UPD;
case ARM::VLDRD: return ARM::VLDMD_UPD;
case ARM::VSTRS: return ARM::VSTMS_UPD;
DebugLoc dl = MI->getDebugLoc();
bool isAM5 = (Opcode == ARM::VLDRD || Opcode == ARM::VLDRS ||
Opcode == ARM::VSTRD || Opcode == ARM::VSTRS);
- bool isAM2 = (Opcode == ARM::LDR || Opcode == ARM::STR);
- if (isAM2 && ARM_AM::getAM2Offset(MI->getOperand(3).getImm()) != 0)
- return false;
- if (isAM5 && ARM_AM::getAM5Offset(MI->getOperand(2).getImm()) != 0)
- return false;
- if (isT2i32Load(Opcode) || isT2i32Store(Opcode))
+ bool isAM2 = (Opcode == ARM::LDRi12 || Opcode == ARM::STRi12);
+ if (isi32Load(Opcode) || isi32Store(Opcode))
if (MI->getOperand(2).getImm() != 0)
return false;
+ if (isAM5 && ARM_AM::getAM5Offset(MI->getOperand(2).getImm()) != 0)
+ return false;
bool isLd = isi32Load(Opcode) || Opcode == ARM::VLDRS || Opcode == ARM::VLDRD;
// Can't do the merge if the destination register is the same as the would-be
if (!DoMerge)
return false;
- bool isDPR = NewOpc == ARM::VLDMD || NewOpc == ARM::VSTMD;
unsigned Offset = 0;
if (isAM5)
- Offset = ARM_AM::getAM5Opc(AddSub == ARM_AM::sub ? ARM_AM::db : ARM_AM::ia,
- (isDPR ? 2 : 1));
+ Offset = ARM_AM::getAM4ModeImm(AddSub == ARM_AM::sub ?
+ ARM_AM::db : ARM_AM::ia);
else if (isAM2)
Offset = ARM_AM::getAM2Opc(AddSub, Bytes, ARM_AM::no_shift);
else
if (isAM5) {
// VLDM[SD}_UPD, VSTM[SD]_UPD
+ // (There are no base-updating versions of VLDR/VSTR instructions, but the
+ // updating load/store-multiple instructions can be used with only one
+ // register.)
MachineOperand &MO = MI->getOperand(0);
BuildMI(MBB, MBBI, dl, TII->get(NewOpc))
.addReg(Base, getDefRegState(true)) // WB base register
/// 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();
+ // When no memory operands are present, conservatively assume unaligned,
+ // volatile, unfoldable.
+ if (!MI->hasOneMemOperand())
+ return false;
- // Don't touch volatile memory accesses - we may be changing their order.
- if (MMO->isVolatile())
- return false;
+ const MachineMemOperand *MMO = *MI->memoperands_begin();
- // Unaligned ldr/str is emulated by some kernels, but unaligned ldm/stm is
- // not.
- if (MMO->getAlignment() < 4)
- return false;
- }
+ // 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.
int Opcode = MI->getOpcode();
switch (Opcode) {
default: break;
- case ARM::LDR:
- case ARM::STR:
- return MI->getOperand(1).isReg() && MI->getOperand(2).getReg() == 0;
case ARM::VLDRS:
case ARM::VSTRS:
return MI->getOperand(1).isReg();
case ARM::VLDRD:
case ARM::VSTRD:
return MI->getOperand(1).isReg();
+ case ARM::LDRi12:
+ case ARM::STRi12:
case ARM::t2LDRi8:
case ARM::t2LDRi12:
case ARM::t2STRi8:
static int getMemoryOpOffset(const MachineInstr *MI) {
int Opcode = MI->getOpcode();
- bool isAM2 = Opcode == ARM::LDR || Opcode == ARM::STR;
bool isAM3 = Opcode == ARM::LDRD || Opcode == ARM::STRD;
unsigned NumOperands = MI->getDesc().getNumOperands();
unsigned OffField = MI->getOperand(NumOperands-3).getImm();
if (Opcode == ARM::t2LDRi12 || Opcode == ARM::t2LDRi8 ||
Opcode == ARM::t2STRi12 || Opcode == ARM::t2STRi8 ||
- Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8)
+ Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8 ||
+ Opcode == ARM::LDRi12 || Opcode == ARM::STRi12)
return OffField;
- int Offset = isAM2
- ? ARM_AM::getAM2Offset(OffField)
- : (isAM3 ? ARM_AM::getAM3Offset(OffField)
- : ARM_AM::getAM5Offset(OffField) * 4);
- if (isAM2) {
- if (ARM_AM::getAM2Op(OffField) == ARM_AM::sub)
- Offset = -Offset;
- } else if (isAM3) {
+ int Offset = isAM3 ? ARM_AM::getAM3Offset(OffField)
+ : ARM_AM::getAM5Offset(OffField) * 4;
+ if (isAM3) {
if (ARM_AM::getAM3Op(OffField) == ARM_AM::sub)
Offset = -Offset;
} else {
static void InsertLDR_STR(MachineBasicBlock &MBB,
MachineBasicBlock::iterator &MBBI,
- int OffImm, bool isDef,
+ int Offset, bool isDef,
DebugLoc dl, unsigned NewOpc,
unsigned Reg, bool RegDeadKill, bool RegUndef,
unsigned BaseReg, bool BaseKill, bool BaseUndef,
- unsigned OffReg, bool OffKill, bool OffUndef,
+ bool OffKill, bool OffUndef,
ARMCC::CondCodes Pred, unsigned PredReg,
const TargetInstrInfo *TII, bool isT2) {
- int Offset = OffImm;
- if (!isT2) {
- if (OffImm < 0)
- Offset = ARM_AM::getAM2Opc(ARM_AM::sub, -OffImm, ARM_AM::no_shift);
- else
- Offset = ARM_AM::getAM2Opc(ARM_AM::add, OffImm, ARM_AM::no_shift);
- }
if (isDef) {
MachineInstrBuilder MIB = BuildMI(MBB, MBBI, MBBI->getDebugLoc(),
TII->get(NewOpc))
.addReg(Reg, getDefRegState(true) | getDeadRegState(RegDeadKill))
.addReg(BaseReg, getKillRegState(BaseKill)|getUndefRegState(BaseUndef));
- if (!isT2)
- MIB.addReg(OffReg, getKillRegState(OffKill)|getUndefRegState(OffUndef));
MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
} else {
MachineInstrBuilder MIB = BuildMI(MBB, MBBI, MBBI->getDebugLoc(),
TII->get(NewOpc))
.addReg(Reg, getKillRegState(RegDeadKill) | getUndefRegState(RegUndef))
.addReg(BaseReg, getKillRegState(BaseKill)|getUndefRegState(BaseUndef));
- if (!isT2)
- MIB.addReg(OffReg, getKillRegState(OffKill)|getUndefRegState(OffUndef));
MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
}
}
if ((EvenRegNum & 1) == 0 && (EvenRegNum + 1) == OddRegNum)
return false;
+ MachineBasicBlock::iterator NewBBI = MBBI;
bool isT2 = Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8;
bool isLd = Opcode == ARM::LDRD || Opcode == ARM::t2LDRDi8;
bool EvenDeadKill = isLd ?
unsigned BaseReg = BaseOp.getReg();
bool BaseKill = BaseOp.isKill();
bool BaseUndef = BaseOp.isUndef();
- unsigned OffReg = isT2 ? 0 : MI->getOperand(3).getReg();
bool OffKill = isT2 ? false : MI->getOperand(3).isKill();
bool OffUndef = isT2 ? false : MI->getOperand(3).isUndef();
int OffImm = getMemoryOpOffset(MI);
unsigned PredReg = 0;
ARMCC::CondCodes Pred = llvm::getInstrPredicate(MI, PredReg);
- if (OddRegNum > EvenRegNum && OffReg == 0 && OffImm == 0) {
+ if (OddRegNum > EvenRegNum && OffImm == 0) {
// Ascending register numbers and no offset. It's safe to change it to a
// ldm or stm.
unsigned NewOpc = (isLd)
getKillRegState(OddDeadKill) | getUndefRegState(OddUndef));
++NumSTRD2STM;
}
+ NewBBI = llvm::prior(MBBI);
} else {
// Split into two instructions.
- assert((!isT2 || !OffReg) &&
- "Thumb2 ldrd / strd does not encode offset register!");
unsigned NewOpc = (isLd)
- ? (isT2 ? (OffImm < 0 ? ARM::t2LDRi8 : ARM::t2LDRi12) : ARM::LDR)
- : (isT2 ? (OffImm < 0 ? ARM::t2STRi8 : ARM::t2STRi12) : ARM::STR);
+ ? (isT2 ? (OffImm < 0 ? ARM::t2LDRi8 : ARM::t2LDRi12) : ARM::LDRi12)
+ : (isT2 ? (OffImm < 0 ? ARM::t2STRi8 : ARM::t2STRi12) : ARM::STRi12);
DebugLoc dl = MBBI->getDebugLoc();
// If this is a load and base register is killed, it may have been
// re-defed by the load, make sure the first load does not clobber it.
if (isLd &&
(BaseKill || OffKill) &&
- (TRI->regsOverlap(EvenReg, BaseReg) ||
- (OffReg && TRI->regsOverlap(EvenReg, OffReg)))) {
- assert(!TRI->regsOverlap(OddReg, BaseReg) &&
- (!OffReg || !TRI->regsOverlap(OddReg, OffReg)));
+ (TRI->regsOverlap(EvenReg, BaseReg))) {
+ assert(!TRI->regsOverlap(OddReg, BaseReg));
InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc,
OddReg, OddDeadKill, false,
- BaseReg, false, BaseUndef, OffReg, false, OffUndef,
+ BaseReg, false, BaseUndef, false, OffUndef,
Pred, PredReg, TII, isT2);
+ NewBBI = llvm::prior(MBBI);
InsertLDR_STR(MBB, MBBI, OffImm, isLd, dl, NewOpc,
EvenReg, EvenDeadKill, false,
- BaseReg, BaseKill, BaseUndef, OffReg, OffKill, OffUndef,
+ BaseReg, BaseKill, BaseUndef, OffKill, OffUndef,
Pred, PredReg, TII, isT2);
} else {
if (OddReg == EvenReg && EvenDeadKill) {
}
InsertLDR_STR(MBB, MBBI, OffImm, isLd, dl, NewOpc,
EvenReg, EvenDeadKill, EvenUndef,
- BaseReg, false, BaseUndef, OffReg, false, OffUndef,
+ BaseReg, false, BaseUndef, false, OffUndef,
Pred, PredReg, TII, isT2);
+ NewBBI = llvm::prior(MBBI);
InsertLDR_STR(MBB, MBBI, OffImm+4, isLd, dl, NewOpc,
OddReg, OddDeadKill, OddUndef,
- BaseReg, BaseKill, BaseUndef, OffReg, OffKill, OffUndef,
+ BaseReg, BaseKill, BaseUndef, OffKill, OffUndef,
Pred, PredReg, TII, isT2);
}
if (isLd)
++NumSTRD2STR;
}
- MBBI = prior(MBBI);
MBB.erase(MI);
+ MBBI = NewBBI;
+ return true;
}
return false;
}
if (isMemOp) {
int Opcode = MBBI->getOpcode();
unsigned Size = getLSMultipleTransferSize(MBBI);
+ const MachineOperand &MO = MBBI->getOperand(0);
+ unsigned Reg = MO.getReg();
+ bool isKill = MO.isDef() ? false : MO.isKill();
unsigned Base = MBBI->getOperand(1).getReg();
unsigned PredReg = 0;
ARMCC::CondCodes Pred = llvm::getInstrPredicate(MBBI, PredReg);
CurrSize = Size;
CurrPred = Pred;
CurrPredReg = PredReg;
- MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
- NumMemOps++;
+ MemOps.push_back(MemOpQueueEntry(Offset, Reg, isKill, Position, MBBI));
+ ++NumMemOps;
Advance = true;
} else {
if (Clobber) {
// No need to match PredReg.
// Continue adding to the queue.
if (Offset > MemOps.back().Offset) {
- MemOps.push_back(MemOpQueueEntry(Offset, Position, MBBI));
- NumMemOps++;
+ MemOps.push_back(MemOpQueueEntry(Offset, Reg, isKill,
+ Position, MBBI));
+ ++NumMemOps;
Advance = true;
} else {
for (MemOpQueueIter I = MemOps.begin(), E = MemOps.end();
I != E; ++I) {
if (Offset < I->Offset) {
- MemOps.insert(I, MemOpQueueEntry(Offset, Position, MBBI));
- NumMemOps++;
+ MemOps.insert(I, MemOpQueueEntry(Offset, Reg, isKill,
+ Position, MBBI));
+ ++NumMemOps;
Advance = true;
break;
} else if (Offset == I->Offset) {
unsigned NewOpc = isThumb2 ? ARM::t2LDM_RET : ARM::LDM_RET;
PrevMI->setDesc(TII->get(NewOpc));
MO.setReg(ARM::PC);
+ PrevMI->copyImplicitOps(&*MBBI);
MBB.erase(MBBI);
return true;
}
namespace {
struct ARMPreAllocLoadStoreOpt : public MachineFunctionPass{
static char ID;
- ARMPreAllocLoadStoreOpt() : MachineFunctionPass(&ID) {}
+ ARMPreAllocLoadStoreOpt() : MachineFunctionPass(ID) {}
const TargetData *TD;
const TargetInstrInfo *TII;
bool CanFormLdStDWord(MachineInstr *Op0, MachineInstr *Op1, DebugLoc &dl,
unsigned &NewOpc, unsigned &EvenReg,
unsigned &OddReg, unsigned &BaseReg,
- unsigned &OffReg, int &Offset,
+ int &Offset,
unsigned &PredReg, ARMCC::CondCodes &Pred,
bool &isT2);
bool RescheduleOps(MachineBasicBlock *MBB,
DebugLoc &dl,
unsigned &NewOpc, unsigned &EvenReg,
unsigned &OddReg, unsigned &BaseReg,
- unsigned &OffReg, int &Offset,
- unsigned &PredReg,
+ int &Offset, unsigned &PredReg,
ARMCC::CondCodes &Pred,
bool &isT2) {
// Make sure we're allowed to generate LDRD/STRD.
// FIXME: VLDRS / VSTRS -> VLDRD / VSTRD
unsigned Scale = 1;
unsigned Opcode = Op0->getOpcode();
- if (Opcode == ARM::LDR)
+ if (Opcode == ARM::LDRi12)
NewOpc = ARM::LDRD;
- else if (Opcode == ARM::STR)
+ else if (Opcode == ARM::STRi12)
NewOpc = ARM::STRD;
else if (Opcode == ARM::t2LDRi8 || Opcode == ARM::t2LDRi12) {
NewOpc = ARM::t2LDRDi8;
} else
return false;
- // Make sure the offset registers match.
- if (!isT2 &&
- (Op0->getOperand(2).getReg() != Op1->getOperand(2).getReg()))
- return false;
-
- // Must sure the base address satisfies i64 ld / st alignment requirement.
+ // Make sure the base address satisfies i64 ld / st alignment requirement.
if (!Op0->hasOneMemOperand() ||
!(*Op0->memoperands_begin())->getValue() ||
(*Op0->memoperands_begin())->isVolatile())
unsigned Align = (*Op0->memoperands_begin())->getAlignment();
const Function *Func = MF->getFunction();
unsigned ReqAlign = STI->hasV6Ops()
- ? TD->getPrefTypeAlignment(Type::getInt64Ty(Func->getContext()))
+ ? TD->getABITypeAlignment(Type::getInt64Ty(Func->getContext()))
: 8; // Pre-v6 need 8-byte align
if (Align < ReqAlign)
return false;
if (EvenReg == OddReg)
return false;
BaseReg = Op0->getOperand(1).getReg();
- if (!isT2)
- OffReg = Op0->getOperand(2).getReg();
Pred = llvm::getInstrPredicate(Op0, PredReg);
dl = Op0->getDebugLoc();
return true;
} else {
// This is the new location for the loads / stores.
MachineBasicBlock::iterator InsertPos = isLd ? FirstOp : LastOp;
- while (InsertPos != MBB->end() && MemOps.count(InsertPos))
+ while (InsertPos != MBB->end()
+ && (MemOps.count(InsertPos) || InsertPos->isDebugValue()))
++InsertPos;
// If we are moving a pair of loads / stores, see if it makes sense
MachineInstr *Op0 = Ops.back();
MachineInstr *Op1 = Ops[Ops.size()-2];
unsigned EvenReg = 0, OddReg = 0;
- unsigned BaseReg = 0, OffReg = 0, PredReg = 0;
+ unsigned BaseReg = 0, PredReg = 0;
ARMCC::CondCodes Pred = ARMCC::AL;
bool isT2 = false;
unsigned NewOpc = 0;
int Offset = 0;
DebugLoc dl;
if (NumMove == 2 && CanFormLdStDWord(Op0, Op1, dl, NewOpc,
- EvenReg, OddReg, BaseReg, OffReg,
+ EvenReg, OddReg, BaseReg,
Offset, PredReg, Pred, isT2)) {
Ops.pop_back();
Ops.pop_back();
.addReg(EvenReg, RegState::Define)
.addReg(OddReg, RegState::Define)
.addReg(BaseReg);
+ // FIXME: We're converting from LDRi12 to an insn that still
+ // uses addrmode2, so we need an explicit offset reg. It should
+ // always by reg0 since we're transforming LDRi12s.
if (!isT2)
- MIB.addReg(OffReg);
+ MIB.addReg(0);
MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
++NumLDRDFormed;
} else {
.addReg(EvenReg)
.addReg(OddReg)
.addReg(BaseReg);
+ // FIXME: We're converting from LDRi12 to an insn that still
+ // uses addrmode2, so we need an explicit offset reg. It should
+ // always by reg0 since we're transforming STRi12s.
if (!isT2)
- MIB.addReg(OffReg);
+ MIB.addReg(0);
MIB.addImm(Offset).addImm(Pred).addReg(PredReg);
++NumSTRDFormed;
}
projects / oota-llvm.git / blobdiff