int ARMTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
const Instruction *I) {
int ISD = TLI->InstructionOpcodeToISD(Opcode);
// On NEON a vector select gets lowered to vbsl.
if (ST->hasNEON() && ValTy->isVectorTy() && ISD == ISD::SELECT) {
// Lowering of some vector selects is currently far from perfect.
static const TypeConversionCostTblEntry NEONVectorSelectTbl[] = {
{ ISD::SELECT, MVT::v4i1, MVT::v4i64, 4*4 + 1*2 + 1 },
{ ISD::SELECT, MVT::v8i1, MVT::v8i64, 50 },
{ ISD::SELECT, MVT::v16i1, MVT::v16i64, 100 }
};
EVT SelCondTy = TLI->getValueType(DL, CondTy);
EVT SelValTy = TLI->getValueType(DL, ValTy);
if (SelCondTy.isSimple() && SelValTy.isSimple()) {
if (const auto *Entry = ConvertCostTableLookup(NEONVectorSelectTbl, ISD,
SelCondTy.getSimpleVT(),
SelValTy.getSimpleVT()))
return Entry->Cost;
}
std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
return LT.first;
}
return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, I);
}
bool WebAssemblyFastISel::selectBitCast(const Instruction *I) {
// Target-independent code can handle this, except it doesn't set the dead
// flag on the ARGUMENTS clobber, so we have to do that manually in order
// to satisfy code that expects this of isBitcast() instructions.
EVT VT = TLI.getValueType(DL, I->getOperand(0)->getType());
EVT RetVT = TLI.getValueType(DL, I->getType());
if (!VT.isSimple() || !RetVT.isSimple())
return false;
if (VT == RetVT) {
// No-op bitcast.
updateValueMap(I, getRegForValue(I->getOperand(0)));
return true;
}
unsigned Reg = fastEmit_ISD_BITCAST_r(VT.getSimpleVT(), RetVT.getSimpleVT(),
getRegForValue(I->getOperand(0)),
I->getOperand(0)->hasOneUse());
if (!Reg)
return false;
MachineBasicBlock::iterator Iter = FuncInfo.InsertPt;
--Iter;
assert(Iter->isBitcast());
Iter->setPhysRegsDeadExcept(ArrayRef<unsigned>(), TRI);
updateValueMap(I, Reg);
return true;
}
unsigned BasicTTI::getMemoryOpCost(unsigned Opcode, Type *Src,
unsigned Alignment,
unsigned AddressSpace) const {
assert(!Src->isVoidTy() && "Invalid type");
std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(Src);
// Assuming that all loads of legal types cost 1.
unsigned Cost = LT.first;
if (Src->isVectorTy() &&
Src->getPrimitiveSizeInBits() < LT.second.getSizeInBits()) {
// This is a vector load that legalizes to a larger type than the vector
// itself. Unless the corresponding extending load or truncating store is
// legal, then this will scalarize.
TargetLowering::LegalizeAction LA = TargetLowering::Expand;
EVT MemVT = getTLI()->getValueType(Src, true);
if (MemVT.isSimple() && MemVT != MVT::Other) {
if (Opcode == Instruction::Store)
LA = getTLI()->getTruncStoreAction(LT.second, MemVT.getSimpleVT());
else
LA = getTLI()->getLoadExtAction(ISD::EXTLOAD, MemVT.getSimpleVT());
}
if (LA != TargetLowering::Legal && LA != TargetLowering::Custom) {
// This is a vector load/store for some illegal type that is scalarized.
// We must account for the cost of building or decomposing the vector.
Cost += getScalarizationOverhead(Src, Opcode != Instruction::Store,
Opcode == Instruction::Store);
}
}
return Cost;
}
bool FastISel::SelectCast(const User *I, unsigned Opcode) {
EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
EVT DstVT = TLI.getValueType(I->getType());
if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
DstVT == MVT::Other || !DstVT.isSimple())
// Unhandled type. Halt "fast" selection and bail.
return false;
// Check if the destination type is legal.
if (!TLI.isTypeLegal(DstVT))
return false;
// Check if the source operand is legal.
if (!TLI.isTypeLegal(SrcVT))
return false;
unsigned InputReg = getRegForValue(I->getOperand(0));
if (!InputReg)
// Unhandled operand. Halt "fast" selection and bail.
return false;
bool InputRegIsKill = hasTrivialKill(I->getOperand(0));
unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
DstVT.getSimpleVT(),
Opcode,
InputReg, InputRegIsKill);
if (!ResultReg)
return false;
UpdateValueMap(I, ResultReg);
return true;
}
unsigned ARMTTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
Type *CondTy) const {
int ISD = TLI->InstructionOpcodeToISD(Opcode);
// On NEON a a vector select gets lowered to vbsl.
if (ST->hasNEON() && ValTy->isVectorTy() && ISD == ISD::SELECT) {
// Lowering of some vector selects is currently far from perfect.
static const TypeConversionCostTblEntry<MVT::SimpleValueType>
NEONVectorSelectTbl[] = {
{ ISD::SELECT, MVT::v16i1, MVT::v16i16, 2*16 + 1 + 3*1 + 4*1 },
{ ISD::SELECT, MVT::v8i1, MVT::v8i32, 4*8 + 1*3 + 1*4 + 1*2 },
{ ISD::SELECT, MVT::v16i1, MVT::v16i32, 4*16 + 1*6 + 1*8 + 1*4 },
{ ISD::SELECT, MVT::v4i1, MVT::v4i64, 4*4 + 1*2 + 1 },
{ ISD::SELECT, MVT::v8i1, MVT::v8i64, 50 },
{ ISD::SELECT, MVT::v16i1, MVT::v16i64, 100 }
};
EVT SelCondTy = TLI->getValueType(CondTy);
EVT SelValTy = TLI->getValueType(ValTy);
if (SelCondTy.isSimple() && SelValTy.isSimple()) {
int Idx = ConvertCostTableLookup(NEONVectorSelectTbl, ISD,
SelCondTy.getSimpleVT(),
SelValTy.getSimpleVT());
if (Idx != -1)
return NEONVectorSelectTbl[Idx].Cost;
}
std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
return LT.first;
}
return TargetTransformInfo::getCmpSelInstrCost(Opcode, ValTy, CondTy);
}
int AArch64TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
Type *CondTy) {
int ISD = TLI->InstructionOpcodeToISD(Opcode);
// We don't lower some vector selects well that are wider than the register
// width.
if (ValTy->isVectorTy() && ISD == ISD::SELECT) {
// We would need this many instructions to hide the scalarization happening.
const int AmortizationCost = 20;
static const TypeConversionCostTblEntry<MVT::SimpleValueType>
VectorSelectTbl[] = {
{ ISD::SELECT, MVT::v16i1, MVT::v16i16, 16 },
{ ISD::SELECT, MVT::v8i1, MVT::v8i32, 8 },
{ ISD::SELECT, MVT::v16i1, MVT::v16i32, 16 },
{ ISD::SELECT, MVT::v4i1, MVT::v4i64, 4 * AmortizationCost },
{ ISD::SELECT, MVT::v8i1, MVT::v8i64, 8 * AmortizationCost },
{ ISD::SELECT, MVT::v16i1, MVT::v16i64, 16 * AmortizationCost }
};
EVT SelCondTy = TLI->getValueType(DL, CondTy);
EVT SelValTy = TLI->getValueType(DL, ValTy);
if (SelCondTy.isSimple() && SelValTy.isSimple()) {
int Idx =
ConvertCostTableLookup(VectorSelectTbl, ISD, SelCondTy.getSimpleVT(),
SelValTy.getSimpleVT());
if (Idx != -1)
return VectorSelectTbl[Idx].Cost;
}
}
return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy);
}
bool FastISel::SelectCast(const User *I, unsigned Opcode) {
EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
EVT DstVT = TLI.getValueType(I->getType());
if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
DstVT == MVT::Other || !DstVT.isSimple())
// Unhandled type. Halt "fast" selection and bail.
return false;
// Check if the destination type is legal. Or as a special case,
// it may be i1 if we're doing a truncate because that's
// easy and somewhat common.
if (!TLI.isTypeLegal(DstVT))
if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE)
// Unhandled type. Halt "fast" selection and bail.
return false;
// Check if the source operand is legal. Or as a special case,
// it may be i1 if we're doing zero-extension because that's
// easy and somewhat common.
if (!TLI.isTypeLegal(SrcVT))
if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND)
// Unhandled type. Halt "fast" selection and bail.
return false;
unsigned InputReg = getRegForValue(I->getOperand(0));
if (!InputReg)
// Unhandled operand. Halt "fast" selection and bail.
return false;
bool InputRegIsKill = hasTrivialKill(I->getOperand(0));
// If the operand is i1, arrange for the high bits in the register to be zero.
if (SrcVT == MVT::i1) {
SrcVT = TLI.getTypeToTransformTo(I->getContext(), SrcVT);
InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg, InputRegIsKill);
if (!InputReg)
return false;
InputRegIsKill = true;
}
// If the result is i1, truncate to the target's type for i1 first.
if (DstVT == MVT::i1)
DstVT = TLI.getTypeToTransformTo(I->getContext(), DstVT);
unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
DstVT.getSimpleVT(),
Opcode,
InputReg, InputRegIsKill);
if (!ResultReg)
return false;
UpdateValueMap(I, ResultReg);
return true;
}
bool FastISel::SelectBitCast(const User *I) {
// If the bitcast doesn't change the type, just use the operand value.
if (I->getType() == I->getOperand(0)->getType()) {
unsigned Reg = getRegForValue(I->getOperand(0));
if (Reg == 0)
return false;
UpdateValueMap(I, Reg);
return true;
}
// Bitcasts of other values become reg-reg copies or BITCAST operators.
EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
EVT DstVT = TLI.getValueType(I->getType());
if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
DstVT == MVT::Other || !DstVT.isSimple() ||
!TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
// Unhandled type. Halt "fast" selection and bail.
return false;
unsigned Op0 = getRegForValue(I->getOperand(0));
if (Op0 == 0)
// Unhandled operand. Halt "fast" selection and bail.
return false;
bool Op0IsKill = hasTrivialKill(I->getOperand(0));
// First, try to perform the bitcast by inserting a reg-reg copy.
unsigned ResultReg = 0;
if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
// Don't attempt a cross-class copy. It will likely fail.
if (SrcClass == DstClass) {
ResultReg = createResultReg(DstClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
ResultReg).addReg(Op0);
}
}
// If the reg-reg copy failed, select a BITCAST opcode.
if (!ResultReg)
ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
ISD::BITCAST, Op0, Op0IsKill);
if (!ResultReg)
return false;
UpdateValueMap(I, ResultReg);
return true;
}
bool FastISel::SelectBitCast(User *I) {
// If the bitcast doesn't change the type, just use the operand value.
if (I->getType() == I->getOperand(0)->getType()) {
unsigned Reg = getRegForValue(I->getOperand(0));
if (Reg == 0)
return false;
UpdateValueMap(I, Reg);
return true;
}
// Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
EVT DstVT = TLI.getValueType(I->getType());
if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
DstVT == MVT::Other || !DstVT.isSimple() ||
!TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
// Unhandled type. Halt "fast" selection and bail.
return false;
unsigned Op0 = getRegForValue(I->getOperand(0));
if (Op0 == 0)
// Unhandled operand. Halt "fast" selection and bail.
return false;
// First, try to perform the bitcast by inserting a reg-reg copy.
unsigned ResultReg = 0;
if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
ResultReg = createResultReg(DstClass);
bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
Op0, DstClass, SrcClass);
if (!InsertedCopy)
ResultReg = 0;
}
// If the reg-reg copy failed, select a BIT_CONVERT opcode.
if (!ResultReg)
ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
ISD::BIT_CONVERT, Op0);
if (!ResultReg)
return false;
UpdateValueMap(I, ResultReg);
return true;
}
// HandleByVal - Allocate a stack slot large enough to pass an argument by
// value. The size and alignment information of the argument is encoded in its
// parameter attribute.
void Hexagon_CCState::HandleByVal(unsigned ValNo, EVT ValVT,
EVT LocVT, CCValAssign::LocInfo LocInfo,
int MinSize, int MinAlign,
ISD::ArgFlagsTy ArgFlags) {
unsigned Align = ArgFlags.getByValAlign();
unsigned Size = ArgFlags.getByValSize();
if (MinSize > (int)Size)
Size = MinSize;
if (MinAlign > (int)Align)
Align = MinAlign;
unsigned Offset = AllocateStack(Size, Align);
addLoc(CCValAssign::getMem(ValNo, ValVT.getSimpleVT(), Offset,
LocVT.getSimpleVT(), LocInfo));
}
unsigned FastISel::getRegForGEPIndex(Value *Idx) {
unsigned IdxN = getRegForValue(Idx);
if (IdxN == 0)
// Unhandled operand. Halt "fast" selection and bail.
return 0;
// If the index is smaller or larger than intptr_t, truncate or extend it.
MVT PtrVT = TLI.getPointerTy();
EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
if (IdxVT.bitsLT(PtrVT))
IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND, IdxN);
else if (IdxVT.bitsGT(PtrVT))
IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE, IdxN);
return IdxN;
}
SDNode *PTXDAGToDAGISel::SelectREADPARAM(SDNode *Node) {
SDValue Chain = Node->getOperand(0);
SDValue Index = Node->getOperand(1);
int OpCode;
// Get the type of parameter we are reading
EVT VT = Node->getValueType(0);
assert(VT.isSimple() && "READ_PARAM only implemented for MVT types");
MVT Type = VT.getSimpleVT();
if (Type == MVT::i1)
OpCode = PTX::READPARAMPRED;
else if (Type == MVT::i16)
OpCode = PTX::READPARAMI16;
else if (Type == MVT::i32)
OpCode = PTX::READPARAMI32;
else if (Type == MVT::i64)
OpCode = PTX::READPARAMI64;
else if (Type == MVT::f32)
OpCode = PTX::READPARAMF32;
else {
assert(Type == MVT::f64 && "Unexpected type!");
OpCode = PTX::READPARAMF64;
}
SDValue Pred = CurDAG->getRegister(PTX::NoRegister, MVT::i1);
SDValue PredOp = CurDAG->getTargetConstant(PTXPredicate::None, MVT::i32);
DebugLoc dl = Node->getDebugLoc();
SDValue Ops[] = { Index, Pred, PredOp, Chain };
return CurDAG->getMachineNode(OpCode, dl, VT, Ops, 4);
}
bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg,
unsigned DstReg, int Offset) {
unsigned StrOpc;
switch (VT.getSimpleVT().SimpleTy) {
default: return false;
case MVT::i1:
case MVT::i8: StrOpc = isThumb ? ARM::tSTRB : ARM::STRB; break;
case MVT::i16: StrOpc = isThumb ? ARM::tSTRH : ARM::STRH; break;
case MVT::i32: StrOpc = isThumb ? ARM::tSTR : ARM::STR; break;
case MVT::f32:
if (!Subtarget->hasVFP2()) return false;
StrOpc = ARM::VSTRS;
break;
case MVT::f64:
if (!Subtarget->hasVFP2()) return false;
StrOpc = ARM::VSTRD;
break;
}
if (isThumb)
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(StrOpc), SrcReg)
.addReg(DstReg).addImm(Offset).addReg(0));
else
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(StrOpc), SrcReg)
.addReg(DstReg).addReg(0).addImm(Offset));
return true;
}
unsigned AArch64TTI::getCastInstrCost(unsigned Opcode, Type *Dst,
Type *Src) const {
int ISD = TLI->InstructionOpcodeToISD(Opcode);
assert(ISD && "Invalid opcode");
EVT SrcTy = TLI->getValueType(Src);
EVT DstTy = TLI->getValueType(Dst);
if (!SrcTy.isSimple() || !DstTy.isSimple())
return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
static const TypeConversionCostTblEntry<MVT> ConversionTbl[] = {
// LowerVectorINT_TO_FP:
{ ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i32, 1 },
{ ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i8, 1 },
{ ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i16, 1 },
{ ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 1 },
{ ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 },
{ ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 1 },
{ ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i8, 1 },
{ ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i16, 1 },
{ ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 1 },
{ ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 },
// LowerVectorFP_TO_INT
{ ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 1 },
{ ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f64, 1 },
{ ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1 },
{ ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f64, 1 },
{ ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f64, 1 },
{ ISD::FP_TO_SINT, MVT::v2i32, MVT::v2f64, 1 },
{ ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f32, 4 },
{ ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f32, 4 },
{ ISD::FP_TO_UINT, MVT::v4i16, MVT::v4f32, 4 },
{ ISD::FP_TO_SINT, MVT::v4i16, MVT::v4f32, 4 },
{ ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f64, 4 },
{ ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f64, 4 },
};
int Idx = ConvertCostTableLookup<MVT>(
ConversionTbl, array_lengthof(ConversionTbl), ISD, DstTy.getSimpleVT(),
SrcTy.getSimpleVT());
if (Idx != -1)
return ConversionTbl[Idx].Cost;
return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
}
/// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
/// to emit an instruction with a floating-point immediate operand using
/// FastEmit_rf. If that fails, it materializes the immediate into a register
/// and try FastEmit_rr instead.
unsigned FastISel::FastEmit_rf_(MVT VT, unsigned Opcode,
unsigned Op0, bool Op0IsKill,
const ConstantFP *FPImm, MVT ImmType) {
// First check if immediate type is legal. If not, we can't use the rf form.
unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, Op0IsKill, FPImm);
if (ResultReg != 0)
return ResultReg;
// Materialize the constant in a register.
unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
if (MaterialReg == 0) {
// If the target doesn't have a way to directly enter a floating-point
// value into a register, use an alternate approach.
// TODO: The current approach only supports floating-point constants
// that can be constructed by conversion from integer values. This should
// be replaced by code that creates a load from a constant-pool entry,
// which will require some target-specific work.
const APFloat &Flt = FPImm->getValueAPF();
EVT IntVT = TLI.getPointerTy();
uint64_t x[2];
uint32_t IntBitWidth = IntVT.getSizeInBits();
bool isExact;
(void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
APFloat::rmTowardZero, &isExact);
if (!isExact)
return 0;
APInt IntVal(IntBitWidth, 2, x);
unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
ISD::Constant, IntVal.getZExtValue());
if (IntegerReg == 0)
return 0;
MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
ISD::SINT_TO_FP, IntegerReg, /*Kill=*/true);
if (MaterialReg == 0)
return 0;
}
return FastEmit_rr(VT, VT, Opcode,
Op0, Op0IsKill,
MaterialReg, /*Kill=*/true);
}
bool MipsFastISel::isTypeLegal(Type *Ty, MVT &VT) {
EVT evt = TLI.getValueType(Ty, true);
// Only handle simple types.
if (evt == MVT::Other || !evt.isSimple())
return false;
VT = evt.getSimpleVT();
// Handle all legal types, i.e. a register that will directly hold this
// value.
return TLI.isTypeLegal(VT);
}
unsigned X86TTI::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const {
int ISD = TLI->InstructionOpcodeToISD(Opcode);
assert(ISD && "Invalid opcode");
EVT SrcTy = TLI->getValueType(Src);
EVT DstTy = TLI->getValueType(Dst);
if (!SrcTy.isSimple() || !DstTy.isSimple())
return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
static const TypeConversionCostTblEntry<MVT> AVXConversionTbl[] = {
{ ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
{ ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
{ ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
{ ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
{ ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 },
{ ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 1 },
{ ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 1 },
{ ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 1 },
{ ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 1 },
{ ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 1 },
{ ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f32, 1 },
{ ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 1 },
{ ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 6 },
{ ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 9 },
{ ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 8 },
{ ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 6 },
{ ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 6 },
{ ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 3 },
};
if (ST->hasAVX()) {
int Idx = ConvertCostTableLookup<MVT>(AVXConversionTbl,
array_lengthof(AVXConversionTbl),
ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT());
if (Idx != -1)
return AVXConversionTbl[Idx].Cost;
}
return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
}
bool Cpu0TargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
switch (SVT) {
case MVT::i64:
case MVT::i32:
case MVT::i16:
return true;
default:
return false;
}
}
bool
MipsSETargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
switch (SVT) {
case MVT::i64:
case MVT::i32:
if (Fast)
*Fast = true;
return true;
default:
return false;
}
}
std::pair<unsigned, bool> FastISel::getRegForGEPIndex(const Value *Idx) {
unsigned IdxN = getRegForValue(Idx);
if (IdxN == 0)
// Unhandled operand. Halt "fast" selection and bail.
return std::pair<unsigned, bool>(0, false);
bool IdxNIsKill = hasTrivialKill(Idx);
// If the index is smaller or larger than intptr_t, truncate or extend it.
MVT PtrVT = TLI.getPointerTy();
EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
if (IdxVT.bitsLT(PtrVT)) {
IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND,
IdxN, IdxNIsKill);
IdxNIsKill = true;
}
else if (IdxVT.bitsGT(PtrVT)) {
IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE,
IdxN, IdxNIsKill);
IdxNIsKill = true;
}
return std::pair<unsigned, bool>(IdxN, IdxNIsKill);
}
unsigned FastISel::getRegForValue(const Value *V) {
EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true);
// Don't handle non-simple values in FastISel.
if (!RealVT.isSimple())
return 0;
// Ignore illegal types. We must do this before looking up the value
// in ValueMap because Arguments are given virtual registers regardless
// of whether FastISel can handle them.
MVT VT = RealVT.getSimpleVT();
if (!TLI.isTypeLegal(VT)) {
// Promote MVT::i1 to a legal type though, because it's common and easy.
if (VT == MVT::i1)
VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT();
else
return 0;
}
// Look up the value to see if we already have a register for it. We
// cache values defined by Instructions across blocks, and other values
// only locally. This is because Instructions already have the SSA
// def-dominates-use requirement enforced.
DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V);
if (I != FuncInfo.ValueMap.end()) {
unsigned Reg = I->second;
return Reg;
}
unsigned Reg = LocalValueMap[V];
if (Reg != 0)
return Reg;
// In bottom-up mode, just create the virtual register which will be used
// to hold the value. It will be materialized later.
if (isa<Instruction>(V) &&
(!isa<AllocaInst>(V) ||
!FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(V))))
return FuncInfo.InitializeRegForValue(V);
SavePoint SaveInsertPt = enterLocalValueArea();
// Materialize the value in a register. Emit any instructions in the
// local value area.
Reg = materializeRegForValue(V, VT);
leaveLocalValueArea(SaveInsertPt);
return Reg;
}
SDValue PTXTargetLowering::
LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
EVT PtrVT = getPointerTy();
DebugLoc dl = Op.getDebugLoc();
const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
assert(PtrVT.isSimple() && "Pointer must be to primitive type.");
SDValue targetGlobal = DAG.getTargetGlobalAddress(GV, dl, PtrVT);
SDValue movInstr = DAG.getNode(PTXISD::COPY_ADDRESS,
dl,
PtrVT.getSimpleVT(),
targetGlobal);
return movInstr;
}
// Materialize a constant into a register, and return the register
// number (or zero if we failed to handle it).
unsigned MipsFastISel::fastMaterializeConstant(const Constant *C) {
EVT CEVT = TLI.getValueType(C->getType(), true);
// Only handle simple types.
if (!CEVT.isSimple())
return 0;
MVT VT = CEVT.getSimpleVT();
if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
return (UnsupportedFPMode) ? 0 : materializeFP(CFP, VT);
else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
return materializeGV(GV, VT);
else if (isa<ConstantInt>(C))
return materializeInt(C, VT);
return 0;
}
/// SelectFNeg - Emit an FNeg operation.
///
bool
FastISel::SelectFNeg(const User *I) {
unsigned OpReg = getRegForValue(BinaryOperator::getFNegArgument(I));
if (OpReg == 0) return false;
bool OpRegIsKill = hasTrivialKill(I);
// If the target has ISD::FNEG, use it.
EVT VT = TLI.getValueType(I->getType());
unsigned ResultReg = FastEmit_r(VT.getSimpleVT(), VT.getSimpleVT(),
ISD::FNEG, OpReg, OpRegIsKill);
if (ResultReg != 0) {
UpdateValueMap(I, ResultReg);
return true;
}
// Bitcast the value to integer, twiddle the sign bit with xor,
// and then bitcast it back to floating-point.
if (VT.getSizeInBits() > 64) return false;
EVT IntVT = EVT::getIntegerVT(I->getContext(), VT.getSizeInBits());
if (!TLI.isTypeLegal(IntVT))
return false;
unsigned IntReg = FastEmit_r(VT.getSimpleVT(), IntVT.getSimpleVT(),
ISD::BITCAST, OpReg, OpRegIsKill);
if (IntReg == 0)
return false;
unsigned IntResultReg = FastEmit_ri_(IntVT.getSimpleVT(), ISD::XOR,
IntReg, /*Kill=*/true,
UINT64_C(1) << (VT.getSizeInBits()-1),
IntVT.getSimpleVT());
if (IntResultReg == 0)
return false;
ResultReg = FastEmit_r(IntVT.getSimpleVT(), VT.getSimpleVT(),
ISD::BITCAST, IntResultReg, /*Kill=*/true);
if (ResultReg == 0)
return false;
UpdateValueMap(I, ResultReg);
return true;
}
SDNode *PTXDAGToDAGISel::SelectWRITEPARAM(SDNode *Node) {
SDValue Chain = Node->getOperand(0);
SDValue Value = Node->getOperand(1);
int OpCode;
//Node->dumpr(CurDAG);
// Get the type of parameter we are writing
EVT VT = Value->getValueType(0);
assert(VT.isSimple() && "WRITE_PARAM only implemented for MVT types");
MVT Type = VT.getSimpleVT();
if (Type == MVT::i1)
OpCode = PTX::WRITEPARAMPRED;
else if (Type == MVT::i16)
OpCode = PTX::WRITEPARAMI16;
else if (Type == MVT::i32)
OpCode = PTX::WRITEPARAMI32;
else if (Type == MVT::i64)
OpCode = PTX::WRITEPARAMI64;
else if (Type == MVT::f32)
OpCode = PTX::WRITEPARAMF32;
else if (Type == MVT::f64)
OpCode = PTX::WRITEPARAMF64;
else
llvm_unreachable("Invalid type in SelectWRITEPARAM");
SDValue Pred = CurDAG->getRegister(PTX::NoRegister, MVT::i1);
SDValue PredOp = CurDAG->getTargetConstant(PTXPredicate::None, MVT::i32);
DebugLoc dl = Node->getDebugLoc();
SDValue Ops[] = { Value, Pred, PredOp, Chain };
SDNode* Ret = CurDAG->getMachineNode(OpCode, dl, MVT::Other, Ops, 4);
//dbgs() << "SelectWRITEPARAM produced:\n\t";
//Ret->dumpr(CurDAG);
return Ret;
}
unsigned FastISel::getRegForValue(const Value *V) {
EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true);
// Don't handle non-simple values in FastISel.
if (!RealVT.isSimple())
return 0;
// Ignore illegal types. We must do this before looking up the value
// in ValueMap because Arguments are given virtual registers regardless
// of whether FastISel can handle them.
MVT VT = RealVT.getSimpleVT();
if (!TLI.isTypeLegal(VT)) {
// Handle integer promotions, though, because they're common and easy.
if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT();
else
return 0;
}
// Look up the value to see if we already have a register for it.
unsigned Reg = lookUpRegForValue(V);
if (Reg != 0)
return Reg;
// In bottom-up mode, just create the virtual register which will be used
// to hold the value. It will be materialized later.
if (isa<Instruction>(V) &&
(!isa<AllocaInst>(V) ||
!FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(V))))
return FuncInfo.InitializeRegForValue(V);
SavePoint SaveInsertPt = enterLocalValueArea();
// Materialize the value in a register. Emit any instructions in the
// local value area.
Reg = materializeRegForValue(V, VT);
leaveLocalValueArea(SaveInsertPt);
return Reg;
}
bool
FastISel::SelectExtractValue(const User *U) {
const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(U);
if (!EVI)
return false;
// Make sure we only try to handle extracts with a legal result. But also
// allow i1 because it's easy.
EVT RealVT = TLI.getValueType(EVI->getType(), /*AllowUnknown=*/true);
if (!RealVT.isSimple())
return false;
MVT VT = RealVT.getSimpleVT();
if (!TLI.isTypeLegal(VT) && VT != MVT::i1)
return false;
const Value *Op0 = EVI->getOperand(0);
const Type *AggTy = Op0->getType();
// Get the base result register.
unsigned ResultReg;
DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(Op0);
if (I != FuncInfo.ValueMap.end())
ResultReg = I->second;
else if (isa<Instruction>(Op0))
ResultReg = FuncInfo.InitializeRegForValue(Op0);
else
return false; // fast-isel can't handle aggregate constants at the moment
// Get the actual result register, which is an offset from the base register.
unsigned VTIndex = ComputeLinearIndex(AggTy, EVI->idx_begin(), EVI->idx_end());
SmallVector<EVT, 4> AggValueVTs;
ComputeValueVTs(TLI, AggTy, AggValueVTs);
for (unsigned i = 0; i < VTIndex; i++)
ResultReg += TLI.getNumRegisters(FuncInfo.Fn->getContext(), AggValueVTs[i]);
UpdateValueMap(EVI, ResultReg);
return true;
}
std::pair<unsigned, MVT>
TargetLoweringBase::getTypeLegalizationCost(Type *Ty) const {
LLVMContext &C = Ty->getContext();
EVT MTy = getValueType(Ty);
unsigned Cost = 1;
// We keep legalizing the type until we find a legal kind. We assume that
// the only operation that costs anything is the split. After splitting
// we need to handle two types.
while (true) {
LegalizeKind LK = getTypeConversion(C, MTy);
if (LK.first == TypeLegal)
return std::make_pair(Cost, MTy.getSimpleVT());
if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
Cost *= 2;
// Keep legalizing the type.
MTy = LK.second;
}
}
bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg,
unsigned Reg, int Offset) {
assert(VT.isSimple() && "Non-simple types are invalid here!");
unsigned Opc;
switch (VT.getSimpleVT().SimpleTy) {
default:
assert(false && "Trying to emit for an unhandled type!");
return false;
case MVT::i16:
Opc = isThumb ? ARM::tLDRH : ARM::LDRH;
VT = MVT::i32;
break;
case MVT::i8:
Opc = isThumb ? ARM::tLDRB : ARM::LDRB;
VT = MVT::i32;
break;
case MVT::i32:
Opc = isThumb ? ARM::tLDR : ARM::LDR;
break;
}
ResultReg = createResultReg(TLI.getRegClassFor(VT));
// TODO: Fix the Addressing modes so that these can share some code.
// Since this is a Thumb1 load this will work in Thumb1 or 2 mode.
if (isThumb)
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ResultReg)
.addReg(Reg).addImm(Offset).addReg(0));
else
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ResultReg)
.addReg(Reg).addReg(0).addImm(Offset));
return true;
}
/// LowerCCCArguments - transform physical registers into virtual registers and
/// generate load operations for arguments places on the stack.
// FIXME: struct return stuff
// FIXME: varargs
SDValue
MSP430TargetLowering::LowerCCCArguments(SDValue Chain,
CallingConv::ID CallConv,
bool isVarArg,
const SmallVectorImpl<ISD::InputArg>
&Ins,
DebugLoc dl,
SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals)
const {
MachineFunction &MF = DAG.getMachineFunction();
MachineFrameInfo *MFI = MF.getFrameInfo();
MachineRegisterInfo &RegInfo = MF.getRegInfo();
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
ArgLocs, *DAG.getContext());
CCInfo.AnalyzeFormalArguments(Ins, CC_MSP430);
assert(!isVarArg && "Varargs not supported yet");
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
if (VA.isRegLoc()) {
// Arguments passed in registers
EVT RegVT = VA.getLocVT();
switch (RegVT.getSimpleVT().SimpleTy) {
default:
{
#ifndef NDEBUG
errs() << "LowerFormalArguments Unhandled argument type: "
<< RegVT.getSimpleVT().SimpleTy << "\n";
#endif
llvm_unreachable(0);
}
case MVT::i16:
unsigned VReg =
RegInfo.createVirtualRegister(MSP430::GR16RegisterClass);
RegInfo.addLiveIn(VA.getLocReg(), VReg);
SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, VReg, RegVT);
// If this is an 8-bit value, it is really passed promoted to 16
// bits. Insert an assert[sz]ext to capture this, then truncate to the
// right size.
if (VA.getLocInfo() == CCValAssign::SExt)
ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
DAG.getValueType(VA.getValVT()));
else if (VA.getLocInfo() == CCValAssign::ZExt)
ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
DAG.getValueType(VA.getValVT()));
if (VA.getLocInfo() != CCValAssign::Full)
ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
InVals.push_back(ArgValue);
}
} else {
// Sanity check
assert(VA.isMemLoc());
// Load the argument to a virtual register
unsigned ObjSize = VA.getLocVT().getSizeInBits()/8;
if (ObjSize > 2) {
errs() << "LowerFormalArguments Unhandled argument type: "
<< EVT(VA.getLocVT()).getEVTString()
<< "\n";
}
// Create the frame index object for this incoming parameter...
int FI = MFI->CreateFixedObject(ObjSize, VA.getLocMemOffset(), true);
// Create the SelectionDAG nodes corresponding to a load
//from this parameter
SDValue FIN = DAG.getFrameIndex(FI, MVT::i16);
InVals.push_back(DAG.getLoad(VA.getLocVT(), dl, Chain, FIN,
MachinePointerInfo::getFixedStack(FI),
false, false, 0));
}
}
return Chain;
}