SDNode *Select(SDNode *N);
+ bool isShifterOpProfitable(const SDValue &Shift,
+ ARM_AM::ShiftOpc ShOpcVal, unsigned ShAmt);
bool SelectShifterOperandReg(SDValue N, SDValue &A,
SDValue &B, SDValue &C);
+ bool SelectShiftShifterOperandReg(SDValue N, SDValue &A,
+ SDValue &B, SDValue &C);
+ bool SelectAddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm);
+ bool SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset, SDValue &Opc);
+
AddrMode2Type SelectAddrMode2Worker(SDValue N, SDValue &Base,
SDValue &Offset, SDValue &Opc);
bool SelectAddrMode2Base(SDValue N, SDValue &Base, SDValue &Offset,
bool SelectAddrMode2(SDValue N, SDValue &Base, SDValue &Offset,
SDValue &Opc) {
SelectAddrMode2Worker(N, Base, Offset, Opc);
+// return SelectAddrMode2ShOp(N, Base, Offset, Opc);
// This always matches one way or another.
return true;
}
SDValue &Offset, SDValue &Opc);
bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
SDValue &Offset, SDValue &Opc);
- bool SelectAddrMode4(SDValue N, SDValue &Addr, SDValue &Mode);
bool SelectAddrMode5(SDValue N, SDValue &Base,
SDValue &Offset);
- bool SelectAddrMode6(SDValue N, SDValue &Addr, SDValue &Align);
+ bool SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,SDValue &Align);
bool SelectAddrModePC(SDValue N, SDValue &Offset,
SDValue &Label);
bool SelectT2AddrModeSoReg(SDValue N, SDValue &Base,
SDValue &OffReg, SDValue &ShImm);
+ inline bool is_so_imm(unsigned Imm) const {
+ return ARM_AM::getSOImmVal(Imm) != -1;
+ }
+
+ inline bool is_so_imm_not(unsigned Imm) const {
+ return ARM_AM::getSOImmVal(~Imm) != -1;
+ }
+
+ inline bool is_t2_so_imm(unsigned Imm) const {
+ return ARM_AM::getT2SOImmVal(Imm) != -1;
+ }
+
+ inline bool is_t2_so_imm_not(unsigned Imm) const {
+ return ARM_AM::getT2SOImmVal(~Imm) != -1;
+ }
+
inline bool Pred_so_imm(SDNode *inN) const {
ConstantSDNode *N = cast<ConstantSDNode>(inN);
- return ARM_AM::getSOImmVal(N->getZExtValue()) != -1;
+ return is_so_imm(N->getZExtValue());
}
inline bool Pred_t2_so_imm(SDNode *inN) const {
ConstantSDNode *N = cast<ConstantSDNode>(inN);
- return ARM_AM::getT2SOImmVal(N->getZExtValue()) != -1;
+ return is_t2_so_imm(N->getZExtValue());
}
// Include the pieces autogenerated from the target description.
SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
ARMCC::CondCodes CCVal, SDValue CCR,
SDValue InFlag);
- SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
+ SDNode *SelectT2CMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
ARMCC::CondCodes CCVal, SDValue CCR,
SDValue InFlag);
- SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
+ SDNode *SelectARMCMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
ARMCC::CondCodes CCVal, SDValue CCR,
SDValue InFlag);
SDNode *QuadSRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
+
+ // Get the alignment operand for a NEON VLD or VST instruction.
+ SDValue GetVLDSTAlign(SDValue Align, unsigned NumVecs, bool is64BitVector);
};
}
}
+bool ARMDAGToDAGISel::isShifterOpProfitable(const SDValue &Shift,
+ ARM_AM::ShiftOpc ShOpcVal,
+ unsigned ShAmt) {
+ if (!Subtarget->isCortexA9())
+ return true;
+ if (Shift.hasOneUse())
+ return true;
+ // R << 2 is free.
+ return ShOpcVal == ARM_AM::lsl && ShAmt == 2;
+}
+
bool ARMDAGToDAGISel::SelectShifterOperandReg(SDValue N,
SDValue &BaseReg,
SDValue &ShReg,
ShImmVal = RHS->getZExtValue() & 31;
} else {
ShReg = N.getOperand(1);
+ if (!isShifterOpProfitable(N, ShOpcVal, ShImmVal))
+ return false;
}
Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
MVT::i32);
return true;
}
+bool ARMDAGToDAGISel::SelectShiftShifterOperandReg(SDValue N,
+ SDValue &BaseReg,
+ SDValue &ShReg,
+ SDValue &Opc) {
+ ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
+
+ // Don't match base register only case. That is matched to a separate
+ // lower complexity pattern with explicit register operand.
+ if (ShOpcVal == ARM_AM::no_shift) return false;
+
+ BaseReg = N.getOperand(0);
+ unsigned ShImmVal = 0;
+ // Do not check isShifterOpProfitable. This must return true.
+ if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
+ ShReg = CurDAG->getRegister(0, MVT::i32);
+ ShImmVal = RHS->getZExtValue() & 31;
+ } else {
+ ShReg = N.getOperand(1);
+ }
+ Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
+ MVT::i32);
+ return true;
+}
+
+bool ARMDAGToDAGISel::SelectAddrModeImm12(SDValue N,
+ SDValue &Base,
+ SDValue &OffImm) {
+ // Match simple R + imm12 operands.
+
+ // Base only.
+ if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
+ if (N.getOpcode() == ISD::FrameIndex) {
+ // Match frame index...
+ int FI = cast<FrameIndexSDNode>(N)->getIndex();
+ Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
+ OffImm = CurDAG->getTargetConstant(0, MVT::i32);
+ return true;
+ } else if (N.getOpcode() == ARMISD::Wrapper &&
+ !(Subtarget->useMovt() &&
+ N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
+ Base = N.getOperand(0);
+ } else
+ Base = N;
+ OffImm = CurDAG->getTargetConstant(0, MVT::i32);
+ return true;
+ }
+
+ if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
+ int RHSC = (int)RHS->getZExtValue();
+ if (N.getOpcode() == ISD::SUB)
+ RHSC = -RHSC;
+
+ if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
+ Base = N.getOperand(0);
+ if (Base.getOpcode() == ISD::FrameIndex) {
+ int FI = cast<FrameIndexSDNode>(Base)->getIndex();
+ Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
+ }
+ OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
+ return true;
+ }
+ }
+
+ // Base only.
+ Base = N;
+ OffImm = CurDAG->getTargetConstant(0, MVT::i32);
+ return true;
+}
+
+
+
+bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset,
+ SDValue &Opc) {
+ if (N.getOpcode() == ISD::MUL &&
+ (!Subtarget->isCortexA9() || N.hasOneUse())) {
+ if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
+ // X * [3,5,9] -> X + X * [2,4,8] etc.
+ int RHSC = (int)RHS->getZExtValue();
+ if (RHSC & 1) {
+ RHSC = RHSC & ~1;
+ ARM_AM::AddrOpc AddSub = ARM_AM::add;
+ if (RHSC < 0) {
+ AddSub = ARM_AM::sub;
+ RHSC = - RHSC;
+ }
+ if (isPowerOf2_32(RHSC)) {
+ unsigned ShAmt = Log2_32(RHSC);
+ Base = Offset = N.getOperand(0);
+ Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
+ ARM_AM::lsl),
+ MVT::i32);
+ return true;
+ }
+ }
+ }
+ }
+
+ if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB)
+ return false;
+
+ // Leave simple R +/- imm12 operands for LDRi12
+ if (N.getOpcode() == ISD::ADD) {
+ if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
+ int RHSC = (int)RHS->getZExtValue();
+ if ((RHSC >= 0 && RHSC < 0x1000) ||
+ (RHSC < 0 && RHSC > -0x1000)) // 12 bits.
+ return false;
+ }
+ }
+
+ if (Subtarget->isCortexA9() && !N.hasOneUse())
+ // Compute R +/- (R << N) and reuse it.
+ return false;
+
+ // Otherwise this is R +/- [possibly shifted] R.
+ ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
+ ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
+ unsigned ShAmt = 0;
+
+ Base = N.getOperand(0);
+ Offset = N.getOperand(1);
+
+ if (ShOpcVal != ARM_AM::no_shift) {
+ // Check to see if the RHS of the shift is a constant, if not, we can't fold
+ // it.
+ if (ConstantSDNode *Sh =
+ dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
+ ShAmt = Sh->getZExtValue();
+ if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt))
+ Offset = N.getOperand(1).getOperand(0);
+ else {
+ ShAmt = 0;
+ ShOpcVal = ARM_AM::no_shift;
+ }
+ } else {
+ ShOpcVal = ARM_AM::no_shift;
+ }
+ }
+
+ // Try matching (R shl C) + (R).
+ if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift &&
+ !(Subtarget->isCortexA9() || N.getOperand(0).hasOneUse())) {
+ ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
+ if (ShOpcVal != ARM_AM::no_shift) {
+ // Check to see if the RHS of the shift is a constant, if not, we can't
+ // fold it.
+ if (ConstantSDNode *Sh =
+ dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
+ ShAmt = Sh->getZExtValue();
+ if (!Subtarget->isCortexA9() ||
+ (N.hasOneUse() &&
+ isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt))) {
+ Offset = N.getOperand(0).getOperand(0);
+ Base = N.getOperand(1);
+ } else {
+ ShAmt = 0;
+ ShOpcVal = ARM_AM::no_shift;
+ }
+ } else {
+ ShOpcVal = ARM_AM::no_shift;
+ }
+ }
+ }
+
+ Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
+ MVT::i32);
+ return true;
+}
+
+
+
+
+//-----
+
AddrMode2Type ARMDAGToDAGISel::SelectAddrMode2Worker(SDValue N,
SDValue &Base,
SDValue &Offset,
SDValue &Opc) {
- if (N.getOpcode() == ISD::MUL) {
+ if (N.getOpcode() == ISD::MUL &&
+ (!Subtarget->isCortexA9() || N.hasOneUse())) {
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
// X * [3,5,9] -> X + X * [2,4,8] etc.
int RHSC = (int)RHS->getZExtValue();
}
}
+ if (Subtarget->isCortexA9() && !N.hasOneUse()) {
+ // Compute R +/- (R << N) and reuse it.
+ Base = N;
+ Offset = CurDAG->getRegister(0, MVT::i32);
+ Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
+ ARM_AM::no_shift),
+ MVT::i32);
+ return AM2_BASE;
+ }
+
// Otherwise this is R +/- [possibly shifted] R.
ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
if (ConstantSDNode *Sh =
dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
ShAmt = Sh->getZExtValue();
- Offset = N.getOperand(1).getOperand(0);
+ if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt))
+ Offset = N.getOperand(1).getOperand(0);
+ else {
+ ShAmt = 0;
+ ShOpcVal = ARM_AM::no_shift;
+ }
} else {
ShOpcVal = ARM_AM::no_shift;
}
}
// Try matching (R shl C) + (R).
- if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
+ if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift &&
+ !(Subtarget->isCortexA9() || N.getOperand(0).hasOneUse())) {
ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
if (ShOpcVal != ARM_AM::no_shift) {
// Check to see if the RHS of the shift is a constant, if not, we can't
if (ConstantSDNode *Sh =
dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
ShAmt = Sh->getZExtValue();
- Offset = N.getOperand(0).getOperand(0);
- Base = N.getOperand(1);
+ if (!Subtarget->isCortexA9() ||
+ (N.hasOneUse() &&
+ isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt))) {
+ Offset = N.getOperand(0).getOperand(0);
+ Base = N.getOperand(1);
+ } else {
+ ShAmt = 0;
+ ShOpcVal = ARM_AM::no_shift;
+ }
} else {
ShOpcVal = ARM_AM::no_shift;
}
// it.
if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
ShAmt = Sh->getZExtValue();
- Offset = N.getOperand(0);
+ if (isShifterOpProfitable(N, ShOpcVal, ShAmt))
+ Offset = N.getOperand(0);
+ else {
+ ShAmt = 0;
+ ShOpcVal = ARM_AM::no_shift;
+ }
} else {
ShOpcVal = ARM_AM::no_shift;
}
return true;
}
-bool ARMDAGToDAGISel::SelectAddrMode4(SDValue N, SDValue &Addr, SDValue &Mode) {
- Addr = N;
- Mode = CurDAG->getTargetConstant(ARM_AM::getAM4ModeImm(ARM_AM::ia), MVT::i32);
- return true;
-}
-
-bool ARMDAGToDAGISel::SelectAddrMode5(SDValue N,
+bool ARMDAGToDAGISel::SelectAddrMode5(SDValue N,
SDValue &Base, SDValue &Offset) {
if (N.getOpcode() != ISD::ADD) {
Base = N;
return true;
}
-bool ARMDAGToDAGISel::SelectAddrMode6(SDValue N, SDValue &Addr, SDValue &Align){
+bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,
+ SDValue &Align) {
Addr = N;
- // Default to no alignment.
- Align = CurDAG->getTargetConstant(0, MVT::i32);
+
+ unsigned Alignment = 0;
+ if (LSBaseSDNode *LSN = dyn_cast<LSBaseSDNode>(Parent)) {
+ // This case occurs only for VLD1-lane/dup and VST1-lane instructions.
+ // The maximum alignment is equal to the memory size being referenced.
+ unsigned LSNAlign = LSN->getAlignment();
+ unsigned MemSize = LSN->getMemoryVT().getSizeInBits() / 8;
+ if (LSNAlign > MemSize && MemSize > 1)
+ Alignment = MemSize;
+ } else {
+ // All other uses of addrmode6 are for intrinsics. For now just record
+ // the raw alignment value; it will be refined later based on the legal
+ // alignment operands for the intrinsic.
+ Alignment = cast<MemIntrinsicSDNode>(Parent)->getAlignment();
+ }
+
+ Align = CurDAG->getTargetConstant(Alignment, MVT::i32);
return true;
}
return false;
}
+ if (Subtarget->isCortexA9() && !N.hasOneUse()) {
+ // Compute R + (R << [1,2,3]) and reuse it.
+ Base = N;
+ return false;
+ }
+
// Look for (R + R) or (R + (R << [1,2,3])).
unsigned ShAmt = 0;
Base = N.getOperand(0);
// it.
if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
ShAmt = Sh->getZExtValue();
- if (ShAmt >= 4) {
+ if (ShAmt < 4 && isShifterOpProfitable(OffReg, ShOpcVal, ShAmt))
+ OffReg = OffReg.getOperand(0);
+ else {
ShAmt = 0;
ShOpcVal = ARM_AM::no_shift;
- } else
- OffReg = OffReg.getOperand(0);
+ }
} else {
ShOpcVal = ARM_AM::no_shift;
}
/// GetVLDSTAlign - Get the alignment (in bytes) for the alignment operand
/// of a NEON VLD or VST instruction. The supported values depend on the
/// number of registers being loaded.
-static unsigned GetVLDSTAlign(SDNode *N, unsigned NumVecs, bool is64BitVector) {
+SDValue ARMDAGToDAGISel::GetVLDSTAlign(SDValue Align, unsigned NumVecs,
+ bool is64BitVector) {
unsigned NumRegs = NumVecs;
if (!is64BitVector && NumVecs < 3)
NumRegs *= 2;
- unsigned Alignment = cast<MemIntrinsicSDNode>(N)->getAlignment();
+ unsigned Alignment = cast<ConstantSDNode>(Align)->getZExtValue();
if (Alignment >= 32 && NumRegs == 4)
- return 32;
- if (Alignment >= 16 && (NumRegs == 2 || NumRegs == 4))
- return 16;
- if (Alignment >= 8)
- return 8;
- return 0;
+ Alignment = 32;
+ else if (Alignment >= 16 && (NumRegs == 2 || NumRegs == 4))
+ Alignment = 16;
+ else if (Alignment >= 8)
+ Alignment = 8;
+ else
+ Alignment = 0;
+
+ return CurDAG->getTargetConstant(Alignment, MVT::i32);
}
SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
DebugLoc dl = N->getDebugLoc();
SDValue MemAddr, Align;
- if (!SelectAddrMode6(N->getOperand(2), MemAddr, Align))
+ if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
return NULL;
SDValue Chain = N->getOperand(0);
EVT VT = N->getValueType(0);
bool is64BitVector = VT.is64BitVector();
-
- unsigned Alignment = GetVLDSTAlign(N, NumVecs, is64BitVector);
- Align = CurDAG->getTargetConstant(Alignment, MVT::i32);
+ Align = GetVLDSTAlign(Align, NumVecs, is64BitVector);
unsigned OpcodeIndex;
switch (VT.getSimpleVT().SimpleTy) {
DebugLoc dl = N->getDebugLoc();
SDValue MemAddr, Align;
- if (!SelectAddrMode6(N->getOperand(2), MemAddr, Align))
+ if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
return NULL;
SDValue Chain = N->getOperand(0);
EVT VT = N->getOperand(3).getValueType();
bool is64BitVector = VT.is64BitVector();
-
- unsigned Alignment = GetVLDSTAlign(N, NumVecs, is64BitVector);
- Align = CurDAG->getTargetConstant(Alignment, MVT::i32);
+ Align = GetVLDSTAlign(Align, NumVecs, is64BitVector);
unsigned OpcodeIndex;
switch (VT.getSimpleVT().SimpleTy) {
RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
else {
SDValue V2 = N->getOperand(2+3);
- // If it's a vld3, form a quad D-register and leave the last part as
+ // If it's a vld3, form a quad D-register and leave the last part as
// an undef.
SDValue V3 = (NumVecs == 3)
? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
DebugLoc dl = N->getDebugLoc();
SDValue MemAddr, Align;
- if (!SelectAddrMode6(N->getOperand(2), MemAddr, Align))
+ if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
return NULL;
SDValue Chain = N->getOperand(0);
EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
bool is64BitVector = VT.is64BitVector();
+ unsigned Alignment = 0;
+ if (NumVecs != 3) {
+ Alignment = cast<ConstantSDNode>(Align)->getZExtValue();
+ unsigned NumBytes = NumVecs * VT.getVectorElementType().getSizeInBits()/8;
+ if (Alignment > NumBytes)
+ Alignment = NumBytes;
+ // Alignment must be a power of two; make sure of that.
+ Alignment = (Alignment & -Alignment);
+ if (Alignment == 1)
+ Alignment = 0;
+ }
+ Align = CurDAG->getTargetConstant(Alignment, MVT::i32);
+
unsigned OpcodeIndex;
switch (VT.getSimpleVT().SimpleTy) {
default: llvm_unreachable("unhandled vld/vst lane type");
Ops.push_back(MemAddr);
Ops.push_back(Align);
- unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
+ unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
QOpcodes[OpcodeIndex]);
SDValue SuperReg;
RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0);
else {
SDValue V2 = N->getOperand(FirstTblReg + 2);
- // If it's a vtbl3, form a quad D-register and leave the last part as
+ // If it's a vtbl3, form a quad D-register and leave the last part as
// an undef.
SDValue V3 = (NumVecs == 3)
? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
}
SDNode *ARMDAGToDAGISel::
-SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
- ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
+SelectT2CMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
+ ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
if (!T)
return 0;
- if (Pred_t2_so_imm(TrueVal.getNode())) {
- SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
+ unsigned Opc = 0;
+ unsigned TrueImm = T->getZExtValue();
+ if (is_t2_so_imm(TrueImm)) {
+ Opc = ARM::t2MOVCCi;
+ } else if (TrueImm <= 0xffff) {
+ Opc = ARM::t2MOVCCi16;
+ } else if (is_t2_so_imm_not(TrueImm)) {
+ TrueImm = ~TrueImm;
+ Opc = ARM::t2MVNCCi;
+ } else if (TrueVal.getNode()->hasOneUse() && Subtarget->hasV6T2Ops()) {
+ // Large immediate.
+ Opc = ARM::t2MOVCCi32imm;
+ }
+
+ if (Opc) {
+ SDValue True = CurDAG->getTargetConstant(TrueImm, MVT::i32);
SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
- return CurDAG->SelectNodeTo(N,
- ARM::t2MOVCCi, MVT::i32, Ops, 5);
+ return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
}
+
return 0;
}
SDNode *ARMDAGToDAGISel::
-SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
- ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
+SelectARMCMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
+ ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
if (!T)
return 0;
- if (Pred_so_imm(TrueVal.getNode())) {
- SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
+ unsigned Opc = 0;
+ unsigned TrueImm = T->getZExtValue();
+ bool isSoImm = is_so_imm(TrueImm);
+ if (isSoImm) {
+ Opc = ARM::MOVCCi;
+ } else if (Subtarget->hasV6T2Ops() && TrueImm <= 0xffff) {
+ Opc = ARM::MOVCCi16;
+ } else if (is_so_imm_not(TrueImm)) {
+ TrueImm = ~TrueImm;
+ Opc = ARM::MVNCCi;
+ } else if (TrueVal.getNode()->hasOneUse() &&
+ (Subtarget->hasV6T2Ops() || ARM_AM::isSOImmTwoPartVal(TrueImm))) {
+ // Large immediate.
+ Opc = ARM::MOVCCi32imm;
+ }
+
+ if (Opc) {
+ SDValue True = CurDAG->getTargetConstant(TrueImm, MVT::i32);
SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
- return CurDAG->SelectNodeTo(N,
- ARM::MOVCCi, MVT::i32, Ops, 5);
+ return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
}
+
return 0;
}
// (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
// Pattern complexity = 10 cost = 1 size = 0
if (Subtarget->isThumb()) {
- SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
+ SDNode *Res = SelectT2CMOVImmOp(N, FalseVal, TrueVal,
CCVal, CCR, InFlag);
if (!Res)
- Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
+ Res = SelectT2CMOVImmOp(N, TrueVal, FalseVal,
ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
if (Res)
return Res;
} else {
- SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
+ SDNode *Res = SelectARMCMOVImmOp(N, FalseVal, TrueVal,
CCVal, CCR, InFlag);
if (!Res)
- Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
+ Res = SelectARMCMOVImmOp(N, TrueVal, FalseVal,
ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
if (Res)
return Res;
} else {
SDValue Ops[] = {
CPIdx,
- CurDAG->getRegister(0, MVT::i32),
CurDAG->getTargetConstant(0, MVT::i32),
getAL(CurDAG),
CurDAG->getRegister(0, MVT::i32),
CurDAG->getEntryNode()
};
ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
- Ops, 6);
+ Ops, 5);
}
ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
return NULL;
EVT VecVT = N->getValueType(0);
EVT EltVT = VecVT.getVectorElementType();
unsigned NumElts = VecVT.getVectorNumElements();
- if (EltVT.getSimpleVT() == MVT::f64) {
+ if (EltVT == MVT::f64) {
assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
return PairDRegs(VecVT, N->getOperand(0), N->getOperand(1));
}
- assert(EltVT.getSimpleVT() == MVT::f32 &&
- "unexpected type for BUILD_VECTOR");
+ assert(EltVT == MVT::f32 && "unexpected type for BUILD_VECTOR");
if (NumElts == 2)
return PairSRegs(VecVT, N->getOperand(0), N->getOperand(1));
assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
projects / oota-llvm.git / blobdiff