[X86][SSE] Attempt to match OP(SHUFFLE(X,Y),SHUFFLE(X,Y)) -> SHUFFLE(HOP(X,Y))

[RKSimon]

Summary

An initial backend patch towards fixing the various poor HADD combines (PR34724, PR41813, PR45747 etc.).

This extends isHorizontalBinOp to check if we have per-element horizontal ops (odd+even element pairs), but not in the expected serial order - in which case we build a "post shuffle mask" that we can apply to the HOP result, assuming we have fast-hops/optsize etc.

The next step will be to extend the SHUFFLE(HOP(X,Y)) combines as suggested on PR41813 - accepting more post-shuffle masks even on slow-hop targets if we can fold it into another shuffle.

[RKSimon]

clang-format

[spatel]

Need to add more tests to make sure integer and 256-bit work as expected.

What do you think about consolidating the h-op creation in 1 helper as a preliminary to avoid the caller code duplication? We can also more easily assert properties of the shuffle mask (eg, only chooses from operand 0?).

I had this:

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 450927aaf5c..3ed354d483f 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -44507,21 +44507,66 @@ static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, SelectionDAG &DAG,
   return true;
 }
 
+/// Try to synthesize horizontal add/sub from adds/subs of shuffles.
+static SDValue getHorizontalBinop(SDNode *N, SelectionDAG &DAG,
+                                  const X86Subtarget &Subtarget) {
+  unsigned Opcode = N->getOpcode();
+  unsigned HorizOpcode;
+  switch (Opcode) {
+  case ISD::ADD: HorizOpcode = X86ISD::HADD; break;
+  case ISD::SUB: HorizOpcode = X86ISD::HSUB; break;
+  case ISD::FADD: HorizOpcode = X86ISD::FHADD; break;
+  case ISD::FSUB: HorizOpcode = X86ISD::FHSUB; break;
+  default:
+    llvm_unreachable("Unexpected opcode for horizontal op");
+  }
+
+  EVT VT = N->getValueType(0);
+  if (!VT.isSimple())
+    return SDValue();
+  switch (VT.getSimpleVT().SimpleTy) {
+  case MVT::v8i16:
+  case MVT::v4i32:
+  case MVT::v16i16:
+  case MVT::v8i32:
+    // 256-bit vectors without AVX2 are handled by splitting below.
+    if (!Subtarget.hasSSSE3())
+      return SDValue();
+    break;
+  case MVT::v4f32:
+  case MVT::v2f64:
+    if (!Subtarget.hasSSE3())
+      return SDValue();
+    break;
+  case MVT::v8f32:
+  case MVT::v4f64:
+    if (!Subtarget.hasAVX())
+      return SDValue();
+    break;
+  default:
+    return SDValue();
+  }
+
+  SDValue Op0 = N->getOperand(0), Op1 = N->getOperand(1);
+  bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
+  if (!isHorizontalBinOp(Op0, Op1, DAG, Subtarget, IsCommutable))
+    return SDValue();
+
+  if (VT.getScalarType().isFloatingPoint())
+    return DAG.getNode(HorizOpcode, SDLoc(N), VT, Op0, Op1);
+
+  auto HopBuilder = [&](SelectionDAG &DAG, const SDLoc &DL,
+                        ArrayRef<SDValue> Ops) {
+    return DAG.getNode(HorizOpcode, DL, Ops[0].getValueType(), Ops);
+  };
+  return SplitOpsAndApply(DAG, Subtarget, SDLoc(N), VT, {Op0, Op1}, HopBuilder);
+}
+
 /// Do target-specific dag combines on floating-point adds/subs.
 static SDValue combineFaddFsub(SDNode *N, SelectionDAG &DAG,
                                const X86Subtarget &Subtarget) {

[RKSimon]

rebase

[RKSimon]

rebase with new tests

[RKSimon]

Relax lane crossing limitations for AVX2+ targets

[RKSimon]

Always permit lane crossing on AVX2 targets, and on non-AVX2 targets, integer hops will always be split to 128-bits so permit lane crossing there as well.

[spatel]

llvm/lib/Target/X86/X86ISelLowering.cpp

44470

Reduce negative logic?
// Avoid 128-bit lane crossing if this is pre-AVX2 and FP (integer will be split).
if (!Subtarget.hasAVX2 && VT.isFloatingPoint() && ...)

[RKSimon]

simplify lane-crossing logic

[spatel]

LGTM

llvm/lib/Target/X86/X86ISelLowering.cpp

 /// and
 ///   B = < float b0, float b1, float b2, float b3 >
 /// then the result of doing a horizontal operation on A and B is
 ///   A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
 /// A horizontal-op B, for some already available A and B, and if so then LHS is
 /// set to A, RHS to B, and the routine returns 'true'.
 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, SelectionDAG &DAG,
-                              const X86Subtarget &Subtarget,
-                              bool IsCommutative) {
+                              const X86Subtarget &Subtarget, bool IsCommutative,
+                              SmallVectorImpl<int> &PostShuffleMask) {
   // If either operand is undef, bail out. The binop should be simplified.
   if (LHS.isUndef() || RHS.isUndef())
     return false;
 
   // Look for the following pattern:
   //   A = < float a0, float a1, float a2, float a3 >
   //   B = < float b0, float b1, float b2, float b3 >
   // and
static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, SelectionDAG &DAG,
   }
 
   if (RMask.empty()) {
     C = RHS;
     for (unsigned i = 0; i != NumElts; ++i)
       RMask.push_back(i);
   }
 
+  // Avoid 128-bit lane crossing if pre-AVX2 and FP (integer will split).
+  if (!Subtarget.hasAVX2() && VT.isFloatingPoint() &&
+      (isLaneCrossingShuffleMask(128, VT.getScalarSizeInBits(), LMask) ||
+       isLaneCrossingShuffleMask(128, VT.getScalarSizeInBits(), RMask)))
+    return false;
+
   // If A and B occur in reverse order in RHS, then canonicalize by commuting
   // RHS operands and shuffle mask.
   if (A != C) {
     std::swap(C, D);
     ShuffleVectorSDNode::commuteMask(RMask);
   }
   // Check that the shuffles are both shuffling the same vectors.
   if (!(A == C && B == D))
     return false;
 
+  PostShuffleMask.clear();
+  PostShuffleMask.append(NumElts, SM_SentinelUndef);
+
   // LHS and RHS are now:
   //   LHS = shuffle A, B, LMask
   //   RHS = shuffle A, B, RMask
   // Check that the masks correspond to performing a horizontal operation.
   // AVX defines horizontal add/sub to operate independently on 128-bit lanes,
   // so we just repeat the inner loop if this is a 256-bit op.
   unsigned Num128BitChunks = VT.getSizeInBits() / 128;
   unsigned NumEltsPer128BitChunk = NumElts / Num128BitChunks;
+  unsigned NumEltsPer64BitChunk = NumEltsPer128BitChunk / 2;
   assert((NumEltsPer128BitChunk % 2 == 0) &&
          "Vector type should have an even number of elements in each lane");
   for (unsigned j = 0; j != NumElts; j += NumEltsPer128BitChunk) {
     for (unsigned i = 0; i != NumEltsPer128BitChunk; ++i) {
       // Ignore undefined components.
       int LIdx = LMask[i + j], RIdx = RMask[i + j];
       if (LIdx < 0 || RIdx < 0 ||
           (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
           (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
         continue;
 
+      // Check that successive odd/even elements are being operated on. If not,
+      // this is not a horizontal operation.
+      if (!((RIdx & 1) == 1 && (LIdx + 1) == RIdx) &&
+          !((LIdx & 1) == 1 && (RIdx + 1) == LIdx && IsCommutative))
+        return false;
+
+      // Compute the post-shuffle mask index based on where the element
+      // is stored in the HOP result, and where it needs to be moved to.
+      int Base = LIdx & ~1u;
+      int Index = ((Base % NumEltsPer128BitChunk) / 2) +
+                  ((Base % NumElts) & ~(NumEltsPer128BitChunk - 1));
+
       // The  low half of the 128-bit result must choose from A.
       // The high half of the 128-bit result must choose from B,
       // unless B is undef. In that case, we are always choosing from A.
-      unsigned NumEltsPer64BitChunk = NumEltsPer128BitChunk / 2;
-      unsigned Src = B.getNode() ? i >= NumEltsPer64BitChunk : 0;
-
-      // Check that successive elements are being operated on. If not, this is
-      // not a horizontal operation.
-      int Index = 2 * (i % NumEltsPer64BitChunk) + NumElts * Src + j;
-      if (!(LIdx == Index && RIdx == Index + 1) &&
-          !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
-        return false;
+      if ((B && Base >= (int)NumElts) || (!B && i >= NumEltsPer64BitChunk))
+        Index += NumEltsPer64BitChunk;
+      PostShuffleMask[i + j] = Index;
     }
   }
 
   LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
   RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
 
-  if (!shouldUseHorizontalOp(LHS == RHS && NumShuffles < 2, DAG, Subtarget))
+  bool IsIdentityPostShuffle =
+      isSequentialOrUndefInRange(PostShuffleMask, 0, NumElts, 0);
+  if (IsIdentityPostShuffle)
+    PostShuffleMask.clear();
+
+  // Assume a SingleSource HOP if we only shuffle one input and don't need to
+  // shuffle the result.
+  if (!shouldUseHorizontalOp(LHS == RHS &&
+                                 (NumShuffles < 2 || !IsIdentityPostShuffle),
+                             DAG, Subtarget))
     return false;
 
   LHS = DAG.getBitcast(VT, LHS);
   RHS = DAG.getBitcast(VT, RHS);
   return true;
 }
 
 /// Do target-specific dag combines on floating-point adds/subs.
 static SDValue combineFaddFsub(SDNode *N, SelectionDAG &DAG,
                                const X86Subtarget &Subtarget) {
   EVT VT = N->getValueType(0);
   SDValue LHS = N->getOperand(0);
   SDValue RHS = N->getOperand(1);
   bool IsFadd = N->getOpcode() == ISD::FADD;
   auto HorizOpcode = IsFadd ? X86ISD::FHADD : X86ISD::FHSUB;
   assert((IsFadd || N->getOpcode() == ISD::FSUB) && "Wrong opcode");
 
   // Try to synthesize horizontal add/sub from adds/subs of shuffles.
+  SmallVector<int, 8> PostShuffleMask;
   if (((Subtarget.hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
        (Subtarget.hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
-      isHorizontalBinOp(LHS, RHS, DAG, Subtarget, IsFadd))
-    return DAG.getNode(HorizOpcode, SDLoc(N), VT, LHS, RHS);
+      isHorizontalBinOp(LHS, RHS, DAG, Subtarget, IsFadd, PostShuffleMask)) {
+    SDValue HorizBinOp = DAG.getNode(HorizOpcode, SDLoc(N), VT, LHS, RHS);
+    if (!PostShuffleMask.empty())
+      HorizBinOp = DAG.getVectorShuffle(VT, SDLoc(HorizBinOp), HorizBinOp,
+                                        DAG.getUNDEF(VT), PostShuffleMask);
+    return HorizBinOp;
+  }
 
   // NOTE: isHorizontalBinOp may have changed LHS/RHS variables.
 
   return SDValue();
 }
 
 /// Attempt to pre-truncate inputs to arithmetic ops if it will simplify
 /// the codegen.

 static SDValue combineAddOrSubToHADDorHSUB(SDNode *N, SelectionDAG &DAG,
                                            const X86Subtarget &Subtarget) {
   EVT VT = N->getValueType(0);
   SDValue Op0 = N->getOperand(0);
   SDValue Op1 = N->getOperand(1);
   bool IsAdd = N->getOpcode() == ISD::ADD;
   assert((IsAdd || N->getOpcode() == ISD::SUB) && "Wrong opcode");
 
+  SmallVector<int, 8> PostShuffleMask;
   if ((VT == MVT::v8i16 || VT == MVT::v4i32 || VT == MVT::v16i16 ||
        VT == MVT::v8i32) &&
       Subtarget.hasSSSE3() &&
-      isHorizontalBinOp(Op0, Op1, DAG, Subtarget, IsAdd)) {
+      isHorizontalBinOp(Op0, Op1, DAG, Subtarget, IsAdd, PostShuffleMask)) {
     auto HOpBuilder = [IsAdd](SelectionDAG &DAG, const SDLoc &DL,
                               ArrayRef<SDValue> Ops) {
-      return DAG.getNode(IsAdd ? X86ISD::HADD : X86ISD::HSUB,
-                         DL, Ops[0].getValueType(), Ops);
+      return DAG.getNode(IsAdd ? X86ISD::HADD : X86ISD::HSUB, DL,
+                         Ops[0].getValueType(), Ops);
     };
-    return SplitOpsAndApply(DAG, Subtarget, SDLoc(N), VT, {Op0, Op1},
-                            HOpBuilder);
+    SDValue HorizBinOp =
+        SplitOpsAndApply(DAG, Subtarget, SDLoc(N), VT, {Op0, Op1}, HOpBuilder);
+    if (!PostShuffleMask.empty())
+      HorizBinOp = DAG.getVectorShuffle(VT, SDLoc(HorizBinOp), HorizBinOp,
+                                        DAG.getUNDEF(VT), PostShuffleMask);
+    return HorizBinOp;
   }
 
   return SDValue();
 }
 
 static SDValue combineAdd(SDNode *N, SelectionDAG &DAG,
                           TargetLowering::DAGCombinerInfo &DCI,
                           const X86Subtarget &Subtarget) {

llvm/test/CodeGen/X86/haddsub-3.ll

 ; SSE2-NEXT:    movaps %xmm0, %xmm1
 ; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,1],xmm0[3,3]
 ; SSE2-NEXT:    addps %xmm0, %xmm1
 ; SSE2-NEXT:    movaps %xmm1, %xmm0
 ; SSE2-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
 ; SSE2-NEXT:    addss %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSSE3-LABEL: pr26491:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    addps %xmm0, %xmm1
-; SSSE3-NEXT:    movaps %xmm1, %xmm0
-; SSSE3-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
-; SSSE3-NEXT:    addss %xmm1, %xmm0
-; SSSE3-NEXT:    retq
-;
-; AVX-LABEL: pr26491:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vaddps %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-NEXT:    vaddss %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    retq
+; SSSE3-SLOW-LABEL: pr26491:
+; SSSE3-SLOW:       # %bb.0:
+; SSSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSSE3-SLOW-NEXT:    addps %xmm0, %xmm1
+; SSSE3-SLOW-NEXT:    movaps %xmm1, %xmm0
+; SSSE3-SLOW-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
+; SSSE3-SLOW-NEXT:    addss %xmm1, %xmm0
+; SSSE3-SLOW-NEXT:    retq
+;
+; SSSE3-FAST-LABEL: pr26491:
+; SSSE3-FAST:       # %bb.0:
+; SSSE3-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSSE3-FAST-NEXT:    movaps %xmm0, %xmm1
+; SSSE3-FAST-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,1],xmm0[2,3]
+; SSSE3-FAST-NEXT:    addss %xmm0, %xmm1
+; SSSE3-FAST-NEXT:    movaps %xmm1, %xmm0
+; SSSE3-FAST-NEXT:    retq
+;
+; AVX1-SLOW-LABEL: pr26491:
+; AVX1-SLOW:       # %bb.0:
+; AVX1-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX1-SLOW-NEXT:    vaddps %xmm0, %xmm1, %xmm0
+; AVX1-SLOW-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX1-SLOW-NEXT:    vaddss %xmm0, %xmm1, %xmm0
+; AVX1-SLOW-NEXT:    retq
+;
+; AVX1-FAST-LABEL: pr26491:
+; AVX1-FAST:       # %bb.0:
+; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vpermilps {{.*#+}} xmm1 = xmm0[3,1,2,3]
+; AVX1-FAST-NEXT:    vaddss %xmm0, %xmm1, %xmm0
+; AVX1-FAST-NEXT:    retq
+;
+; AVX2-LABEL: pr26491:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX2-NEXT:    vaddps %xmm0, %xmm1, %xmm0
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX2-NEXT:    vaddss %xmm0, %xmm1, %xmm0
+; AVX2-NEXT:    retq
   %1 = shufflevector <4 x float> %a0, <4 x float> undef, <4 x i32> <i32 1, i32 1, i32 3, i32 3>
   %2 = fadd <4 x float> %1, %a0
   %3 = extractelement <4 x float> %2, i32 2
   %4 = extractelement <4 x float> %2, i32 0
   %5 = fadd float %3, %4
   ret float %5
 }
 

llvm/test/CodeGen/X86/haddsub-4.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+ssse3           | FileCheck %s --check-prefixes=SSE,SSSE3,SSSE3-SLOW
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+ssse3,fast-hops | FileCheck %s --check-prefixes=SSE,SSSE3,SSSE3-FAST
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx             | FileCheck %s --check-prefixes=AVX,AVX-SLOW,AVX1,AVX1-SLOW
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx,fast-hops   | FileCheck %s --check-prefixes=AVX,AVX-FAST,AVX1,AVX1-FAST
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2            | FileCheck %s --check-prefixes=AVX,AVX-SLOW,AVX2,AVX2-SLOW
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2,fast-hops  | FileCheck %s --check-prefixes=AVX,AVX-FAST,AVX2,AVX2-FAST
 
 define <8 x i16> @hadd_reverse_v8i16(<8 x i16> %a0, <8 x i16> %a1) nounwind {
 ; SSE-LABEL: hadd_reverse_v8i16:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [14,15,10,11,6,7,2,3,14,15,10,11,12,13,14,15]
-; SSE-NEXT:    movdqa %xmm1, %xmm3
-; SSE-NEXT:    pshufb %xmm2, %xmm3
-; SSE-NEXT:    movdqa %xmm0, %xmm4
-; SSE-NEXT:    pshufb %xmm2, %xmm4
-; SSE-NEXT:    punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm3[0]
-; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [12,13,8,9,4,5,0,1,8,9,12,13,12,13,14,15]
-; SSE-NEXT:    pshufb %xmm2, %xmm1
-; SSE-NEXT:    pshufb %xmm2, %xmm0
-; SSE-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE-NEXT:    paddw %xmm4, %xmm0
+; SSE-NEXT:    phaddw %xmm1, %xmm0
+; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
+; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: hadd_reverse_v8i16:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vmovdqa {{.*#+}} xmm2 = [14,15,10,11,6,7,2,3,14,15,10,11,12,13,14,15]
-; AVX-NEXT:    vpshufb %xmm2, %xmm1, %xmm3
-; AVX-NEXT:    vpshufb %xmm2, %xmm0, %xmm2
-; AVX-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
-; AVX-NEXT:    vmovdqa {{.*#+}} xmm3 = [12,13,8,9,4,5,0,1,8,9,12,13,12,13,14,15]
-; AVX-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
-; AVX-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
-; AVX-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX-NEXT:    vpaddw %xmm0, %xmm2, %xmm0
+; AVX-NEXT:    vphaddw %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
+; AVX-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
 ; AVX-NEXT:    retq
   %lhs = shufflevector <8 x i16> %a0, <8 x i16> %a1, <8 x i32> <i32 7, i32 5, i32 3, i32 1, i32 15, i32 13, i32 11, i32 9>
   %rhs = shufflevector <8 x i16> %a0, <8 x i16> %a1, <8 x i32> <i32 6, i32 4, i32 2, i32 0, i32 14, i32 12, i32 10, i32 8>
   %add = add <8 x i16> %lhs, %rhs
   ret <8 x i16> %add
 }
 
 define <8 x i16> @hadd_reverse2_v8i16(<8 x i16> %a0, <8 x i16> %a1) nounwind {
; AVX-NEXT:    retq
   %rhs = shufflevector <8 x i16> %shuf0, <8 x i16> %shuf1, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>
   %add = add <8 x i16> %lhs, %rhs
   ret <8 x i16> %add
 }
 
 define <16 x i16> @hadd_reverse_v16i16(<16 x i16> %a0, <16 x i16> %a1) nounwind {
 ; SSE-LABEL: hadd_reverse_v16i16:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    movdqa %xmm0, %xmm4
-; SSE-NEXT:    movdqa {{.*#+}} xmm0 = [14,15,10,11,6,7,2,3,14,15,10,11,12,13,14,15]
-; SSE-NEXT:    movdqa %xmm2, %xmm5
-; SSE-NEXT:    pshufb %xmm0, %xmm5
-; SSE-NEXT:    movdqa %xmm4, %xmm6
-; SSE-NEXT:    pshufb %xmm0, %xmm6
-; SSE-NEXT:    punpcklqdq {{.*#+}} xmm6 = xmm6[0],xmm5[0]
-; SSE-NEXT:    movdqa %xmm3, %xmm5
-; SSE-NEXT:    pshufb %xmm0, %xmm5
-; SSE-NEXT:    movdqa %xmm1, %xmm7
-; SSE-NEXT:    pshufb %xmm0, %xmm7
-; SSE-NEXT:    punpcklqdq {{.*#+}} xmm7 = xmm7[0],xmm5[0]
-; SSE-NEXT:    movdqa {{.*#+}} xmm0 = [12,13,8,9,4,5,0,1,8,9,12,13,12,13,14,15]
-; SSE-NEXT:    pshufb %xmm0, %xmm2
-; SSE-NEXT:    pshufb %xmm0, %xmm4
-; SSE-NEXT:    punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
-; SSE-NEXT:    paddw %xmm6, %xmm4
-; SSE-NEXT:    pshufb %xmm0, %xmm3
-; SSE-NEXT:    pshufb %xmm0, %xmm1
-; SSE-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm3[0]
-; SSE-NEXT:    paddw %xmm7, %xmm1
-; SSE-NEXT:    movdqa %xmm1, %xmm0
-; SSE-NEXT:    movdqa %xmm4, %xmm1
+; SSE-NEXT:    phaddw %xmm3, %xmm1
+; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[3,2,1,0,4,5,6,7]
+; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,7,6,5,4]
+; SSE-NEXT:    phaddw %xmm2, %xmm0
+; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
+; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,7,6,5,4]
+; SSE-NEXT:    movdqa %xmm3, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: hadd_reverse_v16i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [14,15,10,11,6,7,2,3,14,15,10,11,12,13,14,15]
-; AVX1-NEXT:    vpshufb %xmm3, %xmm2, %xmm4
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm3, %xmm5, %xmm6
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm4 = xmm6[0],xmm4[0]
-; AVX1-NEXT:    vpshufb %xmm3, %xmm1, %xmm6
-; AVX1-NEXT:    vpshufb %xmm3, %xmm0, %xmm3
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm6[0]
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm6 = [12,13,8,9,4,5,0,1,8,9,12,13,12,13,14,15]
-; AVX1-NEXT:    vpshufb %xmm6, %xmm2, %xmm2
-; AVX1-NEXT:    vpshufb %xmm6, %xmm5, %xmm5
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm5[0],xmm2[0]
-; AVX1-NEXT:    vpaddw %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpshufb %xmm6, %xmm1, %xmm1
-; AVX1-NEXT:    vpshufb %xmm6, %xmm0, %xmm0
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX1-NEXT:    vpaddw %xmm0, %xmm3, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
+; AVX1-NEXT:    vphaddw %xmm2, %xmm3, %xmm2
+; AVX1-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[3,2,1,0,4,5,6,7]
+; AVX1-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,6,5,4]
+; AVX1-NEXT:    vphaddw %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
+; AVX1-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
 ; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: hadd_reverse_v16i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm1[3,1,2,3,4,5,6,7,11,9,10,11,12,13,14,15]
-; AVX2-NEXT:    vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,7,5,6,7,8,9,10,11,15,13,14,15]
-; AVX2-NEXT:    vpshuflw {{.*#+}} ymm3 = ymm0[3,1,2,3,4,5,6,7,11,9,10,11,12,13,14,15]
-; AVX2-NEXT:    vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,7,5,6,7,8,9,10,11,15,13,14,15]
-; AVX2-NEXT:    vshufps {{.*#+}} ymm2 = ymm3[2,0],ymm2[2,0],ymm3[6,4],ymm2[6,4]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[2,3,0,1]
-; AVX2-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,12,13,8,9,4,5,0,1,u,u,u,u,u,u,u,u,28,29,24,25,20,21,16,17]
-; AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[12,13,8,9,4,5,0,1,u,u,u,u,u,u,u,u,28,29,24,25,20,21,16,17,u,u,u,u,u,u,u,u]
-; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
+; AVX2-NEXT:    vphaddw %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[3,2,1,0,4,5,6,7,11,10,9,8,12,13,14,15]
+; AVX2-NEXT:    vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,7,6,5,4,8,9,10,11,15,14,13,12]
 ; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1]
-; AVX2-NEXT:    vpaddw %ymm0, %ymm2, %ymm0
 ; AVX2-NEXT:    retq
   %lhs = shufflevector <16 x i16> %a0, <16 x i16> %a1, <16 x i32> <i32 15, i32 13, i32 11, i32 9, i32 31, i32 29, i32 27, i32 25, i32 7, i32 5, i32 3, i32 1, i32 23, i32 21, i32 19, i32 17>
   %rhs = shufflevector <16 x i16> %a0, <16 x i16> %a1, <16 x i32> <i32 14, i32 12, i32 10, i32 8, i32 30, i32 28, i32 26, i32 24, i32 6, i32 4, i32 2, i32 0, i32 22, i32 20, i32 18, i32 16>
   %add = add <16 x i16> %lhs, %rhs
   ret <16 x i16> %add
 }
 
 define <16 x i16> @hadd_reverse2_v16i16(<16 x i16> %a0, <16 x i16> %a1) nounwind {

 ; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm3[2,3,0,1]
 ; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,0,1]
 ; AVX1-NEXT:    vhaddpd %ymm0, %ymm1, %ymm0
 ; AVX1-NEXT:    vmovapd %ymm2, %ymm1
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: hadd_reverse_v8f64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm4 = ymm2[0,3,2,1]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm5 = ymm0[0,3,2,1]
-; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm5 = ymm3[0,3,2,1]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm6 = ymm1[0,3,2,1]
-; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[2,1,0,3]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[2,1,0,3]
-; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[2],ymm2[2]
-; AVX2-NEXT:    vaddpd %ymm0, %ymm4, %ymm2
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm3[2,1,0,3]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[2,1,0,3]
-; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
-; AVX2-NEXT:    vaddpd %ymm0, %ymm5, %ymm0
-; AVX2-NEXT:    vmovapd %ymm2, %ymm1
+; AVX2-NEXT:    vhaddpd %ymm3, %ymm1, %ymm1
+; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm1[2,3,0,1]
+; AVX2-NEXT:    vhaddpd %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm0[2,3,0,1]
+; AVX2-NEXT:    vmovapd %ymm3, %ymm0
 ; AVX2-NEXT:    retq
   %lhs = shufflevector <8 x double> %a0, <8 x double> %a1, <8 x i32> <i32 7, i32 15, i32 5, i32 13, i32 3, i32 11, i32 1, i32 9>
   %rhs = shufflevector <8 x double> %a0, <8 x double> %a1, <8 x i32> <i32 6, i32 14, i32 4, i32 12, i32 2, i32 10, i32 0, i32 8>
   %fadd = fadd <8 x double> %lhs, %rhs
   ret <8 x double> %fadd
 }
 
 define <8 x double> @hadd_reverse2_v8f64(<8 x double> %a0, <8 x double> %a1) nounwind {
; AVX2-NEXT:    retq
   ret <8 x double> %fadd
 }
 
 define <16 x float> @hadd_reverse_v16f32(<16 x float> %a0, <16 x float> %a1) nounwind {
 ; SSE-LABEL: hadd_reverse_v16f32:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    movaps %xmm5, %xmm8
 ; SSE-NEXT:    movaps %xmm1, %xmm5
-; SSE-NEXT:    movaps %xmm8, %xmm1
-; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,1],xmm4[3,1]
-; SSE-NEXT:    movaps %xmm5, %xmm9
-; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[3,1],xmm0[3,1]
-; SSE-NEXT:    movaps %xmm7, %xmm10
-; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[3,1],xmm6[3,1]
-; SSE-NEXT:    movaps %xmm3, %xmm11
-; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[3,1],xmm2[3,1]
-; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,0],xmm4[2,0]
-; SSE-NEXT:    addps %xmm1, %xmm8
-; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[2,0],xmm0[2,0]
-; SSE-NEXT:    addps %xmm9, %xmm5
-; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,0],xmm6[2,0]
-; SSE-NEXT:    addps %xmm10, %xmm7
-; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[2,0],xmm2[2,0]
-; SSE-NEXT:    addps %xmm11, %xmm3
+; SSE-NEXT:    haddps %xmm2, %xmm3
+; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[1,0,3,2]
+; SSE-NEXT:    haddps %xmm6, %xmm7
+; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[1,0,3,2]
+; SSE-NEXT:    haddps %xmm0, %xmm5
+; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,0,3,2]
+; SSE-NEXT:    haddps %xmm4, %xmm8
+; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[1,0,3,2]
 ; SSE-NEXT:    movaps %xmm3, %xmm0
 ; SSE-NEXT:    movaps %xmm7, %xmm1
 ; SSE-NEXT:    movaps %xmm5, %xmm2
 ; SSE-NEXT:    movaps %xmm8, %xmm3
 ; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: hadd_reverse_v16f32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm4 = ymm0[2,3],ymm2[2,3]
 ; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
-; AVX1-NEXT:    vshufps {{.*#+}} ymm2 = ymm4[3,1],ymm0[3,1],ymm4[7,5],ymm0[7,5]
-; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm5 = ymm1[2,3],ymm3[2,3]
-; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm3
-; AVX1-NEXT:    vshufps {{.*#+}} ymm6 = ymm5[3,1],ymm3[3,1],ymm5[7,5],ymm3[7,5]
-; AVX1-NEXT:    vshufps {{.*#+}} ymm0 = ymm4[2,0],ymm0[2,0],ymm4[6,4],ymm0[6,4]
-; AVX1-NEXT:    vaddps %ymm0, %ymm2, %ymm1
-; AVX1-NEXT:    vshufps {{.*#+}} ymm0 = ymm5[2,0],ymm3[2,0],ymm5[6,4],ymm3[6,4]
-; AVX1-NEXT:    vaddps %ymm0, %ymm6, %ymm0
+; AVX1-NEXT:    vhaddps %ymm0, %ymm4, %ymm2
+; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm3[2,3]
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
+; AVX1-NEXT:    vhaddps %ymm1, %ymm0, %ymm0
+; AVX1-NEXT:    vpermilps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]
+; AVX1-NEXT:    vpermilps {{.*#+}} ymm1 = ymm2[1,0,3,2,5,4,7,6]
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: hadd_reverse_v16f32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vshufps {{.*#+}} ymm4 = ymm0[3,1],ymm2[3,1],ymm0[7,5],ymm2[7,5]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm4 = ymm4[2,0,3,1]
-; AVX2-NEXT:    vshufps {{.*#+}} ymm5 = ymm1[3,1],ymm3[3,1],ymm1[7,5],ymm3[7,5]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm5 = ymm5[2,0,3,1]
-; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,0],ymm2[2,0],ymm0[6,4],ymm2[6,4]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[2,0,3,1]
-; AVX2-NEXT:    vaddps %ymm0, %ymm4, %ymm2
-; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[2,0],ymm3[2,0],ymm1[6,4],ymm3[6,4]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[2,0,3,1]
-; AVX2-NEXT:    vaddps %ymm0, %ymm5, %ymm0
-; AVX2-NEXT:    vmovaps %ymm2, %ymm1
+; AVX2-NEXT:    vhaddps %ymm3, %ymm1, %ymm1
+; AVX2-NEXT:    vpermilps {{.*#+}} ymm1 = ymm1[1,0,3,2,5,4,7,6]
+; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm1[2,0,3,1]
+; AVX2-NEXT:    vhaddps %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpermilps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]
+; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm0[2,0,3,1]
+; AVX2-NEXT:    vmovaps %ymm3, %ymm0
 ; AVX2-NEXT:    retq
   %lhs = shufflevector <16 x float> %a0, <16 x float> %a1, <16 x i32> <i32 15, i32 13, i32 11, i32 9, i32 31, i32 29, i32 27, i32 25, i32 7, i32 5, i32 3, i32 1, i32 23, i32 21, i32 19, i32 17>
   %rhs = shufflevector <16 x float> %a0, <16 x float> %a1, <16 x i32> <i32 14, i32 12, i32 10, i32 8, i32 30, i32 28, i32 26, i32 24, i32 6, i32 4, i32 2, i32 0, i32 22, i32 20, i32 18, i32 16>
   %fadd = fadd <16 x float> %lhs, %rhs
   ret <16 x float> %fadd
 }
 
 define <16 x float> @hadd_reverse2_v16f32(<16 x float> %a0, <16 x float> %a1) nounwind {

llvm/test/CodeGen/X86/haddsub-shuf.ll

; AVX2-NEXT:    retq
   ret <4 x float> %2
 }
 
 declare <4 x float> @llvm.x86.sse3.hadd.ps(<4 x float>, <4 x float>)
 
 define <4 x float> @PR34724_1(<4 x float> %a, <4 x float> %b) {
 ; SSSE3_SLOW-LABEL: PR34724_1:
 ; SSSE3_SLOW:       # %bb.0:
-; SSSE3_SLOW-NEXT:    movaps %xmm1, %xmm2
-; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,0],xmm0[3,2]
-; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0],xmm0[2,3]
 ; SSSE3_SLOW-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
-; SSSE3_SLOW-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; SSSE3_SLOW-NEXT:    addps %xmm0, %xmm2
-; SSSE3_SLOW-NEXT:    movsldup {{.*#+}} xmm0 = xmm1[0,0,2,2]
-; SSSE3_SLOW-NEXT:    addps %xmm1, %xmm0
-; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,0],xmm2[1,0]
-; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0],xmm0[2,0]
-; SSSE3_SLOW-NEXT:    movaps %xmm2, %xmm0
+; SSSE3_SLOW-NEXT:    haddps %xmm1, %xmm0
+; SSSE3_SLOW-NEXT:    movsldup {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSSE3_SLOW-NEXT:    addps %xmm1, %xmm2
+; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,0],xmm0[2,0]
+; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm2[2,0]
 ; SSSE3_SLOW-NEXT:    retq
 ;
 ; SSSE3_FAST-LABEL: PR34724_1:
 ; SSSE3_FAST:       # %bb.0:
-; SSSE3_FAST-NEXT:    movaps %xmm1, %xmm2
-; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,0],xmm0[3,2]
-; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0],xmm0[2,3]
 ; SSSE3_FAST-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
-; SSSE3_FAST-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; SSSE3_FAST-NEXT:    addps %xmm0, %xmm2
+; SSSE3_FAST-NEXT:    haddps %xmm1, %xmm0
 ; SSSE3_FAST-NEXT:    haddps %xmm1, %xmm1
-; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,0],xmm2[1,0]
-; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0],xmm1[2,0]
-; SSSE3_FAST-NEXT:    movaps %xmm2, %xmm0
+; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,0],xmm0[2,0]
+; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[2,0]
 ; SSSE3_FAST-NEXT:    retq
 ;
 ; AVX1_SLOW-LABEL: PR34724_1:
 ; AVX1_SLOW:       # %bb.0:
-; AVX1_SLOW-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
-; AVX1_SLOW-NEXT:    vinsertps {{.*#+}} xmm2 = xmm2[0],xmm1[0],zero,zero
-; AVX1_SLOW-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[3],xmm1[1],zero,zero
-; AVX1_SLOW-NEXT:    vaddps %xmm0, %xmm2, %xmm0
+; AVX1_SLOW-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1_SLOW-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
 ; AVX1_SLOW-NEXT:    vmovsldup {{.*#+}} xmm2 = xmm1[0,0,2,2]
 ; AVX1_SLOW-NEXT:    vaddps %xmm1, %xmm2, %xmm1
-; AVX1_SLOW-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
-; AVX1_SLOW-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,1,3]
+; AVX1_SLOW-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3]
+; AVX1_SLOW-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,2,3]
 ; AVX1_SLOW-NEXT:    retq
 ;
 ; AVX1_FAST-LABEL: PR34724_1:
 ; AVX1_FAST:       # %bb.0:
-; AVX1_FAST-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
-; AVX1_FAST-NEXT:    vinsertps {{.*#+}} xmm2 = xmm2[0],xmm1[0],zero,zero
-; AVX1_FAST-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[3],xmm1[1],zero,zero
-; AVX1_FAST-NEXT:    vaddps %xmm0, %xmm2, %xmm0
+; AVX1_FAST-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1_FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
 ; AVX1_FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
-; AVX1_FAST-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
-; AVX1_FAST-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,1,3]
+; AVX1_FAST-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3]
+; AVX1_FAST-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,2,3]
 ; AVX1_FAST-NEXT:    retq
 ;
 ; AVX2_SLOW-LABEL: PR34724_1:
 ; AVX2_SLOW:       # %bb.0:
-; AVX2_SLOW-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
-; AVX2_SLOW-NEXT:    vinsertps {{.*#+}} xmm2 = xmm2[0],xmm1[0],zero,zero
-; AVX2_SLOW-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[3],xmm1[1],zero,zero
-; AVX2_SLOW-NEXT:    vaddps %xmm0, %xmm2, %xmm0
+; AVX2_SLOW-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2_SLOW-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
 ; AVX2_SLOW-NEXT:    vmovsldup {{.*#+}} xmm2 = xmm1[0,0,2,2]
 ; AVX2_SLOW-NEXT:    vaddps %xmm1, %xmm2, %xmm1
-; AVX2_SLOW-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
-; AVX2_SLOW-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,1,3]
+; AVX2_SLOW-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3]
+; AVX2_SLOW-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,2,3]
 ; AVX2_SLOW-NEXT:    retq
 ;
 ; AVX2_FAST-LABEL: PR34724_1:
 ; AVX2_FAST:       # %bb.0:
-; AVX2_FAST-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
-; AVX2_FAST-NEXT:    vinsertps {{.*#+}} xmm2 = xmm2[0],xmm1[0],zero,zero
-; AVX2_FAST-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[3],xmm1[1],zero,zero
-; AVX2_FAST-NEXT:    vaddps %xmm0, %xmm2, %xmm0
+; AVX2_FAST-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2_FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
 ; AVX2_FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
-; AVX2_FAST-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
-; AVX2_FAST-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,1,3]
+; AVX2_FAST-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3]
+; AVX2_FAST-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,2,3]
 ; AVX2_FAST-NEXT:    retq
   %t0 = shufflevector <4 x float> %a, <4 x float> %b, <2 x i32> <i32 2, i32 4>
   %t1 = shufflevector <4 x float> %a, <4 x float> %b, <2 x i32> <i32 3, i32 5>
   %t2 = fadd <2 x float> %t0, %t1
   %vecinit9 = shufflevector <2 x float> %t2, <2 x float> undef, <4 x i32> <i32 undef, i32 0, i32 1, i32 undef>
   %t3 = shufflevector <4 x float> %b, <4 x float> undef, <4 x i32> <i32 undef, i32 undef, i32 undef, i32 2>
   %t4 = fadd <4 x float> %t3, %b
   %vecinit13 = shufflevector <4 x float> %vecinit9, <4 x float> %t4, <4 x i32> <i32 undef, i32 1, i32 2, i32 7>
   ret <4 x float> %vecinit13
 }
 
 define <4 x float> @PR34724_2(<4 x float> %a, <4 x float> %b) {
 ; SSSE3_SLOW-LABEL: PR34724_2:
 ; SSSE3_SLOW:       # %bb.0:
-; SSSE3_SLOW-NEXT:    movaps %xmm1, %xmm2
-; SSSE3_SLOW-NEXT:    movsldup {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSSE3_SLOW-NEXT:    addps %xmm1, %xmm3
-; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[2,0]
-; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[2,3]
-; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,0],xmm0[3,0]
-; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0],xmm0[2,3]
+; SSSE3_SLOW-NEXT:    haddps %xmm1, %xmm0
+; SSSE3_SLOW-NEXT:    movsldup {{.*#+}} xmm2 = xmm1[0,0,2,2]
 ; SSSE3_SLOW-NEXT:    addps %xmm1, %xmm2
-; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm3 = xmm3[3,0],xmm2[1,0]
-; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0],xmm3[2,0]
-; SSSE3_SLOW-NEXT:    movaps %xmm2, %xmm0
+; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,0],xmm0[2,0]
+; SSSE3_SLOW-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,0]
 ; SSSE3_SLOW-NEXT:    retq
 ;
 ; SSSE3_FAST-LABEL: PR34724_2:
 ; SSSE3_FAST:       # %bb.0:
-; SSSE3_FAST-NEXT:    movaps %xmm1, %xmm3
-; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,0],xmm0[2,0]
-; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm3 = xmm3[2,0],xmm0[2,3]
-; SSSE3_FAST-NEXT:    movaps %xmm1, %xmm2
-; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,0],xmm0[3,0]
-; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,0],xmm0[2,3]
-; SSSE3_FAST-NEXT:    addps %xmm3, %xmm2
+; SSSE3_FAST-NEXT:    haddps %xmm1, %xmm0
 ; SSSE3_FAST-NEXT:    haddps %xmm1, %xmm1
-; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,0],xmm2[1,0]
-; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,0],xmm1[2,0]
-; SSSE3_FAST-NEXT:    movaps %xmm2, %xmm0
+; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,0],xmm0[2,0]
+; SSSE3_FAST-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,0]
 ; SSSE3_FAST-NEXT:    retq
 ;
 ; AVX1_SLOW-LABEL: PR34724_2:
 ; AVX1_SLOW:       # %bb.0:
-; AVX1_SLOW-NEXT:    vblendps {{.*#+}} xmm2 = xmm1[0,1],xmm0[2,3]
-; AVX1_SLOW-NEXT:    vpermilps {{.*#+}} xmm2 = xmm2[2,0,2,3]
-; AVX1_SLOW-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[3],xmm1[1],zero,zero
-; AVX1_SLOW-NEXT:    vaddps %xmm0, %xmm2, %xmm0
+; AVX1_SLOW-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
 ; AVX1_SLOW-NEXT:    vmovsldup {{.*#+}} xmm2 = xmm1[0,0,2,2]
 ; AVX1_SLOW-NEXT:    vaddps %xmm1, %xmm2, %xmm1
-; AVX1_SLOW-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
-; AVX1_SLOW-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,1,3]
+; AVX1_SLOW-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3]
 ; AVX1_SLOW-NEXT:    retq
 ;
 ; AVX1_FAST-LABEL: PR34724_2:
 ; AVX1_FAST:       # %bb.0:
-; AVX1_FAST-NEXT:    vblendps {{.*#+}} xmm2 = xmm1[0,1],xmm0[2,3]
-; AVX1_FAST-NEXT:    vpermilps {{.*#+}} xmm2 = xmm2[2,0,2,3]
-; AVX1_FAST-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[3],xmm1[1],zero,zero
-; AVX1_FAST-NEXT:    vaddps %xmm0, %xmm2, %xmm0
+; AVX1_FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
 ; AVX1_FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
-; AVX1_FAST-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
-; AVX1_FAST-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,1,3]
+; AVX1_FAST-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3]
 ; AVX1_FAST-NEXT:    retq
 ;
 ; AVX2_SLOW-LABEL: PR34724_2:
 ; AVX2_SLOW:       # %bb.0:
-; AVX2_SLOW-NEXT:    vblendps {{.*#+}} xmm2 = xmm1[0,1],xmm0[2,3]
-; AVX2_SLOW-NEXT:    vpermilps {{.*#+}} xmm2 = xmm2[2,0,2,3]
-; AVX2_SLOW-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[3],xmm1[1],zero,zero
-; AVX2_SLOW-NEXT:    vaddps %xmm0, %xmm2, %xmm0
+; AVX2_SLOW-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
 ; AVX2_SLOW-NEXT:    vmovsldup {{.*#+}} xmm2 = xmm1[0,0,2,2]
 ; AVX2_SLOW-NEXT:    vaddps %xmm1, %xmm2, %xmm1
-; AVX2_SLOW-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
-; AVX2_SLOW-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,1,3]
+; AVX2_SLOW-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3]
 ; AVX2_SLOW-NEXT:    retq
 ;
 ; AVX2_FAST-LABEL: PR34724_2:
 ; AVX2_FAST:       # %bb.0:
-; AVX2_FAST-NEXT:    vblendps {{.*#+}} xmm2 = xmm1[0,1],xmm0[2,3]
-; AVX2_FAST-NEXT:    vpermilps {{.*#+}} xmm2 = xmm2[2,0,2,3]
-; AVX2_FAST-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[3],xmm1[1],zero,zero
-; AVX2_FAST-NEXT:    vaddps %xmm0, %xmm2, %xmm0
+; AVX2_FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
 ; AVX2_FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
-; AVX2_FAST-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
-; AVX2_FAST-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,1,3]
+; AVX2_FAST-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3]
 ; AVX2_FAST-NEXT:    retq
   %t0 = shufflevector <4 x float> %a, <4 x float> %b, <4 x i32> <i32 2, i32 4, i32 undef, i32 undef>
   %t1 = shufflevector <4 x float> %a, <4 x float> %b, <4 x i32> <i32 3, i32 5, i32 undef, i32 undef>
   %t2 = fadd <4 x float> %t0, %t1
   %vecinit9 = shufflevector <4 x float> %t2, <4 x float> undef, <4 x i32> <i32 undef, i32 0, i32 1, i32 undef>
   %t3 = shufflevector <4 x float> %b, <4 x float> undef, <4 x i32> <i32 undef, i32 undef, i32 undef, i32 2>
   %t4 = fadd <4 x float> %t3, %b
   %vecinit13 = shufflevector <4 x float> %vecinit9, <4 x float> %t4, <4 x i32> <i32 undef, i32 1, i32 2, i32 7>
   ret <4 x float> %vecinit13
 }

llvm/test/CodeGen/X86/haddsub-undef.ll

 ; SSE-FAST-LABEL: PR44694:
 ; SSE-FAST:       # %bb.0:
 ; SSE-FAST-NEXT:    movapd %xmm1, %xmm0
 ; SSE-FAST-NEXT:    haddpd %xmm3, %xmm2
 ; SSE-FAST-NEXT:    haddpd %xmm1, %xmm0
 ; SSE-FAST-NEXT:    movapd %xmm2, %xmm1
 ; SSE-FAST-NEXT:    retq
 ;
-; AVX-LABEL: PR44694:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
-; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
-; AVX-NEXT:    vhaddpd %ymm0, %ymm1, %ymm0
-; AVX-NEXT:    retq
+; AVX1-SLOW-LABEL: PR44694:
+; AVX1-SLOW:       # %bb.0:
+; AVX1-SLOW-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
+; AVX1-SLOW-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
+; AVX1-SLOW-NEXT:    vhaddpd %ymm0, %ymm1, %ymm0
+; AVX1-SLOW-NEXT:    retq
+;
+; AVX1-FAST-LABEL: PR44694:
+; AVX1-FAST:       # %bb.0:
+; AVX1-FAST-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
+; AVX1-FAST-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
+; AVX1-FAST-NEXT:    vhaddpd %ymm0, %ymm1, %ymm0
+; AVX1-FAST-NEXT:    retq
+;
+; AVX512-LABEL: PR44694:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vhaddpd %ymm1, %ymm0, %ymm0
+; AVX512-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
+; AVX512-NEXT:    retq
   %3 = shufflevector <4 x double> %0, <4 x double> %1, <4 x i32> <i32 undef, i32 2, i32 4, i32 6>
   %4 = shufflevector <4 x double> %0, <4 x double> %1, <4 x i32> <i32 undef, i32 3, i32 5, i32 7>
   %5 = fadd <4 x double> %3, %4
   ret <4 x double> %5
 }
 
 define <4 x float> @PR45747_1(<4 x float> %a, <4 x float> %b) nounwind {
-; SSE-LABEL: PR45747_1:
-; SSE:       # %bb.0:
-; SSE-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE-NEXT:    addps %xmm0, %xmm1
-; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2,2,3]
-; SSE-NEXT:    movaps %xmm1, %xmm0
-; SSE-NEXT:    retq
+; SSE-SLOW-LABEL: PR45747_1:
+; SSE-SLOW:       # %bb.0:
+; SSE-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE-SLOW-NEXT:    addps %xmm0, %xmm1
+; SSE-SLOW-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2,2,3]
+; SSE-SLOW-NEXT:    movaps %xmm1, %xmm0
+; SSE-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: PR45747_1:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vaddps %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,2,2,3]
-; AVX-NEXT:    retq
+; SSE-FAST-LABEL: PR45747_1:
+; SSE-FAST:       # %bb.0:
+; SSE-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSE-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: PR45747_1:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vaddps %xmm0, %xmm1, %xmm0
+; AVX-SLOW-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,2,2,3]
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: PR45747_1:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
   %t0 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> <i32 undef, i32 undef, i32 3, i32 undef>
   %t1 = fadd <4 x float> %t0, %a
   %shuffle = shufflevector <4 x float> %t1, <4 x float> undef, <4 x i32> <i32 undef, i32 2, i32 undef, i32 undef>
   ret <4 x float> %shuffle
 }
 
 define <4 x float> @PR45747_2(<4 x float> %a, <4 x float> %b) nounwind {
-; SSE-LABEL: PR45747_2:
-; SSE:       # %bb.0:
-; SSE-NEXT:    movshdup {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE-NEXT:    addps %xmm1, %xmm0
-; SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
-; SSE-NEXT:    retq
+; SSE-SLOW-LABEL: PR45747_2:
+; SSE-SLOW:       # %bb.0:
+; SSE-SLOW-NEXT:    movshdup {{.*#+}} xmm0 = xmm1[1,1,3,3]
+; SSE-SLOW-NEXT:    addps %xmm1, %xmm0
+; SSE-SLOW-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
+; SSE-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: PR45747_2:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; AVX-NEXT:    vaddps %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
-; AVX-NEXT:    retq
+; SSE-FAST-LABEL: PR45747_2:
+; SSE-FAST:       # %bb.0:
+; SSE-FAST-NEXT:    movaps %xmm1, %xmm0
+; SSE-FAST-NEXT:    haddps %xmm1, %xmm0
+; SSE-FAST-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: PR45747_2:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm0 = xmm1[1,1,3,3]
+; AVX-SLOW-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: PR45747_2:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm0
+; AVX-FAST-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; AVX-FAST-NEXT:    retq
   %t0 = shufflevector <4 x float> %b, <4 x float> undef, <4 x i32> <i32 undef, i32 undef, i32 3, i32 undef>
   %t1 = fadd <4 x float> %t0, %b
   %shuffle = shufflevector <4 x float> %t1, <4 x float> undef, <4 x i32> <i32 2, i32 undef, i32 undef, i32 undef>
   ret <4 x float> %shuffle
 }

llvm/test/CodeGen/X86/vector-shuffle-256-v4.ll

 ; AVX512VL-FAST:       # %bb.0:
 ; AVX512VL-FAST-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,ymm0[16,17,18,19,20,21,22,23],zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX512VL-FAST-NEXT:    retq
   %1 = shufflevector <4 x i64> %a, <4 x i64> <i64 0, i64 undef, i64 undef, i64 undef>, <4 x i32> <i32 1, i32 4, i32 2, i32 4>
   ret <4 x i64> %1
 }
 
 define <4 x double> @add_v4f64_0246_1357(<4 x double> %a, <4 x double> %b) {
-; ALL-LABEL: add_v4f64_0246_1357:
-; ALL:       # %bb.0: # %entry
-; ALL-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
-; ALL-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; ALL-NEXT:    vhaddpd %ymm2, %ymm0, %ymm0
-; ALL-NEXT:    retq
+; AVX1-LABEL: add_v4f64_0246_1357:
+; AVX1:       # %bb.0: # %entry
+; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT:    vhaddpd %ymm2, %ymm0, %ymm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: add_v4f64_0246_1357:
+; AVX2:       # %bb.0: # %entry
+; AVX2-NEXT:    vhaddpd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
+; AVX2-NEXT:    retq
+;
+; AVX512VL-LABEL: add_v4f64_0246_1357:
+; AVX512VL:       # %bb.0: # %entry
+; AVX512VL-NEXT:    vhaddpd %ymm1, %ymm0, %ymm0
+; AVX512VL-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
+; AVX512VL-NEXT:    retq
 entry:
   %shuffle = shufflevector <4 x double> %a, <4 x double> %b, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
   %shuffle1 = shufflevector <4 x double> %a, <4 x double> %b, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
   %add = fadd <4 x double> %shuffle, %shuffle1
   ret <4 x double> %add
 }
 
 define <4 x double> @add_v4f64_4602_5713(<4 x double> %a, <4 x double> %b) {
-; ALL-LABEL: add_v4f64_4602_5713:
-; ALL:       # %bb.0: # %entry
-; ALL-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm1[2,3],ymm0[2,3]
-; ALL-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; ALL-NEXT:    vhaddpd %ymm2, %ymm0, %ymm0
-; ALL-NEXT:    retq
+; AVX1-LABEL: add_v4f64_4602_5713:
+; AVX1:       # %bb.0: # %entry
+; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm1[2,3],ymm0[2,3]
+; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vhaddpd %ymm2, %ymm0, %ymm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: add_v4f64_4602_5713:
+; AVX2:       # %bb.0: # %entry
+; AVX2-NEXT:    vhaddpd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[1,3,0,2]
+; AVX2-NEXT:    retq
+;
+; AVX512VL-LABEL: add_v4f64_4602_5713:
+; AVX512VL:       # %bb.0: # %entry
+; AVX512VL-NEXT:    vhaddpd %ymm1, %ymm0, %ymm0
+; AVX512VL-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[1,3,0,2]
+; AVX512VL-NEXT:    retq
 entry:
   %shuffle = shufflevector <4 x double> %a, <4 x double> %b, <4 x i32> <i32 4, i32 6, i32 0, i32 2>
   %shuffle1 = shufflevector <4 x double> %a, <4 x double> %b, <4 x i32> <i32 5, i32 7, i32 1, i32 3>
   %add = fadd <4 x double> %shuffle, %shuffle1
   ret <4 x double> %add
 }
 
 define <4 x i64> @add_v4i64_0246_1357(<4 x i64> %a, <4 x i64> %b) {

llvm/test/CodeGen/X86/vector-shuffle-256-v8.ll

 define <8 x float> @add_v8f32_02468ACE_13579BDF(<8 x float> %a, <8 x float> %b) {
 ; AVX1-LABEL: add_v8f32_02468ACE_13579BDF:
 ; AVX1:       # %bb.0: # %entry
 ; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
 ; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    vhaddps %ymm2, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: add_v8f32_02468ACE_13579BDF:
-; AVX2:       # %bb.0: # %entry
-; AVX2-NEXT:    vshufps {{.*#+}} ymm2 = ymm0[0,2],ymm1[0,2],ymm0[4,6],ymm1[4,6]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]
-; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm1[1,3],ymm0[5,7],ymm1[5,7]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX2-NEXT:    vaddps %ymm0, %ymm2, %ymm0
-; AVX2-NEXT:    retq
-;
-; AVX512VL-SLOW-LABEL: add_v8f32_02468ACE_13579BDF:
-; AVX512VL-SLOW:       # %bb.0: # %entry
-; AVX512VL-SLOW-NEXT:    vshufps {{.*#+}} ymm2 = ymm0[0,2],ymm1[0,2],ymm0[4,6],ymm1[4,6]
-; AVX512VL-SLOW-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]
-; AVX512VL-SLOW-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm1[1,3],ymm0[5,7],ymm1[5,7]
-; AVX512VL-SLOW-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX512VL-SLOW-NEXT:    vaddps %ymm0, %ymm2, %ymm0
-; AVX512VL-SLOW-NEXT:    retq
-;
-; AVX512VL-FAST-LABEL: add_v8f32_02468ACE_13579BDF:
-; AVX512VL-FAST:       # %bb.0: # %entry
-; AVX512VL-FAST-NEXT:    vmovaps {{.*#+}} ymm2 = [0,2,4,6,8,10,12,14]
-; AVX512VL-FAST-NEXT:    vpermi2ps %ymm1, %ymm0, %ymm2
-; AVX512VL-FAST-NEXT:    vmovaps {{.*#+}} ymm3 = [1,3,5,7,9,11,13,15]
-; AVX512VL-FAST-NEXT:    vpermi2ps %ymm1, %ymm0, %ymm3
-; AVX512VL-FAST-NEXT:    vaddps %ymm3, %ymm2, %ymm0
-; AVX512VL-FAST-NEXT:    retq
+; AVX2OR512VL-LABEL: add_v8f32_02468ACE_13579BDF:
+; AVX2OR512VL:       # %bb.0: # %entry
+; AVX2OR512VL-NEXT:    vhaddps %ymm1, %ymm0, %ymm0
+; AVX2OR512VL-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
+; AVX2OR512VL-NEXT:    retq
 entry:
   %shuffle = shufflevector <8 x float> %a, <8 x float> %b, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14>
   %shuffle1 = shufflevector <8 x float> %a, <8 x float> %b, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>
   %add = fadd <8 x float> %shuffle, %shuffle1
   ret <8 x float> %add
 }
 
 define <8 x float> @add_v8f32_8ACE0246_9BDF1357(<8 x float> %a, <8 x float> %b) {
 ; AVX1-LABEL: add_v8f32_8ACE0246_9BDF1357:
 ; AVX1:       # %bb.0: # %entry
 ; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm1[2,3],ymm0[2,3]
 ; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
 ; AVX1-NEXT:    vhaddps %ymm2, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: add_v8f32_8ACE0246_9BDF1357:
-; AVX2:       # %bb.0: # %entry
-; AVX2-NEXT:    vshufps {{.*#+}} ymm2 = ymm1[0,2],ymm0[0,2],ymm1[4,6],ymm0[4,6]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]
-; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[1,3],ymm0[1,3],ymm1[5,7],ymm0[5,7]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX2-NEXT:    vaddps %ymm0, %ymm2, %ymm0
-; AVX2-NEXT:    retq
-;
-; AVX512VL-SLOW-LABEL: add_v8f32_8ACE0246_9BDF1357:
-; AVX512VL-SLOW:       # %bb.0: # %entry
-; AVX512VL-SLOW-NEXT:    vshufps {{.*#+}} ymm2 = ymm1[0,2],ymm0[0,2],ymm1[4,6],ymm0[4,6]
-; AVX512VL-SLOW-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]
-; AVX512VL-SLOW-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[1,3],ymm0[1,3],ymm1[5,7],ymm0[5,7]
-; AVX512VL-SLOW-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX512VL-SLOW-NEXT:    vaddps %ymm0, %ymm2, %ymm0
-; AVX512VL-SLOW-NEXT:    retq
-;
-; AVX512VL-FAST-LABEL: add_v8f32_8ACE0246_9BDF1357:
-; AVX512VL-FAST:       # %bb.0: # %entry
-; AVX512VL-FAST-NEXT:    vmovaps {{.*#+}} ymm2 = [0,2,4,6,8,10,12,14]
-; AVX512VL-FAST-NEXT:    vpermi2ps %ymm0, %ymm1, %ymm2
-; AVX512VL-FAST-NEXT:    vmovaps {{.*#+}} ymm3 = [1,3,5,7,9,11,13,15]
-; AVX512VL-FAST-NEXT:    vpermi2ps %ymm0, %ymm1, %ymm3
-; AVX512VL-FAST-NEXT:    vaddps %ymm3, %ymm2, %ymm0
-; AVX512VL-FAST-NEXT:    retq
+; AVX2OR512VL-LABEL: add_v8f32_8ACE0246_9BDF1357:
+; AVX2OR512VL:       # %bb.0: # %entry
+; AVX2OR512VL-NEXT:    vhaddps %ymm1, %ymm0, %ymm0
+; AVX2OR512VL-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[1,3,0,2]
+; AVX2OR512VL-NEXT:    retq
 entry:
   %shuffle = shufflevector <8 x float> %a, <8 x float> %b, <8 x i32> <i32 8, i32 10, i32 12, i32 14, i32 0, i32 2, i32 4, i32 6>
   %shuffle1 = shufflevector <8 x float> %a, <8 x float> %b, <8 x i32> <i32 9, i32 11, i32 13, i32 15, i32 1, i32 3, i32 5, i32 7>
   %add = fadd <8 x float> %shuffle, %shuffle1
   ret <8 x float> %add
 }
 
 define <8 x i32> @add_v8i32_02468ACE_13579BDF(<8 x i32> %a, <8 x i32> %b) {
 ; AVX1-LABEL: add_v8i32_02468ACE_13579BDF:
 ; AVX1:       # %bb.0: # %entry
-; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
-; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1-NEXT:    vshufps {{.*#+}} ymm1 = ymm0[0,2],ymm2[0,2],ymm0[4,6],ymm2[4,6]
-; AVX1-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm2[1,3],ymm0[5,7],ymm2[5,7]
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3
-; AVX1-NEXT:    vpaddd %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpaddd %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
+; AVX1-NEXT:    vphaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm2, %ymm2
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3]
+; AVX1-NEXT:    vshufpd {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[3],ymm0[3]
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: add_v8i32_02468ACE_13579BDF:
-; AVX2:       # %bb.0: # %entry
-; AVX2-NEXT:    vshufps {{.*#+}} ymm2 = ymm0[0,2],ymm1[0,2],ymm0[4,6],ymm1[4,6]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]
-; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm1[1,3],ymm0[5,7],ymm1[5,7]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX2-NEXT:    vpaddd %ymm0, %ymm2, %ymm0
-; AVX2-NEXT:    retq
-;
-; AVX512VL-SLOW-LABEL: add_v8i32_02468ACE_13579BDF:
-; AVX512VL-SLOW:       # %bb.0: # %entry
-; AVX512VL-SLOW-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
-; AVX512VL-SLOW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm2
-; AVX512VL-SLOW-NEXT:    vpmovqd %zmm2, %ymm2
-; AVX512VL-SLOW-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm1[1,3],ymm0[5,7],ymm1[5,7]
-; AVX512VL-SLOW-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX512VL-SLOW-NEXT:    vpaddd %ymm0, %ymm2, %ymm0
-; AVX512VL-SLOW-NEXT:    retq
-;
-; AVX512VL-FAST-LABEL: add_v8i32_02468ACE_13579BDF:
-; AVX512VL-FAST:       # %bb.0: # %entry
-; AVX512VL-FAST-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
-; AVX512VL-FAST-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm2
-; AVX512VL-FAST-NEXT:    vpmovqd %zmm2, %ymm2
-; AVX512VL-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [1,3,5,7,9,11,13,15]
-; AVX512VL-FAST-NEXT:    vpermi2d %ymm1, %ymm0, %ymm3
-; AVX512VL-FAST-NEXT:    vpaddd %ymm3, %ymm2, %ymm0
-; AVX512VL-FAST-NEXT:    retq
+; AVX2OR512VL-LABEL: add_v8i32_02468ACE_13579BDF:
+; AVX2OR512VL:       # %bb.0: # %entry
+; AVX2OR512VL-NEXT:    vphaddd %ymm1, %ymm0, %ymm0
+; AVX2OR512VL-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
+; AVX2OR512VL-NEXT:    retq
 entry:
   %shuffle = shufflevector <8 x i32> %a, <8 x i32> %b, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14>
   %shuffle1 = shufflevector <8 x i32> %a, <8 x i32> %b, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>
   %add = add <8 x i32> %shuffle, %shuffle1
   ret <8 x i32> %add
 }
 
 define <8 x i32> @add_v8i32_8ACE0246_9BDF1357(<8 x i32> %a, <8 x i32> %b) {
 ; AVX1-LABEL: add_v8i32_8ACE0246_9BDF1357:
 ; AVX1:       # %bb.0: # %entry
-; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm1[2,3],ymm0[2,3]
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT:    vshufps {{.*#+}} ymm1 = ymm0[0,2],ymm2[0,2],ymm0[4,6],ymm2[4,6]
-; AVX1-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm2[1,3],ymm0[5,7],ymm2[5,7]
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3
-; AVX1-NEXT:    vpaddd %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpaddd %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
+; AVX1-NEXT:    vphaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm2, %ymm2
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3]
+; AVX1-NEXT:    vshufpd {{.*#+}} ymm0 = ymm2[1],ymm0[1],ymm2[2],ymm0[2]
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: add_v8i32_8ACE0246_9BDF1357:
-; AVX2:       # %bb.0: # %entry
-; AVX2-NEXT:    vshufps {{.*#+}} ymm2 = ymm1[0,2],ymm0[0,2],ymm1[4,6],ymm0[4,6]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]
-; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[1,3],ymm0[1,3],ymm1[5,7],ymm0[5,7]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX2-NEXT:    vpaddd %ymm0, %ymm2, %ymm0
-; AVX2-NEXT:    retq
-;
-; AVX512VL-SLOW-LABEL: add_v8i32_8ACE0246_9BDF1357:
-; AVX512VL-SLOW:       # %bb.0: # %entry
-; AVX512VL-SLOW-NEXT:    # kill: def $ymm1 killed $ymm1 def $zmm1
-; AVX512VL-SLOW-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm2
-; AVX512VL-SLOW-NEXT:    vpmovqd %zmm2, %ymm2
-; AVX512VL-SLOW-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[1,3],ymm0[1,3],ymm1[5,7],ymm0[5,7]
-; AVX512VL-SLOW-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX512VL-SLOW-NEXT:    vpaddd %ymm0, %ymm2, %ymm0
-; AVX512VL-SLOW-NEXT:    retq
-;
-; AVX512VL-FAST-LABEL: add_v8i32_8ACE0246_9BDF1357:
-; AVX512VL-FAST:       # %bb.0: # %entry
-; AVX512VL-FAST-NEXT:    # kill: def $ymm1 killed $ymm1 def $zmm1
-; AVX512VL-FAST-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm2
-; AVX512VL-FAST-NEXT:    vpmovqd %zmm2, %ymm2
-; AVX512VL-FAST-NEXT:    vmovdqa {{.*#+}} ymm3 = [1,3,5,7,9,11,13,15]
-; AVX512VL-FAST-NEXT:    vpermi2d %ymm0, %ymm1, %ymm3
-; AVX512VL-FAST-NEXT:    vpaddd %ymm3, %ymm2, %ymm0
-; AVX512VL-FAST-NEXT:    retq
+; AVX2OR512VL-LABEL: add_v8i32_8ACE0246_9BDF1357:
+; AVX2OR512VL:       # %bb.0: # %entry
+; AVX2OR512VL-NEXT:    vphaddd %ymm1, %ymm0, %ymm0
+; AVX2OR512VL-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[1,3,0,2]
+; AVX2OR512VL-NEXT:    retq
 entry:
   %shuffle = shufflevector <8 x i32> %a, <8 x i32> %b, <8 x i32> <i32 8, i32 10, i32 12, i32 14, i32 0, i32 2, i32 4, i32 6>
   %shuffle1 = shufflevector <8 x i32> %a, <8 x i32> %b, <8 x i32> <i32 9, i32 11, i32 13, i32 15, i32 1, i32 3, i32 5, i32 7>
   %add = add <8 x i32> %shuffle, %shuffle1
   ret <8 x i32> %add
 }
 
 ; This test used to crash due to bad handling of concat_vectors after a bitcast