bool SelectThumbAddrModeSP(SDValue N, SDValue &Base, SDValue &OffImm);
// Thumb 2 Addressing Modes:
- bool SelectT2ShifterOperandReg(SDValue N,
- SDValue &BaseReg, SDValue &Opc);
bool SelectT2AddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm);
bool SelectT2AddrModeImm8(SDValue N, SDValue &Base,
SDValue &OffImm);
// Get the alignment operand for a NEON VLD or VST instruction.
SDValue GetVLDSTAlign(SDValue Align, SDLoc dl, unsigned NumVecs,
bool is64BitVector);
+
+ /// Returns the number of instructions required to materialize the given
+ /// constant in a register, or 3 if a literal pool load is needed.
+ unsigned ConstantMaterializationCost(unsigned Val) const;
+
+ /// Checks if N is a multiplication by a constant where we can extract out a
+ /// power of two from the constant so that it can be used in a shift, but only
+ /// if it simplifies the materialization of the constant. Returns true if it
+ /// is, and assigns to PowerOfTwo the power of two that should be extracted
+ /// out and to NewMulConst the new constant to be multiplied by.
+ bool canExtractShiftFromMul(const SDValue &N, unsigned MaxShift,
+ unsigned &PowerOfTwo, SDValue &NewMulConst) const;
+
+ /// Replace N with M in CurDAG, in a way that also ensures that M gets
+ /// selected when N would have been selected.
+ void replaceDAGValue(const SDValue &N, SDValue M);
};
}
bool isThumb2 = Subtarget->isThumb();
for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
E = CurDAG->allnodes_end(); I != E; ) {
- SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
+ SDNode *N = &*I++; // Preincrement iterator to avoid invalidation issues.
if (N->getOpcode() != ISD::ADD)
continue;
SDValue CPTmp1;
SDValue CPTmp2;
if (isThumb2) {
- if (SelectT2ShifterOperandReg(N0, CPTmp0, CPTmp1))
+ if (SelectImmShifterOperand(N0, CPTmp0, CPTmp1))
continue;
} else {
if (SelectImmShifterOperand(N0, CPTmp0, CPTmp1) ||
(ShAmt == 2 || (Subtarget->isSwift() && ShAmt == 1));
}
+unsigned ARMDAGToDAGISel::ConstantMaterializationCost(unsigned Val) const {
+ if (Subtarget->isThumb()) {
+ if (Val <= 255) return 1; // MOV
+ if (Subtarget->hasV6T2Ops() && Val <= 0xffff) return 1; // MOVW
+ if (~Val <= 255) return 2; // MOV + MVN
+ if (ARM_AM::isThumbImmShiftedVal(Val)) return 2; // MOV + LSL
+ } else {
+ if (ARM_AM::getSOImmVal(Val) != -1) return 1; // MOV
+ if (ARM_AM::getSOImmVal(~Val) != -1) return 1; // MVN
+ if (Subtarget->hasV6T2Ops() && Val <= 0xffff) return 1; // MOVW
+ if (ARM_AM::isSOImmTwoPartVal(Val)) return 2; // two instrs
+ }
+ if (Subtarget->useMovt(*MF)) return 2; // MOVW + MOVT
+ return 3; // Literal pool load
+}
+
+bool ARMDAGToDAGISel::canExtractShiftFromMul(const SDValue &N,
+ unsigned MaxShift,
+ unsigned &PowerOfTwo,
+ SDValue &NewMulConst) const {
+ assert(N.getOpcode() == ISD::MUL);
+ assert(MaxShift > 0);
+
+ // If the multiply is used in more than one place then changing the constant
+ // will make other uses incorrect, so don't.
+ if (!N.hasOneUse()) return false;
+ // Check if the multiply is by a constant
+ ConstantSDNode *MulConst = dyn_cast<ConstantSDNode>(N.getOperand(1));
+ if (!MulConst) return false;
+ // If the constant is used in more than one place then modifying it will mean
+ // we need to materialize two constants instead of one, which is a bad idea.
+ if (!MulConst->hasOneUse()) return false;
+ unsigned MulConstVal = MulConst->getZExtValue();
+ if (MulConstVal == 0) return false;
+
+ // Find the largest power of 2 that MulConstVal is a multiple of
+ PowerOfTwo = MaxShift;
+ while ((MulConstVal % (1 << PowerOfTwo)) != 0) {
+ --PowerOfTwo;
+ if (PowerOfTwo == 0) return false;
+ }
+
+ // Only optimise if the new cost is better
+ unsigned NewMulConstVal = MulConstVal / (1 << PowerOfTwo);
+ NewMulConst = CurDAG->getConstant(NewMulConstVal, SDLoc(N), MVT::i32);
+ unsigned OldCost = ConstantMaterializationCost(MulConstVal);
+ unsigned NewCost = ConstantMaterializationCost(NewMulConstVal);
+ return NewCost < OldCost;
+}
+
+void ARMDAGToDAGISel::replaceDAGValue(const SDValue &N, SDValue M) {
+ CurDAG->RepositionNode(N.getNode()->getIterator(), M.getNode());
+ CurDAG->ReplaceAllUsesWith(N, M);
+}
+
bool ARMDAGToDAGISel::SelectImmShifterOperand(SDValue N,
SDValue &BaseReg,
SDValue &Opc,
if (DisableShifterOp)
return false;
+ // If N is a multiply-by-constant and it's profitable to extract a shift and
+ // use it in a shifted operand do so.
+ if (N.getOpcode() == ISD::MUL) {
+ unsigned PowerOfTwo = 0;
+ SDValue NewMulConst;
+ if (canExtractShiftFromMul(N, 31, PowerOfTwo, NewMulConst)) {
+ replaceDAGValue(N.getOperand(1), NewMulConst);
+ BaseReg = N;
+ Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ARM_AM::lsl,
+ PowerOfTwo),
+ SDLoc(N), MVT::i32);
+ return true;
+ }
+ }
+
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
// Don't match base register only case. That is matched to a separate
}
}
+ // If Offset is a multiply-by-constant and it's profitable to extract a shift
+ // and use it in a shifted operand do so.
+ if (Offset.getOpcode() == ISD::MUL) {
+ unsigned PowerOfTwo = 0;
+ SDValue NewMulConst;
+ if (canExtractShiftFromMul(Offset, 31, PowerOfTwo, NewMulConst)) {
+ replaceDAGValue(Offset.getOperand(1), NewMulConst);
+ ShAmt = PowerOfTwo;
+ ShOpcVal = ARM_AM::lsl;
+ }
+ }
+
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
SDLoc(N), MVT::i32);
return true;
//===----------------------------------------------------------------------===//
-bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDValue N, SDValue &BaseReg,
- SDValue &Opc) {
- if (DisableShifterOp)
- return false;
-
- ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
-
- // 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;
- if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
- ShImmVal = RHS->getZExtValue() & 31;
- Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal), SDLoc(N));
- return true;
- }
-
- return false;
-}
-
bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDValue N,
SDValue &Base, SDValue &OffImm) {
// Match simple R + imm12 operands.
}
}
+ // If OffReg is a multiply-by-constant and it's profitable to extract a shift
+ // and use it in a shifted operand do so.
+ if (OffReg.getOpcode() == ISD::MUL) {
+ unsigned PowerOfTwo = 0;
+ SDValue NewMulConst;
+ if (canExtractShiftFromMul(OffReg, 3, PowerOfTwo, NewMulConst)) {
+ replaceDAGValue(OffReg.getOperand(1), NewMulConst);
+ ShAmt = PowerOfTwo;
+ }
+ }
+
ShImm = CurDAG->getTargetConstant(ShAmt, SDLoc(N), MVT::i32);
return true;
}
case ISD::Constant: {
unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
- bool UseCP = true;
- if (Subtarget->useMovt(*MF))
- // Thumb2-aware targets have the MOVT instruction, so all immediates can
- // be done with MOV + MOVT, at worst.
- UseCP = false;
- else {
- if (Subtarget->isThumb()) {
- UseCP = (Val > 255 && // MOV
- ~Val > 255 && // MOV + MVN
- !ARM_AM::isThumbImmShiftedVal(Val) && // MOV + LSL
- !(Subtarget->hasV6T2Ops() && Val <= 0xffff)); // MOVW
- } else
- UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV
- ARM_AM::getSOImmVal(~Val) == -1 && // MVN
- !ARM_AM::isSOImmTwoPartVal(Val) && // two instrs.
- !(Subtarget->hasV6T2Ops() && Val <= 0xffff)); // MOVW
- }
-
- if (UseCP) {
+ // If we can't materialize the constant we need to use a literal pool
+ if (ConstantMaterializationCost(Val) > 2) {
SDValue CPIdx = CurDAG->getTargetConstantPool(
ConstantInt::get(Type::getInt32Ty(*CurDAG->getContext()), Val),
TLI->getPointerTy(CurDAG->getDataLayout()));
SelectionDAG *CurDAG, SDLoc DL,
std::vector<SDValue>& Ops) {
SmallVector<StringRef, 5> Fields;
- RegString.split(Fields, ":");
+ RegString.split(Fields, ':');
if (Fields.size() > 1) {
bool AllIntFields = true;
// The flags here are common to those allowed for apsr in the A class cores and
// those allowed for the special registers in the M class cores. Returns a
// value representing which flags were present, -1 if invalid.
-static inline int getMClassFlagsMask(StringRef Flags, bool hasThumb2DSP) {
+static inline int getMClassFlagsMask(StringRef Flags, bool hasDSP) {
if (Flags.empty())
- return 0x2 | (int)hasThumb2DSP;
+ return 0x2 | (int)hasDSP;
return StringSwitch<int>(Flags)
.Case("g", 0x1)
}
// We know we are now handling a write so need to get the mask for the flags.
- int Mask = getMClassFlagsMask(Flags, Subtarget->hasThumb2DSP());
+ int Mask = getMClassFlagsMask(Flags, Subtarget->hasDSP());
// Only apsr, iapsr, eapsr, xpsr can have flags. The other register values
// shouldn't have flags present.
// The _g and _nzcvqg versions are only valid if the DSP extension is
// available.
- if (!Subtarget->hasThumb2DSP() && (Mask & 0x1))
+ if (!Subtarget->hasDSP() && (Mask & 0x1))
return -1;
// The register was valid so need to put the mask in the correct place
}
SmallVector<StringRef, 5> Fields;
- StringRef(SpecialReg).split(Fields, "_", 1, false);
+ StringRef(SpecialReg).split(Fields, '_', 1, false);
std::string Reg = Fields[0].str();
StringRef Flags = Fields.size() == 2 ? Fields[1] : "";
// be an immediate and not a memory constraint.
// Fallthrough.
case InlineAsm::Constraint_m:
+ case InlineAsm::Constraint_o:
case InlineAsm::Constraint_Q:
case InlineAsm::Constraint_Um:
case InlineAsm::Constraint_Un:
projects / oota-llvm.git / blobdiff