SDValue
StatepointLoweringState::allocateStackSlot(EVT ValueType,
SelectionDAGBuilder &Builder) {
NumSlotsAllocatedForStatepoints++;
MachineFrameInfo &MFI = Builder.DAG.getMachineFunction().getFrameInfo();
unsigned SpillSize = ValueType.getSizeInBits() / 8;
assert((SpillSize * 8) == ValueType.getSizeInBits() && "Size not in bytes?");
// First look for a previously created stack slot which is not in
// use (accounting for the fact arbitrary slots may already be
// reserved), or to create a new stack slot and use it.
const size_t NumSlots = AllocatedStackSlots.size();
assert(NextSlotToAllocate <= NumSlots && "Broken invariant");
assert(AllocatedStackSlots.size() ==
Builder.FuncInfo.StatepointStackSlots.size() &&
"Broken invariant");
for (; NextSlotToAllocate < NumSlots; NextSlotToAllocate++) {
if (!AllocatedStackSlots.test(NextSlotToAllocate)) {
const int FI = Builder.FuncInfo.StatepointStackSlots[NextSlotToAllocate];
if (MFI.getObjectSize(FI) == SpillSize) {
AllocatedStackSlots.set(NextSlotToAllocate);
// TODO: Is ValueType the right thing to use here?
return Builder.DAG.getFrameIndex(FI, ValueType);
}
}
}
// Couldn't find a free slot, so create a new one:
SDValue SpillSlot = Builder.DAG.CreateStackTemporary(ValueType);
const unsigned FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
MFI.markAsStatepointSpillSlotObjectIndex(FI);
Builder.FuncInfo.StatepointStackSlots.push_back(FI);
AllocatedStackSlots.resize(AllocatedStackSlots.size()+1, true);
assert(AllocatedStackSlots.size() ==
Builder.FuncInfo.StatepointStackSlots.size() &&
"Broken invariant");
StatepointMaxSlotsRequired = std::max<unsigned long>(
StatepointMaxSlotsRequired, Builder.FuncInfo.StatepointStackSlots.size());
return SpillSlot;
}
/// Get the iX type with the same bitwidth as T.
IntegerType *AtomicExpand::getCorrespondingIntegerType(Type *T,
const DataLayout &DL) {
EVT VT = TLI->getValueType(DL, T);
unsigned BitWidth = VT.getStoreSizeInBits();
assert(BitWidth == VT.getSizeInBits() && "must be a power of two");
return IntegerType::get(T->getContext(), BitWidth);
}
/// ComplexPattern used on Cpu0InstrInfo
/// Used on Cpu0 Load/Store instructions
bool Cpu0DAGToDAGISel::
SelectAddr(SDNode *Parent, SDValue Addr, SDValue &Base, SDValue &Offset) {
EVT ValTy = Addr.getValueType();
// If Parent is an unaligned f32 load or store, select a (base + index)
// floating point load/store instruction (luxc1 or suxc1).
const LSBaseSDNode* LS = 0;
if (Parent && (LS = dyn_cast<LSBaseSDNode>(Parent))) {
EVT VT = LS->getMemoryVT();
if (VT.getSizeInBits() / 8 > LS->getAlignment()) {
assert(TLI.allowsUnalignedMemoryAccesses(VT) &&
"Unaligned loads/stores not supported for this type.");
if (VT == MVT::f32)
return false;
}
}
// if Address is FI, get the TargetFrameIndex.
if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), ValTy);
Offset = CurDAG->getTargetConstant(0, ValTy);
return true;
}
// on PIC code Load GA
if (Addr.getOpcode() == Cpu0ISD::Wrapper) {
Base = Addr.getOperand(0);
Offset = Addr.getOperand(1);
return true;
}
if (TM.getRelocationModel() != Reloc::PIC_) {
if ((Addr.getOpcode() == ISD::TargetExternalSymbol ||
Addr.getOpcode() == ISD::TargetGlobalAddress))
return false;
}
// Addresses of the form FI+const or FI|const
if (CurDAG->isBaseWithConstantOffset(Addr)) {
ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Addr.getOperand(1));
if (isInt<16>(CN->getSExtValue())) {
// If the first operand is a FI, get the TargetFI Node
if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>
(Addr.getOperand(0)))
Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), ValTy);
else
Base = Addr.getOperand(0);
Offset = CurDAG->getTargetConstant(CN->getZExtValue(), ValTy);
return true;
}
}
Base = Addr;
Offset = CurDAG->getTargetConstant(0, ValTy);
return true;
}
SDValue DAGTypeLegalizer::ExpandOp_NormalStore(SDNode *N, unsigned OpNo) {
assert(ISD::isNormalStore(N) && "This routine only for normal stores!");
assert(OpNo == 1 && "Can only expand the stored value so far");
DebugLoc dl = N->getDebugLoc();
StoreSDNode *St = cast<StoreSDNode>(N);
EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), St->getValue().getValueType());
SDValue Chain = St->getChain();
SDValue Ptr = St->getBasePtr();
int SVOffset = St->getSrcValueOffset();
unsigned Alignment = St->getAlignment();
bool isVolatile = St->isVolatile();
assert(NVT.isByteSized() && "Expanded type not byte sized!");
unsigned IncrementSize = NVT.getSizeInBits() / 8;
SDValue Lo, Hi;
GetExpandedOp(St->getValue(), Lo, Hi);
if (TLI.isBigEndian())
std::swap(Lo, Hi);
Lo = DAG.getStore(Chain, dl, Lo, Ptr, St->getSrcValue(), SVOffset,
isVolatile, Alignment);
Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
DAG.getIntPtrConstant(IncrementSize));
assert(isTypeLegal(Ptr.getValueType()) && "Pointers must be legal!");
Hi = DAG.getStore(Chain, dl, Hi, Ptr, St->getSrcValue(),
SVOffset + IncrementSize,
isVolatile, MinAlign(Alignment, IncrementSize));
return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
}
static SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) {
SDNode *Node = Op.getNode();
EVT VT = Node->getValueType(0);
SDValue InChain = Node->getOperand(0);
SDValue VAListPtr = Node->getOperand(1);
const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
DebugLoc dl = Node->getDebugLoc();
SDValue VAList = DAG.getLoad(MVT::i32, dl, InChain, VAListPtr, SV, 0,
false, false, 0);
// Increment the pointer, VAList, to the next vaarg
SDValue NextPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, VAList,
DAG.getConstant(VT.getSizeInBits()/8,
MVT::i32));
// Store the incremented VAList to the legalized pointer
InChain = DAG.getStore(VAList.getValue(1), dl, NextPtr,
VAListPtr, SV, 0, false, false, 0);
// Load the actual argument out of the pointer VAList, unless this is an
// f64 load.
if (VT != MVT::f64)
return DAG.getLoad(VT, dl, InChain, VAList, NULL, 0, false, false, 0);
// Otherwise, load it as i64, then do a bitconvert.
SDValue V = DAG.getLoad(MVT::i64, dl, InChain, VAList, NULL, 0,
false, false, 0);
// Bit-Convert the value to f64.
SDValue Ops[2] = {
DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, V),
V.getValue(1)
};
return DAG.getMergeValues(Ops, 2, dl);
}
MVT WebAssemblyTargetLowering::getScalarShiftAmountTy(const DataLayout & /*DL*/,
EVT VT) const {
unsigned BitWidth = NextPowerOf2(VT.getSizeInBits() - 1);
if (BitWidth > 1 && BitWidth < 8) BitWidth = 8;
if (BitWidth > 64) {
BitWidth = 64;
assert(BitWidth >= Log2_32_Ceil(VT.getSizeInBits()) &&
"64-bit shift counts ought to be enough for anyone");
}
MVT Result = MVT::getIntegerVT(BitWidth);
assert(Result != MVT::INVALID_SIMPLE_VALUE_TYPE &&
"Unable to represent scalar shift amount type");
return Result;
}
/// ComplexPattern used on Cpu0InstrInfo
/// Used on Cpu0 Load/Store instructions
bool Cpu0DAGToDAGISel::
SelectAddr(SDNode *Parent, SDValue Addr, SDValue &Base, SDValue &Offset) {
EVT ValTy = Addr.getValueType();
// If Parent is an unaligned f32 load or store, select a (base + index)
// floating point load/store instruction (luxc1 or suxc1).
const LSBaseSDNode* LS = 0;
if (Parent && (LS = dyn_cast<LSBaseSDNode>(Parent))) {
EVT VT = LS->getMemoryVT();
if (VT.getSizeInBits() / 8 > LS->getAlignment()) {
assert(TLI.allowsUnalignedMemoryAccesses(VT) &&
"Unaligned loads/stores not supported for this type.");
if (VT == MVT::f32)
return false;
}
}
// if Address is FI, get the TargetFrameIndex.
if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), ValTy);
Offset = CurDAG->getTargetConstant(0, ValTy);
return true;
}
Base = Addr;
Offset = CurDAG->getTargetConstant(0, ValTy);
return true;
}
MVT WebAssemblyTargetLowering::getScalarShiftAmountTy(const DataLayout & /*DL*/,
EVT VT) const {
unsigned BitWidth = NextPowerOf2(VT.getSizeInBits() - 1);
if (BitWidth > 1 && BitWidth < 8) BitWidth = 8;
if (BitWidth > 64) {
// The shift will be lowered to a libcall, and compiler-rt libcalls expect
// the count to be an i32.
BitWidth = 32;
assert(BitWidth >= Log2_32_Ceil(VT.getSizeInBits()) &&
"32-bit shift counts ought to be enough for anyone");
}
MVT Result = MVT::getIntegerVT(BitWidth);
assert(Result != MVT::INVALID_SIMPLE_VALUE_TYPE &&
"Unable to represent scalar shift amount type");
return Result;
}
// Select constant vector splats whose value is a power of 2.
//
// In addition to the requirements of selectVSplat(), this function returns
// true and sets Imm if:
// * The splat value is the same width as the elements of the vector
// * The splat value is a power of two.
//
// This function looks through ISD::BITCAST nodes.
// TODO: This might not be appropriate for big-endian MSA since BITCAST is
// sometimes a shuffle in big-endian mode.
bool MipsSEDAGToDAGISel::selectVSplatUimmPow2(SDValue N, SDValue &Imm) const {
APInt ImmValue;
EVT EltTy = N->getValueType(0).getVectorElementType();
if (N->getOpcode() == ISD::BITCAST)
N = N->getOperand(0);
if (selectVSplat(N.getNode(), ImmValue, EltTy.getSizeInBits()) &&
ImmValue.getBitWidth() == EltTy.getSizeInBits()) {
int32_t Log2 = ImmValue.exactLogBase2();
if (Log2 != -1) {
Imm = CurDAG->getTargetConstant(Log2, SDLoc(N), EltTy);
return true;
}
}
return false;
}
/// 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;
}
// Select constant vector splats whose value only has a consecutive sequence
// of right-most bits set (e.g. 0b00...0011...11).
//
// In addition to the requirements of selectVSplat(), this function returns
// true and sets Imm if:
// * The splat value is the same width as the elements of the vector
// * The splat value is a consecutive sequence of right-most bits.
//
// This function looks through ISD::BITCAST nodes.
// TODO: This might not be appropriate for big-endian MSA since BITCAST is
// sometimes a shuffle in big-endian mode.
bool MipsSEDAGToDAGISel::selectVSplatMaskR(SDValue N, SDValue &Imm) const {
APInt ImmValue;
EVT EltTy = N->getValueType(0).getVectorElementType();
if (N->getOpcode() == ISD::BITCAST)
N = N->getOperand(0);
if (selectVSplat(N.getNode(), ImmValue, EltTy.getSizeInBits()) &&
ImmValue.getBitWidth() == EltTy.getSizeInBits()) {
// Extract the run of set bits starting with bit zero, and test that the
// result is the same as the original value
if (ImmValue == (ImmValue & ~(ImmValue + 1))) {
Imm = CurDAG->getTargetConstant(ImmValue.countPopulation(), SDLoc(N),
EltTy);
return true;
}
}
return false;
}
// Select constant vector splats.
//
// In addition to the requirements of selectVSplat(), this function returns
// true and sets Imm if:
// * The splat value is the same width as the elements of the vector
// * The splat value fits in an integer with the specified signed-ness and
// width.
//
// This function looks through ISD::BITCAST nodes.
// TODO: This might not be appropriate for big-endian MSA since BITCAST is
// sometimes a shuffle in big-endian mode.
//
// It's worth noting that this function is not used as part of the selection
// of ldi.[bhwd] since it does not permit using the wrong-typed ldi.[bhwd]
// instruction to achieve the desired bit pattern. ldi.[bhwd] is selected in
// MipsSEDAGToDAGISel::selectNode.
bool MipsSEDAGToDAGISel::
selectVSplatCommon(SDValue N, SDValue &Imm, bool Signed,
unsigned ImmBitSize) const {
APInt ImmValue;
EVT EltTy = N->getValueType(0).getVectorElementType();
if (N->getOpcode() == ISD::BITCAST)
N = N->getOperand(0);
if (selectVSplat(N.getNode(), ImmValue, EltTy.getSizeInBits()) &&
ImmValue.getBitWidth() == EltTy.getSizeInBits()) {
if (( Signed && ImmValue.isSignedIntN(ImmBitSize)) ||
(!Signed && ImmValue.isIntN(ImmBitSize))) {
Imm = CurDAG->getTargetConstant(ImmValue, SDLoc(N), EltTy);
return true;
}
}
return false;
}
SDNode *AArch64DAGToDAGISel::TrySelectToMoveImm(SDNode *Node) {
SDNode *ResNode;
SDLoc dl(Node);
EVT DestType = Node->getValueType(0);
unsigned DestWidth = DestType.getSizeInBits();
unsigned MOVOpcode;
EVT MOVType;
int UImm16, Shift;
uint32_t LogicalBits;
uint64_t BitPat = cast<ConstantSDNode>(Node)->getZExtValue();
if (A64Imms::isMOVZImm(DestWidth, BitPat, UImm16, Shift)) {
MOVType = DestType;
MOVOpcode = DestWidth == 64 ? AArch64::MOVZxii : AArch64::MOVZwii;
} else if (A64Imms::isMOVNImm(DestWidth, BitPat, UImm16, Shift)) {
MOVType = DestType;
MOVOpcode = DestWidth == 64 ? AArch64::MOVNxii : AArch64::MOVNwii;
} else if (DestWidth == 64 && A64Imms::isMOVNImm(32, BitPat, UImm16, Shift)) {
// To get something like 0x0000_0000_ffff_1234 into a 64-bit register we can
// use a 32-bit instruction: "movn w0, 0xedbc".
MOVType = MVT::i32;
MOVOpcode = AArch64::MOVNwii;
} else if (A64Imms::isLogicalImm(DestWidth, BitPat, LogicalBits)) {
MOVOpcode = DestWidth == 64 ? AArch64::ORRxxi : AArch64::ORRwwi;
uint16_t ZR = DestWidth == 64 ? AArch64::XZR : AArch64::WZR;
return CurDAG->getMachineNode(MOVOpcode, dl, DestType,
CurDAG->getRegister(ZR, DestType),
CurDAG->getTargetConstant(LogicalBits, MVT::i32));
} else {
// Can't handle it in one instruction. There's scope for permitting two (or
// more) instructions, but that'll need more thought.
return NULL;
}
ResNode = CurDAG->getMachineNode(MOVOpcode, dl, MOVType,
CurDAG->getTargetConstant(UImm16, MVT::i32),
CurDAG->getTargetConstant(Shift, MVT::i32));
if (MOVType != DestType) {
ResNode = CurDAG->getMachineNode(TargetOpcode::SUBREG_TO_REG, dl,
MVT::i64, MVT::i32, MVT::Other,
CurDAG->getTargetConstant(0, MVT::i64),
SDValue(ResNode, 0),
CurDAG->getTargetConstant(AArch64::sub_32, MVT::i32));
}
return ResNode;
}
SDNode *AArch64DAGToDAGISel::SelectToLitPool(SDNode *Node) {
DebugLoc DL = Node->getDebugLoc();
uint64_t UnsignedVal = cast<ConstantSDNode>(Node)->getZExtValue();
int64_t SignedVal = cast<ConstantSDNode>(Node)->getSExtValue();
EVT DestType = Node->getValueType(0);
EVT PtrVT = TLI.getPointerTy();
// Since we may end up loading a 64-bit constant from a 32-bit entry the
// constant in the pool may have a different type to the eventual node.
ISD::LoadExtType Extension;
EVT MemType;
assert((DestType == MVT::i64 || DestType == MVT::i32)
&& "Only expect integer constants at the moment");
if (DestType == MVT::i32) {
Extension = ISD::NON_EXTLOAD;
MemType = MVT::i32;
} else if (UnsignedVal <= UINT32_MAX) {
Extension = ISD::ZEXTLOAD;
MemType = MVT::i32;
} else if (SignedVal >= INT32_MIN && SignedVal <= INT32_MAX) {
Extension = ISD::SEXTLOAD;
MemType = MVT::i32;
} else {
Extension = ISD::NON_EXTLOAD;
MemType = MVT::i64;
}
Constant *CV = ConstantInt::get(Type::getIntNTy(*CurDAG->getContext(),
MemType.getSizeInBits()),
UnsignedVal);
SDValue PoolAddr;
unsigned Alignment = TLI.getDataLayout()->getABITypeAlignment(CV->getType());
PoolAddr = CurDAG->getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0,
AArch64II::MO_NO_FLAG),
CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0,
AArch64II::MO_LO12),
CurDAG->getConstant(Alignment, MVT::i32));
return CurDAG->getExtLoad(Extension, DL, DestType, CurDAG->getEntryNode(),
PoolAddr,
MachinePointerInfo::getConstantPool(), MemType,
/* isVolatile = */ false,
/* isNonTemporal = */ false,
Alignment).getNode();
}
/// DecodeUNPCKLMask - This decodes the shuffle masks for unpcklps/unpcklpd
/// and punpckl*. VT indicates the type of the vector allowing it to handle
/// different datatypes and vector widths.
void DecodeUNPCKLMask(EVT VT, SmallVectorImpl<unsigned> &ShuffleMask) {
unsigned NumElts = VT.getVectorNumElements();
// Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
// independently on 128-bit lanes.
unsigned NumLanes = VT.getSizeInBits() / 128;
if (NumLanes == 0 ) NumLanes = 1; // Handle MMX
unsigned NumLaneElts = NumElts / NumLanes;
for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
for (unsigned i = l, e = l + NumLaneElts/2; i != e; ++i) {
ShuffleMask.push_back(i); // Reads from dest/src1
ShuffleMask.push_back(i+NumElts); // Reads from src/src2
}
}
}
/// DecodePSHUFMask - This decodes the shuffle masks for pshufd, and vpermilp*.
/// VT indicates the type of the vector allowing it to handle different
/// datatypes and vector widths.
void DecodePSHUFMask(EVT VT, unsigned Imm,
SmallVectorImpl<unsigned> &ShuffleMask) {
unsigned NumElts = VT.getVectorNumElements();
unsigned NumLanes = VT.getSizeInBits() / 128;
unsigned NumLaneElts = NumElts / NumLanes;
int NewImm = Imm;
for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
for (unsigned i = 0; i != NumLaneElts; ++i) {
ShuffleMask.push_back(NewImm % NumLaneElts + l);
NewImm /= NumLaneElts;
}
if (NumLaneElts == 4) NewImm = Imm; // reload imm
}
}
void DAGTypeLegalizer::ExpandRes_NormalLoad(SDNode *N, SDValue &Lo,
SDValue &Hi) {
assert(ISD::isNormalLoad(N) && "This routine only for normal loads!");
SDLoc dl(N);
LoadSDNode *LD = cast<LoadSDNode>(N);
EVT ValueVT = LD->getValueType(0);
EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), ValueVT);
SDValue Chain = LD->getChain();
SDValue Ptr = LD->getBasePtr();
unsigned Alignment = LD->getAlignment();
bool isVolatile = LD->isVolatile();
bool isNonTemporal = LD->isNonTemporal();
bool isInvariant = LD->isInvariant();
AAMDNodes AAInfo = LD->getAAInfo();
assert(NVT.isByteSized() && "Expanded type not byte sized!");
Lo = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getPointerInfo(),
isVolatile, isNonTemporal, isInvariant, Alignment,
AAInfo);
// Increment the pointer to the other half.
unsigned IncrementSize = NVT.getSizeInBits() / 8;
Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
DAG.getConstant(IncrementSize, dl, Ptr.getValueType()));
Hi = DAG.getLoad(NVT, dl, Chain, Ptr,
LD->getPointerInfo().getWithOffset(IncrementSize),
isVolatile, isNonTemporal, isInvariant,
MinAlign(Alignment, IncrementSize), AAInfo);
// Build a factor node to remember that this load is independent of the
// other one.
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
Hi.getValue(1));
// Handle endianness of the load.
if (TLI.hasBigEndianPartOrdering(ValueVT))
std::swap(Lo, Hi);
// Modified the chain - switch anything that used the old chain to use
// the new one.
ReplaceValueWith(SDValue(N, 1), Chain);
}
/// 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 MSP430TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
if (!VT1.isInteger() || !VT2.isInteger())
return false;
return (VT1.getSizeInBits() > VT2.getSizeInBits());
}
SDValue
NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
SmallVectorImpl<SDValue> &InVals) const {
SelectionDAG &DAG = CLI.DAG;
DebugLoc &dl = CLI.DL;
SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
SDValue Chain = CLI.Chain;
SDValue Callee = CLI.Callee;
bool &isTailCall = CLI.IsTailCall;
ArgListTy &Args = CLI.Args;
Type *retTy = CLI.RetTy;
ImmutableCallSite *CS = CLI.CS;
bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
SDValue tempChain = Chain;
Chain = DAG.getCALLSEQ_START(Chain,
DAG.getIntPtrConstant(uniqueCallSite, true));
SDValue InFlag = Chain.getValue(1);
assert((Outs.size() == Args.size()) &&
"Unexpected number of arguments to function call");
unsigned paramCount = 0;
// Declare the .params or .reg need to pass values
// to the function
for (unsigned i=0, e=Outs.size(); i!=e; ++i) {
EVT VT = Outs[i].VT;
if (Outs[i].Flags.isByVal() == false) {
// Plain scalar
// for ABI, declare .param .b<size> .param<n>;
// for nonABI, declare .reg .b<size> .param<n>;
unsigned isReg = 1;
if (isABI)
isReg = 0;
unsigned sz = VT.getSizeInBits();
if (VT.isInteger() && (sz < 32)) sz = 32;
SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue DeclareParamOps[] = { Chain,
DAG.getConstant(paramCount, MVT::i32),
DAG.getConstant(sz, MVT::i32),
DAG.getConstant(isReg, MVT::i32),
InFlag };
Chain = DAG.getNode(NVPTXISD::DeclareScalarParam, dl, DeclareParamVTs,
DeclareParamOps, 5);
InFlag = Chain.getValue(1);
SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue CopyParamOps[] = { Chain, DAG.getConstant(paramCount, MVT::i32),
DAG.getConstant(0, MVT::i32), OutVals[i], InFlag };
unsigned opcode = NVPTXISD::StoreParam;
if (isReg)
opcode = NVPTXISD::MoveToParam;
else {
if (Outs[i].Flags.isZExt())
opcode = NVPTXISD::StoreParamU32;
else if (Outs[i].Flags.isSExt())
opcode = NVPTXISD::StoreParamS32;
}
Chain = DAG.getNode(opcode, dl, CopyParamVTs, CopyParamOps, 5);
InFlag = Chain.getValue(1);
++paramCount;
continue;
}
// struct or vector
SmallVector<EVT, 16> vtparts;
const PointerType *PTy = dyn_cast<PointerType>(Args[i].Ty);
assert(PTy &&
"Type of a byval parameter should be pointer");
ComputeValueVTs(*this, PTy->getElementType(), vtparts);
if (isABI) {
// declare .param .align 16 .b8 .param<n>[<size>];
unsigned sz = Outs[i].Flags.getByValSize();
SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
// The ByValAlign in the Outs[i].Flags is alway set at this point, so we
// don't need to
// worry about natural alignment or not. See TargetLowering::LowerCallTo()
SDValue DeclareParamOps[] = { Chain,
DAG.getConstant(Outs[i].Flags.getByValAlign(), MVT::i32),
DAG.getConstant(paramCount, MVT::i32),
DAG.getConstant(sz, MVT::i32),
InFlag };
Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
DeclareParamOps, 5);
InFlag = Chain.getValue(1);
unsigned curOffset = 0;
for (unsigned j=0,je=vtparts.size(); j!=je; ++j) {
unsigned elems = 1;
EVT elemtype = vtparts[j];
if (vtparts[j].isVector()) {
elems = vtparts[j].getVectorNumElements();
elemtype = vtparts[j].getVectorElementType();
}
for (unsigned k=0,ke=elems; k!=ke; ++k) {
unsigned sz = elemtype.getSizeInBits();
if (elemtype.isInteger() && (sz < 8)) sz = 8;
SDValue srcAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(),
OutVals[i],
DAG.getConstant(curOffset,
getPointerTy()));
SDValue theVal = DAG.getLoad(elemtype, dl, tempChain, srcAddr,
MachinePointerInfo(), false, false, false, 0);
SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue CopyParamOps[] = { Chain, DAG.getConstant(paramCount,
MVT::i32),
DAG.getConstant(curOffset, MVT::i32),
theVal, InFlag };
Chain = DAG.getNode(NVPTXISD::StoreParam, dl, CopyParamVTs,
CopyParamOps, 5);
InFlag = Chain.getValue(1);
curOffset += sz/8;
}
}
++paramCount;
continue;
}
// Non-abi, struct or vector
// Declare a bunch or .reg .b<size> .param<n>
unsigned curOffset = 0;
for (unsigned j=0,je=vtparts.size(); j!=je; ++j) {
unsigned elems = 1;
EVT elemtype = vtparts[j];
if (vtparts[j].isVector()) {
elems = vtparts[j].getVectorNumElements();
elemtype = vtparts[j].getVectorElementType();
}
for (unsigned k=0,ke=elems; k!=ke; ++k) {
unsigned sz = elemtype.getSizeInBits();
if (elemtype.isInteger() && (sz < 32)) sz = 32;
SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue DeclareParamOps[] = { Chain, DAG.getConstant(paramCount,
MVT::i32),
DAG.getConstant(sz, MVT::i32),
DAG.getConstant(1, MVT::i32),
InFlag };
Chain = DAG.getNode(NVPTXISD::DeclareScalarParam, dl, DeclareParamVTs,
DeclareParamOps, 5);
InFlag = Chain.getValue(1);
SDValue srcAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), OutVals[i],
DAG.getConstant(curOffset,
getPointerTy()));
SDValue theVal = DAG.getLoad(elemtype, dl, tempChain, srcAddr,
MachinePointerInfo(), false, false, false, 0);
SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue CopyParamOps[] = { Chain, DAG.getConstant(paramCount, MVT::i32),
DAG.getConstant(0, MVT::i32), theVal,
InFlag };
Chain = DAG.getNode(NVPTXISD::MoveToParam, dl, CopyParamVTs,
CopyParamOps, 5);
InFlag = Chain.getValue(1);
++paramCount;
}
}
}
GlobalAddressSDNode *Func = dyn_cast<GlobalAddressSDNode>(Callee.getNode());
unsigned retAlignment = 0;
// Handle Result
unsigned retCount = 0;
if (Ins.size() > 0) {
SmallVector<EVT, 16> resvtparts;
ComputeValueVTs(*this, retTy, resvtparts);
// Declare one .param .align 16 .b8 func_retval0[<size>] for ABI or
// individual .reg .b<size> func_retval<0..> for non ABI
unsigned resultsz = 0;
for (unsigned i=0,e=resvtparts.size(); i!=e; ++i) {
unsigned elems = 1;
EVT elemtype = resvtparts[i];
if (resvtparts[i].isVector()) {
elems = resvtparts[i].getVectorNumElements();
elemtype = resvtparts[i].getVectorElementType();
}
for (unsigned j=0,je=elems; j!=je; ++j) {
unsigned sz = elemtype.getSizeInBits();
if (isABI == false) {
if (elemtype.isInteger() && (sz < 32)) sz = 32;
}
else {
if (elemtype.isInteger() && (sz < 8)) sz = 8;
}
if (isABI == false) {
SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue DeclareRetOps[] = { Chain, DAG.getConstant(2, MVT::i32),
DAG.getConstant(sz, MVT::i32),
DAG.getConstant(retCount, MVT::i32),
InFlag };
Chain = DAG.getNode(NVPTXISD::DeclareRet, dl, DeclareRetVTs,
DeclareRetOps, 5);
InFlag = Chain.getValue(1);
++retCount;
}
resultsz += sz;
}
}
if (isABI) {
if (retTy->isPrimitiveType() || retTy->isIntegerTy() ||
retTy->isPointerTy() ) {
// Scalar needs to be at least 32bit wide
if (resultsz < 32)
resultsz = 32;
SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue DeclareRetOps[] = { Chain, DAG.getConstant(1, MVT::i32),
DAG.getConstant(resultsz, MVT::i32),
DAG.getConstant(0, MVT::i32), InFlag };
Chain = DAG.getNode(NVPTXISD::DeclareRet, dl, DeclareRetVTs,
DeclareRetOps, 5);
InFlag = Chain.getValue(1);
}
else {
if (Func) { // direct call
if (!llvm::getAlign(*(CS->getCalledFunction()), 0, retAlignment))
retAlignment = getDataLayout()->getABITypeAlignment(retTy);
} else { // indirect call
const CallInst *CallI = dyn_cast<CallInst>(CS->getInstruction());
if (!llvm::getAlign(*CallI, 0, retAlignment))
retAlignment = getDataLayout()->getABITypeAlignment(retTy);
}
SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue DeclareRetOps[] = { Chain, DAG.getConstant(retAlignment,
MVT::i32),
DAG.getConstant(resultsz/8, MVT::i32),
DAG.getConstant(0, MVT::i32), InFlag };
Chain = DAG.getNode(NVPTXISD::DeclareRetParam, dl, DeclareRetVTs,
DeclareRetOps, 5);
InFlag = Chain.getValue(1);
}
}
}
if (!Func) {
// This is indirect function call case : PTX requires a prototype of the
// form
// proto_0 : .callprototype(.param .b32 _) _ (.param .b32 _);
// to be emitted, and the label has to used as the last arg of call
// instruction.
// The prototype is embedded in a string and put as the operand for an
// INLINEASM SDNode.
SDVTList InlineAsmVTs = DAG.getVTList(MVT::Other, MVT::Glue);
std::string proto_string = getPrototype(retTy, Args, Outs, retAlignment);
const char *asmstr = nvTM->getManagedStrPool()->
getManagedString(proto_string.c_str())->c_str();
SDValue InlineAsmOps[] = { Chain,
DAG.getTargetExternalSymbol(asmstr,
getPointerTy()),
DAG.getMDNode(0),
DAG.getTargetConstant(0, MVT::i32), InFlag };
Chain = DAG.getNode(ISD::INLINEASM, dl, InlineAsmVTs, InlineAsmOps, 5);
InFlag = Chain.getValue(1);
}
// Op to just print "call"
SDVTList PrintCallVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue PrintCallOps[] = { Chain,
DAG.getConstant(isABI ? ((Ins.size()==0) ? 0 : 1)
: retCount, MVT::i32),
InFlag };
Chain = DAG.getNode(Func?(NVPTXISD::PrintCallUni):(NVPTXISD::PrintCall), dl,
PrintCallVTs, PrintCallOps, 3);
InFlag = Chain.getValue(1);
// Ops to print out the function name
SDVTList CallVoidVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue CallVoidOps[] = { Chain, Callee, InFlag };
Chain = DAG.getNode(NVPTXISD::CallVoid, dl, CallVoidVTs, CallVoidOps, 3);
InFlag = Chain.getValue(1);
// Ops to print out the param list
SDVTList CallArgBeginVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue CallArgBeginOps[] = { Chain, InFlag };
Chain = DAG.getNode(NVPTXISD::CallArgBegin, dl, CallArgBeginVTs,
CallArgBeginOps, 2);
InFlag = Chain.getValue(1);
for (unsigned i=0, e=paramCount; i!=e; ++i) {
unsigned opcode;
if (i==(e-1))
opcode = NVPTXISD::LastCallArg;
else
opcode = NVPTXISD::CallArg;
SDVTList CallArgVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue CallArgOps[] = { Chain, DAG.getConstant(1, MVT::i32),
DAG.getConstant(i, MVT::i32),
InFlag };
Chain = DAG.getNode(opcode, dl, CallArgVTs, CallArgOps, 4);
InFlag = Chain.getValue(1);
}
SDVTList CallArgEndVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue CallArgEndOps[] = { Chain,
DAG.getConstant(Func ? 1 : 0, MVT::i32),
InFlag };
Chain = DAG.getNode(NVPTXISD::CallArgEnd, dl, CallArgEndVTs, CallArgEndOps,
3);
InFlag = Chain.getValue(1);
if (!Func) {
SDVTList PrototypeVTs = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue PrototypeOps[] = { Chain,
DAG.getConstant(uniqueCallSite, MVT::i32),
InFlag };
Chain = DAG.getNode(NVPTXISD::Prototype, dl, PrototypeVTs, PrototypeOps, 3);
InFlag = Chain.getValue(1);
}
// Generate loads from param memory/moves from registers for result
if (Ins.size() > 0) {
if (isABI) {
unsigned resoffset = 0;
for (unsigned i=0,e=Ins.size(); i!=e; ++i) {
unsigned sz = Ins[i].VT.getSizeInBits();
if (Ins[i].VT.isInteger() && (sz < 8)) sz = 8;
std::vector<EVT> LoadRetVTs;
LoadRetVTs.push_back(Ins[i].VT);
LoadRetVTs.push_back(MVT::Other); LoadRetVTs.push_back(MVT::Glue);
std::vector<SDValue> LoadRetOps;
LoadRetOps.push_back(Chain);
LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
LoadRetOps.push_back(DAG.getConstant(resoffset, MVT::i32));
LoadRetOps.push_back(InFlag);
SDValue retval = DAG.getNode(NVPTXISD::LoadParam, dl, LoadRetVTs,
&LoadRetOps[0], LoadRetOps.size());
Chain = retval.getValue(1);
InFlag = retval.getValue(2);
InVals.push_back(retval);
resoffset += sz/8;
}
}
else {
SmallVector<EVT, 16> resvtparts;
ComputeValueVTs(*this, retTy, resvtparts);
assert(Ins.size() == resvtparts.size() &&
"Unexpected number of return values in non-ABI case");
unsigned paramNum = 0;
for (unsigned i=0,e=Ins.size(); i!=e; ++i) {
assert(EVT(Ins[i].VT) == resvtparts[i] &&
"Unexpected EVT type in non-ABI case");
unsigned numelems = 1;
EVT elemtype = Ins[i].VT;
if (Ins[i].VT.isVector()) {
numelems = Ins[i].VT.getVectorNumElements();
elemtype = Ins[i].VT.getVectorElementType();
}
std::vector<SDValue> tempRetVals;
for (unsigned j=0; j<numelems; ++j) {
std::vector<EVT> MoveRetVTs;
MoveRetVTs.push_back(elemtype);
MoveRetVTs.push_back(MVT::Other); MoveRetVTs.push_back(MVT::Glue);
std::vector<SDValue> MoveRetOps;
MoveRetOps.push_back(Chain);
MoveRetOps.push_back(DAG.getConstant(0, MVT::i32));
MoveRetOps.push_back(DAG.getConstant(paramNum, MVT::i32));
MoveRetOps.push_back(InFlag);
SDValue retval = DAG.getNode(NVPTXISD::LoadParam, dl, MoveRetVTs,
&MoveRetOps[0], MoveRetOps.size());
Chain = retval.getValue(1);
InFlag = retval.getValue(2);
tempRetVals.push_back(retval);
++paramNum;
}
if (Ins[i].VT.isVector())
InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, Ins[i].VT,
&tempRetVals[0], tempRetVals.size()));
else
InVals.push_back(tempRetVals[0]);
}
}
}
Chain = DAG.getCALLSEQ_END(Chain,
DAG.getIntPtrConstant(uniqueCallSite, true),
DAG.getIntPtrConstant(uniqueCallSite+1, true),
InFlag);
uniqueCallSite++;
// set isTailCall to false for now, until we figure out how to express
// tail call optimization in PTX
isTailCall = false;
return Chain;
}
std::string NVPTXTargetLowering::getPrototype(Type *retTy,
const ArgListTy &Args,
const SmallVectorImpl<ISD::OutputArg> &Outs,
unsigned retAlignment) const {
bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
std::stringstream O;
O << "prototype_" << uniqueCallSite << " : .callprototype ";
if (retTy->getTypeID() == Type::VoidTyID)
O << "()";
else {
O << "(";
if (isABI) {
if (retTy->isPrimitiveType() || retTy->isIntegerTy()) {
unsigned size = 0;
if (const IntegerType *ITy = dyn_cast<IntegerType>(retTy)) {
size = ITy->getBitWidth();
if (size < 32) size = 32;
}
else {
assert(retTy->isFloatingPointTy() &&
"Floating point type expected here");
size = retTy->getPrimitiveSizeInBits();
}
O << ".param .b" << size << " _";
}
else if (isa<PointerType>(retTy))
O << ".param .b" << getPointerTy().getSizeInBits()
<< " _";
else {
if ((retTy->getTypeID() == Type::StructTyID) ||
isa<VectorType>(retTy)) {
SmallVector<EVT, 16> vtparts;
ComputeValueVTs(*this, retTy, vtparts);
unsigned totalsz = 0;
for (unsigned i=0,e=vtparts.size(); i!=e; ++i) {
unsigned elems = 1;
EVT elemtype = vtparts[i];
if (vtparts[i].isVector()) {
elems = vtparts[i].getVectorNumElements();
elemtype = vtparts[i].getVectorElementType();
}
for (unsigned j=0, je=elems; j!=je; ++j) {
unsigned sz = elemtype.getSizeInBits();
if (elemtype.isInteger() && (sz < 8)) sz = 8;
totalsz += sz/8;
}
}
O << ".param .align "
<< retAlignment
<< " .b8 _["
<< totalsz << "]";
}
else {
assert(false &&
"Unknown return type");
}
}
}
else {
SmallVector<EVT, 16> vtparts;
ComputeValueVTs(*this, retTy, vtparts);
unsigned idx = 0;
for (unsigned i=0,e=vtparts.size(); i!=e; ++i) {
unsigned elems = 1;
EVT elemtype = vtparts[i];
if (vtparts[i].isVector()) {
elems = vtparts[i].getVectorNumElements();
elemtype = vtparts[i].getVectorElementType();
}
for (unsigned j=0, je=elems; j!=je; ++j) {
unsigned sz = elemtype.getSizeInBits();
if (elemtype.isInteger() && (sz < 32)) sz = 32;
O << ".reg .b" << sz << " _";
if (j<je-1) O << ", ";
++idx;
}
if (i < e-1)
O << ", ";
}
}
O << ") ";
}
O << "_ (";
bool first = true;
MVT thePointerTy = getPointerTy();
for (unsigned i=0,e=Args.size(); i!=e; ++i) {
const Type *Ty = Args[i].Ty;
if (!first) {
O << ", ";
}
first = false;
if (Outs[i].Flags.isByVal() == false) {
unsigned sz = 0;
if (isa<IntegerType>(Ty)) {
sz = cast<IntegerType>(Ty)->getBitWidth();
if (sz < 32) sz = 32;
}
else if (isa<PointerType>(Ty))
sz = thePointerTy.getSizeInBits();
else
sz = Ty->getPrimitiveSizeInBits();
if (isABI)
O << ".param .b" << sz << " ";
else
O << ".reg .b" << sz << " ";
O << "_";
continue;
}
const PointerType *PTy = dyn_cast<PointerType>(Ty);
assert(PTy &&
"Param with byval attribute should be a pointer type");
Type *ETy = PTy->getElementType();
if (isABI) {
unsigned align = Outs[i].Flags.getByValAlign();
unsigned sz = getDataLayout()->getTypeAllocSize(ETy);
O << ".param .align " << align
<< " .b8 ";
O << "_";
O << "[" << sz << "]";
continue;
}
else {
SmallVector<EVT, 16> vtparts;
ComputeValueVTs(*this, ETy, vtparts);
for (unsigned i=0,e=vtparts.size(); i!=e; ++i) {
unsigned elems = 1;
EVT elemtype = vtparts[i];
if (vtparts[i].isVector()) {
elems = vtparts[i].getVectorNumElements();
elemtype = vtparts[i].getVectorElementType();
}
for (unsigned j=0,je=elems; j!=je; ++j) {
unsigned sz = elemtype.getSizeInBits();
if (elemtype.isInteger() && (sz < 32)) sz = 32;
O << ".reg .b" << sz << " ";
O << "_";
if (j<je-1) O << ", ";
}
if (i<e-1)
O << ", ";
}
continue;
}
}
O << ");";
return O.str();
}
/// LowerCCCArguments - transform physical registers into virtual registers and
/// generate load operations for arguments places on the stack.
// FIXME: struct return stuff
// FIXME: varargs
SDValue
SystemZTargetLowering::LowerCCCArguments(SDValue Chain,
CallingConv::ID CallConv,
bool isVarArg,
const SmallVectorImpl<ISD::InputArg>
&Ins,
DebugLoc dl,
SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) {
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_SystemZ);
if (isVarArg)
report_fatal_error("Varargs not supported yet");
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
SDValue ArgValue;
CCValAssign &VA = ArgLocs[i];
EVT LocVT = VA.getLocVT();
if (VA.isRegLoc()) {
// Arguments passed in registers
TargetRegisterClass *RC;
switch (LocVT.getSimpleVT().SimpleTy) {
default:
#ifndef NDEBUG
errs() << "LowerFormalArguments Unhandled argument type: "
<< LocVT.getSimpleVT().SimpleTy
<< "\n";
#endif
llvm_unreachable(0);
case MVT::i64:
RC = SystemZ::GR64RegisterClass;
break;
case MVT::f32:
RC = SystemZ::FP32RegisterClass;
break;
case MVT::f64:
RC = SystemZ::FP64RegisterClass;
break;
}
unsigned VReg = RegInfo.createVirtualRegister(RC);
RegInfo.addLiveIn(VA.getLocReg(), VReg);
ArgValue = DAG.getCopyFromReg(Chain, dl, VReg, LocVT);
} else {
// Sanity check
assert(VA.isMemLoc());
// Create the nodes corresponding to a load from this parameter slot.
// Create the frame index object for this incoming parameter...
int FI = MFI->CreateFixedObject(LocVT.getSizeInBits()/8,
VA.getLocMemOffset(), true, false);
// Create the SelectionDAG nodes corresponding to a load
// from this parameter
SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
ArgValue = DAG.getLoad(LocVT, dl, Chain, FIN,
PseudoSourceValue::getFixedStack(FI), 0,
false, false, 0);
}
// If this is an 8/16/32-bit value, it is really passed promoted to 64
// 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, LocVT, ArgValue,
DAG.getValueType(VA.getValVT()));
else if (VA.getLocInfo() == CCValAssign::ZExt)
ArgValue = DAG.getNode(ISD::AssertZext, dl, LocVT, ArgValue,
DAG.getValueType(VA.getValVT()));
if (VA.getLocInfo() != CCValAssign::Full)
ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
InVals.push_back(ArgValue);
}
return Chain;
}
void DAGTypeLegalizer::ExpandRes_BIT_CONVERT(SDNode *N, SDValue &Lo,
SDValue &Hi) {
EVT OutVT = N->getValueType(0);
EVT NOutVT = TLI.getTypeToTransformTo(*DAG.getContext(), OutVT);
SDValue InOp = N->getOperand(0);
EVT InVT = InOp.getValueType();
DebugLoc dl = N->getDebugLoc();
// Handle some special cases efficiently.
switch (getTypeAction(InVT)) {
default:
assert(false && "Unknown type action!");
case Legal:
case PromoteInteger:
break;
case SoftenFloat:
// Convert the integer operand instead.
SplitInteger(GetSoftenedFloat(InOp), Lo, Hi);
Lo = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Hi);
return;
case ExpandInteger:
case ExpandFloat:
// Convert the expanded pieces of the input.
GetExpandedOp(InOp, Lo, Hi);
Lo = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Hi);
return;
case SplitVector:
GetSplitVector(InOp, Lo, Hi);
if (TLI.isBigEndian())
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Hi);
return;
case ScalarizeVector:
// Convert the element instead.
SplitInteger(BitConvertToInteger(GetScalarizedVector(InOp)), Lo, Hi);
Lo = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Hi);
return;
case WidenVector: {
assert(!(InVT.getVectorNumElements() & 1) && "Unsupported BIT_CONVERT");
InOp = GetWidenedVector(InOp);
EVT InNVT = EVT::getVectorVT(*DAG.getContext(), InVT.getVectorElementType(),
InVT.getVectorNumElements()/2);
Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, InOp,
DAG.getIntPtrConstant(0));
Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, InOp,
DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
if (TLI.isBigEndian())
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BIT_CONVERT, dl, NOutVT, Hi);
return;
}
}
if (InVT.isVector() && OutVT.isInteger()) {
// Handle cases like i64 = BIT_CONVERT v1i64 on x86, where the operand
// is legal but the result is not.
EVT NVT = EVT::getVectorVT(*DAG.getContext(), NOutVT, 2);
if (isTypeLegal(NVT)) {
SDValue CastInOp = DAG.getNode(ISD::BIT_CONVERT, dl, NVT, InOp);
Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NOutVT, CastInOp,
DAG.getIntPtrConstant(0));
Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NOutVT, CastInOp,
DAG.getIntPtrConstant(1));
if (TLI.isBigEndian())
std::swap(Lo, Hi);
return;
}
}
// Lower the bit-convert to a store/load from the stack.
assert(NOutVT.isByteSized() && "Expanded type not byte sized!");
// Create the stack frame object. Make sure it is aligned for both
// the source and expanded destination types.
unsigned Alignment =
TLI.getTargetData()->getPrefTypeAlignment(NOutVT.getTypeForEVT(*DAG.getContext()));
SDValue StackPtr = DAG.CreateStackTemporary(InVT, Alignment);
int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
const Value *SV = PseudoSourceValue::getFixedStack(SPFI);
// Emit a store to the stack slot.
SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, InOp, StackPtr, SV, 0);
// Load the first half from the stack slot.
Lo = DAG.getLoad(NOutVT, dl, Store, StackPtr, SV, 0);
// Increment the pointer to the other half.
unsigned IncrementSize = NOutVT.getSizeInBits() / 8;
StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
DAG.getIntPtrConstant(IncrementSize));
// Load the second half from the stack slot.
Hi = DAG.getLoad(NOutVT, dl, Store, StackPtr, SV, IncrementSize, false,
MinAlign(Alignment, IncrementSize));
// Handle endianness of the load.
if (TLI.isBigEndian())
std::swap(Lo, Hi);
}
/// materializeRegForValue - Helper for getRegForValue. This function is
/// called when the value isn't already available in a register and must
/// be materialized with new instructions.
unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) {
unsigned Reg = 0;
if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
if (CI->getValue().getActiveBits() <= 64)
Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
} else if (isa<AllocaInst>(V)) {
Reg = TargetMaterializeAlloca(cast<AllocaInst>(V));
} else if (isa<ConstantPointerNull>(V)) {
// Translate this as an integer zero so that it can be
// local-CSE'd with actual integer zeros.
Reg =
getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext())));
} else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
// Try to emit the constant directly.
Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
if (!Reg) {
// Try to emit the constant by using an integer constant with a cast.
const APFloat &Flt = CF->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) {
APInt IntVal(IntBitWidth, 2, x);
unsigned IntegerReg =
getRegForValue(ConstantInt::get(V->getContext(), IntVal));
if (IntegerReg != 0)
Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP,
IntegerReg, /*Kill=*/false);
}
}
} else if (const Operator *Op = dyn_cast<Operator>(V)) {
if (!SelectOperator(Op, Op->getOpcode()))
if (!isa<Instruction>(Op) ||
!TargetSelectInstruction(cast<Instruction>(Op)))
return 0;
Reg = lookUpRegForValue(Op);
} else if (isa<UndefValue>(V)) {
Reg = createResultReg(TLI.getRegClassFor(VT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(TargetOpcode::IMPLICIT_DEF), Reg);
}
// If target-independent code couldn't handle the value, give target-specific
// code a try.
if (!Reg && isa<Constant>(V))
Reg = TargetMaterializeConstant(cast<Constant>(V));
// Don't cache constant materializations in the general ValueMap.
// To do so would require tracking what uses they dominate.
if (Reg != 0) {
LocalValueMap[V] = Reg;
LastLocalValue = MRI.getVRegDef(Reg);
}
return Reg;
}
SDValue
X86SelectionDAGInfo::EmitTargetCodeForMemset(SelectionDAG &DAG, SDLoc dl,
SDValue Chain,
SDValue Dst, SDValue Src,
SDValue Size, unsigned Align,
bool isVolatile,
MachinePointerInfo DstPtrInfo) const {
ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
// If to a segment-relative address space, use the default lowering.
if (DstPtrInfo.getAddrSpace() >= 256)
return SDValue();
// If not DWORD aligned or size is more than the threshold, call the library.
// The libc version is likely to be faster for these cases. It can use the
// address value and run time information about the CPU.
if ((Align & 3) != 0 ||
!ConstantSize ||
ConstantSize->getZExtValue() >
Subtarget->getMaxInlineSizeThreshold()) {
// Check to see if there is a specialized entry-point for memory zeroing.
ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
if (const char *bzeroEntry = V &&
V->isNullValue() ? Subtarget->getBZeroEntry() : nullptr) {
EVT IntPtr = TLI.getPointerTy();
Type *IntPtrTy = getDataLayout()->getIntPtrType(*DAG.getContext());
TargetLowering::ArgListTy Args;
TargetLowering::ArgListEntry Entry;
Entry.Node = Dst;
Entry.Ty = IntPtrTy;
Args.push_back(Entry);
Entry.Node = Size;
Args.push_back(Entry);
TargetLowering::CallLoweringInfo CLI(DAG);
CLI.setDebugLoc(dl).setChain(Chain)
.setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
DAG.getExternalSymbol(bzeroEntry, IntPtr), &Args, 0)
.setDiscardResult();
std::pair<SDValue,SDValue> CallResult = TLI.LowerCallTo(CLI);
return CallResult.second;
}
// Otherwise have the target-independent code call memset.
return SDValue();
}
uint64_t SizeVal = ConstantSize->getZExtValue();
SDValue InFlag;
EVT AVT;
SDValue Count;
ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
unsigned BytesLeft = 0;
bool TwoRepStos = false;
if (ValC) {
unsigned ValReg;
uint64_t Val = ValC->getZExtValue() & 255;
// If the value is a constant, then we can potentially use larger sets.
switch (Align & 3) {
case 2: // WORD aligned
AVT = MVT::i16;
ValReg = X86::AX;
Val = (Val << 8) | Val;
break;
case 0: // DWORD aligned
AVT = MVT::i32;
ValReg = X86::EAX;
Val = (Val << 8) | Val;
Val = (Val << 16) | Val;
if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
AVT = MVT::i64;
ValReg = X86::RAX;
Val = (Val << 32) | Val;
}
break;
default: // Byte aligned
AVT = MVT::i8;
ValReg = X86::AL;
Count = DAG.getIntPtrConstant(SizeVal);
break;
}
if (AVT.bitsGT(MVT::i8)) {
unsigned UBytes = AVT.getSizeInBits() / 8;
Count = DAG.getIntPtrConstant(SizeVal / UBytes);
BytesLeft = SizeVal % UBytes;
}
Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
InFlag);
InFlag = Chain.getValue(1);
} else {
AVT = MVT::i8;
Count = DAG.getIntPtrConstant(SizeVal);
Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
InFlag = Chain.getValue(1);
}
Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
X86::ECX,
Count, InFlag);
InFlag = Chain.getValue(1);
Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
X86::EDI,
Dst, InFlag);
InFlag = Chain.getValue(1);
SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops);
if (TwoRepStos) {
InFlag = Chain.getValue(1);
Count = Size;
EVT CVT = Count.getValueType();
SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
X86::ECX,
Left, InFlag);
InFlag = Chain.getValue(1);
Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops);
} else if (BytesLeft) {
// Handle the last 1 - 7 bytes.
unsigned Offset = SizeVal - BytesLeft;
EVT AddrVT = Dst.getValueType();
EVT SizeVT = Size.getValueType();
Chain = DAG.getMemset(Chain, dl,
DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
DAG.getConstant(Offset, AddrVT)),
Src,
DAG.getConstant(BytesLeft, SizeVT),
Align, isVolatile, DstPtrInfo.getWithOffset(Offset));
}
// TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
return Chain;
}
SDValue Cpu0TargetLowering::
LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
{
// If the relocation model is PIC, use the General Dynamic TLS Model or
// Local Dynamic TLS model, otherwise use the Initial Exec or
// Local Exec TLS Model.
GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
DebugLoc dl = GA->getDebugLoc();
const GlobalValue *GV = GA->getGlobal();
EVT PtrVT = getPointerTy();
if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
// General Dynamic TLS Model
bool LocalDynamic = GV->hasInternalLinkage();
unsigned Flag = LocalDynamic ? Cpu0II::MO_TLSLDM :Cpu0II::MO_TLSGD;
SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, Flag);
SDValue Argument = DAG.getNode(Cpu0ISD::Wrapper, dl, PtrVT,
GetGlobalReg(DAG, PtrVT), TGA);
unsigned PtrSize = PtrVT.getSizeInBits();
IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
ArgListTy Args;
ArgListEntry Entry;
Entry.Node = Argument;
Entry.Ty = PtrTy;
Args.push_back(Entry);
std::pair<SDValue, SDValue> CallResult =
LowerCallTo(DAG.getEntryNode(), PtrTy,
false, false, false, false, 0, CallingConv::C,
/*isTailCall=*/false, /*doesNotRet=*/false,
/*isReturnValueUsed=*/true,
TlsGetAddr, Args, DAG, dl);
SDValue Ret = CallResult.first;
if (!LocalDynamic)
return Ret;
SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
Cpu0II::MO_DTPREL_HI);
SDValue Hi = DAG.getNode(Cpu0ISD::Hi, dl, PtrVT, TGAHi);
SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
Cpu0II::MO_DTPREL_LO);
SDValue Lo = DAG.getNode(Cpu0ISD::Lo, dl, PtrVT, TGALo);
SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Ret);
return DAG.getNode(ISD::ADD, dl, PtrVT, Add, Lo);
}
SDValue Offset;
if (GV->isDeclaration()) {
// Initial Exec TLS Model
SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
Cpu0II::MO_GOTTPREL);
TGA = DAG.getNode(Cpu0ISD::Wrapper, dl, PtrVT, GetGlobalReg(DAG, PtrVT),
TGA);
Offset = DAG.getLoad(PtrVT, dl,
DAG.getEntryNode(), TGA, MachinePointerInfo(),
false, false, false, 0);
} else {
// Local Exec TLS Model
SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
Cpu0II::MO_TPREL_HI);
SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
Cpu0II::MO_TPREL_LO);
SDValue Hi = DAG.getNode(Cpu0ISD::Hi, dl, PtrVT, TGAHi);
SDValue Lo = DAG.getNode(Cpu0ISD::Lo, dl, PtrVT, TGALo);
Offset = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
}
SDValue ThreadPointer = DAG.getNode(Cpu0ISD::ThreadPointer, dl, PtrVT);
return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
}
/// LowerFormalArguments - transform physical registers into virtual registers
/// and generate load operations for arguments places on the stack.
SDValue
Cpu0TargetLowering::LowerFormalArguments(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();
Cpu0FunctionInfo *Cpu0FI = MF.getInfo<Cpu0FunctionInfo>();
Cpu0FI->setVarArgsFrameIndex(0);
// Used with vargs to acumulate store chains.
std::vector<SDValue> OutChains;
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
getTargetMachine(), ArgLocs, *DAG.getContext());
CCInfo.AnalyzeFormalArguments(Ins, CC_Cpu0);
Function::const_arg_iterator FuncArg =
DAG.getMachineFunction().getFunction()->arg_begin();
int LastFI = 0;// Cpu0FI->LastInArgFI is 0 at the entry of this function.
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i, ++FuncArg) {
CCValAssign &VA = ArgLocs[i];
EVT ValVT = VA.getValVT();
ISD::ArgFlagsTy Flags = Ins[i].Flags;
bool IsRegLoc = VA.isRegLoc();
if (Flags.isByVal()) {
assert(Flags.getByValSize() &&
"ByVal args of size 0 should have been ignored by front-end.");
continue;
}
{ // VA.isRegLoc()
// sanity check
assert(VA.isMemLoc());
// The stack pointer offset is relative to the caller stack frame.
LastFI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
VA.getLocMemOffset(), true);
// Create load nodes to retrieve arguments from the stack
SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
InVals.push_back(DAG.getLoad(ValVT, dl, Chain, FIN,
MachinePointerInfo::getFixedStack(LastFI),
false, false, false, 0));
}
}
Cpu0FI->setLastInArgFI(LastFI);
#if 1
// All stores are grouped in one node to allow the matching between
// the size of Ins and InVals. This only happens when on varg functions
if (!OutChains.empty()) {
OutChains.push_back(Chain);
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&OutChains[0], OutChains.size());
}
#endif
return Chain;
}
/// ComputePHILiveOutRegInfo - Compute LiveOutInfo for a PHI's destination
/// register based on the LiveOutInfo of its operands.
void FunctionLoweringInfo::ComputePHILiveOutRegInfo(const PHINode *PN) {
Type *Ty = PN->getType();
if (!Ty->isIntegerTy() || Ty->isVectorTy())
return;
SmallVector<EVT, 1> ValueVTs;
ComputeValueVTs(*TLI, MF->getDataLayout(), Ty, ValueVTs);
assert(ValueVTs.size() == 1 &&
"PHIs with non-vector integer types should have a single VT.");
EVT IntVT = ValueVTs[0];
if (TLI->getNumRegisters(PN->getContext(), IntVT) != 1)
return;
IntVT = TLI->getTypeToTransformTo(PN->getContext(), IntVT);
unsigned BitWidth = IntVT.getSizeInBits();
unsigned DestReg = ValueMap[PN];
if (!TargetRegisterInfo::isVirtualRegister(DestReg))
return;
LiveOutRegInfo.grow(DestReg);
LiveOutInfo &DestLOI = LiveOutRegInfo[DestReg];
Value *V = PN->getIncomingValue(0);
if (isa<UndefValue>(V) || isa<ConstantExpr>(V)) {
DestLOI.NumSignBits = 1;
DestLOI.Known = KnownBits(BitWidth);
return;
}
if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
APInt Val = CI->getValue().zextOrTrunc(BitWidth);
DestLOI.NumSignBits = Val.getNumSignBits();
DestLOI.Known.Zero = ~Val;
DestLOI.Known.One = Val;
} else {
assert(ValueMap.count(V) && "V should have been placed in ValueMap when its"
"CopyToReg node was created.");
unsigned SrcReg = ValueMap[V];
if (!TargetRegisterInfo::isVirtualRegister(SrcReg)) {
DestLOI.IsValid = false;
return;
}
const LiveOutInfo *SrcLOI = GetLiveOutRegInfo(SrcReg, BitWidth);
if (!SrcLOI) {
DestLOI.IsValid = false;
return;
}
DestLOI = *SrcLOI;
}
assert(DestLOI.Known.Zero.getBitWidth() == BitWidth &&
DestLOI.Known.One.getBitWidth() == BitWidth &&
"Masks should have the same bit width as the type.");
for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
Value *V = PN->getIncomingValue(i);
if (isa<UndefValue>(V) || isa<ConstantExpr>(V)) {
DestLOI.NumSignBits = 1;
DestLOI.Known = KnownBits(BitWidth);
return;
}
if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
APInt Val = CI->getValue().zextOrTrunc(BitWidth);
DestLOI.NumSignBits = std::min(DestLOI.NumSignBits, Val.getNumSignBits());
DestLOI.Known.Zero &= ~Val;
DestLOI.Known.One &= Val;
continue;
}
assert(ValueMap.count(V) && "V should have been placed in ValueMap when "
"its CopyToReg node was created.");
unsigned SrcReg = ValueMap[V];
if (!TargetRegisterInfo::isVirtualRegister(SrcReg)) {
DestLOI.IsValid = false;
return;
}
const LiveOutInfo *SrcLOI = GetLiveOutRegInfo(SrcReg, BitWidth);
if (!SrcLOI) {
DestLOI.IsValid = false;
return;
}
DestLOI.NumSignBits = std::min(DestLOI.NumSignBits, SrcLOI->NumSignBits);
DestLOI.Known.Zero &= SrcLOI->Known.Zero;
DestLOI.Known.One &= SrcLOI->Known.One;
}
}
void DAGTypeLegalizer::ExpandRes_BITCAST(SDNode *N, SDValue &Lo, SDValue &Hi) {
EVT OutVT = N->getValueType(0);
EVT NOutVT = TLI.getTypeToTransformTo(*DAG.getContext(), OutVT);
SDValue InOp = N->getOperand(0);
EVT InVT = InOp.getValueType();
SDLoc dl(N);
// Handle some special cases efficiently.
switch (getTypeAction(InVT)) {
case TargetLowering::TypeLegal:
case TargetLowering::TypePromoteInteger:
break;
case TargetLowering::TypePromoteFloat:
llvm_unreachable("Bitcast of a promotion-needing float should never need"
"expansion");
case TargetLowering::TypeSoftenFloat:
// Convert the integer operand instead.
SplitInteger(GetSoftenedFloat(InOp), Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeExpandInteger:
case TargetLowering::TypeExpandFloat:
// Convert the expanded pieces of the input.
GetExpandedOp(InOp, Lo, Hi);
if (TLI.hasBigEndianPartOrdering(InVT) !=
TLI.hasBigEndianPartOrdering(OutVT))
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeSplitVector:
GetSplitVector(InOp, Lo, Hi);
if (TLI.hasBigEndianPartOrdering(OutVT))
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeScalarizeVector:
// Convert the element instead.
SplitInteger(BitConvertToInteger(GetScalarizedVector(InOp)), Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeWidenVector: {
assert(!(InVT.getVectorNumElements() & 1) && "Unsupported BITCAST");
InOp = GetWidenedVector(InOp);
EVT LoVT, HiVT;
std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(InVT);
std::tie(Lo, Hi) = DAG.SplitVector(InOp, dl, LoVT, HiVT);
if (TLI.hasBigEndianPartOrdering(OutVT))
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
}
}
if (InVT.isVector() && OutVT.isInteger()) {
// Handle cases like i64 = BITCAST v1i64 on x86, where the operand
// is legal but the result is not.
unsigned NumElems = 2;
EVT ElemVT = NOutVT;
EVT NVT = EVT::getVectorVT(*DAG.getContext(), ElemVT, NumElems);
// If <ElemVT * N> is not a legal type, try <ElemVT/2 * (N*2)>.
while (!isTypeLegal(NVT)) {
unsigned NewSizeInBits = ElemVT.getSizeInBits() / 2;
// If the element size is smaller than byte, bail.
if (NewSizeInBits < 8)
break;
NumElems *= 2;
ElemVT = EVT::getIntegerVT(*DAG.getContext(), NewSizeInBits);
NVT = EVT::getVectorVT(*DAG.getContext(), ElemVT, NumElems);
}
if (isTypeLegal(NVT)) {
SDValue CastInOp = DAG.getNode(ISD::BITCAST, dl, NVT, InOp);
SmallVector<SDValue, 8> Vals;
for (unsigned i = 0; i < NumElems; ++i)
Vals.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ElemVT,
CastInOp, DAG.getConstant(i, dl,
TLI.getVectorIdxTy())));
// Build Lo, Hi pair by pairing extracted elements if needed.
unsigned Slot = 0;
for (unsigned e = Vals.size(); e - Slot > 2; Slot += 2, e += 1) {
// Each iteration will BUILD_PAIR two nodes and append the result until
// there are only two nodes left, i.e. Lo and Hi.
SDValue LHS = Vals[Slot];
SDValue RHS = Vals[Slot + 1];
if (TLI.isBigEndian())
std::swap(LHS, RHS);
Vals.push_back(DAG.getNode(ISD::BUILD_PAIR, dl,
EVT::getIntegerVT(
*DAG.getContext(),
LHS.getValueType().getSizeInBits() << 1),
LHS, RHS));
}
Lo = Vals[Slot++];
Hi = Vals[Slot++];
if (TLI.isBigEndian())
std::swap(Lo, Hi);
return;
}
}
// Lower the bit-convert to a store/load from the stack.
assert(NOutVT.isByteSized() && "Expanded type not byte sized!");
// Create the stack frame object. Make sure it is aligned for both
// the source and expanded destination types.
unsigned Alignment =
TLI.getDataLayout()->getPrefTypeAlignment(NOutVT.
getTypeForEVT(*DAG.getContext()));
SDValue StackPtr = DAG.CreateStackTemporary(InVT, Alignment);
int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
// Emit a store to the stack slot.
SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, InOp, StackPtr, PtrInfo,
false, false, 0);
// Load the first half from the stack slot.
Lo = DAG.getLoad(NOutVT, dl, Store, StackPtr, PtrInfo,
false, false, false, 0);
// Increment the pointer to the other half.
unsigned IncrementSize = NOutVT.getSizeInBits() / 8;
StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
DAG.getConstant(IncrementSize, dl,
StackPtr.getValueType()));
// Load the second half from the stack slot.
Hi = DAG.getLoad(NOutVT, dl, Store, StackPtr,
PtrInfo.getWithOffset(IncrementSize), false,
false, false, MinAlign(Alignment, IncrementSize));
// Handle endianness of the load.
if (TLI.hasBigEndianPartOrdering(OutVT))
std::swap(Lo, Hi);
}
unsigned FastISel::getRegForValue(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-dominatess-use requirement enforced.
if (ValueMap.count(V))
return ValueMap[V];
unsigned Reg = LocalValueMap[V];
if (Reg != 0)
return Reg;
if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
if (CI->getValue().getActiveBits() <= 64)
Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
} else if (isa<AllocaInst>(V)) {
Reg = TargetMaterializeAlloca(cast<AllocaInst>(V));
} else if (isa<ConstantPointerNull>(V)) {
// Translate this as an integer zero so that it can be
// local-CSE'd with actual integer zeros.
Reg =
getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext())));
} else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
if (!Reg) {
const APFloat &Flt = CF->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) {
APInt IntVal(IntBitWidth, 2, x);
unsigned IntegerReg =
getRegForValue(ConstantInt::get(V->getContext(), IntVal));
if (IntegerReg != 0)
Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP, IntegerReg);
}
}
} else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
if (!SelectOperator(CE, CE->getOpcode())) return 0;
Reg = LocalValueMap[CE];
} else if (isa<UndefValue>(V)) {
Reg = createResultReg(TLI.getRegClassFor(VT));
BuildMI(MBB, DL, TII.get(TargetInstrInfo::IMPLICIT_DEF), Reg);
}
// If target-independent code couldn't handle the value, give target-specific
// code a try.
if (!Reg && isa<Constant>(V))
Reg = TargetMaterializeConstant(cast<Constant>(V));
// Don't cache constant materializations in the general ValueMap.
// To do so would require tracking what uses they dominate.
if (Reg != 0)
LocalValueMap[V] = Reg;
return Reg;
}
