namespace {
class VectorLegalizer {
SelectionDAG& DAG;
- TargetLowering& TLI;
+ const TargetLowering &TLI;
bool Changed; // Keep track of whether anything changed
/// LegalizedNodes - For nodes that are of legal width, and that have more
/// than one use, this map indicates what regularized operand to use. This
/// allows us to avoid legalizing the same thing more than once.
- DenseMap<SDValue, SDValue> LegalizedNodes;
+ SmallDenseMap<SDValue, SDValue, 64> LegalizedNodes;
// Adds a node to the translation cache
void AddLegalizedOperand(SDValue From, SDValue To) {
SDValue UnrollVSETCC(SDValue Op);
// Implements expansion for FNEG; falls back to UnrollVectorOp if FSUB
// isn't legal.
+ // Implements expansion for UINT_TO_FLOAT; falls back to UnrollVectorOp if
+ // SINT_TO_FLOAT and SHR on vectors isn't legal.
+ SDValue ExpandUINT_TO_FLOAT(SDValue Op);
+ // Implement expansion for SIGN_EXTEND_INREG using SRL and SRA.
+ SDValue ExpandSEXTINREG(SDValue Op);
+ // Implement vselect in terms of XOR, AND, OR when blend is not supported
+ // by the target.
+ SDValue ExpandVSELECT(SDValue Op);
+ SDValue ExpandSELECT(SDValue Op);
+ SDValue ExpandLoad(SDValue Op);
+ SDValue ExpandStore(SDValue Op);
SDValue ExpandFNEG(SDValue Op);
// Implements vector promotion; this is essentially just bitcasting the
// operands to a different type and bitcasting the result back to the
// original type.
SDValue PromoteVectorOp(SDValue Op);
+ // Implements [SU]INT_TO_FP vector promotion; this is a [zs]ext of the input
+ // operand to the next size up.
+ SDValue PromoteVectorOpINT_TO_FP(SDValue Op);
+ // Implements FP_TO_[SU]INT vector promotion of the result type; it is
+ // promoted to the next size up integer type. The result is then truncated
+ // back to the original type.
+ SDValue PromoteVectorOpFP_TO_INT(SDValue Op, bool isSigned);
public:
bool Run();
};
bool VectorLegalizer::Run() {
+ // Before we start legalizing vector nodes, check if there are any vectors.
+ bool HasVectors = false;
+ for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
+ E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I) {
+ // Check if the values of the nodes contain vectors. We don't need to check
+ // the operands because we are going to check their values at some point.
+ for (SDNode::value_iterator J = I->value_begin(), E = I->value_end();
+ J != E; ++J)
+ HasVectors |= J->isVector();
+
+ // If we found a vector node we can start the legalization.
+ if (HasVectors)
+ break;
+ }
+
+ // If this basic block has no vectors then no need to legalize vectors.
+ if (!HasVectors)
+ return false;
+
// The legalize process is inherently a bottom-up recursive process (users
// legalize their uses before themselves). Given infinite stack space, we
// could just start legalizing on the root and traverse the whole graph. In
// node is only legalized after all of its operands are legalized.
DAG.AssignTopologicalOrder();
for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
- E = prior(DAG.allnodes_end()); I != llvm::next(E); ++I)
+ E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I)
LegalizeOp(SDValue(I, 0));
// Finally, it's possible the root changed. Get the new root.
Ops.push_back(LegalizeOp(Node->getOperand(i)));
SDValue Result =
- DAG.UpdateNodeOperands(Op.getValue(0), Ops.data(), Ops.size());
+ SDValue(DAG.UpdateNodeOperands(Op.getNode(), Ops.data(), Ops.size()), 0);
+
+ if (Op.getOpcode() == ISD::LOAD) {
+ LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
+ ISD::LoadExtType ExtType = LD->getExtensionType();
+ if (LD->getMemoryVT().isVector() && ExtType != ISD::NON_EXTLOAD) {
+ if (TLI.isLoadExtLegal(LD->getExtensionType(), LD->getMemoryVT()))
+ return TranslateLegalizeResults(Op, Result);
+ Changed = true;
+ return LegalizeOp(ExpandLoad(Op));
+ }
+ } else if (Op.getOpcode() == ISD::STORE) {
+ StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
+ EVT StVT = ST->getMemoryVT();
+ MVT ValVT = ST->getValue().getSimpleValueType();
+ if (StVT.isVector() && ST->isTruncatingStore())
+ switch (TLI.getTruncStoreAction(ValVT, StVT.getSimpleVT())) {
+ default: llvm_unreachable("This action is not supported yet!");
+ case TargetLowering::Legal:
+ return TranslateLegalizeResults(Op, Result);
+ case TargetLowering::Custom:
+ Changed = true;
+ return TranslateLegalizeResults(Op, TLI.LowerOperation(Result, DAG));
+ case TargetLowering::Expand:
+ Changed = true;
+ return LegalizeOp(ExpandStore(Op));
+ }
+ }
bool HasVectorValue = false;
for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end();
case ISD::SRL:
case ISD::ROTL:
case ISD::ROTR:
- case ISD::CTTZ:
+ case ISD::BSWAP:
case ISD::CTLZ:
+ case ISD::CTTZ:
+ case ISD::CTLZ_ZERO_UNDEF:
+ case ISD::CTTZ_ZERO_UNDEF:
case ISD::CTPOP:
case ISD::SELECT:
+ case ISD::VSELECT:
case ISD::SELECT_CC:
- case ISD::VSETCC:
+ case ISD::SETCC:
case ISD::ZERO_EXTEND:
case ISD::ANY_EXTEND:
case ISD::TRUNCATE:
case ISD::FP_TO_UINT:
case ISD::FNEG:
case ISD::FABS:
+ case ISD::FCOPYSIGN:
case ISD::FSQRT:
case ISD::FSIN:
case ISD::FCOS:
case ISD::FTRUNC:
case ISD::FRINT:
case ISD::FNEARBYINT:
+ case ISD::FROUND:
case ISD::FFLOOR:
+ case ISD::FP_ROUND:
+ case ISD::FP_EXTEND:
+ case ISD::FMA:
+ case ISD::SIGN_EXTEND_INREG:
QueryType = Node->getValueType(0);
break;
+ case ISD::FP_ROUND_INREG:
+ QueryType = cast<VTSDNode>(Node->getOperand(1))->getVT();
+ break;
case ISD::SINT_TO_FP:
case ISD::UINT_TO_FP:
QueryType = Node->getOperand(0).getValueType();
switch (TLI.getOperationAction(Node->getOpcode(), QueryType)) {
case TargetLowering::Promote:
- // "Promote" the operation by bitcasting
- Result = PromoteVectorOp(Op);
- Changed = true;
+ switch (Op.getOpcode()) {
+ default:
+ // "Promote" the operation by bitcasting
+ Result = PromoteVectorOp(Op);
+ Changed = true;
+ break;
+ case ISD::SINT_TO_FP:
+ case ISD::UINT_TO_FP:
+ // "Promote" the operation by extending the operand.
+ Result = PromoteVectorOpINT_TO_FP(Op);
+ Changed = true;
+ break;
+ case ISD::FP_TO_UINT:
+ case ISD::FP_TO_SINT:
+ // Promote the operation by extending the operand.
+ Result = PromoteVectorOpFP_TO_INT(Op, Op->getOpcode() == ISD::FP_TO_SINT);
+ Changed = true;
+ break;
+ }
break;
case TargetLowering::Legal: break;
case TargetLowering::Custom: {
// FALL THROUGH
}
case TargetLowering::Expand:
- if (Node->getOpcode() == ISD::FNEG)
+ if (Node->getOpcode() == ISD::SIGN_EXTEND_INREG)
+ Result = ExpandSEXTINREG(Op);
+ else if (Node->getOpcode() == ISD::VSELECT)
+ Result = ExpandVSELECT(Op);
+ else if (Node->getOpcode() == ISD::SELECT)
+ Result = ExpandSELECT(Op);
+ else if (Node->getOpcode() == ISD::UINT_TO_FP)
+ Result = ExpandUINT_TO_FLOAT(Op);
+ else if (Node->getOpcode() == ISD::FNEG)
Result = ExpandFNEG(Op);
- else if (Node->getOpcode() == ISD::VSETCC)
+ else if (Node->getOpcode() == ISD::SETCC)
Result = UnrollVSETCC(Op);
else
Result = DAG.UnrollVectorOp(Op.getNode());
// Vector "promotion" is basically just bitcasting and doing the operation
// in a different type. For example, x86 promotes ISD::AND on v2i32 to
// v1i64.
- EVT VT = Op.getValueType();
+ MVT VT = Op.getSimpleValueType();
assert(Op.getNode()->getNumValues() == 1 &&
"Can't promote a vector with multiple results!");
- EVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
- DebugLoc dl = Op.getDebugLoc();
+ MVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
+ SDLoc dl(Op);
SmallVector<SDValue, 4> Operands(Op.getNumOperands());
for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
if (Op.getOperand(j).getValueType().isVector())
- Operands[j] = DAG.getNode(ISD::BIT_CONVERT, dl, NVT, Op.getOperand(j));
+ Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT, Op.getOperand(j));
else
Operands[j] = Op.getOperand(j);
}
Op = DAG.getNode(Op.getOpcode(), dl, NVT, &Operands[0], Operands.size());
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Op);
+ return DAG.getNode(ISD::BITCAST, dl, VT, Op);
+}
+
+SDValue VectorLegalizer::PromoteVectorOpINT_TO_FP(SDValue Op) {
+ // INT_TO_FP operations may require the input operand be promoted even
+ // when the type is otherwise legal.
+ EVT VT = Op.getOperand(0).getValueType();
+ assert(Op.getNode()->getNumValues() == 1 &&
+ "Can't promote a vector with multiple results!");
+
+ // Normal getTypeToPromoteTo() doesn't work here, as that will promote
+ // by widening the vector w/ the same element width and twice the number
+ // of elements. We want the other way around, the same number of elements,
+ // each twice the width.
+ //
+ // Increase the bitwidth of the element to the next pow-of-two
+ // (which is greater than 8 bits).
+
+ EVT NVT = VT.widenIntegerVectorElementType(*DAG.getContext());
+ assert(NVT.isSimple() && "Promoting to a non-simple vector type!");
+ SDLoc dl(Op);
+ SmallVector<SDValue, 4> Operands(Op.getNumOperands());
+
+ unsigned Opc = Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND :
+ ISD::SIGN_EXTEND;
+ for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
+ if (Op.getOperand(j).getValueType().isVector())
+ Operands[j] = DAG.getNode(Opc, dl, NVT, Op.getOperand(j));
+ else
+ Operands[j] = Op.getOperand(j);
+ }
+
+ return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), &Operands[0],
+ Operands.size());
}
+// For FP_TO_INT we promote the result type to a vector type with wider
+// elements and then truncate the result. This is different from the default
+// PromoteVector which uses bitcast to promote thus assumning that the
+// promoted vector type has the same overall size.
+SDValue VectorLegalizer::PromoteVectorOpFP_TO_INT(SDValue Op, bool isSigned) {
+ assert(Op.getNode()->getNumValues() == 1 &&
+ "Can't promote a vector with multiple results!");
+ EVT VT = Op.getValueType();
+
+ EVT NewVT;
+ unsigned NewOpc;
+ while (1) {
+ NewVT = VT.widenIntegerVectorElementType(*DAG.getContext());
+ assert(NewVT.isSimple() && "Promoting to a non-simple vector type!");
+ if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewVT)) {
+ NewOpc = ISD::FP_TO_SINT;
+ break;
+ }
+ if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewVT)) {
+ NewOpc = ISD::FP_TO_UINT;
+ break;
+ }
+ }
+
+ SDLoc loc(Op);
+ SDValue promoted = DAG.getNode(NewOpc, SDLoc(Op), NewVT, Op.getOperand(0));
+ return DAG.getNode(ISD::TRUNCATE, SDLoc(Op), VT, promoted);
+}
+
+
+SDValue VectorLegalizer::ExpandLoad(SDValue Op) {
+ SDLoc dl(Op);
+ LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
+ SDValue Chain = LD->getChain();
+ SDValue BasePTR = LD->getBasePtr();
+ EVT SrcVT = LD->getMemoryVT();
+ ISD::LoadExtType ExtType = LD->getExtensionType();
+
+ SmallVector<SDValue, 8> Vals;
+ SmallVector<SDValue, 8> LoadChains;
+ unsigned NumElem = SrcVT.getVectorNumElements();
+
+ EVT SrcEltVT = SrcVT.getScalarType();
+ EVT DstEltVT = Op.getNode()->getValueType(0).getScalarType();
+
+ if (SrcVT.getVectorNumElements() > 1 && !SrcEltVT.isByteSized()) {
+ // When elements in a vector is not byte-addressable, we cannot directly
+ // load each element by advancing pointer, which could only address bytes.
+ // Instead, we load all significant words, mask bits off, and concatenate
+ // them to form each element. Finally, they are extended to destination
+ // scalar type to build the destination vector.
+ EVT WideVT = TLI.getPointerTy();
+
+ assert(WideVT.isRound() &&
+ "Could not handle the sophisticated case when the widest integer is"
+ " not power of 2.");
+ assert(WideVT.bitsGE(SrcEltVT) &&
+ "Type is not legalized?");
+
+ unsigned WideBytes = WideVT.getStoreSize();
+ unsigned Offset = 0;
+ unsigned RemainingBytes = SrcVT.getStoreSize();
+ SmallVector<SDValue, 8> LoadVals;
+
+ while (RemainingBytes > 0) {
+ SDValue ScalarLoad;
+ unsigned LoadBytes = WideBytes;
+
+ if (RemainingBytes >= LoadBytes) {
+ ScalarLoad = DAG.getLoad(WideVT, dl, Chain, BasePTR,
+ LD->getPointerInfo().getWithOffset(Offset),
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->isInvariant(), LD->getAlignment(),
+ LD->getTBAAInfo());
+ } else {
+ EVT LoadVT = WideVT;
+ while (RemainingBytes < LoadBytes) {
+ LoadBytes >>= 1; // Reduce the load size by half.
+ LoadVT = EVT::getIntegerVT(*DAG.getContext(), LoadBytes << 3);
+ }
+ ScalarLoad = DAG.getExtLoad(ISD::EXTLOAD, dl, WideVT, Chain, BasePTR,
+ LD->getPointerInfo().getWithOffset(Offset),
+ LoadVT, LD->isVolatile(),
+ LD->isNonTemporal(), LD->getAlignment(),
+ LD->getTBAAInfo());
+ }
+
+ RemainingBytes -= LoadBytes;
+ Offset += LoadBytes;
+ BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
+ DAG.getConstant(LoadBytes, BasePTR.getValueType()));
+
+ LoadVals.push_back(ScalarLoad.getValue(0));
+ LoadChains.push_back(ScalarLoad.getValue(1));
+ }
+
+ // Extract bits, pack and extend/trunc them into destination type.
+ unsigned SrcEltBits = SrcEltVT.getSizeInBits();
+ SDValue SrcEltBitMask = DAG.getConstant((1U << SrcEltBits) - 1, WideVT);
+
+ unsigned BitOffset = 0;
+ unsigned WideIdx = 0;
+ unsigned WideBits = WideVT.getSizeInBits();
+
+ for (unsigned Idx = 0; Idx != NumElem; ++Idx) {
+ SDValue Lo, Hi, ShAmt;
+
+ if (BitOffset < WideBits) {
+ ShAmt = DAG.getConstant(BitOffset, TLI.getShiftAmountTy(WideVT));
+ Lo = DAG.getNode(ISD::SRL, dl, WideVT, LoadVals[WideIdx], ShAmt);
+ Lo = DAG.getNode(ISD::AND, dl, WideVT, Lo, SrcEltBitMask);
+ }
+
+ BitOffset += SrcEltBits;
+ if (BitOffset >= WideBits) {
+ WideIdx++;
+ Offset -= WideBits;
+ if (Offset > 0) {
+ ShAmt = DAG.getConstant(SrcEltBits - Offset,
+ TLI.getShiftAmountTy(WideVT));
+ Hi = DAG.getNode(ISD::SHL, dl, WideVT, LoadVals[WideIdx], ShAmt);
+ Hi = DAG.getNode(ISD::AND, dl, WideVT, Hi, SrcEltBitMask);
+ }
+ }
+
+ if (Hi.getNode())
+ Lo = DAG.getNode(ISD::OR, dl, WideVT, Lo, Hi);
+
+ switch (ExtType) {
+ default: llvm_unreachable("Unknown extended-load op!");
+ case ISD::EXTLOAD:
+ Lo = DAG.getAnyExtOrTrunc(Lo, dl, DstEltVT);
+ break;
+ case ISD::ZEXTLOAD:
+ Lo = DAG.getZExtOrTrunc(Lo, dl, DstEltVT);
+ break;
+ case ISD::SEXTLOAD:
+ ShAmt = DAG.getConstant(WideBits - SrcEltBits,
+ TLI.getShiftAmountTy(WideVT));
+ Lo = DAG.getNode(ISD::SHL, dl, WideVT, Lo, ShAmt);
+ Lo = DAG.getNode(ISD::SRA, dl, WideVT, Lo, ShAmt);
+ Lo = DAG.getSExtOrTrunc(Lo, dl, DstEltVT);
+ break;
+ }
+ Vals.push_back(Lo);
+ }
+ } else {
+ unsigned Stride = SrcVT.getScalarType().getSizeInBits()/8;
+
+ for (unsigned Idx=0; Idx<NumElem; Idx++) {
+ SDValue ScalarLoad = DAG.getExtLoad(ExtType, dl,
+ Op.getNode()->getValueType(0).getScalarType(),
+ Chain, BasePTR, LD->getPointerInfo().getWithOffset(Idx * Stride),
+ SrcVT.getScalarType(),
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->getAlignment(), LD->getTBAAInfo());
+
+ BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
+ DAG.getConstant(Stride, BasePTR.getValueType()));
+
+ Vals.push_back(ScalarLoad.getValue(0));
+ LoadChains.push_back(ScalarLoad.getValue(1));
+ }
+ }
+
+ SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &LoadChains[0], LoadChains.size());
+ SDValue Value = DAG.getNode(ISD::BUILD_VECTOR, dl,
+ Op.getNode()->getValueType(0), &Vals[0], Vals.size());
+
+ AddLegalizedOperand(Op.getValue(0), Value);
+ AddLegalizedOperand(Op.getValue(1), NewChain);
+
+ return (Op.getResNo() ? NewChain : Value);
+}
+
+SDValue VectorLegalizer::ExpandStore(SDValue Op) {
+ SDLoc dl(Op);
+ StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
+ SDValue Chain = ST->getChain();
+ SDValue BasePTR = ST->getBasePtr();
+ SDValue Value = ST->getValue();
+ EVT StVT = ST->getMemoryVT();
+
+ unsigned Alignment = ST->getAlignment();
+ bool isVolatile = ST->isVolatile();
+ bool isNonTemporal = ST->isNonTemporal();
+ const MDNode *TBAAInfo = ST->getTBAAInfo();
+
+ unsigned NumElem = StVT.getVectorNumElements();
+ // The type of the data we want to save
+ EVT RegVT = Value.getValueType();
+ EVT RegSclVT = RegVT.getScalarType();
+ // The type of data as saved in memory.
+ EVT MemSclVT = StVT.getScalarType();
+
+ // Cast floats into integers
+ unsigned ScalarSize = MemSclVT.getSizeInBits();
+
+ // Round odd types to the next pow of two.
+ if (!isPowerOf2_32(ScalarSize))
+ ScalarSize = NextPowerOf2(ScalarSize);
+
+ // Store Stride in bytes
+ unsigned Stride = ScalarSize/8;
+ // Extract each of the elements from the original vector
+ // and save them into memory individually.
+ SmallVector<SDValue, 8> Stores;
+ for (unsigned Idx = 0; Idx < NumElem; Idx++) {
+ SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
+ RegSclVT, Value, DAG.getConstant(Idx, TLI.getVectorIdxTy()));
+
+ // This scalar TruncStore may be illegal, but we legalize it later.
+ SDValue Store = DAG.getTruncStore(Chain, dl, Ex, BasePTR,
+ ST->getPointerInfo().getWithOffset(Idx*Stride), MemSclVT,
+ isVolatile, isNonTemporal, Alignment, TBAAInfo);
+
+ BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
+ DAG.getConstant(Stride, BasePTR.getValueType()));
+
+ Stores.push_back(Store);
+ }
+ SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
+ &Stores[0], Stores.size());
+ AddLegalizedOperand(Op, TF);
+ return TF;
+}
+
+SDValue VectorLegalizer::ExpandSELECT(SDValue Op) {
+ // Lower a select instruction where the condition is a scalar and the
+ // operands are vectors. Lower this select to VSELECT and implement it
+ // using XOR AND OR. The selector bit is broadcasted.
+ EVT VT = Op.getValueType();
+ SDLoc DL(Op);
+
+ SDValue Mask = Op.getOperand(0);
+ SDValue Op1 = Op.getOperand(1);
+ SDValue Op2 = Op.getOperand(2);
+
+ assert(VT.isVector() && !Mask.getValueType().isVector()
+ && Op1.getValueType() == Op2.getValueType() && "Invalid type");
+
+ unsigned NumElem = VT.getVectorNumElements();
+
+ // If we can't even use the basic vector operations of
+ // AND,OR,XOR, we will have to scalarize the op.
+ // Notice that the operation may be 'promoted' which means that it is
+ // 'bitcasted' to another type which is handled.
+ // Also, we need to be able to construct a splat vector using BUILD_VECTOR.
+ if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
+ TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
+ TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
+ TLI.getOperationAction(ISD::BUILD_VECTOR, VT) == TargetLowering::Expand)
+ return DAG.UnrollVectorOp(Op.getNode());
+
+ // Generate a mask operand.
+ EVT MaskTy = VT.changeVectorElementTypeToInteger();
+
+ // What is the size of each element in the vector mask.
+ EVT BitTy = MaskTy.getScalarType();
+
+ Mask = DAG.getSelect(DL, BitTy, Mask,
+ DAG.getConstant(APInt::getAllOnesValue(BitTy.getSizeInBits()), BitTy),
+ DAG.getConstant(0, BitTy));
+
+ // Broadcast the mask so that the entire vector is all-one or all zero.
+ SmallVector<SDValue, 8> Ops(NumElem, Mask);
+ Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskTy, &Ops[0], Ops.size());
+
+ // Bitcast the operands to be the same type as the mask.
+ // This is needed when we select between FP types because
+ // the mask is a vector of integers.
+ Op1 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op1);
+ Op2 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op2);
+
+ SDValue AllOnes = DAG.getConstant(
+ APInt::getAllOnesValue(BitTy.getSizeInBits()), MaskTy);
+ SDValue NotMask = DAG.getNode(ISD::XOR, DL, MaskTy, Mask, AllOnes);
+
+ Op1 = DAG.getNode(ISD::AND, DL, MaskTy, Op1, Mask);
+ Op2 = DAG.getNode(ISD::AND, DL, MaskTy, Op2, NotMask);
+ SDValue Val = DAG.getNode(ISD::OR, DL, MaskTy, Op1, Op2);
+ return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
+}
+
+SDValue VectorLegalizer::ExpandSEXTINREG(SDValue Op) {
+ EVT VT = Op.getValueType();
+
+ // Make sure that the SRA and SHL instructions are available.
+ if (TLI.getOperationAction(ISD::SRA, VT) == TargetLowering::Expand ||
+ TLI.getOperationAction(ISD::SHL, VT) == TargetLowering::Expand)
+ return DAG.UnrollVectorOp(Op.getNode());
+
+ SDLoc DL(Op);
+ EVT OrigTy = cast<VTSDNode>(Op->getOperand(1))->getVT();
+
+ unsigned BW = VT.getScalarType().getSizeInBits();
+ unsigned OrigBW = OrigTy.getScalarType().getSizeInBits();
+ SDValue ShiftSz = DAG.getConstant(BW - OrigBW, VT);
+
+ Op = Op.getOperand(0);
+ Op = DAG.getNode(ISD::SHL, DL, VT, Op, ShiftSz);
+ return DAG.getNode(ISD::SRA, DL, VT, Op, ShiftSz);
+}
+
+SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) {
+ // Implement VSELECT in terms of XOR, AND, OR
+ // on platforms which do not support blend natively.
+ SDLoc DL(Op);
+
+ SDValue Mask = Op.getOperand(0);
+ SDValue Op1 = Op.getOperand(1);
+ SDValue Op2 = Op.getOperand(2);
+
+ EVT VT = Mask.getValueType();
+
+ // If we can't even use the basic vector operations of
+ // AND,OR,XOR, we will have to scalarize the op.
+ // Notice that the operation may be 'promoted' which means that it is
+ // 'bitcasted' to another type which is handled.
+ // This operation also isn't safe with AND, OR, XOR when the boolean
+ // type is 0/1 as we need an all ones vector constant to mask with.
+ // FIXME: Sign extend 1 to all ones if thats legal on the target.
+ if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
+ TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
+ TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
+ TLI.getBooleanContents(true) !=
+ TargetLowering::ZeroOrNegativeOneBooleanContent)
+ return DAG.UnrollVectorOp(Op.getNode());
+
+ // If the mask and the type are different sizes, unroll the vector op. This
+ // can occur when getSetCCResultType returns something that is different in
+ // size from the operand types. For example, v4i8 = select v4i32, v4i8, v4i8.
+ if (VT.getSizeInBits() != Op1.getValueType().getSizeInBits())
+ return DAG.UnrollVectorOp(Op.getNode());
+
+ // Bitcast the operands to be the same type as the mask.
+ // This is needed when we select between FP types because
+ // the mask is a vector of integers.
+ Op1 = DAG.getNode(ISD::BITCAST, DL, VT, Op1);
+ Op2 = DAG.getNode(ISD::BITCAST, DL, VT, Op2);
+
+ SDValue AllOnes = DAG.getConstant(
+ APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()), VT);
+ SDValue NotMask = DAG.getNode(ISD::XOR, DL, VT, Mask, AllOnes);
+
+ Op1 = DAG.getNode(ISD::AND, DL, VT, Op1, Mask);
+ Op2 = DAG.getNode(ISD::AND, DL, VT, Op2, NotMask);
+ SDValue Val = DAG.getNode(ISD::OR, DL, VT, Op1, Op2);
+ return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
+}
+
+SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) {
+ EVT VT = Op.getOperand(0).getValueType();
+ SDLoc DL(Op);
+
+ // Make sure that the SINT_TO_FP and SRL instructions are available.
+ if (TLI.getOperationAction(ISD::SINT_TO_FP, VT) == TargetLowering::Expand ||
+ TLI.getOperationAction(ISD::SRL, VT) == TargetLowering::Expand)
+ return DAG.UnrollVectorOp(Op.getNode());
+
+ EVT SVT = VT.getScalarType();
+ assert((SVT.getSizeInBits() == 64 || SVT.getSizeInBits() == 32) &&
+ "Elements in vector-UINT_TO_FP must be 32 or 64 bits wide");
+
+ unsigned BW = SVT.getSizeInBits();
+ SDValue HalfWord = DAG.getConstant(BW/2, VT);
+
+ // Constants to clear the upper part of the word.
+ // Notice that we can also use SHL+SHR, but using a constant is slightly
+ // faster on x86.
+ uint64_t HWMask = (SVT.getSizeInBits()==64)?0x00000000FFFFFFFF:0x0000FFFF;
+ SDValue HalfWordMask = DAG.getConstant(HWMask, VT);
+
+ // Two to the power of half-word-size.
+ SDValue TWOHW = DAG.getConstantFP((1<<(BW/2)), Op.getValueType());
+
+ // Clear upper part of LO, lower HI
+ SDValue HI = DAG.getNode(ISD::SRL, DL, VT, Op.getOperand(0), HalfWord);
+ SDValue LO = DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), HalfWordMask);
+
+ // Convert hi and lo to floats
+ // Convert the hi part back to the upper values
+ SDValue fHI = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), HI);
+ fHI = DAG.getNode(ISD::FMUL, DL, Op.getValueType(), fHI, TWOHW);
+ SDValue fLO = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), LO);
+
+ // Add the two halves
+ return DAG.getNode(ISD::FADD, DL, Op.getValueType(), fHI, fLO);
+}
+
+
SDValue VectorLegalizer::ExpandFNEG(SDValue Op) {
if (TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) {
SDValue Zero = DAG.getConstantFP(-0.0, Op.getValueType());
- return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
+ return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
Zero, Op.getOperand(0));
}
return DAG.UnrollVectorOp(Op.getNode());
EVT EltVT = VT.getVectorElementType();
SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1), CC = Op.getOperand(2);
EVT TmpEltVT = LHS.getValueType().getVectorElementType();
- DebugLoc dl = Op.getDebugLoc();
+ SDLoc dl(Op);
SmallVector<SDValue, 8> Ops(NumElems);
for (unsigned i = 0; i < NumElems; ++i) {
SDValue LHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS,
- DAG.getIntPtrConstant(i));
+ DAG.getConstant(i, TLI.getVectorIdxTy()));
SDValue RHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS,
- DAG.getIntPtrConstant(i));
- Ops[i] = DAG.getNode(ISD::SETCC, dl, TLI.getSetCCResultType(TmpEltVT),
+ DAG.getConstant(i, TLI.getVectorIdxTy()));
+ Ops[i] = DAG.getNode(ISD::SETCC, dl,
+ TLI.getSetCCResultType(*DAG.getContext(), TmpEltVT),
LHSElem, RHSElem, CC);
- Ops[i] = DAG.getNode(ISD::SELECT, dl, EltVT, Ops[i],
- DAG.getConstant(APInt::getAllOnesValue
- (EltVT.getSizeInBits()), EltVT),
- DAG.getConstant(0, EltVT));
+ Ops[i] = DAG.getSelect(dl, EltVT, Ops[i],
+ DAG.getConstant(APInt::getAllOnesValue
+ (EltVT.getSizeInBits()), EltVT),
+ DAG.getConstant(0, EltVT));
}
return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], NumElems);
}