[X86] Support `X86ISD::PCMPEQ` and `X86ISD::PCMPGT` in ComputeKnownBits

[goldstein.w.n]

Summary

These functions where missing support but are used enough that it
makes sense to track them.

[RKSimon]

Nice! LGTM with one minor

llvm/lib/Target/X86/X86ISelLowering.cpp

38573

(style) Don't use auto

[goldstein.w.n]

remove auto

[nikic]

Shouldn't this be a DAGCombine instead? It seems pointless to make computeKnownBits() recurse past an instruction that can only return either unknown or constant. If it returns constant, we want to just replace the instruction, and otherwise we get zero information.

[RKSimon]

In D148963#4304551, @nikic wrote:

Shouldn't this be a DAGCombine instead? It seems pointless to make computeKnownBits() recurse past an instruction that can only return either unknown or constant. If it returns constant, we want to just replace the instruction, and otherwise we get zero information.

At least part of this is only matching cases because it has access to the DemandedElts mask

[nikic]

In D148963#4304568, @RKSimon wrote:

In D148963#4304551, @nikic wrote:

Shouldn't this be a DAGCombine instead? It seems pointless to make computeKnownBits() recurse past an instruction that can only return either unknown or constant. If it returns constant, we want to just replace the instruction, and otherwise we get zero information.

At least part of this is only matching cases because it has access to the DemandedElts mask

Hm, through which pathway does this actually end up folding in the non-trivial DemandedElts case? I'd have guessed SimplifyDemandedBits, but it doesn't seem to have a generic constant known bits to constant combine.

[RKSimon]

@goldstein.w.n In case you missed it, this has somehow exposed a miscompile: https://github.com/llvm/llvm-project/issues/67287

[goldstein.w.n]

In D148963#4653118, @RKSimon wrote:

@goldstein.w.n In case you missed it, this has somehow exposed a miscompile: https://github.com/llvm/llvm-project/issues/67287

Oh, had missed it. Looking into it.

llvm/lib/Target/X86/X86ISelLowering.cpp

if (Op.getResNo() == 0) {
       Known = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
       Known2 = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
       Known &= Known2;
     }
     break;
   }
   case X86ISD::ANDNP: {
     KnownBits Known2;
     Known = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
     Known2 = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
 
     // ANDNP = (~X & Y);
     Known.One &= Known2.Zero;
     Known.Zero |= Known2.One;
     break;
   }
   case X86ISD::FOR: {
     KnownBits Known2;
     Known = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
     Known2 = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
 
     Known |= Known2;
     break;
   }
   case X86ISD::PSADBW: {
     assert(VT.getScalarType() == MVT::i64 &&
            Op.getOperand(0).getValueType().getScalarType() == MVT::i8 &&
            "Unexpected PSADBW types");
 
     // PSADBW - fills low 16 bits and zeros upper 48 bits of each i64 result.
     Known.Zero.setBitsFrom(16);
     break;
   }
+  case X86ISD::PCMPGT:
+  case X86ISD::PCMPEQ: {
+    KnownBits KnownLhs =
+        DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
+    KnownBits KnownRhs =
+        DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
+    std::optional<bool> Res = Opc == X86ISD::PCMPEQ
+                                  ? KnownBits::eq(KnownLhs, KnownRhs)
+                                  : KnownBits::sgt(KnownLhs, KnownRhs);
+    if (Res) {
+      if (*Res)
+        Known.setAllOnes();
+      else
+        Known.setAllZero();
+    }
+    break;
+  }
   case X86ISD::PMULUDQ: {
     KnownBits Known2;
     Known = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
     Known2 = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
 
     Known = Known.trunc(BitWidth / 2).zext(BitWidth);
     Known2 = Known2.trunc(BitWidth / 2).zext(BitWidth);
     Known = KnownBits::mul(Known, Known2);

llvm/test/CodeGen/X86/combine-srl.ll

 
 define <4 x i32> @combine_vec_lshr_lzcnt_bit1(<4 x i32> %x) {
 ; SSE-LABEL: combine_vec_lshr_lzcnt_bit1:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE-NEXT:    movdqa {{.*#+}} xmm1 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
 ; SSE-NEXT:    movdqa %xmm1, %xmm2
 ; SSE-NEXT:    pshufb %xmm0, %xmm2
-; SSE-NEXT:    movdqa %xmm0, %xmm3
-; SSE-NEXT:    psrlw $4, %xmm3
-; SSE-NEXT:    pxor %xmm4, %xmm4
-; SSE-NEXT:    pshufb %xmm3, %xmm1
-; SSE-NEXT:    pcmpeqb %xmm4, %xmm3
-; SSE-NEXT:    pand %xmm2, %xmm3
-; SSE-NEXT:    paddb %xmm1, %xmm3
-; SSE-NEXT:    movdqa %xmm0, %xmm1
-; SSE-NEXT:    pcmpeqb %xmm4, %xmm1
-; SSE-NEXT:    psrlw $8, %xmm1
-; SSE-NEXT:    pand %xmm3, %xmm1
-; SSE-NEXT:    psrlw $8, %xmm3
-; SSE-NEXT:    paddw %xmm1, %xmm3
-; SSE-NEXT:    pcmpeqw %xmm4, %xmm0
+; SSE-NEXT:    psrlw $4, %xmm0
+; SSE-NEXT:    pxor %xmm3, %xmm3
+; SSE-NEXT:    pshufb %xmm0, %xmm1
+; SSE-NEXT:    pcmpeqb %xmm3, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    paddb %xmm1, %xmm0
+; SSE-NEXT:    movdqa {{.*#+}} xmm1 = [255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0]
+; SSE-NEXT:    pand %xmm0, %xmm1
+; SSE-NEXT:    psrlw $8, %xmm0
+; SSE-NEXT:    paddw %xmm1, %xmm0
+; SSE-NEXT:    pblendw {{.*#+}} xmm3 = xmm0[0],xmm3[1],xmm0[2],xmm3[3],xmm0[4],xmm3[5],xmm0[6],xmm3[7]
 ; SSE-NEXT:    psrld $16, %xmm0
-; SSE-NEXT:    pand %xmm3, %xmm0
-; SSE-NEXT:    psrld $16, %xmm3
 ; SSE-NEXT:    paddd %xmm3, %xmm0
 ; SSE-NEXT:    psrld $5, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: combine_vec_lshr_lzcnt_bit1:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
 ; AVX-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
 ; AVX-NEXT:    vpshufb %xmm0, %xmm1, %xmm2
-; AVX-NEXT:    vpsrlw $4, %xmm0, %xmm3
-; AVX-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX-NEXT:    vpcmpeqb %xmm4, %xmm3, %xmm5
-; AVX-NEXT:    vpand %xmm5, %xmm2, %xmm2
-; AVX-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
-; AVX-NEXT:    vpaddb %xmm1, %xmm2, %xmm1
-; AVX-NEXT:    vpcmpeqb %xmm4, %xmm0, %xmm2
-; AVX-NEXT:    vpsrlw $8, %xmm2, %xmm2
-; AVX-NEXT:    vpand %xmm2, %xmm1, %xmm2
-; AVX-NEXT:    vpsrlw $8, %xmm1, %xmm1
-; AVX-NEXT:    vpaddw %xmm2, %xmm1, %xmm1
-; AVX-NEXT:    vpcmpeqw %xmm4, %xmm0, %xmm0
+; AVX-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX-NEXT:    vpcmpeqb %xmm3, %xmm0, %xmm4
+; AVX-NEXT:    vpand %xmm4, %xmm2, %xmm2
+; AVX-NEXT:    vpshufb %xmm0, %xmm1, %xmm0
+; AVX-NEXT:    vpaddb %xmm0, %xmm2, %xmm0
+; AVX-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1
+; AVX-NEXT:    vpsrlw $8, %xmm0, %xmm0
+; AVX-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],xmm3[1],xmm0[2],xmm3[3],xmm0[4],xmm3[5],xmm0[6],xmm3[7]
 ; AVX-NEXT:    vpsrld $16, %xmm0, %xmm0
-; AVX-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    vpsrld $16, %xmm1, %xmm1
-; AVX-NEXT:    vpaddd %xmm0, %xmm1, %xmm0
+; AVX-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    vpsrld $5, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %1 = and <4 x i32> %x, <i32 4, i32 32, i32 64, i32 128>
   %2 = call <4 x i32> @llvm.ctlz.v4i32(<4 x i32> %1, i1 0)
   %3 = lshr <4 x i32> %2, <i32 5, i32 5, i32 5, i32 5>
   ret <4 x i32> %3
 }
 declare <4 x i32> @llvm.ctlz.v4i32(<4 x i32>, i1)

llvm/test/CodeGen/X86/fpclamptosat_vec.ll

 ; CHECK-NEXT:    movq %rax, %rdx
 ; CHECK-NEXT:    sarq $63, %rdx
 ; CHECK-NEXT:    andq %rcx, %rdx
 ; CHECK-NEXT:    orq %rax, %rdx
 ; CHECK-NEXT:    movq %rdx, %xmm0
 ; CHECK-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
 ; CHECK-NEXT:    movdqa {{.*#+}} xmm0 = [9223372039002259456,9223372039002259456]
 ; CHECK-NEXT:    pxor %xmm1, %xmm0
-; CHECK-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259455,9223372039002259455]
-; CHECK-NEXT:    pcmpgtd %xmm0, %xmm2
-; CHECK-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; CHECK-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
+; CHECK-NEXT:    movdqa {{.*#+}} xmm3 = [2147483647,2147483647,2147483647,2147483647]
+; CHECK-NEXT:    pcmpgtd %xmm2, %xmm3
 ; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
 ; CHECK-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; CHECK-NEXT:    pand %xmm3, %xmm0
-; CHECK-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; CHECK-NEXT:    por %xmm0, %xmm2
-; CHECK-NEXT:    pand %xmm2, %xmm1
-; CHECK-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; CHECK-NEXT:    por %xmm1, %xmm2
-; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
+; CHECK-NEXT:    pand %xmm0, %xmm1
+; CHECK-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; CHECK-NEXT:    por %xmm1, %xmm0
+; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; CHECK-NEXT:    retq
 entry:
   %conv = fptoui <2 x double> %x to <2 x i64>
   %0 = icmp ult <2 x i64> %conv, <i64 4294967295, i64 4294967295>
   %spec.store.select = select <2 x i1> %0, <2 x i64> %conv, <2 x i64> <i64 4294967295, i64 4294967295>
   %conv6 = trunc <2 x i64> %spec.store.select to <2 x i32>
   ret <2 x i32> %conv6
 }

 ; CHECK-NEXT:    andq %rcx, %rdx
 ; CHECK-NEXT:    orq %rax, %rdx
 ; CHECK-NEXT:    movq %rdx, %xmm0
 ; CHECK-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm3[0]
 ; CHECK-NEXT:    movdqa {{.*#+}} xmm2 = [4294967295,4294967295]
 ; CHECK-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; CHECK-NEXT:    movdqa %xmm0, %xmm4
 ; CHECK-NEXT:    pxor %xmm3, %xmm4
-; CHECK-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
-; CHECK-NEXT:    pcmpeqd %xmm3, %xmm5
-; CHECK-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259455,9223372039002259455]
+; CHECK-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; CHECK-NEXT:    movdqa {{.*#+}} xmm6 = [2147483647,2147483647,2147483647,2147483647]
 ; CHECK-NEXT:    movdqa %xmm6, %xmm7
-; CHECK-NEXT:    pcmpgtd %xmm4, %xmm7
-; CHECK-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[0,0,2,2]
-; CHECK-NEXT:    pand %xmm5, %xmm4
-; CHECK-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
-; CHECK-NEXT:    por %xmm4, %xmm5
-; CHECK-NEXT:    pand %xmm5, %xmm0
-; CHECK-NEXT:    pandn %xmm2, %xmm5
-; CHECK-NEXT:    por %xmm0, %xmm5
+; CHECK-NEXT:    pcmpgtd %xmm5, %xmm7
+; CHECK-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; CHECK-NEXT:    pcmpeqd %xmm3, %xmm4
+; CHECK-NEXT:    pand %xmm7, %xmm4
+; CHECK-NEXT:    pand %xmm4, %xmm0
+; CHECK-NEXT:    pandn %xmm2, %xmm4
+; CHECK-NEXT:    por %xmm0, %xmm4
 ; CHECK-NEXT:    movdqa %xmm1, %xmm0
 ; CHECK-NEXT:    pxor %xmm3, %xmm0
-; CHECK-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
-; CHECK-NEXT:    pcmpeqd %xmm3, %xmm4
-; CHECK-NEXT:    pcmpgtd %xmm0, %xmm6
-; CHECK-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[0,0,2,2]
-; CHECK-NEXT:    pand %xmm4, %xmm3
-; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
-; CHECK-NEXT:    por %xmm3, %xmm0
+; CHECK-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,0,2,2]
+; CHECK-NEXT:    pcmpgtd %xmm5, %xmm6
+; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; CHECK-NEXT:    pcmpeqd %xmm3, %xmm0
+; CHECK-NEXT:    pand %xmm6, %xmm0
 ; CHECK-NEXT:    pand %xmm0, %xmm1
 ; CHECK-NEXT:    pandn %xmm2, %xmm0
 ; CHECK-NEXT:    por %xmm1, %xmm0
-; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm5[0,2]
+; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm4[0,2]
 ; CHECK-NEXT:    retq
 entry:
   %conv = fptoui <4 x float> %x to <4 x i64>
   %0 = icmp ult <4 x i64> %conv, <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295>
   %spec.store.select = select <4 x i1> %0, <4 x i64> %conv, <4 x i64> <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295>
   %conv6 = trunc <4 x i64> %spec.store.select to <4 x i32>
   ret <4 x i32> %conv6
 }

 ; CHECK-NEXT:    orq %rax, %rdx
 ; CHECK-NEXT:    movq %rdx, %xmm0
 ; CHECK-NEXT:    punpcklqdq (%rsp), %xmm0 # 16-byte Folded Reload
 ; CHECK-NEXT:    # xmm0 = xmm0[0],mem[0]
 ; CHECK-NEXT:    movdqa {{.*#+}} xmm1 = [4294967295,4294967295]
 ; CHECK-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259456,9223372039002259456]
 ; CHECK-NEXT:    movdqa %xmm0, %xmm3
 ; CHECK-NEXT:    pxor %xmm2, %xmm3
-; CHECK-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
-; CHECK-NEXT:    pcmpeqd %xmm2, %xmm4
-; CHECK-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259455,9223372039002259455]
+; CHECK-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
+; CHECK-NEXT:    movdqa {{.*#+}} xmm5 = [2147483647,2147483647,2147483647,2147483647]
 ; CHECK-NEXT:    movdqa %xmm5, %xmm6
-; CHECK-NEXT:    pcmpgtd %xmm3, %xmm6
-; CHECK-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[0,0,2,2]
-; CHECK-NEXT:    pand %xmm4, %xmm3
-; CHECK-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
-; CHECK-NEXT:    por %xmm3, %xmm4
-; CHECK-NEXT:    pand %xmm4, %xmm0
-; CHECK-NEXT:    pandn %xmm1, %xmm4
-; CHECK-NEXT:    por %xmm0, %xmm4
+; CHECK-NEXT:    pcmpgtd %xmm4, %xmm6
+; CHECK-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; CHECK-NEXT:    pcmpeqd %xmm2, %xmm3
+; CHECK-NEXT:    pand %xmm6, %xmm3
+; CHECK-NEXT:    pand %xmm3, %xmm0
+; CHECK-NEXT:    pandn %xmm1, %xmm3
+; CHECK-NEXT:    por %xmm0, %xmm3
 ; CHECK-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
 ; CHECK-NEXT:    movdqa %xmm6, %xmm0
 ; CHECK-NEXT:    pxor %xmm2, %xmm0
-; CHECK-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
-; CHECK-NEXT:    pcmpeqd %xmm2, %xmm3
-; CHECK-NEXT:    pcmpgtd %xmm0, %xmm5
-; CHECK-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[0,0,2,2]
-; CHECK-NEXT:    pand %xmm3, %xmm2
-; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3]
-; CHECK-NEXT:    por %xmm2, %xmm0
-; CHECK-NEXT:    movdqa %xmm6, %xmm2
-; CHECK-NEXT:    pand %xmm0, %xmm2
+; CHECK-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
+; CHECK-NEXT:    pcmpgtd %xmm4, %xmm5
+; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; CHECK-NEXT:    pcmpeqd %xmm2, %xmm0
+; CHECK-NEXT:    pand %xmm5, %xmm0
+; CHECK-NEXT:    pand %xmm0, %xmm6
 ; CHECK-NEXT:    pandn %xmm1, %xmm0
-; CHECK-NEXT:    por %xmm2, %xmm0
-; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm4[0,2]
+; CHECK-NEXT:    por %xmm6, %xmm0
+; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm3[0,2]
 ; CHECK-NEXT:    addq $72, %rsp
 ; CHECK-NEXT:    .cfi_def_cfa_offset 8
 ; CHECK-NEXT:    retq
 entry:
   %conv = fptoui <4 x half> %x to <4 x i64>
   %0 = icmp ult <4 x i64> %conv, <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295>
   %spec.store.select = select <4 x i1> %0, <4 x i64> %conv, <4 x i64> <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295>
   %conv6 = trunc <4 x i64> %spec.store.select to <4 x i32>

 ; CHECK-NEXT:    movq %rax, %rdx
 ; CHECK-NEXT:    sarq $63, %rdx
 ; CHECK-NEXT:    andq %rcx, %rdx
 ; CHECK-NEXT:    orq %rax, %rdx
 ; CHECK-NEXT:    movq %rdx, %xmm0
 ; CHECK-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
 ; CHECK-NEXT:    movdqa {{.*#+}} xmm0 = [9223372039002259456,9223372039002259456]
 ; CHECK-NEXT:    pxor %xmm1, %xmm0
-; CHECK-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259455,9223372039002259455]
-; CHECK-NEXT:    pcmpgtd %xmm0, %xmm2
-; CHECK-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; CHECK-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
+; CHECK-NEXT:    movdqa {{.*#+}} xmm3 = [2147483647,2147483647,2147483647,2147483647]
+; CHECK-NEXT:    pcmpgtd %xmm2, %xmm3
 ; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
 ; CHECK-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; CHECK-NEXT:    pand %xmm3, %xmm0
-; CHECK-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; CHECK-NEXT:    por %xmm0, %xmm2
-; CHECK-NEXT:    pand %xmm2, %xmm1
-; CHECK-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; CHECK-NEXT:    por %xmm1, %xmm2
-; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
+; CHECK-NEXT:    pand %xmm0, %xmm1
+; CHECK-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; CHECK-NEXT:    por %xmm1, %xmm0
+; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; CHECK-NEXT:    retq
 entry:
   %conv = fptoui <2 x double> %x to <2 x i64>
   %spec.store.select = call <2 x i64> @llvm.umin.v2i64(<2 x i64> %conv, <2 x i64> <i64 4294967295, i64 4294967295>)
   %conv6 = trunc <2 x i64> %spec.store.select to <2 x i32>
   ret <2 x i32> %conv6
 }
 

 ; CHECK-NEXT:    sarq $63, %rdx
 ; CHECK-NEXT:    andq %rcx, %rdx
 ; CHECK-NEXT:    orq %rax, %rdx
 ; CHECK-NEXT:    movq %rdx, %xmm0
 ; CHECK-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm3[0]
 ; CHECK-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259456,9223372039002259456]
 ; CHECK-NEXT:    movdqa %xmm0, %xmm3
 ; CHECK-NEXT:    pxor %xmm2, %xmm3
-; CHECK-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
-; CHECK-NEXT:    pcmpeqd %xmm2, %xmm4
-; CHECK-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259455,9223372039002259455]
+; CHECK-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
+; CHECK-NEXT:    movdqa {{.*#+}} xmm5 = [2147483647,2147483647,2147483647,2147483647]
 ; CHECK-NEXT:    movdqa %xmm5, %xmm6
-; CHECK-NEXT:    pcmpgtd %xmm3, %xmm6
-; CHECK-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[0,0,2,2]
-; CHECK-NEXT:    pand %xmm4, %xmm3
-; CHECK-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
-; CHECK-NEXT:    por %xmm3, %xmm4
-; CHECK-NEXT:    movdqa {{.*#+}} xmm3 = [4294967295,4294967295]
-; CHECK-NEXT:    pand %xmm4, %xmm0
-; CHECK-NEXT:    pandn %xmm3, %xmm4
-; CHECK-NEXT:    por %xmm0, %xmm4
+; CHECK-NEXT:    pcmpgtd %xmm4, %xmm6
+; CHECK-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; CHECK-NEXT:    pcmpeqd %xmm2, %xmm3
+; CHECK-NEXT:    pand %xmm6, %xmm3
+; CHECK-NEXT:    movdqa {{.*#+}} xmm4 = [4294967295,4294967295]
+; CHECK-NEXT:    pand %xmm3, %xmm0
+; CHECK-NEXT:    pandn %xmm4, %xmm3
+; CHECK-NEXT:    por %xmm0, %xmm3
 ; CHECK-NEXT:    movdqa %xmm1, %xmm0
 ; CHECK-NEXT:    pxor %xmm2, %xmm0
-; CHECK-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
-; CHECK-NEXT:    pcmpeqd %xmm2, %xmm6
-; CHECK-NEXT:    pcmpgtd %xmm0, %xmm5
-; CHECK-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[0,0,2,2]
-; CHECK-NEXT:    pand %xmm6, %xmm2
-; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3]
-; CHECK-NEXT:    por %xmm2, %xmm0
+; CHECK-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[0,0,2,2]
+; CHECK-NEXT:    pcmpgtd %xmm6, %xmm5
+; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; CHECK-NEXT:    pcmpeqd %xmm2, %xmm0
+; CHECK-NEXT:    pand %xmm5, %xmm0
 ; CHECK-NEXT:    pand %xmm0, %xmm1
-; CHECK-NEXT:    pandn %xmm3, %xmm0
+; CHECK-NEXT:    pandn %xmm4, %xmm0
 ; CHECK-NEXT:    por %xmm1, %xmm0
-; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm4[0,2]
+; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm3[0,2]
 ; CHECK-NEXT:    retq
 entry:
   %conv = fptoui <4 x float> %x to <4 x i64>
   %spec.store.select = call <4 x i64> @llvm.umin.v4i64(<4 x i64> %conv, <4 x i64> <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295>)
   %conv6 = trunc <4 x i64> %spec.store.select to <4 x i32>
   ret <4 x i32> %conv6
 }
 

 ; CHECK-NEXT:    andq %rcx, %rdx
 ; CHECK-NEXT:    orq %rax, %rdx
 ; CHECK-NEXT:    movq %rdx, %xmm0
 ; CHECK-NEXT:    punpcklqdq (%rsp), %xmm0 # 16-byte Folded Reload
 ; CHECK-NEXT:    # xmm0 = xmm0[0],mem[0]
 ; CHECK-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; CHECK-NEXT:    movdqa %xmm0, %xmm2
 ; CHECK-NEXT:    pxor %xmm1, %xmm2
-; CHECK-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[1,1,3,3]
-; CHECK-NEXT:    pcmpeqd %xmm1, %xmm3
-; CHECK-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259455,9223372039002259455]
+; CHECK-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; CHECK-NEXT:    movdqa {{.*#+}} xmm4 = [2147483647,2147483647,2147483647,2147483647]
 ; CHECK-NEXT:    movdqa %xmm4, %xmm5
-; CHECK-NEXT:    pcmpgtd %xmm2, %xmm5
-; CHECK-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[0,0,2,2]
-; CHECK-NEXT:    pand %xmm3, %xmm2
-; CHECK-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]
-; CHECK-NEXT:    por %xmm2, %xmm3
-; CHECK-NEXT:    movdqa {{.*#+}} xmm2 = [4294967295,4294967295]
-; CHECK-NEXT:    pand %xmm3, %xmm0
-; CHECK-NEXT:    pandn %xmm2, %xmm3
-; CHECK-NEXT:    por %xmm0, %xmm3
+; CHECK-NEXT:    pcmpgtd %xmm3, %xmm5
+; CHECK-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; CHECK-NEXT:    pcmpeqd %xmm1, %xmm2
+; CHECK-NEXT:    pand %xmm5, %xmm2
+; CHECK-NEXT:    movdqa {{.*#+}} xmm3 = [4294967295,4294967295]
+; CHECK-NEXT:    pand %xmm2, %xmm0
+; CHECK-NEXT:    pandn %xmm3, %xmm2
+; CHECK-NEXT:    por %xmm0, %xmm2
 ; CHECK-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
 ; CHECK-NEXT:    movdqa %xmm6, %xmm0
 ; CHECK-NEXT:    pxor %xmm1, %xmm0
-; CHECK-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3]
-; CHECK-NEXT:    pcmpeqd %xmm1, %xmm5
-; CHECK-NEXT:    pcmpgtd %xmm0, %xmm4
-; CHECK-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,0,2,2]
-; CHECK-NEXT:    pand %xmm5, %xmm1
-; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
-; CHECK-NEXT:    por %xmm1, %xmm0
-; CHECK-NEXT:    movdqa %xmm6, %xmm1
-; CHECK-NEXT:    pand %xmm0, %xmm1
-; CHECK-NEXT:    pandn %xmm2, %xmm0
-; CHECK-NEXT:    por %xmm1, %xmm0
-; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm3[0,2]
+; CHECK-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,0,2,2]
+; CHECK-NEXT:    pcmpgtd %xmm5, %xmm4
+; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; CHECK-NEXT:    pcmpeqd %xmm1, %xmm0
+; CHECK-NEXT:    pand %xmm4, %xmm0
+; CHECK-NEXT:    pand %xmm0, %xmm6
+; CHECK-NEXT:    pandn %xmm3, %xmm0
+; CHECK-NEXT:    por %xmm6, %xmm0
+; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm2[0,2]
 ; CHECK-NEXT:    addq $72, %rsp
 ; CHECK-NEXT:    .cfi_def_cfa_offset 8
 ; CHECK-NEXT:    retq
 entry:
   %conv = fptoui <4 x half> %x to <4 x i64>
   %spec.store.select = call <4 x i64> @llvm.umin.v4i64(<4 x i64> %conv, <4 x i64> <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295>)
   %conv6 = trunc <4 x i64> %spec.store.select to <4 x i32>
   ret <4 x i32> %conv6

llvm/test/CodeGen/X86/masked_store_trunc_usat.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=sse2    | FileCheck %s --check-prefix=SSE2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=sse4.2  | FileCheck %s --check-prefix=SSE4
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx     | FileCheck %s --check-prefixes=AVX,AVX1
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx2    | FileCheck %s --check-prefixes=AVX,AVX2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512f | FileCheck %s --check-prefix=AVX512F
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512vl | FileCheck %s --check-prefix=AVX512VL
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512bw | FileCheck %s --check-prefixes=AVX512BW
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512vl,avx512bw | FileCheck %s --check-prefixes=AVX512VL,AVX512BWVL
 
 define void @truncstore_v8i64_v8i32(<8 x i64> %x, ptr %p, <8 x i32> %mask) {
 ; SSE2-LABEL: truncstore_v8i64_v8i32:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pxor %xmm7, %xmm7
+; SSE2-NEXT:    pxor %xmm8, %xmm8
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm6 = [4294967295,4294967295]
-; SSE2-NEXT:    movdqa {{.*#+}} xmm8 = [9223372039002259456,9223372039002259456]
-; SSE2-NEXT:    movdqa %xmm1, %xmm10
-; SSE2-NEXT:    pxor %xmm8, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm10[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm8, %xmm11
-; SSE2-NEXT:    movdqa {{.*#+}} xmm9 = [9223372039002259455,9223372039002259455]
-; SSE2-NEXT:    movdqa %xmm9, %xmm12
-; SSE2-NEXT:    pcmpgtd %xmm10, %xmm12
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm12[0,0,2,2]
-; SSE2-NEXT:    pand %xmm11, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm12[1,1,3,3]
-; SSE2-NEXT:    por %xmm10, %xmm11
-; SSE2-NEXT:    pand %xmm11, %xmm1
-; SSE2-NEXT:    pandn %xmm6, %xmm11
-; SSE2-NEXT:    por %xmm1, %xmm11
+; SSE2-NEXT:    movdqa {{.*#+}} xmm10 = [9223372039002259456,9223372039002259456]
+; SSE2-NEXT:    movdqa %xmm1, %xmm9
+; SSE2-NEXT:    pxor %xmm10, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm9[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm7 = [2147483647,2147483647,2147483647,2147483647]
+; SSE2-NEXT:    movdqa %xmm7, %xmm12
+; SSE2-NEXT:    pcmpgtd %xmm11, %xmm12
+; SSE2-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm9
+; SSE2-NEXT:    pand %xmm12, %xmm9
+; SSE2-NEXT:    pand %xmm9, %xmm1
+; SSE2-NEXT:    pandn %xmm6, %xmm9
+; SSE2-NEXT:    por %xmm1, %xmm9
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    pxor %xmm8, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm8, %xmm10
-; SSE2-NEXT:    movdqa %xmm9, %xmm12
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm12
-; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm12[0,0,2,2]
-; SSE2-NEXT:    pand %xmm10, %xmm13
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm12[1,1,3,3]
-; SSE2-NEXT:    por %xmm13, %xmm1
+; SSE2-NEXT:    pxor %xmm10, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm1[0,0,2,2]
+; SSE2-NEXT:    movdqa %xmm7, %xmm12
+; SSE2-NEXT:    pcmpgtd %xmm11, %xmm12
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm1
+; SSE2-NEXT:    pand %xmm12, %xmm1
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    pandn %xmm6, %xmm1
 ; SSE2-NEXT:    por %xmm0, %xmm1
-; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm11[0,2]
+; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm9[0,2]
 ; SSE2-NEXT:    movdqa %xmm3, %xmm0
-; SSE2-NEXT:    pxor %xmm8, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm8, %xmm10
-; SSE2-NEXT:    movdqa %xmm9, %xmm11
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm11
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm11[0,0,2,2]
-; SSE2-NEXT:    pand %xmm10, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm11[1,1,3,3]
-; SSE2-NEXT:    movdqa %xmm2, %xmm11
-; SSE2-NEXT:    pxor %xmm8, %xmm11
-; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm11[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm8, %xmm12
-; SSE2-NEXT:    pcmpgtd %xmm11, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm12, %xmm8
+; SSE2-NEXT:    pxor %xmm10, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm9 = xmm0[0,0,2,2]
+; SSE2-NEXT:    movdqa %xmm7, %xmm11
+; SSE2-NEXT:    pcmpgtd %xmm9, %xmm11
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE2-NEXT:    movdqa %xmm2, %xmm9
+; SSE2-NEXT:    pxor %xmm10, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm9[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm12, %xmm7
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm7, %xmm5
-; SSE2-NEXT:    pcmpeqd %xmm11, %xmm11
-; SSE2-NEXT:    pxor %xmm11, %xmm5
-; SSE2-NEXT:    pcmpeqd %xmm7, %xmm4
-; SSE2-NEXT:    pxor %xmm11, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm9
+; SSE2-NEXT:    pcmpeqd %xmm8, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm10
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm8, %xmm4
+; SSE2-NEXT:    pxor %xmm10, %xmm4
 ; SSE2-NEXT:    packssdw %xmm5, %xmm4
 ; SSE2-NEXT:    packsswb %xmm4, %xmm4
 ; SSE2-NEXT:    pmovmskb %xmm4, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    je .LBB0_2
 ; SSE2-NEXT:  # %bb.1: # %cond.store
 ; SSE2-NEXT:    movss %xmm1, (%rdi)
 ; SSE2-NEXT:  .LBB0_2: # %else
-; SSE2-NEXT:    por %xmm10, %xmm0
-; SSE2-NEXT:    por %xmm9, %xmm8
+; SSE2-NEXT:    pand %xmm11, %xmm0
+; SSE2-NEXT:    pand %xmm7, %xmm9
 ; SSE2-NEXT:    testb $2, %al
 ; SSE2-NEXT:    je .LBB0_4
 ; SSE2-NEXT:  # %bb.3: # %cond.store1
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[1,1,1,1]
 ; SSE2-NEXT:    movd %xmm4, 4(%rdi)
 ; SSE2-NEXT:  .LBB0_4: # %else2
 ; SSE2-NEXT:    pand %xmm0, %xmm3
 ; SSE2-NEXT:    pandn %xmm6, %xmm0
-; SSE2-NEXT:    pand %xmm8, %xmm2
-; SSE2-NEXT:    pandn %xmm6, %xmm8
+; SSE2-NEXT:    pand %xmm9, %xmm2
+; SSE2-NEXT:    pandn %xmm6, %xmm9
 ; SSE2-NEXT:    testb $4, %al
 ; SSE2-NEXT:    je .LBB0_6
 ; SSE2-NEXT:  # %bb.5: # %cond.store3
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[2,3,2,3]
 ; SSE2-NEXT:    movd %xmm4, 8(%rdi)
 ; SSE2-NEXT:  .LBB0_6: # %else4
 ; SSE2-NEXT:    por %xmm0, %xmm3
-; SSE2-NEXT:    por %xmm8, %xmm2
+; SSE2-NEXT:    por %xmm9, %xmm2
 ; SSE2-NEXT:    testb $8, %al
 ; SSE2-NEXT:    je .LBB0_8
 ; SSE2-NEXT:  # %bb.7: # %cond.store5
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3]
 ; SSE2-NEXT:    movd %xmm0, 12(%rdi)
 ; SSE2-NEXT:  .LBB0_8: # %else6
 ; SSE2-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,2],xmm3[0,2]
 ; SSE2-NEXT:    testb $16, %al
; AVX512BW-NEXT:    retq
   call void @llvm.masked.store.v8i32.p0(<8 x i32> %d, ptr %p, i32 1, <8 x i1> %a)
   ret void
 }
 
 define void @truncstore_v8i64_v8i16(<8 x i64> %x, ptr %p, <8 x i32> %mask) {
 ; SSE2-LABEL: truncstore_v8i64_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm6, %xmm6
-; SSE2-NEXT:    movdqa {{.*#+}} xmm7 = [65535,65535]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm8 = [65535,65535]
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm9 = [9223372039002259456,9223372039002259456]
-; SSE2-NEXT:    movdqa %xmm2, %xmm8
-; SSE2-NEXT:    pxor %xmm9, %xmm8
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm8[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm9, %xmm11
-; SSE2-NEXT:    movdqa {{.*#+}} xmm10 = [9223372039002324991,9223372039002324991]
+; SSE2-NEXT:    movdqa %xmm2, %xmm7
+; SSE2-NEXT:    pxor %xmm9, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm7[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm10 = [2147549183,2147549183,2147549183,2147549183]
 ; SSE2-NEXT:    movdqa %xmm10, %xmm12
-; SSE2-NEXT:    pcmpgtd %xmm8, %xmm12
-; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm12[0,0,2,2]
-; SSE2-NEXT:    pand %xmm11, %xmm13
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm12[1,1,3,3]
-; SSE2-NEXT:    por %xmm13, %xmm8
-; SSE2-NEXT:    pand %xmm8, %xmm2
-; SSE2-NEXT:    pandn %xmm7, %xmm8
-; SSE2-NEXT:    por %xmm2, %xmm8
+; SSE2-NEXT:    pcmpgtd %xmm11, %xmm12
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm7
+; SSE2-NEXT:    pand %xmm12, %xmm7
+; SSE2-NEXT:    pand %xmm7, %xmm2
+; SSE2-NEXT:    pandn %xmm8, %xmm7
+; SSE2-NEXT:    por %xmm2, %xmm7
 ; SSE2-NEXT:    movdqa %xmm3, %xmm2
 ; SSE2-NEXT:    pxor %xmm9, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm2[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm9, %xmm11
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm2[0,0,2,2]
 ; SSE2-NEXT:    movdqa %xmm10, %xmm12
-; SSE2-NEXT:    pcmpgtd %xmm2, %xmm12
-; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm12[0,0,2,2]
-; SSE2-NEXT:    pand %xmm11, %xmm13
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm12[1,1,3,3]
-; SSE2-NEXT:    por %xmm13, %xmm2
+; SSE2-NEXT:    pcmpgtd %xmm11, %xmm12
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm2
+; SSE2-NEXT:    pand %xmm12, %xmm2
 ; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pandn %xmm7, %xmm2
+; SSE2-NEXT:    pandn %xmm8, %xmm2
 ; SSE2-NEXT:    por %xmm3, %xmm2
 ; SSE2-NEXT:    movdqa %xmm0, %xmm3
 ; SSE2-NEXT:    pxor %xmm9, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm3[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm9, %xmm11
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm3[0,0,2,2]
 ; SSE2-NEXT:    movdqa %xmm10, %xmm12
-; SSE2-NEXT:    pcmpgtd %xmm3, %xmm12
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm12[0,0,2,2]
-; SSE2-NEXT:    pand %xmm11, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm12[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm11
-; SSE2-NEXT:    pand %xmm11, %xmm0
-; SSE2-NEXT:    pandn %xmm7, %xmm11
-; SSE2-NEXT:    por %xmm0, %xmm11
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    pxor %xmm9, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpgtd %xmm11, %xmm12
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd %xmm9, %xmm3
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm10[0,0,2,2]
+; SSE2-NEXT:    pand %xmm12, %xmm3
 ; SSE2-NEXT:    pand %xmm3, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm10[1,1,3,3]
+; SSE2-NEXT:    pandn %xmm8, %xmm3
 ; SSE2-NEXT:    por %xmm0, %xmm3
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pandn %xmm7, %xmm3
-; SSE2-NEXT:    por %xmm1, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,2,2,3]
+; SSE2-NEXT:    movdqa %xmm1, %xmm0
+; SSE2-NEXT:    pxor %xmm9, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm11, %xmm10
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSE2-NEXT:    pand %xmm10, %xmm0
+; SSE2-NEXT:    pand %xmm0, %xmm1
+; SSE2-NEXT:    pandn %xmm8, %xmm0
+; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm11[0,2,2,3]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
 ; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm2 = xmm0[0,1,0,2,4,5,6,7]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm8[0,2,2,3]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,2,4,5,6,7]
 ; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
 ; SSE2-NEXT:    movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
 ; SSE2-NEXT:    pcmpeqd %xmm6, %xmm5
 ; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE2-NEXT:    pxor %xmm1, %xmm5
 ; SSE2-NEXT:    pcmpeqd %xmm6, %xmm4
 ; SSE2-NEXT:    pxor %xmm1, %xmm4

 define void @truncstore_v8i64_v8i8(<8 x i64> %x, ptr %p, <8 x i32> %mask) {
 ; SSE2-LABEL: truncstore_v8i64_v8i8:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm6, %xmm6
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm7 = [255,255]
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm8 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm1, %xmm10
 ; SSE2-NEXT:    pxor %xmm8, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm10[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm8, %xmm11
-; SSE2-NEXT:    movdqa {{.*#+}} xmm9 = [9223372039002259711,9223372039002259711]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm10[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm9 = [2147483903,2147483903,2147483903,2147483903]
 ; SSE2-NEXT:    movdqa %xmm9, %xmm12
-; SSE2-NEXT:    pcmpgtd %xmm10, %xmm12
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm12[0,0,2,2]
-; SSE2-NEXT:    pand %xmm11, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm12[1,1,3,3]
-; SSE2-NEXT:    por %xmm10, %xmm11
-; SSE2-NEXT:    pand %xmm11, %xmm1
-; SSE2-NEXT:    pandn %xmm7, %xmm11
-; SSE2-NEXT:    por %xmm1, %xmm11
+; SSE2-NEXT:    pcmpgtd %xmm11, %xmm12
+; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm8, %xmm10
+; SSE2-NEXT:    pand %xmm12, %xmm10
+; SSE2-NEXT:    pand %xmm10, %xmm1
+; SSE2-NEXT:    pandn %xmm7, %xmm10
+; SSE2-NEXT:    por %xmm1, %xmm10
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    pxor %xmm8, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm8, %xmm10
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm1[0,0,2,2]
 ; SSE2-NEXT:    movdqa %xmm9, %xmm12
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm12
-; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm12[0,0,2,2]
-; SSE2-NEXT:    pand %xmm10, %xmm13
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm12[1,1,3,3]
-; SSE2-NEXT:    por %xmm13, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm11, %xmm12
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm8, %xmm1
+; SSE2-NEXT:    pand %xmm12, %xmm1
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    pandn %xmm7, %xmm1
 ; SSE2-NEXT:    por %xmm0, %xmm1
-; SSE2-NEXT:    packuswb %xmm11, %xmm1
+; SSE2-NEXT:    packuswb %xmm10, %xmm1
 ; SSE2-NEXT:    movdqa %xmm3, %xmm0
 ; SSE2-NEXT:    pxor %xmm8, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm8, %xmm10
+; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    movdqa %xmm9, %xmm11
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm11
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm11[0,0,2,2]
-; SSE2-NEXT:    pand %xmm10, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm11[1,1,3,3]
-; SSE2-NEXT:    por %xmm0, %xmm10
-; SSE2-NEXT:    pand %xmm10, %xmm3
-; SSE2-NEXT:    pandn %xmm7, %xmm10
-; SSE2-NEXT:    por %xmm3, %xmm10
-; SSE2-NEXT:    movdqa %xmm2, %xmm0
-; SSE2-NEXT:    pxor %xmm8, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpgtd %xmm10, %xmm11
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm8, %xmm0
+; SSE2-NEXT:    pand %xmm11, %xmm0
+; SSE2-NEXT:    pand %xmm0, %xmm3
+; SSE2-NEXT:    pandn %xmm7, %xmm0
+; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    movdqa %xmm2, %xmm3
+; SSE2-NEXT:    pxor %xmm8, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm3[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm10, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd %xmm8, %xmm3
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm3, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm0, %xmm3
+; SSE2-NEXT:    pand %xmm9, %xmm3
 ; SSE2-NEXT:    pand %xmm3, %xmm2
 ; SSE2-NEXT:    pandn %xmm7, %xmm3
 ; SSE2-NEXT:    por %xmm2, %xmm3
-; SSE2-NEXT:    packuswb %xmm10, %xmm3
+; SSE2-NEXT:    packuswb %xmm0, %xmm3
 ; SSE2-NEXT:    packuswb %xmm3, %xmm1
 ; SSE2-NEXT:    packuswb %xmm1, %xmm1
 ; SSE2-NEXT:    pcmpeqd %xmm6, %xmm5
 ; SSE2-NEXT:    pcmpeqd %xmm0, %xmm0
 ; SSE2-NEXT:    pxor %xmm0, %xmm5
 ; SSE2-NEXT:    pcmpeqd %xmm6, %xmm4
 ; SSE2-NEXT:    pxor %xmm0, %xmm4
 ; SSE2-NEXT:    packssdw %xmm5, %xmm4

 define void @truncstore_v4i64_v4i32(<4 x i64> %x, ptr %p, <4 x i32> %mask) {
 ; SSE2-LABEL: truncstore_v4i64_v4i32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm3, %xmm3
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [4294967295,4294967295]
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm1, %xmm6
 ; SSE2-NEXT:    pxor %xmm5, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm5, %xmm7
-; SSE2-NEXT:    movdqa {{.*#+}} xmm8 = [9223372039002259455,9223372039002259455]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm8 = [2147483647,2147483647,2147483647,2147483647]
 ; SSE2-NEXT:    movdqa %xmm8, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm6, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm7, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm6, %xmm7
-; SSE2-NEXT:    pand %xmm7, %xmm1
-; SSE2-NEXT:    pandn %xmm4, %xmm7
-; SSE2-NEXT:    por %xmm1, %xmm7
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm6
+; SSE2-NEXT:    pand %xmm9, %xmm6
+; SSE2-NEXT:    pand %xmm6, %xmm1
+; SSE2-NEXT:    pandn %xmm4, %xmm6
+; SSE2-NEXT:    por %xmm1, %xmm6
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    pxor %xmm5, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm5, %xmm6
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm8
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm8[0,0,2,2]
-; SSE2-NEXT:    pand %xmm6, %xmm5
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm8[1,1,3,3]
-; SSE2-NEXT:    por %xmm5, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm8
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm1
+; SSE2-NEXT:    pand %xmm8, %xmm1
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    pandn %xmm4, %xmm1
 ; SSE2-NEXT:    por %xmm0, %xmm1
-; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm7[0,2]
+; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm6[0,2]
 ; SSE2-NEXT:    pcmpeqd %xmm2, %xmm3
 ; SSE2-NEXT:    movmskps %xmm3, %eax
 ; SSE2-NEXT:    xorl $15, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    jne .LBB3_1
 ; SSE2-NEXT:  # %bb.2: # %else
 ; SSE2-NEXT:    testb $2, %al
 ; SSE2-NEXT:    jne .LBB3_3

 define void @truncstore_v4i64_v4i16(<4 x i64> %x, ptr %p, <4 x i32> %mask) {
 ; SSE2-LABEL: truncstore_v4i64_v4i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm3, %xmm3
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [65535,65535]
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm0, %xmm6
 ; SSE2-NEXT:    pxor %xmm5, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm5, %xmm7
-; SSE2-NEXT:    movdqa {{.*#+}} xmm8 = [9223372039002324991,9223372039002324991]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm8 = [2147549183,2147549183,2147549183,2147549183]
 ; SSE2-NEXT:    movdqa %xmm8, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm6, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm7, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm6, %xmm7
-; SSE2-NEXT:    pand %xmm7, %xmm0
-; SSE2-NEXT:    pandn %xmm4, %xmm7
-; SSE2-NEXT:    por %xmm0, %xmm7
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    pxor %xmm5, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd %xmm5, %xmm6
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm8
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm8[0,0,2,2]
+; SSE2-NEXT:    pand %xmm9, %xmm6
 ; SSE2-NEXT:    pand %xmm6, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm8[1,1,3,3]
-; SSE2-NEXT:    por %xmm0, %xmm5
-; SSE2-NEXT:    pand %xmm5, %xmm1
-; SSE2-NEXT:    pandn %xmm4, %xmm5
-; SSE2-NEXT:    por %xmm1, %xmm5
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,2,2,3]
+; SSE2-NEXT:    pandn %xmm4, %xmm6
+; SSE2-NEXT:    por %xmm0, %xmm6
+; SSE2-NEXT:    movdqa %xmm1, %xmm0
+; SSE2-NEXT:    pxor %xmm5, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm8
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm0
+; SSE2-NEXT:    pand %xmm8, %xmm0
+; SSE2-NEXT:    pand %xmm0, %xmm1
+; SSE2-NEXT:    pandn %xmm4, %xmm0
+; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,2,2,3]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
 ; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; SSE2-NEXT:    pcmpeqd %xmm2, %xmm3
 ; SSE2-NEXT:    movmskps %xmm3, %eax
 ; SSE2-NEXT:    xorl $15, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    jne .LBB4_1
 ; SSE2-NEXT:  # %bb.2: # %else

 define void @truncstore_v4i64_v4i8(<4 x i64> %x, ptr %p, <4 x i32> %mask) {
 ; SSE2-LABEL: truncstore_v4i64_v4i8:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm3, %xmm3
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [255,255]
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm0, %xmm4
 ; SSE2-NEXT:    pxor %xmm6, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm6, %xmm7
-; SSE2-NEXT:    movdqa {{.*#+}} xmm8 = [9223372039002259711,9223372039002259711]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm8 = [2147483903,2147483903,2147483903,2147483903]
 ; SSE2-NEXT:    movdqa %xmm8, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm4, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm7, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm10, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm4
+; SSE2-NEXT:    pand %xmm9, %xmm4
 ; SSE2-NEXT:    pand %xmm4, %xmm0
 ; SSE2-NEXT:    pandn %xmm5, %xmm4
 ; SSE2-NEXT:    por %xmm0, %xmm4
 ; SSE2-NEXT:    movdqa %xmm1, %xmm0
 ; SSE2-NEXT:    pxor %xmm6, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm6, %xmm7
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm8
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm8[0,0,2,2]
-; SSE2-NEXT:    pand %xmm7, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm8[1,1,3,3]
-; SSE2-NEXT:    por %xmm0, %xmm6
-; SSE2-NEXT:    pand %xmm6, %xmm1
-; SSE2-NEXT:    pandn %xmm5, %xmm6
-; SSE2-NEXT:    por %xmm1, %xmm6
-; SSE2-NEXT:    pand %xmm5, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm8
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm0
+; SSE2-NEXT:    pand %xmm8, %xmm0
+; SSE2-NEXT:    pand %xmm0, %xmm1
+; SSE2-NEXT:    pandn %xmm5, %xmm0
+; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    pand %xmm5, %xmm0
 ; SSE2-NEXT:    pand %xmm5, %xmm4
-; SSE2-NEXT:    packuswb %xmm6, %xmm4
+; SSE2-NEXT:    packuswb %xmm0, %xmm4
 ; SSE2-NEXT:    packuswb %xmm4, %xmm4
 ; SSE2-NEXT:    packuswb %xmm4, %xmm4
 ; SSE2-NEXT:    pcmpeqd %xmm2, %xmm3
 ; SSE2-NEXT:    movmskps %xmm3, %ecx
 ; SSE2-NEXT:    xorl $15, %ecx
 ; SSE2-NEXT:    testb $1, %cl
 ; SSE2-NEXT:    movd %xmm4, %eax
 ; SSE2-NEXT:    jne .LBB5_1

 }
 
 define void @truncstore_v2i64_v2i32(<2 x i64> %x, ptr %p, <2 x i64> %mask) {
 ; SSE2-LABEL: truncstore_v2i64_v2i32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm2, %xmm2
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    pxor %xmm0, %xmm3
-; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259455,9223372039002259455]
-; SSE2-NEXT:    pcmpgtd %xmm3, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [2147483647,2147483647,2147483647,2147483647]
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm5
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
 ; SSE2-NEXT:    pand %xmm5, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm4
-; SSE2-NEXT:    pand %xmm4, %xmm0
-; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm4
-; SSE2-NEXT:    por %xmm0, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,2,2,3]
+; SSE2-NEXT:    pand %xmm3, %xmm0
+; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
+; SSE2-NEXT:    por %xmm0, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,2,2,3]
 ; SSE2-NEXT:    pcmpeqd %xmm1, %xmm2
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,0,3,2]
 ; SSE2-NEXT:    pand %xmm2, %xmm1
 ; SSE2-NEXT:    movmskpd %xmm1, %eax
 ; SSE2-NEXT:    xorl $3, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    jne .LBB6_1
 ; SSE2-NEXT:  # %bb.2: # %else

 }
 
 define void @truncstore_v2i64_v2i16(<2 x i64> %x, ptr %p, <2 x i64> %mask) {
 ; SSE2-LABEL: truncstore_v2i64_v2i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm2, %xmm2
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    pxor %xmm0, %xmm3
-; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002324991,9223372039002324991]
-; SSE2-NEXT:    pcmpgtd %xmm3, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [2147549183,2147549183,2147549183,2147549183]
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm5
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
 ; SSE2-NEXT:    pand %xmm5, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm4
-; SSE2-NEXT:    pand %xmm4, %xmm0
-; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm4
-; SSE2-NEXT:    por %xmm0, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,2,2,3]
+; SSE2-NEXT:    pand %xmm3, %xmm0
+; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
+; SSE2-NEXT:    por %xmm0, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
 ; SSE2-NEXT:    pcmpeqd %xmm1, %xmm2
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,0,3,2]
 ; SSE2-NEXT:    pand %xmm2, %xmm1
 ; SSE2-NEXT:    movmskpd %xmm1, %eax
 ; SSE2-NEXT:    xorl $3, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    jne .LBB7_1

 }
 
 define void @truncstore_v2i64_v2i8(<2 x i64> %x, ptr %p, <2 x i64> %mask) {
 ; SSE2-LABEL: truncstore_v2i64_v2i8:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    pxor %xmm2, %xmm2
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    pxor %xmm0, %xmm3
-; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259711,9223372039002259711]
-; SSE2-NEXT:    pcmpgtd %xmm3, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [2147483903,2147483903,2147483903,2147483903]
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm5
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
 ; SSE2-NEXT:    pand %xmm5, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm4
-; SSE2-NEXT:    pand %xmm4, %xmm0
-; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm4
-; SSE2-NEXT:    por %xmm0, %xmm4
-; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm4
-; SSE2-NEXT:    packuswb %xmm4, %xmm4
-; SSE2-NEXT:    packuswb %xmm4, %xmm4
-; SSE2-NEXT:    packuswb %xmm4, %xmm4
+; SSE2-NEXT:    pand %xmm3, %xmm0
+; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
+; SSE2-NEXT:    por %xmm0, %xmm3
+; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
+; SSE2-NEXT:    packuswb %xmm3, %xmm3
+; SSE2-NEXT:    packuswb %xmm3, %xmm3
+; SSE2-NEXT:    packuswb %xmm3, %xmm3
 ; SSE2-NEXT:    pcmpeqd %xmm1, %xmm2
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
 ; SSE2-NEXT:    pand %xmm2, %xmm0
 ; SSE2-NEXT:    movmskpd %xmm0, %eax
 ; SSE2-NEXT:    xorl $3, %eax
 ; SSE2-NEXT:    testb $1, %al
-; SSE2-NEXT:    movd %xmm4, %ecx
+; SSE2-NEXT:    movd %xmm3, %ecx
 ; SSE2-NEXT:    jne .LBB8_1
 ; SSE2-NEXT:  # %bb.2: # %else
 ; SSE2-NEXT:    testb $2, %al
 ; SSE2-NEXT:    jne .LBB8_3
 ; SSE2-NEXT:  .LBB8_4: # %else2
 ; SSE2-NEXT:    retq
 ; SSE2-NEXT:  .LBB8_1: # %cond.store
 ; SSE2-NEXT:    movb %cl, (%rdi)

llvm/test/CodeGen/X86/psubus.ll

 }
 
 define <8 x i16> @psubus_8i64_max(<8 x i16> %x, <8 x i64> %y) nounwind {
 ; SSE2OR3-LABEL: psubus_8i64_max:
 ; SSE2OR3:       # %bb.0: # %vector.ph
 ; SSE2OR3-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259456,9223372039002259456]
 ; SSE2OR3-NEXT:    movdqa %xmm2, %xmm7
 ; SSE2OR3-NEXT:    pxor %xmm5, %xmm7
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
+; SSE2OR3-NEXT:    movdqa {{.*#+}} xmm6 = [2147549183,2147549183,2147549183,2147549183]
+; SSE2OR3-NEXT:    movdqa %xmm6, %xmm9
+; SSE2OR3-NEXT:    pcmpgtd %xmm8, %xmm9
 ; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,1,3,3]
 ; SSE2OR3-NEXT:    pcmpeqd %xmm5, %xmm8
-; SSE2OR3-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002324991,9223372039002324991]
-; SSE2OR3-NEXT:    movdqa %xmm6, %xmm9
-; SSE2OR3-NEXT:    pcmpgtd %xmm7, %xmm9
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm7 = xmm9[0,0,2,2]
-; SSE2OR3-NEXT:    pand %xmm8, %xmm7
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,1,3,3]
-; SSE2OR3-NEXT:    por %xmm7, %xmm8
+; SSE2OR3-NEXT:    pand %xmm9, %xmm8
 ; SSE2OR3-NEXT:    movdqa {{.*#+}} xmm7 = [65535,65535]
 ; SSE2OR3-NEXT:    pand %xmm8, %xmm2
 ; SSE2OR3-NEXT:    pandn %xmm7, %xmm8
 ; SSE2OR3-NEXT:    por %xmm2, %xmm8
 ; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm2 = xmm8[0,2,2,3]
 ; SSE2OR3-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7]
 ; SSE2OR3-NEXT:    movdqa %xmm1, %xmm8
 ; SSE2OR3-NEXT:    pxor %xmm5, %xmm8
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm9 = xmm8[1,1,3,3]
-; SSE2OR3-NEXT:    pcmpeqd %xmm5, %xmm9
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm9 = xmm8[0,0,2,2]
 ; SSE2OR3-NEXT:    movdqa %xmm6, %xmm10
-; SSE2OR3-NEXT:    pcmpgtd %xmm8, %xmm10
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm8 = xmm10[0,0,2,2]
-; SSE2OR3-NEXT:    pand %xmm9, %xmm8
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm9 = xmm10[1,1,3,3]
-; SSE2OR3-NEXT:    por %xmm8, %xmm9
-; SSE2OR3-NEXT:    pand %xmm9, %xmm1
-; SSE2OR3-NEXT:    pandn %xmm7, %xmm9
-; SSE2OR3-NEXT:    por %xmm1, %xmm9
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm1 = xmm9[0,2,2,3]
+; SSE2OR3-NEXT:    pcmpgtd %xmm9, %xmm10
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[1,1,3,3]
+; SSE2OR3-NEXT:    pcmpeqd %xmm5, %xmm8
+; SSE2OR3-NEXT:    pand %xmm10, %xmm8
+; SSE2OR3-NEXT:    pand %xmm8, %xmm1
+; SSE2OR3-NEXT:    pandn %xmm7, %xmm8
+; SSE2OR3-NEXT:    por %xmm1, %xmm8
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm1 = xmm8[0,2,2,3]
 ; SSE2OR3-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
 ; SSE2OR3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
 ; SSE2OR3-NEXT:    movdqa %xmm4, %xmm2
 ; SSE2OR3-NEXT:    pxor %xmm5, %xmm2
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm8 = xmm2[1,1,3,3]
-; SSE2OR3-NEXT:    pcmpeqd %xmm5, %xmm8
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm8 = xmm2[0,0,2,2]
 ; SSE2OR3-NEXT:    movdqa %xmm6, %xmm9
-; SSE2OR3-NEXT:    pcmpgtd %xmm2, %xmm9
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm2 = xmm9[0,0,2,2]
-; SSE2OR3-NEXT:    pand %xmm8, %xmm2
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,1,3,3]
-; SSE2OR3-NEXT:    por %xmm2, %xmm8
-; SSE2OR3-NEXT:    pand %xmm8, %xmm4
-; SSE2OR3-NEXT:    pandn %xmm7, %xmm8
-; SSE2OR3-NEXT:    por %xmm4, %xmm8
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm2 = xmm8[0,2,2,3]
+; SSE2OR3-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2OR3-NEXT:    pcmpeqd %xmm5, %xmm2
+; SSE2OR3-NEXT:    pand %xmm9, %xmm2
+; SSE2OR3-NEXT:    pand %xmm2, %xmm4
+; SSE2OR3-NEXT:    pandn %xmm7, %xmm2
+; SSE2OR3-NEXT:    por %xmm4, %xmm2
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
 ; SSE2OR3-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[0,1,0,2,4,5,6,7]
 ; SSE2OR3-NEXT:    movdqa %xmm3, %xmm4
 ; SSE2OR3-NEXT:    pxor %xmm5, %xmm4
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm8 = xmm4[1,1,3,3]
-; SSE2OR3-NEXT:    pcmpeqd %xmm5, %xmm8
-; SSE2OR3-NEXT:    pcmpgtd %xmm4, %xmm6
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[0,0,2,2]
-; SSE2OR3-NEXT:    pand %xmm8, %xmm4
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3]
-; SSE2OR3-NEXT:    por %xmm4, %xmm5
-; SSE2OR3-NEXT:    pand %xmm5, %xmm3
-; SSE2OR3-NEXT:    pandn %xmm7, %xmm5
-; SSE2OR3-NEXT:    por %xmm3, %xmm5
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[0,2,2,3]
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm8 = xmm4[0,0,2,2]
+; SSE2OR3-NEXT:    pcmpgtd %xmm8, %xmm6
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2OR3-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSE2OR3-NEXT:    pand %xmm6, %xmm4
+; SSE2OR3-NEXT:    pand %xmm4, %xmm3
+; SSE2OR3-NEXT:    pandn %xmm7, %xmm4
+; SSE2OR3-NEXT:    por %xmm3, %xmm4
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[0,2,2,3]
 ; SSE2OR3-NEXT:    pshuflw {{.*#+}} xmm3 = xmm3[0,1,0,2,4,5,6,7]
 ; SSE2OR3-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
 ; SSE2OR3-NEXT:    movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]
 ; SSE2OR3-NEXT:    psubusw %xmm3, %xmm0
 ; SSE2OR3-NEXT:    retq
 ;
 ; SSE41-LABEL: psubus_8i64_max:
 ; SSE41:       # %bb.0: # %vector.ph
-; SSE41-NEXT:    movdqa %xmm0, %xmm5
-; SSE41-NEXT:    movdqa {{.*#+}} xmm8 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movdqa %xmm0, %xmm6
+; SSE41-NEXT:    movdqa {{.*#+}} xmm9 = [9223372039002259456,9223372039002259456]
 ; SSE41-NEXT:    movdqa %xmm4, %xmm0
-; SSE41-NEXT:    pxor %xmm8, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002324991,9223372039002324991]
-; SSE41-NEXT:    movdqa %xmm6, %xmm7
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm7
-; SSE41-NEXT:    movdqa %xmm6, %xmm9
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
-; SSE41-NEXT:    pand %xmm7, %xmm0
-; SSE41-NEXT:    por %xmm9, %xmm0
-; SSE41-NEXT:    movapd {{.*#+}} xmm7 = [65535,65535]
-; SSE41-NEXT:    movapd %xmm7, %xmm9
-; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm9
+; SSE41-NEXT:    pxor %xmm9, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm7 = [9223372039002324991,9223372039002324991]
+; SSE41-NEXT:    movdqa %xmm7, %xmm8
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm8
+; SSE41-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [2147549183,2147549183,2147549183,2147549183]
+; SSE41-NEXT:    movdqa %xmm5, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm10, %xmm0
+; SSE41-NEXT:    pand %xmm8, %xmm0
+; SSE41-NEXT:    movapd {{.*#+}} xmm8 = [65535,65535]
+; SSE41-NEXT:    movapd %xmm8, %xmm10
+; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm10
 ; SSE41-NEXT:    movdqa %xmm3, %xmm0
-; SSE41-NEXT:    pxor %xmm8, %xmm0
-; SSE41-NEXT:    movdqa %xmm6, %xmm4
+; SSE41-NEXT:    pxor %xmm9, %xmm0
+; SSE41-NEXT:    movdqa %xmm7, %xmm4
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
-; SSE41-NEXT:    movdqa %xmm6, %xmm10
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm10
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm10[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm5, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm11, %xmm0
 ; SSE41-NEXT:    pand %xmm4, %xmm0
-; SSE41-NEXT:    por %xmm10, %xmm0
-; SSE41-NEXT:    movapd %xmm7, %xmm4
+; SSE41-NEXT:    movapd %xmm8, %xmm4
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm4
-; SSE41-NEXT:    packusdw %xmm9, %xmm4
+; SSE41-NEXT:    packusdw %xmm10, %xmm4
 ; SSE41-NEXT:    movdqa %xmm2, %xmm0
-; SSE41-NEXT:    pxor %xmm8, %xmm0
-; SSE41-NEXT:    movdqa %xmm6, %xmm3
+; SSE41-NEXT:    pxor %xmm9, %xmm0
+; SSE41-NEXT:    movdqa %xmm7, %xmm3
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
-; SSE41-NEXT:    movdqa %xmm6, %xmm9
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm5, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm10, %xmm0
 ; SSE41-NEXT:    pand %xmm3, %xmm0
-; SSE41-NEXT:    por %xmm9, %xmm0
-; SSE41-NEXT:    movapd %xmm7, %xmm3
+; SSE41-NEXT:    movapd %xmm8, %xmm3
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm3
-; SSE41-NEXT:    pxor %xmm1, %xmm8
-; SSE41-NEXT:    movdqa %xmm6, %xmm2
-; SSE41-NEXT:    pcmpeqd %xmm8, %xmm2
-; SSE41-NEXT:    pcmpgtd %xmm8, %xmm6
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
-; SSE41-NEXT:    pand %xmm2, %xmm0
-; SSE41-NEXT:    por %xmm6, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm7
-; SSE41-NEXT:    packusdw %xmm3, %xmm7
-; SSE41-NEXT:    packusdw %xmm4, %xmm7
-; SSE41-NEXT:    psubusw %xmm7, %xmm5
+; SSE41-NEXT:    pxor %xmm1, %xmm9
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm7
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm5
+; SSE41-NEXT:    pand %xmm7, %xmm5
 ; SSE41-NEXT:    movdqa %xmm5, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm8
+; SSE41-NEXT:    packusdw %xmm3, %xmm8
+; SSE41-NEXT:    packusdw %xmm4, %xmm8
+; SSE41-NEXT:    psubusw %xmm8, %xmm6
+; SSE41-NEXT:    movdqa %xmm6, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: psubus_8i64_max:
 ; AVX1:       # %bb.0: # %vector.ph
 ; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [9223372036854775808,9223372036854775808]
 ; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm5
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm6 = [9223372036854841343,9223372036854841343]

 
 ; v8i32/v8i64 - sub(x,trunc(umin(y,zext(x))))
 define <8 x i32> @test33(<8 x i32> %a0, <8 x i64> %a1) {
 ; SSE2OR3-LABEL: test33:
 ; SSE2OR3:       # %bb.0:
 ; SSE2OR3-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259456,9223372039002259456]
 ; SSE2OR3-NEXT:    movdqa %xmm3, %xmm8
 ; SSE2OR3-NEXT:    pxor %xmm6, %xmm8
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm9 = xmm8[0,0,2,2]
+; SSE2OR3-NEXT:    movdqa {{.*#+}} xmm7 = [2147483647,2147483647,2147483647,2147483647]
+; SSE2OR3-NEXT:    movdqa %xmm7, %xmm10
+; SSE2OR3-NEXT:    pcmpgtd %xmm9, %xmm10
 ; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm9 = xmm8[1,1,3,3]
 ; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm9
-; SSE2OR3-NEXT:    movdqa {{.*#+}} xmm7 = [9223372039002259455,9223372039002259455]
-; SSE2OR3-NEXT:    movdqa %xmm7, %xmm10
-; SSE2OR3-NEXT:    pcmpgtd %xmm8, %xmm10
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm8 = xmm10[0,0,2,2]
-; SSE2OR3-NEXT:    pand %xmm9, %xmm8
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm9 = xmm10[1,1,3,3]
-; SSE2OR3-NEXT:    por %xmm8, %xmm9
+; SSE2OR3-NEXT:    pand %xmm10, %xmm9
 ; SSE2OR3-NEXT:    pcmpeqd %xmm8, %xmm8
 ; SSE2OR3-NEXT:    pand %xmm9, %xmm3
 ; SSE2OR3-NEXT:    pxor %xmm8, %xmm9
 ; SSE2OR3-NEXT:    por %xmm3, %xmm9
 ; SSE2OR3-NEXT:    movdqa %xmm2, %xmm3
 ; SSE2OR3-NEXT:    pxor %xmm6, %xmm3
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm10 = xmm3[1,1,3,3]
-; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm10
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm10 = xmm3[0,0,2,2]
 ; SSE2OR3-NEXT:    movdqa %xmm7, %xmm11
-; SSE2OR3-NEXT:    pcmpgtd %xmm3, %xmm11
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm11[0,0,2,2]
-; SSE2OR3-NEXT:    pand %xmm10, %xmm3
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm10 = xmm11[1,1,3,3]
-; SSE2OR3-NEXT:    por %xmm3, %xmm10
-; SSE2OR3-NEXT:    pand %xmm10, %xmm2
-; SSE2OR3-NEXT:    pxor %xmm8, %xmm10
-; SSE2OR3-NEXT:    por %xmm2, %xmm10
-; SSE2OR3-NEXT:    shufps {{.*#+}} xmm10 = xmm10[0,2],xmm9[0,2]
+; SSE2OR3-NEXT:    pcmpgtd %xmm10, %xmm11
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm3
+; SSE2OR3-NEXT:    pand %xmm11, %xmm3
+; SSE2OR3-NEXT:    pand %xmm3, %xmm2
+; SSE2OR3-NEXT:    pxor %xmm8, %xmm3
+; SSE2OR3-NEXT:    por %xmm2, %xmm3
+; SSE2OR3-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,2],xmm9[0,2]
 ; SSE2OR3-NEXT:    movdqa %xmm0, %xmm2
-; SSE2OR3-NEXT:    psubd %xmm10, %xmm2
-; SSE2OR3-NEXT:    pxor %xmm6, %xmm10
+; SSE2OR3-NEXT:    psubd %xmm3, %xmm2
+; SSE2OR3-NEXT:    pxor %xmm6, %xmm3
 ; SSE2OR3-NEXT:    pxor %xmm6, %xmm0
-; SSE2OR3-NEXT:    pcmpgtd %xmm10, %xmm0
+; SSE2OR3-NEXT:    pcmpgtd %xmm3, %xmm0
 ; SSE2OR3-NEXT:    pand %xmm2, %xmm0
 ; SSE2OR3-NEXT:    movdqa %xmm5, %xmm2
 ; SSE2OR3-NEXT:    pxor %xmm6, %xmm2
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[1,1,3,3]
-; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm3
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
 ; SSE2OR3-NEXT:    movdqa %xmm7, %xmm9
-; SSE2OR3-NEXT:    pcmpgtd %xmm2, %xmm9
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm2 = xmm9[0,0,2,2]
-; SSE2OR3-NEXT:    pand %xmm3, %xmm2
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm9[1,1,3,3]
-; SSE2OR3-NEXT:    por %xmm2, %xmm3
-; SSE2OR3-NEXT:    pand %xmm3, %xmm5
-; SSE2OR3-NEXT:    pxor %xmm8, %xmm3
-; SSE2OR3-NEXT:    por %xmm5, %xmm3
-; SSE2OR3-NEXT:    movdqa %xmm4, %xmm2
-; SSE2OR3-NEXT:    pxor %xmm6, %xmm2
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[1,1,3,3]
-; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm5
-; SSE2OR3-NEXT:    pcmpgtd %xmm2, %xmm7
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[0,0,2,2]
-; SSE2OR3-NEXT:    pand %xmm5, %xmm2
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
-; SSE2OR3-NEXT:    por %xmm2, %xmm5
-; SSE2OR3-NEXT:    pxor %xmm5, %xmm8
-; SSE2OR3-NEXT:    pand %xmm4, %xmm5
-; SSE2OR3-NEXT:    por %xmm8, %xmm5
-; SSE2OR3-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,2],xmm3[0,2]
+; SSE2OR3-NEXT:    pcmpgtd %xmm3, %xmm9
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm2
+; SSE2OR3-NEXT:    pand %xmm9, %xmm2
+; SSE2OR3-NEXT:    pand %xmm2, %xmm5
+; SSE2OR3-NEXT:    pxor %xmm8, %xmm2
+; SSE2OR3-NEXT:    por %xmm5, %xmm2
+; SSE2OR3-NEXT:    movdqa %xmm4, %xmm3
+; SSE2OR3-NEXT:    pxor %xmm6, %xmm3
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
+; SSE2OR3-NEXT:    pcmpgtd %xmm5, %xmm7
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm3
+; SSE2OR3-NEXT:    pand %xmm7, %xmm3
+; SSE2OR3-NEXT:    pxor %xmm3, %xmm8
+; SSE2OR3-NEXT:    pand %xmm4, %xmm3
+; SSE2OR3-NEXT:    por %xmm8, %xmm3
+; SSE2OR3-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,2],xmm2[0,2]
 ; SSE2OR3-NEXT:    movdqa %xmm1, %xmm2
-; SSE2OR3-NEXT:    psubd %xmm5, %xmm2
-; SSE2OR3-NEXT:    pxor %xmm6, %xmm5
+; SSE2OR3-NEXT:    psubd %xmm3, %xmm2
+; SSE2OR3-NEXT:    pxor %xmm6, %xmm3
 ; SSE2OR3-NEXT:    pxor %xmm6, %xmm1
-; SSE2OR3-NEXT:    pcmpgtd %xmm5, %xmm1
+; SSE2OR3-NEXT:    pcmpgtd %xmm3, %xmm1
 ; SSE2OR3-NEXT:    pand %xmm2, %xmm1
 ; SSE2OR3-NEXT:    retq
 ;
 ; SSE41-LABEL: test33:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa %xmm0, %xmm6
-; SSE41-NEXT:    movdqa {{.*#+}} xmm9 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movdqa %xmm0, %xmm7
+; SSE41-NEXT:    movdqa {{.*#+}} xmm10 = [9223372039002259456,9223372039002259456]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm0
-; SSE41-NEXT:    pxor %xmm9, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm7 = [9223372039002259455,9223372039002259455]
-; SSE41-NEXT:    movdqa %xmm7, %xmm8
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm8
-; SSE41-NEXT:    movdqa %xmm7, %xmm10
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm10
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm10[0,0,2,2]
-; SSE41-NEXT:    pand %xmm8, %xmm0
-; SSE41-NEXT:    por %xmm10, %xmm0
-; SSE41-NEXT:    movapd {{.*#+}} xmm8 = [4294967295,4294967295]
-; SSE41-NEXT:    movapd %xmm8, %xmm10
-; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm10
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm8 = [9223372039002259455,9223372039002259455]
+; SSE41-NEXT:    movdqa %xmm8, %xmm9
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm9
+; SSE41-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = [2147483647,2147483647,2147483647,2147483647]
+; SSE41-NEXT:    movdqa %xmm6, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm11, %xmm0
+; SSE41-NEXT:    pand %xmm9, %xmm0
+; SSE41-NEXT:    movapd {{.*#+}} xmm9 = [4294967295,4294967295]
+; SSE41-NEXT:    movapd %xmm9, %xmm11
+; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm11
 ; SSE41-NEXT:    movdqa %xmm2, %xmm0
-; SSE41-NEXT:    pxor %xmm9, %xmm0
-; SSE41-NEXT:    movdqa %xmm7, %xmm3
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm8, %xmm3
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
-; SSE41-NEXT:    movdqa %xmm7, %xmm11
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm11
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm11[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm12 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm6, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm12, %xmm0
 ; SSE41-NEXT:    pand %xmm3, %xmm0
-; SSE41-NEXT:    por %xmm11, %xmm0
-; SSE41-NEXT:    movapd %xmm8, %xmm3
+; SSE41-NEXT:    movapd %xmm9, %xmm3
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm3
-; SSE41-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,2],xmm10[0,2]
-; SSE41-NEXT:    pmaxud %xmm3, %xmm6
-; SSE41-NEXT:    psubd %xmm3, %xmm6
+; SSE41-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,2],xmm11[0,2]
+; SSE41-NEXT:    pmaxud %xmm3, %xmm7
+; SSE41-NEXT:    psubd %xmm3, %xmm7
 ; SSE41-NEXT:    movdqa %xmm5, %xmm0
-; SSE41-NEXT:    pxor %xmm9, %xmm0
-; SSE41-NEXT:    movdqa %xmm7, %xmm2
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm8, %xmm2
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE41-NEXT:    movdqa %xmm7, %xmm3
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm6, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
 ; SSE41-NEXT:    pand %xmm2, %xmm0
-; SSE41-NEXT:    por %xmm3, %xmm0
-; SSE41-NEXT:    movapd %xmm8, %xmm2
+; SSE41-NEXT:    movapd %xmm9, %xmm2
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm5, %xmm2
-; SSE41-NEXT:    pxor %xmm4, %xmm9
-; SSE41-NEXT:    movdqa %xmm7, %xmm3
-; SSE41-NEXT:    pcmpeqd %xmm9, %xmm3
-; SSE41-NEXT:    pcmpgtd %xmm9, %xmm7
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
-; SSE41-NEXT:    pand %xmm3, %xmm0
-; SSE41-NEXT:    por %xmm7, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm8
-; SSE41-NEXT:    shufps {{.*#+}} xmm8 = xmm8[0,2],xmm2[0,2]
-; SSE41-NEXT:    pmaxud %xmm8, %xmm1
-; SSE41-NEXT:    psubd %xmm8, %xmm1
+; SSE41-NEXT:    pxor %xmm4, %xmm10
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm8
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm10[0,0,2,2]
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm6
+; SSE41-NEXT:    pand %xmm8, %xmm6
 ; SSE41-NEXT:    movdqa %xmm6, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm9
+; SSE41-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,2],xmm2[0,2]
+; SSE41-NEXT:    pmaxud %xmm9, %xmm1
+; SSE41-NEXT:    psubd %xmm9, %xmm1
+; SSE41-NEXT:    movdqa %xmm7, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: test33:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [9223372036854775808,9223372036854775808]
 ; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm5
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm6 = [9223372041149743103,9223372041149743103]

 ; SSE2OR3-LABEL: test34:
 ; SSE2OR3:       # %bb.0:
 ; SSE2OR3-NEXT:    movdqa {{.*#+}} xmm6 = [1,1,1,1]
 ; SSE2OR3-NEXT:    pand %xmm6, %xmm1
 ; SSE2OR3-NEXT:    pand %xmm6, %xmm0
 ; SSE2OR3-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259456,9223372039002259456]
 ; SSE2OR3-NEXT:    movdqa %xmm3, %xmm8
 ; SSE2OR3-NEXT:    pxor %xmm6, %xmm8
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm9 = xmm8[0,0,2,2]
+; SSE2OR3-NEXT:    movdqa {{.*#+}} xmm7 = [2147483647,2147483647,2147483647,2147483647]
+; SSE2OR3-NEXT:    movdqa %xmm7, %xmm10
+; SSE2OR3-NEXT:    pcmpgtd %xmm9, %xmm10
 ; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm9 = xmm8[1,1,3,3]
 ; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm9
-; SSE2OR3-NEXT:    movdqa {{.*#+}} xmm7 = [9223372039002259455,9223372039002259455]
-; SSE2OR3-NEXT:    movdqa %xmm7, %xmm10
-; SSE2OR3-NEXT:    pcmpgtd %xmm8, %xmm10
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm8 = xmm10[0,0,2,2]
-; SSE2OR3-NEXT:    pand %xmm9, %xmm8
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm9 = xmm10[1,1,3,3]
-; SSE2OR3-NEXT:    por %xmm8, %xmm9
+; SSE2OR3-NEXT:    pand %xmm10, %xmm9
 ; SSE2OR3-NEXT:    pcmpeqd %xmm8, %xmm8
 ; SSE2OR3-NEXT:    pand %xmm9, %xmm3
 ; SSE2OR3-NEXT:    pxor %xmm8, %xmm9
 ; SSE2OR3-NEXT:    por %xmm3, %xmm9
 ; SSE2OR3-NEXT:    movdqa %xmm2, %xmm3
 ; SSE2OR3-NEXT:    pxor %xmm6, %xmm3
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm10 = xmm3[1,1,3,3]
-; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm10
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm10 = xmm3[0,0,2,2]
 ; SSE2OR3-NEXT:    movdqa %xmm7, %xmm11
-; SSE2OR3-NEXT:    pcmpgtd %xmm3, %xmm11
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm11[0,0,2,2]
-; SSE2OR3-NEXT:    pand %xmm10, %xmm3
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm10 = xmm11[1,1,3,3]
-; SSE2OR3-NEXT:    por %xmm3, %xmm10
-; SSE2OR3-NEXT:    pand %xmm10, %xmm2
-; SSE2OR3-NEXT:    pxor %xmm8, %xmm10
-; SSE2OR3-NEXT:    por %xmm2, %xmm10
-; SSE2OR3-NEXT:    shufps {{.*#+}} xmm10 = xmm10[0,2],xmm9[0,2]
+; SSE2OR3-NEXT:    pcmpgtd %xmm10, %xmm11
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm3
+; SSE2OR3-NEXT:    pand %xmm11, %xmm3
+; SSE2OR3-NEXT:    pand %xmm3, %xmm2
+; SSE2OR3-NEXT:    pxor %xmm8, %xmm3
+; SSE2OR3-NEXT:    por %xmm2, %xmm3
+; SSE2OR3-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,2],xmm9[0,2]
 ; SSE2OR3-NEXT:    movdqa %xmm0, %xmm2
-; SSE2OR3-NEXT:    psubd %xmm10, %xmm2
-; SSE2OR3-NEXT:    pxor %xmm6, %xmm10
+; SSE2OR3-NEXT:    psubd %xmm3, %xmm2
+; SSE2OR3-NEXT:    pxor %xmm6, %xmm3
 ; SSE2OR3-NEXT:    por %xmm6, %xmm0
-; SSE2OR3-NEXT:    pcmpgtd %xmm10, %xmm0
+; SSE2OR3-NEXT:    pcmpgtd %xmm3, %xmm0
 ; SSE2OR3-NEXT:    pand %xmm2, %xmm0
 ; SSE2OR3-NEXT:    movdqa %xmm5, %xmm2
 ; SSE2OR3-NEXT:    pxor %xmm6, %xmm2
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[1,1,3,3]
-; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm3
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
 ; SSE2OR3-NEXT:    movdqa %xmm7, %xmm9
-; SSE2OR3-NEXT:    pcmpgtd %xmm2, %xmm9
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm2 = xmm9[0,0,2,2]
-; SSE2OR3-NEXT:    pand %xmm3, %xmm2
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm9[1,1,3,3]
-; SSE2OR3-NEXT:    por %xmm2, %xmm3
-; SSE2OR3-NEXT:    pand %xmm3, %xmm5
-; SSE2OR3-NEXT:    pxor %xmm8, %xmm3
-; SSE2OR3-NEXT:    por %xmm5, %xmm3
-; SSE2OR3-NEXT:    movdqa %xmm4, %xmm2
-; SSE2OR3-NEXT:    pxor %xmm6, %xmm2
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[1,1,3,3]
-; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm5
-; SSE2OR3-NEXT:    pcmpgtd %xmm2, %xmm7
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[0,0,2,2]
-; SSE2OR3-NEXT:    pand %xmm5, %xmm2
-; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
-; SSE2OR3-NEXT:    por %xmm2, %xmm5
-; SSE2OR3-NEXT:    pxor %xmm5, %xmm8
-; SSE2OR3-NEXT:    pand %xmm4, %xmm5
-; SSE2OR3-NEXT:    por %xmm8, %xmm5
-; SSE2OR3-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,2],xmm3[0,2]
+; SSE2OR3-NEXT:    pcmpgtd %xmm3, %xmm9
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm2
+; SSE2OR3-NEXT:    pand %xmm9, %xmm2
+; SSE2OR3-NEXT:    pand %xmm2, %xmm5
+; SSE2OR3-NEXT:    pxor %xmm8, %xmm2
+; SSE2OR3-NEXT:    por %xmm5, %xmm2
+; SSE2OR3-NEXT:    movdqa %xmm4, %xmm3
+; SSE2OR3-NEXT:    pxor %xmm6, %xmm3
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
+; SSE2OR3-NEXT:    pcmpgtd %xmm5, %xmm7
+; SSE2OR3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2OR3-NEXT:    pcmpeqd %xmm6, %xmm3
+; SSE2OR3-NEXT:    pand %xmm7, %xmm3
+; SSE2OR3-NEXT:    pxor %xmm3, %xmm8
+; SSE2OR3-NEXT:    pand %xmm4, %xmm3
+; SSE2OR3-NEXT:    por %xmm8, %xmm3
+; SSE2OR3-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,2],xmm2[0,2]
 ; SSE2OR3-NEXT:    movdqa %xmm1, %xmm2
-; SSE2OR3-NEXT:    psubd %xmm5, %xmm2
-; SSE2OR3-NEXT:    pxor %xmm6, %xmm5
+; SSE2OR3-NEXT:    psubd %xmm3, %xmm2
+; SSE2OR3-NEXT:    pxor %xmm6, %xmm3
 ; SSE2OR3-NEXT:    por %xmm6, %xmm1
-; SSE2OR3-NEXT:    pcmpgtd %xmm5, %xmm1
+; SSE2OR3-NEXT:    pcmpgtd %xmm3, %xmm1
 ; SSE2OR3-NEXT:    pand %xmm2, %xmm1
 ; SSE2OR3-NEXT:    retq
 ;
 ; SSE41-LABEL: test34:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm6
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [1,1,1,1]
 ; SSE41-NEXT:    pand %xmm0, %xmm1
 ; SSE41-NEXT:    pand %xmm0, %xmm6
-; SSE41-NEXT:    movdqa {{.*#+}} xmm9 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm10 = [9223372039002259456,9223372039002259456]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm0
-; SSE41-NEXT:    pxor %xmm9, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm7 = [9223372039002259455,9223372039002259455]
-; SSE41-NEXT:    movdqa %xmm7, %xmm8
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm8
-; SSE41-NEXT:    movdqa %xmm7, %xmm10
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm10
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm10[0,0,2,2]
-; SSE41-NEXT:    pand %xmm8, %xmm0
-; SSE41-NEXT:    por %xmm10, %xmm0
-; SSE41-NEXT:    movapd {{.*#+}} xmm8 = [4294967295,4294967295]
-; SSE41-NEXT:    movapd %xmm8, %xmm10
-; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm10
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm8 = [9223372039002259455,9223372039002259455]
+; SSE41-NEXT:    movdqa %xmm8, %xmm9
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm9
+; SSE41-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm7 = [2147483647,2147483647,2147483647,2147483647]
+; SSE41-NEXT:    movdqa %xmm7, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm11, %xmm0
+; SSE41-NEXT:    pand %xmm9, %xmm0
+; SSE41-NEXT:    movapd {{.*#+}} xmm9 = [4294967295,4294967295]
+; SSE41-NEXT:    movapd %xmm9, %xmm11
+; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm11
 ; SSE41-NEXT:    movdqa %xmm2, %xmm0
-; SSE41-NEXT:    pxor %xmm9, %xmm0
-; SSE41-NEXT:    movdqa %xmm7, %xmm3
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm8, %xmm3
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
-; SSE41-NEXT:    movdqa %xmm7, %xmm11
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm11
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm11[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm12 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm7, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm12, %xmm0
 ; SSE41-NEXT:    pand %xmm3, %xmm0
-; SSE41-NEXT:    por %xmm11, %xmm0
-; SSE41-NEXT:    movapd %xmm8, %xmm3
+; SSE41-NEXT:    movapd %xmm9, %xmm3
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm3
-; SSE41-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,2],xmm10[0,2]
+; SSE41-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,2],xmm11[0,2]
 ; SSE41-NEXT:    pmaxud %xmm3, %xmm6
 ; SSE41-NEXT:    psubd %xmm3, %xmm6
 ; SSE41-NEXT:    movdqa %xmm5, %xmm0
-; SSE41-NEXT:    pxor %xmm9, %xmm0
-; SSE41-NEXT:    movdqa %xmm7, %xmm2
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm8, %xmm2
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE41-NEXT:    movdqa %xmm7, %xmm3
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm7, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
 ; SSE41-NEXT:    pand %xmm2, %xmm0
-; SSE41-NEXT:    por %xmm3, %xmm0
-; SSE41-NEXT:    movapd %xmm8, %xmm2
+; SSE41-NEXT:    movapd %xmm9, %xmm2
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm5, %xmm2
-; SSE41-NEXT:    pxor %xmm4, %xmm9
-; SSE41-NEXT:    movdqa %xmm7, %xmm3
-; SSE41-NEXT:    pcmpeqd %xmm9, %xmm3
-; SSE41-NEXT:    pcmpgtd %xmm9, %xmm7
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
-; SSE41-NEXT:    pand %xmm3, %xmm0
-; SSE41-NEXT:    por %xmm7, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm8
-; SSE41-NEXT:    shufps {{.*#+}} xmm8 = xmm8[0,2],xmm2[0,2]
-; SSE41-NEXT:    pmaxud %xmm8, %xmm1
-; SSE41-NEXT:    psubd %xmm8, %xmm1
+; SSE41-NEXT:    pxor %xmm4, %xmm10
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm8
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm10[0,0,2,2]
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm7
+; SSE41-NEXT:    pand %xmm8, %xmm7
+; SSE41-NEXT:    movdqa %xmm7, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm9
+; SSE41-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,2],xmm2[0,2]
+; SSE41-NEXT:    pmaxud %xmm9, %xmm1
+; SSE41-NEXT:    psubd %xmm9, %xmm1
 ; SSE41-NEXT:    movdqa %xmm6, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: test34:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
 ; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [9223372036854775808,9223372036854775808]

llvm/test/CodeGen/X86/sat-add.ll

 }
 
 define <2 x i64> @unsigned_sat_constant_v2i64_using_cmp_notval(<2 x i64> %x) {
 ; SSE2-LABEL: unsigned_sat_constant_v2i64_using_cmp_notval:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [42,42]
 ; SSE2-NEXT:    paddq %xmm0, %xmm1
 ; SSE2-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSE2-NEXT:    pand %xmm3, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pand %xmm2, %xmm0
 ; SSE2-NEXT:    por %xmm1, %xmm0
-; SSE2-NEXT:    por %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: unsigned_sat_constant_v2i64_using_cmp_notval:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [42,42]
 ; SSE41-NEXT:    paddq %xmm0, %xmm1
 ; SSE41-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSE41-NEXT:    pand %xmm3, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pand %xmm2, %xmm0
 ; SSE41-NEXT:    por %xmm1, %xmm0
-; SSE41-NEXT:    por %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; SSE42-LABEL: unsigned_sat_constant_v2i64_using_cmp_notval:
 ; SSE42:       # %bb.0:
 ; SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
 ; SSE42-NEXT:    movdqa %xmm0, %xmm2
 ; SSE42-NEXT:    pxor %xmm1, %xmm2
 ; SSE42-NEXT:    paddq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0

llvm/test/CodeGen/X86/vec_setcc-2.ll

; SSE41-NEXT:    retq
   %cmp = icmp uge <4 x i32> %x, <i32 -42, i32 -42, i32 -42, i32 -42>
   ret <4 x i1> %cmp
 }
 
 define <2 x i1> @uge_v2i64_splat(<2 x i64> %x) {
 ; SSE2-LABEL: uge_v2i64_splat:
 ; SSE2:       ## %bb.0:
 ; SSE2-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259898,9223372039002259898]
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT:    pand %xmm2, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm0, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147484090,2147484090,2147484090,2147484090]
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE2-NEXT:    pand %xmm2, %xmm1
 ; SSE2-NEXT:    pcmpeqd %xmm0, %xmm0
 ; SSE2-NEXT:    pxor %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: uge_v2i64_splat:
 ; SSE41:       ## %bb.0:
 ; SSE41-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [9223372036854776250,9223372036854776250]
; SSE41-NEXT:    retq
   %cmp = icmp ult <4 x i32> %x, <i32 -42, i32 -42, i32 -42, i32 -42>
   ret <4 x i1> %cmp
 }
 
 define <2 x i1> @ult_v2i64_splat(<2 x i64> %x) {
 ; SSE2-LABEL: ult_v2i64_splat:
 ; SSE2:       ## %bb.0:
 ; SSE2-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259898,9223372039002259898]
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147484090,2147484090,2147484090,2147484090]
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pand %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_v2i64_splat:
 ; SSE41:       ## %bb.0:
 ; SSE41-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [9223372036854776250,9223372036854776250]
 ; SSE41-NEXT:    pcmpgtq %xmm0, %xmm1
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0

llvm/test/CodeGen/X86/vector-popcnt-128-ult-ugt.ll

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_2_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_2_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_2_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_2_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483651,2147483651]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483651,2147483651,2147483651,2147483651]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_3_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483651,2147483651]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483651,2147483651,2147483651,2147483651]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_3_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483651,2147483651]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483651,2147483651,2147483651,2147483651]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_3_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483651,2147483651]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483651,2147483651,2147483651,2147483651]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_3_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_3_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_3_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_3_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_3_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483652,2147483652]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483652,2147483652,2147483652,2147483652]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_4_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483652,2147483652]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483652,2147483652,2147483652,2147483652]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_4_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483652,2147483652]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483652,2147483652,2147483652,2147483652]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_4_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483652,2147483652]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483652,2147483652,2147483652,2147483652]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_4_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_4_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_4_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_4_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_4_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483653,2147483653]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483653,2147483653,2147483653,2147483653]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_5_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483653,2147483653]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483653,2147483653,2147483653,2147483653]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_5_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483653,2147483653]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483653,2147483653,2147483653,2147483653]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_5_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483653,2147483653]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483653,2147483653,2147483653,2147483653]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_5_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_5_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_5_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_5_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_5_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483654,2147483654]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483654,2147483654,2147483654,2147483654]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_6_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483654,2147483654]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483654,2147483654,2147483654,2147483654]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_6_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483654,2147483654]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483654,2147483654,2147483654,2147483654]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_6_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483654,2147483654]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483654,2147483654,2147483654,2147483654]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_6_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_6_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_6_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_6_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_6_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483655,2147483655]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483655,2147483655,2147483655,2147483655]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_7_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483655,2147483655]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483655,2147483655,2147483655,2147483655]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_7_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483655,2147483655]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483655,2147483655,2147483655,2147483655]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_7_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483655,2147483655]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483655,2147483655,2147483655,2147483655]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_7_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_7_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_7_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_7_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_7_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483656,2147483656]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483656,2147483656,2147483656,2147483656]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_8_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483656,2147483656]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483656,2147483656,2147483656,2147483656]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_8_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483656,2147483656]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483656,2147483656,2147483656,2147483656]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_8_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483656,2147483656]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483656,2147483656,2147483656,2147483656]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_8_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_8_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_8_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_8_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_8_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483657,2147483657]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483657,2147483657,2147483657,2147483657]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_9_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483657,2147483657]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483657,2147483657,2147483657,2147483657]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_9_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483657,2147483657]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483657,2147483657,2147483657,2147483657]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_9_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483657,2147483657]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483657,2147483657,2147483657,2147483657]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_9_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_9_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_9_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_9_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_9_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483658,2147483658]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483658,2147483658,2147483658,2147483658]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_10_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483658,2147483658]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483658,2147483658,2147483658,2147483658]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_10_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483658,2147483658]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483658,2147483658,2147483658,2147483658]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_10_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483658,2147483658]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483658,2147483658,2147483658,2147483658]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_10_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_10_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_10_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_10_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_10_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483659,2147483659]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483659,2147483659,2147483659,2147483659]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_11_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483659,2147483659]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483659,2147483659,2147483659,2147483659]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_11_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483659,2147483659]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483659,2147483659,2147483659,2147483659]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_11_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483659,2147483659]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483659,2147483659,2147483659,2147483659]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_11_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_11_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_11_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_11_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_11_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483660,2147483660]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483660,2147483660,2147483660,2147483660]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_12_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483660,2147483660]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483660,2147483660,2147483660,2147483660]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_12_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483660,2147483660]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483660,2147483660,2147483660,2147483660]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_12_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483660,2147483660]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483660,2147483660,2147483660,2147483660]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_12_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_12_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_12_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_12_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_12_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483661,2147483661]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483661,2147483661,2147483661,2147483661]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_13_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483661,2147483661]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483661,2147483661,2147483661,2147483661]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_13_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483661,2147483661]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483661,2147483661,2147483661,2147483661]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_13_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483661,2147483661]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483661,2147483661,2147483661,2147483661]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_13_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_13_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_13_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_13_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_13_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483662,2147483662]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483662,2147483662,2147483662,2147483662]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_14_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483662,2147483662]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483662,2147483662,2147483662,2147483662]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_14_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483662,2147483662]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483662,2147483662,2147483662,2147483662]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_14_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483662,2147483662]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483662,2147483662,2147483662,2147483662]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_14_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_14_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_14_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_14_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_14_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483663,2147483663]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483663,2147483663,2147483663,2147483663]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_15_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483663,2147483663]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483663,2147483663,2147483663,2147483663]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_15_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483663,2147483663]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483663,2147483663,2147483663,2147483663]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_15_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483663,2147483663]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483663,2147483663,2147483663,2147483663]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_15_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_15_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_15_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_15_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_15_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483664,2147483664]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483664,2147483664,2147483664,2147483664]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_16_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483664,2147483664]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483664,2147483664,2147483664,2147483664]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_16_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483664,2147483664]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483664,2147483664,2147483664,2147483664]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_16_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483664,2147483664]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483664,2147483664,2147483664,2147483664]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_16_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_16_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_16_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_16_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_16_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483665,2147483665]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483665,2147483665,2147483665,2147483665]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_17_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483665,2147483665]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483665,2147483665,2147483665,2147483665]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_17_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483665,2147483665]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483665,2147483665,2147483665,2147483665]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_17_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483665,2147483665]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483665,2147483665,2147483665,2147483665]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_17_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_17_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_17_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_17_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_17_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483666,2147483666]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483666,2147483666,2147483666,2147483666]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_18_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483666,2147483666]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483666,2147483666,2147483666,2147483666]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_18_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483666,2147483666]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483666,2147483666,2147483666,2147483666]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_18_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483666,2147483666]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483666,2147483666,2147483666,2147483666]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_18_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_18_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_18_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_18_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_18_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483667,2147483667]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483667,2147483667,2147483667,2147483667]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_19_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483667,2147483667]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483667,2147483667,2147483667,2147483667]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_19_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483667,2147483667]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483667,2147483667,2147483667,2147483667]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_19_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483667,2147483667]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483667,2147483667,2147483667,2147483667]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_19_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_19_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_19_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_19_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_19_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483668,2147483668]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483668,2147483668,2147483668,2147483668]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_20_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483668,2147483668]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483668,2147483668,2147483668,2147483668]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_20_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483668,2147483668]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483668,2147483668,2147483668,2147483668]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_20_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483668,2147483668]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483668,2147483668,2147483668,2147483668]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_20_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_20_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_20_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_20_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_20_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483669,2147483669]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483669,2147483669,2147483669,2147483669]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_21_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483669,2147483669]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483669,2147483669,2147483669,2147483669]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_21_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483669,2147483669]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483669,2147483669,2147483669,2147483669]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_21_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483669,2147483669]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483669,2147483669,2147483669,2147483669]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_21_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_21_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_21_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_21_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_21_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483670,2147483670]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483670,2147483670,2147483670,2147483670]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_22_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483670,2147483670]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483670,2147483670,2147483670,2147483670]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_22_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483670,2147483670]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483670,2147483670,2147483670,2147483670]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_22_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483670,2147483670]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483670,2147483670,2147483670,2147483670]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_22_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_22_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_22_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_22_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_22_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483671,2147483671]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483671,2147483671,2147483671,2147483671]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_23_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483671,2147483671]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483671,2147483671,2147483671,2147483671]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_23_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483671,2147483671]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483671,2147483671,2147483671,2147483671]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_23_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483671,2147483671]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483671,2147483671,2147483671,2147483671]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_23_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_23_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_23_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_23_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_23_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483672,2147483672]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483672,2147483672,2147483672,2147483672]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_24_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483672,2147483672]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483672,2147483672,2147483672,2147483672]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_24_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483672,2147483672]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483672,2147483672,2147483672,2147483672]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_24_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483672,2147483672]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483672,2147483672,2147483672,2147483672]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_24_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_24_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_24_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_24_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_24_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483673,2147483673]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483673,2147483673,2147483673,2147483673]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_25_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483673,2147483673]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483673,2147483673,2147483673,2147483673]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_25_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483673,2147483673]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483673,2147483673,2147483673,2147483673]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_25_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483673,2147483673]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483673,2147483673,2147483673,2147483673]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_25_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_25_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_25_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_25_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_25_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483674,2147483674]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483674,2147483674,2147483674,2147483674]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_26_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483674,2147483674]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483674,2147483674,2147483674,2147483674]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_26_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483674,2147483674]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483674,2147483674,2147483674,2147483674]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_26_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483674,2147483674]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483674,2147483674,2147483674,2147483674]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_26_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_26_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_26_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_26_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_26_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483675,2147483675]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483675,2147483675,2147483675,2147483675]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_27_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483675,2147483675]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483675,2147483675,2147483675,2147483675]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_27_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483675,2147483675]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483675,2147483675,2147483675,2147483675]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_27_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483675,2147483675]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483675,2147483675,2147483675,2147483675]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_27_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_27_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_27_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_27_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_27_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483676,2147483676]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483676,2147483676,2147483676,2147483676]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_28_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483676,2147483676]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483676,2147483676,2147483676,2147483676]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_28_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483676,2147483676]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483676,2147483676,2147483676,2147483676]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_28_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483676,2147483676]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483676,2147483676,2147483676,2147483676]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_28_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_28_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_28_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_28_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_28_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483677,2147483677]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483677,2147483677,2147483677,2147483677]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_29_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483677,2147483677]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483677,2147483677,2147483677,2147483677]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_29_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483677,2147483677]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483677,2147483677,2147483677,2147483677]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_29_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483677,2147483677]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483677,2147483677,2147483677,2147483677]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_29_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_29_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_29_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_29_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_29_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483678,2147483678]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483678,2147483678,2147483678,2147483678]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_30_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483678,2147483678]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483678,2147483678,2147483678,2147483678]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_30_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483678,2147483678]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483678,2147483678,2147483678,2147483678]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_30_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483678,2147483678]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483678,2147483678,2147483678,2147483678]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_30_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_30_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_30_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_30_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_30_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483679,2147483679]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483679,2147483679,2147483679,2147483679]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_31_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483679,2147483679]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483679,2147483679,2147483679,2147483679]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_31_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483679,2147483679]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483679,2147483679,2147483679,2147483679]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_31_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483679,2147483679]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483679,2147483679,2147483679,2147483679]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_31_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_31_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_31_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_31_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_31_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483680,2147483680]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483680,2147483680,2147483680,2147483680]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_32_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483680,2147483680]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483680,2147483680,2147483680,2147483680]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_32_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483680,2147483680]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483680,2147483680,2147483680,2147483680]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_32_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483680,2147483680]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483680,2147483680,2147483680,2147483680]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_32_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_32_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_32_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_32_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_32_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483681,2147483681]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483681,2147483681,2147483681,2147483681]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_33_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483681,2147483681]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483681,2147483681,2147483681,2147483681]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_33_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483681,2147483681]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483681,2147483681,2147483681,2147483681]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_33_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483681,2147483681]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483681,2147483681,2147483681,2147483681]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_33_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_33_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_33_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_33_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_33_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483682,2147483682]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483682,2147483682,2147483682,2147483682]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_34_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483682,2147483682]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483682,2147483682,2147483682,2147483682]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_34_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483682,2147483682]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483682,2147483682,2147483682,2147483682]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_34_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483682,2147483682]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483682,2147483682,2147483682,2147483682]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_34_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_34_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_34_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_34_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_34_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483683,2147483683]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483683,2147483683,2147483683,2147483683]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_35_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483683,2147483683]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483683,2147483683,2147483683,2147483683]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_35_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483683,2147483683]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483683,2147483683,2147483683,2147483683]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_35_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483683,2147483683]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483683,2147483683,2147483683,2147483683]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_35_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_35_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_35_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_35_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_35_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483684,2147483684]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483684,2147483684,2147483684,2147483684]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_36_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483684,2147483684]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483684,2147483684,2147483684,2147483684]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_36_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483684,2147483684]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483684,2147483684,2147483684,2147483684]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_36_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483684,2147483684]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483684,2147483684,2147483684,2147483684]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_36_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_36_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_36_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_36_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_36_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483685,2147483685]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483685,2147483685,2147483685,2147483685]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_37_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483685,2147483685]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483685,2147483685,2147483685,2147483685]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_37_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483685,2147483685]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483685,2147483685,2147483685,2147483685]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_37_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483685,2147483685]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483685,2147483685,2147483685,2147483685]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_37_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_37_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_37_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_37_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_37_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483686,2147483686]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483686,2147483686,2147483686,2147483686]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_38_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483686,2147483686]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483686,2147483686,2147483686,2147483686]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_38_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483686,2147483686]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483686,2147483686,2147483686,2147483686]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_38_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483686,2147483686]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483686,2147483686,2147483686,2147483686]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_38_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_38_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_38_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_38_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_38_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483687,2147483687]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483687,2147483687,2147483687,2147483687]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_39_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483687,2147483687]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483687,2147483687,2147483687,2147483687]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_39_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483687,2147483687]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483687,2147483687,2147483687,2147483687]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_39_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483687,2147483687]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483687,2147483687,2147483687,2147483687]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_39_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_39_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_39_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_39_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_39_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483688,2147483688]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483688,2147483688,2147483688,2147483688]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_40_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483688,2147483688]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483688,2147483688,2147483688,2147483688]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_40_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483688,2147483688]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483688,2147483688,2147483688,2147483688]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_40_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483688,2147483688]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483688,2147483688,2147483688,2147483688]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_40_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_40_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_40_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_40_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_40_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483689,2147483689]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483689,2147483689,2147483689,2147483689]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_41_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483689,2147483689]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483689,2147483689,2147483689,2147483689]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_41_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483689,2147483689]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483689,2147483689,2147483689,2147483689]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_41_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483689,2147483689]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483689,2147483689,2147483689,2147483689]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_41_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_41_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_41_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_41_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_41_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483690,2147483690]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483690,2147483690,2147483690,2147483690]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_42_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483690,2147483690]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483690,2147483690,2147483690,2147483690]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_42_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483690,2147483690]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483690,2147483690,2147483690,2147483690]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_42_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483690,2147483690]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483690,2147483690,2147483690,2147483690]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_42_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_42_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_42_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_42_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_42_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483691,2147483691]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483691,2147483691,2147483691,2147483691]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_43_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483691,2147483691]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483691,2147483691,2147483691,2147483691]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_43_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483691,2147483691]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483691,2147483691,2147483691,2147483691]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_43_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483691,2147483691]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483691,2147483691,2147483691,2147483691]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_43_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_43_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_43_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_43_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_43_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483692,2147483692]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483692,2147483692,2147483692,2147483692]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_44_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483692,2147483692]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483692,2147483692,2147483692,2147483692]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_44_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483692,2147483692]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483692,2147483692,2147483692,2147483692]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_44_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483692,2147483692]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483692,2147483692,2147483692,2147483692]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_44_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_44_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_44_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_44_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_44_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483693,2147483693]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483693,2147483693,2147483693,2147483693]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_45_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483693,2147483693]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483693,2147483693,2147483693,2147483693]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_45_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483693,2147483693]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483693,2147483693,2147483693,2147483693]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_45_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483693,2147483693]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483693,2147483693,2147483693,2147483693]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_45_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_45_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_45_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_45_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_45_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483694,2147483694]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483694,2147483694,2147483694,2147483694]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_46_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483694,2147483694]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483694,2147483694,2147483694,2147483694]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_46_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483694,2147483694]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483694,2147483694,2147483694,2147483694]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_46_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483694,2147483694]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483694,2147483694,2147483694,2147483694]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_46_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_46_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_46_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_46_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_46_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483695,2147483695]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483695,2147483695,2147483695,2147483695]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_47_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483695,2147483695]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483695,2147483695,2147483695,2147483695]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_47_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483695,2147483695]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483695,2147483695,2147483695,2147483695]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_47_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483695,2147483695]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483695,2147483695,2147483695,2147483695]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_47_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_47_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_47_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_47_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_47_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483696,2147483696]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483696,2147483696,2147483696,2147483696]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_48_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483696,2147483696]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483696,2147483696,2147483696,2147483696]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_48_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483696,2147483696]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483696,2147483696,2147483696,2147483696]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_48_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483696,2147483696]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483696,2147483696,2147483696,2147483696]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_48_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_48_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_48_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_48_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_48_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483697,2147483697]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483697,2147483697,2147483697,2147483697]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_49_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483697,2147483697]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483697,2147483697,2147483697,2147483697]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_49_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483697,2147483697]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483697,2147483697,2147483697,2147483697]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_49_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483697,2147483697]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483697,2147483697,2147483697,2147483697]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_49_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_49_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_49_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_49_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_49_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483698,2147483698]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483698,2147483698,2147483698,2147483698]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_50_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483698,2147483698]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483698,2147483698,2147483698,2147483698]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_50_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483698,2147483698]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483698,2147483698,2147483698,2147483698]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_50_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483698,2147483698]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483698,2147483698,2147483698,2147483698]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_50_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_50_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_50_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_50_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_50_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483699,2147483699]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483699,2147483699,2147483699,2147483699]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_51_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483699,2147483699]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483699,2147483699,2147483699,2147483699]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_51_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483699,2147483699]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483699,2147483699,2147483699,2147483699]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_51_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483699,2147483699]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483699,2147483699,2147483699,2147483699]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_51_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_51_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_51_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_51_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_51_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483700,2147483700]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483700,2147483700,2147483700,2147483700]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_52_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483700,2147483700]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483700,2147483700,2147483700,2147483700]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_52_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483700,2147483700]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483700,2147483700,2147483700,2147483700]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_52_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483700,2147483700]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483700,2147483700,2147483700,2147483700]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_52_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_52_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_52_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_52_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_52_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483701,2147483701]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483701,2147483701,2147483701,2147483701]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_53_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483701,2147483701]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483701,2147483701,2147483701,2147483701]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_53_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483701,2147483701]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483701,2147483701,2147483701,2147483701]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_53_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483701,2147483701]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483701,2147483701,2147483701,2147483701]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_53_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_53_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_53_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_53_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_53_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483702,2147483702]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483702,2147483702,2147483702,2147483702]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_54_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483702,2147483702]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483702,2147483702,2147483702,2147483702]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_54_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483702,2147483702]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483702,2147483702,2147483702,2147483702]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_54_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483702,2147483702]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483702,2147483702,2147483702,2147483702]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_54_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_54_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_54_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_54_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_54_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483703,2147483703]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483703,2147483703,2147483703,2147483703]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_55_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483703,2147483703]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483703,2147483703,2147483703,2147483703]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_55_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483703,2147483703]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483703,2147483703,2147483703,2147483703]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_55_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483703,2147483703]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483703,2147483703,2147483703,2147483703]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_55_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_55_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_55_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_55_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_55_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483704,2147483704]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483704,2147483704,2147483704,2147483704]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_56_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483704,2147483704]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483704,2147483704,2147483704,2147483704]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_56_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483704,2147483704]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483704,2147483704,2147483704,2147483704]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_56_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483704,2147483704]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483704,2147483704,2147483704,2147483704]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_56_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_56_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_56_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_56_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_56_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483705,2147483705]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483705,2147483705,2147483705,2147483705]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_57_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483705,2147483705]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483705,2147483705,2147483705,2147483705]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_57_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483705,2147483705]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483705,2147483705,2147483705,2147483705]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_57_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483705,2147483705]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483705,2147483705,2147483705,2147483705]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_57_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_57_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_57_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_57_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_57_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483706,2147483706]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483706,2147483706,2147483706,2147483706]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_58_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483706,2147483706]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483706,2147483706,2147483706,2147483706]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_58_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483706,2147483706]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483706,2147483706,2147483706,2147483706]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_58_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483706,2147483706]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483706,2147483706,2147483706,2147483706]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_58_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_58_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_58_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_58_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_58_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483707,2147483707]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483707,2147483707,2147483707,2147483707]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_59_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483707,2147483707]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483707,2147483707,2147483707,2147483707]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_59_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483707,2147483707]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483707,2147483707,2147483707,2147483707]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_59_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483707,2147483707]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483707,2147483707,2147483707,2147483707]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_59_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_59_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_59_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_59_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_59_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483708,2147483708]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483708,2147483708,2147483708,2147483708]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_60_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483708,2147483708]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483708,2147483708,2147483708,2147483708]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_60_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483708,2147483708]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483708,2147483708,2147483708,2147483708]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_60_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483708,2147483708]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483708,2147483708,2147483708,2147483708]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_60_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_60_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_60_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_60_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_60_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483709,2147483709]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483709,2147483709,2147483709,2147483709]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_61_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483709,2147483709]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483709,2147483709,2147483709,2147483709]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_61_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483709,2147483709]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483709,2147483709,2147483709,2147483709]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_61_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483709,2147483709]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483709,2147483709,2147483709,2147483709]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_61_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_61_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_61_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_61_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_61_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483710,2147483710]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483710,2147483710,2147483710,2147483710]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_62_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483710,2147483710]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483710,2147483710,2147483710,2147483710]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_62_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483710,2147483710]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483710,2147483710,2147483710,2147483710]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_62_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483710,2147483710]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483710,2147483710,2147483710,2147483710]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_62_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE2-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_62_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm3
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE3-NEXT:    por %xmm3, %xmm0
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_62_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSSE3-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_62_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; SSE41-NEXT:    pcmpgtd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_62_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    psrlw $2, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    paddb %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $4, %xmm1
 ; SSE2-NEXT:    paddb %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    psadbw %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483711,2147483711]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [2147483711,2147483711,2147483711,2147483711]
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_63_v2i64:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    psubb %xmm1, %xmm0
 ; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $2, %xmm0
 ; SSE3-NEXT:    pand %xmm1, %xmm0
 ; SSE3-NEXT:    paddb %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $4, %xmm1
 ; SSE3-NEXT:    paddb %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    psadbw %xmm1, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483711,2147483711]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483711,2147483711,2147483711,2147483711]
 ; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
-; SSE3-NEXT:    pand %xmm2, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_63_v2i64:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb %xmm0, %xmm3
 ; SSSE3-NEXT:    paddb %xmm4, %xmm3
 ; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483711,2147483711]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    psadbw %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2147483711,2147483711,2147483711,2147483711]
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_63_v2i64:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    pand %xmm1, %xmm2
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
 ; SSE41-NEXT:    pshufb %xmm2, %xmm4
 ; SSE41-NEXT:    psrlw $4, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483711,2147483711]
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    psadbw %xmm3, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]
+; SSE41-NEXT:    por {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483711,2147483711,2147483711,2147483711]
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_63_v2i64:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2

llvm/test/CodeGen/X86/vector-trunc-usat.ll

 ; Unsigned saturation truncation to vXi32
 ;
 
 define <2 x i32> @trunc_usat_v2i64_v2i32(<2 x i64> %a0) {
 ; SSE2-LABEL: trunc_usat_v2i64_v2i32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    pxor %xmm0, %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259455,9223372039002259455]
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [2147483647,2147483647,2147483647,2147483647]
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm3
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm2
-; SSE2-NEXT:    pand %xmm2, %xmm0
-; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSE2-NEXT:    por %xmm0, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
+; SSE2-NEXT:    pand %xmm1, %xmm0
+; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE2-NEXT:    por %xmm0, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v2i64_v2i32:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    pxor %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259455,9223372039002259455]
-; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [2147483647,2147483647,2147483647,2147483647]
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm3
 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSSE3-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    por %xmm1, %xmm2
-; SSSE3-NEXT:    pand %xmm2, %xmm0
-; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSSE3-NEXT:    por %xmm0, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
+; SSSE3-NEXT:    pand %xmm1, %xmm0
+; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    por %xmm0, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v2i64_v2i32:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm1
 ; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [4294967295,4294967295]
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [9223372039002259456,9223372039002259456]
 ; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259455,9223372039002259455]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [2147483647,2147483647,2147483647,2147483647]
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE41-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    pand %xmm4, %xmm0
-; SSE41-NEXT:    por %xmm3, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
 ; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: trunc_usat_v2i64_v2i32:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vmovapd {{.*#+}} xmm1 = [4294967295,4294967295]
 ; AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm2
; SKX-NEXT:    retq
   ret <2 x i32> %3
 }
 
 define void @trunc_usat_v2i64_v2i32_store(<2 x i64> %a0, ptr %p1) {
 ; SSE2-LABEL: trunc_usat_v2i64_v2i32_store:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    pxor %xmm0, %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259455,9223372039002259455]
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [2147483647,2147483647,2147483647,2147483647]
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm3
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm2
-; SSE2-NEXT:    pand %xmm2, %xmm0
-; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSE2-NEXT:    por %xmm0, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
+; SSE2-NEXT:    pand %xmm1, %xmm0
+; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE2-NEXT:    por %xmm0, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
 ; SSE2-NEXT:    movq %xmm0, (%rdi)
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v2i64_v2i32_store:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    pxor %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259455,9223372039002259455]
-; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [2147483647,2147483647,2147483647,2147483647]
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm3
 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSSE3-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    por %xmm1, %xmm2
-; SSSE3-NEXT:    pand %xmm2, %xmm0
-; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSSE3-NEXT:    por %xmm0, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
+; SSSE3-NEXT:    pand %xmm1, %xmm0
+; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    por %xmm0, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
 ; SSSE3-NEXT:    movq %xmm0, (%rdi)
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v2i64_v2i32_store:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm1
 ; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [4294967295,4294967295]
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [9223372039002259456,9223372039002259456]
 ; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259455,9223372039002259455]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [2147483647,2147483647,2147483647,2147483647]
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE41-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    pand %xmm4, %xmm0
-; SSE41-NEXT:    por %xmm3, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
 ; SSE41-NEXT:    movq %xmm0, (%rdi)
 ; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: trunc_usat_v2i64_v2i32_store:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vmovapd {{.*#+}} xmm1 = [4294967295,4294967295]

 }
 
 define <4 x i32> @trunc_usat_v4i64_v4i32(<4 x i64> %a0) {
 ; SSE2-LABEL: trunc_usat_v4i64_v4i32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm0, %xmm3
 ; SSE2-NEXT:    pxor %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm2, %xmm4
-; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259455,9223372039002259455]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [2147483647,2147483647,2147483647,2147483647]
 ; SSE2-NEXT:    movdqa %xmm5, %xmm6
-; SSE2-NEXT:    pcmpgtd %xmm3, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[0,0,2,2]
-; SSE2-NEXT:    pand %xmm4, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm4
-; SSE2-NEXT:    movdqa %xmm1, %xmm3
-; SSE2-NEXT:    pxor %xmm2, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm2, %xmm6
-; SSE2-NEXT:    pcmpgtd %xmm3, %xmm5
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[0,0,2,2]
-; SSE2-NEXT:    pand %xmm6, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]
-; SSE2-NEXT:    por %xmm2, %xmm3
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSE2-NEXT:    por %xmm1, %xmm3
-; SSE2-NEXT:    pand %xmm4, %xmm0
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSE2-NEXT:    pand %xmm6, %xmm3
+; SSE2-NEXT:    movdqa %xmm1, %xmm4
+; SSE2-NEXT:    pxor %xmm2, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm6, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm4
+; SSE2-NEXT:    pand %xmm5, %xmm4
+; SSE2-NEXT:    pand %xmm4, %xmm1
 ; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm4
-; SSE2-NEXT:    por %xmm4, %xmm0
-; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm3[0,2]
+; SSE2-NEXT:    por %xmm1, %xmm4
+; SSE2-NEXT:    pand %xmm3, %xmm0
+; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
+; SSE2-NEXT:    por %xmm3, %xmm0
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm4[0,2]
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v4i64_v4i32:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm3
 ; SSSE3-NEXT:    pxor %xmm2, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm4
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259455,9223372039002259455]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm5 = [2147483647,2147483647,2147483647,2147483647]
 ; SSSE3-NEXT:    movdqa %xmm5, %xmm6
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm4, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
-; SSSE3-NEXT:    por %xmm3, %xmm4
-; SSSE3-NEXT:    movdqa %xmm1, %xmm3
-; SSSE3-NEXT:    pxor %xmm2, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm6
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm5
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm6, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm3
-; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
-; SSSE3-NEXT:    por %xmm1, %xmm3
-; SSSE3-NEXT:    pand %xmm4, %xmm0
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSSE3-NEXT:    pand %xmm6, %xmm3
+; SSSE3-NEXT:    movdqa %xmm1, %xmm4
+; SSSE3-NEXT:    pxor %xmm2, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpgtd %xmm6, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm4
+; SSSE3-NEXT:    pand %xmm5, %xmm4
+; SSSE3-NEXT:    pand %xmm4, %xmm1
 ; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm4
-; SSSE3-NEXT:    por %xmm4, %xmm0
-; SSSE3-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm3[0,2]
+; SSSE3-NEXT:    por %xmm1, %xmm4
+; SSSE3-NEXT:    pand %xmm3, %xmm0
+; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
+; SSSE3-NEXT:    por %xmm3, %xmm0
+; SSSE3-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm4[0,2]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v4i64_v4i32:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [9223372039002259456,9223372039002259456]
-; SSE41-NEXT:    movdqa %xmm2, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm3
-; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259455,9223372039002259455]
-; SSE41-NEXT:    movdqa %xmm4, %xmm5
-; SSE41-NEXT:    pcmpeqd %xmm3, %xmm5
-; SSE41-NEXT:    movdqa %xmm4, %xmm6
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm6
-; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[0,0,2,2]
-; SSE41-NEXT:    pand %xmm5, %xmm3
-; SSE41-NEXT:    por %xmm6, %xmm3
-; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    movdqa %xmm4, %xmm5
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm5
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    pxor %xmm4, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259455,9223372039002259455]
+; SSE41-NEXT:    movdqa %xmm5, %xmm6
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483647,2147483647,2147483647,2147483647]
+; SSE41-NEXT:    movdqa %xmm0, %xmm3
+; SSE41-NEXT:    pcmpgtd %xmm7, %xmm3
+; SSE41-NEXT:    pand %xmm6, %xmm3
+; SSE41-NEXT:    pxor %xmm1, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm4, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpgtd %xmm4, %xmm0
 ; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
 ; SSE41-NEXT:    movapd {{.*#+}} xmm4 = [4294967295,4294967295]
 ; SSE41-NEXT:    movapd {{.*#+}} xmm5 = [4294967295,429496729]
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm5
 ; SSE41-NEXT:    movdqa %xmm3, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm4
 ; SSE41-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,2],xmm5[0,2]
 ; SSE41-NEXT:    movaps %xmm4, %xmm0
 ; SSE41-NEXT:    retq
; SKX-NEXT:    retq
   %3 = trunc <4 x i64> %2 to <4 x i32>
   ret <4 x i32> %3
 }
 
 define <8 x i32> @trunc_usat_v8i64_v8i32(ptr %p0) {
 ; SSE2-LABEL: trunc_usat_v8i64_v8i32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa (%rdi), %xmm2
-; SSE2-NEXT:    movdqa 16(%rdi), %xmm5
+; SSE2-NEXT:    movdqa 16(%rdi), %xmm0
 ; SSE2-NEXT:    movdqa 32(%rdi), %xmm6
 ; SSE2-NEXT:    movdqa 48(%rdi), %xmm1
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [4294967295,4294967295]
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [9223372039002259456,9223372039002259456]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm1, %xmm7
-; SSE2-NEXT:    pxor %xmm0, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm8
-; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259455,9223372039002259455]
-; SSE2-NEXT:    movdqa %xmm4, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm7, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm8, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm7, %xmm8
-; SSE2-NEXT:    pand %xmm8, %xmm1
-; SSE2-NEXT:    pandn %xmm3, %xmm8
-; SSE2-NEXT:    por %xmm1, %xmm8
+; SSE2-NEXT:    pxor %xmm4, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [2147483647,2147483647,2147483647,2147483647]
+; SSE2-NEXT:    movdqa %xmm5, %xmm9
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm7
+; SSE2-NEXT:    pand %xmm9, %xmm7
+; SSE2-NEXT:    pand %xmm7, %xmm1
+; SSE2-NEXT:    pandn %xmm3, %xmm7
+; SSE2-NEXT:    por %xmm1, %xmm7
 ; SSE2-NEXT:    movdqa %xmm6, %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm7
-; SSE2-NEXT:    movdqa %xmm4, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm7, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm10, %xmm1
+; SSE2-NEXT:    pxor %xmm4, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[0,0,2,2]
+; SSE2-NEXT:    movdqa %xmm5, %xmm9
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm1
+; SSE2-NEXT:    pand %xmm9, %xmm1
 ; SSE2-NEXT:    pand %xmm1, %xmm6
 ; SSE2-NEXT:    pandn %xmm3, %xmm1
 ; SSE2-NEXT:    por %xmm6, %xmm1
-; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm8[0,2]
-; SSE2-NEXT:    movdqa %xmm5, %xmm6
-; SSE2-NEXT:    pxor %xmm0, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm7
-; SSE2-NEXT:    movdqa %xmm4, %xmm8
-; SSE2-NEXT:    pcmpgtd %xmm6, %xmm8
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm8[0,0,2,2]
-; SSE2-NEXT:    pand %xmm7, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm8[1,1,3,3]
-; SSE2-NEXT:    por %xmm6, %xmm7
-; SSE2-NEXT:    pand %xmm7, %xmm5
-; SSE2-NEXT:    pandn %xmm3, %xmm7
-; SSE2-NEXT:    por %xmm5, %xmm7
-; SSE2-NEXT:    movdqa %xmm2, %xmm5
-; SSE2-NEXT:    pxor %xmm0, %xmm5
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm6
-; SSE2-NEXT:    pcmpgtd %xmm5, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSE2-NEXT:    pand %xmm6, %xmm5
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
-; SSE2-NEXT:    por %xmm5, %xmm0
+; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm7[0,2]
+; SSE2-NEXT:    movdqa %xmm0, %xmm6
+; SSE2-NEXT:    pxor %xmm4, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSE2-NEXT:    movdqa %xmm5, %xmm8
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm8
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm6
+; SSE2-NEXT:    pand %xmm8, %xmm6
+; SSE2-NEXT:    pand %xmm6, %xmm0
+; SSE2-NEXT:    pandn %xmm3, %xmm6
+; SSE2-NEXT:    por %xmm0, %xmm6
+; SSE2-NEXT:    movdqa %xmm2, %xmm0
+; SSE2-NEXT:    pxor %xmm4, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm0
+; SSE2-NEXT:    pand %xmm5, %xmm0
 ; SSE2-NEXT:    pand %xmm0, %xmm2
 ; SSE2-NEXT:    pandn %xmm3, %xmm0
 ; SSE2-NEXT:    por %xmm2, %xmm0
-; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm7[0,2]
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm6[0,2]
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v8i64_v8i32:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa (%rdi), %xmm2
-; SSSE3-NEXT:    movdqa 16(%rdi), %xmm5
+; SSSE3-NEXT:    movdqa 16(%rdi), %xmm0
 ; SSSE3-NEXT:    movdqa 32(%rdi), %xmm6
 ; SSSE3-NEXT:    movdqa 48(%rdi), %xmm1
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [4294967295,4294967295]
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [9223372039002259456,9223372039002259456]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    movdqa %xmm1, %xmm7
-; SSSE3-NEXT:    pxor %xmm0, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm8
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259455,9223372039002259455]
-; SSSE3-NEXT:    movdqa %xmm4, %xmm9
-; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm9
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm9[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm8, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,1,3,3]
-; SSSE3-NEXT:    por %xmm7, %xmm8
-; SSSE3-NEXT:    pand %xmm8, %xmm1
-; SSSE3-NEXT:    pandn %xmm3, %xmm8
-; SSSE3-NEXT:    por %xmm1, %xmm8
+; SSSE3-NEXT:    pxor %xmm4, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm5 = [2147483647,2147483647,2147483647,2147483647]
+; SSSE3-NEXT:    movdqa %xmm5, %xmm9
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm7
+; SSSE3-NEXT:    pand %xmm9, %xmm7
+; SSSE3-NEXT:    pand %xmm7, %xmm1
+; SSSE3-NEXT:    pandn %xmm3, %xmm7
+; SSSE3-NEXT:    por %xmm1, %xmm7
 ; SSSE3-NEXT:    movdqa %xmm6, %xmm1
-; SSSE3-NEXT:    pxor %xmm0, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm7
-; SSSE3-NEXT:    movdqa %xmm4, %xmm9
-; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm9
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm7, %xmm10
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm9[1,1,3,3]
-; SSSE3-NEXT:    por %xmm10, %xmm1
+; SSSE3-NEXT:    pxor %xmm4, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    movdqa %xmm5, %xmm9
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm1
+; SSSE3-NEXT:    pand %xmm9, %xmm1
 ; SSSE3-NEXT:    pand %xmm1, %xmm6
 ; SSSE3-NEXT:    pandn %xmm3, %xmm1
 ; SSSE3-NEXT:    por %xmm6, %xmm1
-; SSSE3-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm8[0,2]
-; SSSE3-NEXT:    movdqa %xmm5, %xmm6
-; SSSE3-NEXT:    pxor %xmm0, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm7
-; SSSE3-NEXT:    movdqa %xmm4, %xmm8
-; SSSE3-NEXT:    pcmpgtd %xmm6, %xmm8
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm8[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm7, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm8[1,1,3,3]
-; SSSE3-NEXT:    por %xmm6, %xmm7
-; SSSE3-NEXT:    pand %xmm7, %xmm5
-; SSSE3-NEXT:    pandn %xmm3, %xmm7
-; SSSE3-NEXT:    por %xmm5, %xmm7
-; SSSE3-NEXT:    movdqa %xmm2, %xmm5
-; SSSE3-NEXT:    pxor %xmm0, %xmm5
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm6
-; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm4
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm6, %xmm5
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
-; SSSE3-NEXT:    por %xmm5, %xmm0
+; SSSE3-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm7[0,2]
+; SSSE3-NEXT:    movdqa %xmm0, %xmm6
+; SSSE3-NEXT:    pxor %xmm4, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSSE3-NEXT:    movdqa %xmm5, %xmm8
+; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm8
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm6
+; SSSE3-NEXT:    pand %xmm8, %xmm6
+; SSSE3-NEXT:    pand %xmm6, %xmm0
+; SSSE3-NEXT:    pandn %xmm3, %xmm6
+; SSSE3-NEXT:    por %xmm0, %xmm6
+; SSSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSSE3-NEXT:    pxor %xmm4, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm0
+; SSSE3-NEXT:    pand %xmm5, %xmm0
 ; SSSE3-NEXT:    pand %xmm0, %xmm2
 ; SSSE3-NEXT:    pandn %xmm3, %xmm0
 ; SSSE3-NEXT:    por %xmm2, %xmm0
-; SSSE3-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm7[0,2]
+; SSSE3-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm6[0,2]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v8i64_v8i32:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa (%rdi), %xmm3
-; SSE41-NEXT:    movdqa 16(%rdi), %xmm6
-; SSE41-NEXT:    movdqa 32(%rdi), %xmm7
+; SSE41-NEXT:    movdqa (%rdi), %xmm4
+; SSE41-NEXT:    movdqa 16(%rdi), %xmm7
+; SSE41-NEXT:    movdqa 32(%rdi), %xmm8
 ; SSE41-NEXT:    movdqa 48(%rdi), %xmm1
-; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [4294967295,4294967295]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movapd {{.*#+}} xmm3 = [4294967295,4294967295]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259456,9223372039002259456]
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
-; SSE41-NEXT:    pxor %xmm5, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259455,9223372039002259455]
-; SSE41-NEXT:    movdqa %xmm4, %xmm8
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm8
-; SSE41-NEXT:    movdqa %xmm4, %xmm9
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
-; SSE41-NEXT:    pand %xmm8, %xmm0
-; SSE41-NEXT:    por %xmm9, %xmm0
-; SSE41-NEXT:    movapd %xmm2, %xmm8
-; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm8
-; SSE41-NEXT:    movdqa %xmm7, %xmm0
-; SSE41-NEXT:    pxor %xmm5, %xmm0
-; SSE41-NEXT:    movdqa %xmm4, %xmm1
+; SSE41-NEXT:    pxor %xmm6, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259455,9223372039002259455]
+; SSE41-NEXT:    movdqa %xmm5, %xmm9
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm9
+; SSE41-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [2147483647,2147483647,2147483647,2147483647]
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm10, %xmm0
+; SSE41-NEXT:    pand %xmm9, %xmm0
+; SSE41-NEXT:    movapd %xmm3, %xmm9
+; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm9
+; SSE41-NEXT:    movdqa %xmm8, %xmm0
+; SSE41-NEXT:    pxor %xmm6, %xmm0
+; SSE41-NEXT:    movdqa %xmm5, %xmm1
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa %xmm4, %xmm9
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm10, %xmm0
 ; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    por %xmm9, %xmm0
-; SSE41-NEXT:    movapd %xmm2, %xmm1
-; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm1
-; SSE41-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm8[0,2]
-; SSE41-NEXT:    movdqa %xmm6, %xmm0
-; SSE41-NEXT:    pxor %xmm5, %xmm0
-; SSE41-NEXT:    movdqa %xmm4, %xmm7
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm7
-; SSE41-NEXT:    movdqa %xmm4, %xmm8
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm8
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm8[0,0,2,2]
-; SSE41-NEXT:    pand %xmm7, %xmm0
-; SSE41-NEXT:    por %xmm8, %xmm0
-; SSE41-NEXT:    movapd %xmm2, %xmm7
-; SSE41-NEXT:    blendvpd %xmm0, %xmm6, %xmm7
-; SSE41-NEXT:    pxor %xmm3, %xmm5
-; SSE41-NEXT:    movdqa %xmm4, %xmm6
-; SSE41-NEXT:    pcmpeqd %xmm5, %xmm6
-; SSE41-NEXT:    pcmpgtd %xmm5, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pand %xmm6, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm2
-; SSE41-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,2],xmm7[0,2]
-; SSE41-NEXT:    movaps %xmm2, %xmm0
+; SSE41-NEXT:    movapd %xmm3, %xmm1
+; SSE41-NEXT:    blendvpd %xmm0, %xmm8, %xmm1
+; SSE41-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm9[0,2]
+; SSE41-NEXT:    movdqa %xmm7, %xmm0
+; SSE41-NEXT:    pxor %xmm6, %xmm0
+; SSE41-NEXT:    movdqa %xmm5, %xmm8
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm8
+; SSE41-NEXT:    pshufd {{.*#+}} xmm9 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm9, %xmm0
+; SSE41-NEXT:    pand %xmm8, %xmm0
+; SSE41-NEXT:    movapd %xmm3, %xmm8
+; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm8
+; SSE41-NEXT:    pxor %xmm4, %xmm6
+; SSE41-NEXT:    pcmpeqd %xmm6, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSE41-NEXT:    pand %xmm5, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm3
+; SSE41-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,2],xmm8[0,2]
+; SSE41-NEXT:    movaps %xmm3, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_usat_v8i64_v8i32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa (%rdi), %xmm0
 ; AVX1-NEXT:    vmovdqa 16(%rdi), %xmm1
 ; AVX1-NEXT:    vmovdqa 32(%rdi), %xmm2
 ; AVX1-NEXT:    vmovdqa 48(%rdi), %xmm3

 ; Unsigned saturation truncation to vXi16
 ;
 
 define <2 x i16> @trunc_usat_v2i64_v2i16(<2 x i64> %a0) {
 ; SSE2-LABEL: trunc_usat_v2i64_v2i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    pxor %xmm0, %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002324991,9223372039002324991]
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [2147549183,2147549183,2147549183,2147549183]
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm3
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm2
-; SSE2-NEXT:    pand %xmm2, %xmm0
-; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSE2-NEXT:    por %xmm0, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
+; SSE2-NEXT:    pand %xmm1, %xmm0
+; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE2-NEXT:    por %xmm0, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v2i64_v2i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    pxor %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002324991,9223372039002324991]
-; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [2147549183,2147549183,2147549183,2147549183]
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm3
 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSSE3-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    por %xmm1, %xmm2
-; SSSE3-NEXT:    pand %xmm2, %xmm0
-; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSSE3-NEXT:    por %xmm0, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
+; SSSE3-NEXT:    pand %xmm1, %xmm0
+; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    por %xmm0, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
 ; SSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v2i64_v2i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm1
 ; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [65535,65535]
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [9223372039002259456,9223372039002259456]
 ; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002324991,9223372039002324991]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [2147549183,2147549183,2147549183,2147549183]
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE41-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    pand %xmm4, %xmm0
-; SSE41-NEXT:    por %xmm3, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
 ; SSE41-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_usat_v2i64_v2i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovapd {{.*#+}} xmm1 = [65535,65535]
; SKX-NEXT:    retq
   ret <2 x i16> %3
 }
 
 define void @trunc_usat_v2i64_v2i16_store(<2 x i64> %a0, ptr %p1) {
 ; SSE2-LABEL: trunc_usat_v2i64_v2i16_store:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    pxor %xmm0, %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002324991,9223372039002324991]
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [2147549183,2147549183,2147549183,2147549183]
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm3
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm2
-; SSE2-NEXT:    pand %xmm2, %xmm0
-; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSE2-NEXT:    por %xmm0, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
+; SSE2-NEXT:    pand %xmm1, %xmm0
+; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE2-NEXT:    por %xmm0, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
 ; SSE2-NEXT:    movd %xmm0, (%rdi)
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v2i64_v2i16_store:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    pxor %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002324991,9223372039002324991]
-; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [2147549183,2147549183,2147549183,2147549183]
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm3
 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSSE3-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    por %xmm1, %xmm2
-; SSSE3-NEXT:    pand %xmm2, %xmm0
-; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSSE3-NEXT:    por %xmm0, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
+; SSSE3-NEXT:    pand %xmm1, %xmm0
+; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    por %xmm0, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
 ; SSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
 ; SSSE3-NEXT:    movd %xmm0, (%rdi)
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v2i64_v2i16_store:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm1
 ; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [65535,65535]
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [9223372039002259456,9223372039002259456]
 ; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002324991,9223372039002324991]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [2147549183,2147549183,2147549183,2147549183]
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE41-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    pand %xmm4, %xmm0
-; SSE41-NEXT:    por %xmm3, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
 ; SSE41-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
 ; SSE41-NEXT:    movd %xmm0, (%rdi)
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_usat_v2i64_v2i16_store:
 ; AVX1:       # %bb.0:

 
 define <4 x i16> @trunc_usat_v4i64_v4i16(<4 x i64> %a0) {
 ; SSE2-LABEL: trunc_usat_v4i64_v4i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [65535,65535]
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm0, %xmm4
 ; SSE2-NEXT:    pxor %xmm3, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm5
-; SSE2-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002324991,9223372039002324991]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm6 = [2147549183,2147549183,2147549183,2147549183]
 ; SSE2-NEXT:    movdqa %xmm6, %xmm7
-; SSE2-NEXT:    pcmpgtd %xmm4, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[0,0,2,2]
-; SSE2-NEXT:    pand %xmm5, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
-; SSE2-NEXT:    por %xmm4, %xmm5
-; SSE2-NEXT:    pand %xmm5, %xmm0
-; SSE2-NEXT:    pandn %xmm2, %xmm5
-; SSE2-NEXT:    por %xmm0, %xmm5
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    pxor %xmm3, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd %xmm3, %xmm4
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
+; SSE2-NEXT:    pand %xmm7, %xmm4
 ; SSE2-NEXT:    pand %xmm4, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
-; SSE2-NEXT:    por %xmm0, %xmm3
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pandn %xmm2, %xmm3
-; SSE2-NEXT:    por %xmm1, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,2,2,3]
+; SSE2-NEXT:    pandn %xmm2, %xmm4
+; SSE2-NEXT:    por %xmm0, %xmm4
+; SSE2-NEXT:    movdqa %xmm1, %xmm0
+; SSE2-NEXT:    pxor %xmm3, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm0
+; SSE2-NEXT:    pand %xmm6, %xmm0
+; SSE2-NEXT:    pand %xmm0, %xmm1
+; SSE2-NEXT:    pandn %xmm2, %xmm0
+; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,2,2,3]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
 ; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v4i64_v4i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [65535,65535]
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm4
 ; SSSE3-NEXT:    pxor %xmm3, %xmm4
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm5
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002324991,9223372039002324991]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm6 = [2147549183,2147549183,2147549183,2147549183]
 ; SSSE3-NEXT:    movdqa %xmm6, %xmm7
-; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm5, %xmm4
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
-; SSSE3-NEXT:    por %xmm4, %xmm5
-; SSSE3-NEXT:    pand %xmm5, %xmm0
-; SSSE3-NEXT:    pandn %xmm2, %xmm5
-; SSSE3-NEXT:    por %xmm0, %xmm5
-; SSSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSSE3-NEXT:    pxor %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
 ; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm4
-; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
+; SSSE3-NEXT:    pand %xmm7, %xmm4
 ; SSSE3-NEXT:    pand %xmm4, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
-; SSSE3-NEXT:    por %xmm0, %xmm3
-; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pandn %xmm2, %xmm3
-; SSSE3-NEXT:    por %xmm1, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,2,2,3]
+; SSSE3-NEXT:    pandn %xmm2, %xmm4
+; SSSE3-NEXT:    por %xmm0, %xmm4
+; SSSE3-NEXT:    movdqa %xmm1, %xmm0
+; SSSE3-NEXT:    pxor %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm0
+; SSSE3-NEXT:    pand %xmm6, %xmm0
+; SSSE3-NEXT:    pand %xmm0, %xmm1
+; SSSE3-NEXT:    pandn %xmm2, %xmm0
+; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; SSSE3-NEXT:    pshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7]
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,2,2,3]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,2,2,3]
 ; SSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
 ; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v4i64_v4i16:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa %xmm0, %xmm3
-; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [65535,65535]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movdqa %xmm0, %xmm2
+; SSE41-NEXT:    movapd {{.*#+}} xmm3 = [65535,65535]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259456,9223372039002259456]
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
-; SSE41-NEXT:    pxor %xmm4, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002324991,9223372039002324991]
-; SSE41-NEXT:    movdqa %xmm5, %xmm6
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm6
-; SSE41-NEXT:    movdqa %xmm5, %xmm7
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm7
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
-; SSE41-NEXT:    pand %xmm6, %xmm0
-; SSE41-NEXT:    por %xmm7, %xmm0
-; SSE41-NEXT:    movapd %xmm2, %xmm6
-; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm6
-; SSE41-NEXT:    pxor %xmm3, %xmm4
-; SSE41-NEXT:    movdqa %xmm5, %xmm1
-; SSE41-NEXT:    pcmpeqd %xmm4, %xmm1
-; SSE41-NEXT:    pcmpgtd %xmm4, %xmm5
+; SSE41-NEXT:    pxor %xmm5, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002324991,9223372039002324991]
+; SSE41-NEXT:    movdqa %xmm6, %xmm7
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm7
+; SSE41-NEXT:    pshufd {{.*#+}} xmm8 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [2147549183,2147549183,2147549183,2147549183]
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm8, %xmm0
+; SSE41-NEXT:    pand %xmm7, %xmm0
+; SSE41-NEXT:    movapd %xmm3, %xmm7
+; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm7
+; SSE41-NEXT:    pxor %xmm2, %xmm5
+; SSE41-NEXT:    pcmpeqd %xmm5, %xmm6
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    por %xmm5, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm2
-; SSE41-NEXT:    packusdw %xmm6, %xmm2
-; SSE41-NEXT:    packusdw %xmm2, %xmm2
-; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSE41-NEXT:    pand %xmm6, %xmm4
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm3
+; SSE41-NEXT:    packusdw %xmm7, %xmm3
+; SSE41-NEXT:    packusdw %xmm3, %xmm3
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_usat_v4i64_v4i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
 ; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm3
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [9223372036854841343,9223372036854841343]

 
 define void @trunc_usat_v4i64_v4i16_store(<4 x i64> %a0, ptr%p1) {
 ; SSE2-LABEL: trunc_usat_v4i64_v4i16_store:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [65535,65535]
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm0, %xmm4
 ; SSE2-NEXT:    pxor %xmm3, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm5
-; SSE2-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002324991,9223372039002324991]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm6 = [2147549183,2147549183,2147549183,2147549183]
 ; SSE2-NEXT:    movdqa %xmm6, %xmm7
-; SSE2-NEXT:    pcmpgtd %xmm4, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[0,0,2,2]
-; SSE2-NEXT:    pand %xmm5, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
-; SSE2-NEXT:    por %xmm4, %xmm5
-; SSE2-NEXT:    pand %xmm5, %xmm0
-; SSE2-NEXT:    pandn %xmm2, %xmm5
-; SSE2-NEXT:    por %xmm0, %xmm5
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    pxor %xmm3, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd %xmm3, %xmm4
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
+; SSE2-NEXT:    pand %xmm7, %xmm4
 ; SSE2-NEXT:    pand %xmm4, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
-; SSE2-NEXT:    por %xmm0, %xmm3
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pandn %xmm2, %xmm3
-; SSE2-NEXT:    por %xmm1, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,2,2,3]
+; SSE2-NEXT:    pandn %xmm2, %xmm4
+; SSE2-NEXT:    por %xmm0, %xmm4
+; SSE2-NEXT:    movdqa %xmm1, %xmm0
+; SSE2-NEXT:    pxor %xmm3, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm0
+; SSE2-NEXT:    pand %xmm6, %xmm0
+; SSE2-NEXT:    pand %xmm0, %xmm1
+; SSE2-NEXT:    pandn %xmm2, %xmm0
+; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[0,2,2,3]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
 ; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 ; SSE2-NEXT:    movq %xmm1, (%rdi)
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v4i64_v4i16_store:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [65535,65535]
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm4
 ; SSSE3-NEXT:    pxor %xmm3, %xmm4
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm5
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002324991,9223372039002324991]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm6 = [2147549183,2147549183,2147549183,2147549183]
 ; SSSE3-NEXT:    movdqa %xmm6, %xmm7
-; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm5, %xmm4
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
-; SSSE3-NEXT:    por %xmm4, %xmm5
-; SSSE3-NEXT:    pand %xmm5, %xmm0
-; SSSE3-NEXT:    pandn %xmm2, %xmm5
-; SSSE3-NEXT:    por %xmm0, %xmm5
-; SSSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSSE3-NEXT:    pxor %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
 ; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm4
-; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
+; SSSE3-NEXT:    pand %xmm7, %xmm4
 ; SSSE3-NEXT:    pand %xmm4, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
-; SSSE3-NEXT:    por %xmm0, %xmm3
-; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pandn %xmm2, %xmm3
-; SSSE3-NEXT:    por %xmm1, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,2,2,3]
+; SSSE3-NEXT:    pandn %xmm2, %xmm4
+; SSSE3-NEXT:    por %xmm0, %xmm4
+; SSSE3-NEXT:    movdqa %xmm1, %xmm0
+; SSSE3-NEXT:    pxor %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm0
+; SSSE3-NEXT:    pand %xmm6, %xmm0
+; SSSE3-NEXT:    pand %xmm0, %xmm1
+; SSSE3-NEXT:    pandn %xmm2, %xmm0
+; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; SSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[0,2,2,3]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,2,2,3]
 ; SSSE3-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
 ; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 ; SSSE3-NEXT:    movq %xmm1, (%rdi)
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v4i64_v4i16_store:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    movapd {{.*#+}} xmm3 = [65535,65535]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movapd {{.*#+}} xmm4 = [65535,65535]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259456,9223372039002259456]
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
-; SSE41-NEXT:    pxor %xmm4, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002324991,9223372039002324991]
-; SSE41-NEXT:    movdqa %xmm5, %xmm6
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm6
-; SSE41-NEXT:    movdqa %xmm5, %xmm7
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm7
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
-; SSE41-NEXT:    pand %xmm6, %xmm0
-; SSE41-NEXT:    por %xmm7, %xmm0
-; SSE41-NEXT:    movapd %xmm3, %xmm6
-; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm6
-; SSE41-NEXT:    pxor %xmm2, %xmm4
-; SSE41-NEXT:    movdqa %xmm5, %xmm1
-; SSE41-NEXT:    pcmpeqd %xmm4, %xmm1
-; SSE41-NEXT:    pcmpgtd %xmm4, %xmm5
+; SSE41-NEXT:    pxor %xmm5, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002324991,9223372039002324991]
+; SSE41-NEXT:    movdqa %xmm6, %xmm7
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm7
+; SSE41-NEXT:    pshufd {{.*#+}} xmm8 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [2147549183,2147549183,2147549183,2147549183]
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm8, %xmm0
+; SSE41-NEXT:    pand %xmm7, %xmm0
+; SSE41-NEXT:    movapd %xmm4, %xmm7
+; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm7
+; SSE41-NEXT:    pxor %xmm2, %xmm5
+; SSE41-NEXT:    pcmpeqd %xmm5, %xmm6
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    por %xmm5, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm3
-; SSE41-NEXT:    packusdw %xmm6, %xmm3
-; SSE41-NEXT:    packusdw %xmm3, %xmm3
-; SSE41-NEXT:    movq %xmm3, (%rdi)
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE41-NEXT:    pand %xmm6, %xmm3
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm4
+; SSE41-NEXT:    packusdw %xmm7, %xmm4
+; SSE41-NEXT:    packusdw %xmm4, %xmm4
+; SSE41-NEXT:    movq %xmm4, (%rdi)
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_usat_v4i64_v4i16_store:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
 ; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm3
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [9223372036854841343,9223372036854841343]
; SKX-NEXT:    retq
   %3 = trunc <4 x i64> %2 to <4 x i16>
   store <4 x i16> %3, ptr%p1
   ret void
 }
 
 define <8 x i16> @trunc_usat_v8i64_v8i16(ptr %p0) {
 ; SSE2-LABEL: trunc_usat_v8i64_v8i16:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa (%rdi), %xmm4
+; SSE2-NEXT:    movdqa (%rdi), %xmm5
 ; SSE2-NEXT:    movdqa 16(%rdi), %xmm0
 ; SSE2-NEXT:    movdqa 32(%rdi), %xmm6
 ; SSE2-NEXT:    movdqa 48(%rdi), %xmm7
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [65535,65535]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [65535,65535]
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
-; SSE2-NEXT:    movdqa %xmm6, %xmm2
-; SSE2-NEXT:    pxor %xmm3, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm2[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm8
-; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002324991,9223372039002324991]
-; SSE2-NEXT:    movdqa %xmm5, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm2, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm8, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm10, %xmm2
-; SSE2-NEXT:    pand %xmm2, %xmm6
-; SSE2-NEXT:    pandn %xmm1, %xmm2
-; SSE2-NEXT:    por %xmm6, %xmm2
+; SSE2-NEXT:    movdqa %xmm6, %xmm1
+; SSE2-NEXT:    pxor %xmm3, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147549183,2147549183,2147549183,2147549183]
+; SSE2-NEXT:    movdqa %xmm4, %xmm9
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm1
+; SSE2-NEXT:    pand %xmm9, %xmm1
+; SSE2-NEXT:    pand %xmm1, %xmm6
+; SSE2-NEXT:    pandn %xmm2, %xmm1
+; SSE2-NEXT:    por %xmm6, %xmm1
 ; SSE2-NEXT:    movdqa %xmm7, %xmm6
 ; SSE2-NEXT:    pxor %xmm3, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm8
-; SSE2-NEXT:    movdqa %xmm5, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm6, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm8, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm10, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2]
+; SSE2-NEXT:    movdqa %xmm4, %xmm9
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm6
+; SSE2-NEXT:    pand %xmm9, %xmm6
 ; SSE2-NEXT:    pand %xmm6, %xmm7
-; SSE2-NEXT:    pandn %xmm1, %xmm6
+; SSE2-NEXT:    pandn %xmm2, %xmm6
 ; SSE2-NEXT:    por %xmm7, %xmm6
-; SSE2-NEXT:    movdqa %xmm4, %xmm7
+; SSE2-NEXT:    movdqa %xmm5, %xmm7
 ; SSE2-NEXT:    pxor %xmm3, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm8
-; SSE2-NEXT:    movdqa %xmm5, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm7, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm8, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm7, %xmm8
-; SSE2-NEXT:    pand %xmm8, %xmm4
-; SSE2-NEXT:    pandn %xmm1, %xmm8
-; SSE2-NEXT:    por %xmm4, %xmm8
-; SSE2-NEXT:    movdqa %xmm0, %xmm4
-; SSE2-NEXT:    pxor %xmm3, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
+; SSE2-NEXT:    movdqa %xmm4, %xmm9
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd %xmm3, %xmm7
-; SSE2-NEXT:    pcmpgtd %xmm4, %xmm5
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[0,0,2,2]
-; SSE2-NEXT:    pand %xmm7, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm4
-; SSE2-NEXT:    pand %xmm4, %xmm0
-; SSE2-NEXT:    pandn %xmm1, %xmm4
-; SSE2-NEXT:    por %xmm0, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,2,2,3]
+; SSE2-NEXT:    pand %xmm9, %xmm7
+; SSE2-NEXT:    pand %xmm7, %xmm5
+; SSE2-NEXT:    pandn %xmm2, %xmm7
+; SSE2-NEXT:    por %xmm5, %xmm7
+; SSE2-NEXT:    movdqa %xmm0, %xmm5
+; SSE2-NEXT:    pxor %xmm3, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm5
+; SSE2-NEXT:    pand %xmm4, %xmm5
+; SSE2-NEXT:    pand %xmm5, %xmm0
+; SSE2-NEXT:    pandn %xmm2, %xmm5
+; SSE2-NEXT:    por %xmm0, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm8[0,2,2,3]
-; SSE2-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3]
+; SSE2-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7]
+; SSE2-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm3 = xmm0[0,1,0,2,4,5,6,7]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
 ; SSE2-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,2,4,5,6,7]
 ; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]
-; SSE2-NEXT:    movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
+; SSE2-NEXT:    movsd {{.*#+}} xmm0 = xmm2[0],xmm0[1]
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v8i64_v8i16:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa (%rdi), %xmm4
+; SSSE3-NEXT:    movdqa (%rdi), %xmm5
 ; SSSE3-NEXT:    movdqa 16(%rdi), %xmm0
 ; SSSE3-NEXT:    movdqa 32(%rdi), %xmm6
 ; SSSE3-NEXT:    movdqa 48(%rdi), %xmm7
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [65535,65535]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [65535,65535]
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
-; SSSE3-NEXT:    movdqa %xmm6, %xmm2
-; SSSE3-NEXT:    pxor %xmm3, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm8
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002324991,9223372039002324991]
-; SSSE3-NEXT:    movdqa %xmm5, %xmm9
-; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm9
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm8, %xmm10
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm9[1,1,3,3]
-; SSSE3-NEXT:    por %xmm10, %xmm2
-; SSSE3-NEXT:    pand %xmm2, %xmm6
-; SSSE3-NEXT:    pandn %xmm1, %xmm2
-; SSSE3-NEXT:    por %xmm6, %xmm2
+; SSSE3-NEXT:    movdqa %xmm6, %xmm1
+; SSSE3-NEXT:    pxor %xmm3, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm4 = [2147549183,2147549183,2147549183,2147549183]
+; SSSE3-NEXT:    movdqa %xmm4, %xmm9
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm1
+; SSSE3-NEXT:    pand %xmm9, %xmm1
+; SSSE3-NEXT:    pand %xmm1, %xmm6
+; SSSE3-NEXT:    pandn %xmm2, %xmm1
+; SSSE3-NEXT:    por %xmm6, %xmm1
 ; SSSE3-NEXT:    movdqa %xmm7, %xmm6
 ; SSSE3-NEXT:    pxor %xmm3, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm8
-; SSSE3-NEXT:    movdqa %xmm5, %xmm9
-; SSSE3-NEXT:    pcmpgtd %xmm6, %xmm9
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm8, %xmm10
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm9[1,1,3,3]
-; SSSE3-NEXT:    por %xmm10, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2]
+; SSSE3-NEXT:    movdqa %xmm4, %xmm9
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm6
+; SSSE3-NEXT:    pand %xmm9, %xmm6
 ; SSSE3-NEXT:    pand %xmm6, %xmm7
-; SSSE3-NEXT:    pandn %xmm1, %xmm6
+; SSSE3-NEXT:    pandn %xmm2, %xmm6
 ; SSSE3-NEXT:    por %xmm7, %xmm6
-; SSSE3-NEXT:    movdqa %xmm4, %xmm7
+; SSSE3-NEXT:    movdqa %xmm5, %xmm7
 ; SSSE3-NEXT:    pxor %xmm3, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm8
-; SSSE3-NEXT:    movdqa %xmm5, %xmm9
-; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm9
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm9[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm8, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,1,3,3]
-; SSSE3-NEXT:    por %xmm7, %xmm8
-; SSSE3-NEXT:    pand %xmm8, %xmm4
-; SSSE3-NEXT:    pandn %xmm1, %xmm8
-; SSSE3-NEXT:    por %xmm4, %xmm8
-; SSSE3-NEXT:    movdqa %xmm0, %xmm4
-; SSSE3-NEXT:    pxor %xmm3, %xmm4
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
+; SSSE3-NEXT:    movdqa %xmm4, %xmm9
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
 ; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm7
-; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm5
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm7, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
-; SSSE3-NEXT:    por %xmm3, %xmm4
-; SSSE3-NEXT:    pand %xmm4, %xmm0
-; SSSE3-NEXT:    pandn %xmm1, %xmm4
-; SSSE3-NEXT:    por %xmm0, %xmm4
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,2,2,3]
+; SSSE3-NEXT:    pand %xmm9, %xmm7
+; SSSE3-NEXT:    pand %xmm7, %xmm5
+; SSSE3-NEXT:    pandn %xmm2, %xmm7
+; SSSE3-NEXT:    por %xmm5, %xmm7
+; SSSE3-NEXT:    movdqa %xmm0, %xmm5
+; SSSE3-NEXT:    pxor %xmm3, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[0,0,2,2]
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm5
+; SSSE3-NEXT:    pand %xmm4, %xmm5
+; SSSE3-NEXT:    pand %xmm5, %xmm0
+; SSSE3-NEXT:    pandn %xmm2, %xmm5
+; SSSE3-NEXT:    por %xmm0, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,2,2,3]
 ; SSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm8[0,2,2,3]
-; SSSE3-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3]
+; SSSE3-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7]
+; SSSE3-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,2,2,3]
 ; SSSE3-NEXT:    pshuflw {{.*#+}} xmm3 = xmm0[0,1,0,2,4,5,6,7]
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
 ; SSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,0,2,4,5,6,7]
 ; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]
-; SSSE3-NEXT:    movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
+; SSSE3-NEXT:    movsd {{.*#+}} xmm0 = xmm2[0],xmm0[1]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v8i64_v8i16:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa (%rdi), %xmm7
-; SSE41-NEXT:    movdqa 16(%rdi), %xmm1
-; SSE41-NEXT:    movdqa 32(%rdi), %xmm3
-; SSE41-NEXT:    movdqa 48(%rdi), %xmm6
-; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [65535,65535]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movdqa (%rdi), %xmm8
+; SSE41-NEXT:    movdqa 16(%rdi), %xmm2
+; SSE41-NEXT:    movdqa 32(%rdi), %xmm4
+; SSE41-NEXT:    movdqa 48(%rdi), %xmm7
+; SSE41-NEXT:    movapd {{.*#+}} xmm3 = [65535,65535]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm6, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002324991,9223372039002324991]
+; SSE41-NEXT:    movdqa %xmm5, %xmm9
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm9
+; SSE41-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [2147549183,2147549183,2147549183,2147549183]
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
-; SSE41-NEXT:    pxor %xmm5, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002324991,9223372039002324991]
-; SSE41-NEXT:    movdqa %xmm4, %xmm8
+; SSE41-NEXT:    pcmpgtd %xmm10, %xmm0
+; SSE41-NEXT:    pand %xmm9, %xmm0
+; SSE41-NEXT:    movapd %xmm3, %xmm9
+; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm9
+; SSE41-NEXT:    movdqa %xmm8, %xmm0
+; SSE41-NEXT:    pxor %xmm6, %xmm0
+; SSE41-NEXT:    movdqa %xmm5, %xmm2
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm10, %xmm0
+; SSE41-NEXT:    pand %xmm2, %xmm0
+; SSE41-NEXT:    movapd %xmm3, %xmm2
+; SSE41-NEXT:    blendvpd %xmm0, %xmm8, %xmm2
+; SSE41-NEXT:    packusdw %xmm9, %xmm2
+; SSE41-NEXT:    movdqa %xmm7, %xmm0
+; SSE41-NEXT:    pxor %xmm6, %xmm0
+; SSE41-NEXT:    movdqa %xmm5, %xmm8
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm8
-; SSE41-NEXT:    movdqa %xmm4, %xmm9
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm9 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm9, %xmm0
 ; SSE41-NEXT:    pand %xmm8, %xmm0
-; SSE41-NEXT:    por %xmm9, %xmm0
-; SSE41-NEXT:    movapd %xmm2, %xmm8
-; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm8
-; SSE41-NEXT:    movdqa %xmm7, %xmm0
-; SSE41-NEXT:    pxor %xmm5, %xmm0
-; SSE41-NEXT:    movdqa %xmm4, %xmm1
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa %xmm4, %xmm9
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    por %xmm9, %xmm0
-; SSE41-NEXT:    movapd %xmm2, %xmm1
-; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm1
-; SSE41-NEXT:    packusdw %xmm8, %xmm1
-; SSE41-NEXT:    movdqa %xmm6, %xmm0
-; SSE41-NEXT:    pxor %xmm5, %xmm0
-; SSE41-NEXT:    movdqa %xmm4, %xmm7
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm7
-; SSE41-NEXT:    movdqa %xmm4, %xmm8
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm8
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm8[0,0,2,2]
-; SSE41-NEXT:    pand %xmm7, %xmm0
-; SSE41-NEXT:    por %xmm8, %xmm0
-; SSE41-NEXT:    movapd %xmm2, %xmm7
-; SSE41-NEXT:    blendvpd %xmm0, %xmm6, %xmm7
-; SSE41-NEXT:    pxor %xmm3, %xmm5
-; SSE41-NEXT:    movdqa %xmm4, %xmm6
-; SSE41-NEXT:    pcmpeqd %xmm5, %xmm6
-; SSE41-NEXT:    pcmpgtd %xmm5, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pand %xmm6, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm2
-; SSE41-NEXT:    packusdw %xmm7, %xmm2
-; SSE41-NEXT:    packusdw %xmm2, %xmm1
+; SSE41-NEXT:    movapd %xmm3, %xmm8
+; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm8
+; SSE41-NEXT:    pxor %xmm4, %xmm6
+; SSE41-NEXT:    pcmpeqd %xmm6, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm1
+; SSE41-NEXT:    pand %xmm5, %xmm1
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm3
+; SSE41-NEXT:    packusdw %xmm8, %xmm3
+; SSE41-NEXT:    packusdw %xmm3, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_usat_v8i64_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa (%rdi), %xmm0
 ; AVX1-NEXT:    vmovdqa 16(%rdi), %xmm1
 ; AVX1-NEXT:    vmovdqa 32(%rdi), %xmm2
 ; AVX1-NEXT:    vmovdqa 48(%rdi), %xmm3

 ; Unsigned saturation truncation to vXi8
 ;
 
 define <2 x i8> @trunc_usat_v2i64_v2i8(<2 x i64> %a0) {
 ; SSE2-LABEL: trunc_usat_v2i64_v2i8:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    pxor %xmm0, %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259711,9223372039002259711]
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [2147483903,2147483903,2147483903,2147483903]
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm3
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm2
-; SSE2-NEXT:    pand %xmm2, %xmm0
-; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSE2-NEXT:    por %xmm2, %xmm0
+; SSE2-NEXT:    pand %xmm1, %xmm0
+; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE2-NEXT:    packuswb %xmm0, %xmm0
 ; SSE2-NEXT:    packuswb %xmm0, %xmm0
 ; SSE2-NEXT:    packuswb %xmm0, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v2i64_v2i8:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    pxor %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259711,9223372039002259711]
-; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [2147483903,2147483903,2147483903,2147483903]
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm3
 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSSE3-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    por %xmm1, %xmm2
-; SSSE3-NEXT:    pand %xmm2, %xmm0
-; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    pand %xmm1, %xmm0
+; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    por %xmm1, %xmm0
 ; SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v2i64_v2i8:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm1
 ; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [255,255]
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [9223372039002259456,9223372039002259456]
 ; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259711,9223372039002259711]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [2147483903,2147483903,2147483903,2147483903]
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE41-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    pand %xmm4, %xmm0
-; SSE41-NEXT:    por %xmm3, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
 ; SSE41-NEXT:    pshufb {{.*#+}} xmm2 = xmm2[0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
 ; SSE41-NEXT:    movdqa %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: trunc_usat_v2i64_v2i8:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vmovapd {{.*#+}} xmm1 = [255,255]
; SKX-NEXT:    retq
   ret <2 x i8> %3
 }
 
 define void @trunc_usat_v2i64_v2i8_store(<2 x i64> %a0, ptr %p1) {
 ; SSE2-LABEL: trunc_usat_v2i64_v2i8_store:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    pxor %xmm0, %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259711,9223372039002259711]
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [2147483903,2147483903,2147483903,2147483903]
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm3
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSE2-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm2
-; SSE2-NEXT:    pand %xmm2, %xmm0
-; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSE2-NEXT:    por %xmm0, %xmm2
-; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSE2-NEXT:    packuswb %xmm2, %xmm2
-; SSE2-NEXT:    packuswb %xmm2, %xmm2
-; SSE2-NEXT:    packuswb %xmm2, %xmm2
-; SSE2-NEXT:    movd %xmm2, %eax
+; SSE2-NEXT:    pand %xmm1, %xmm0
+; SSE2-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE2-NEXT:    por %xmm0, %xmm1
+; SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSE2-NEXT:    packuswb %xmm1, %xmm1
+; SSE2-NEXT:    packuswb %xmm1, %xmm1
+; SSE2-NEXT:    packuswb %xmm1, %xmm1
+; SSE2-NEXT:    movd %xmm1, %eax
 ; SSE2-NEXT:    movw %ax, (%rdi)
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v2i64_v2i8_store:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    pxor %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259711,9223372039002259711]
-; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [2147483903,2147483903,2147483903,2147483903]
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm3
 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSSE3-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    por %xmm1, %xmm2
-; SSSE3-NEXT:    pand %xmm2, %xmm0
-; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
-; SSSE3-NEXT:    por %xmm0, %xmm2
-; SSSE3-NEXT:    pshufb {{.*#+}} xmm2 = xmm2[0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
-; SSSE3-NEXT:    movd %xmm2, %eax
+; SSSE3-NEXT:    pand %xmm1, %xmm0
+; SSSE3-NEXT:    pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; SSSE3-NEXT:    por %xmm0, %xmm1
+; SSSE3-NEXT:    pshufb {{.*#+}} xmm1 = xmm1[0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
+; SSSE3-NEXT:    movd %xmm1, %eax
 ; SSSE3-NEXT:    movw %ax, (%rdi)
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v2i64_v2i8_store:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm1
 ; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [255,255]
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [9223372039002259456,9223372039002259456]
 ; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259711,9223372039002259711]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [2147483903,2147483903,2147483903,2147483903]
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE41-NEXT:    pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
 ; SSE41-NEXT:    pand %xmm4, %xmm0
-; SSE41-NEXT:    por %xmm3, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
 ; SSE41-NEXT:    pshufb {{.*#+}} xmm2 = xmm2[0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
 ; SSE41-NEXT:    pextrw $0, %xmm2, (%rdi)
 ; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: trunc_usat_v2i64_v2i8_store:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vmovapd {{.*#+}} xmm1 = [255,255]

 
 define <4 x i8> @trunc_usat_v4i64_v4i8(<4 x i64> %a0) {
 ; SSE2-LABEL: trunc_usat_v4i64_v4i8:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [255,255]
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm0, %xmm4
 ; SSE2-NEXT:    pxor %xmm3, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm5
-; SSE2-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259711,9223372039002259711]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm6 = [2147483903,2147483903,2147483903,2147483903]
 ; SSE2-NEXT:    movdqa %xmm6, %xmm7
-; SSE2-NEXT:    pcmpgtd %xmm4, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[0,0,2,2]
-; SSE2-NEXT:    pand %xmm5, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
-; SSE2-NEXT:    por %xmm4, %xmm5
-; SSE2-NEXT:    pand %xmm5, %xmm0
-; SSE2-NEXT:    pandn %xmm2, %xmm5
-; SSE2-NEXT:    por %xmm5, %xmm0
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm4
+; SSE2-NEXT:    pand %xmm7, %xmm4
+; SSE2-NEXT:    pand %xmm4, %xmm0
+; SSE2-NEXT:    pandn %xmm2, %xmm4
+; SSE2-NEXT:    por %xmm4, %xmm0
 ; SSE2-NEXT:    movdqa %xmm1, %xmm4
 ; SSE2-NEXT:    pxor %xmm3, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm5
-; SSE2-NEXT:    pcmpgtd %xmm4, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[0,0,2,2]
-; SSE2-NEXT:    pand %xmm5, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm4
+; SSE2-NEXT:    pand %xmm6, %xmm4
 ; SSE2-NEXT:    pand %xmm4, %xmm1
 ; SSE2-NEXT:    pandn %xmm2, %xmm4
 ; SSE2-NEXT:    por %xmm1, %xmm4
 ; SSE2-NEXT:    pand %xmm2, %xmm4
 ; SSE2-NEXT:    pand %xmm2, %xmm0
 ; SSE2-NEXT:    packuswb %xmm4, %xmm0
 ; SSE2-NEXT:    packuswb %xmm0, %xmm0
 ; SSE2-NEXT:    packuswb %xmm0, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v4i64_v4i8:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [255,255]
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm4
 ; SSSE3-NEXT:    pxor %xmm3, %xmm4
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm5
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259711,9223372039002259711]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm6 = [2147483903,2147483903,2147483903,2147483903]
 ; SSSE3-NEXT:    movdqa %xmm6, %xmm7
-; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm5, %xmm4
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
-; SSSE3-NEXT:    por %xmm4, %xmm5
-; SSSE3-NEXT:    pand %xmm5, %xmm0
-; SSSE3-NEXT:    pandn %xmm2, %xmm5
-; SSSE3-NEXT:    por %xmm5, %xmm0
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm4
+; SSSE3-NEXT:    pand %xmm7, %xmm4
+; SSSE3-NEXT:    pand %xmm4, %xmm0
+; SSSE3-NEXT:    pandn %xmm2, %xmm4
+; SSSE3-NEXT:    por %xmm4, %xmm0
 ; SSSE3-NEXT:    movdqa %xmm1, %xmm4
 ; SSSE3-NEXT:    pxor %xmm3, %xmm4
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm5
-; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm5, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
-; SSSE3-NEXT:    por %xmm3, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm4
+; SSSE3-NEXT:    pand %xmm6, %xmm4
 ; SSSE3-NEXT:    pand %xmm4, %xmm1
 ; SSSE3-NEXT:    pandn %xmm2, %xmm4
 ; SSSE3-NEXT:    por %xmm1, %xmm4
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
 ; SSSE3-NEXT:    pshufb %xmm1, %xmm4
 ; SSSE3-NEXT:    pshufb %xmm1, %xmm0
 ; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v4i64_v4i8:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    movapd {{.*#+}} xmm4 = [255,255]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259456,9223372039002259456]
-; SSE41-NEXT:    pxor %xmm5, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259711,9223372039002259711]
-; SSE41-NEXT:    movdqa %xmm6, %xmm3
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
-; SSE41-NEXT:    movdqa %xmm6, %xmm7
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm7
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
-; SSE41-NEXT:    pand %xmm3, %xmm0
-; SSE41-NEXT:    por %xmm7, %xmm0
-; SSE41-NEXT:    movapd %xmm4, %xmm3
-; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm3
-; SSE41-NEXT:    pxor %xmm1, %xmm5
-; SSE41-NEXT:    movdqa %xmm6, %xmm2
-; SSE41-NEXT:    pcmpeqd %xmm5, %xmm2
-; SSE41-NEXT:    pcmpgtd %xmm5, %xmm6
+; SSE41-NEXT:    movapd {{.*#+}} xmm5 = [255,255]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    pxor %xmm6, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm7 = [9223372039002259711,9223372039002259711]
+; SSE41-NEXT:    movdqa %xmm7, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm8 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [2147483903,2147483903,2147483903,2147483903]
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm8, %xmm0
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    movapd %xmm5, %xmm4
+; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm4
+; SSE41-NEXT:    pxor %xmm1, %xmm6
+; SSE41-NEXT:    pcmpeqd %xmm6, %xmm7
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
-; SSE41-NEXT:    pand %xmm2, %xmm0
-; SSE41-NEXT:    por %xmm6, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE41-NEXT:    pand %xmm7, %xmm3
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm5
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
+; SSE41-NEXT:    pshufb %xmm0, %xmm5
 ; SSE41-NEXT:    pshufb %xmm0, %xmm4
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
-; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_usat_v4i64_v4i8:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
 ; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [9223372036854776063,9223372036854776063]
 ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2

 
 define void @trunc_usat_v4i64_v4i8_store(<4 x i64> %a0, ptr%p1) {
 ; SSE2-LABEL: trunc_usat_v4i64_v4i8_store:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [255,255]
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm0, %xmm3
 ; SSE2-NEXT:    pxor %xmm4, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm4, %xmm5
-; SSE2-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259711,9223372039002259711]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm6 = [2147483903,2147483903,2147483903,2147483903]
 ; SSE2-NEXT:    movdqa %xmm6, %xmm7
-; SSE2-NEXT:    pcmpgtd %xmm3, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
-; SSE2-NEXT:    pand %xmm5, %xmm8
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm7[1,1,3,3]
-; SSE2-NEXT:    por %xmm8, %xmm3
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm3
+; SSE2-NEXT:    pand %xmm7, %xmm3
 ; SSE2-NEXT:    pand %xmm3, %xmm0
 ; SSE2-NEXT:    pandn %xmm2, %xmm3
 ; SSE2-NEXT:    por %xmm0, %xmm3
 ; SSE2-NEXT:    movdqa %xmm1, %xmm0
 ; SSE2-NEXT:    pxor %xmm4, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm4, %xmm5
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
-; SSE2-NEXT:    pand %xmm5, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
-; SSE2-NEXT:    por %xmm0, %xmm4
-; SSE2-NEXT:    pand %xmm4, %xmm1
-; SSE2-NEXT:    pandn %xmm2, %xmm4
-; SSE2-NEXT:    por %xmm1, %xmm4
-; SSE2-NEXT:    pand %xmm2, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm0
+; SSE2-NEXT:    pand %xmm6, %xmm0
+; SSE2-NEXT:    pand %xmm0, %xmm1
+; SSE2-NEXT:    pandn %xmm2, %xmm0
+; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    pand %xmm2, %xmm0
 ; SSE2-NEXT:    pand %xmm2, %xmm3
-; SSE2-NEXT:    packuswb %xmm4, %xmm3
+; SSE2-NEXT:    packuswb %xmm0, %xmm3
 ; SSE2-NEXT:    packuswb %xmm3, %xmm3
 ; SSE2-NEXT:    packuswb %xmm3, %xmm3
 ; SSE2-NEXT:    movd %xmm3, (%rdi)
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v4i64_v4i8_store:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [255,255]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [255,255]
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm3
-; SSSE3-NEXT:    pxor %xmm4, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm5
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259711,9223372039002259711]
+; SSSE3-NEXT:    movdqa %xmm0, %xmm2
+; SSSE3-NEXT:    pxor %xmm4, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm6 = [2147483903,2147483903,2147483903,2147483903]
 ; SSSE3-NEXT:    movdqa %xmm6, %xmm7
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm5, %xmm8
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm7[1,1,3,3]
-; SSSE3-NEXT:    por %xmm8, %xmm3
-; SSSE3-NEXT:    pand %xmm3, %xmm0
-; SSSE3-NEXT:    pandn %xmm2, %xmm3
-; SSSE3-NEXT:    por %xmm0, %xmm3
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm2
+; SSSE3-NEXT:    pand %xmm7, %xmm2
+; SSSE3-NEXT:    pand %xmm2, %xmm0
+; SSSE3-NEXT:    pandn %xmm3, %xmm2
+; SSSE3-NEXT:    por %xmm0, %xmm2
 ; SSSE3-NEXT:    movdqa %xmm1, %xmm0
 ; SSSE3-NEXT:    pxor %xmm4, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm5
-; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm5, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
-; SSSE3-NEXT:    por %xmm0, %xmm4
-; SSSE3-NEXT:    pand %xmm4, %xmm1
-; SSSE3-NEXT:    pandn %xmm2, %xmm4
-; SSSE3-NEXT:    por %xmm1, %xmm4
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
-; SSSE3-NEXT:    pshufb %xmm0, %xmm4
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
-; SSSE3-NEXT:    movd %xmm3, (%rdi)
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,0,2,2]
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm0
+; SSSE3-NEXT:    pand %xmm6, %xmm0
+; SSSE3-NEXT:    pand %xmm0, %xmm1
+; SSSE3-NEXT:    pandn %xmm3, %xmm0
+; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
+; SSSE3-NEXT:    pshufb %xmm1, %xmm0
+; SSSE3-NEXT:    pshufb %xmm1, %xmm2
+; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; SSSE3-NEXT:    movd %xmm2, (%rdi)
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v4i64_v4i8_store:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    movapd {{.*#+}} xmm3 = [255,255]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]
-; SSE41-NEXT:    pxor %xmm4, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259711,9223372039002259711]
-; SSE41-NEXT:    movdqa %xmm5, %xmm6
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm6
-; SSE41-NEXT:    movdqa %xmm5, %xmm7
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm7
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
-; SSE41-NEXT:    pand %xmm6, %xmm0
-; SSE41-NEXT:    por %xmm7, %xmm0
-; SSE41-NEXT:    movapd %xmm3, %xmm6
-; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm6
-; SSE41-NEXT:    pxor %xmm1, %xmm4
-; SSE41-NEXT:    movdqa %xmm5, %xmm2
-; SSE41-NEXT:    pcmpeqd %xmm4, %xmm2
-; SSE41-NEXT:    pcmpgtd %xmm4, %xmm5
+; SSE41-NEXT:    movapd {{.*#+}} xmm4 = [255,255]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    pxor %xmm5, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259711,9223372039002259711]
+; SSE41-NEXT:    movdqa %xmm6, %xmm7
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm7
+; SSE41-NEXT:    pshufd {{.*#+}} xmm8 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [2147483903,2147483903,2147483903,2147483903]
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm8, %xmm0
+; SSE41-NEXT:    pand %xmm7, %xmm0
+; SSE41-NEXT:    movapd %xmm4, %xmm7
+; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm7
+; SSE41-NEXT:    pxor %xmm1, %xmm5
+; SSE41-NEXT:    pcmpeqd %xmm5, %xmm6
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
-; SSE41-NEXT:    pand %xmm2, %xmm0
-; SSE41-NEXT:    por %xmm5, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm3
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE41-NEXT:    pand %xmm6, %xmm3
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm4
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = <0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u>
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    pshufb %xmm0, %xmm6
-; SSE41-NEXT:    punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
-; SSE41-NEXT:    movd %xmm6, (%rdi)
+; SSE41-NEXT:    pshufb %xmm0, %xmm4
+; SSE41-NEXT:    pshufb %xmm0, %xmm7
+; SSE41-NEXT:    punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm4[0],xmm7[1],xmm4[1],xmm7[2],xmm4[2],xmm7[3],xmm4[3]
+; SSE41-NEXT:    movd %xmm7, (%rdi)
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_usat_v4i64_v4i8_store:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
 ; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [9223372036854776063,9223372036854776063]
 ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2

 ; SSE2-NEXT:    movdqa (%rdi), %xmm6
 ; SSE2-NEXT:    movdqa 16(%rdi), %xmm0
 ; SSE2-NEXT:    movdqa 32(%rdi), %xmm1
 ; SSE2-NEXT:    movdqa 48(%rdi), %xmm5
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [255,255]
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm0, %xmm7
 ; SSE2-NEXT:    pxor %xmm3, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm8
-; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259711,9223372039002259711]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483903,2147483903,2147483903,2147483903]
 ; SSE2-NEXT:    movdqa %xmm4, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm7, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm8, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm7, %xmm8
-; SSE2-NEXT:    pand %xmm8, %xmm0
-; SSE2-NEXT:    pandn %xmm2, %xmm8
-; SSE2-NEXT:    por %xmm0, %xmm8
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm7
+; SSE2-NEXT:    pand %xmm9, %xmm7
+; SSE2-NEXT:    pand %xmm7, %xmm0
+; SSE2-NEXT:    pandn %xmm2, %xmm7
+; SSE2-NEXT:    por %xmm0, %xmm7
 ; SSE2-NEXT:    movdqa %xmm6, %xmm0
 ; SSE2-NEXT:    pxor %xmm3, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    movdqa %xmm4, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm7, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm10, %xmm0
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm0
+; SSE2-NEXT:    pand %xmm9, %xmm0
 ; SSE2-NEXT:    pand %xmm0, %xmm6
 ; SSE2-NEXT:    pandn %xmm2, %xmm0
 ; SSE2-NEXT:    por %xmm6, %xmm0
-; SSE2-NEXT:    packuswb %xmm8, %xmm0
+; SSE2-NEXT:    packuswb %xmm7, %xmm0
 ; SSE2-NEXT:    movdqa %xmm5, %xmm6
 ; SSE2-NEXT:    pxor %xmm3, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
 ; SSE2-NEXT:    movdqa %xmm4, %xmm8
-; SSE2-NEXT:    pcmpgtd %xmm6, %xmm8
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm8[0,0,2,2]
-; SSE2-NEXT:    pand %xmm7, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm8[1,1,3,3]
-; SSE2-NEXT:    por %xmm6, %xmm7
-; SSE2-NEXT:    pand %xmm7, %xmm5
-; SSE2-NEXT:    pandn %xmm2, %xmm7
-; SSE2-NEXT:    por %xmm5, %xmm7
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm8
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm6
+; SSE2-NEXT:    pand %xmm8, %xmm6
+; SSE2-NEXT:    pand %xmm6, %xmm5
+; SSE2-NEXT:    pandn %xmm2, %xmm6
+; SSE2-NEXT:    por %xmm5, %xmm6
 ; SSE2-NEXT:    movdqa %xmm1, %xmm5
 ; SSE2-NEXT:    pxor %xmm3, %xmm5
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm6
-; SSE2-NEXT:    pcmpgtd %xmm5, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[0,0,2,2]
-; SSE2-NEXT:    pand %xmm6, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm4
-; SSE2-NEXT:    pand %xmm4, %xmm1
-; SSE2-NEXT:    pandn %xmm2, %xmm4
-; SSE2-NEXT:    por %xmm1, %xmm4
-; SSE2-NEXT:    packuswb %xmm7, %xmm4
-; SSE2-NEXT:    packuswb %xmm4, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm5
+; SSE2-NEXT:    pand %xmm4, %xmm5
+; SSE2-NEXT:    pand %xmm5, %xmm1
+; SSE2-NEXT:    pandn %xmm2, %xmm5
+; SSE2-NEXT:    por %xmm1, %xmm5
+; SSE2-NEXT:    packuswb %xmm6, %xmm5
+; SSE2-NEXT:    packuswb %xmm5, %xmm0
 ; SSE2-NEXT:    packuswb %xmm0, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v8i64_v8i8:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa (%rdi), %xmm6
 ; SSSE3-NEXT:    movdqa 16(%rdi), %xmm0
 ; SSSE3-NEXT:    movdqa 32(%rdi), %xmm1
 ; SSSE3-NEXT:    movdqa 48(%rdi), %xmm5
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [255,255]
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm7
 ; SSSE3-NEXT:    pxor %xmm3, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm8
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259711,9223372039002259711]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm4 = [2147483903,2147483903,2147483903,2147483903]
 ; SSSE3-NEXT:    movdqa %xmm4, %xmm9
-; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm9
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm9[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm8, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,1,3,3]
-; SSSE3-NEXT:    por %xmm7, %xmm8
-; SSSE3-NEXT:    pand %xmm8, %xmm0
-; SSSE3-NEXT:    pandn %xmm2, %xmm8
-; SSSE3-NEXT:    por %xmm0, %xmm8
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm7
+; SSSE3-NEXT:    pand %xmm9, %xmm7
+; SSSE3-NEXT:    pand %xmm7, %xmm0
+; SSSE3-NEXT:    pandn %xmm2, %xmm7
+; SSSE3-NEXT:    por %xmm0, %xmm7
 ; SSSE3-NEXT:    movdqa %xmm6, %xmm0
 ; SSSE3-NEXT:    pxor %xmm3, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm0[0,0,2,2]
 ; SSSE3-NEXT:    movdqa %xmm4, %xmm9
-; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm7, %xmm10
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[1,1,3,3]
-; SSSE3-NEXT:    por %xmm10, %xmm0
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm0
+; SSSE3-NEXT:    pand %xmm9, %xmm0
 ; SSSE3-NEXT:    pand %xmm0, %xmm6
 ; SSSE3-NEXT:    pandn %xmm2, %xmm0
 ; SSSE3-NEXT:    por %xmm6, %xmm0
-; SSSE3-NEXT:    packuswb %xmm8, %xmm0
+; SSSE3-NEXT:    packuswb %xmm7, %xmm0
 ; SSSE3-NEXT:    movdqa %xmm5, %xmm6
 ; SSSE3-NEXT:    pxor %xmm3, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
 ; SSSE3-NEXT:    movdqa %xmm4, %xmm8
-; SSSE3-NEXT:    pcmpgtd %xmm6, %xmm8
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm8[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm7, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm8[1,1,3,3]
-; SSSE3-NEXT:    por %xmm6, %xmm7
-; SSSE3-NEXT:    pand %xmm7, %xmm5
-; SSSE3-NEXT:    pandn %xmm2, %xmm7
-; SSSE3-NEXT:    por %xmm5, %xmm7
+; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm8
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm6
+; SSSE3-NEXT:    pand %xmm8, %xmm6
+; SSSE3-NEXT:    pand %xmm6, %xmm5
+; SSSE3-NEXT:    pandn %xmm2, %xmm6
+; SSSE3-NEXT:    por %xmm5, %xmm6
 ; SSSE3-NEXT:    movdqa %xmm1, %xmm5
 ; SSSE3-NEXT:    pxor %xmm3, %xmm5
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm6
-; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm4
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm6, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
-; SSSE3-NEXT:    por %xmm3, %xmm4
-; SSSE3-NEXT:    pand %xmm4, %xmm1
-; SSSE3-NEXT:    pandn %xmm2, %xmm4
-; SSSE3-NEXT:    por %xmm1, %xmm4
-; SSSE3-NEXT:    packuswb %xmm7, %xmm4
-; SSSE3-NEXT:    packuswb %xmm4, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2]
+; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm5
+; SSSE3-NEXT:    pand %xmm4, %xmm5
+; SSSE3-NEXT:    pand %xmm5, %xmm1
+; SSSE3-NEXT:    pandn %xmm2, %xmm5
+; SSSE3-NEXT:    por %xmm1, %xmm5
+; SSSE3-NEXT:    packuswb %xmm6, %xmm5
+; SSSE3-NEXT:    packuswb %xmm5, %xmm0
 ; SSSE3-NEXT:    packuswb %xmm0, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v8i64_v8i8:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa (%rdi), %xmm7
-; SSE41-NEXT:    movdqa 16(%rdi), %xmm1
-; SSE41-NEXT:    movdqa 32(%rdi), %xmm3
-; SSE41-NEXT:    movdqa 48(%rdi), %xmm6
-; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [255,255]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movdqa (%rdi), %xmm8
+; SSE41-NEXT:    movdqa 16(%rdi), %xmm2
+; SSE41-NEXT:    movdqa 32(%rdi), %xmm4
+; SSE41-NEXT:    movdqa 48(%rdi), %xmm7
+; SSE41-NEXT:    movapd {{.*#+}} xmm3 = [255,255]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm6, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259711,9223372039002259711]
+; SSE41-NEXT:    movdqa %xmm5, %xmm9
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm9
+; SSE41-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [2147483903,2147483903,2147483903,2147483903]
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
-; SSE41-NEXT:    pxor %xmm5, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259711,9223372039002259711]
-; SSE41-NEXT:    movdqa %xmm4, %xmm8
+; SSE41-NEXT:    pcmpgtd %xmm10, %xmm0
+; SSE41-NEXT:    pand %xmm9, %xmm0
+; SSE41-NEXT:    movapd %xmm3, %xmm9
+; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm9
+; SSE41-NEXT:    movdqa %xmm8, %xmm0
+; SSE41-NEXT:    pxor %xmm6, %xmm0
+; SSE41-NEXT:    movdqa %xmm5, %xmm2
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm10, %xmm0
+; SSE41-NEXT:    pand %xmm2, %xmm0
+; SSE41-NEXT:    movapd %xmm3, %xmm2
+; SSE41-NEXT:    blendvpd %xmm0, %xmm8, %xmm2
+; SSE41-NEXT:    packusdw %xmm9, %xmm2
+; SSE41-NEXT:    movdqa %xmm7, %xmm0
+; SSE41-NEXT:    pxor %xmm6, %xmm0
+; SSE41-NEXT:    movdqa %xmm5, %xmm8
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm8
-; SSE41-NEXT:    movdqa %xmm4, %xmm9
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm9 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm9, %xmm0
 ; SSE41-NEXT:    pand %xmm8, %xmm0
-; SSE41-NEXT:    por %xmm9, %xmm0
-; SSE41-NEXT:    movapd %xmm2, %xmm8
-; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm8
-; SSE41-NEXT:    movdqa %xmm7, %xmm0
-; SSE41-NEXT:    pxor %xmm5, %xmm0
-; SSE41-NEXT:    movdqa %xmm4, %xmm1
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa %xmm4, %xmm9
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    por %xmm9, %xmm0
-; SSE41-NEXT:    movapd %xmm2, %xmm1
-; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm1
-; SSE41-NEXT:    packusdw %xmm8, %xmm1
-; SSE41-NEXT:    movdqa %xmm6, %xmm0
-; SSE41-NEXT:    pxor %xmm5, %xmm0
-; SSE41-NEXT:    movdqa %xmm4, %xmm7
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm7
-; SSE41-NEXT:    movdqa %xmm4, %xmm8
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm8
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm8[0,0,2,2]
-; SSE41-NEXT:    pand %xmm7, %xmm0
-; SSE41-NEXT:    por %xmm8, %xmm0
-; SSE41-NEXT:    movapd %xmm2, %xmm7
-; SSE41-NEXT:    blendvpd %xmm0, %xmm6, %xmm7
-; SSE41-NEXT:    pxor %xmm3, %xmm5
-; SSE41-NEXT:    movdqa %xmm4, %xmm6
-; SSE41-NEXT:    pcmpeqd %xmm5, %xmm6
-; SSE41-NEXT:    pcmpgtd %xmm5, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pand %xmm6, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm2
-; SSE41-NEXT:    packusdw %xmm7, %xmm2
-; SSE41-NEXT:    packusdw %xmm2, %xmm1
-; SSE41-NEXT:    packuswb %xmm1, %xmm1
+; SSE41-NEXT:    movapd %xmm3, %xmm8
+; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm8
+; SSE41-NEXT:    pxor %xmm4, %xmm6
+; SSE41-NEXT:    pcmpeqd %xmm6, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm1
+; SSE41-NEXT:    pand %xmm5, %xmm1
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm3
+; SSE41-NEXT:    packusdw %xmm8, %xmm3
+; SSE41-NEXT:    packusdw %xmm3, %xmm2
+; SSE41-NEXT:    packuswb %xmm2, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_usat_v8i64_v8i8:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa (%rdi), %xmm0
 ; AVX1-NEXT:    vmovdqa 16(%rdi), %xmm1
 ; AVX1-NEXT:    vmovdqa 32(%rdi), %xmm2
 ; AVX1-NEXT:    vmovdqa 48(%rdi), %xmm3

 ; SSE2-NEXT:    movdqa (%rdi), %xmm6
 ; SSE2-NEXT:    movdqa 16(%rdi), %xmm5
 ; SSE2-NEXT:    movdqa 32(%rdi), %xmm0
 ; SSE2-NEXT:    movdqa 48(%rdi), %xmm4
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [255,255]
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm5, %xmm7
 ; SSE2-NEXT:    pxor %xmm2, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm2, %xmm8
-; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259711,9223372039002259711]
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [2147483903,2147483903,2147483903,2147483903]
 ; SSE2-NEXT:    movdqa %xmm3, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm7, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm8, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm7, %xmm8
-; SSE2-NEXT:    pand %xmm8, %xmm5
-; SSE2-NEXT:    pandn %xmm1, %xmm8
-; SSE2-NEXT:    por %xmm5, %xmm8
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm7
+; SSE2-NEXT:    pand %xmm9, %xmm7
+; SSE2-NEXT:    pand %xmm7, %xmm5
+; SSE2-NEXT:    pandn %xmm1, %xmm7
+; SSE2-NEXT:    por %xmm5, %xmm7
 ; SSE2-NEXT:    movdqa %xmm6, %xmm5
 ; SSE2-NEXT:    pxor %xmm2, %xmm5
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm2, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[0,0,2,2]
 ; SSE2-NEXT:    movdqa %xmm3, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm5, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm7, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm10, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm5
+; SSE2-NEXT:    pand %xmm9, %xmm5
 ; SSE2-NEXT:    pand %xmm5, %xmm6
 ; SSE2-NEXT:    pandn %xmm1, %xmm5
 ; SSE2-NEXT:    por %xmm6, %xmm5
-; SSE2-NEXT:    packuswb %xmm8, %xmm5
+; SSE2-NEXT:    packuswb %xmm7, %xmm5
 ; SSE2-NEXT:    movdqa %xmm4, %xmm6
 ; SSE2-NEXT:    pxor %xmm2, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm2, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
 ; SSE2-NEXT:    movdqa %xmm3, %xmm8
-; SSE2-NEXT:    pcmpgtd %xmm6, %xmm8
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm8[0,0,2,2]
-; SSE2-NEXT:    pand %xmm7, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm8[1,1,3,3]
-; SSE2-NEXT:    por %xmm6, %xmm7
-; SSE2-NEXT:    pand %xmm7, %xmm4
-; SSE2-NEXT:    pandn %xmm1, %xmm7
-; SSE2-NEXT:    por %xmm4, %xmm7
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm8
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm6
+; SSE2-NEXT:    pand %xmm8, %xmm6
+; SSE2-NEXT:    pand %xmm6, %xmm4
+; SSE2-NEXT:    pandn %xmm1, %xmm6
+; SSE2-NEXT:    por %xmm4, %xmm6
 ; SSE2-NEXT:    movdqa %xmm0, %xmm4
 ; SSE2-NEXT:    pxor %xmm2, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm2, %xmm6
-; SSE2-NEXT:    pcmpgtd %xmm4, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSE2-NEXT:    pand %xmm6, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
-; SSE2-NEXT:    por %xmm2, %xmm3
-; SSE2-NEXT:    pand %xmm3, %xmm0
-; SSE2-NEXT:    pandn %xmm1, %xmm3
-; SSE2-NEXT:    por %xmm0, %xmm3
-; SSE2-NEXT:    packuswb %xmm7, %xmm3
-; SSE2-NEXT:    packuswb %xmm3, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm4
+; SSE2-NEXT:    pand %xmm3, %xmm4
+; SSE2-NEXT:    pand %xmm4, %xmm0
+; SSE2-NEXT:    pandn %xmm1, %xmm4
+; SSE2-NEXT:    por %xmm0, %xmm4
+; SSE2-NEXT:    packuswb %xmm6, %xmm4
+; SSE2-NEXT:    packuswb %xmm4, %xmm5
 ; SSE2-NEXT:    packuswb %xmm5, %xmm5
 ; SSE2-NEXT:    movq %xmm5, (%rsi)
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v8i64_v8i8_store:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa (%rdi), %xmm6
 ; SSSE3-NEXT:    movdqa 16(%rdi), %xmm5
 ; SSSE3-NEXT:    movdqa 32(%rdi), %xmm0
 ; SSSE3-NEXT:    movdqa 48(%rdi), %xmm4
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [255,255]
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    movdqa %xmm5, %xmm7
 ; SSSE3-NEXT:    pxor %xmm2, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm8
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259711,9223372039002259711]
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [2147483903,2147483903,2147483903,2147483903]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm9
-; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm9
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm9[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm8, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,1,3,3]
-; SSSE3-NEXT:    por %xmm7, %xmm8
-; SSSE3-NEXT:    pand %xmm8, %xmm5
-; SSSE3-NEXT:    pandn %xmm1, %xmm8
-; SSSE3-NEXT:    por %xmm5, %xmm8
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm7
+; SSSE3-NEXT:    pand %xmm9, %xmm7
+; SSSE3-NEXT:    pand %xmm7, %xmm5
+; SSSE3-NEXT:    pandn %xmm1, %xmm7
+; SSSE3-NEXT:    por %xmm5, %xmm7
 ; SSSE3-NEXT:    movdqa %xmm6, %xmm5
 ; SSSE3-NEXT:    pxor %xmm2, %xmm5
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[0,0,2,2]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm9
-; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm9
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm7, %xmm10
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm9[1,1,3,3]
-; SSSE3-NEXT:    por %xmm10, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm5
+; SSSE3-NEXT:    pand %xmm9, %xmm5
 ; SSSE3-NEXT:    pand %xmm5, %xmm6
 ; SSSE3-NEXT:    pandn %xmm1, %xmm5
 ; SSSE3-NEXT:    por %xmm6, %xmm5
-; SSSE3-NEXT:    packuswb %xmm8, %xmm5
+; SSSE3-NEXT:    packuswb %xmm7, %xmm5
 ; SSSE3-NEXT:    movdqa %xmm4, %xmm6
 ; SSSE3-NEXT:    pxor %xmm2, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
 ; SSSE3-NEXT:    movdqa %xmm3, %xmm8
-; SSSE3-NEXT:    pcmpgtd %xmm6, %xmm8
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm8[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm7, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm8[1,1,3,3]
-; SSSE3-NEXT:    por %xmm6, %xmm7
-; SSSE3-NEXT:    pand %xmm7, %xmm4
-; SSSE3-NEXT:    pandn %xmm1, %xmm7
-; SSSE3-NEXT:    por %xmm4, %xmm7
+; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm8
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm6
+; SSSE3-NEXT:    pand %xmm8, %xmm6
+; SSSE3-NEXT:    pand %xmm6, %xmm4
+; SSSE3-NEXT:    pandn %xmm1, %xmm6
+; SSSE3-NEXT:    por %xmm4, %xmm6
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm4
 ; SSSE3-NEXT:    pxor %xmm2, %xmm4
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm6
-; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm6, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    por %xmm2, %xmm3
-; SSSE3-NEXT:    pand %xmm3, %xmm0
-; SSSE3-NEXT:    pandn %xmm1, %xmm3
-; SSSE3-NEXT:    por %xmm0, %xmm3
-; SSSE3-NEXT:    packuswb %xmm7, %xmm3
-; SSSE3-NEXT:    packuswb %xmm3, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm4
+; SSSE3-NEXT:    pand %xmm3, %xmm4
+; SSSE3-NEXT:    pand %xmm4, %xmm0
+; SSSE3-NEXT:    pandn %xmm1, %xmm4
+; SSSE3-NEXT:    por %xmm0, %xmm4
+; SSSE3-NEXT:    packuswb %xmm6, %xmm4
+; SSSE3-NEXT:    packuswb %xmm4, %xmm5
 ; SSSE3-NEXT:    packuswb %xmm5, %xmm5
 ; SSSE3-NEXT:    movq %xmm5, (%rsi)
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v8i64_v8i8_store:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa (%rdi), %xmm7
-; SSE41-NEXT:    movdqa 16(%rdi), %xmm6
-; SSE41-NEXT:    movdqa 32(%rdi), %xmm2
-; SSE41-NEXT:    movdqa 48(%rdi), %xmm5
-; SSE41-NEXT:    movapd {{.*#+}} xmm1 = [255,255]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movdqa (%rdi), %xmm8
+; SSE41-NEXT:    movdqa 16(%rdi), %xmm7
+; SSE41-NEXT:    movdqa 32(%rdi), %xmm3
+; SSE41-NEXT:    movdqa 48(%rdi), %xmm6
+; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [255,255]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movdqa %xmm7, %xmm0
+; SSE41-NEXT:    pxor %xmm5, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259711,9223372039002259711]
+; SSE41-NEXT:    movdqa %xmm4, %xmm9
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm9
+; SSE41-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [2147483903,2147483903,2147483903,2147483903]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm10, %xmm0
+; SSE41-NEXT:    pand %xmm9, %xmm0
+; SSE41-NEXT:    movapd %xmm2, %xmm9
+; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm9
+; SSE41-NEXT:    movdqa %xmm8, %xmm0
+; SSE41-NEXT:    pxor %xmm5, %xmm0
+; SSE41-NEXT:    movdqa %xmm4, %xmm7
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm7
+; SSE41-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm10, %xmm0
+; SSE41-NEXT:    pand %xmm7, %xmm0
+; SSE41-NEXT:    movapd %xmm2, %xmm7
+; SSE41-NEXT:    blendvpd %xmm0, %xmm8, %xmm7
+; SSE41-NEXT:    packusdw %xmm9, %xmm7
 ; SSE41-NEXT:    movdqa %xmm6, %xmm0
-; SSE41-NEXT:    pxor %xmm4, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259711,9223372039002259711]
-; SSE41-NEXT:    movdqa %xmm3, %xmm8
+; SSE41-NEXT:    pxor %xmm5, %xmm0
+; SSE41-NEXT:    movdqa %xmm4, %xmm8
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm8
-; SSE41-NEXT:    movdqa %xmm3, %xmm9
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm9 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm9, %xmm0
 ; SSE41-NEXT:    pand %xmm8, %xmm0
-; SSE41-NEXT:    por %xmm9, %xmm0
-; SSE41-NEXT:    movapd %xmm1, %xmm8
+; SSE41-NEXT:    movapd %xmm2, %xmm8
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm6, %xmm8
-; SSE41-NEXT:    movdqa %xmm7, %xmm0
-; SSE41-NEXT:    pxor %xmm4, %xmm0
-; SSE41-NEXT:    movdqa %xmm3, %xmm6
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm6
-; SSE41-NEXT:    movdqa %xmm3, %xmm9
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
-; SSE41-NEXT:    pand %xmm6, %xmm0
-; SSE41-NEXT:    por %xmm9, %xmm0
-; SSE41-NEXT:    movapd %xmm1, %xmm6
-; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm6
-; SSE41-NEXT:    packusdw %xmm8, %xmm6
-; SSE41-NEXT:    movdqa %xmm5, %xmm0
-; SSE41-NEXT:    pxor %xmm4, %xmm0
-; SSE41-NEXT:    movdqa %xmm3, %xmm7
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm7
-; SSE41-NEXT:    movdqa %xmm3, %xmm8
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm8
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm8[0,0,2,2]
-; SSE41-NEXT:    pand %xmm7, %xmm0
-; SSE41-NEXT:    por %xmm8, %xmm0
-; SSE41-NEXT:    movapd %xmm1, %xmm7
-; SSE41-NEXT:    blendvpd %xmm0, %xmm5, %xmm7
-; SSE41-NEXT:    pxor %xmm2, %xmm4
-; SSE41-NEXT:    movdqa %xmm3, %xmm5
-; SSE41-NEXT:    pcmpeqd %xmm4, %xmm5
-; SSE41-NEXT:    pcmpgtd %xmm4, %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
-; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm3, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm1
-; SSE41-NEXT:    packusdw %xmm7, %xmm1
-; SSE41-NEXT:    packusdw %xmm1, %xmm6
-; SSE41-NEXT:    packuswb %xmm6, %xmm6
-; SSE41-NEXT:    movq %xmm6, (%rsi)
+; SSE41-NEXT:    pxor %xmm3, %xmm5
+; SSE41-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm1
+; SSE41-NEXT:    pand %xmm4, %xmm1
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm2
+; SSE41-NEXT:    packusdw %xmm8, %xmm2
+; SSE41-NEXT:    packusdw %xmm2, %xmm7
+; SSE41-NEXT:    packuswb %xmm7, %xmm7
+; SSE41-NEXT:    movq %xmm7, (%rsi)
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_usat_v8i64_v8i8_store:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa (%rdi), %xmm0
 ; AVX1-NEXT:    vmovdqa 16(%rdi), %xmm1
 ; AVX1-NEXT:    vmovdqa 32(%rdi), %xmm2
 ; AVX1-NEXT:    vmovdqa 48(%rdi), %xmm3
; SKX-NEXT:    retq
   store <8 x i8> %3, ptr%p1
   ret void
 }
 
 define <16 x i8> @trunc_usat_v16i64_v16i8(ptr %p0) {
 ; SSE2-LABEL: trunc_usat_v16i64_v16i8:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa 96(%rdi), %xmm1
-; SSE2-NEXT:    movdqa 112(%rdi), %xmm3
+; SSE2-NEXT:    movdqa 112(%rdi), %xmm4
 ; SSE2-NEXT:    movdqa 64(%rdi), %xmm6
 ; SSE2-NEXT:    movdqa 80(%rdi), %xmm7
 ; SSE2-NEXT:    movdqa (%rdi), %xmm10
 ; SSE2-NEXT:    movdqa 16(%rdi), %xmm0
 ; SSE2-NEXT:    movdqa 32(%rdi), %xmm8
 ; SSE2-NEXT:    movdqa 48(%rdi), %xmm9
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [255,255]
-; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSE2-NEXT:    movdqa %xmm0, %xmm11
-; SSE2-NEXT:    pxor %xmm4, %xmm11
-; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm11[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm4, %xmm12
-; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259711,9223372039002259711]
+; SSE2-NEXT:    pxor %xmm3, %xmm11
+; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm11[0,0,2,2]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [2147483903,2147483903,2147483903,2147483903]
 ; SSE2-NEXT:    movdqa %xmm5, %xmm13
-; SSE2-NEXT:    pcmpgtd %xmm11, %xmm13
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm13[0,0,2,2]
-; SSE2-NEXT:    pand %xmm12, %xmm11
-; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm13[1,1,3,3]
-; SSE2-NEXT:    por %xmm11, %xmm12
-; SSE2-NEXT:    pand %xmm12, %xmm0
-; SSE2-NEXT:    pandn %xmm2, %xmm12
-; SSE2-NEXT:    por %xmm0, %xmm12
+; SSE2-NEXT:    pcmpgtd %xmm12, %xmm13
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm11[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm11
+; SSE2-NEXT:    pand %xmm13, %xmm11
+; SSE2-NEXT:    pand %xmm11, %xmm0
+; SSE2-NEXT:    pandn %xmm2, %xmm11
+; SSE2-NEXT:    por %xmm0, %xmm11
 ; SSE2-NEXT:    movdqa %xmm10, %xmm0
-; SSE2-NEXT:    pxor %xmm4, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm4, %xmm11
+; SSE2-NEXT:    pxor %xmm3, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    movdqa %xmm5, %xmm13
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm13
-; SSE2-NEXT:    pshufd {{.*#+}} xmm14 = xmm13[0,0,2,2]
-; SSE2-NEXT:    pand %xmm11, %xmm14
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm13[1,1,3,3]
-; SSE2-NEXT:    por %xmm14, %xmm0
+; SSE2-NEXT:    pcmpgtd %xmm12, %xmm13
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm0
+; SSE2-NEXT:    pand %xmm13, %xmm0
 ; SSE2-NEXT:    pand %xmm0, %xmm10
 ; SSE2-NEXT:    pandn %xmm2, %xmm0
 ; SSE2-NEXT:    por %xmm10, %xmm0
-; SSE2-NEXT:    packuswb %xmm12, %xmm0
+; SSE2-NEXT:    packuswb %xmm11, %xmm0
 ; SSE2-NEXT:    movdqa %xmm9, %xmm10
-; SSE2-NEXT:    pxor %xmm4, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm10[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm4, %xmm11
-; SSE2-NEXT:    movdqa %xmm5, %xmm12
-; SSE2-NEXT:    pcmpgtd %xmm10, %xmm12
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm12[0,0,2,2]
-; SSE2-NEXT:    pand %xmm11, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm12[1,1,3,3]
-; SSE2-NEXT:    por %xmm10, %xmm11
-; SSE2-NEXT:    pand %xmm11, %xmm9
-; SSE2-NEXT:    pandn %xmm2, %xmm11
-; SSE2-NEXT:    por %xmm9, %xmm11
-; SSE2-NEXT:    movdqa %xmm8, %xmm9
-; SSE2-NEXT:    pxor %xmm4, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm4, %xmm10
+; SSE2-NEXT:    pxor %xmm3, %xmm10
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm10[0,0,2,2]
 ; SSE2-NEXT:    movdqa %xmm5, %xmm12
-; SSE2-NEXT:    pcmpgtd %xmm9, %xmm12
-; SSE2-NEXT:    pshufd {{.*#+}} xmm9 = xmm12[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm11, %xmm12
+; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm10
+; SSE2-NEXT:    pand %xmm12, %xmm10
 ; SSE2-NEXT:    pand %xmm10, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm12[1,1,3,3]
-; SSE2-NEXT:    por %xmm9, %xmm10
-; SSE2-NEXT:    pand %xmm10, %xmm8
 ; SSE2-NEXT:    pandn %xmm2, %xmm10
-; SSE2-NEXT:    por %xmm8, %xmm10
-; SSE2-NEXT:    packuswb %xmm11, %xmm10
-; SSE2-NEXT:    packuswb %xmm10, %xmm0
-; SSE2-NEXT:    movdqa %xmm7, %xmm8
-; SSE2-NEXT:    pxor %xmm4, %xmm8
-; SSE2-NEXT:    pshufd {{.*#+}} xmm9 = xmm8[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm4, %xmm9
-; SSE2-NEXT:    movdqa %xmm5, %xmm10
-; SSE2-NEXT:    pcmpgtd %xmm8, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm10[0,0,2,2]
+; SSE2-NEXT:    por %xmm9, %xmm10
+; SSE2-NEXT:    movdqa %xmm8, %xmm9
+; SSE2-NEXT:    pxor %xmm3, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm9[0,0,2,2]
+; SSE2-NEXT:    movdqa %xmm5, %xmm12
+; SSE2-NEXT:    pcmpgtd %xmm11, %xmm12
+; SSE2-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm9
+; SSE2-NEXT:    pand %xmm12, %xmm9
 ; SSE2-NEXT:    pand %xmm9, %xmm8
-; SSE2-NEXT:    pshufd {{.*#+}} xmm9 = xmm10[1,1,3,3]
-; SSE2-NEXT:    por %xmm8, %xmm9
-; SSE2-NEXT:    pand %xmm9, %xmm7
 ; SSE2-NEXT:    pandn %xmm2, %xmm9
-; SSE2-NEXT:    por %xmm7, %xmm9
+; SSE2-NEXT:    por %xmm8, %xmm9
+; SSE2-NEXT:    packuswb %xmm10, %xmm9
+; SSE2-NEXT:    packuswb %xmm9, %xmm0
+; SSE2-NEXT:    movdqa %xmm7, %xmm8
+; SSE2-NEXT:    pxor %xmm3, %xmm8
+; SSE2-NEXT:    pshufd {{.*#+}} xmm9 = xmm8[0,0,2,2]
+; SSE2-NEXT:    movdqa %xmm5, %xmm10
+; SSE2-NEXT:    pcmpgtd %xmm9, %xmm10
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm8
+; SSE2-NEXT:    pand %xmm10, %xmm8
+; SSE2-NEXT:    pand %xmm8, %xmm7
+; SSE2-NEXT:    pandn %xmm2, %xmm8
+; SSE2-NEXT:    por %xmm7, %xmm8
 ; SSE2-NEXT:    movdqa %xmm6, %xmm7
-; SSE2-NEXT:    pxor %xmm4, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm4, %xmm8
+; SSE2-NEXT:    pxor %xmm3, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm9 = xmm7[0,0,2,2]
 ; SSE2-NEXT:    movdqa %xmm5, %xmm10
-; SSE2-NEXT:    pcmpgtd %xmm7, %xmm10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm10[0,0,2,2]
-; SSE2-NEXT:    pand %xmm8, %xmm11
-; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm10[1,1,3,3]
-; SSE2-NEXT:    por %xmm11, %xmm7
+; SSE2-NEXT:    pcmpgtd %xmm9, %xmm10
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm7
+; SSE2-NEXT:    pand %xmm10, %xmm7
 ; SSE2-NEXT:    pand %xmm7, %xmm6
 ; SSE2-NEXT:    pandn %xmm2, %xmm7
 ; SSE2-NEXT:    por %xmm6, %xmm7
-; SSE2-NEXT:    packuswb %xmm9, %xmm7
-; SSE2-NEXT:    movdqa %xmm3, %xmm6
-; SSE2-NEXT:    pxor %xmm4, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm4, %xmm8
+; SSE2-NEXT:    packuswb %xmm8, %xmm7
+; SSE2-NEXT:    movdqa %xmm4, %xmm6
+; SSE2-NEXT:    pxor %xmm3, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2]
 ; SSE2-NEXT:    movdqa %xmm5, %xmm9
-; SSE2-NEXT:    pcmpgtd %xmm6, %xmm9
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm9[0,0,2,2]
-; SSE2-NEXT:    pand %xmm8, %xmm6
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,1,3,3]
-; SSE2-NEXT:    por %xmm6, %xmm8
-; SSE2-NEXT:    pand %xmm8, %xmm3
-; SSE2-NEXT:    pandn %xmm2, %xmm8
-; SSE2-NEXT:    por %xmm3, %xmm8
-; SSE2-NEXT:    movdqa %xmm1, %xmm3
-; SSE2-NEXT:    pxor %xmm4, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[1,1,3,3]
-; SSE2-NEXT:    pcmpeqd %xmm4, %xmm6
-; SSE2-NEXT:    pcmpgtd %xmm3, %xmm5
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[0,0,2,2]
-; SSE2-NEXT:    pand %xmm6, %xmm3
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm6
+; SSE2-NEXT:    pand %xmm9, %xmm6
+; SSE2-NEXT:    pand %xmm6, %xmm4
+; SSE2-NEXT:    pandn %xmm2, %xmm6
+; SSE2-NEXT:    por %xmm4, %xmm6
+; SSE2-NEXT:    movdqa %xmm1, %xmm4
+; SSE2-NEXT:    pxor %xmm3, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm4
+; SSE2-NEXT:    pand %xmm5, %xmm4
 ; SSE2-NEXT:    pand %xmm4, %xmm1
 ; SSE2-NEXT:    pandn %xmm2, %xmm4
 ; SSE2-NEXT:    por %xmm1, %xmm4
-; SSE2-NEXT:    packuswb %xmm8, %xmm4
+; SSE2-NEXT:    packuswb %xmm6, %xmm4
 ; SSE2-NEXT:    packuswb %xmm4, %xmm7
 ; SSE2-NEXT:    packuswb %xmm7, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc_usat_v16i64_v16i8:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa 96(%rdi), %xmm1
-; SSSE3-NEXT:    movdqa 112(%rdi), %xmm3
+; SSSE3-NEXT:    movdqa 112(%rdi), %xmm4
 ; SSSE3-NEXT:    movdqa 64(%rdi), %xmm6
 ; SSSE3-NEXT:    movdqa 80(%rdi), %xmm7
 ; SSSE3-NEXT:    movdqa (%rdi), %xmm10
 ; SSSE3-NEXT:    movdqa 16(%rdi), %xmm0
 ; SSSE3-NEXT:    movdqa 32(%rdi), %xmm8
 ; SSSE3-NEXT:    movdqa 48(%rdi), %xmm9
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [255,255]
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm4 = [9223372039002259456,9223372039002259456]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [9223372039002259456,9223372039002259456]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm11
-; SSSE3-NEXT:    pxor %xmm4, %xmm11
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm12 = xmm11[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm12
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259711,9223372039002259711]
+; SSSE3-NEXT:    pxor %xmm3, %xmm11
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm12 = xmm11[0,0,2,2]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm5 = [2147483903,2147483903,2147483903,2147483903]
 ; SSSE3-NEXT:    movdqa %xmm5, %xmm13
-; SSSE3-NEXT:    pcmpgtd %xmm11, %xmm13
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm13[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm12, %xmm11
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm12 = xmm13[1,1,3,3]
-; SSSE3-NEXT:    por %xmm11, %xmm12
-; SSSE3-NEXT:    pand %xmm12, %xmm0
-; SSSE3-NEXT:    pandn %xmm2, %xmm12
-; SSSE3-NEXT:    por %xmm0, %xmm12
+; SSSE3-NEXT:    pcmpgtd %xmm12, %xmm13
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm11[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm11
+; SSSE3-NEXT:    pand %xmm13, %xmm11
+; SSSE3-NEXT:    pand %xmm11, %xmm0
+; SSSE3-NEXT:    pandn %xmm2, %xmm11
+; SSSE3-NEXT:    por %xmm0, %xmm11
 ; SSSE3-NEXT:    movdqa %xmm10, %xmm0
-; SSSE3-NEXT:    pxor %xmm4, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm11
+; SSSE3-NEXT:    pxor %xmm3, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm12 = xmm0[0,0,2,2]
 ; SSSE3-NEXT:    movdqa %xmm5, %xmm13
-; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm13
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm14 = xmm13[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm11, %xmm14
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm13[1,1,3,3]
-; SSSE3-NEXT:    por %xmm14, %xmm0
+; SSSE3-NEXT:    pcmpgtd %xmm12, %xmm13
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm0
+; SSSE3-NEXT:    pand %xmm13, %xmm0
 ; SSSE3-NEXT:    pand %xmm0, %xmm10
 ; SSSE3-NEXT:    pandn %xmm2, %xmm0
 ; SSSE3-NEXT:    por %xmm10, %xmm0
-; SSSE3-NEXT:    packuswb %xmm12, %xmm0
+; SSSE3-NEXT:    packuswb %xmm11, %xmm0
 ; SSSE3-NEXT:    movdqa %xmm9, %xmm10
-; SSSE3-NEXT:    pxor %xmm4, %xmm10
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm10[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm11
-; SSSE3-NEXT:    movdqa %xmm5, %xmm12
-; SSSE3-NEXT:    pcmpgtd %xmm10, %xmm12
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm12[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm11, %xmm10
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm12[1,1,3,3]
-; SSSE3-NEXT:    por %xmm10, %xmm11
-; SSSE3-NEXT:    pand %xmm11, %xmm9
-; SSSE3-NEXT:    pandn %xmm2, %xmm11
-; SSSE3-NEXT:    por %xmm9, %xmm11
-; SSSE3-NEXT:    movdqa %xmm8, %xmm9
-; SSSE3-NEXT:    pxor %xmm4, %xmm9
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm9[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm10
+; SSSE3-NEXT:    pxor %xmm3, %xmm10
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm10[0,0,2,2]
 ; SSSE3-NEXT:    movdqa %xmm5, %xmm12
-; SSSE3-NEXT:    pcmpgtd %xmm9, %xmm12
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm9 = xmm12[0,0,2,2]
+; SSSE3-NEXT:    pcmpgtd %xmm11, %xmm12
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm10
+; SSSE3-NEXT:    pand %xmm12, %xmm10
 ; SSSE3-NEXT:    pand %xmm10, %xmm9
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm12[1,1,3,3]
-; SSSE3-NEXT:    por %xmm9, %xmm10
-; SSSE3-NEXT:    pand %xmm10, %xmm8
 ; SSSE3-NEXT:    pandn %xmm2, %xmm10
-; SSSE3-NEXT:    por %xmm8, %xmm10
-; SSSE3-NEXT:    packuswb %xmm11, %xmm10
-; SSSE3-NEXT:    packuswb %xmm10, %xmm0
-; SSSE3-NEXT:    movdqa %xmm7, %xmm8
-; SSSE3-NEXT:    pxor %xmm4, %xmm8
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm9 = xmm8[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm9
-; SSSE3-NEXT:    movdqa %xmm5, %xmm10
-; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm10
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm10[0,0,2,2]
+; SSSE3-NEXT:    por %xmm9, %xmm10
+; SSSE3-NEXT:    movdqa %xmm8, %xmm9
+; SSSE3-NEXT:    pxor %xmm3, %xmm9
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm9[0,0,2,2]
+; SSSE3-NEXT:    movdqa %xmm5, %xmm12
+; SSSE3-NEXT:    pcmpgtd %xmm11, %xmm12
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm9
+; SSSE3-NEXT:    pand %xmm12, %xmm9
 ; SSSE3-NEXT:    pand %xmm9, %xmm8
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm9 = xmm10[1,1,3,3]
-; SSSE3-NEXT:    por %xmm8, %xmm9
-; SSSE3-NEXT:    pand %xmm9, %xmm7
 ; SSSE3-NEXT:    pandn %xmm2, %xmm9
-; SSSE3-NEXT:    por %xmm7, %xmm9
+; SSSE3-NEXT:    por %xmm8, %xmm9
+; SSSE3-NEXT:    packuswb %xmm10, %xmm9
+; SSSE3-NEXT:    packuswb %xmm9, %xmm0
+; SSSE3-NEXT:    movdqa %xmm7, %xmm8
+; SSSE3-NEXT:    pxor %xmm3, %xmm8
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm9 = xmm8[0,0,2,2]
+; SSSE3-NEXT:    movdqa %xmm5, %xmm10
+; SSSE3-NEXT:    pcmpgtd %xmm9, %xmm10
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm8
+; SSSE3-NEXT:    pand %xmm10, %xmm8
+; SSSE3-NEXT:    pand %xmm8, %xmm7
+; SSSE3-NEXT:    pandn %xmm2, %xmm8
+; SSSE3-NEXT:    por %xmm7, %xmm8
 ; SSSE3-NEXT:    movdqa %xmm6, %xmm7
-; SSSE3-NEXT:    pxor %xmm4, %xmm7
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm8
+; SSSE3-NEXT:    pxor %xmm3, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm9 = xmm7[0,0,2,2]
 ; SSSE3-NEXT:    movdqa %xmm5, %xmm10
-; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm10
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm10[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm8, %xmm11
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm10[1,1,3,3]
-; SSSE3-NEXT:    por %xmm11, %xmm7
+; SSSE3-NEXT:    pcmpgtd %xmm9, %xmm10
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm7
+; SSSE3-NEXT:    pand %xmm10, %xmm7
 ; SSSE3-NEXT:    pand %xmm7, %xmm6
 ; SSSE3-NEXT:    pandn %xmm2, %xmm7
 ; SSSE3-NEXT:    por %xmm6, %xmm7
-; SSSE3-NEXT:    packuswb %xmm9, %xmm7
-; SSSE3-NEXT:    movdqa %xmm3, %xmm6
-; SSSE3-NEXT:    pxor %xmm4, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm8
+; SSSE3-NEXT:    packuswb %xmm8, %xmm7
+; SSSE3-NEXT:    movdqa %xmm4, %xmm6
+; SSSE3-NEXT:    pxor %xmm3, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2]
 ; SSSE3-NEXT:    movdqa %xmm5, %xmm9
-; SSSE3-NEXT:    pcmpgtd %xmm6, %xmm9
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm9[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm8, %xmm6
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm9[1,1,3,3]
-; SSSE3-NEXT:    por %xmm6, %xmm8
-; SSSE3-NEXT:    pand %xmm8, %xmm3
-; SSSE3-NEXT:    pandn %xmm2, %xmm8
-; SSSE3-NEXT:    por %xmm3, %xmm8
-; SSSE3-NEXT:    movdqa %xmm1, %xmm3
-; SSSE3-NEXT:    pxor %xmm4, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[1,1,3,3]
-; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm6
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm5
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[0,0,2,2]
-; SSSE3-NEXT:    pand %xmm6, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
-; SSSE3-NEXT:    por %xmm3, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm9
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm6
+; SSSE3-NEXT:    pand %xmm9, %xmm6
+; SSSE3-NEXT:    pand %xmm6, %xmm4
+; SSSE3-NEXT:    pandn %xmm2, %xmm6
+; SSSE3-NEXT:    por %xmm4, %xmm6
+; SSSE3-NEXT:    movdqa %xmm1, %xmm4
+; SSSE3-NEXT:    pxor %xmm3, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm4
+; SSSE3-NEXT:    pand %xmm5, %xmm4
 ; SSSE3-NEXT:    pand %xmm4, %xmm1
 ; SSSE3-NEXT:    pandn %xmm2, %xmm4
 ; SSSE3-NEXT:    por %xmm1, %xmm4
-; SSSE3-NEXT:    packuswb %xmm8, %xmm4
+; SSSE3-NEXT:    packuswb %xmm6, %xmm4
 ; SSSE3-NEXT:    packuswb %xmm4, %xmm7
 ; SSSE3-NEXT:    packuswb %xmm7, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_usat_v16i64_v16i8:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa 96(%rdi), %xmm2
-; SSE41-NEXT:    movdqa 112(%rdi), %xmm4
-; SSE41-NEXT:    movdqa 64(%rdi), %xmm7
-; SSE41-NEXT:    movdqa 80(%rdi), %xmm8
-; SSE41-NEXT:    movdqa (%rdi), %xmm11
-; SSE41-NEXT:    movdqa 16(%rdi), %xmm1
-; SSE41-NEXT:    movdqa 32(%rdi), %xmm9
-; SSE41-NEXT:    movdqa 48(%rdi), %xmm10
-; SSE41-NEXT:    movapd {{.*#+}} xmm3 = [255,255]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movdqa 96(%rdi), %xmm3
+; SSE41-NEXT:    movdqa 112(%rdi), %xmm5
+; SSE41-NEXT:    movdqa 64(%rdi), %xmm8
+; SSE41-NEXT:    movdqa 80(%rdi), %xmm9
+; SSE41-NEXT:    movdqa (%rdi), %xmm12
+; SSE41-NEXT:    movdqa 16(%rdi), %xmm2
+; SSE41-NEXT:    movdqa 32(%rdi), %xmm10
+; SSE41-NEXT:    movdqa 48(%rdi), %xmm11
+; SSE41-NEXT:    movapd {{.*#+}} xmm4 = [255,255]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm7 = [9223372039002259456,9223372039002259456]
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm7, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = [9223372039002259711,9223372039002259711]
+; SSE41-NEXT:    movdqa %xmm6, %xmm13
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm13
+; SSE41-NEXT:    pshufd {{.*#+}} xmm14 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [2147483903,2147483903,2147483903,2147483903]
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
-; SSE41-NEXT:    pxor %xmm6, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [9223372039002259711,9223372039002259711]
-; SSE41-NEXT:    movdqa %xmm5, %xmm12
+; SSE41-NEXT:    pcmpgtd %xmm14, %xmm0
+; SSE41-NEXT:    pand %xmm13, %xmm0
+; SSE41-NEXT:    movapd %xmm4, %xmm13
+; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm13
+; SSE41-NEXT:    movdqa %xmm12, %xmm0
+; SSE41-NEXT:    pxor %xmm7, %xmm0
+; SSE41-NEXT:    movdqa %xmm6, %xmm2
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm14 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm14, %xmm0
+; SSE41-NEXT:    pand %xmm2, %xmm0
+; SSE41-NEXT:    movapd %xmm4, %xmm2
+; SSE41-NEXT:    blendvpd %xmm0, %xmm12, %xmm2
+; SSE41-NEXT:    packusdw %xmm13, %xmm2
+; SSE41-NEXT:    movdqa %xmm11, %xmm0
+; SSE41-NEXT:    pxor %xmm7, %xmm0
+; SSE41-NEXT:    movdqa %xmm6, %xmm12
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm12
-; SSE41-NEXT:    movdqa %xmm5, %xmm13
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm13
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm13[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm13, %xmm0
 ; SSE41-NEXT:    pand %xmm12, %xmm0
-; SSE41-NEXT:    por %xmm13, %xmm0
-; SSE41-NEXT:    movapd %xmm3, %xmm12
-; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm12
-; SSE41-NEXT:    movdqa %xmm11, %xmm0
-; SSE41-NEXT:    pxor %xmm6, %xmm0
-; SSE41-NEXT:    movdqa %xmm5, %xmm1
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    movdqa %xmm5, %xmm13
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm13
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm13[0,0,2,2]
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    por %xmm13, %xmm0
-; SSE41-NEXT:    movapd %xmm3, %xmm1
-; SSE41-NEXT:    blendvpd %xmm0, %xmm11, %xmm1
-; SSE41-NEXT:    packusdw %xmm12, %xmm1
+; SSE41-NEXT:    movapd %xmm4, %xmm12
+; SSE41-NEXT:    blendvpd %xmm0, %xmm11, %xmm12
 ; SSE41-NEXT:    movdqa %xmm10, %xmm0
-; SSE41-NEXT:    pxor %xmm6, %xmm0
-; SSE41-NEXT:    movdqa %xmm5, %xmm11
+; SSE41-NEXT:    pxor %xmm7, %xmm0
+; SSE41-NEXT:    movdqa %xmm6, %xmm11
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm11
-; SSE41-NEXT:    movdqa %xmm5, %xmm12
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm12
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm12[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm13, %xmm0
 ; SSE41-NEXT:    pand %xmm11, %xmm0
-; SSE41-NEXT:    por %xmm12, %xmm0
-; SSE41-NEXT:    movapd %xmm3, %xmm11
+; SSE41-NEXT:    movapd %xmm4, %xmm11
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm10, %xmm11
+; SSE41-NEXT:    packusdw %xmm12, %xmm11
+; SSE41-NEXT:    packusdw %xmm11, %xmm2
 ; SSE41-NEXT:    movdqa %xmm9, %xmm0
-; SSE41-NEXT:    pxor %xmm6, %xmm0
-; SSE41-NEXT:    movdqa %xmm5, %xmm10
+; SSE41-NEXT:    pxor %xmm7, %xmm0
+; SSE41-NEXT:    movdqa %xmm6, %xmm10
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm10
-; SSE41-NEXT:    movdqa %xmm5, %xmm12
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm12
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm12[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm11, %xmm0
 ; SSE41-NEXT:    pand %xmm10, %xmm0
-; SSE41-NEXT:    por %xmm12, %xmm0
-; SSE41-NEXT:    movapd %xmm3, %xmm10
+; SSE41-NEXT:    movapd %xmm4, %xmm10
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm9, %xmm10
-; SSE41-NEXT:    packusdw %xmm11, %xmm10
-; SSE41-NEXT:    packusdw %xmm10, %xmm1
 ; SSE41-NEXT:    movdqa %xmm8, %xmm0
-; SSE41-NEXT:    pxor %xmm6, %xmm0
-; SSE41-NEXT:    movdqa %xmm5, %xmm9
+; SSE41-NEXT:    pxor %xmm7, %xmm0
+; SSE41-NEXT:    movdqa %xmm6, %xmm9
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm9
-; SSE41-NEXT:    movdqa %xmm5, %xmm10
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm10
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm10[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm11, %xmm0
 ; SSE41-NEXT:    pand %xmm9, %xmm0
-; SSE41-NEXT:    por %xmm10, %xmm0
-; SSE41-NEXT:    movapd %xmm3, %xmm9
+; SSE41-NEXT:    movapd %xmm4, %xmm9
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm8, %xmm9
-; SSE41-NEXT:    movdqa %xmm7, %xmm0
-; SSE41-NEXT:    pxor %xmm6, %xmm0
-; SSE41-NEXT:    movdqa %xmm5, %xmm8
+; SSE41-NEXT:    packusdw %xmm10, %xmm9
+; SSE41-NEXT:    movdqa %xmm5, %xmm0
+; SSE41-NEXT:    pxor %xmm7, %xmm0
+; SSE41-NEXT:    movdqa %xmm6, %xmm8
 ; SSE41-NEXT:    pcmpeqd %xmm0, %xmm8
-; SSE41-NEXT:    movdqa %xmm5, %xmm10
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm10
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm10[0,0,2,2]
+; SSE41-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm10, %xmm0
 ; SSE41-NEXT:    pand %xmm8, %xmm0
-; SSE41-NEXT:    por %xmm10, %xmm0
-; SSE41-NEXT:    movapd %xmm3, %xmm8
-; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm8
-; SSE41-NEXT:    packusdw %xmm9, %xmm8
-; SSE41-NEXT:    movdqa %xmm4, %xmm0
-; SSE41-NEXT:    pxor %xmm6, %xmm0
-; SSE41-NEXT:    movdqa %xmm5, %xmm7
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm7
-; SSE41-NEXT:    movdqa %xmm5, %xmm9
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm9
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,0,2,2]
-; SSE41-NEXT:    pand %xmm7, %xmm0
-; SSE41-NEXT:    por %xmm9, %xmm0
-; SSE41-NEXT:    movapd %xmm3, %xmm7
-; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm7
-; SSE41-NEXT:    pxor %xmm2, %xmm6
-; SSE41-NEXT:    movdqa %xmm5, %xmm4
-; SSE41-NEXT:    pcmpeqd %xmm6, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm6, %xmm5
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
-; SSE41-NEXT:    pand %xmm4, %xmm0
-; SSE41-NEXT:    por %xmm5, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm3
-; SSE41-NEXT:    packusdw %xmm7, %xmm3
-; SSE41-NEXT:    packusdw %xmm3, %xmm8
-; SSE41-NEXT:    packuswb %xmm8, %xmm1
+; SSE41-NEXT:    movapd %xmm4, %xmm8
+; SSE41-NEXT:    blendvpd %xmm0, %xmm5, %xmm8
+; SSE41-NEXT:    pxor %xmm3, %xmm7
+; SSE41-NEXT:    pcmpeqd %xmm7, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm1
+; SSE41-NEXT:    pand %xmm6, %xmm1
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm4
+; SSE41-NEXT:    packusdw %xmm8, %xmm4
+; SSE41-NEXT:    packusdw %xmm4, %xmm9
+; SSE41-NEXT:    packuswb %xmm9, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc_usat_v16i64_v16i8:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa (%rdi), %xmm0
 ; AVX1-NEXT:    vmovdqa 16(%rdi), %xmm1
 ; AVX1-NEXT:    vmovdqa 32(%rdi), %xmm2
 ; AVX1-NEXT:    vmovdqa 48(%rdi), %xmm3