[x86] Begin a significant overhaul of how vector lowering is done in the

x86 backend.

This sketches out a new code path for vector lowering, hidden behind an
off-by-default flag while it is under development. The fundamental idea
behind the new code path is to aggressively break down the problem space
in ways that ease selecting the odd set of instructions available on
x86, and carefully avoid scalarizing code even when forced to use older
ISAs. Notably, this starts off restricting itself to SSE2 and implements
the complete vector shuffle and blend space for 128-bit vectors in SSE2
without scalarizing. The plan is to layer on top of this ISA extensions
where we can bail out of the complex SSE2 lowering and opt for
a cheaper, specialized instruction (or set of instructions). It also
needs to be generalized to AVX and AVX512 vector widths.

Currently, this does a decent but not perfect job for SSE2. There are
some specific shortcomings that I plan to address:
- We need a peephole combine to fold together shuffles where possible.
  There are cases where a previous shuffle could be modified slightly to
  arrange for elements to be in the correct position and a later shuffle
  eliminated. Doing this eagerly added quite a bit of complexity, and
  so my plan is to combine away these redundancies afterward.
- There are a lot more clever ways to use unpck and pack that need to be
  added. This is essential for real world shuffles as it turns out...

Once SSE2 is polished a bit I should be able to get interesting numbers
on performance improvements on benchmarks conducive to vectorization.
All of this will be off by default until it is functionally equivalent
of course.

Differential Revision: http://reviews.llvm.org/D4225

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@211888 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 1029 insertions, 0 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index d7e63e39ec..3ceeac3b34 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -44,11 +44,13 @@
 #include "llvm/MC/MCContext.h"
 #include "llvm/MC/MCExpr.h"
 #include "llvm/MC/MCSymbol.h"
+#include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/MathExtras.h"
 #include "llvm/Target/TargetOptions.h"
 #include <bitset>
+#include <numeric>
 #include <cctype>
 using namespace llvm;
 
@@ -56,6 +58,11 @@ using namespace llvm;
 
 STATISTIC(NumTailCalls, "Number of tail calls");
 
+static cl::opt<bool> ExperimentalVectorShuffleLowering(
+    "x86-experimental-vector-shuffle-lowering", cl::init(false),
+    cl::desc("Enable an experimental vector shuffle lowering code path."),
+    cl::Hidden);
+
 // Forward declarations.
 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
                        SDValue V2);
@@ -6875,6 +6882,1023 @@ static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
   return LowerAVXCONCAT_VECTORS(Op, DAG);
 }
 
+
+//===----------------------------------------------------------------------===//
+// Vector shuffle lowering
+//
+// This is an experimental code path for lowering vector shuffles on x86. It is
+// designed to handle arbitrary vector shuffles and blends, gracefully
+// degrading performance as necessary. It works hard to recognize idiomatic
+// shuffles and lower them to optimal instruction patterns without leaving
+// a framework that allows reasonably efficient handling of all vector shuffle
+// patterns.
+//===----------------------------------------------------------------------===//
+
+/// \brief Tiny helper function to identify a no-op mask.
+///
+/// This is a somewhat boring predicate function. It checks whether the mask
+/// array input, which is assumed to be a single-input shuffle mask of the kind
+/// used by the X86 shuffle instructions (not a fully general
+/// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
+/// in-place shuffle are 'no-op's.
+static bool isNoopShuffleMask(ArrayRef<int> Mask) {
+  for (int i = 0, Size = Mask.size(); i < Size; ++i)
+    if (Mask[i] != -1 && Mask[i] != i)
+      return false;
+  return true;
+}
+
+/// \brief Helper function to classify a mask as a single-input mask.
+///
+/// This isn't a generic single-input test because in the vector shuffle
+/// lowering we canonicalize single inputs to be the first input operand. This
+/// means we can more quickly test for a single input by only checking whether
+/// an input from the second operand exists. We also assume that the size of
+/// mask corresponds to the size of the input vectors which isn't true in the
+/// fully general case.
+static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
+  for (int M : Mask)
+    if (M >= (int)Mask.size())
+      return false;
+  return true;
+}
+
+/// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
+///
+/// This helper function produces an 8-bit shuffle immediate corresponding to
+/// the ubiquitous shuffle encoding scheme used in x86 instructions for
+/// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
+/// example.
+///
+/// NB: We rely heavily on "undef" masks preserving the input lane.
+static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask,
+                                          SelectionDAG &DAG) {
+  assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
+  assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
+  assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
+  assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
+  assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
+
+  unsigned Imm = 0;
+  Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
+  Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
+  Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
+  Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
+  return DAG.getConstant(Imm, MVT::i8);
+}
+
+/// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
+///
+/// This is the basis function for the 2-lane 64-bit shuffles as we have full
+/// support for floating point shuffles but not integer shuffles. These
+/// instructions will incur a domain crossing penalty on some chips though so
+/// it is better to avoid lowering through this for integer vectors where
+/// possible.
+static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+                                       const X86Subtarget *Subtarget,
+                                       SelectionDAG &DAG) {
+  SDLoc DL(Op);
+  assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
+  assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
+  assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
+  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+  ArrayRef<int> Mask = SVOp->getMask();
+  assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
+
+  if (isSingleInputShuffleMask(Mask)) {
+    // Straight shuffle of a single input vector. Simulate this by using the
+    // single input as both of the "inputs" to this instruction..
+    unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
+    return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1,
+                       DAG.getConstant(SHUFPDMask, MVT::i8));
+  }
+  assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
+  assert(Mask[1] >= 2 && "Non-canonicalized blend!");
+
+  unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
+  return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V2,
+                     DAG.getConstant(SHUFPDMask, MVT::i8));
+}
+
+/// \brief Handle lowering of 2-lane 64-bit integer shuffles.
+///
+/// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
+/// the integer unit to minimize domain crossing penalties. However, for blends
+/// it falls back to the floating point shuffle operation with appropriate bit
+/// casting.
+static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+                                       const X86Subtarget *Subtarget,
+                                       SelectionDAG &DAG) {
+  SDLoc DL(Op);
+  assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
+  assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
+  assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
+  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+  ArrayRef<int> Mask = SVOp->getMask();
+  assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
+
+  if (isSingleInputShuffleMask(Mask)) {
+    // Straight shuffle of a single input vector. For everything from SSE2
+    // onward this has a single fast instruction with no scary immediates.
+    // We have to map the mask as it is actually a v4i32 shuffle instruction.
+    V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
+    int WidenedMask[4] = {
+        std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
+        std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
+    return DAG.getNode(
+        ISD::BITCAST, DL, MVT::v2i64,
+        DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1,
+                    getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
+  }
+
+  // We implement this with SHUFPD which is pretty lame because it will likely
+  // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
+  // However, all the alternatives are still more cycles and newer chips don't
+  // have this problem. It would be really nice if x86 had better shuffles here.
+  V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
+  V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
+  return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
+                     DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
+}
+
+/// \brief Lower 4-lane 32-bit floating point shuffles.
+///
+/// Uses instructions exclusively from the floating point unit to minimize
+/// domain crossing penalties, as these are sufficient to implement all v4f32
+/// shuffles.
+static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+                                       const X86Subtarget *Subtarget,
+                                       SelectionDAG &DAG) {
+  SDLoc DL(Op);
+  assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
+  assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
+  assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
+  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+  ArrayRef<int> Mask = SVOp->getMask();
+  assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
+
+  SDValue LowV = V1, HighV = V2;
+  int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
+
+  int NumV2Elements =
+      std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
+
+  if (NumV2Elements == 0)
+    // Straight shuffle of a single input vector. We pass the input vector to
+    // both operands to simulate this with a SHUFPS.
+    return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
+                       getV4X86ShuffleImm8ForMask(Mask, DAG));
+
+  if (NumV2Elements == 1) {
+    int V2Index =
+        std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
+        Mask.begin();
+    // Compute the index adjacent to V2Index and in the same half by toggling
+    // the low bit.
+    int V2AdjIndex = V2Index ^ 1;
+
+    if (Mask[V2AdjIndex] == -1) {
+      // Handles all the cases where we have a single V2 element and an undef.
+      // This will only ever happen in the high lanes because we commute the
+      // vector otherwise.
+      if (V2Index < 2)
+        std::swap(LowV, HighV);
+      NewMask[V2Index] -= 4;
+    } else {
+      // Handle the case where the V2 element ends up adjacent to a V1 element.
+      // To make this work, blend them together as the first step.
+      int V1Index = V2AdjIndex;
+      int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
+      V2 = DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V2, V1,
+                       getV4X86ShuffleImm8ForMask(BlendMask, DAG));
+
+      // Now proceed to reconstruct the final blend as we have the necessary
+      // high or low half formed.
+      if (V2Index < 2) {
+        LowV = V2;
+        HighV = V1;
+      } else {
+        HighV = V2;
+      }
+      NewMask[V1Index] = 2; // We put the V1 element in V2[2].
+      NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
+    }
+  } else if (NumV2Elements == 2) {
+    if (Mask[0] < 4 && Mask[1] < 4) {
+      // Handle the easy case where we have V1 in the low lanes and V2 in the
+      // high lanes. We never see this reversed because we sort the shuffle.
+      NewMask[2] -= 4;
+      NewMask[3] -= 4;
+    } else {
+      // We have a mixture of V1 and V2 in both low and high lanes. Rather than
+      // trying to place elements directly, just blend them and set up the final
+      // shuffle to place them.
+
+      // The first two blend mask elements are for V1, the second two are for
+      // V2.
+      int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
+                          Mask[2] < 4 ? Mask[2] : Mask[3],
+                          (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
+                          (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
+      V1 = DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V2,
+                       getV4X86ShuffleImm8ForMask(BlendMask, DAG));
+
+      // Now we do a normal shuffle of V1 by giving V1 as both operands to
+      // a blend.
+      HighV = V1;
+      NewMask[0] = Mask[0] < 4 ? 0 : 2;
+      NewMask[1] = Mask[0] < 4 ? 2 : 0;
+      NewMask[2] = Mask[2] < 4 ? 1 : 3;
+      NewMask[3] = Mask[2] < 4 ? 3 : 1;
+    }
+  }
+  return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, LowV, HighV,
+                     getV4X86ShuffleImm8ForMask(NewMask, DAG));
+}
+
+/// \brief Lower 4-lane i32 vector shuffles.
+///
+/// We try to handle these with integer-domain shuffles where we can, but for
+/// blends we use the floating point domain blend instructions.
+static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+                                       const X86Subtarget *Subtarget,
+                                       SelectionDAG &DAG) {
+  SDLoc DL(Op);
+  assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
+  assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
+  assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
+  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+  ArrayRef<int> Mask = SVOp->getMask();
+  assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
+
+  if (isSingleInputShuffleMask(Mask))
+    // Straight shuffle of a single input vector. For everything from SSE2
+    // onward this has a single fast instruction with no scary immediates.
+    return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
+                       getV4X86ShuffleImm8ForMask(Mask, DAG));
+
+  // We implement this with SHUFPS because it can blend from two vectors.
+  // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
+  // up the inputs, bypassing domain shift penalties that we would encur if we
+  // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
+  // relevant.
+  return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
+                     DAG.getVectorShuffle(
+                         MVT::v4f32, DL,
+                         DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
+                         DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
+}
+
+/// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
+/// shuffle lowering, and the most complex part.
+///
+/// The lowering strategy is to try to form pairs of input lanes which are
+/// targeted at the same half of the final vector, and then use a dword shuffle
+/// to place them onto the right half, and finally unpack the paired lanes into
+/// their final position.
+///
+/// The exact breakdown of how to form these dword pairs and align them on the
+/// correct sides is really tricky. See the comments within the function for
+/// more of the details.
+static SDValue lowerV8I16SingleInputVectorShuffle(
+    SDLoc DL, SDValue V, MutableArrayRef<int> Mask,
+    const X86Subtarget *Subtarget, SelectionDAG &DAG) {
+  assert(V.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
+  MutableArrayRef<int> LoMask = Mask.slice(0, 4);
+  MutableArrayRef<int> HiMask = Mask.slice(4, 4);
+
+  auto isLo = [](int M) { return M >= 0 && M < 4; };
+  auto isHi = [](int M) { return M >= 4; };
+
+  SmallVector<int, 4> LoInputs;
+  std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
+               [](int M) { return M >= 0; });
+  std::sort(LoInputs.begin(), LoInputs.end());
+  LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
+  SmallVector<int, 4> HiInputs;
+  std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
+               [](int M) { return M >= 0; });
+  std::sort(HiInputs.begin(), HiInputs.end());
+  HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
+  int NumLToL =
+      std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
+  int NumHToL = LoInputs.size() - NumLToL;
+  int NumLToH =
+      std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
+  int NumHToH = HiInputs.size() - NumLToH;
+  MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
+  MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
+  MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
+  MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
+
+  // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
+  // such inputs we can swap two of the dwords across the half mark and end up
+  // with <=2 inputs to each half in each half. Once there, we can fall through
+  // to the generic code below. For example:
+  //
+  // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
+  // Mask:  [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
+  //
+  // Before we had 3-1 in the low half and 3-1 in the high half. Afterward, 2-2
+  // and 2-2.
+  auto balanceSides = [&](ArrayRef<int> ThreeInputs, int OneInput,
+                          int ThreeInputHalfSum, int OneInputHalfOffset) {
+    // Compute the index of dword with only one word among the three inputs in
+    // a half by taking the sum of the half with three inputs and subtracting
+    // the sum of the actual three inputs. The difference is the remaining
+    // slot.
+    int DWordA = (ThreeInputHalfSum -
+                  std::accumulate(ThreeInputs.begin(), ThreeInputs.end(), 0)) /
+                 2;
+    int DWordB = OneInputHalfOffset / 2 + (OneInput / 2 + 1) % 2;
+
+    int PSHUFDMask[] = {0, 1, 2, 3};
+    PSHUFDMask[DWordA] = DWordB;
+    PSHUFDMask[DWordB] = DWordA;
+    V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
+                    DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
+                                DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
+                                getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
+
+    // Adjust the mask to match the new locations of A and B.
+    for (int &M : Mask)
+      if (M != -1 && M/2 == DWordA)
+        M = 2 * DWordB + M % 2;
+      else if (M != -1 && M/2 == DWordB)
+        M = 2 * DWordA + M % 2;
+
+    // Recurse back into this routine to re-compute state now that this isn't
+    // a 3 and 1 problem.
+    return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
+                                Mask);
+  };
+  if (NumLToL == 3 && NumHToL == 1)
+    return balanceSides(LToLInputs, HToLInputs[0], 0 + 1 + 2 + 3, 4);
+  else if (NumLToL == 1 && NumHToL == 3)
+    return balanceSides(HToLInputs, LToLInputs[0], 4 + 5 + 6 + 7, 0);
+  else if (NumLToH == 1 && NumHToH == 3)
+    return balanceSides(HToHInputs, LToHInputs[0], 4 + 5 + 6 + 7, 0);
+  else if (NumLToH == 3 && NumHToH == 1)
+    return balanceSides(LToHInputs, HToHInputs[0], 0 + 1 + 2 + 3, 4);
+
+  // At this point there are at most two inputs to the low and high halves from
+  // each half. That means the inputs can always be grouped into dwords and
+  // those dwords can then be moved to the correct half with a dword shuffle.
+  // We use at most one low and one high word shuffle to collect these paired
+  // inputs into dwords, and finally a dword shuffle to place them.
+  int PSHUFLMask[4] = {-1, -1, -1, -1};
+  int PSHUFHMask[4] = {-1, -1, -1, -1};
+  int PSHUFDMask[4] = {-1, -1, -1, -1};
+
+  // First fix the masks for all the inputs that are staying in their
+  // original halves. This will then dictate the targets of the cross-half
+  // shuffles.
+  auto fixInPlaceInputs = [&PSHUFDMask](
+      ArrayRef<int> InPlaceInputs, MutableArrayRef<int> SourceHalfMask,
+      MutableArrayRef<int> HalfMask, int HalfOffset) {
+    if (InPlaceInputs.empty())
+      return;
+    if (InPlaceInputs.size() == 1) {
+      SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
+          InPlaceInputs[0] - HalfOffset;
+      PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
+      return;
+    }
+
+    assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
+    SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
+        InPlaceInputs[0] - HalfOffset;
+    // Put the second input next to the first so that they are packed into
+    // a dword. We find the adjacent index by toggling the low bit.
+    int AdjIndex = InPlaceInputs[0] ^ 1;
+    SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
+    std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
+    PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
+  };
+  if (!HToLInputs.empty())
+    fixInPlaceInputs(LToLInputs, PSHUFLMask, LoMask, 0);
+  if (!LToHInputs.empty())
+    fixInPlaceInputs(HToHInputs, PSHUFHMask, HiMask, 4);
+
+  // Now gather the cross-half inputs and place them into a free dword of
+  // their target half.
+  // FIXME: This operation could almost certainly be simplified dramatically to
+  // look more like the 3-1 fixing operation.
+  auto moveInputsToRightHalf = [&PSHUFDMask](
+      MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
+      MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
+      int SourceOffset, int DestOffset) {
+    auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
+      return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
+    };
+    auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
+                                               int Word) {
+      int LowWord = Word & ~1;
+      int HighWord = Word | 1;
+      return isWordClobbered(SourceHalfMask, LowWord) ||
+             isWordClobbered(SourceHalfMask, HighWord);
+    };
+
+    if (IncomingInputs.empty())
+      return;
+
+    if (ExistingInputs.empty()) {
+      // Map any dwords with inputs from them into the right half.
+      for (int Input : IncomingInputs) {
+        // If the source half mask maps over the inputs, turn those into
+        // swaps and use the swapped lane.
+        if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
+          if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
+            SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
+                Input - SourceOffset;
+            // We have to swap the uses in our half mask in one sweep.
+            for (int &M : HalfMask)
+              if (M == SourceHalfMask[Input - SourceOffset])
+                M = Input;
+              else if (M == Input)
+                M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
+          } else {
+            assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
+                       Input - SourceOffset &&
+                   "Previous placement doesn't match!");
+          }
+          // Note that this correctly re-maps both when we do a swap and when
+          // we observe the other side of the swap above. We rely on that to
+          // avoid swapping the members of the input list directly.
+          Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
+        }
+
+        // Map the input's dword into the correct half.
+        if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
+          PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
+        else
+          assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
+                     Input / 2 &&
+                 "Previous placement doesn't match!");
+      }
+
+      // And just directly shift any other-half mask elements to be same-half
+      // as we will have mirrored the dword containing the element into the
+      // same position within that half.
+      for (int &M : HalfMask)
+        if (M >= SourceOffset && M < SourceOffset + 4) {
+          M = M - SourceOffset + DestOffset;
+          assert(M >= 0 && "This should never wrap below zero!");
+        }
+      return;
+    }
+
+    // Ensure we have the input in a viable dword of its current half. This
+    // is particularly tricky because the original position may be clobbered
+    // by inputs being moved and *staying* in that half.
+    if (IncomingInputs.size() == 1) {
+      if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
+        int InputFixed = std::find(std::begin(SourceHalfMask),
+                                   std::end(SourceHalfMask), -1) -
+                         std::begin(SourceHalfMask) + SourceOffset;
+        SourceHalfMask[InputFixed - SourceOffset] =
+            IncomingInputs[0] - SourceOffset;
+        std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
+                     InputFixed);
+        IncomingInputs[0] = InputFixed;
+      }
+    } else if (IncomingInputs.size() == 2) {
+      if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
+          isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
+        int SourceDWordBase = !isDWordClobbered(SourceHalfMask, 0) ? 0 : 2;
+        assert(!isDWordClobbered(SourceHalfMask, SourceDWordBase) &&
+               "Not all dwords can be clobbered!");
+        SourceHalfMask[SourceDWordBase] = IncomingInputs[0] - SourceOffset;
+        SourceHalfMask[SourceDWordBase + 1] = IncomingInputs[1] - SourceOffset;
+        for (int &M : HalfMask)
+          if (M == IncomingInputs[0])
+            M = SourceDWordBase + SourceOffset;
+          else if (M == IncomingInputs[1])
+            M = SourceDWordBase + 1 + SourceOffset;
+        IncomingInputs[0] = SourceDWordBase + SourceOffset;
+        IncomingInputs[1] = SourceDWordBase + 1 + SourceOffset;
+      }
+    } else {
+      llvm_unreachable("Unhandled input size!");
+    }
+
+    // Now hoist the DWord down to the right half.
+    int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
+    assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
+    PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
+    for (int Input : IncomingInputs)
+      std::replace(HalfMask.begin(), HalfMask.end(), Input,
+                   FreeDWord * 2 + Input % 2);
+  };
+  moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask,
+                        /*SourceOffset*/ 4, /*DestOffset*/ 0);
+  moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask,
+                        /*SourceOffset*/ 0, /*DestOffset*/ 4);
+
+  // Now enact all the shuffles we've computed to move the inputs into their
+  // target half.
+  if (!isNoopShuffleMask(PSHUFLMask))
+    V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
+                    getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG));
+  if (!isNoopShuffleMask(PSHUFHMask))
+    V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
+                    getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG));
+  if (!isNoopShuffleMask(PSHUFDMask))
+    V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
+                    DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
+                                DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
+                                getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
+
+  // At this point, each half should contain all its inputs, and we can then
+  // just shuffle them into their final position.
+  assert(std::count_if(LoMask.begin(), LoMask.end(), isHi) == 0 &&
+         "Failed to lift all the high half inputs to the low mask!");
+  assert(std::count_if(HiMask.begin(), HiMask.end(), isLo) == 0 &&
+         "Failed to lift all the low half inputs to the high mask!");
+
+  // Do a half shuffle for the low mask.
+  if (!isNoopShuffleMask(LoMask))
+    V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
+                    getV4X86ShuffleImm8ForMask(LoMask, DAG));
+
+  // Do a half shuffle with the high mask after shifting its values down.
+  for (int &M : HiMask)
+    if (M >= 0)
+      M -= 4;
+  if (!isNoopShuffleMask(HiMask))
+    V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
+                    getV4X86ShuffleImm8ForMask(HiMask, DAG));
+
+  return V;
+}
+
+/// \brief Detect whether the mask pattern should be lowered through
+/// interleaving.
+///
+/// This essentially tests whether viewing the mask as an interleaving of two
+/// sub-sequences reduces the cross-input traffic of a blend operation. If so,
+/// lowering it through interleaving is a significantly better strategy.
+static bool shouldLowerAsInterleaving(ArrayRef<int> Mask) {
+  int NumEvenInputs[2] = {0, 0};
+  int NumOddInputs[2] = {0, 0};
+  int NumLoInputs[2] = {0, 0};
+  int NumHiInputs[2] = {0, 0};
+  for (int i = 0, Size = Mask.size(); i < Size; ++i) {
+    if (Mask[i] < 0)
+      continue;
+
+    int InputIdx = Mask[i] >= Size;
+
+    if (i < Size / 2)
+      ++NumLoInputs[InputIdx];
+    else
+      ++NumHiInputs[InputIdx];
+
+    if ((i % 2) == 0)
+      ++NumEvenInputs[InputIdx];
+    else
+      ++NumOddInputs[InputIdx];
+  }
+
+  // The minimum number of cross-input results for both the interleaved and
+  // split cases. If interleaving results in fewer cross-input results, return
+  // true.
+  int InterleavedCrosses = std::min(NumEvenInputs[1] + NumOddInputs[0],
+                                    NumEvenInputs[0] + NumOddInputs[1]);
+  int SplitCrosses = std::min(NumLoInputs[1] + NumHiInputs[0],
+                              NumLoInputs[0] + NumHiInputs[1]);
+  return InterleavedCrosses < SplitCrosses;
+}
+
+/// \brief Blend two v8i16 vectors using a naive unpack strategy.
+///
+/// This strategy only works when the inputs from each vector fit into a single
+/// half of that vector, and generally there are not so many inputs as to leave
+/// the in-place shuffles required highly constrained (and thus expensive). It
+/// shifts all the inputs into a single side of both input vectors and then
+/// uses an unpack to interleave these inputs in a single vector. At that
+/// point, we will fall back on the generic single input shuffle lowering.
+static SDValue lowerV8I16BasicBlendVectorShuffle(SDLoc DL, SDValue V1,
+                                                 SDValue V2,
+                                                 MutableArrayRef<int> Mask,
+                                                 const X86Subtarget *Subtarget,
+                                                 SelectionDAG &DAG) {
+  assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
+  assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
+  SmallVector<int, 3> LoV1Inputs, HiV1Inputs, LoV2Inputs, HiV2Inputs;
+  for (int i = 0; i < 8; ++i)
+    if (Mask[i] >= 0 && Mask[i] < 4)
+      LoV1Inputs.push_back(i);
+    else if (Mask[i] >= 4 && Mask[i] < 8)
+      HiV1Inputs.push_back(i);
+    else if (Mask[i] >= 8 && Mask[i] < 12)
+      LoV2Inputs.push_back(i);
+    else if (Mask[i] >= 12)
+      HiV2Inputs.push_back(i);
+
+  int NumV1Inputs = LoV1Inputs.size() + HiV1Inputs.size();
+  int NumV2Inputs = LoV2Inputs.size() + HiV2Inputs.size();
+
+  assert(NumV1Inputs > 0 && NumV1Inputs <= 3 && "At most 3 inputs supported");
+  assert(NumV2Inputs > 0 && NumV2Inputs <= 3 && "At most 3 inputs supported");
+  assert(NumV1Inputs + NumV2Inputs <= 4 && "At most 4 combined inputs");
+
+  bool MergeFromLo = LoV1Inputs.size() + LoV2Inputs.size() >=
+                     HiV1Inputs.size() + HiV2Inputs.size();
+
+  auto moveInputsToHalf = [&](SDValue V, ArrayRef<int> LoInputs,
+                              ArrayRef<int> HiInputs, bool MoveToLo,
+                              int MaskOffset) {
+    ArrayRef<int> GoodInputs = MoveToLo ? LoInputs : HiInputs;
+    ArrayRef<int> BadInputs = MoveToLo ? HiInputs : LoInputs;
+    if (BadInputs.empty())
+      return V;
+
+    int MoveMask[] = {-1, -1, -1, -1, -1, -1, -1, -1};
+    int MoveOffset = MoveToLo ? 0 : 4;
+
+    if (GoodInputs.empty()) {
+      for (int BadInput : BadInputs) {
+        MoveMask[Mask[BadInput] % 4 + MoveOffset] = Mask[BadInput] - MaskOffset;
+        Mask[BadInput] = Mask[BadInput] % 4 + MoveOffset + MaskOffset;
+      }
+    } else {
+      if (GoodInputs.size() == 2) {
+        // If the low inputs are spread across two dwords, pack them into
+        // a single dword.
+        MoveMask[Mask[GoodInputs[0]] % 2 + MoveOffset] =
+            Mask[GoodInputs[0]] - MaskOffset;
+        MoveMask[Mask[GoodInputs[1]] % 2 + MoveOffset] =
+            Mask[GoodInputs[1]] - MaskOffset;
+        Mask[GoodInputs[0]] = Mask[GoodInputs[0]] % 2 + MoveOffset + MaskOffset;
+        Mask[GoodInputs[1]] = Mask[GoodInputs[0]] % 2 + MoveOffset + MaskOffset;
+      } else {
+        // Otherwise pin the low inputs.
+        for (int GoodInput : GoodInputs)
+          MoveMask[Mask[GoodInput]] = Mask[GoodInput] - MaskOffset;
+      }
+
+      int MoveMaskIdx =
+          std::find(std::begin(MoveMask) + MoveOffset, std::end(MoveMask), -1) -
+          std::begin(MoveMask);
+      assert(MoveMaskIdx >= MoveOffset && "Established above");
+
+      if (BadInputs.size() == 2) {
+        assert(MoveMask[MoveMaskIdx] == -1 && "Expected empty slot");
+        assert(MoveMask[MoveMaskIdx + 1] == -1 && "Expected empty slot");
+        MoveMask[MoveMaskIdx + Mask[BadInputs[0]] % 2] =
+            Mask[BadInputs[0]] - MaskOffset;
+        MoveMask[MoveMaskIdx + Mask[BadInputs[1]] % 2] =
+            Mask[BadInputs[1]] - MaskOffset;
+        Mask[BadInputs[0]] = MoveMaskIdx + Mask[BadInputs[0]] % 2 + MaskOffset;
+        Mask[BadInputs[1]] = MoveMaskIdx + Mask[BadInputs[1]] % 2 + MaskOffset;
+      } else {
+        assert(BadInputs.size() == 1 && "All sizes handled");
+        MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
+        Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
+      }
+    }
+
+    return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
+                                MoveMask);
+  };
+  V1 = moveInputsToHalf(V1, LoV1Inputs, HiV1Inputs, MergeFromLo,
+                        /*MaskOffset*/ 0);
+  V2 = moveInputsToHalf(V2, LoV2Inputs, HiV2Inputs, MergeFromLo,
+                        /*MaskOffset*/ 8);
+
+  // FIXME: Select an interleaving of the merge of V1 and V2 that minimizes
+  // cross-half traffic in the final shuffle.
+
+  // Munge the mask to be a single-input mask after the unpack merges the
+  // results.
+  for (int &M : Mask)
+    if (M != -1)
+      M = 2 * (M % 4) + (M / 8);
+
+  return DAG.getVectorShuffle(
+      MVT::v8i16, DL, DAG.getNode(MergeFromLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
+                                  DL, MVT::v8i16, V1, V2),
+      DAG.getUNDEF(MVT::v8i16), Mask);
+}
+
+/// \brief Generic lowering of 8-lane i16 shuffles.
+///
+/// This handles both single-input shuffles and combined shuffle/blends with
+/// two inputs. The single input shuffles are immediately delegated to
+/// a dedicated lowering routine.
+///
+/// The blends are lowered in one of three fundamental ways. If there are few
+/// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
+/// of the input is significantly cheaper when lowered as an interleaving of
+/// the two inputs, try to interleave them. Otherwise, blend the low and high
+/// halves of the inputs separately (making them have relatively few inputs)
+/// and then concatenate them.
+static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+                                       const X86Subtarget *Subtarget,
+                                       SelectionDAG &DAG) {
+  SDLoc DL(Op);
+  assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
+  assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
+  assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
+  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+  ArrayRef<int> OrigMask = SVOp->getMask();
+  int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
+                        OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
+  MutableArrayRef<int> Mask(MaskStorage);
+
+  assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
+
+  int Size = Mask.size();
+  assert(Size == 8 && "Unexpected mask size for v8 shuffle!");
+
+  auto isV1 = [](int M) { return M >= 0 && M < 8; };
+  auto isV2 = [](int M) { return M >= 8; };
+
+  int NumV1Inputs = std::count_if(Mask.begin(), Mask.end(), isV1);
+  int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
+
+  if (NumV2Inputs == 0)
+    return lowerV8I16SingleInputVectorShuffle(DL, V1, Mask, Subtarget, DAG);
+
+  assert(NumV1Inputs > 0 && "All single-input shuffles should be canonicalized "
+                            "to be V1-input shuffles.");
+
+  if (NumV1Inputs + NumV2Inputs <= 4)
+    return lowerV8I16BasicBlendVectorShuffle(DL, V1, V2, Mask, Subtarget, DAG);
+
+  // Check whether an interleaving lowering is likely to be more efficient.
+  // This isn't perfect but it is a strong heuristic that tends to work well on
+  // the kinds of shuffles that show up in practice.
+  //
+  // FIXME: Handle 1x, 2x, and 4x interleaving.
+  if (shouldLowerAsInterleaving(Mask)) {
+    // FIXME: Figure out whether we should pack these into the low or high
+    // halves.
+
+    int EMask[8], OMask[8];
+    for (int i = 0; i < 4; ++i) {
+      EMask[i] = Mask[2*i];
+      OMask[i] = Mask[2*i + 1];
+      EMask[i + 4] = -1;
+      OMask[i + 4] = -1;
+    }
+
+    SDValue Evens = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, EMask);
+    SDValue Odds = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, OMask);
+
+    return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, Evens, Odds);
+  }
+
+  int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
+  int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
+
+  for (int i = 0; i < 4; ++i) {
+    LoBlendMask[i] = Mask[i];
+    HiBlendMask[i] = Mask[i + 4];
+  }
+
+  SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask);
+  SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask);
+  LoV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, LoV);
+  HiV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, HiV);
+
+  return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
+                     DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, LoV, HiV));
+}
+
+/// \brief Generic lowering of v16i8 shuffles.
+///
+/// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
+/// detect any complexity reducing interleaving. If that doesn't help, it uses
+/// UNPCK to spread the i8 elements across two i16-element vectors, and uses
+/// the existing lowering for v8i16 blends on each half, finally PACK-ing them
+/// back together.
+static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+                                       const X86Subtarget *Subtarget,
+                                       SelectionDAG &DAG) {
+  SDLoc DL(Op);
+  assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
+  assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
+  assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
+  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+  ArrayRef<int> OrigMask = SVOp->getMask();
+  assert(OrigMask.size() == 16 && "Unexpected mask size for v16 shuffle!");
+  int MaskStorage[16] = {
+      OrigMask[0],  OrigMask[1],  OrigMask[2],  OrigMask[3],
+      OrigMask[4],  OrigMask[5],  OrigMask[6],  OrigMask[7],
+      OrigMask[8],  OrigMask[9],  OrigMask[10], OrigMask[11],
+      OrigMask[12], OrigMask[13], OrigMask[14], OrigMask[15]};
+  MutableArrayRef<int> Mask(MaskStorage);
+  MutableArrayRef<int> LoMask = Mask.slice(0, 8);
+  MutableArrayRef<int> HiMask = Mask.slice(8, 8);
+
+  // Check whether an interleaving lowering is likely to be more efficient.
+  // This isn't perfect but it is a strong heuristic that tends to work well on
+  // the kinds of shuffles that show up in practice.
+  //
+  // FIXME: We need to handle other interleaving widths (i16, i32, ...).
+  if (shouldLowerAsInterleaving(Mask)) {
+    // FIXME: Figure out whether we should pack these into the low or high
+    // halves.
+
+    int EMask[16], OMask[16];
+    for (int i = 0; i < 8; ++i) {
+      EMask[i] = Mask[2*i];
+      OMask[i] = Mask[2*i + 1];
+      EMask[i + 8] = -1;
+      OMask[i + 8] = -1;
+    }
+
+    SDValue Evens = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, EMask);
+    SDValue Odds = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, OMask);
+
+    return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, Evens, Odds);
+  }
+
+  SDValue LoV1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
+                             DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1,
+                                         DAG.getUNDEF(MVT::v8i16)));
+  SDValue HiV1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
+                             DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1,
+                                         DAG.getUNDEF(MVT::v8i16)));
+  SDValue LoV2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
+                             DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V2,
+                                         DAG.getUNDEF(MVT::v8i16)));
+  SDValue HiV2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
+                             DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V2,
+                                         DAG.getUNDEF(MVT::v8i16)));
+
+  int V1LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
+  int V1HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
+  int V2LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
+  int V2HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
+
+  auto buildBlendMasks = [](MutableArrayRef<int> HalfMask,
+                            MutableArrayRef<int> V1HalfBlendMask,
+                            MutableArrayRef<int> V2HalfBlendMask) {
+    for (int i = 0; i < 8; ++i)
+      if (HalfMask[i] >= 0 && HalfMask[i] < 16) {
+        V1HalfBlendMask[i] = HalfMask[i];
+        HalfMask[i] = i;
+      } else if (HalfMask[i] >= 16) {
+        V2HalfBlendMask[i] = HalfMask[i] - 16;
+        HalfMask[i] = i + 8;
+      }
+  };
+  buildBlendMasks(LoMask, V1LoBlendMask, V2LoBlendMask);
+  buildBlendMasks(HiMask, V1HiBlendMask, V2HiBlendMask);
+
+  SDValue V1Lo = DAG.getVectorShuffle(MVT::v8i16, DL, LoV1, HiV1, V1LoBlendMask);
+  SDValue V2Lo = DAG.getVectorShuffle(MVT::v8i16, DL, LoV2, HiV2, V2LoBlendMask);
+  SDValue V1Hi = DAG.getVectorShuffle(MVT::v8i16, DL, LoV1, HiV1, V1HiBlendMask);
+  SDValue V2Hi = DAG.getVectorShuffle(MVT::v8i16, DL, LoV2, HiV2, V2HiBlendMask);
+
+  SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Lo, V2Lo, LoMask);
+  SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Hi, V2Hi, HiMask);
+
+  return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
+}
+
+/// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
+///
+/// This routine breaks down the specific type of 128-bit shuffle and
+/// dispatches to the lowering routines accordingly.
+static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+                                        MVT VT, const X86Subtarget *Subtarget,
+                                        SelectionDAG &DAG) {
+  switch (VT.SimpleTy) {
+  case MVT::v2i64:
+    return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
+  case MVT::v2f64:
+    return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
+  case MVT::v4i32:
+    return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
+  case MVT::v4f32:
+    return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
+  case MVT::v8i16:
+    return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
+  case MVT::v16i8:
+    return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
+
+  default:
+    llvm_unreachable("Unimplemented!");
+  }
+}
+
+/// \brief Tiny helper function to test whether adjacent masks are sequential.
+static bool areAdjacentMasksSequential(ArrayRef<int> Mask) {
+  for (int i = 0, Size = Mask.size(); i < Size; i += 2)
+    if (Mask[i] + 1 != Mask[i+1])
+      return false;
+
+  return true;
+}
+
+/// \brief Top-level lowering for x86 vector shuffles.
+///
+/// This handles decomposition, canonicalization, and lowering of all x86
+/// vector shuffles. Most of the specific lowering strategies are encapsulated
+/// above in helper routines. The canonicalization attempts to widen shuffles
+/// to involve fewer lanes of wider elements, consolidate symmetric patterns
+/// s.t. only one of the two inputs needs to be tested, etc.
+static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
+                                  SelectionDAG &DAG) {
+  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+  ArrayRef<int> Mask = SVOp->getMask();
+  SDValue V1 = Op.getOperand(0);
+  SDValue V2 = Op.getOperand(1);
+  MVT VT = Op.getSimpleValueType();
+  int NumElements = VT.getVectorNumElements();
+  SDLoc dl(Op);
+
+  assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
+
+  bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
+  bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
+  if (V1IsUndef && V2IsUndef)
+    return DAG.getUNDEF(VT);
+
+  // When we create a shuffle node we put the UNDEF node to second operand,
+  // but in some cases the first operand may be transformed to UNDEF.
+  // In this case we should just commute the node.
+  if (V1IsUndef)
+    return CommuteVectorShuffle(SVOp, DAG);
+
+  // Check for non-undef masks pointing at an undef vector and make the masks
+  // undef as well. This makes it easier to match the shuffle based solely on
+  // the mask.
+  if (V2IsUndef)
+    for (int M : Mask)
+      if (M >= NumElements) {
+        SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
+        for (int &M : NewMask)
+          if (M >= NumElements)
+            M = -1;
+        return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
+      }
+
+  // For integer vector shuffles, try to collapse them into a shuffle of fewer
+  // lanes but wider integers. We cap this to not form integers larger than i64
+  // but it might be interesting to form i128 integers to handle flipping the
+  // low and high halves of AVX 256-bit vectors.
+  if (VT.isInteger() && VT.getScalarSizeInBits() < 64 &&
+      areAdjacentMasksSequential(Mask)) {
+    SmallVector<int, 8> NewMask;
+    for (int i = 0, Size = Mask.size(); i < Size; i += 2)
+      NewMask.push_back(Mask[i] / 2);
+    MVT NewVT =
+        MVT::getVectorVT(MVT::getIntegerVT(VT.getScalarSizeInBits() * 2),
+                         VT.getVectorNumElements() / 2);
+    V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
+    V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
+    return DAG.getNode(ISD::BITCAST, dl, VT,
+                       DAG.getVectorShuffle(NewVT, dl, V1, V2, NewMask));
+  }
+
+  int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
+  for (int M : SVOp->getMask())
+    if (M < 0)
+      ++NumUndefElements;
+    else if (M < NumElements)
+      ++NumV1Elements;
+    else
+      ++NumV2Elements;
+
+  // Commute the shuffle as needed such that more elements come from V1 than
+  // V2. This allows us to match the shuffle pattern strictly on how many
+  // elements come from V1 without handling the symmetric cases.
+  if (NumV2Elements > NumV1Elements)
+    return CommuteVectorShuffle(SVOp, DAG);
+
+  // When the number of V1 and V2 elements are the same, try to minimize the
+  // number of uses of V2 in the low half of the vector.
+  if (NumV1Elements == NumV2Elements) {
+    int LowV1Elements = 0, LowV2Elements = 0;
+    for (int M : SVOp->getMask().slice(0, NumElements / 2))
+      if (M >= NumElements)
+        ++LowV2Elements;
+      else if (M >= 0)
+        ++LowV1Elements;
+    if (LowV2Elements > LowV1Elements)
+      return CommuteVectorShuffle(SVOp, DAG);
+  }
+
+  // For each vector width, delegate to a specialized lowering routine.
+  if (VT.getSizeInBits() == 128)
+    return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
+
+  llvm_unreachable("Unimplemented!");
+}
+
+
+//===----------------------------------------------------------------------===//
+// Legacy vector shuffle lowering
+//
+// This code is the legacy code handling vector shuffles until the above
+// replaces its functionality and performance.
+//===----------------------------------------------------------------------===//
+
 static bool isBlendMask(ArrayRef<int> MaskVals, MVT VT, bool hasSSE41,
                         bool hasInt256, unsigned *MaskOut = nullptr) {
   MVT EltVT = VT.getVectorElementType();
@@ -8127,6 +9151,11 @@ X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
   bool OptForSize = MF.getFunction()->getAttributes().
     hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
 
+  // Check if we should use the experimental vector shuffle lowering. If so,
+  // delegate completely to that code path.
+  if (ExperimentalVectorShuffleLowering)
+    return lowerVectorShuffle(Op, Subtarget, DAG);
+
   assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
 
   if (V1IsUndef && V2IsUndef)