[DAGCombiner] sink target-supported FP<->int cast op after concat vectors

Try to combine N short vector cast ops into 1 wide vector cast op:
concat (cast X), (cast Y)... -> cast (concat X, Y...)

This is part of solving PR45794:
https://bugs.llvm.org/show_bug.cgi?id=45794

As noted in the code comment, this is uglier than I was hoping because
the opcode determines whether we pass the source or destination type
to isOperationLegalOrCustom(). Also IIUC, there's no way to validate
what the other (dest or src) type is. Without the extra legality check
on that, there's an ARM regression test in:
test/CodeGen/ARM/isel-v8i32-crash.ll
...that will crash trying to lower an unsupported v8f32 to v8i16.

Differential Revision: https://reviews.llvm.org/D79360

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index d8c6b51..d044c7b 100644 (file)
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -18546,6 +18546,69 @@ static SDValue combineConcatVectorOfExtracts(SDNode *N, SelectionDAG &DAG) {
                                      DAG.getBitcast(VT, SV1), Mask, DAG);
 }
 
+static SDValue combineConcatVectorOfCasts(SDNode *N, SelectionDAG &DAG) {
+  unsigned CastOpcode = N->getOperand(0).getOpcode();
+  switch (CastOpcode) {
+  case ISD::SINT_TO_FP:
+  case ISD::UINT_TO_FP:
+  case ISD::FP_TO_SINT:
+  case ISD::FP_TO_UINT:
+    // TODO: Allow more opcodes?
+    //  case ISD::BITCAST:
+    //  case ISD::TRUNCATE:
+    //  case ISD::ZERO_EXTEND:
+    //  case ISD::SIGN_EXTEND:
+    //  case ISD::FP_EXTEND:
+    break;
+  default:
+    return SDValue();
+  }
+
+  EVT SrcVT = N->getOperand(0).getOperand(0).getValueType();
+  if (!SrcVT.isVector())
+    return SDValue();
+
+  // All operands of the concat must be the same kind of cast from the same
+  // source type.
+  SmallVector<SDValue, 4> SrcOps;
+  for (SDValue Op : N->ops()) {
+    if (Op.getOpcode() != CastOpcode || !Op.hasOneUse() ||
+        Op.getOperand(0).getValueType() != SrcVT)
+      return SDValue();
+    SrcOps.push_back(Op.getOperand(0));
+  }
+
+  // The wider cast must be supported by the target. This is unusual because
+  // the operation support type parameter depends on the opcode. In addition,
+  // check the other type in the cast to make sure this is really legal.
+  EVT VT = N->getValueType(0);
+  EVT SrcEltVT = SrcVT.getVectorElementType();
+  unsigned NumElts = SrcVT.getVectorElementCount().Min * N->getNumOperands();
+  EVT ConcatSrcVT = EVT::getVectorVT(*DAG.getContext(), SrcEltVT, NumElts);
+  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+  switch (CastOpcode) {
+  case ISD::SINT_TO_FP:
+  case ISD::UINT_TO_FP:
+    if (!TLI.isOperationLegalOrCustom(CastOpcode, ConcatSrcVT) ||
+        !TLI.isTypeLegal(VT))
+      return SDValue();
+    break;
+  case ISD::FP_TO_SINT:
+  case ISD::FP_TO_UINT:
+    if (!TLI.isOperationLegalOrCustom(CastOpcode, VT) ||
+        !TLI.isTypeLegal(ConcatSrcVT))
+      return SDValue();
+    break;
+  default:
+    llvm_unreachable("Unexpected cast opcode");
+  }
+
+  // concat (cast X), (cast Y)... -> cast (concat X, Y...)
+  SDLoc DL(N);
+  SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, DL, ConcatSrcVT, SrcOps);
+  return DAG.getNode(CastOpcode, DL, VT, NewConcat);
+}
+
 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
   // If we only have one input vector, we don't need to do any concatenation.
   if (N->getNumOperands() == 1)
@@ -18673,6 +18736,9 @@ SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
     if (SDValue V = combineConcatVectorOfExtracts(N, DAG))
       return V;
 
+  if (SDValue V = combineConcatVectorOfCasts(N, DAG))
+    return V;
+
   // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
   // nodes often generate nop CONCAT_VECTOR nodes.
   // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that

diff --git a/llvm/test/CodeGen/X86/avx-shift.ll b/llvm/test/CodeGen/X86/avx-shift.ll
index 443204c..1d5ec2d 100644 (file)
--- a/llvm/test/CodeGen/X86/avx-shift.ll
+++ b/llvm/test/CodeGen/X86/avx-shift.ll
@@ -167,15 +167,14 @@ define <32 x i8> @vshift12(<32 x i8> %a) {
 define <8 x i32> @vshift08(<8 x i32> %a)  {
 ; CHECK-LABEL: vshift08:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpslld $23, %xmm0, %xmm1
+; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; CHECK-NEXT:    vpslld $23, %xmm1, %xmm1
 ; CHECK-NEXT:    vmovdqa {{.*#+}} xmm2 = [1065353216,1065353216,1065353216,1065353216]
 ; CHECK-NEXT:    vpaddd %xmm2, %xmm1, %xmm1
-; CHECK-NEXT:    vcvttps2dq %xmm1, %xmm1
-; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; CHECK-NEXT:    vpslld $23, %xmm0, %xmm0
 ; CHECK-NEXT:    vpaddd %xmm2, %xmm0, %xmm0
-; CHECK-NEXT:    vcvttps2dq %xmm0, %xmm0
-; CHECK-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; CHECK-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; CHECK-NEXT:    vcvttps2dq %ymm0, %ymm0
 ; CHECK-NEXT:    retq
   %bitop = shl <8 x i32> <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>, %a
   ret <8 x i32> %bitop

diff --git a/llvm/test/CodeGen/X86/concat-cast.ll b/llvm/test/CodeGen/X86/concat-cast.ll
index 6162baf..8f8dee1 100644 (file)
--- a/llvm/test/CodeGen/X86/concat-cast.ll
+++ b/llvm/test/CodeGen/X86/concat-cast.ll
@@ -8,16 +8,14 @@
 define <4 x float> @sitofp_v4i32_v4f32(<2 x i32> %x, <2 x i32> %y) {
 ; SSE-LABEL: sitofp_v4i32_v4f32:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    cvtdq2ps %xmm0, %xmm0
-; SSE-NEXT:    cvtdq2ps %xmm1, %xmm1
 ; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE-NEXT:    cvtdq2ps %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: sitofp_v4i32_v4f32:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vcvtdq2ps %xmm0, %xmm0
-; AVX-NEXT:    vcvtdq2ps %xmm1, %xmm1
 ; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX-NEXT:    vcvtdq2ps %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %s0 = sitofp <2 x i32> %x to <2 x float>
   %s1 = sitofp <2 x i32> %y to <2 x float>
@@ -28,68 +26,55 @@ define <4 x float> @sitofp_v4i32_v4f32(<2 x i32> %x, <2 x i32> %y) {
 define <4 x float> @uitofp_v4i32_v4f32(<2 x i32> %x, <2 x i32> %y) {
 ; SSE2-LABEL: uitofp_v4i32_v4f32:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    xorpd %xmm2, %xmm2
-; SSE2-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSE2-NEXT:    movapd {{.*#+}} xmm3 = [4.503599627370496E+15,4.503599627370496E+15]
-; SSE2-NEXT:    orpd %xmm3, %xmm0
-; SSE2-NEXT:    subpd %xmm3, %xmm0
-; SSE2-NEXT:    cvtpd2ps %xmm0, %xmm0
-; SSE2-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
-; SSE2-NEXT:    orpd %xmm3, %xmm1
-; SSE2-NEXT:    subpd %xmm3, %xmm1
-; SSE2-NEXT:    cvtpd2ps %xmm1, %xmm1
-; SSE2-NEXT:    unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535]
+; SSE2-NEXT:    pand %xmm0, %xmm1
+; SSE2-NEXT:    por {{.*}}(%rip), %xmm1
+; SSE2-NEXT:    psrld $16, %xmm0
+; SSE2-NEXT:    por {{.*}}(%rip), %xmm0
+; SSE2-NEXT:    subps {{.*}}(%rip), %xmm0
+; SSE2-NEXT:    addps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: uitofp_v4i32_v4f32:
 ; SSE4:       # %bb.0:
-; SSE4-NEXT:    pmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; SSE4-NEXT:    movdqa {{.*#+}} xmm2 = [4.503599627370496E+15,4.503599627370496E+15]
-; SSE4-NEXT:    por %xmm2, %xmm0
-; SSE4-NEXT:    subpd %xmm2, %xmm0
-; SSE4-NEXT:    cvtpd2ps %xmm0, %xmm0
-; SSE4-NEXT:    pmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
-; SSE4-NEXT:    por %xmm2, %xmm1
-; SSE4-NEXT:    subpd %xmm2, %xmm1
-; SSE4-NEXT:    cvtpd2ps %xmm1, %xmm1
-; SSE4-NEXT:    unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE4-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE4-NEXT:    movdqa {{.*#+}} xmm1 = [1258291200,1258291200,1258291200,1258291200]
+; SSE4-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7]
+; SSE4-NEXT:    psrld $16, %xmm0
+; SSE4-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
+; SSE4-NEXT:    subps {{.*}}(%rip), %xmm0
+; SSE4-NEXT:    addps %xmm1, %xmm0
 ; SSE4-NEXT:    retq
 ;
 ; AVX1-LABEL: uitofp_v4i32_v4f32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [4.503599627370496E+15,4.503599627370496E+15]
-; AVX1-NEXT:    vpor %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vsubpd %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vcvtpd2ps %xmm0, %xmm0
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
-; AVX1-NEXT:    vpor %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vsubpd %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vcvtpd2ps %xmm1, %xmm1
-; AVX1-NEXT:    vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
+; AVX1-NEXT:    vsubps {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: uitofp_v4i32_v4f32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm2 = [4.503599627370496E+15,4.503599627370496E+15]
-; AVX2-NEXT:    vpor %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vsubpd %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vcvtpd2ps %xmm0, %xmm0
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
-; AVX2-NEXT:    vpor %xmm2, %xmm1, %xmm1
-; AVX2-NEXT:    vsubpd %xmm2, %xmm1, %xmm1
-; AVX2-NEXT:    vcvtpd2ps %xmm1, %xmm1
-; AVX2-NEXT:    vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [1258291200,1258291200,1258291200,1258291200]
+; AVX2-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7]
+; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [1392508928,1392508928,1392508928,1392508928]
+; AVX2-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3],xmm0[4],xmm2[5],xmm0[6],xmm2[7]
+; AVX2-NEXT:    vbroadcastss {{.*#+}} xmm2 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; AVX2-NEXT:    vsubps %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: uitofp_v4i32_v4f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    # kill: def $xmm1 killed $xmm1 def $zmm1
-; AVX512-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; AVX512-NEXT:    vcvtudq2ps %zmm0, %zmm0
-; AVX512-NEXT:    vcvtudq2ps %zmm1, %zmm1
 ; AVX512-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX512-NEXT:    vcvtudq2ps %zmm0, %zmm0
+; AVX512-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
   %s0 = uitofp <2 x i32> %x to <2 x float>
@@ -101,16 +86,14 @@ define <4 x float> @uitofp_v4i32_v4f32(<2 x i32> %x, <2 x i32> %y) {
 define <4 x i32> @fptosi_v4f32_v4i32(<2 x float> %x, <2 x float> %y) {
 ; SSE-LABEL: fptosi_v4f32_v4i32:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    cvttps2dq %xmm0, %xmm0
-; SSE-NEXT:    cvttps2dq %xmm1, %xmm1
 ; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE-NEXT:    cvttps2dq %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: fptosi_v4f32_v4i32:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vcvttps2dq %xmm0, %xmm0
-; AVX-NEXT:    vcvttps2dq %xmm1, %xmm1
 ; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX-NEXT:    vcvttps2dq %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %s0 = fptosi <2 x float> %x to <2 x i32>
   %s1 = fptosi <2 x float> %y to <2 x i32>
@@ -208,11 +191,9 @@ define <4 x i32> @fptoui_v4f32_v4i32(<2 x float> %x, <2 x float> %y) {
 ;
 ; AVX512-LABEL: fptoui_v4f32_v4i32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    # kill: def $xmm1 killed $xmm1 def $zmm1
-; AVX512-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; AVX512-NEXT:    vcvttps2udq %zmm0, %zmm0
-; AVX512-NEXT:    vcvttps2udq %zmm1, %zmm1
 ; AVX512-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX512-NEXT:    vcvttps2udq %zmm0, %zmm0
+; AVX512-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
   %s0 = fptoui <2 x float> %x to <2 x i32>
@@ -230,9 +211,8 @@ define <4 x double> @sitofp_v4i32_v4f64(<2 x i32> %x, <2 x i32> %y) {
 ;
 ; AVX-LABEL: sitofp_v4i32_v4f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vcvtdq2pd %xmm0, %xmm0
-; AVX-NEXT:    vcvtdq2pd %xmm1, %xmm1
-; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX-NEXT:    vcvtdq2pd %xmm0, %ymm0
 ; AVX-NEXT:    retq
   %s0 = sitofp <2 x i32> %x to <2 x double>
   %s1 = sitofp <2 x i32> %y to <2 x double>
@@ -267,34 +247,27 @@ define <4 x double> @uitofp_v4i32_v4f64(<2 x i32> %x, <2 x i32> %y) {
 ; AVX1-LABEL: uitofp_v4i32_v4f64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [4.503599627370496E+15,4.503599627370496E+15]
-; AVX1-NEXT:    vpor %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vsubpd %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
-; AVX1-NEXT:    vpor %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vsubpd %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT:    vbroadcastsd {{.*#+}} ymm1 = [4.503599627370496E+15,4.503599627370496E+15,4.503599627370496E+15,4.503599627370496E+15]
+; AVX1-NEXT:    vorpd %ymm1, %ymm0, %ymm0
+; AVX1-NEXT:    vsubpd %ymm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: uitofp_v4i32_v4f64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm2 = [4.503599627370496E+15,4.503599627370496E+15]
-; AVX2-NEXT:    vpor %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vsubpd %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero
-; AVX2-NEXT:    vpor %xmm2, %xmm1, %xmm1
-; AVX2-NEXT:    vsubpd %xmm2, %xmm1, %xmm1
-; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX2-NEXT:    vpmovzxdq {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [4.503599627370496E+15,4.503599627370496E+15,4.503599627370496E+15,4.503599627370496E+15]
+; AVX2-NEXT:    vpor %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vsubpd %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: uitofp_v4i32_v4f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    # kill: def $xmm1 killed $xmm1 def $ymm1
-; AVX512-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX512-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; AVX512-NEXT:    vcvtudq2pd %ymm0, %zmm0
-; AVX512-NEXT:    vcvtudq2pd %ymm1, %zmm1
-; AVX512-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 killed $zmm0
 ; AVX512-NEXT:    retq
   %s0 = uitofp <2 x i32> %x to <2 x double>
   %s1 = uitofp <2 x i32> %y to <2 x double>
@@ -312,9 +285,10 @@ define <4 x i32> @fptosi_v4f64_v4i32(<2 x double> %x, <2 x double> %y) {
 ;
 ; AVX-LABEL: fptosi_v4f64_v4i32:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vcvttpd2dq %xmm0, %xmm0
-; AVX-NEXT:    vcvttpd2dq %xmm1, %xmm1
-; AVX-NEXT:    vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vcvttpd2dq %ymm0, %xmm0
+; AVX-NEXT:    vzeroupper
 ; AVX-NEXT:    retq
   %s0 = fptosi <2 x double> %x to <2 x i32>
   %s1 = fptosi <2 x double> %y to <2 x i32>
@@ -404,11 +378,10 @@ define <4 x i32> @fptoui_v4f64_v4i32(<2 x double> %x, <2 x double> %y) {
 ;
 ; AVX512-LABEL: fptoui_v4f64_v4i32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    # kill: def $xmm1 killed $xmm1 def $zmm1
-; AVX512-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
+; AVX512-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX512-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX512-NEXT:    vcvttpd2udq %zmm0, %ymm0
-; AVX512-NEXT:    vcvttpd2udq %zmm1, %ymm1
-; AVX512-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; AVX512-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
   %s0 = fptoui <2 x double> %x to <2 x i32>
@@ -417,6 +390,8 @@ define <4 x i32> @fptoui_v4f64_v4i32(<2 x double> %x, <2 x double> %y) {
   ret <4 x i32> %r
 }
 
+; Negative test
+
 define <4 x float> @mismatch_tofp_v4i32_v4f32(<2 x i32> %x, <2 x i32> %y) {
 ; SSE2-LABEL: mismatch_tofp_v4i32_v4f32:
 ; SSE2:       # %bb.0:
@@ -477,6 +452,8 @@ define <4 x float> @mismatch_tofp_v4i32_v4f32(<2 x i32> %x, <2 x i32> %y) {
   ret <4 x float> %r
 }
 
+; Negative test
+
 define <4 x float> @sitofp_v4i32_v4f32_extra_use(<2 x i32> %x, <2 x i32> %y, <2 x float>* %p) {
 ; SSE-LABEL: sitofp_v4i32_v4f32_extra_use:
 ; SSE:       # %bb.0: