X86ISelLowering.cpp

    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 inserts into undef.  
    if (getSubtarget()->hasSSE41()) {
      V[0] = DAG.getUNDEF(VT);
      for (unsigned i = 0; i < NumElems; ++i)
        if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
          V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
                             Op.getOperand(i), DAG.getIntPtrConstant(i));
      return V[0];
    }
    
    // Otherwise, expand into a number of unpckl*
    // 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>
    for (unsigned i = 0; i < NumElems; ++i)
      V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
    NumElems >>= 1;
    while (NumElems != 0) {
      for (unsigned i = 0; i < NumElems; ++i)
        V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
      NumElems >>= 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::BIT_CONVERT,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::BIT_CONVERT, 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::BIT_CONVERT, 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::BIT_CONVERT, 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)
static
SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
                                 SelectionDAG &DAG,
                                 const X86TargetLowering &TLI) {
  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 (TLI.getSubtarget()->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::BIT_CONVERT, dl, MVT::v2i64, V1),
                  DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
    NewV = DAG.getNode(ISD::BIT_CONVERT, 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])) {
      return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
                                  DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
    }
  }

  // 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 (TLI.getSubtarget()->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::BIT_CONVERT, 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::BIT_CONVERT, 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::BIT_CONVERT, 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::BIT_CONVERT, 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 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]);
  }

  // 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::BIT_CONVERT, dl, MVT::v8i16, V1);
  V2 = DAG.getNode(ISD::BIT_CONVERT, 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::BIT_CONVERT, dl, MVT::v16i8, NewV);
}

/// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
/// ones, or rewriting v4i32 / v2i32 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 <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
static
SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
                                 SelectionDAG &DAG,
                                 const TargetLowering &TLI, 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 MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
  EVT NewVT = MaskVT;
  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;
  }

  if (NewWidth == 2) {
    if (VT.isInteger())
      NewVT = MVT::v2i64;
    else
      NewVT = MVT::v2f64;
  }
  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::BIT_CONVERT, dl, NewVT, V1);
  V2 = DAG.getNode(ISD::BIT_CONVERT, 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.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
          SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
          SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
          SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
        // PR2108
        OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
        return DAG.getNode(ISD::BIT_CONVERT, 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::BIT_CONVERT, dl, VT,
                     DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
                                 DAG.getNode(ISD::BIT_CONVERT, 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
    // containing the element from Y from the intermediate, and the other half
    // from X.
    if (NumHi == 3) {
      // Normalize it so the 3 elements come from V1.
      CommuteVectorShuffleMask(PermMask, VT);
      std::swap(V1, V2);
    }

    // Find the element from V2.
    unsigned HiIndex;
    for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
      int Val = PermMask[HiIndex];
      if (Val < 0)
        continue;
      if (Val >= 4)
        break;
    }

    Mask1[0] = PermMask[HiIndex];
    Mask1[1] = -1;
    Mask1[2] = PermMask[HiIndex^1];
    Mask1[3] = -1;
    V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);

    if (HiIndex >= 2) {
      Mask1[0] = PermMask[0];
      Mask1[1] = PermMask[1];
      Mask1[2] = HiIndex & 1 ? 6 : 4;
      Mask1[3] = HiIndex & 1 ? 4 : 6;
      return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
    } else {
      Mask1[0] = HiIndex & 1 ? 2 : 0;
      Mask1[1] = HiIndex & 1 ? 0 : 2;
      Mask1[2] = PermMask[2];
      Mask1[3] = PermMask[3];
      if (Mask1[2] >= 0)
        Mask1[2] += 4;
      if (Mask1[3] >= 0)
        Mask1[3] += 4;
      return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
    }
  }

  // Break it into (shuffle shuffle_hi, shuffle_lo).
  Locs.clear();
  SmallVector<int,8> LoMask(4U, -1);
  SmallVector<int,8> HiMask(4U, -1);

  SmallVector<int,8> *MaskPtr = &LoMask;
  unsigned MaskIdx = 0;
  unsigned LoIdx = 0;
  unsigned HiIdx = 2;
  for (unsigned i = 0; i != 4; ++i) {
    if (i == 2) {
      MaskPtr = &HiMask;
      MaskIdx = 1;
      LoIdx = 0;
      HiIdx = 2;
    }
    int Idx = PermMask[i];
    if (Idx < 0) {
      Locs[i] = std::make_pair(-1, -1);
    } else if (Idx < 4) {
      Locs[i] = std::make_pair(MaskIdx, LoIdx);
      (*MaskPtr)[LoIdx] = Idx;
      LoIdx++;
    } else {
      Locs[i] = std::make_pair(MaskIdx, HiIdx);
      (*MaskPtr)[HiIdx] = Idx;
      HiIdx++;
    }
  }

  SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
  SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
  SmallVector<int, 8> MaskOps;
  for (unsigned i = 0; i != 4; ++i) {
    if (Locs[i].first == -1) {
      MaskOps.push_back(-1);
    } else {
      unsigned Idx = Locs[i].first * 4 + Locs[i].second;
      MaskOps.push_back(Idx);
    }
  }
  return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
}

SDValue
X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
  SDValue V1 = Op.getOperand(0);
  SDValue V2 = Op.getOperand(1);
  EVT VT = Op.getValueType();
  DebugLoc dl = Op.getDebugLoc();
  unsigned NumElems = VT.getVectorNumElements();
  bool isMMX = VT.getSizeInBits() == 64;
  bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
  bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
  bool V1IsSplat = false;
  bool V2IsSplat = false;

  if (isZeroShuffle(SVOp))
    return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);

  // Promote splats to v4f32.
  if (SVOp->isSplat()) {
    if (isMMX || NumElems < 4)
      return Op;
    return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
  }

  // If the shuffle can be profitably rewritten as a narrower shuffle, then
  // do it!
  if (VT == MVT::v8i16 || VT == MVT::v16i8) {
    SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
    if (NewOp.getNode())
      return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
                         LowerVECTOR_SHUFFLE(NewOp, DAG));
  } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
    // FIXME: Figure out a cleaner way to do this.
    // Try to make use of movq to zero out the top part.
    if (ISD::isBuildVectorAllZeros(V2.getNode())) {
      SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
      if (NewOp.getNode()) {
        if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
          return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
                              DAG, Subtarget, dl);
      }
    } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
      SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
      if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
        return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
                            DAG, Subtarget, dl);
    }
  }

  if (X86::isPSHUFDMask(SVOp))
    return Op;

  // Check if this can be converted into a logical shift.
  bool isLeft = false;
  unsigned ShAmt = 0;
  SDValue ShVal;
  bool isShift = getSubtarget()->hasSSE2() &&
    isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
  if (isShift && ShVal.hasOneUse()) {
    // If the shifted value has multiple uses, it may be cheaper to use
    // v_set0 + movlhps or movhlps, etc.
    EVT EltVT = VT.getVectorElementType();
    ShAmt *= EltVT.getSizeInBits();
    return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
  }

  if (X86::isMOVLMask(SVOp)) {
    if (V1IsUndef)
      return V2;
    if (ISD::isBuildVectorAllZeros(V1.getNode()))
      return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
    if (!isMMX)
      return Op;
  }

  // FIXME: fold these into legal mask.
  if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
                 X86::isMOVSLDUPMask(SVOp) ||
                 X86::isMOVHLPSMask(SVOp) ||
                 X86::isMOVLHPSMask(SVOp) ||
                 X86::isMOVLPMask(SVOp)))
    return Op;

  if (ShouldXformToMOVHLPS(SVOp) ||
      ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
    return CommuteVectorShuffle(SVOp, DAG);

  if (isShift) {
    // No better options. Use a vshl / vsrl.
    EVT EltVT = VT.getVectorElementType();
    ShAmt *= EltVT.getSizeInBits();
    return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
  }

  bool Commuted = false;
  // FIXME: This should also accept a bitcast of a splat?  Be careful, not
  // 1,1,1,1 -> v8i16 though.
  V1IsSplat = isSplatVector(V1.getNode());
  V2IsSplat = isSplatVector(V2.getNode());

  // Canonicalize the splat or undef, if present, to be on the RHS.
  if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
    Op = CommuteVectorShuffle(SVOp, DAG);
    SVOp = cast<ShuffleVectorSDNode>(Op);
    V1 = SVOp->getOperand(0);
    V2 = SVOp->getOperand(1);
    std::swap(V1IsSplat, V2IsSplat);
    std::swap(V1IsUndef, V2IsUndef);
    Commuted = true;
  }

  if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
    // Shuffling low element of v1 into undef, just return v1.
    if (V2IsUndef)
      return V1;
    // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
    // the instruction selector will not match, so get a canonical MOVL with
    // swapped operands to undo the commute.
    return getMOVL(DAG, dl, VT, V2, V1);
  }

  if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
      X86::isUNPCKH_v_undef_Mask(SVOp) ||
      X86::isUNPCKLMask(SVOp) ||
      X86::isUNPCKHMask(SVOp))
    return Op;

  if (V2IsSplat) {
    // Normalize mask so all entries that point to V2 points to its first
    // element then try to match unpck{h|l} again. If match, return a
    // new vector_shuffle with the corrected mask.
    SDValue NewMask = NormalizeMask(SVOp, DAG);
    ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
    if (NSVOp != SVOp) {
      if (X86::isUNPCKLMask(NSVOp, true)) {
        return NewMask;
      } else if (X86::isUNPCKHMask(NSVOp, true)) {
        return NewMask;
      }
    }
  }

  if (Commuted) {
    // Commute is back and try unpck* again.
    // FIXME: this seems wrong.
    SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
    ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
    if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
        X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
        X86::isUNPCKLMask(NewSVOp) ||
        X86::isUNPCKHMask(NewSVOp))
      return NewOp;
  }

  // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.

  // Normalize the node to match x86 shuffle ops if needed
  if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
    return CommuteVectorShuffle(SVOp, DAG);

  // Check for legal shuffle and return?
  SmallVector<int, 16> PermMask;
  SVOp->getMask(PermMask);
  if (isShuffleMaskLegal(PermMask, VT))
    return Op;

  // Handle v8i16 specifically since SSE can do byte extraction and insertion.
  if (VT == MVT::v8i16) {
    SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
    if (NewOp.getNode())
      return NewOp;
  }

  if (VT == MVT::v16i8) {
    SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
    if (NewOp.getNode())
      return NewOp;
  }

  // Handle all 4 wide cases with a number of shuffles except for MMX.
  if (NumElems == 4 && !isMMX)
    return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);

  return SDValue();
}

SDValue
X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
                                                SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  DebugLoc dl = Op.getDebugLoc();
  if (VT.getSizeInBits() == 8) {
    SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
                                    Op.getOperand(0), Op.getOperand(1));
    SDValue Assert  = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
                                    DAG.getValueType(VT));
    return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
  } else if (VT.getSizeInBits() == 16) {
    unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
    // If Idx is 0, it's cheaper to do a move instead of a pextrw.
    if (Idx == 0)
      return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
                         DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
                                     DAG.getNode(ISD::BIT_CONVERT, dl,
                                                 MVT::v4i32,
                                                 Op.getOperand(0)),
                                     Op.getOperand(1)));
    SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
                                    Op.getOperand(0), Op.getOperand(1));
    SDValue Assert  = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
                                    DAG.getValueType(VT));
    return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
  } else if (VT == MVT::f32) {
    // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
    // the result back to FR32 register. It's only worth matching if the
    // result has a single use which is a store or a bitcast to i32.  And in
    // the case of a store, it's not worth it if the index is a constant 0,
    // because a MOVSSmr can be used instead, which is smaller and faster.
    if (!Op.hasOneUse())
      return SDValue();
    SDNode *User = *Op.getNode()->use_begin();
    if ((User->getOpcode() != ISD::STORE ||
         (isa<ConstantSDNode>(Op.getOperand(1)) &&
          cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
        (User->getOpcode() != ISD::BIT_CONVERT ||
         User->getValueType(0) != MVT::i32))
      return SDValue();
    SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
                                  DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
                                              Op.getOperand(0)),
                                              Op.getOperand(1));
    return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
  } else if (VT == MVT::i32) {
    // ExtractPS works with constant index.
    if (isa<ConstantSDNode>(Op.getOperand(1)))
      return Op;
  }
  return SDValue();
}


SDValue
X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
                                           SelectionDAG &DAG) const {
  if (!isa<ConstantSDNode>(Op.getOperand(1)))
    return SDValue();

  if (Subtarget->hasSSE41()) {
    SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
    if (Res.getNode())
      return Res;
  }

  EVT VT = Op.getValueType();
  DebugLoc dl = Op.getDebugLoc();
  // TODO: handle v16i8.
  if (VT.getSizeInBits() == 16) {
    SDValue Vec = Op.getOperand(0);
    unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
    if (Idx == 0)
      return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
                         DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
                                     DAG.getNode(ISD::BIT_CONVERT, dl,
                                                 MVT::v4i32, Vec),
                                     Op.getOperand(1)));
    // Transform it so it match pextrw which produces a 32-bit result.
    EVT EltVT = MVT::i32;
    SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
                                    Op.getOperand(0), Op.getOperand(1));
    SDValue Assert  = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
                                    DAG.getValueType(VT));
    return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
  } else if (VT.getSizeInBits() == 32) {
    unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
    if (Idx == 0)
      return Op;

    // SHUFPS the element to the lowest double word, then movss.
    int Mask[4] = { Idx, -1, -1, -1 };
    EVT VVT = Op.getOperand(0).getValueType();
    SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
                                       DAG.getUNDEF(VVT), Mask);
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
                       DAG.getIntPtrConstant(0));
  } else if (VT.getSizeInBits() == 64) {
    // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
    // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
    //        to match extract_elt for f64.
    unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
    if (Idx == 0)
      return Op;

    // UNPCKHPD the element to the lowest double word, then movsd.
    // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
    // to a f64mem, the whole operation is folded into a single MOVHPDmr.
    int Mask[2] = { 1, -1 };
    EVT VVT = Op.getOperand(0).getValueType();
    SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
                                       DAG.getUNDEF(VVT), Mask);
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
                       DAG.getIntPtrConstant(0));