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;
+ 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))
.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, false, isDPR ? NumRegs<<1 : NumRegs))
+ .addImm(ARM_AM::getAM5Opc(Mode, isDPR ? NumRegs<<1 : NumRegs))
.addImm(Pred).addReg(PredReg);
- MIB.addReg(0); // Add optional writeback (0 for now).
for (unsigned i = 0; i != NumRegs; ++i)
MIB = MIB.addReg(Regs[i].first, getDefRegState(isDef)
| getKillRegState(Regs[i].second));
unsigned PReg = PMO.getReg();
unsigned PRegNum = PMO.isUndef() ? UINT_MAX
: ARMRegisterInfo::getRegisterNumbering(PReg);
+ unsigned Count = 1;
for (unsigned i = SIndex+1, e = MemOps.size(); i != e; ++i) {
int NewOffset = MemOps[i].Offset;
: 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.
if (Reg != ARM::SP &&
NewOffset == Offset + (int)Size &&
- ((isAM4 && RegNum > PRegNum) || RegNum == PRegNum+1)) {
+ ((isAM4 && RegNum > PRegNum)
+ || ((Size < 8 || Count < 16) && RegNum == PRegNum+1))) {
Offset += Size;
PRegNum = RegNum;
+ ++Count;
} else {
// Can't merge this in. Try merge the earlier ones first.
MergeOpsUpdate(MBB, MemOps, SIndex, i, insertAfter, SOffset,
case ARM::STM:
case ARM::t2LDM:
case ARM::t2STM:
- return (MI->getNumOperands() - 5) * 4;
+ return (MI->getNumOperands() - 4) * 4;
case ARM::VLDMS:
case ARM::VSTMS:
case ARM::VLDMD:
}
}
+static unsigned getUpdatingLSMultipleOpcode(unsigned Opc) {
+ switch (Opc) {
+ case ARM::LDM: return ARM::LDM_UPD;
+ case ARM::STM: return ARM::STM_UPD;
+ case ARM::t2LDM: return ARM::t2LDM_UPD;
+ case ARM::t2STM: return ARM::t2STM_UPD;
+ case ARM::VLDMS: return ARM::VLDMS_UPD;
+ case ARM::VLDMD: return ARM::VLDMD_UPD;
+ case ARM::VSTMS: return ARM::VSTMS_UPD;
+ case ARM::VSTMD: return ARM::VSTMD_UPD;
+ default: llvm_unreachable("Unhandled opcode!");
+ }
+ return 0;
+}
+
/// MergeBaseUpdateLSMultiple - Fold proceeding/trailing inc/dec of base
/// register into the LDM/STM/VLDM{D|S}/VSTM{D|S} op when possible:
///
MachineBasicBlock::iterator &I) {
MachineInstr *MI = MBBI;
unsigned Base = MI->getOperand(0).getReg();
+ bool BaseKill = MI->getOperand(0).isKill();
unsigned Bytes = getLSMultipleTransferSize(MI);
unsigned PredReg = 0;
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);
- if (isAM4) {
- if (ARM_AM::getAM4WBFlag(MI->getOperand(1).getImm()))
- return false;
+ bool DoMerge = false;
+ ARM_AM::AMSubMode Mode = ARM_AM::ia;
+ unsigned Offset = 0;
- // Can't use the updating AM4 sub-mode if the base register is also a dest
+ 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());
+ }
- ARM_AM::AMSubMode Mode = ARM_AM::getAM4SubMode(MI->getOperand(1).getImm());
- if (MBBI != MBB.begin()) {
- MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
+ // Try merging with the previous instruction.
+ if (MBBI != MBB.begin()) {
+ MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
+ if (isAM4) {
if (Mode == ARM_AM::ia &&
isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
- MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::db, true));
- MI->getOperand(4).setReg(Base);
- MI->getOperand(4).setIsDef();
- MBB.erase(PrevMBBI);
- return true;
- } else if (Mode == ARM_AM::ib &&
+ DoMerge = true;
+ Mode = ARM_AM::db;
+ } else if (isAM4 && Mode == ARM_AM::ib &&
isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
- MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(ARM_AM::da, true));
- MI->getOperand(4).setReg(Base); // WB to base
- MI->getOperand(4).setIsDef();
- MBB.erase(PrevMBBI);
- return true;
+ 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 (DoMerge)
+ MBB.erase(PrevMBBI);
+ }
- if (MBBI != MBB.end()) {
- MachineBasicBlock::iterator NextMBBI = llvm::next(MBBI);
+ // Try merging with the next instruction.
+ if (!DoMerge && MBBI != MBB.end()) {
+ MachineBasicBlock::iterator NextMBBI = llvm::next(MBBI);
+ if (isAM4) {
if ((Mode == ARM_AM::ia || Mode == ARM_AM::ib) &&
isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
- MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
- MI->getOperand(4).setReg(Base); // WB to base
- MI->getOperand(4).setIsDef();
- if (NextMBBI == I) {
- Advance = true;
- ++I;
- }
- MBB.erase(NextMBBI);
- return true;
+ DoMerge = true;
} else if ((Mode == ARM_AM::da || Mode == ARM_AM::db) &&
isMatchingDecrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
- MI->getOperand(1).setImm(ARM_AM::getAM4ModeImm(Mode, true));
- MI->getOperand(4).setReg(Base); // WB to base
- MI->getOperand(4).setIsDef();
- if (NextMBBI == I) {
- Advance = true;
- ++I;
- }
- MBB.erase(NextMBBI);
- return true;
+ DoMerge = true;
}
- }
- } else {
- // VLDM{D|S}, VSTM{D|S} addressing mode 5 ops.
- if (ARM_AM::getAM5WBFlag(MI->getOperand(1).getImm()))
- return false;
-
- ARM_AM::AMSubMode Mode = ARM_AM::getAM5SubMode(MI->getOperand(1).getImm());
- unsigned Offset = ARM_AM::getAM5Offset(MI->getOperand(1).getImm());
- if (MBBI != MBB.begin()) {
- MachineBasicBlock::iterator PrevMBBI = prior(MBBI);
+ } else {
if (Mode == ARM_AM::ia &&
- isMatchingDecrement(PrevMBBI, Base, Bytes, 0, Pred, PredReg)) {
- MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::db, true, Offset));
- MI->getOperand(4).setReg(Base); // WB to base
- MI->getOperand(4).setIsDef();
- MBB.erase(PrevMBBI);
- return true;
+ isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
+ DoMerge = true;
}
}
-
- if (MBBI != MBB.end()) {
- MachineBasicBlock::iterator NextMBBI = llvm::next(MBBI);
- if (Mode == ARM_AM::ia &&
- isMatchingIncrement(NextMBBI, Base, Bytes, 0, Pred, PredReg)) {
- MI->getOperand(1).setImm(ARM_AM::getAM5Opc(ARM_AM::ia, true, Offset));
- MI->getOperand(4).setReg(Base); // WB to base
- MI->getOperand(4).setIsDef();
- if (NextMBBI == I) {
- Advance = true;
- ++I;
- }
- MBB.erase(NextMBBI);
+ if (DoMerge) {
+ if (NextMBBI == I) {
+ Advance = true;
+ ++I;
}
- return true;
+ MBB.erase(NextMBBI);
}
}
- return false;
+ if (!DoMerge)
+ return false;
+
+ 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);
+ }
+ // Transfer the rest of operands.
+ for (unsigned OpNum = 4, e = MI->getNumOperands(); OpNum != e; ++OpNum)
+ MIB.addOperand(MI->getOperand(OpNum));
+ // Transfer memoperands.
+ (*MIB).setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
+
+ MBB.erase(MBBI);
+ return true;
}
static unsigned getPreIndexedLoadStoreOpcode(unsigned Opc) {
switch (Opc) {
case ARM::LDR: return ARM::LDR_PRE;
case ARM::STR: return ARM::STR_PRE;
- case ARM::VLDRS: return ARM::VLDMS;
- case ARM::VLDRD: return ARM::VLDMD;
- case ARM::VSTRS: return ARM::VSTMS;
- case ARM::VSTRD: return ARM::VSTMD;
+ case ARM::VLDRS: return ARM::VLDMS_UPD;
+ case ARM::VLDRD: return ARM::VLDMD_UPD;
+ case ARM::VSTRS: return ARM::VSTMS_UPD;
+ case ARM::VSTRD: return ARM::VSTMD_UPD;
case ARM::t2LDRi8:
case ARM::t2LDRi12:
return ARM::t2LDR_PRE;
switch (Opc) {
case ARM::LDR: return ARM::LDR_POST;
case ARM::STR: return ARM::STR_POST;
- case ARM::VLDRS: return ARM::VLDMS;
- case ARM::VLDRD: return ARM::VLDMD;
- case ARM::VSTRS: return ARM::VSTMS;
- case ARM::VSTRD: return ARM::VSTMD;
+ case ARM::VLDRS: return ARM::VLDMS_UPD;
+ case ARM::VLDRD: return ARM::VLDMD_UPD;
+ case ARM::VSTRS: return ARM::VSTMS_UPD;
+ case ARM::VSTRD: return ARM::VSTMD_UPD;
case ARM::t2LDRi8:
case ARM::t2LDRi12:
return ARM::t2LDR_POST;
unsigned Offset = 0;
if (isAM5)
Offset = ARM_AM::getAM5Opc(AddSub == ARM_AM::sub ? ARM_AM::db : ARM_AM::ia,
- true, (isDPR ? 2 : 1));
+ (isDPR ? 2 : 1));
else if (isAM2)
Offset = ARM_AM::getAM2Opc(AddSub, Bytes, ARM_AM::no_shift);
else
Offset = AddSub == ARM_AM::sub ? -Bytes : Bytes;
if (isAM5) {
- // VLDM[SD}, VSTM[SD]
+ // VLDM[SD}_UPD, VSTM[SD]_UPD
MachineOperand &MO = MI->getOperand(0);
BuildMI(MBB, MBBI, dl, TII->get(NewOpc))
+ .addReg(Base, getDefRegState(true)) // WB base register
.addReg(Base, getKillRegState(isLd ? BaseKill : false))
.addImm(Offset)
.addImm(Pred).addReg(PredReg)
- .addReg(Base, getDefRegState(true)) // WB base register
.addReg(MO.getReg(), (isLd ? getDefRegState(true) :
getKillRegState(MO.isKill())));
} else if (isLd) {
.addReg(BaseReg, getKillRegState(BaseKill))
.addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia))
.addImm(Pred).addReg(PredReg)
- .addReg(0)
.addReg(EvenReg, getDefRegState(isLd) | getDeadRegState(EvenDeadKill))
.addReg(OddReg, getDefRegState(isLd) | getDeadRegState(OddDeadKill));
++NumLDRD2LDM;
.addReg(BaseReg, getKillRegState(BaseKill))
.addImm(ARM_AM::getAM4ModeImm(ARM_AM::ia))
.addImm(Pred).addReg(PredReg)
- .addReg(0)
.addReg(EvenReg,
getKillRegState(EvenDeadKill) | getUndefRegState(EvenUndef))
.addReg(OddReg,
};
}
-/// MergeReturnIntoLDM - If this is a exit BB, try merging the return op
-/// (bx lr) into the preceeding stack restore so it directly restore the value
-/// of LR into pc.
-/// ldmfd sp!, {r7, lr}
+/// MergeReturnIntoLDM - If this is a exit BB, try merging the return ops
+/// ("bx lr" and "mov pc, lr") into the preceeding stack restore so it
+/// directly restore the value of LR into pc.
+/// ldmfd sp!, {..., lr}
/// bx lr
+/// or
+/// ldmfd sp!, {..., lr}
+/// mov pc, lr
/// =>
-/// ldmfd sp!, {r7, pc}
+/// ldmfd sp!, {..., pc}
bool ARMLoadStoreOpt::MergeReturnIntoLDM(MachineBasicBlock &MBB) {
if (MBB.empty()) return false;
MachineBasicBlock::iterator MBBI = prior(MBB.end());
if (MBBI != MBB.begin() &&
- (MBBI->getOpcode() == ARM::BX_RET || MBBI->getOpcode() == ARM::tBX_RET)) {
+ (MBBI->getOpcode() == ARM::BX_RET ||
+ MBBI->getOpcode() == ARM::tBX_RET ||
+ MBBI->getOpcode() == ARM::MOVPCLR)) {
MachineInstr *PrevMI = prior(MBBI);
- if (PrevMI->getOpcode() == ARM::LDM || PrevMI->getOpcode() == ARM::t2LDM) {
+ if (PrevMI->getOpcode() == ARM::LDM_UPD ||
+ PrevMI->getOpcode() == ARM::t2LDM_UPD) {
MachineOperand &MO = PrevMI->getOperand(PrevMI->getNumOperands()-1);
if (MO.getReg() != ARM::LR)
return false;
return false;
unsigned Align = (*Op0->memoperands_begin())->getAlignment();
- Function *Func = MF->getFunction();
+ const Function *Func = MF->getFunction();
unsigned ReqAlign = STI->hasV6Ops()
? TD->getPrefTypeAlignment(Type::getInt64Ty(Func->getContext()))
: 8; // Pre-v6 need 8-byte align
projects / oota-llvm.git / blobdiff