X86ISelLowering.cpp

    SDValue Ptr = LD->getBasePtr();
    if (!ISD::isNormalLoad(LD) || LD->isVolatile())
      return SDValue();
    EVT PVT = LD->getValueType(0);
    if (PVT != MVT::i32 && PVT != MVT::f32)
      return SDValue();

    int FI = -1;
    int64_t Offset = 0;
    if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
      FI = FINode->getIndex();
      Offset = 0;
    } else if (Ptr.getOpcode() == ISD::ADD &&
               isa<ConstantSDNode>(Ptr.getOperand(1)) &&
               isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
      FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
      Offset = Ptr.getConstantOperandVal(1);
      Ptr = Ptr.getOperand(0);
    } else {
      return SDValue();
    }

    SDValue Chain = LD->getChain();
    // Make sure the stack object alignment is at least 16.
    MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
    if (DAG.InferPtrAlignment(Ptr) < 16) {
      if (MFI->isFixedObjectIndex(FI)) {
        // Can't change the alignment. FIXME: It's possible to compute
        // the exact stack offset and reference FI + adjust offset instead.
        // If someone *really* cares about this. That's the way to implement it.
        return SDValue();
      } else {
        MFI->setObjectAlignment(FI, 16);
      }
    }

    // (Offset % 16) must be multiple of 4. Then address is then
    // Ptr + (Offset & ~15).
    if (Offset < 0)
      return SDValue();
    if ((Offset % 16) & 3)
      return SDValue();
    int64_t StartOffset = Offset & ~15;
    if (StartOffset)
      Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
                        Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));

    int EltNo = (Offset - StartOffset) >> 2;
    int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
    EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
    SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,
                             LD->getPointerInfo().getWithOffset(StartOffset),
                             false, false, 0);
    // Canonicalize it to a v4i32 shuffle.
    V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
    return DAG.getNode(ISD::BITCAST, dl, VT,
                       DAG.getVectorShuffle(MVT::v4i32, dl, V1,
                                            DAG.getUNDEF(MVT::v4i32),&Mask[0]));
  }

  return SDValue();
}

/// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
/// vector of type 'VT', see if the elements can be replaced by a single large
/// load which has the same value as a build_vector whose operands are 'elts'.
///
/// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
///
/// FIXME: we'd also like to handle the case where the last elements are zero
/// rather than undef via VZEXT_LOAD, but we do not detect that case today.
/// There's even a handy isZeroNode for that purpose.
static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
                                        DebugLoc &DL, SelectionDAG &DAG) {
  EVT EltVT = VT.getVectorElementType();
  unsigned NumElems = Elts.size();

  LoadSDNode *LDBase = NULL;
  unsigned LastLoadedElt = -1U;

  // For each element in the initializer, see if we've found a load or an undef.
  // If we don't find an initial load element, or later load elements are
  // non-consecutive, bail out.
  for (unsigned i = 0; i < NumElems; ++i) {
    SDValue Elt = Elts[i];

    if (!Elt.getNode() ||
        (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
      return SDValue();
    if (!LDBase) {
      if (Elt.getNode()->getOpcode() == ISD::UNDEF)
        return SDValue();
      LDBase = cast<LoadSDNode>(Elt.getNode());
      LastLoadedElt = i;
      continue;
    }
    if (Elt.getOpcode() == ISD::UNDEF)
      continue;

    LoadSDNode *LD = cast<LoadSDNode>(Elt);
    if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
      return SDValue();
    LastLoadedElt = i;
  }

  // If we have found an entire vector of loads and undefs, then return a large
  // load of the entire vector width starting at the base pointer.  If we found
  // consecutive loads for the low half, generate a vzext_load node.
  if (LastLoadedElt == NumElems - 1) {
    if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
      return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
                         LDBase->getPointerInfo(),
                         LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
    return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
                       LDBase->getPointerInfo(),
                       LDBase->isVolatile(), LDBase->isNonTemporal(),
                       LDBase->getAlignment());
  } else if (NumElems == 4 && LastLoadedElt == 1) {
    SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
    SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
    SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
                                              Ops, 2, MVT::i32,
                                              LDBase->getMemOperand());
    return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
  }
  return SDValue();
}

SDValue
X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
  DebugLoc dl = Op.getDebugLoc();
  // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
  // All one's are handled with pcmpeqd. In AVX, zero's are handled with
  // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
  // is present, so AllOnes is ignored.
  if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
      (Op.getValueType().getSizeInBits() != 256 &&
       ISD::isBuildVectorAllOnes(Op.getNode()))) {
    // Canonicalize this to <4 x i32> (SSE) to
    // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
    // eliminated on x86-32 hosts.
    if (Op.getValueType() == MVT::v4i32)
      return Op;

    if (ISD::isBuildVectorAllOnes(Op.getNode()))
      return getOnesVector(Op.getValueType(), DAG, dl);
    return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
  }

  EVT VT = Op.getValueType();
  EVT ExtVT = VT.getVectorElementType();
  unsigned EVTBits = ExtVT.getSizeInBits();

  unsigned NumElems = Op.getNumOperands();
  unsigned NumZero  = 0;
  unsigned NumNonZero = 0;
  unsigned NonZeros = 0;
  bool IsAllConstants = true;
  SmallSet<SDValue, 8> Values;
  for (unsigned i = 0; i < NumElems; ++i) {
    SDValue Elt = Op.getOperand(i);
    if (Elt.getOpcode() == ISD::UNDEF)
      continue;
    Values.insert(Elt);
    if (Elt.getOpcode() != ISD::Constant &&
        Elt.getOpcode() != ISD::ConstantFP)
      IsAllConstants = false;
    if (X86::isZeroNode(Elt))
      NumZero++;
    else {
      NonZeros |= (1 << i);
      NumNonZero++;
    }
  }

  // All undef vector. Return an UNDEF.  All zero vectors were handled above.
  if (NumNonZero == 0)
    return DAG.getUNDEF(VT);

  // Special case for single non-zero, non-undef, element.
  if (NumNonZero == 1) {
    unsigned Idx = CountTrailingZeros_32(NonZeros);
    SDValue Item = Op.getOperand(Idx);

    // If this is an insertion of an i64 value on x86-32, and if the top bits of
    // the value are obviously zero, truncate the value to i32 and do the
    // insertion that way.  Only do this if the value is non-constant or if the
    // value is a constant being inserted into element 0.  It is cheaper to do
    // a constant pool load than it is to do a movd + shuffle.
    if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
        (!IsAllConstants || Idx == 0)) {
      if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
        // Handle SSE only.
        assert(VT == MVT::v2i64 && "Expected an SSE value type!");
        EVT VecVT = MVT::v4i32;
        unsigned VecElts = 4;

        // Truncate the value (which may itself be a constant) to i32, and
        // convert it to a vector with movd (S2V+shuffle to zero extend).
        Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
        Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
        Item = getShuffleVectorZeroOrUndef(Item, 0, true,
                                           Subtarget->hasSSE2(), DAG);

        // Now we have our 32-bit value zero extended in the low element of
        // a vector.  If Idx != 0, swizzle it into place.
        if (Idx != 0) {
          SmallVector<int, 4> Mask;
          Mask.push_back(Idx);
          for (unsigned i = 1; i != VecElts; ++i)
            Mask.push_back(i);
          Item = DAG.getVectorShuffle(VecVT, dl, Item,
                                      DAG.getUNDEF(Item.getValueType()),
                                      &Mask[0]);
        }
        return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
      }
    }

    // If we have a constant or non-constant insertion into the low element of
    // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
    // the rest of the elements.  This will be matched as movd/movq/movss/movsd
    // depending on what the source datatype is.
    if (Idx == 0) {
      if (NumZero == 0) {
        return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
      } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
          (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
        Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
        // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
        return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
                                           DAG);
      } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
        Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
        assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
        EVT MiddleVT = MVT::v4i32;
        Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
        Item = getShuffleVectorZeroOrUndef(Item, 0, true,
                                           Subtarget->hasSSE2(), DAG);
        return DAG.getNode(ISD::BITCAST, dl, VT, Item);
      }
    }

    // Is it a vector logical left shift?
    if (NumElems == 2 && Idx == 1 &&
        X86::isZeroNode(Op.getOperand(0)) &&
        !X86::isZeroNode(Op.getOperand(1))) {
      unsigned NumBits = VT.getSizeInBits();
      return getVShift(true, VT,
                       DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
                                   VT, Op.getOperand(1)),
                       NumBits/2, DAG, *this, dl);
    }

    if (IsAllConstants) // Otherwise, it's better to do a constpool load.
      return SDValue();

    // Otherwise, if this is a vector with i32 or f32 elements, and the element
    // is a non-constant being inserted into an element other than the low one,
    // we can't use a constant pool load.  Instead, use SCALAR_TO_VECTOR (aka
    // movd/movss) to move this into the low element, then shuffle it into
    // place.
    if (EVTBits == 32) {
      Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);

      // Turn it into a shuffle of zero and zero-extended scalar to vector.
      Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
                                         Subtarget->hasSSE2(), DAG);
      SmallVector<int, 8> MaskVec;
      for (unsigned i = 0; i < NumElems; i++)
        MaskVec.push_back(i == Idx ? 0 : 1);
      return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
    }
  }

  // Splat is obviously ok. Let legalizer expand it to a shuffle.
  if (Values.size() == 1) {
    if (EVTBits == 32) {
      // Instead of a shuffle like this:
      // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
      // Check if it's possible to issue this instead.
      // shuffle (vload ptr)), undef, <1, 1, 1, 1>
      unsigned Idx = CountTrailingZeros_32(NonZeros);
      SDValue Item = Op.getOperand(Idx);
      if (Op.getNode()->isOnlyUserOf(Item.getNode()))
        return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
    }
    return SDValue();
  }

  // A vector full of immediates; various special cases are already
  // handled, so this is best done with a single constant-pool load.
  if (IsAllConstants)
    return SDValue();

  // Let legalizer expand 2-wide build_vectors.
  if (EVTBits == 64) {
    if (NumNonZero == 1) {
      // One half is zero or undef.
      unsigned Idx = CountTrailingZeros_32(NonZeros);
      SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
                                 Op.getOperand(Idx));
      return getShuffleVectorZeroOrUndef(V2, Idx, true,
                                         Subtarget->hasSSE2(), DAG);
    }
    return SDValue();
  }

  // If element VT is < 32 bits, convert it to inserts into a zero vector.
  if (EVTBits == 8 && NumElems == 16) {
    SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
                                        *this);
    if (V.getNode()) return V;
  }

  if (EVTBits == 16 && NumElems == 8) {
    SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
                                      *this);
    if (V.getNode()) return V;
  }

  // If element VT is == 32 bits, turn it into a number of shuffles.
  SmallVector<SDValue, 8> V;
  V.resize(NumElems);
  if (NumElems == 4 && NumZero > 0) {
    for (unsigned i = 0; i < 4; ++i) {
      bool isZero = !(NonZeros & (1 << i));
      if (isZero)
        V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
      else
        V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
    }

    for (unsigned i = 0; i < 2; ++i) {
      switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
        default: break;
        case 0:
          V[i] = V[i*2];  // Must be a zero vector.
          break;
        case 1:
          V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
          break;
        case 2:
          V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
          break;
        case 3:
          V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
          break;
      }
    }

    SmallVector<int, 8> MaskVec;
    bool Reverse = (NonZeros & 0x3) == 2;
    for (unsigned i = 0; i < 2; ++i)
      MaskVec.push_back(Reverse ? 1-i : i);
    Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
    for (unsigned i = 0; i < 2; ++i)
      MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
    return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
  }

  if (Values.size() > 1 && VT.getSizeInBits() == 128) {
    // Check for a build vector of consecutive loads.
    for (unsigned i = 0; i < NumElems; ++i)
      V[i] = Op.getOperand(i);

    // Check for elements which are consecutive loads.
    SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
    if (LD.getNode())
      return LD;

    // For SSE 4.1, use insertps to put the high elements into the low element.
    if (getSubtarget()->hasSSE41()) {
      SDValue Result;
      if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
        Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
      else
        Result = DAG.getUNDEF(VT);

      for (unsigned i = 1; i < NumElems; ++i) {
        if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
        Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
                             Op.getOperand(i), DAG.getIntPtrConstant(i));
      }
      return Result;
    }

    // Otherwise, expand into a number of unpckl*, start by extending each of
    // our (non-undef) elements to the full vector width with the element in the
    // bottom slot of the vector (which generates no code for SSE).
    for (unsigned i = 0; i < NumElems; ++i) {
      if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
        V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
      else
        V[i] = DAG.getUNDEF(VT);
    }

    // Next, we iteratively mix elements, e.g. for v4f32:
    //   Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
    //         : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
    //   Step 2: unpcklps X, Y ==>    <3, 2, 1, 0>
    unsigned EltStride = NumElems >> 1;
    while (EltStride != 0) {
      for (unsigned i = 0; i < EltStride; ++i) {
        // If V[i+EltStride] is undef and this is the first round of mixing,
        // then it is safe to just drop this shuffle: V[i] is already in the
        // right place, the one element (since it's the first round) being
        // inserted as undef can be dropped.  This isn't safe for successive
        // rounds because they will permute elements within both vectors.
        if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
            EltStride == NumElems/2)
          continue;

        V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
      }
      EltStride >>= 1;
    }
    return V[0];
  }
  return SDValue();
}

SDValue
X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
  // We support concatenate two MMX registers and place them in a MMX
  // register.  This is better than doing a stack convert.
  DebugLoc dl = Op.getDebugLoc();
  EVT ResVT = Op.getValueType();
  assert(Op.getNumOperands() == 2);
  assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
         ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
  int Mask[2];
  SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
  SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
  InVec = Op.getOperand(1);
  if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
    unsigned NumElts = ResVT.getVectorNumElements();
    VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
    VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
                       InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
  } else {
    InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
    SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
    Mask[0] = 0; Mask[1] = 2;
    VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
  }
  return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
}

// v8i16 shuffles - Prefer shuffles in the following order:
// 1. [all]   pshuflw, pshufhw, optional move
// 2. [ssse3] 1 x pshufb
// 3. [ssse3] 2 x pshufb + 1 x por
// 4. [all]   mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
SDValue
X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
                                            SelectionDAG &DAG) const {
  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
  SDValue V1 = SVOp->getOperand(0);
  SDValue V2 = SVOp->getOperand(1);
  DebugLoc dl = SVOp->getDebugLoc();
  SmallVector<int, 8> MaskVals;

  // Determine if more than 1 of the words in each of the low and high quadwords
  // of the result come from the same quadword of one of the two inputs.  Undef
  // mask values count as coming from any quadword, for better codegen.
  SmallVector<unsigned, 4> LoQuad(4);
  SmallVector<unsigned, 4> HiQuad(4);
  BitVector InputQuads(4);
  for (unsigned i = 0; i < 8; ++i) {
    SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
    int EltIdx = SVOp->getMaskElt(i);
    MaskVals.push_back(EltIdx);
    if (EltIdx < 0) {
      ++Quad[0];
      ++Quad[1];
      ++Quad[2];
      ++Quad[3];
      continue;
    }
    ++Quad[EltIdx / 4];
    InputQuads.set(EltIdx / 4);
  }

  int BestLoQuad = -1;
  unsigned MaxQuad = 1;
  for (unsigned i = 0; i < 4; ++i) {
    if (LoQuad[i] > MaxQuad) {
      BestLoQuad = i;
      MaxQuad = LoQuad[i];
    }
  }

  int BestHiQuad = -1;
  MaxQuad = 1;
  for (unsigned i = 0; i < 4; ++i) {
    if (HiQuad[i] > MaxQuad) {
      BestHiQuad = i;
      MaxQuad = HiQuad[i];
    }
  }

  // For SSSE3, If all 8 words of the result come from only 1 quadword of each
  // of the two input vectors, shuffle them into one input vector so only a
  // single pshufb instruction is necessary. If There are more than 2 input
  // quads, disable the next transformation since it does not help SSSE3.
  bool V1Used = InputQuads[0] || InputQuads[1];
  bool V2Used = InputQuads[2] || InputQuads[3];
  if (Subtarget->hasSSSE3()) {
    if (InputQuads.count() == 2 && V1Used && V2Used) {
      BestLoQuad = InputQuads.find_first();
      BestHiQuad = InputQuads.find_next(BestLoQuad);
    }
    if (InputQuads.count() > 2) {
      BestLoQuad = -1;
      BestHiQuad = -1;
    }
  }

  // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
  // the shuffle mask.  If a quad is scored as -1, that means that it contains
  // words from all 4 input quadwords.
  SDValue NewV;
  if (BestLoQuad >= 0 || BestHiQuad >= 0) {
    SmallVector<int, 8> MaskV;
    MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
    MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
    NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
                  DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
                  DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
    NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);

    // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
    // source words for the shuffle, to aid later transformations.
    bool AllWordsInNewV = true;
    bool InOrder[2] = { true, true };
    for (unsigned i = 0; i != 8; ++i) {
      int idx = MaskVals[i];
      if (idx != (int)i)
        InOrder[i/4] = false;
      if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
        continue;
      AllWordsInNewV = false;
      break;
    }

    bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
    if (AllWordsInNewV) {
      for (int i = 0; i != 8; ++i) {
        int idx = MaskVals[i];
        if (idx < 0)
          continue;
        idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
        if ((idx != i) && idx < 4)
          pshufhw = false;
        if ((idx != i) && idx > 3)
          pshuflw = false;
      }
      V1 = NewV;
      V2Used = false;
      BestLoQuad = 0;
      BestHiQuad = 1;
    }

    // If we've eliminated the use of V2, and the new mask is a pshuflw or
    // pshufhw, that's as cheap as it gets.  Return the new shuffle.
    if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
      unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
      unsigned TargetMask = 0;
      NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
                                  DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
      TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
                             X86::getShufflePSHUFLWImmediate(NewV.getNode());
      V1 = NewV.getOperand(0);
      return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
    }
  }

  // If we have SSSE3, and all words of the result are from 1 input vector,
  // case 2 is generated, otherwise case 3 is generated.  If no SSSE3
  // is present, fall back to case 4.
  if (Subtarget->hasSSSE3()) {
    SmallVector<SDValue,16> pshufbMask;

    // If we have elements from both input vectors, set the high bit of the
    // shuffle mask element to zero out elements that come from V2 in the V1
    // mask, and elements that come from V1 in the V2 mask, so that the two
    // results can be OR'd together.
    bool TwoInputs = V1Used && V2Used;
    for (unsigned i = 0; i != 8; ++i) {
      int EltIdx = MaskVals[i] * 2;
      if (TwoInputs && (EltIdx >= 16)) {
        pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
        pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
        continue;
      }
      pshufbMask.push_back(DAG.getConstant(EltIdx,   MVT::i8));
      pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
    }
    V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
    V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
                     DAG.getNode(ISD::BUILD_VECTOR, dl,
                                 MVT::v16i8, &pshufbMask[0], 16));
    if (!TwoInputs)
      return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);

    // Calculate the shuffle mask for the second input, shuffle it, and
    // OR it with the first shuffled input.
    pshufbMask.clear();
    for (unsigned i = 0; i != 8; ++i) {
      int EltIdx = MaskVals[i] * 2;
      if (EltIdx < 16) {
        pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
        pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
        continue;
      }
      pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
      pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
    }
    V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
    V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
                     DAG.getNode(ISD::BUILD_VECTOR, dl,
                                 MVT::v16i8, &pshufbMask[0], 16));
    V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
    return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
  }

  // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
  // and update MaskVals with new element order.
  BitVector InOrder(8);
  if (BestLoQuad >= 0) {
    SmallVector<int, 8> MaskV;
    for (int i = 0; i != 4; ++i) {
      int idx = MaskVals[i];
      if (idx < 0) {
        MaskV.push_back(-1);
        InOrder.set(i);
      } else if ((idx / 4) == BestLoQuad) {
        MaskV.push_back(idx & 3);
        InOrder.set(i);
      } else {
        MaskV.push_back(-1);
      }
    }
    for (unsigned i = 4; i != 8; ++i)
      MaskV.push_back(i);
    NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
                                &MaskV[0]);

    if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
      NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
                               NewV.getOperand(0),
                               X86::getShufflePSHUFLWImmediate(NewV.getNode()),
                               DAG);
  }

  // If BestHi >= 0, generate a pshufhw to put the high elements in order,
  // and update MaskVals with the new element order.
  if (BestHiQuad >= 0) {
    SmallVector<int, 8> MaskV;
    for (unsigned i = 0; i != 4; ++i)
      MaskV.push_back(i);
    for (unsigned i = 4; i != 8; ++i) {
      int idx = MaskVals[i];
      if (idx < 0) {
        MaskV.push_back(-1);
        InOrder.set(i);
      } else if ((idx / 4) == BestHiQuad) {
        MaskV.push_back((idx & 3) + 4);
        InOrder.set(i);
      } else {
        MaskV.push_back(-1);
      }
    }
    NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
                                &MaskV[0]);

    if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
      NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
                              NewV.getOperand(0),
                              X86::getShufflePSHUFHWImmediate(NewV.getNode()),
                              DAG);
  }

  // In case BestHi & BestLo were both -1, which means each quadword has a word
  // from each of the four input quadwords, calculate the InOrder bitvector now
  // before falling through to the insert/extract cleanup.
  if (BestLoQuad == -1 && BestHiQuad == -1) {
    NewV = V1;
    for (int i = 0; i != 8; ++i)
      if (MaskVals[i] < 0 || MaskVals[i] == i)
        InOrder.set(i);
  }

  // The other elements are put in the right place using pextrw and pinsrw.
  for (unsigned i = 0; i != 8; ++i) {
    if (InOrder[i])
      continue;
    int EltIdx = MaskVals[i];
    if (EltIdx < 0)
      continue;
    SDValue ExtOp = (EltIdx < 8)
    ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
                  DAG.getIntPtrConstant(EltIdx))
    : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
                  DAG.getIntPtrConstant(EltIdx - 8));
    NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
                       DAG.getIntPtrConstant(i));
  }
  return NewV;
}

// v16i8 shuffles - Prefer shuffles in the following order:
// 1. [ssse3] 1 x pshufb
// 2. [ssse3] 2 x pshufb + 1 x por
// 3. [all]   v8i16 shuffle + N x pextrw + rotate + pinsrw
static
SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
                                 SelectionDAG &DAG,
                                 const X86TargetLowering &TLI) {
  SDValue V1 = SVOp->getOperand(0);
  SDValue V2 = SVOp->getOperand(1);
  DebugLoc dl = SVOp->getDebugLoc();
  SmallVector<int, 16> MaskVals;
  SVOp->getMask(MaskVals);

  // If we have SSSE3, case 1 is generated when all result bytes come from
  // one of  the inputs.  Otherwise, case 2 is generated.  If no SSSE3 is
  // present, fall back to case 3.
  // FIXME: kill V2Only once shuffles are canonizalized by getNode.
  bool V1Only = true;
  bool V2Only = true;
  for (unsigned i = 0; i < 16; ++i) {
    int EltIdx = MaskVals[i];
    if (EltIdx < 0)
      continue;
    if (EltIdx < 16)
      V2Only = false;
    else
      V1Only = false;
  }

  // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
  if (TLI.getSubtarget()->hasSSSE3()) {
    SmallVector<SDValue,16> pshufbMask;

    // If all result elements are from one input vector, then only translate
    // undef mask values to 0x80 (zero out result) in the pshufb mask.
    //
    // Otherwise, we have elements from both input vectors, and must zero out
    // elements that come from V2 in the first mask, and V1 in the second mask
    // so that we can OR them together.
    bool TwoInputs = !(V1Only || V2Only);
    for (unsigned i = 0; i != 16; ++i) {
      int EltIdx = MaskVals[i];
      if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
        pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
        continue;
      }
      pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
    }
    // If all the elements are from V2, assign it to V1 and return after
    // building the first pshufb.
    if (V2Only)
      V1 = V2;
    V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
                     DAG.getNode(ISD::BUILD_VECTOR, dl,
                                 MVT::v16i8, &pshufbMask[0], 16));
    if (!TwoInputs)
      return V1;

    // Calculate the shuffle mask for the second input, shuffle it, and
    // OR it with the first shuffled input.
    pshufbMask.clear();
    for (unsigned i = 0; i != 16; ++i) {
      int EltIdx = MaskVals[i];
      if (EltIdx < 16) {
        pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
        continue;
      }
      pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
    }
    V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
                     DAG.getNode(ISD::BUILD_VECTOR, dl,
                                 MVT::v16i8, &pshufbMask[0], 16));
    return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
  }

  // No SSSE3 - Calculate in place words and then fix all out of place words
  // With 0-16 extracts & inserts.  Worst case is 16 bytes out of order from
  // the 16 different words that comprise the two doublequadword input vectors.
  V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
  V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
  SDValue NewV = V2Only ? V2 : V1;
  for (int i = 0; i != 8; ++i) {
    int Elt0 = MaskVals[i*2];
    int Elt1 = MaskVals[i*2+1];

    // This word of the result is all undef, skip it.
    if (Elt0 < 0 && Elt1 < 0)
      continue;

    // This word of the result is already in the correct place, skip it.
    if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
      continue;
    if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
      continue;

    SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
    SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
    SDValue InsElt;

    // If Elt0 and Elt1 are defined, are consecutive, and can be load
    // using a single extract together, load it and store it.
    if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
      InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
                           DAG.getIntPtrConstant(Elt1 / 2));
      NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
                        DAG.getIntPtrConstant(i));
      continue;
    }

    // If Elt1 is defined, extract it from the appropriate source.  If the
    // source byte is not also odd, shift the extracted word left 8 bits
    // otherwise clear the bottom 8 bits if we need to do an or.
    if (Elt1 >= 0) {
      InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
                           DAG.getIntPtrConstant(Elt1 / 2));
      if ((Elt1 & 1) == 0)
        InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
                             DAG.getConstant(8, TLI.getShiftAmountTy()));
      else if (Elt0 >= 0)
        InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
                             DAG.getConstant(0xFF00, MVT::i16));
    }
    // If Elt0 is defined, extract it from the appropriate source.  If the
    // source byte is not also even, shift the extracted word right 8 bits. If
    // Elt1 was also defined, OR the extracted values together before
    // inserting them in the result.
    if (Elt0 >= 0) {
      SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
                                    Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
      if ((Elt0 & 1) != 0)
        InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
                              DAG.getConstant(8, TLI.getShiftAmountTy()));
      else if (Elt1 >= 0)
        InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
                             DAG.getConstant(0x00FF, MVT::i16));
      InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
                         : InsElt0;
    }
    NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
                       DAG.getIntPtrConstant(i));
  }
  return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
}

/// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
/// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
/// done when every pair / quad of shuffle mask elements point to elements in
/// the right sequence. e.g.
/// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
static
SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
                                 SelectionDAG &DAG, DebugLoc dl) {
  EVT VT = SVOp->getValueType(0);
  SDValue V1 = SVOp->getOperand(0);
  SDValue V2 = SVOp->getOperand(1);
  unsigned NumElems = VT.getVectorNumElements();
  unsigned NewWidth = (NumElems == 4) ? 2 : 4;
  EVT NewVT;
  switch (VT.getSimpleVT().SimpleTy) {
  default: assert(false && "Unexpected!");
  case MVT::v4f32: NewVT = MVT::v2f64; break;
  case MVT::v4i32: NewVT = MVT::v2i64; break;
  case MVT::v8i16: NewVT = MVT::v4i32; break;
  case MVT::v16i8: NewVT = MVT::v4i32; break;
  }

  int Scale = NumElems / NewWidth;
  SmallVector<int, 8> MaskVec;
  for (unsigned i = 0; i < NumElems; i += Scale) {
    int StartIdx = -1;
    for (int j = 0; j < Scale; ++j) {
      int EltIdx = SVOp->getMaskElt(i+j);
      if (EltIdx < 0)
        continue;
      if (StartIdx == -1)
        StartIdx = EltIdx - (EltIdx % Scale);
      if (EltIdx != StartIdx + j)
        return SDValue();
    }
    if (StartIdx == -1)
      MaskVec.push_back(-1);
    else
      MaskVec.push_back(StartIdx / Scale);
  }

  V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
  V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
  return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
}

/// getVZextMovL - Return a zero-extending vector move low node.
///
static SDValue getVZextMovL(EVT VT, EVT OpVT,
                            SDValue SrcOp, SelectionDAG &DAG,
                            const X86Subtarget *Subtarget, DebugLoc dl) {
  if (VT == MVT::v2f64 || VT == MVT::v4f32) {
    LoadSDNode *LD = NULL;
    if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
      LD = dyn_cast<LoadSDNode>(SrcOp);
    if (!LD) {
      // movssrr and movsdrr do not clear top bits. Try to use movd, movq
      // instead.
      MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
      if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
          SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
          SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
          SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
        // PR2108
        OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
        return DAG.getNode(ISD::BITCAST, dl, VT,
                           DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
                                       DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
                                                   OpVT,
                                                   SrcOp.getOperand(0)
                                                          .getOperand(0))));
      }
    }
  }

  return DAG.getNode(ISD::BITCAST, dl, VT,
                     DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
                                 DAG.getNode(ISD::BITCAST, dl,
                                             OpVT, SrcOp)));
}

/// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
/// shuffles.
static SDValue
LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
  SDValue V1 = SVOp->getOperand(0);
  SDValue V2 = SVOp->getOperand(1);
  DebugLoc dl = SVOp->getDebugLoc();
  EVT VT = SVOp->getValueType(0);

  SmallVector<std::pair<int, int>, 8> Locs;
  Locs.resize(4);
  SmallVector<int, 8> Mask1(4U, -1);
  SmallVector<int, 8> PermMask;
  SVOp->getMask(PermMask);

  unsigned NumHi = 0;
  unsigned NumLo = 0;
  for (unsigned i = 0; i != 4; ++i) {
    int Idx = PermMask[i];
    if (Idx < 0) {
      Locs[i] = std::make_pair(-1, -1);
    } else {
      assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
      if (Idx < 4) {
        Locs[i] = std::make_pair(0, NumLo);
        Mask1[NumLo] = Idx;
        NumLo++;
      } else {
        Locs[i] = std::make_pair(1, NumHi);
        if (2+NumHi < 4)
          Mask1[2+NumHi] = Idx;
        NumHi++;
      }
    }
  }

  if (NumLo <= 2 && NumHi <= 2) {
    // If no more than two elements come from either vector. This can be
    // implemented with two shuffles. First shuffle gather the elements.
    // The second shuffle, which takes the first shuffle as both of its
    // vector operands, put the elements into the right order.
    V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);

    SmallVector<int, 8> Mask2(4U, -1);

    for (unsigned i = 0; i != 4; ++i) {
      if (Locs[i].first == -1)
        continue;
      else {
        unsigned Idx = (i < 2) ? 0 : 4;
        Idx += Locs[i].first * 2 + Locs[i].second;
        Mask2[i] = Idx;
      }
    }

    return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
  } else if (NumLo == 3 || NumHi == 3) {
    // Otherwise, we must have three elements from one vector, call it X, and
    // one element from the other, call it Y.  First, use a shufps to build an
    // intermediate vector with the one element from Y and the element from X
    // that will be in the same half in the final destination (the indexes don't
    // matter). Then, use a shufps to build the final vector, taking the half