[DAGCombiner] narrow vector binop with 2 insert subvector operands

[spatel]

Summary

bo (ins undef, X, Z), (ins undef, Y, Z) --> ins undef, (bo X, Y), Z

This is another step in generic vector narrowing. It's also a step towards more horizontal op formation specifically for x86 (although we still failed to match those in the affected tests).

The scalarization cases are also not optimal (we should be scalarizing those), but it's still an improvement to use a narrower vector op when we know part of the result must be undef because both inputs are undef in some vector lanes.

I think a similar match but checking for a constant operand might help some of the cases in D51553.

[RKSimon]

LGTM with one minor

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

18214

(very minor) Move this comment to the outer if() and use the same terms as we used in the shuffle fold above.

Also, explain that this is likely to occur in reduction patterns.

[spatel]

Actually, this patch isn't correct as-is. We can't insert into an undef base vector because at least 'xor undef, undef --> 0' (not undef).
I know I avoided or fixed the similar problem in IR somewhere along the way.
We need to actually compute the constant vector for the specified binop's opcode.

[efriedma]

'xor undef, undef --> 0' (not undef).

Strictly speaking, xor undef, undef ---> undef is correct. It's just that as a practical matter, we try to fold obvious cases to zero so we don't have to argue with people who write silly constructs like __m128i a = _mm_xor_si128(a, a);. The fold here seems unlikely to cause problems in practice.

[spatel]

In D56875#1365621, @efriedma wrote:

'xor undef, undef --> 0' (not undef).

Strictly speaking, xor undef, undef ---> undef is correct. It's just that as a practical matter, we try to fold obvious cases to zero so we don't have to argue with people who write silly constructs like __m128i a = _mm_xor_si128(a, a);. The fold here seems unlikely to cause problems in practice.

Ah, interesting. Either way, I think we want a test to document the behavior, so I added 1 here:
rL351763

Now, the funny thing about this particular case is that we'll generate the optimal code with x86 AVX either way because 'vxorps' with 128-bit operands zeros the upper half...oops! So I added another test here:
rL351764

I think I should check in the more conservative version of the patch first since I already wrote it. Then, if we decide it's worth loosening to the form shown here currently, I can make that a follow-up.

After looking a bit more closely at other patterns that I was hoping to fix, I'm actually not sure if we will end up keeping this code. I think we might be better off trying more general vector-demanded-elements enhancements for binops which would make this explicit pattern-matching unnecessary.

[spatel]

Patch updated:

1. Compute the constant vector that we're inserting into rather than assuming it is undef.
2. Moved code comment as suggested.
3. Rebased with diffs for 'xor' tests.

[RKSimon]

LGTM!

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

 /// Visit a binary vector operation, like ADD.
 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
   assert(N->getValueType(0).isVector() &&
          "SimplifyVBinOp only works on vectors!");
 
   SDValue LHS = N->getOperand(0);
   SDValue RHS = N->getOperand(1);
   SDValue Ops[] = {LHS, RHS};
+  EVT VT = N->getValueType(0);
 
   // See if we can constant fold the vector operation.
   if (SDValue Fold = DAG.FoldConstantVectorArithmetic(
           N->getOpcode(), SDLoc(LHS), LHS.getValueType(), Ops, N->getFlags()))
     return Fold;
 
   // Type legalization might introduce new shuffles in the DAG.
   // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
   //   -> (shuffle (VBinOp (A, B)), Undef, Mask).
   if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
       isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
       LHS.getOperand(1).isUndef() &&
       RHS.getOperand(1).isUndef()) {
     ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
     ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
 
     if (SVN0->getMask().equals(SVN1->getMask())) {
-      EVT VT = N->getValueType(0);
       SDValue UndefVector = LHS.getOperand(1);
       SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
                                      LHS.getOperand(0), RHS.getOperand(0),
                                      N->getFlags());
       AddUsersToWorklist(N);
       return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
                                   SVN0->getMask());
     }
   }
 
+  // VBinOp (ins undef, X, Z), (ins undef, Y, Z) --> ins VecC, (VBinOp X, Y), Z
+  if (LHS.getOpcode() == ISD::INSERT_SUBVECTOR && LHS.getOperand(0).isUndef() &&
+      RHS.getOpcode() == ISD::INSERT_SUBVECTOR && RHS.getOperand(0).isUndef() &&
+      LHS.getOperand(2) == RHS.getOperand(2) &&
+      (LHS.hasOneUse() || RHS.hasOneUse())) {
+    SDValue X = LHS.getOperand(1);
+    SDValue Y = RHS.getOperand(1);
+    SDValue Z = LHS.getOperand(2);
+    EVT NarrowVT = X.getValueType();
+    if (NarrowVT == Y.getValueType() &&
+        TLI.isOperationLegalOrCustomOrPromote(N->getOpcode(), NarrowVT)) {
+      // (binop undef, undef) may not return undef, so compute that result.
+      SDLoc DL(N);
+      SDValue VecC = DAG.getNode(N->getOpcode(), DL, VT, DAG.getUNDEF(VT),
+                                 DAG.getUNDEF(VT));
+      SDValue NarrowBO = DAG.getNode(N->getOpcode(), DL, NarrowVT, X, Y);
+      return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, VecC, NarrowBO, Z);
+    }
+  }
+
   return SDValue();
 }
 
 SDValue DAGCombiner::SimplifySelect(const SDLoc &DL, SDValue N0, SDValue N1,
                                     SDValue N2) {
   assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
 
   SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,

test/CodeGen/X86/avx512-hadd-hsub.ll

; SKX-NEXT:    retq
   ret float %x230
 }
 
 define <16 x i32> @hadd_16_3(<16 x i32> %x225, <16 x i32> %x227) {
 ; KNL-LABEL: hadd_16_3:
 ; KNL:       # %bb.0:
 ; KNL-NEXT:    vshufps {{.*#+}} ymm2 = ymm0[0,2],ymm1[0,2],ymm0[4,6],ymm1[4,6]
 ; KNL-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm1[1,3],ymm0[5,7],ymm1[5,7]
-; KNL-NEXT:    vpaddd %zmm0, %zmm2, %zmm0
+; KNL-NEXT:    vpaddd %ymm0, %ymm2, %ymm0
 ; KNL-NEXT:    retq
 ;
 ; SKX-LABEL: hadd_16_3:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vshufps {{.*#+}} ymm2 = ymm0[0,2],ymm1[0,2],ymm0[4,6],ymm1[4,6]
 ; SKX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm1[1,3],ymm0[5,7],ymm1[5,7]
-; SKX-NEXT:    vpaddd %zmm0, %zmm2, %zmm0
+; SKX-NEXT:    vpaddd %ymm0, %ymm2, %ymm0
 ; SKX-NEXT:    retq
   %x226 = shufflevector <16 x i32> %x225, <16 x i32> %x227, <16 x i32> <i32 0, i32 2, i32 16, i32 18
 , i32 4, i32 6, i32 20, i32 22, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %x228 = shufflevector <16 x i32> %x225, <16 x i32> %x227, <16 x i32> <i32 1, i32 3, i32 17, i32 19
 , i32 5 , i32 7, i32 21,   i32 23, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef,
  i32 undef, i32 undef>
   %x229 = add <16 x i32> %x226, %x228
   ret <16 x i32> %x229
 }
 
 define <16 x float> @fhadd_16_3(<16 x float> %x225, <16 x float> %x227) {
 ; KNL-LABEL: fhadd_16_3:
 ; KNL:       # %bb.0:
 ; KNL-NEXT:    vshufps {{.*#+}} ymm2 = ymm0[0,2],ymm1[0,2],ymm0[4,6],ymm1[4,6]
 ; KNL-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm1[1,3],ymm0[5,7],ymm1[5,7]
-; KNL-NEXT:    vaddps %zmm0, %zmm2, %zmm0
+; KNL-NEXT:    vaddps %ymm0, %ymm2, %ymm0
 ; KNL-NEXT:    retq
 ;
 ; SKX-LABEL: fhadd_16_3:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vshufps {{.*#+}} ymm2 = ymm0[0,2],ymm1[0,2],ymm0[4,6],ymm1[4,6]
 ; SKX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm1[1,3],ymm0[5,7],ymm1[5,7]
-; SKX-NEXT:    vaddps %zmm0, %zmm2, %zmm0
+; SKX-NEXT:    vaddps %ymm0, %ymm2, %ymm0
 ; SKX-NEXT:    retq
   %x226 = shufflevector <16 x float> %x225, <16 x float> %x227, <16 x i32> <i32 0, i32 2, i32 16, i32 18
 , i32 4, i32 6, i32 20, i32 22, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %x228 = shufflevector <16 x float> %x225, <16 x float> %x227, <16 x i32> <i32 1, i32 3, i32 17, i32 19
 , i32 5 , i32 7, i32 21,   i32 23, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %x229 = fadd <16 x float> %x226, %x228
   ret <16 x float> %x229
 }
 
 define <8 x double> @fhadd_16_4(<8 x double> %x225, <8 x double> %x227) {
 ; KNL-LABEL: fhadd_16_4:
 ; KNL:       # %bb.0:
 ; KNL-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
 ; KNL-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
-; KNL-NEXT:    vaddpd %zmm0, %zmm2, %zmm0
+; KNL-NEXT:    vaddpd %ymm0, %ymm2, %ymm0
 ; KNL-NEXT:    retq
 ;
 ; SKX-LABEL: fhadd_16_4:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
 ; SKX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
-; SKX-NEXT:    vaddpd %zmm0, %zmm2, %zmm0
+; SKX-NEXT:    vaddpd %ymm0, %ymm2, %ymm0
 ; SKX-NEXT:    retq
   %x226 = shufflevector <8 x double> %x225, <8 x double> %x227, <8 x i32> <i32 0, i32 8, i32 2, i32 10, i32 undef, i32 undef, i32 undef, i32 undef>
   %x228 = shufflevector <8 x double> %x225, <8 x double> %x227, <8 x i32> <i32 1, i32 9, i32 3, i32 11, i32 undef ,i32 undef, i32 undef, i32 undef>
   %x229 = fadd <8 x double> %x226, %x228
   ret <8 x double> %x229
 }
 
 define <4 x double> @fadd_noundef_low(<8 x double> %x225, <8 x double> %x227) {

test/CodeGen/X86/scalarize-fp.ll

 ; SSE-LABEL: fadd_op1_constant_v4f64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
 ; SSE-NEXT:    addpd %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: fadd_op1_constant_v4f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT:    vaddpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vaddpd %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %v = insertelement <4 x double> undef, double %x, i32 0
   %b = fadd <4 x double> %v, <double 42.0, double undef, double undef, double undef>
   ret <4 x double> %b
 }
 
 define <4 x double> @load_fadd_op1_constant_v4f64(double* %p) nounwind {
 ; SSE-LABEL: load_fadd_op1_constant_v4f64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
 ; SSE-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
 ; SSE-NEXT:    addpd %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: load_fadd_op1_constant_v4f64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT:    vaddpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vaddpd %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %x = load double, double* %p
   %v = insertelement <4 x double> undef, double %x, i32 0
   %b = fadd <4 x double> %v, <double 42.0, double undef, double undef, double undef>
   ret <4 x double> %b
 }
 
 define <4 x double> @fsub_op0_constant_v4f64(double %x) nounwind {
 ; SSE-LABEL: fsub_op0_constant_v4f64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
 ; SSE-NEXT:    subpd %xmm0, %xmm1
 ; SSE-NEXT:    movapd %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: fsub_op0_constant_v4f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT:    vsubpd %ymm0, %ymm1, %ymm0
+; AVX-NEXT:    vsubpd %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %v = insertelement <4 x double> undef, double %x, i32 0
   %b = fsub <4 x double> <double 42.0, double undef, double undef, double undef>, %v
   ret <4 x double> %b
 }
 
 define <4 x double> @load_fsub_op0_constant_v4f64(double* %p) nounwind {
 ; SSE-LABEL: load_fsub_op0_constant_v4f64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
 ; SSE-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
 ; SSE-NEXT:    subpd %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: load_fsub_op0_constant_v4f64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT:    vsubpd %ymm0, %ymm1, %ymm0
+; AVX-NEXT:    vsubpd %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %x = load double, double* %p
   %v = insertelement <4 x double> undef, double %x, i32 0
   %b = fsub <4 x double> <double 42.0, double undef, double undef, double undef>, %v
   ret <4 x double> %b
 }
 
 define <4 x double> @fmul_op1_constant_v4f64(double %x) nounwind {
 ; SSE-LABEL: fmul_op1_constant_v4f64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
 ; SSE-NEXT:    mulpd %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: fmul_op1_constant_v4f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT:    vmulpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vmulpd %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %v = insertelement <4 x double> undef, double %x, i32 0
   %b = fmul <4 x double> %v, <double 42.0, double undef, double undef, double undef>
   ret <4 x double> %b
 }
 
 define <4 x double> @load_fmul_op1_constant_v4f64(double* %p) nounwind {
 ; SSE-LABEL: load_fmul_op1_constant_v4f64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
 ; SSE-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
 ; SSE-NEXT:    mulpd %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: load_fmul_op1_constant_v4f64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT:    vmulpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vmulpd %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %x = load double, double* %p
   %v = insertelement <4 x double> undef, double %x, i32 0
   %b = fmul <4 x double> %v, <double 42.0, double undef, double undef, double undef>
   ret <4 x double> %b
 }
 
 define <4 x double> @fdiv_op1_constant_v4f64(double %x) nounwind {
 ; SSE-LABEL: fdiv_op1_constant_v4f64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
 ; SSE-NEXT:    divpd %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: fdiv_op1_constant_v4f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT:    vdivpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vdivpd %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %v = insertelement <4 x double> undef, double %x, i32 0
   %b = fdiv <4 x double> %v, <double 42.0, double undef, double undef, double undef>
   ret <4 x double> %b
 }
 
 define <4 x double> @load_fdiv_op1_constant_v4f64(double* %p) nounwind {
 ; SSE-LABEL: load_fdiv_op1_constant_v4f64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
 ; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
 ; SSE-NEXT:    divpd %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: load_fdiv_op1_constant_v4f64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT:    vdivpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vdivpd %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %x = load double, double* %p
   %v = insertelement <4 x double> undef, double %x, i32 0
   %b = fdiv <4 x double> %v, <double 42.0, double undef, double undef, double undef>
   ret <4 x double> %b
 }
 
 define <4 x double> @fdiv_op0_constant_v4f64(double %x) nounwind {
 ; SSE-LABEL: fdiv_op0_constant_v4f64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
 ; SSE-NEXT:    divpd %xmm0, %xmm1
 ; SSE-NEXT:    movapd %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: fdiv_op0_constant_v4f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT:    vdivpd %ymm0, %ymm1, %ymm0
+; AVX-NEXT:    vdivpd %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %v = insertelement <4 x double> undef, double %x, i32 0
   %b = fdiv <4 x double> <double 42.0, double undef, double undef, double undef>, %v
   ret <4 x double> %b
 }
 
 define <4 x double> @load_fdiv_op0_constant_v4f64(double* %p) nounwind {
 ; SSE-LABEL: load_fdiv_op0_constant_v4f64:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
 ; SSE-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
 ; SSE-NEXT:    divpd %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: load_fdiv_op0_constant_v4f64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; AVX-NEXT:    vdivpd %ymm0, %ymm1, %ymm0
+; AVX-NEXT:    vdivpd %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %x = load double, double* %p
   %v = insertelement <4 x double> undef, double %x, i32 0
   %b = fdiv <4 x double> <double 42.0, double undef, double undef, double undef>, %v
   ret <4 x double> %b
 }
 

test/CodeGen/X86/vector-partial-undef.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefix=SSE
 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefix=AVX
 
 ; xor undef, undef --> 0 because it's not worth fighting to make that return undef?
 
 define <4 x i64> @xor_insert_insert(<2 x i64> %x, <2 x i64> %y) {
 ; SSE-LABEL: xor_insert_insert:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    xorps %xmm1, %xmm0
 ; SSE-NEXT:    xorps %xmm1, %xmm1
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: xor_insert_insert:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    # kill: def $xmm1 killed $xmm1 def $ymm1
-; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
-; AVX-NEXT:    vxorps %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vxorps %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %xw = shufflevector <2 x i64> %x, <2 x i64> undef, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
   %yw = shufflevector <2 x i64> %y, <2 x i64> undef, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
   %r = xor <4 x i64> %xw, %yw
   ret <4 x i64> %r
 }
 
 define <4 x i64> @xor_insert_insert_high_half(<2 x i64> %x, <2 x i64> %y) {
 ; SSE-LABEL: xor_insert_insert_high_half:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    xorps %xmm0, %xmm1
 ; SSE-NEXT:    xorps %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: xor_insert_insert_high_half:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
-; AVX-NEXT:    vxorps %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vxorps %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
 ; AVX-NEXT:    retq
   %xw = shufflevector <2 x i64> %x, <2 x i64> undef, <4 x i32> <i32 undef, i32 undef, i32 0, i32 1>
   %yw = shufflevector <2 x i64> %y, <2 x i64> undef, <4 x i32> <i32 undef, i32 undef, i32 0, i32 1>
   %r = xor <4 x i64> %xw, %yw
   ret <4 x i64> %r
 }
 
 ; All elements of the add are undefined: