diff --git a/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 370e0d5..26c4955 100644
--- a/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -5385,6 +5385,11 @@
   EVT VT = LHS.getValueType();
   SDLoc dl(Node);
 
+  assert(LHS.getValueType().isInteger() && "Expected operands to be integers");
+  assert(RHS.getValueType().isInteger() && "Expected operands to be integers");
+  assert(LHS.getValueType() == RHS.getValueType() &&
+         "Expected both operands to be the same type");
+
   // usub.sat(a, b) -> umax(a, b) - b
   if (Opcode == ISD::USUBSAT && isOperationLegalOrCustom(ISD::UMAX, VT)) {
     SDValue Max = DAG.getNode(ISD::UMAX, dl, VT, LHS, RHS);
@@ -5397,11 +5402,6 @@
     return DAG.getNode(ISD::ADD, dl, VT, Min, RHS);
   }
 
-  if (VT.isVector()) {
-    // TODO: Consider not scalarizing here.
-    return SDValue();
-  }
-
   unsigned OverflowOp;
   switch (Opcode) {
   case ISD::SADDSAT:
@@ -5421,16 +5421,7 @@
                      "addition or subtraction node.");
   }
 
-  assert(LHS.getValueType().isScalarInteger() &&
-         "Expected operands to be integers. Vector of int arguments should "
-         "already be unrolled.");
-  assert(RHS.getValueType().isScalarInteger() &&
-         "Expected operands to be integers. Vector of int arguments should "
-         "already be unrolled.");
-  assert(LHS.getValueType() == RHS.getValueType() &&
-         "Expected both operands to be the same type");
-
-  unsigned BitWidth = LHS.getValueSizeInBits();
+  unsigned BitWidth = LHS.getScalarValueSizeInBits();
   EVT ResultType = LHS.getValueType();
   EVT BoolVT =
       getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), ResultType);
diff --git a/test/CodeGen/X86/sadd_sat.ll b/test/CodeGen/X86/sadd_sat.ll
index 3f91b66..b5524e1 100644
--- a/test/CodeGen/X86/sadd_sat.ll
+++ b/test/CodeGen/X86/sadd_sat.ll
@@ -185,55 +185,28 @@
 ;
 ; X64-LABEL: vec:
 ; X64:       # %bb.0:
-; X64-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,2,3]
-; X64-NEXT:    movd %xmm2, %ecx
-; X64-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
-; X64-NEXT:    movd %xmm2, %r8d
-; X64-NEXT:    xorl %edx, %edx
-; X64-NEXT:    movl %r8d, %esi
-; X64-NEXT:    addl %ecx, %esi
-; X64-NEXT:    setns %dl
-; X64-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; X64-NEXT:    addl %ecx, %r8d
-; X64-NEXT:    cmovol %edx, %r8d
-; X64-NEXT:    movd %xmm1, %edx
-; X64-NEXT:    movd %xmm0, %ecx
-; X64-NEXT:    xorl %esi, %esi
-; X64-NEXT:    movl %ecx, %edi
-; X64-NEXT:    addl %edx, %edi
-; X64-NEXT:    setns %sil
-; X64-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; X64-NEXT:    addl %edx, %ecx
-; X64-NEXT:    cmovol %esi, %ecx
-; X64-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; X64-NEXT:    movd %xmm2, %edx
-; X64-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; X64-NEXT:    movd %xmm2, %eax
-; X64-NEXT:    xorl %edi, %edi
-; X64-NEXT:    movl %eax, %esi
-; X64-NEXT:    addl %edx, %esi
-; X64-NEXT:    setns %dil
-; X64-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; X64-NEXT:    addl %edx, %eax
-; X64-NEXT:    cmovol %edi, %eax
-; X64-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; X64-NEXT:    movd %xmm1, %r9d
-; X64-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; X64-NEXT:    movd %xmm0, %edx
-; X64-NEXT:    xorl %edi, %edi
-; X64-NEXT:    movl %edx, %esi
-; X64-NEXT:    addl %r9d, %esi
-; X64-NEXT:    setns %dil
-; X64-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; X64-NEXT:    addl %r9d, %edx
-; X64-NEXT:    cmovol %edi, %edx
-; X64-NEXT:    movd %edx, %xmm0
-; X64-NEXT:    movd %eax, %xmm1
-; X64-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; X64-NEXT:    movd %ecx, %xmm0
-; X64-NEXT:    movd %r8d, %xmm2
-; X64-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; X64-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; X64-NEXT:    pxor %xmm3, %xmm3
+; X64-NEXT:    pxor %xmm4, %xmm4
+; X64-NEXT:    pcmpgtd %xmm1, %xmm4
+; X64-NEXT:    pcmpeqd %xmm2, %xmm2
+; X64-NEXT:    pxor %xmm2, %xmm4
+; X64-NEXT:    pxor %xmm5, %xmm5
+; X64-NEXT:    pcmpgtd %xmm0, %xmm5
+; X64-NEXT:    pxor %xmm2, %xmm5
+; X64-NEXT:    pcmpeqd %xmm5, %xmm4
+; X64-NEXT:    paddd %xmm1, %xmm0
+; X64-NEXT:    pcmpgtd %xmm0, %xmm3
+; X64-NEXT:    pxor %xmm3, %xmm2
+; X64-NEXT:    pcmpeqd %xmm5, %xmm2
+; X64-NEXT:    pandn %xmm4, %xmm2
+; X64-NEXT:    movdqa %xmm3, %xmm1
+; X64-NEXT:    pandn {{.*}}(%rip), %xmm1
+; X64-NEXT:    psrld $1, %xmm3
+; X64-NEXT:    por %xmm1, %xmm3
+; X64-NEXT:    pand %xmm2, %xmm3
+; X64-NEXT:    pandn %xmm0, %xmm2
+; X64-NEXT:    por %xmm3, %xmm2
+; X64-NEXT:    movdqa %xmm2, %xmm0
 ; X64-NEXT:    retq
   %tmp = call <4 x i32> @llvm.sadd.sat.v4i32(<4 x i32> %x, <4 x i32> %y);
   ret <4 x i32> %tmp;
diff --git a/test/CodeGen/X86/sadd_sat_vec.ll b/test/CodeGen/X86/sadd_sat_vec.ll
index 8fe56b8..320cc07 100644
--- a/test/CodeGen/X86/sadd_sat_vec.ll
+++ b/test/CodeGen/X86/sadd_sat_vec.ll
@@ -661,128 +661,222 @@
 ; SSE2-LABEL: v2i32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    psllq $32, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movq %xmm2, %rax
+; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
 ; SSE2-NEXT:    psllq $32, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movq %xmm2, %rcx
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movq %rcx, %rsi
-; SSE2-NEXT:    addq %rax, %rsi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE2-NEXT:    addq %r8, %rdx
-; SSE2-NEXT:    addq %rax, %rcx
-; SSE2-NEXT:    cmovoq %rdx, %rcx
-; SSE2-NEXT:    movq %xmm1, %rax
-; SSE2-NEXT:    movq %xmm0, %rsi
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %rsi, %rdx
-; SSE2-NEXT:    addq %rax, %rdx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    addq %rax, %rsi
-; SSE2-NEXT:    cmovoq %rdi, %rsi
-; SSE2-NEXT:    movq %rsi, %xmm1
-; SSE2-NEXT:    movq %rcx, %xmm0
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,3,2,3]
-; SSE2-NEXT:    psrad $31, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,3,2,3]
+; SSE2-NEXT:    movdqa %xmm0, %xmm3
+; SSE2-NEXT:    paddq %xmm1, %xmm0
+; SSE2-NEXT:    pxor %xmm2, %xmm1
+; SSE2-NEXT:    movdqa %xmm2, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm1, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE2-NEXT:    pxor %xmm1, %xmm4
+; SSE2-NEXT:    pxor %xmm2, %xmm3
+; SSE2-NEXT:    movdqa %xmm2, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2-NEXT:    pand %xmm6, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE2-NEXT:    por %xmm3, %xmm5
+; SSE2-NEXT:    pxor %xmm1, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm3
+; SSE2-NEXT:    movdqa %xmm0, %xmm4
+; SSE2-NEXT:    pxor %xmm2, %xmm4
+; SSE2-NEXT:    movdqa %xmm2, %xmm6
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pand %xmm7, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE2-NEXT:    por %xmm4, %xmm6
+; SSE2-NEXT:    pxor %xmm1, %xmm6
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,0,3,2]
+; SSE2-NEXT:    pand %xmm6, %xmm1
+; SSE2-NEXT:    pandn %xmm3, %xmm1
+; SSE2-NEXT:    movdqa %xmm1, %xmm3
+; SSE2-NEXT:    pandn %xmm0, %xmm3
+; SSE2-NEXT:    por %xmm2, %xmm0
+; SSE2-NEXT:    movdqa %xmm2, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm0, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm0
+; SSE2-NEXT:    pandn {{.*}}(%rip), %xmm0
+; SSE2-NEXT:    pand {{.*}}(%rip), %xmm2
+; SSE2-NEXT:    por %xmm0, %xmm2
+; SSE2-NEXT:    pand %xmm1, %xmm2
+; SSE2-NEXT:    por %xmm3, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,3,2,3]
+; SSE2-NEXT:    psrad $31, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,3,2,3]
 ; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v2i32:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    psllq $32, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm2, %rax
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
 ; SSSE3-NEXT:    psllq $32, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm2, %rcx
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movq %rcx, %rsi
-; SSSE3-NEXT:    addq %rax, %rsi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSSE3-NEXT:    addq %r8, %rdx
-; SSSE3-NEXT:    addq %rax, %rcx
-; SSSE3-NEXT:    cmovoq %rdx, %rcx
-; SSSE3-NEXT:    movq %xmm1, %rax
-; SSSE3-NEXT:    movq %xmm0, %rsi
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %rsi, %rdx
-; SSSE3-NEXT:    addq %rax, %rdx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    addq %rax, %rsi
-; SSSE3-NEXT:    cmovoq %rdi, %rsi
-; SSSE3-NEXT:    movq %rsi, %xmm1
-; SSSE3-NEXT:    movq %rcx, %xmm0
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,3,2,3]
-; SSSE3-NEXT:    psrad $31, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,3,2,3]
+; SSSE3-NEXT:    movdqa %xmm0, %xmm3
+; SSSE3-NEXT:    paddq %xmm1, %xmm0
+; SSSE3-NEXT:    pxor %xmm2, %xmm1
+; SSSE3-NEXT:    movdqa %xmm2, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm5, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm1, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSSE3-NEXT:    pxor %xmm1, %xmm4
+; SSSE3-NEXT:    pxor %xmm2, %xmm3
+; SSSE3-NEXT:    movdqa %xmm2, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm6, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    por %xmm3, %xmm5
+; SSSE3-NEXT:    pxor %xmm1, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm3
+; SSSE3-NEXT:    movdqa %xmm0, %xmm4
+; SSSE3-NEXT:    pxor %xmm2, %xmm4
+; SSSE3-NEXT:    movdqa %xmm2, %xmm6
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm7, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    por %xmm4, %xmm6
+; SSSE3-NEXT:    pxor %xmm1, %xmm6
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm6, %xmm1
+; SSSE3-NEXT:    pandn %xmm3, %xmm1
+; SSSE3-NEXT:    movdqa %xmm1, %xmm3
+; SSSE3-NEXT:    pandn %xmm0, %xmm3
+; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    movdqa %xmm2, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm5, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm0, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSSE3-NEXT:    pandn {{.*}}(%rip), %xmm0
+; SSSE3-NEXT:    pand {{.*}}(%rip), %xmm2
+; SSSE3-NEXT:    por %xmm0, %xmm2
+; SSSE3-NEXT:    pand %xmm1, %xmm2
+; SSSE3-NEXT:    por %xmm3, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,3,2,3]
+; SSSE3-NEXT:    psrad $31, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,3,2,3]
 ; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v2i32:
 ; SSE41:       # %bb.0:
+; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    psllq $32, %xmm1
-; SSE41-NEXT:    movq %xmm1, %rax
-; SSE41-NEXT:    psllq $32, %xmm0
-; SSE41-NEXT:    movq %xmm0, %rcx
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movq %rcx, %rsi
-; SSE41-NEXT:    addq %rax, %rsi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE41-NEXT:    addq %r8, %rdx
-; SSE41-NEXT:    addq %rax, %rcx
-; SSE41-NEXT:    cmovoq %rdx, %rcx
-; SSE41-NEXT:    pextrq $1, %xmm1, %rax
-; SSE41-NEXT:    pextrq $1, %xmm0, %rsi
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %rsi, %rdx
-; SSE41-NEXT:    addq %rax, %rdx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    addq %rax, %rsi
-; SSE41-NEXT:    cmovoq %rdi, %rsi
-; SSE41-NEXT:    movq %rsi, %xmm1
-; SSE41-NEXT:    movq %rcx, %xmm0
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE41-NEXT:    psrad $31, %xmm0
-; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5],xmm0[6,7]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
+; SSE41-NEXT:    psllq $32, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm3
+; SSE41-NEXT:    paddq %xmm1, %xmm2
+; SSE41-NEXT:    pxor %xmm0, %xmm1
+; SSE41-NEXT:    movdqa %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm1
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm1, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm8, %xmm8
+; SSE41-NEXT:    pxor %xmm8, %xmm4
+; SSE41-NEXT:    pxor %xmm0, %xmm3
+; SSE41-NEXT:    movdqa %xmm0, %xmm1
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm1
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE41-NEXT:    pand %xmm6, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[1,1,3,3]
+; SSE41-NEXT:    por %xmm3, %xmm6
+; SSE41-NEXT:    pxor %xmm8, %xmm6
+; SSE41-NEXT:    pcmpeqq %xmm6, %xmm4
+; SSE41-NEXT:    movdqa %xmm2, %xmm1
+; SSE41-NEXT:    pxor %xmm0, %xmm1
+; SSE41-NEXT:    movdqa %xmm0, %xmm3
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[1,1,3,3]
+; SSE41-NEXT:    pand %xmm7, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
+; SSE41-NEXT:    por %xmm5, %xmm1
+; SSE41-NEXT:    pxor %xmm8, %xmm1
+; SSE41-NEXT:    pcmpeqq %xmm6, %xmm1
+; SSE41-NEXT:    pandn %xmm4, %xmm1
+; SSE41-NEXT:    movdqa %xmm2, %xmm3
+; SSE41-NEXT:    por %xmm0, %xmm3
+; SSE41-NEXT:    movdqa %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm0
+; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    movapd {{.*#+}} xmm3 = [9223372036854775808,9223372036854775808]
+; SSE41-NEXT:    blendvpd %xmm0, {{.*}}(%rip), %xmm3
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
+; SSE41-NEXT:    psrad $31, %xmm2
+; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: v2i32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpsllq $32, %xmm1, %xmm1
-; AVX1-NEXT:    vmovq %xmm1, %rax
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
 ; AVX1-NEXT:    vpsllq $32, %xmm0, %xmm0
-; AVX1-NEXT:    vmovq %xmm0, %rcx
-; AVX1-NEXT:    xorl %edx, %edx
-; AVX1-NEXT:    movq %rcx, %rsi
-; AVX1-NEXT:    addq %rax, %rsi
-; AVX1-NEXT:    setns %dl
-; AVX1-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX1-NEXT:    addq %r8, %rdx
-; AVX1-NEXT:    addq %rax, %rcx
-; AVX1-NEXT:    cmovoq %rdx, %rcx
-; AVX1-NEXT:    vpextrq $1, %xmm1, %rax
-; AVX1-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movq %rsi, %rdx
-; AVX1-NEXT:    addq %rax, %rdx
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addq %r8, %rdi
-; AVX1-NEXT:    addq %rax, %rsi
-; AVX1-NEXT:    cmovoq %rdi, %rsi
-; AVX1-NEXT:    vmovq %rsi, %xmm0
-; AVX1-NEXT:    vmovq %rcx, %xmm1
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm5, %xmm3
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm1
+; AVX1-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqq %xmm1, %xmm5, %xmm1
+; AVX1-NEXT:    vpandn %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vmovapd {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
+; AVX1-NEXT:    vblendvpd %xmm0, {{.*}}(%rip), %xmm2, %xmm2
+; AVX1-NEXT:    vblendvpd %xmm1, %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpsrad $31, %xmm0, %xmm1
 ; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
@@ -791,29 +885,22 @@
 ; AVX2-LABEL: v2i32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpsllq $32, %xmm1, %xmm1
-; AVX2-NEXT:    vmovq %xmm1, %rax
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm3
+; AVX2-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX2-NEXT:    vpxor %xmm4, %xmm3, %xmm3
 ; AVX2-NEXT:    vpsllq $32, %xmm0, %xmm0
-; AVX2-NEXT:    vmovq %xmm0, %rcx
-; AVX2-NEXT:    xorl %edx, %edx
-; AVX2-NEXT:    movq %rcx, %rsi
-; AVX2-NEXT:    addq %rax, %rsi
-; AVX2-NEXT:    setns %dl
-; AVX2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX2-NEXT:    addq %r8, %rdx
-; AVX2-NEXT:    addq %rax, %rcx
-; AVX2-NEXT:    cmovoq %rdx, %rcx
-; AVX2-NEXT:    vpextrq $1, %xmm1, %rax
-; AVX2-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movq %rsi, %rdx
-; AVX2-NEXT:    addq %rax, %rdx
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addq %r8, %rdi
-; AVX2-NEXT:    addq %rax, %rsi
-; AVX2-NEXT:    cmovoq %rdi, %rsi
-; AVX2-NEXT:    vmovq %rsi, %xmm0
-; AVX2-NEXT:    vmovq %rcx, %xmm1
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX2-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm5
+; AVX2-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX2-NEXT:    vpcmpeqq %xmm3, %xmm5, %xmm3
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm1
+; AVX2-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqq %xmm1, %xmm5, %xmm1
+; AVX2-NEXT:    vpandn %xmm3, %xmm1, %xmm1
+; AVX2-NEXT:    vmovapd {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
+; AVX2-NEXT:    vblendvpd %xmm0, {{.*}}(%rip), %xmm2, %xmm2
+; AVX2-NEXT:    vblendvpd %xmm1, %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpsrad $31, %xmm0, %xmm1
 ; AVX2-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
 ; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3]
@@ -822,29 +909,19 @@
 ; AVX512-LABEL: v2i32:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vpsllq $32, %xmm1, %xmm1
-; AVX512-NEXT:    vmovq %xmm1, %rax
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltq %xmm2, %xmm1, %k0
 ; AVX512-NEXT:    vpsllq $32, %xmm0, %xmm0
-; AVX512-NEXT:    vmovq %xmm0, %rcx
-; AVX512-NEXT:    xorl %edx, %edx
-; AVX512-NEXT:    movq %rcx, %rsi
-; AVX512-NEXT:    addq %rax, %rsi
-; AVX512-NEXT:    setns %dl
-; AVX512-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX512-NEXT:    addq %r8, %rdx
-; AVX512-NEXT:    addq %rax, %rcx
-; AVX512-NEXT:    cmovoq %rdx, %rcx
-; AVX512-NEXT:    vpextrq $1, %xmm1, %rax
-; AVX512-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movq %rsi, %rdx
-; AVX512-NEXT:    addq %rax, %rdx
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addq %r8, %rdi
-; AVX512-NEXT:    addq %rax, %rsi
-; AVX512-NEXT:    cmovoq %rdi, %rsi
-; AVX512-NEXT:    vmovq %rsi, %xmm0
-; AVX512-NEXT:    vmovq %rcx, %xmm1
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX512-NEXT:    vpcmpnltq %xmm2, %xmm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpaddq %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vpcmpnltq %xmm2, %xmm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandnw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtq %xmm0, %xmm2, %k2
+; AVX512-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
+; AVX512-NEXT:    vmovdqa64 {{.*}}(%rip), %xmm1 {%k2}
+; AVX512-NEXT:    vmovdqa64 %xmm1, %xmm0 {%k1}
 ; AVX512-NEXT:    vpsraq $32, %xmm0, %xmm0
 ; AVX512-NEXT:    retq
   %z = call <2 x i32> @llvm.sadd.sat.v2i32(<2 x i32> %x, <2 x i32> %y)
@@ -854,197 +931,135 @@
 define <4 x i32> @v4i32(<4 x i32> %x, <4 x i32> %y) nounwind {
 ; SSE2-LABEL: v4i32:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,2,3]
-; SSE2-NEXT:    movd %xmm2, %ecx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
-; SSE2-NEXT:    movd %xmm2, %r8d
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movl %r8d, %esi
-; SSE2-NEXT:    addl %ecx, %esi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %ecx, %r8d
-; SSE2-NEXT:    cmovol %edx, %r8d
-; SSE2-NEXT:    movd %xmm1, %edx
-; SSE2-NEXT:    movd %xmm0, %ecx
-; SSE2-NEXT:    xorl %esi, %esi
-; SSE2-NEXT:    movl %ecx, %edi
-; SSE2-NEXT:    addl %edx, %edi
-; SSE2-NEXT:    setns %sil
-; SSE2-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %edx, %ecx
-; SSE2-NEXT:    cmovol %esi, %ecx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movd %xmm2, %edx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movd %xmm2, %eax
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %eax, %esi
-; SSE2-NEXT:    addl %edx, %esi
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %edx, %eax
-; SSE2-NEXT:    cmovol %edi, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r9d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSE2-NEXT:    movd %xmm0, %edx
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %edx, %esi
-; SSE2-NEXT:    addl %r9d, %esi
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %r9d, %edx
-; SSE2-NEXT:    cmovol %edi, %edx
-; SSE2-NEXT:    movd %edx, %xmm0
-; SSE2-NEXT:    movd %eax, %xmm1
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE2-NEXT:    movd %ecx, %xmm0
-; SSE2-NEXT:    movd %r8d, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE2-NEXT:    pxor %xmm3, %xmm3
+; SSE2-NEXT:    pxor %xmm4, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm2
+; SSE2-NEXT:    pxor %xmm2, %xmm4
+; SSE2-NEXT:    pxor %xmm5, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm5
+; SSE2-NEXT:    pxor %xmm2, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSE2-NEXT:    paddd %xmm1, %xmm0
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE2-NEXT:    pxor %xmm3, %xmm2
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm2
+; SSE2-NEXT:    pandn %xmm4, %xmm2
+; SSE2-NEXT:    movdqa %xmm3, %xmm1
+; SSE2-NEXT:    pandn {{.*}}(%rip), %xmm1
+; SSE2-NEXT:    psrld $1, %xmm3
+; SSE2-NEXT:    por %xmm1, %xmm3
+; SSE2-NEXT:    pand %xmm2, %xmm3
+; SSE2-NEXT:    pandn %xmm0, %xmm2
+; SSE2-NEXT:    por %xmm3, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v4i32:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm2, %ecx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm2, %r8d
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movl %r8d, %esi
-; SSSE3-NEXT:    addl %ecx, %esi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %ecx, %r8d
-; SSSE3-NEXT:    cmovol %edx, %r8d
-; SSSE3-NEXT:    movd %xmm1, %edx
-; SSSE3-NEXT:    movd %xmm0, %ecx
-; SSSE3-NEXT:    xorl %esi, %esi
-; SSSE3-NEXT:    movl %ecx, %edi
-; SSSE3-NEXT:    addl %edx, %edi
-; SSSE3-NEXT:    setns %sil
-; SSSE3-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %edx, %ecx
-; SSSE3-NEXT:    cmovol %esi, %ecx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm2, %edx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm2, %eax
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %eax, %esi
-; SSSE3-NEXT:    addl %edx, %esi
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %edx, %eax
-; SSSE3-NEXT:    cmovol %edi, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r9d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm0, %edx
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %edx, %esi
-; SSSE3-NEXT:    addl %r9d, %esi
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %r9d, %edx
-; SSSE3-NEXT:    cmovol %edi, %edx
-; SSSE3-NEXT:    movd %edx, %xmm0
-; SSSE3-NEXT:    movd %eax, %xmm1
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSSE3-NEXT:    movd %ecx, %xmm0
-; SSSE3-NEXT:    movd %r8d, %xmm2
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSSE3-NEXT:    pxor %xmm3, %xmm3
+; SSSE3-NEXT:    pxor %xmm4, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm2
+; SSSE3-NEXT:    pxor %xmm2, %xmm4
+; SSSE3-NEXT:    pxor %xmm5, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm5
+; SSSE3-NEXT:    pxor %xmm2, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSSE3-NEXT:    paddd %xmm1, %xmm0
+; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSSE3-NEXT:    pxor %xmm3, %xmm2
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm2
+; SSSE3-NEXT:    pandn %xmm4, %xmm2
+; SSSE3-NEXT:    movdqa %xmm3, %xmm1
+; SSSE3-NEXT:    pandn {{.*}}(%rip), %xmm1
+; SSSE3-NEXT:    psrld $1, %xmm3
+; SSSE3-NEXT:    por %xmm1, %xmm3
+; SSSE3-NEXT:    pand %xmm2, %xmm3
+; SSSE3-NEXT:    pandn %xmm0, %xmm2
+; SSSE3-NEXT:    por %xmm3, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v4i32:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    pextrd $3, %xmm1, %ecx
-; SSE41-NEXT:    pextrd $3, %xmm0, %r8d
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movl %r8d, %esi
-; SSE41-NEXT:    addl %ecx, %esi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %ecx, %r8d
-; SSE41-NEXT:    cmovol %edx, %r8d
-; SSE41-NEXT:    pextrd $2, %xmm1, %edx
-; SSE41-NEXT:    pextrd $2, %xmm0, %ecx
-; SSE41-NEXT:    xorl %esi, %esi
-; SSE41-NEXT:    movl %ecx, %edi
-; SSE41-NEXT:    addl %edx, %edi
-; SSE41-NEXT:    setns %sil
-; SSE41-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %edx, %ecx
-; SSE41-NEXT:    cmovol %esi, %ecx
-; SSE41-NEXT:    movd %xmm1, %edx
-; SSE41-NEXT:    movd %xmm0, %eax
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %eax, %esi
-; SSE41-NEXT:    addl %edx, %esi
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %edx, %eax
-; SSE41-NEXT:    cmovol %edi, %eax
-; SSE41-NEXT:    pextrd $1, %xmm1, %r9d
-; SSE41-NEXT:    pextrd $1, %xmm0, %edx
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %edx, %esi
-; SSE41-NEXT:    addl %r9d, %esi
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %r9d, %edx
-; SSE41-NEXT:    cmovol %edi, %edx
-; SSE41-NEXT:    movd %eax, %xmm0
-; SSE41-NEXT:    pinsrd $1, %edx, %xmm0
-; SSE41-NEXT:    pinsrd $2, %ecx, %xmm0
-; SSE41-NEXT:    pinsrd $3, %r8d, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm2
+; SSE41-NEXT:    pxor %xmm3, %xmm3
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm4, %xmm4
+; SSE41-NEXT:    pxor %xmm4, %xmm0
+; SSE41-NEXT:    pxor %xmm5, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSE41-NEXT:    pxor %xmm4, %xmm5
+; SSE41-NEXT:    pcmpeqd %xmm5, %xmm0
+; SSE41-NEXT:    paddd %xmm1, %xmm2
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm3
+; SSE41-NEXT:    pxor %xmm4, %xmm3
+; SSE41-NEXT:    pcmpeqd %xmm5, %xmm3
+; SSE41-NEXT:    pandn %xmm0, %xmm3
+; SSE41-NEXT:    movaps {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, {{.*}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm1, %xmm2
+; SSE41-NEXT:    movaps %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX-LABEL: v4i32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpextrd $3, %xmm1, %ecx
-; AVX-NEXT:    vpextrd $3, %xmm0, %r9d
-; AVX-NEXT:    xorl %edx, %edx
-; AVX-NEXT:    movl %r9d, %esi
-; AVX-NEXT:    addl %ecx, %esi
-; AVX-NEXT:    setns %dl
-; AVX-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; AVX-NEXT:    addl %ecx, %r9d
-; AVX-NEXT:    cmovol %edx, %r9d
-; AVX-NEXT:    vpextrd $2, %xmm1, %edx
-; AVX-NEXT:    vpextrd $2, %xmm0, %ecx
-; AVX-NEXT:    xorl %esi, %esi
-; AVX-NEXT:    movl %ecx, %edi
-; AVX-NEXT:    addl %edx, %edi
-; AVX-NEXT:    setns %sil
-; AVX-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX-NEXT:    addl %edx, %ecx
-; AVX-NEXT:    cmovol %esi, %ecx
-; AVX-NEXT:    vmovd %xmm1, %r8d
-; AVX-NEXT:    vmovd %xmm0, %edx
-; AVX-NEXT:    xorl %edi, %edi
-; AVX-NEXT:    movl %edx, %esi
-; AVX-NEXT:    addl %r8d, %esi
-; AVX-NEXT:    setns %dil
-; AVX-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX-NEXT:    addl %r8d, %edx
-; AVX-NEXT:    cmovol %edi, %edx
-; AVX-NEXT:    vpextrd $1, %xmm1, %r8d
-; AVX-NEXT:    vpextrd $1, %xmm0, %eax
-; AVX-NEXT:    xorl %esi, %esi
-; AVX-NEXT:    movl %eax, %edi
-; AVX-NEXT:    addl %r8d, %edi
-; AVX-NEXT:    setns %sil
-; AVX-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX-NEXT:    addl %r8d, %eax
-; AVX-NEXT:    cmovol %esi, %eax
-; AVX-NEXT:    vmovd %edx, %xmm0
-; AVX-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
-; AVX-NEXT:    vpinsrd $2, %ecx, %xmm0, %xmm0
-; AVX-NEXT:    vpinsrd $3, %r9d, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; AVX1-LABEL: v4i32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm2, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm2, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm5, %xmm3
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm2, %xmm1
+; AVX1-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm5, %xmm1
+; AVX1-NEXT:    vpandn %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vmovaps {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
+; AVX1-NEXT:    vblendvps %xmm0, {{.*}}(%rip), %xmm2, %xmm2
+; AVX1-NEXT:    vblendvps %xmm1, %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: v4i32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpgtd %xmm1, %xmm2, %xmm3
+; AVX2-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX2-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX2-NEXT:    vpcmpgtd %xmm0, %xmm2, %xmm5
+; AVX2-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX2-NEXT:    vpcmpeqd %xmm3, %xmm5, %xmm3
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpgtd %xmm0, %xmm2, %xmm1
+; AVX2-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm5, %xmm1
+; AVX2-NEXT:    vpandn %xmm3, %xmm1, %xmm1
+; AVX2-NEXT:    vbroadcastss {{.*#+}} xmm2 = [2147483647,2147483647,2147483647,2147483647]
+; AVX2-NEXT:    vbroadcastss {{.*#+}} xmm3 = [2147483648,2147483648,2147483648,2147483648]
+; AVX2-NEXT:    vblendvps %xmm0, %xmm2, %xmm3, %xmm2
+; AVX2-NEXT:    vblendvps %xmm1, %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    retq
+;
+; AVX512-LABEL: v4i32:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltd %xmm2, %xmm1, %k0
+; AVX512-NEXT:    vpcmpnltd %xmm2, %xmm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vpcmpnltd %xmm2, %xmm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandnw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtd %xmm0, %xmm2, %k2
+; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
+; AVX512-NEXT:    vpbroadcastd {{.*}}(%rip), %xmm1 {%k2}
+; AVX512-NEXT:    vmovdqa32 %xmm1, %xmm0 {%k1}
+; AVX512-NEXT:    retq
   %z = call <4 x i32> @llvm.sadd.sat.v4i32(<4 x i32> %x, <4 x i32> %y)
   ret <4 x i32> %z
 }
@@ -1052,575 +1067,216 @@
 define <8 x i32> @v8i32(<8 x i32> %x, <8 x i32> %y) nounwind {
 ; SSE2-LABEL: v8i32:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pushq %rbp
-; SSE2-NEXT:    pushq %rbx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,2,3]
-; SSE2-NEXT:    movd %xmm4, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[1,1,2,3]
-; SSE2-NEXT:    movd %xmm4, %r8d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r8d, %edx
-; SSE2-NEXT:    addl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %r8d
-; SSE2-NEXT:    cmovol %ecx, %r8d
-; SSE2-NEXT:    movd %xmm3, %eax
-; SSE2-NEXT:    movd %xmm1, %r9d
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movl %r9d, %esi
-; SSE2-NEXT:    addl %eax, %esi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %r9d
-; SSE2-NEXT:    cmovol %edx, %r9d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[2,3,0,1]
-; SSE2-NEXT:    movd %xmm4, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movd %xmm4, %r10d
-; SSE2-NEXT:    xorl %esi, %esi
-; SSE2-NEXT:    movl %r10d, %edi
-; SSE2-NEXT:    addl %eax, %edi
-; SSE2-NEXT:    setns %sil
-; SSE2-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %r10d
-; SSE2-NEXT:    cmovol %esi, %r10d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[3,1,2,3]
-; SSE2-NEXT:    movd %xmm3, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r11d
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %r11d, %esi
-; SSE2-NEXT:    addl %eax, %esi
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %r11d
-; SSE2-NEXT:    cmovol %edi, %r11d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %edx
-; SSE2-NEXT:    xorl %esi, %esi
-; SSE2-NEXT:    movl %edx, %edi
-; SSE2-NEXT:    addl %eax, %edi
-; SSE2-NEXT:    setns %sil
-; SSE2-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %edx
-; SSE2-NEXT:    cmovol %esi, %edx
-; SSE2-NEXT:    movd %xmm2, %esi
-; SSE2-NEXT:    movd %xmm0, %ecx
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %ecx, %eax
-; SSE2-NEXT:    addl %esi, %eax
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %esi, %ecx
-; SSE2-NEXT:    cmovol %edi, %ecx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %esi
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %esi, %ebx
-; SSE2-NEXT:    addl %eax, %ebx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %esi
-; SSE2-NEXT:    cmovol %edi, %esi
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSE2-NEXT:    movd %xmm0, %edi
-; SSE2-NEXT:    xorl %ebx, %ebx
-; SSE2-NEXT:    movl %edi, %ebp
-; SSE2-NEXT:    addl %eax, %ebp
-; SSE2-NEXT:    setns %bl
-; SSE2-NEXT:    addl $2147483647, %ebx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %edi
-; SSE2-NEXT:    cmovol %ebx, %edi
-; SSE2-NEXT:    movd %edi, %xmm0
-; SSE2-NEXT:    movd %esi, %xmm1
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE2-NEXT:    movd %ecx, %xmm0
-; SSE2-NEXT:    movd %edx, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE2-NEXT:    movd %r11d, %xmm1
-; SSE2-NEXT:    movd %r10d, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
-; SSE2-NEXT:    movd %r9d, %xmm1
-; SSE2-NEXT:    movd %r8d, %xmm3
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; SSE2-NEXT:    popq %rbx
-; SSE2-NEXT:    popq %rbp
+; SSE2-NEXT:    movdqa %xmm0, %xmm8
+; SSE2-NEXT:    pxor %xmm6, %xmm6
+; SSE2-NEXT:    pxor %xmm7, %xmm7
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm7
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm5
+; SSE2-NEXT:    pxor %xmm5, %xmm7
+; SSE2-NEXT:    pxor %xmm4, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSE2-NEXT:    pxor %xmm5, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm7
+; SSE2-NEXT:    paddd %xmm2, %xmm8
+; SSE2-NEXT:    pxor %xmm2, %xmm2
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm0
+; SSE2-NEXT:    pxor %xmm5, %xmm0
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm0
+; SSE2-NEXT:    pandn %xmm7, %xmm0
+; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,2147483648,2147483648,2147483648]
+; SSE2-NEXT:    movdqa %xmm2, %xmm7
+; SSE2-NEXT:    pandn %xmm4, %xmm7
+; SSE2-NEXT:    psrld $1, %xmm2
+; SSE2-NEXT:    por %xmm7, %xmm2
+; SSE2-NEXT:    pand %xmm0, %xmm2
+; SSE2-NEXT:    pandn %xmm8, %xmm0
+; SSE2-NEXT:    por %xmm2, %xmm0
+; SSE2-NEXT:    pxor %xmm2, %xmm2
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm2
+; SSE2-NEXT:    pxor %xmm5, %xmm2
+; SSE2-NEXT:    pxor %xmm7, %xmm7
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm7
+; SSE2-NEXT:    pxor %xmm5, %xmm7
+; SSE2-NEXT:    pcmpeqd %xmm7, %xmm2
+; SSE2-NEXT:    paddd %xmm3, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm6
+; SSE2-NEXT:    pxor %xmm6, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm7, %xmm5
+; SSE2-NEXT:    pandn %xmm2, %xmm5
+; SSE2-NEXT:    movdqa %xmm6, %xmm2
+; SSE2-NEXT:    pandn %xmm4, %xmm2
+; SSE2-NEXT:    psrld $1, %xmm6
+; SSE2-NEXT:    por %xmm2, %xmm6
+; SSE2-NEXT:    pand %xmm5, %xmm6
+; SSE2-NEXT:    pandn %xmm1, %xmm5
+; SSE2-NEXT:    por %xmm6, %xmm5
+; SSE2-NEXT:    movdqa %xmm5, %xmm1
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v8i32:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pushq %rbp
-; SSSE3-NEXT:    pushq %rbx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm4, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm4, %r8d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r8d, %edx
-; SSSE3-NEXT:    addl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %r8d
-; SSSE3-NEXT:    cmovol %ecx, %r8d
-; SSSE3-NEXT:    movd %xmm3, %eax
-; SSSE3-NEXT:    movd %xmm1, %r9d
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movl %r9d, %esi
-; SSSE3-NEXT:    addl %eax, %esi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %r9d
-; SSSE3-NEXT:    cmovol %edx, %r9d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm4, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm4, %r10d
-; SSSE3-NEXT:    xorl %esi, %esi
-; SSSE3-NEXT:    movl %r10d, %edi
-; SSSE3-NEXT:    addl %eax, %edi
-; SSSE3-NEXT:    setns %sil
-; SSSE3-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %r10d
-; SSSE3-NEXT:    cmovol %esi, %r10d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm3, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r11d
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %r11d, %esi
-; SSSE3-NEXT:    addl %eax, %esi
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %r11d
-; SSSE3-NEXT:    cmovol %edi, %r11d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %edx
-; SSSE3-NEXT:    xorl %esi, %esi
-; SSSE3-NEXT:    movl %edx, %edi
-; SSSE3-NEXT:    addl %eax, %edi
-; SSSE3-NEXT:    setns %sil
-; SSSE3-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %edx
-; SSSE3-NEXT:    cmovol %esi, %edx
-; SSSE3-NEXT:    movd %xmm2, %esi
-; SSSE3-NEXT:    movd %xmm0, %ecx
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %ecx, %eax
-; SSSE3-NEXT:    addl %esi, %eax
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %esi, %ecx
-; SSSE3-NEXT:    cmovol %edi, %ecx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %esi
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %esi, %ebx
-; SSSE3-NEXT:    addl %eax, %ebx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %esi
-; SSSE3-NEXT:    cmovol %edi, %esi
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm0, %edi
-; SSSE3-NEXT:    xorl %ebx, %ebx
-; SSSE3-NEXT:    movl %edi, %ebp
-; SSSE3-NEXT:    addl %eax, %ebp
-; SSSE3-NEXT:    setns %bl
-; SSSE3-NEXT:    addl $2147483647, %ebx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %edi
-; SSSE3-NEXT:    cmovol %ebx, %edi
-; SSSE3-NEXT:    movd %edi, %xmm0
-; SSSE3-NEXT:    movd %esi, %xmm1
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSSE3-NEXT:    movd %ecx, %xmm0
-; SSSE3-NEXT:    movd %edx, %xmm2
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSSE3-NEXT:    movd %r11d, %xmm1
-; SSSE3-NEXT:    movd %r10d, %xmm2
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
-; SSSE3-NEXT:    movd %r9d, %xmm1
-; SSSE3-NEXT:    movd %r8d, %xmm3
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; SSSE3-NEXT:    popq %rbx
-; SSSE3-NEXT:    popq %rbp
+; SSSE3-NEXT:    movdqa %xmm0, %xmm8
+; SSSE3-NEXT:    pxor %xmm6, %xmm6
+; SSSE3-NEXT:    pxor %xmm7, %xmm7
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm7
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm5
+; SSSE3-NEXT:    pxor %xmm5, %xmm7
+; SSSE3-NEXT:    pxor %xmm4, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSSE3-NEXT:    pxor %xmm5, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm7
+; SSSE3-NEXT:    paddd %xmm2, %xmm8
+; SSSE3-NEXT:    pxor %xmm2, %xmm2
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSSE3-NEXT:    pxor %xmm5, %xmm0
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm0
+; SSSE3-NEXT:    pandn %xmm7, %xmm0
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,2147483648,2147483648,2147483648]
+; SSSE3-NEXT:    movdqa %xmm2, %xmm7
+; SSSE3-NEXT:    pandn %xmm4, %xmm7
+; SSSE3-NEXT:    psrld $1, %xmm2
+; SSSE3-NEXT:    por %xmm7, %xmm2
+; SSSE3-NEXT:    pand %xmm0, %xmm2
+; SSSE3-NEXT:    pandn %xmm8, %xmm0
+; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    pxor %xmm2, %xmm2
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm2
+; SSSE3-NEXT:    pxor %xmm5, %xmm2
+; SSSE3-NEXT:    pxor %xmm7, %xmm7
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm7
+; SSSE3-NEXT:    pxor %xmm5, %xmm7
+; SSSE3-NEXT:    pcmpeqd %xmm7, %xmm2
+; SSSE3-NEXT:    paddd %xmm3, %xmm1
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm6
+; SSSE3-NEXT:    pxor %xmm6, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm7, %xmm5
+; SSSE3-NEXT:    pandn %xmm2, %xmm5
+; SSSE3-NEXT:    movdqa %xmm6, %xmm2
+; SSSE3-NEXT:    pandn %xmm4, %xmm2
+; SSSE3-NEXT:    psrld $1, %xmm6
+; SSSE3-NEXT:    por %xmm2, %xmm6
+; SSSE3-NEXT:    pand %xmm5, %xmm6
+; SSSE3-NEXT:    pandn %xmm1, %xmm5
+; SSSE3-NEXT:    por %xmm6, %xmm5
+; SSSE3-NEXT:    movdqa %xmm5, %xmm1
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v8i32:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    pushq %rbp
-; SSE41-NEXT:    pushq %rbx
-; SSE41-NEXT:    pextrd $3, %xmm3, %eax
-; SSE41-NEXT:    pextrd $3, %xmm1, %r8d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r8d, %edx
-; SSE41-NEXT:    addl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %r8d
-; SSE41-NEXT:    cmovol %ecx, %r8d
-; SSE41-NEXT:    pextrd $2, %xmm3, %eax
-; SSE41-NEXT:    pextrd $2, %xmm1, %r9d
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movl %r9d, %esi
-; SSE41-NEXT:    addl %eax, %esi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %r9d
-; SSE41-NEXT:    cmovol %edx, %r9d
-; SSE41-NEXT:    movd %xmm3, %eax
-; SSE41-NEXT:    movd %xmm1, %r10d
-; SSE41-NEXT:    xorl %esi, %esi
-; SSE41-NEXT:    movl %r10d, %edi
-; SSE41-NEXT:    addl %eax, %edi
-; SSE41-NEXT:    setns %sil
-; SSE41-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %r10d
-; SSE41-NEXT:    cmovol %esi, %r10d
-; SSE41-NEXT:    pextrd $1, %xmm3, %eax
-; SSE41-NEXT:    pextrd $1, %xmm1, %r11d
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %r11d, %esi
-; SSE41-NEXT:    addl %eax, %esi
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %r11d
-; SSE41-NEXT:    cmovol %edi, %r11d
-; SSE41-NEXT:    pextrd $3, %xmm2, %eax
-; SSE41-NEXT:    pextrd $3, %xmm0, %edx
-; SSE41-NEXT:    xorl %esi, %esi
-; SSE41-NEXT:    movl %edx, %edi
-; SSE41-NEXT:    addl %eax, %edi
-; SSE41-NEXT:    setns %sil
-; SSE41-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %edx
-; SSE41-NEXT:    cmovol %esi, %edx
-; SSE41-NEXT:    pextrd $2, %xmm2, %esi
-; SSE41-NEXT:    pextrd $2, %xmm0, %ecx
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %ecx, %eax
-; SSE41-NEXT:    addl %esi, %eax
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %esi, %ecx
-; SSE41-NEXT:    cmovol %edi, %ecx
-; SSE41-NEXT:    movd %xmm2, %eax
-; SSE41-NEXT:    movd %xmm0, %esi
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %esi, %ebx
-; SSE41-NEXT:    addl %eax, %ebx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %esi
-; SSE41-NEXT:    cmovol %edi, %esi
-; SSE41-NEXT:    pextrd $1, %xmm2, %eax
-; SSE41-NEXT:    pextrd $1, %xmm0, %edi
-; SSE41-NEXT:    xorl %ebx, %ebx
-; SSE41-NEXT:    movl %edi, %ebp
-; SSE41-NEXT:    addl %eax, %ebp
-; SSE41-NEXT:    setns %bl
-; SSE41-NEXT:    addl $2147483647, %ebx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %edi
-; SSE41-NEXT:    cmovol %ebx, %edi
-; SSE41-NEXT:    movd %esi, %xmm0
-; SSE41-NEXT:    pinsrd $1, %edi, %xmm0
-; SSE41-NEXT:    pinsrd $2, %ecx, %xmm0
-; SSE41-NEXT:    pinsrd $3, %edx, %xmm0
-; SSE41-NEXT:    movd %r10d, %xmm1
-; SSE41-NEXT:    pinsrd $1, %r11d, %xmm1
-; SSE41-NEXT:    pinsrd $2, %r9d, %xmm1
-; SSE41-NEXT:    pinsrd $3, %r8d, %xmm1
-; SSE41-NEXT:    popq %rbx
-; SSE41-NEXT:    popq %rbp
+; SSE41-NEXT:    movdqa %xmm0, %xmm5
+; SSE41-NEXT:    pxor %xmm4, %xmm4
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm9
+; SSE41-NEXT:    pxor %xmm9, %xmm0
+; SSE41-NEXT:    pxor %xmm7, %xmm7
+; SSE41-NEXT:    pcmpgtd %xmm5, %xmm7
+; SSE41-NEXT:    pxor %xmm9, %xmm7
+; SSE41-NEXT:    pcmpeqd %xmm7, %xmm0
+; SSE41-NEXT:    paddd %xmm2, %xmm5
+; SSE41-NEXT:    pxor %xmm2, %xmm2
+; SSE41-NEXT:    pcmpgtd %xmm5, %xmm2
+; SSE41-NEXT:    pxor %xmm9, %xmm2
+; SSE41-NEXT:    pcmpeqd %xmm7, %xmm2
+; SSE41-NEXT:    pandn %xmm0, %xmm2
+; SSE41-NEXT:    movaps {{.*#+}} xmm8 = [2147483647,2147483647,2147483647,2147483647]
+; SSE41-NEXT:    movaps {{.*#+}} xmm6 = [2147483648,2147483648,2147483648,2147483648]
+; SSE41-NEXT:    movaps %xmm6, %xmm7
+; SSE41-NEXT:    movdqa %xmm5, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm8, %xmm7
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm7, %xmm5
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
+; SSE41-NEXT:    pxor %xmm9, %xmm0
+; SSE41-NEXT:    pxor %xmm2, %xmm2
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm2
+; SSE41-NEXT:    pxor %xmm9, %xmm2
+; SSE41-NEXT:    pcmpeqd %xmm2, %xmm0
+; SSE41-NEXT:    paddd %xmm3, %xmm1
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE41-NEXT:    pxor %xmm9, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm2, %xmm4
+; SSE41-NEXT:    pandn %xmm0, %xmm4
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm8, %xmm6
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm6, %xmm1
+; SSE41-NEXT:    movaps %xmm5, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: v8i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    pushq %rbp
-; AVX1-NEXT:    pushq %rbx
-; AVX1-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX1-NEXT:    vpextrd $3, %xmm0, %r8d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r8d, %edx
-; AVX1-NEXT:    addl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %r8d
-; AVX1-NEXT:    cmovol %ecx, %r8d
-; AVX1-NEXT:    vpextrd $2, %xmm1, %eax
-; AVX1-NEXT:    vpextrd $2, %xmm0, %r9d
-; AVX1-NEXT:    xorl %edx, %edx
-; AVX1-NEXT:    movl %r9d, %esi
-; AVX1-NEXT:    addl %eax, %esi
-; AVX1-NEXT:    setns %dl
-; AVX1-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %r9d
-; AVX1-NEXT:    cmovol %edx, %r9d
-; AVX1-NEXT:    vmovd %xmm1, %eax
-; AVX1-NEXT:    vmovd %xmm0, %r10d
-; AVX1-NEXT:    xorl %esi, %esi
-; AVX1-NEXT:    movl %r10d, %edi
-; AVX1-NEXT:    addl %eax, %edi
-; AVX1-NEXT:    setns %sil
-; AVX1-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %r10d
-; AVX1-NEXT:    cmovol %esi, %r10d
-; AVX1-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX1-NEXT:    vpextrd $1, %xmm0, %r11d
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %r11d, %esi
-; AVX1-NEXT:    addl %eax, %esi
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %r11d
-; AVX1-NEXT:    cmovol %edi, %r11d
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpextrd $3, %xmm0, %edx
-; AVX1-NEXT:    xorl %esi, %esi
-; AVX1-NEXT:    movl %edx, %edi
-; AVX1-NEXT:    addl %eax, %edi
-; AVX1-NEXT:    setns %sil
-; AVX1-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %edx
-; AVX1-NEXT:    cmovol %esi, %edx
-; AVX1-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX1-NEXT:    vpextrd $2, %xmm0, %ecx
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %ecx, %eax
-; AVX1-NEXT:    addl %esi, %eax
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %esi, %ecx
-; AVX1-NEXT:    cmovol %edi, %ecx
-; AVX1-NEXT:    vmovd %xmm1, %eax
-; AVX1-NEXT:    vmovd %xmm0, %esi
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %esi, %ebx
-; AVX1-NEXT:    addl %eax, %ebx
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %esi
-; AVX1-NEXT:    cmovol %edi, %esi
-; AVX1-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX1-NEXT:    vpextrd $1, %xmm0, %edi
-; AVX1-NEXT:    xorl %ebx, %ebx
-; AVX1-NEXT:    movl %edi, %ebp
-; AVX1-NEXT:    addl %eax, %ebp
-; AVX1-NEXT:    setns %bl
-; AVX1-NEXT:    addl $2147483647, %ebx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %edi
-; AVX1-NEXT:    cmovol %ebx, %edi
-; AVX1-NEXT:    vmovd %esi, %xmm0
-; AVX1-NEXT:    vpinsrd $1, %edi, %xmm0, %xmm0
-; AVX1-NEXT:    vpinsrd $2, %ecx, %xmm0, %xmm0
-; AVX1-NEXT:    vpinsrd $3, %edx, %xmm0, %xmm0
-; AVX1-NEXT:    vmovd %r10d, %xmm1
-; AVX1-NEXT:    vpinsrd $1, %r11d, %xmm1, %xmm1
-; AVX1-NEXT:    vpinsrd $2, %r9d, %xmm1, %xmm1
-; AVX1-NEXT:    vpinsrd $3, %r8d, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT:    popq %rbx
-; AVX1-NEXT:    popq %rbp
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpgtd %xmm2, %xmm3, %xmm4
+; AVX1-NEXT:    vpcmpeqd %xmm5, %xmm5, %xmm5
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm4
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm6
+; AVX1-NEXT:    vpcmpgtd %xmm6, %xmm3, %xmm7
+; AVX1-NEXT:    vpxor %xmm5, %xmm7, %xmm7
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm7, %xmm8
+; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm3, %xmm4
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm9
+; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm3, %xmm4
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm4
+; AVX1-NEXT:    vpcmpeqd %xmm9, %xmm4, %xmm9
+; AVX1-NEXT:    vinsertf128 $1, %xmm8, %ymm9, %ymm8
+; AVX1-NEXT:    vpaddd %xmm2, %xmm6, %xmm9
+; AVX1-NEXT:    vpcmpgtd %xmm9, %xmm3, %xmm6
+; AVX1-NEXT:    vpxor %xmm5, %xmm6, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm7, %xmm2
+; AVX1-NEXT:    vpxor %xmm5, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm3, %xmm1
+; AVX1-NEXT:    vpxor %xmm5, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm4, %xmm1
+; AVX1-NEXT:    vpxor %xmm5, %xmm1, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
+; AVX1-NEXT:    vandps %ymm1, %ymm8, %ymm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm2
+; AVX1-NEXT:    vmovaps {{.*#+}} ymm3 = [2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648]
+; AVX1-NEXT:    vblendvps %ymm2, {{.*}}(%rip), %ymm3, %ymm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm9, %ymm0, %ymm0
+; AVX1-NEXT:    vblendvps %ymm1, %ymm2, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: v8i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    pushq %rbp
-; AVX2-NEXT:    pushq %rbx
-; AVX2-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX2-NEXT:    vpextrd $3, %xmm0, %r8d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r8d, %edx
-; AVX2-NEXT:    addl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %r8d
-; AVX2-NEXT:    cmovol %ecx, %r8d
-; AVX2-NEXT:    vpextrd $2, %xmm1, %eax
-; AVX2-NEXT:    vpextrd $2, %xmm0, %r9d
-; AVX2-NEXT:    xorl %edx, %edx
-; AVX2-NEXT:    movl %r9d, %esi
-; AVX2-NEXT:    addl %eax, %esi
-; AVX2-NEXT:    setns %dl
-; AVX2-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %r9d
-; AVX2-NEXT:    cmovol %edx, %r9d
-; AVX2-NEXT:    vmovd %xmm1, %eax
-; AVX2-NEXT:    vmovd %xmm0, %r10d
-; AVX2-NEXT:    xorl %esi, %esi
-; AVX2-NEXT:    movl %r10d, %edi
-; AVX2-NEXT:    addl %eax, %edi
-; AVX2-NEXT:    setns %sil
-; AVX2-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %r10d
-; AVX2-NEXT:    cmovol %esi, %r10d
-; AVX2-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX2-NEXT:    vpextrd $1, %xmm0, %r11d
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %r11d, %esi
-; AVX2-NEXT:    addl %eax, %esi
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %r11d
-; AVX2-NEXT:    cmovol %edi, %r11d
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; AVX2-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; AVX2-NEXT:    vpextrd $3, %xmm0, %edx
-; AVX2-NEXT:    xorl %esi, %esi
-; AVX2-NEXT:    movl %edx, %edi
-; AVX2-NEXT:    addl %eax, %edi
-; AVX2-NEXT:    setns %sil
-; AVX2-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %edx
-; AVX2-NEXT:    cmovol %esi, %edx
-; AVX2-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX2-NEXT:    vpextrd $2, %xmm0, %ecx
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %ecx, %eax
-; AVX2-NEXT:    addl %esi, %eax
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %esi, %ecx
-; AVX2-NEXT:    cmovol %edi, %ecx
-; AVX2-NEXT:    vmovd %xmm1, %eax
-; AVX2-NEXT:    vmovd %xmm0, %esi
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %esi, %ebx
-; AVX2-NEXT:    addl %eax, %ebx
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %esi
-; AVX2-NEXT:    cmovol %edi, %esi
-; AVX2-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX2-NEXT:    vpextrd $1, %xmm0, %edi
-; AVX2-NEXT:    xorl %ebx, %ebx
-; AVX2-NEXT:    movl %edi, %ebp
-; AVX2-NEXT:    addl %eax, %ebp
-; AVX2-NEXT:    setns %bl
-; AVX2-NEXT:    addl $2147483647, %ebx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %edi
-; AVX2-NEXT:    cmovol %ebx, %edi
-; AVX2-NEXT:    vmovd %esi, %xmm0
-; AVX2-NEXT:    vpinsrd $1, %edi, %xmm0, %xmm0
-; AVX2-NEXT:    vpinsrd $2, %ecx, %xmm0, %xmm0
-; AVX2-NEXT:    vpinsrd $3, %edx, %xmm0, %xmm0
-; AVX2-NEXT:    vmovd %r10d, %xmm1
-; AVX2-NEXT:    vpinsrd $1, %r11d, %xmm1, %xmm1
-; AVX2-NEXT:    vpinsrd $2, %r9d, %xmm1, %xmm1
-; AVX2-NEXT:    vpinsrd $3, %r8d, %xmm1, %xmm1
-; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX2-NEXT:    popq %rbx
-; AVX2-NEXT:    popq %rbp
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpgtd %ymm1, %ymm2, %ymm3
+; AVX2-NEXT:    vpcmpeqd %ymm4, %ymm4, %ymm4
+; AVX2-NEXT:    vpxor %ymm4, %ymm3, %ymm3
+; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm2, %ymm5
+; AVX2-NEXT:    vpxor %ymm4, %ymm5, %ymm5
+; AVX2-NEXT:    vpcmpeqd %ymm3, %ymm5, %ymm3
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm2, %ymm1
+; AVX2-NEXT:    vpxor %ymm4, %ymm1, %ymm1
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm5, %ymm1
+; AVX2-NEXT:    vpandn %ymm3, %ymm1, %ymm1
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [2147483647,2147483647,2147483647,2147483647,2147483647,2147483647,2147483647,2147483647]
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm3 = [2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648]
+; AVX2-NEXT:    vblendvps %ymm0, %ymm2, %ymm3, %ymm2
+; AVX2-NEXT:    vblendvps %ymm1, %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: v8i32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    pushq %rbp
-; AVX512-NEXT:    pushq %rbx
-; AVX512-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $3, %xmm0, %r8d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r8d, %edx
-; AVX512-NEXT:    addl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %r8d
-; AVX512-NEXT:    cmovol %ecx, %r8d
-; AVX512-NEXT:    vpextrd $2, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $2, %xmm0, %r9d
-; AVX512-NEXT:    xorl %edx, %edx
-; AVX512-NEXT:    movl %r9d, %esi
-; AVX512-NEXT:    addl %eax, %esi
-; AVX512-NEXT:    setns %dl
-; AVX512-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %r9d
-; AVX512-NEXT:    cmovol %edx, %r9d
-; AVX512-NEXT:    vmovd %xmm1, %eax
-; AVX512-NEXT:    vmovd %xmm0, %r10d
-; AVX512-NEXT:    xorl %esi, %esi
-; AVX512-NEXT:    movl %r10d, %edi
-; AVX512-NEXT:    addl %eax, %edi
-; AVX512-NEXT:    setns %sil
-; AVX512-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %r10d
-; AVX512-NEXT:    cmovol %esi, %r10d
-; AVX512-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $1, %xmm0, %r11d
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %r11d, %esi
-; AVX512-NEXT:    addl %eax, %esi
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %r11d
-; AVX512-NEXT:    cmovol %edi, %r11d
-; AVX512-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; AVX512-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; AVX512-NEXT:    vpextrd $3, %xmm0, %edx
-; AVX512-NEXT:    xorl %esi, %esi
-; AVX512-NEXT:    movl %edx, %edi
-; AVX512-NEXT:    addl %eax, %edi
-; AVX512-NEXT:    setns %sil
-; AVX512-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %edx
-; AVX512-NEXT:    cmovol %esi, %edx
-; AVX512-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX512-NEXT:    vpextrd $2, %xmm0, %ecx
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %ecx, %eax
-; AVX512-NEXT:    addl %esi, %eax
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %esi, %ecx
-; AVX512-NEXT:    cmovol %edi, %ecx
-; AVX512-NEXT:    vmovd %xmm1, %eax
-; AVX512-NEXT:    vmovd %xmm0, %esi
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %esi, %ebx
-; AVX512-NEXT:    addl %eax, %ebx
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %esi
-; AVX512-NEXT:    cmovol %edi, %esi
-; AVX512-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $1, %xmm0, %edi
-; AVX512-NEXT:    xorl %ebx, %ebx
-; AVX512-NEXT:    movl %edi, %ebp
-; AVX512-NEXT:    addl %eax, %ebp
-; AVX512-NEXT:    setns %bl
-; AVX512-NEXT:    addl $2147483647, %ebx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %edi
-; AVX512-NEXT:    cmovol %ebx, %edi
-; AVX512-NEXT:    vmovd %esi, %xmm0
-; AVX512-NEXT:    vpinsrd $1, %edi, %xmm0, %xmm0
-; AVX512-NEXT:    vpinsrd $2, %ecx, %xmm0, %xmm0
-; AVX512-NEXT:    vpinsrd $3, %edx, %xmm0, %xmm0
-; AVX512-NEXT:    vmovd %r10d, %xmm1
-; AVX512-NEXT:    vpinsrd $1, %r11d, %xmm1, %xmm1
-; AVX512-NEXT:    vpinsrd $2, %r9d, %xmm1, %xmm1
-; AVX512-NEXT:    vpinsrd $3, %r8d, %xmm1, %xmm1
-; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX512-NEXT:    popq %rbx
-; AVX512-NEXT:    popq %rbp
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltd %ymm2, %ymm1, %k0
+; AVX512-NEXT:    vpcmpnltd %ymm2, %ymm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpaddd %ymm1, %ymm0, %ymm0
+; AVX512-NEXT:    vpcmpnltd %ymm2, %ymm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandnw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtd %ymm0, %ymm2, %k2
+; AVX512-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648]
+; AVX512-NEXT:    vpbroadcastd {{.*}}(%rip), %ymm1 {%k2}
+; AVX512-NEXT:    vmovdqa32 %ymm1, %ymm0 {%k1}
 ; AVX512-NEXT:    retq
   %z = call <8 x i32> @llvm.sadd.sat.v8i32(<8 x i32> %x, <8 x i32> %y)
   ret <8 x i32> %z
@@ -1629,1187 +1285,382 @@
 define <16 x i32> @v16i32(<16 x i32> %x, <16 x i32> %y) nounwind {
 ; SSE2-LABEL: v16i32:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pushq %rbp
-; SSE2-NEXT:    pushq %r15
-; SSE2-NEXT:    pushq %r14
-; SSE2-NEXT:    pushq %r13
-; SSE2-NEXT:    pushq %r12
-; SSE2-NEXT:    pushq %rbx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[1,1,2,3]
-; SSE2-NEXT:    movd %xmm8, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[1,1,2,3]
-; SSE2-NEXT:    movd %xmm8, %esi
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %esi, %edx
-; SSE2-NEXT:    addl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %esi
-; SSE2-NEXT:    cmovol %ecx, %esi
-; SSE2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE2-NEXT:    movd %xmm5, %eax
-; SSE2-NEXT:    movd %xmm1, %esi
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %esi, %edx
-; SSE2-NEXT:    addl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %esi
-; SSE2-NEXT:    cmovol %ecx, %esi
-; SSE2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[2,3,0,1]
-; SSE2-NEXT:    movd %xmm8, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movd %xmm8, %esi
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %esi, %edx
-; SSE2-NEXT:    addl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %esi
-; SSE2-NEXT:    cmovol %ecx, %esi
-; SSE2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[3,1,2,3]
-; SSE2-NEXT:    movd %xmm5, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %esi
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %esi, %edx
-; SSE2-NEXT:    addl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %esi
-; SSE2-NEXT:    cmovol %ecx, %esi
-; SSE2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r14d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r14d, %edx
-; SSE2-NEXT:    addl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %r14d
-; SSE2-NEXT:    cmovol %ecx, %r14d
-; SSE2-NEXT:    movd %xmm6, %eax
-; SSE2-NEXT:    movd %xmm2, %r15d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r15d, %edx
-; SSE2-NEXT:    addl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %r15d
-; SSE2-NEXT:    cmovol %ecx, %r15d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %r12d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r12d, %edx
-; SSE2-NEXT:    addl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %r12d
-; SSE2-NEXT:    cmovol %ecx, %r12d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r13d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r13d, %esi
-; SSE2-NEXT:    addl %eax, %esi
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %r13d
-; SSE2-NEXT:    cmovol %ecx, %r13d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r11d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r11d, %edi
-; SSE2-NEXT:    addl %eax, %edi
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %r11d
-; SSE2-NEXT:    cmovol %ecx, %r11d
-; SSE2-NEXT:    movd %xmm7, %eax
-; SSE2-NEXT:    movd %xmm3, %ebx
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %ebx, %edi
-; SSE2-NEXT:    addl %eax, %edi
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %ebx
-; SSE2-NEXT:    cmovol %ecx, %ebx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %ebp
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %ebp, %edi
-; SSE2-NEXT:    addl %eax, %edi
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %ebp
-; SSE2-NEXT:    cmovol %ecx, %ebp
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r10d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r10d, %edi
-; SSE2-NEXT:    addl %eax, %edi
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %r10d
-; SSE2-NEXT:    cmovol %ecx, %r10d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %ecx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r9d
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %r9d, %eax
-; SSE2-NEXT:    addl %ecx, %eax
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %ecx, %r9d
-; SSE2-NEXT:    cmovol %edi, %r9d
-; SSE2-NEXT:    movd %xmm4, %edx
-; SSE2-NEXT:    movd %xmm0, %esi
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %esi, %ecx
-; SSE2-NEXT:    addl %edx, %ecx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %edx, %esi
-; SSE2-NEXT:    cmovol %edi, %esi
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %ecx
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %ecx, %edx
-; SSE2-NEXT:    addl %eax, %edx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %eax, %ecx
-; SSE2-NEXT:    cmovol %edi, %ecx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[3,1,2,3]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r8d
-; SSE2-NEXT:    movd %xmm0, %eax
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %eax, %edx
-; SSE2-NEXT:    addl %r8d, %edx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    addl %r8d, %eax
-; SSE2-NEXT:    cmovol %edi, %eax
-; SSE2-NEXT:    movd %eax, %xmm8
-; SSE2-NEXT:    movd %ecx, %xmm5
-; SSE2-NEXT:    movd %esi, %xmm0
-; SSE2-NEXT:    movd %r9d, %xmm9
-; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 4-byte Reload
-; SSE2-NEXT:    # xmm10 = mem[0],zero,zero,zero
-; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 4-byte Reload
-; SSE2-NEXT:    # xmm4 = mem[0],zero,zero,zero
-; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
-; SSE2-NEXT:    # xmm1 = mem[0],zero,zero,zero
-; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 4-byte Reload
-; SSE2-NEXT:    # xmm11 = mem[0],zero,zero,zero
-; SSE2-NEXT:    movd %r13d, %xmm12
-; SSE2-NEXT:    movd %r12d, %xmm6
-; SSE2-NEXT:    movd %r15d, %xmm2
-; SSE2-NEXT:    movd %r14d, %xmm13
-; SSE2-NEXT:    movd %r10d, %xmm14
-; SSE2-NEXT:    movd %ebp, %xmm7
-; SSE2-NEXT:    movd %ebx, %xmm3
-; SSE2-NEXT:    movd %r11d, %xmm15
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm5 = xmm5[0],xmm8[0],xmm5[1],xmm8[1]
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1]
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm5[0]
-; SSE2-NEXT:    unpcklps {{.*#+}} xmm4 = xmm4[0],xmm10[0],xmm4[1],xmm10[1]
-; SSE2-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm11[0],xmm1[1],xmm11[1]
-; SSE2-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm4[0]
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm6 = xmm6[0],xmm12[0],xmm6[1],xmm12[1]
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm13[0],xmm2[1],xmm13[1]
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm6[0]
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm7 = xmm7[0],xmm14[0],xmm7[1],xmm14[1]
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm15[0],xmm3[1],xmm15[1]
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm7[0]
-; SSE2-NEXT:    popq %rbx
-; SSE2-NEXT:    popq %r12
-; SSE2-NEXT:    popq %r13
-; SSE2-NEXT:    popq %r14
-; SSE2-NEXT:    popq %r15
-; SSE2-NEXT:    popq %rbp
+; SSE2-NEXT:    movdqa %xmm1, %xmm8
+; SSE2-NEXT:    movdqa %xmm0, %xmm11
+; SSE2-NEXT:    pxor %xmm10, %xmm10
+; SSE2-NEXT:    pxor %xmm12, %xmm12
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm12
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm9
+; SSE2-NEXT:    pxor %xmm9, %xmm12
+; SSE2-NEXT:    pxor %xmm1, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm1
+; SSE2-NEXT:    pxor %xmm9, %xmm1
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm12
+; SSE2-NEXT:    paddd %xmm4, %xmm11
+; SSE2-NEXT:    pxor %xmm4, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm11, %xmm4
+; SSE2-NEXT:    movdqa %xmm4, %xmm0
+; SSE2-NEXT:    pxor %xmm9, %xmm0
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm0
+; SSE2-NEXT:    pandn %xmm12, %xmm0
+; SSE2-NEXT:    movdqa {{.*#+}} xmm12 = [2147483648,2147483648,2147483648,2147483648]
+; SSE2-NEXT:    movdqa %xmm4, %xmm1
+; SSE2-NEXT:    pandn %xmm12, %xmm1
+; SSE2-NEXT:    psrld $1, %xmm4
+; SSE2-NEXT:    por %xmm1, %xmm4
+; SSE2-NEXT:    pand %xmm0, %xmm4
+; SSE2-NEXT:    pandn %xmm11, %xmm0
+; SSE2-NEXT:    por %xmm4, %xmm0
+; SSE2-NEXT:    pxor %xmm11, %xmm11
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm11
+; SSE2-NEXT:    pxor %xmm9, %xmm11
+; SSE2-NEXT:    pxor %xmm4, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm4
+; SSE2-NEXT:    pxor %xmm9, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm11
+; SSE2-NEXT:    paddd %xmm5, %xmm8
+; SSE2-NEXT:    pxor %xmm5, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm5
+; SSE2-NEXT:    movdqa %xmm5, %xmm1
+; SSE2-NEXT:    pxor %xmm9, %xmm1
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm1
+; SSE2-NEXT:    pandn %xmm11, %xmm1
+; SSE2-NEXT:    movdqa %xmm5, %xmm4
+; SSE2-NEXT:    pandn %xmm12, %xmm4
+; SSE2-NEXT:    psrld $1, %xmm5
+; SSE2-NEXT:    por %xmm4, %xmm5
+; SSE2-NEXT:    pand %xmm1, %xmm5
+; SSE2-NEXT:    pandn %xmm8, %xmm1
+; SSE2-NEXT:    por %xmm5, %xmm1
+; SSE2-NEXT:    pxor %xmm4, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm6, %xmm4
+; SSE2-NEXT:    pxor %xmm9, %xmm4
+; SSE2-NEXT:    pxor %xmm5, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSE2-NEXT:    pxor %xmm9, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSE2-NEXT:    paddd %xmm6, %xmm2
+; SSE2-NEXT:    pxor %xmm6, %xmm6
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm6
+; SSE2-NEXT:    movdqa %xmm6, %xmm8
+; SSE2-NEXT:    pxor %xmm9, %xmm8
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm8
+; SSE2-NEXT:    pandn %xmm4, %xmm8
+; SSE2-NEXT:    movdqa %xmm6, %xmm4
+; SSE2-NEXT:    pandn %xmm12, %xmm4
+; SSE2-NEXT:    psrld $1, %xmm6
+; SSE2-NEXT:    por %xmm4, %xmm6
+; SSE2-NEXT:    pand %xmm8, %xmm6
+; SSE2-NEXT:    pandn %xmm2, %xmm8
+; SSE2-NEXT:    por %xmm6, %xmm8
+; SSE2-NEXT:    pxor %xmm2, %xmm2
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm2
+; SSE2-NEXT:    pxor %xmm9, %xmm2
+; SSE2-NEXT:    pxor %xmm4, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE2-NEXT:    pxor %xmm9, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm2
+; SSE2-NEXT:    paddd %xmm7, %xmm3
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm10
+; SSE2-NEXT:    pxor %xmm10, %xmm9
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm9
+; SSE2-NEXT:    pandn %xmm2, %xmm9
+; SSE2-NEXT:    movdqa %xmm10, %xmm2
+; SSE2-NEXT:    pandn %xmm12, %xmm2
+; SSE2-NEXT:    psrld $1, %xmm10
+; SSE2-NEXT:    por %xmm2, %xmm10
+; SSE2-NEXT:    pand %xmm9, %xmm10
+; SSE2-NEXT:    pandn %xmm3, %xmm9
+; SSE2-NEXT:    por %xmm10, %xmm9
+; SSE2-NEXT:    movdqa %xmm8, %xmm2
+; SSE2-NEXT:    movdqa %xmm9, %xmm3
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v16i32:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pushq %rbp
-; SSSE3-NEXT:    pushq %r15
-; SSSE3-NEXT:    pushq %r14
-; SSSE3-NEXT:    pushq %r13
-; SSSE3-NEXT:    pushq %r12
-; SSSE3-NEXT:    pushq %rbx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm8, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm8, %esi
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %esi, %edx
-; SSSE3-NEXT:    addl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %esi
-; SSSE3-NEXT:    cmovol %ecx, %esi
-; SSSE3-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSSE3-NEXT:    movd %xmm5, %eax
-; SSSE3-NEXT:    movd %xmm1, %esi
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %esi, %edx
-; SSSE3-NEXT:    addl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %esi
-; SSSE3-NEXT:    cmovol %ecx, %esi
-; SSSE3-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm8, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm8, %esi
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %esi, %edx
-; SSSE3-NEXT:    addl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %esi
-; SSSE3-NEXT:    cmovol %ecx, %esi
-; SSSE3-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm5, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %esi
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %esi, %edx
-; SSSE3-NEXT:    addl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %esi
-; SSSE3-NEXT:    cmovol %ecx, %esi
-; SSSE3-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r14d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r14d, %edx
-; SSSE3-NEXT:    addl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %r14d
-; SSSE3-NEXT:    cmovol %ecx, %r14d
-; SSSE3-NEXT:    movd %xmm6, %eax
-; SSSE3-NEXT:    movd %xmm2, %r15d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r15d, %edx
-; SSSE3-NEXT:    addl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %r15d
-; SSSE3-NEXT:    cmovol %ecx, %r15d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %r12d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r12d, %edx
-; SSSE3-NEXT:    addl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %r12d
-; SSSE3-NEXT:    cmovol %ecx, %r12d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r13d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r13d, %esi
-; SSSE3-NEXT:    addl %eax, %esi
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %r13d
-; SSSE3-NEXT:    cmovol %ecx, %r13d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r11d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r11d, %edi
-; SSSE3-NEXT:    addl %eax, %edi
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %r11d
-; SSSE3-NEXT:    cmovol %ecx, %r11d
-; SSSE3-NEXT:    movd %xmm7, %eax
-; SSSE3-NEXT:    movd %xmm3, %ebx
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %ebx, %edi
-; SSSE3-NEXT:    addl %eax, %edi
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %ebx
-; SSSE3-NEXT:    cmovol %ecx, %ebx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %ebp
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %ebp, %edi
-; SSSE3-NEXT:    addl %eax, %edi
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %ebp
-; SSSE3-NEXT:    cmovol %ecx, %ebp
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r10d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r10d, %edi
-; SSSE3-NEXT:    addl %eax, %edi
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %r10d
-; SSSE3-NEXT:    cmovol %ecx, %r10d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %ecx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r9d
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %r9d, %eax
-; SSSE3-NEXT:    addl %ecx, %eax
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %ecx, %r9d
-; SSSE3-NEXT:    cmovol %edi, %r9d
-; SSSE3-NEXT:    movd %xmm4, %edx
-; SSSE3-NEXT:    movd %xmm0, %esi
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %esi, %ecx
-; SSSE3-NEXT:    addl %edx, %ecx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %edx, %esi
-; SSSE3-NEXT:    cmovol %edi, %esi
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %ecx
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %ecx, %edx
-; SSSE3-NEXT:    addl %eax, %edx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %eax, %ecx
-; SSSE3-NEXT:    cmovol %edi, %ecx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[3,1,2,3]
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r8d
-; SSSE3-NEXT:    movd %xmm0, %eax
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %eax, %edx
-; SSSE3-NEXT:    addl %r8d, %edx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    addl %r8d, %eax
-; SSSE3-NEXT:    cmovol %edi, %eax
-; SSSE3-NEXT:    movd %eax, %xmm8
-; SSSE3-NEXT:    movd %ecx, %xmm5
-; SSSE3-NEXT:    movd %esi, %xmm0
-; SSSE3-NEXT:    movd %r9d, %xmm9
-; SSSE3-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 4-byte Reload
-; SSSE3-NEXT:    # xmm10 = mem[0],zero,zero,zero
-; SSSE3-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 4-byte Reload
-; SSSE3-NEXT:    # xmm4 = mem[0],zero,zero,zero
-; SSSE3-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
-; SSSE3-NEXT:    # xmm1 = mem[0],zero,zero,zero
-; SSSE3-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 4-byte Reload
-; SSSE3-NEXT:    # xmm11 = mem[0],zero,zero,zero
-; SSSE3-NEXT:    movd %r13d, %xmm12
-; SSSE3-NEXT:    movd %r12d, %xmm6
-; SSSE3-NEXT:    movd %r15d, %xmm2
-; SSSE3-NEXT:    movd %r14d, %xmm13
-; SSSE3-NEXT:    movd %r10d, %xmm14
-; SSSE3-NEXT:    movd %ebp, %xmm7
-; SSSE3-NEXT:    movd %ebx, %xmm3
-; SSSE3-NEXT:    movd %r11d, %xmm15
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm5 = xmm5[0],xmm8[0],xmm5[1],xmm8[1]
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1]
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm5[0]
-; SSSE3-NEXT:    unpcklps {{.*#+}} xmm4 = xmm4[0],xmm10[0],xmm4[1],xmm10[1]
-; SSSE3-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm11[0],xmm1[1],xmm11[1]
-; SSSE3-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm4[0]
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm6 = xmm6[0],xmm12[0],xmm6[1],xmm12[1]
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm13[0],xmm2[1],xmm13[1]
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm6[0]
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm7 = xmm7[0],xmm14[0],xmm7[1],xmm14[1]
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm15[0],xmm3[1],xmm15[1]
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm7[0]
-; SSSE3-NEXT:    popq %rbx
-; SSSE3-NEXT:    popq %r12
-; SSSE3-NEXT:    popq %r13
-; SSSE3-NEXT:    popq %r14
-; SSSE3-NEXT:    popq %r15
-; SSSE3-NEXT:    popq %rbp
+; SSSE3-NEXT:    movdqa %xmm1, %xmm8
+; SSSE3-NEXT:    movdqa %xmm0, %xmm11
+; SSSE3-NEXT:    pxor %xmm10, %xmm10
+; SSSE3-NEXT:    pxor %xmm12, %xmm12
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm12
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm9
+; SSSE3-NEXT:    pxor %xmm9, %xmm12
+; SSSE3-NEXT:    pxor %xmm1, %xmm1
+; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm1
+; SSSE3-NEXT:    pxor %xmm9, %xmm1
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm12
+; SSSE3-NEXT:    paddd %xmm4, %xmm11
+; SSSE3-NEXT:    pxor %xmm4, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm11, %xmm4
+; SSSE3-NEXT:    movdqa %xmm4, %xmm0
+; SSSE3-NEXT:    pxor %xmm9, %xmm0
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm0
+; SSSE3-NEXT:    pandn %xmm12, %xmm0
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm12 = [2147483648,2147483648,2147483648,2147483648]
+; SSSE3-NEXT:    movdqa %xmm4, %xmm1
+; SSSE3-NEXT:    pandn %xmm12, %xmm1
+; SSSE3-NEXT:    psrld $1, %xmm4
+; SSSE3-NEXT:    por %xmm1, %xmm4
+; SSSE3-NEXT:    pand %xmm0, %xmm4
+; SSSE3-NEXT:    pandn %xmm11, %xmm0
+; SSSE3-NEXT:    por %xmm4, %xmm0
+; SSSE3-NEXT:    pxor %xmm11, %xmm11
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm11
+; SSSE3-NEXT:    pxor %xmm9, %xmm11
+; SSSE3-NEXT:    pxor %xmm4, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm4
+; SSSE3-NEXT:    pxor %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm11
+; SSSE3-NEXT:    paddd %xmm5, %xmm8
+; SSSE3-NEXT:    pxor %xmm5, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm5
+; SSSE3-NEXT:    movdqa %xmm5, %xmm1
+; SSSE3-NEXT:    pxor %xmm9, %xmm1
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm1
+; SSSE3-NEXT:    pandn %xmm11, %xmm1
+; SSSE3-NEXT:    movdqa %xmm5, %xmm4
+; SSSE3-NEXT:    pandn %xmm12, %xmm4
+; SSSE3-NEXT:    psrld $1, %xmm5
+; SSSE3-NEXT:    por %xmm4, %xmm5
+; SSSE3-NEXT:    pand %xmm1, %xmm5
+; SSSE3-NEXT:    pandn %xmm8, %xmm1
+; SSSE3-NEXT:    por %xmm5, %xmm1
+; SSSE3-NEXT:    pxor %xmm4, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm6, %xmm4
+; SSSE3-NEXT:    pxor %xmm9, %xmm4
+; SSSE3-NEXT:    pxor %xmm5, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSSE3-NEXT:    pxor %xmm9, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSSE3-NEXT:    paddd %xmm6, %xmm2
+; SSSE3-NEXT:    pxor %xmm6, %xmm6
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm6
+; SSSE3-NEXT:    movdqa %xmm6, %xmm8
+; SSSE3-NEXT:    pxor %xmm9, %xmm8
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm8
+; SSSE3-NEXT:    pandn %xmm4, %xmm8
+; SSSE3-NEXT:    movdqa %xmm6, %xmm4
+; SSSE3-NEXT:    pandn %xmm12, %xmm4
+; SSSE3-NEXT:    psrld $1, %xmm6
+; SSSE3-NEXT:    por %xmm4, %xmm6
+; SSSE3-NEXT:    pand %xmm8, %xmm6
+; SSSE3-NEXT:    pandn %xmm2, %xmm8
+; SSSE3-NEXT:    por %xmm6, %xmm8
+; SSSE3-NEXT:    pxor %xmm2, %xmm2
+; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm2
+; SSSE3-NEXT:    pxor %xmm9, %xmm2
+; SSSE3-NEXT:    pxor %xmm4, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSSE3-NEXT:    pxor %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm2
+; SSSE3-NEXT:    paddd %xmm7, %xmm3
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm10
+; SSSE3-NEXT:    pxor %xmm10, %xmm9
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm9
+; SSSE3-NEXT:    pandn %xmm2, %xmm9
+; SSSE3-NEXT:    movdqa %xmm10, %xmm2
+; SSSE3-NEXT:    pandn %xmm12, %xmm2
+; SSSE3-NEXT:    psrld $1, %xmm10
+; SSSE3-NEXT:    por %xmm2, %xmm10
+; SSSE3-NEXT:    pand %xmm9, %xmm10
+; SSSE3-NEXT:    pandn %xmm3, %xmm9
+; SSSE3-NEXT:    por %xmm10, %xmm9
+; SSSE3-NEXT:    movdqa %xmm8, %xmm2
+; SSSE3-NEXT:    movdqa %xmm9, %xmm3
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v16i32:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    pushq %rbp
-; SSE41-NEXT:    pushq %r15
-; SSE41-NEXT:    pushq %r14
-; SSE41-NEXT:    pushq %r13
-; SSE41-NEXT:    pushq %r12
-; SSE41-NEXT:    pushq %rbx
-; SSE41-NEXT:    pextrd $3, %xmm5, %eax
-; SSE41-NEXT:    pextrd $3, %xmm1, %esi
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %esi, %edx
-; SSE41-NEXT:    addl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %esi
-; SSE41-NEXT:    cmovol %ecx, %esi
-; SSE41-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE41-NEXT:    pextrd $2, %xmm5, %eax
-; SSE41-NEXT:    pextrd $2, %xmm1, %esi
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %esi, %edx
-; SSE41-NEXT:    addl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %esi
-; SSE41-NEXT:    cmovol %ecx, %esi
-; SSE41-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE41-NEXT:    movd %xmm5, %eax
-; SSE41-NEXT:    movd %xmm1, %esi
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %esi, %edx
-; SSE41-NEXT:    addl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %esi
-; SSE41-NEXT:    cmovol %ecx, %esi
-; SSE41-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE41-NEXT:    pextrd $1, %xmm5, %eax
-; SSE41-NEXT:    pextrd $1, %xmm1, %esi
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %esi, %edx
-; SSE41-NEXT:    addl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %esi
-; SSE41-NEXT:    cmovol %ecx, %esi
-; SSE41-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE41-NEXT:    pextrd $3, %xmm6, %eax
-; SSE41-NEXT:    pextrd $3, %xmm2, %r14d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r14d, %edx
-; SSE41-NEXT:    addl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %r14d
-; SSE41-NEXT:    cmovol %ecx, %r14d
-; SSE41-NEXT:    pextrd $2, %xmm6, %eax
-; SSE41-NEXT:    pextrd $2, %xmm2, %r15d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r15d, %edx
-; SSE41-NEXT:    addl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %r15d
-; SSE41-NEXT:    cmovol %ecx, %r15d
-; SSE41-NEXT:    movd %xmm6, %eax
-; SSE41-NEXT:    movd %xmm2, %r12d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r12d, %edx
-; SSE41-NEXT:    addl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %r12d
-; SSE41-NEXT:    cmovol %ecx, %r12d
-; SSE41-NEXT:    pextrd $1, %xmm6, %eax
-; SSE41-NEXT:    pextrd $1, %xmm2, %r13d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r13d, %esi
-; SSE41-NEXT:    addl %eax, %esi
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %r13d
-; SSE41-NEXT:    cmovol %ecx, %r13d
-; SSE41-NEXT:    pextrd $3, %xmm7, %eax
-; SSE41-NEXT:    pextrd $3, %xmm3, %r11d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r11d, %edi
-; SSE41-NEXT:    addl %eax, %edi
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %r11d
-; SSE41-NEXT:    cmovol %ecx, %r11d
-; SSE41-NEXT:    pextrd $2, %xmm7, %eax
-; SSE41-NEXT:    pextrd $2, %xmm3, %ebx
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %ebx, %edi
-; SSE41-NEXT:    addl %eax, %edi
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %ebx
-; SSE41-NEXT:    cmovol %ecx, %ebx
-; SSE41-NEXT:    movd %xmm7, %eax
-; SSE41-NEXT:    movd %xmm3, %ebp
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %ebp, %edi
-; SSE41-NEXT:    addl %eax, %edi
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %ebp
-; SSE41-NEXT:    cmovol %ecx, %ebp
-; SSE41-NEXT:    pextrd $1, %xmm7, %eax
-; SSE41-NEXT:    pextrd $1, %xmm3, %r10d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r10d, %edi
-; SSE41-NEXT:    addl %eax, %edi
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %r10d
-; SSE41-NEXT:    cmovol %ecx, %r10d
-; SSE41-NEXT:    pextrd $3, %xmm4, %ecx
-; SSE41-NEXT:    pextrd $3, %xmm0, %r9d
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %r9d, %eax
-; SSE41-NEXT:    addl %ecx, %eax
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %ecx, %r9d
-; SSE41-NEXT:    cmovol %edi, %r9d
-; SSE41-NEXT:    pextrd $2, %xmm4, %edx
-; SSE41-NEXT:    pextrd $2, %xmm0, %esi
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %esi, %ecx
-; SSE41-NEXT:    addl %edx, %ecx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %edx, %esi
-; SSE41-NEXT:    cmovol %edi, %esi
-; SSE41-NEXT:    movd %xmm4, %eax
-; SSE41-NEXT:    movd %xmm0, %ecx
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %ecx, %edx
-; SSE41-NEXT:    addl %eax, %edx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %eax, %ecx
-; SSE41-NEXT:    cmovol %edi, %ecx
-; SSE41-NEXT:    pextrd $1, %xmm4, %r8d
-; SSE41-NEXT:    pextrd $1, %xmm0, %eax
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %eax, %edx
-; SSE41-NEXT:    addl %r8d, %edx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    addl %r8d, %eax
-; SSE41-NEXT:    cmovol %edi, %eax
-; SSE41-NEXT:    movd %ecx, %xmm0
-; SSE41-NEXT:    pinsrd $1, %eax, %xmm0
-; SSE41-NEXT:    pinsrd $2, %esi, %xmm0
-; SSE41-NEXT:    pinsrd $3, %r9d, %xmm0
-; SSE41-NEXT:    movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Folded Reload
-; SSE41-NEXT:    # xmm1 = mem[0],zero,zero,zero
-; SSE41-NEXT:    pinsrd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Folded Reload
-; SSE41-NEXT:    pinsrd $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Folded Reload
-; SSE41-NEXT:    pinsrd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Folded Reload
-; SSE41-NEXT:    movd %r12d, %xmm2
-; SSE41-NEXT:    pinsrd $1, %r13d, %xmm2
-; SSE41-NEXT:    pinsrd $2, %r15d, %xmm2
-; SSE41-NEXT:    pinsrd $3, %r14d, %xmm2
-; SSE41-NEXT:    movd %ebp, %xmm3
-; SSE41-NEXT:    pinsrd $1, %r10d, %xmm3
-; SSE41-NEXT:    pinsrd $2, %ebx, %xmm3
-; SSE41-NEXT:    pinsrd $3, %r11d, %xmm3
-; SSE41-NEXT:    popq %rbx
-; SSE41-NEXT:    popq %r12
-; SSE41-NEXT:    popq %r13
-; SSE41-NEXT:    popq %r14
-; SSE41-NEXT:    popq %r15
-; SSE41-NEXT:    popq %rbp
+; SSE41-NEXT:    movdqa %xmm0, %xmm8
+; SSE41-NEXT:    pxor %xmm9, %xmm9
+; SSE41-NEXT:    pxor %xmm11, %xmm11
+; SSE41-NEXT:    pcmpgtd %xmm4, %xmm11
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm10
+; SSE41-NEXT:    pxor %xmm10, %xmm11
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm8, %xmm0
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm11
+; SSE41-NEXT:    paddd %xmm4, %xmm8
+; SSE41-NEXT:    pxor %xmm4, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm8, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pandn %xmm11, %xmm4
+; SSE41-NEXT:    movaps {{.*#+}} xmm12 = [2147483647,2147483647,2147483647,2147483647]
+; SSE41-NEXT:    movaps {{.*#+}} xmm11 = [2147483648,2147483648,2147483648,2147483648]
+; SSE41-NEXT:    movaps %xmm11, %xmm13
+; SSE41-NEXT:    movdqa %xmm8, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm12, %xmm13
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm13, %xmm8
+; SSE41-NEXT:    xorps %xmm13, %xmm13
+; SSE41-NEXT:    pcmpgtd %xmm5, %xmm13
+; SSE41-NEXT:    pxor %xmm10, %xmm13
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm13
+; SSE41-NEXT:    paddd %xmm5, %xmm1
+; SSE41-NEXT:    pxor %xmm4, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pandn %xmm13, %xmm4
+; SSE41-NEXT:    movaps %xmm11, %xmm5
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm12, %xmm5
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm5, %xmm1
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm6, %xmm0
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    xorps %xmm5, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSE41-NEXT:    pxor %xmm10, %xmm5
+; SSE41-NEXT:    pcmpeqd %xmm5, %xmm0
+; SSE41-NEXT:    paddd %xmm6, %xmm2
+; SSE41-NEXT:    pxor %xmm4, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSE41-NEXT:    pandn %xmm0, %xmm4
+; SSE41-NEXT:    movaps %xmm11, %xmm5
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm12, %xmm5
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm5, %xmm2
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm7, %xmm0
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    pxor %xmm4, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm4, %xmm0
+; SSE41-NEXT:    paddd %xmm7, %xmm3
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm9
+; SSE41-NEXT:    pxor %xmm10, %xmm9
+; SSE41-NEXT:    pcmpeqd %xmm4, %xmm9
+; SSE41-NEXT:    pandn %xmm0, %xmm9
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm12, %xmm11
+; SSE41-NEXT:    movdqa %xmm9, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm11, %xmm3
+; SSE41-NEXT:    movaps %xmm8, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: v16i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    pushq %rbp
-; AVX1-NEXT:    pushq %r15
-; AVX1-NEXT:    pushq %r14
-; AVX1-NEXT:    pushq %r13
-; AVX1-NEXT:    pushq %r12
-; AVX1-NEXT:    pushq %rbx
-; AVX1-NEXT:    vpextrd $3, %xmm3, %eax
-; AVX1-NEXT:    vpextrd $3, %xmm1, %esi
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %esi, %edx
-; AVX1-NEXT:    addl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %esi
-; AVX1-NEXT:    cmovol %ecx, %esi
-; AVX1-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX1-NEXT:    vpextrd $2, %xmm3, %eax
-; AVX1-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %esi, %edx
-; AVX1-NEXT:    addl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %esi
-; AVX1-NEXT:    cmovol %ecx, %esi
-; AVX1-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX1-NEXT:    vmovd %xmm3, %eax
-; AVX1-NEXT:    vmovd %xmm1, %esi
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %esi, %edx
-; AVX1-NEXT:    addl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %esi
-; AVX1-NEXT:    cmovol %ecx, %esi
-; AVX1-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX1-NEXT:    vpextrd $1, %xmm3, %eax
-; AVX1-NEXT:    vpextrd $1, %xmm1, %esi
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %esi, %edx
-; AVX1-NEXT:    addl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %esi
-; AVX1-NEXT:    cmovol %ecx, %esi
-; AVX1-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm3
-; AVX1-NEXT:    vpextrd $3, %xmm3, %eax
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vpextrd $3, %xmm1, %r14d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r14d, %edx
-; AVX1-NEXT:    addl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %r14d
-; AVX1-NEXT:    cmovol %ecx, %r14d
-; AVX1-NEXT:    vpextrd $2, %xmm3, %eax
-; AVX1-NEXT:    vpextrd $2, %xmm1, %r15d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r15d, %edx
-; AVX1-NEXT:    addl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %r15d
-; AVX1-NEXT:    cmovol %ecx, %r15d
-; AVX1-NEXT:    vmovd %xmm3, %eax
-; AVX1-NEXT:    vmovd %xmm1, %r12d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r12d, %edx
-; AVX1-NEXT:    addl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %r12d
-; AVX1-NEXT:    cmovol %ecx, %r12d
-; AVX1-NEXT:    vpextrd $1, %xmm3, %eax
-; AVX1-NEXT:    vpextrd $1, %xmm1, %r13d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r13d, %esi
-; AVX1-NEXT:    addl %eax, %esi
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %r13d
-; AVX1-NEXT:    cmovol %ecx, %r13d
-; AVX1-NEXT:    vpextrd $3, %xmm2, %eax
-; AVX1-NEXT:    vpextrd $3, %xmm0, %r11d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r11d, %edi
-; AVX1-NEXT:    addl %eax, %edi
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %r11d
-; AVX1-NEXT:    cmovol %ecx, %r11d
-; AVX1-NEXT:    vpextrd $2, %xmm2, %eax
-; AVX1-NEXT:    vpextrd $2, %xmm0, %ebx
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %ebx, %edi
-; AVX1-NEXT:    addl %eax, %edi
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %ebx
-; AVX1-NEXT:    cmovol %ecx, %ebx
-; AVX1-NEXT:    vmovd %xmm2, %eax
-; AVX1-NEXT:    vmovd %xmm0, %ebp
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %ebp, %edi
-; AVX1-NEXT:    addl %eax, %edi
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %ebp
-; AVX1-NEXT:    cmovol %ecx, %ebp
-; AVX1-NEXT:    vpextrd $1, %xmm2, %eax
-; AVX1-NEXT:    vpextrd $1, %xmm0, %r10d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r10d, %edi
-; AVX1-NEXT:    addl %eax, %edi
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %r10d
-; AVX1-NEXT:    cmovol %ecx, %r10d
-; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm1
-; AVX1-NEXT:    vpextrd $3, %xmm1, %ecx
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpextrd $3, %xmm0, %r8d
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %r8d, %eax
-; AVX1-NEXT:    addl %ecx, %eax
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %ecx, %r8d
-; AVX1-NEXT:    cmovol %edi, %r8d
-; AVX1-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX1-NEXT:    vpextrd $2, %xmm0, %edx
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %edx, %ecx
-; AVX1-NEXT:    addl %esi, %ecx
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %esi, %edx
-; AVX1-NEXT:    cmovol %edi, %edx
-; AVX1-NEXT:    vmovd %xmm1, %eax
-; AVX1-NEXT:    vmovd %xmm0, %ecx
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %ecx, %esi
-; AVX1-NEXT:    addl %eax, %esi
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %eax, %ecx
-; AVX1-NEXT:    cmovol %edi, %ecx
-; AVX1-NEXT:    vpextrd $1, %xmm1, %r9d
-; AVX1-NEXT:    vpextrd $1, %xmm0, %eax
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %eax, %esi
-; AVX1-NEXT:    addl %r9d, %esi
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    addl %r9d, %eax
-; AVX1-NEXT:    cmovol %edi, %eax
-; AVX1-NEXT:    vmovd %ecx, %xmm0
-; AVX1-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
-; AVX1-NEXT:    vpinsrd $2, %edx, %xmm0, %xmm0
-; AVX1-NEXT:    vpinsrd $3, %r8d, %xmm0, %xmm0
-; AVX1-NEXT:    vmovd %ebp, %xmm1
-; AVX1-NEXT:    vpinsrd $1, %r10d, %xmm1, %xmm1
-; AVX1-NEXT:    vpinsrd $2, %ebx, %xmm1, %xmm1
-; AVX1-NEXT:    vpinsrd $3, %r11d, %xmm1, %xmm1
-; AVX1-NEXT:    vmovd %r12d, %xmm2
-; AVX1-NEXT:    vpinsrd $1, %r13d, %xmm2, %xmm2
-; AVX1-NEXT:    vpinsrd $2, %r15d, %xmm2, %xmm2
-; AVX1-NEXT:    vpinsrd $3, %r14d, %xmm2, %xmm2
-; AVX1-NEXT:    vmovd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 4-byte Folded Reload
-; AVX1-NEXT:    # xmm3 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vpinsrd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX1-NEXT:    vpinsrd $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX1-NEXT:    vpinsrd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm1
-; AVX1-NEXT:    popq %rbx
-; AVX1-NEXT:    popq %r12
-; AVX1-NEXT:    popq %r13
-; AVX1-NEXT:    popq %r14
-; AVX1-NEXT:    popq %r15
-; AVX1-NEXT:    popq %rbp
+; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm9
+; AVX1-NEXT:    vpxor %xmm12, %xmm12, %xmm12
+; AVX1-NEXT:    vpcmpgtd %xmm9, %xmm12, %xmm7
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm8
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm7
+; AVX1-NEXT:    vpcmpgtd %xmm7, %xmm12, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm10
+; AVX1-NEXT:    vpcmpeqd %xmm8, %xmm10, %xmm8
+; AVX1-NEXT:    vpcmpgtd %xmm2, %xmm12, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm11
+; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm12, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm6
+; AVX1-NEXT:    vpcmpeqd %xmm11, %xmm6, %xmm11
+; AVX1-NEXT:    vinsertf128 $1, %xmm8, %ymm11, %ymm8
+; AVX1-NEXT:    vpaddd %xmm9, %xmm7, %xmm9
+; AVX1-NEXT:    vpcmpgtd %xmm9, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm5
+; AVX1-NEXT:    vpcmpeqd %xmm5, %xmm10, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm12, %xmm2
+; AVX1-NEXT:    vpxor %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm6, %xmm2
+; AVX1-NEXT:    vpxor %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm5, %ymm2, %ymm2
+; AVX1-NEXT:    vandps %ymm2, %ymm8, %ymm5
+; AVX1-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm7
+; AVX1-NEXT:    vmovaps {{.*#+}} ymm8 = [2147483647,2147483647,2147483647,2147483647,2147483647,2147483647,2147483647,2147483647]
+; AVX1-NEXT:    vmovaps {{.*#+}} ymm10 = [2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648]
+; AVX1-NEXT:    vblendvps %ymm7, %ymm8, %ymm10, %ymm7
+; AVX1-NEXT:    vinsertf128 $1, %xmm9, %ymm0, %ymm0
+; AVX1-NEXT:    vblendvps %ymm5, %ymm7, %ymm0, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm5
+; AVX1-NEXT:    vpcmpgtd %xmm5, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm7
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpcmpgtd %xmm2, %xmm12, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm6
+; AVX1-NEXT:    vpcmpeqd %xmm7, %xmm6, %xmm9
+; AVX1-NEXT:    vpcmpgtd %xmm3, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm11
+; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm7
+; AVX1-NEXT:    vpcmpeqd %xmm11, %xmm7, %xmm11
+; AVX1-NEXT:    vinsertf128 $1, %xmm9, %ymm11, %ymm9
+; AVX1-NEXT:    vpaddd %xmm5, %xmm2, %xmm11
+; AVX1-NEXT:    vpcmpgtd %xmm11, %xmm12, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm6, %xmm2
+; AVX1-NEXT:    vpxor %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddd %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm12, %xmm3
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm7, %xmm3
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
+; AVX1-NEXT:    vandps %ymm2, %ymm9, %ymm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm5, %ymm1, %ymm3
+; AVX1-NEXT:    vblendvps %ymm3, %ymm8, %ymm10, %ymm3
+; AVX1-NEXT:    vinsertf128 $1, %xmm11, %ymm1, %ymm1
+; AVX1-NEXT:    vblendvps %ymm2, %ymm3, %ymm1, %ymm1
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: v16i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    pushq %rbp
-; AVX2-NEXT:    pushq %r15
-; AVX2-NEXT:    pushq %r14
-; AVX2-NEXT:    pushq %r13
-; AVX2-NEXT:    pushq %r12
-; AVX2-NEXT:    pushq %rbx
-; AVX2-NEXT:    vpextrd $3, %xmm3, %eax
-; AVX2-NEXT:    vpextrd $3, %xmm1, %esi
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %esi, %edx
-; AVX2-NEXT:    addl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %esi
-; AVX2-NEXT:    cmovol %ecx, %esi
-; AVX2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX2-NEXT:    vpextrd $2, %xmm3, %eax
-; AVX2-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %esi, %edx
-; AVX2-NEXT:    addl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %esi
-; AVX2-NEXT:    cmovol %ecx, %esi
-; AVX2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX2-NEXT:    vmovd %xmm3, %eax
-; AVX2-NEXT:    vmovd %xmm1, %esi
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %esi, %edx
-; AVX2-NEXT:    addl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %esi
-; AVX2-NEXT:    cmovol %ecx, %esi
-; AVX2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX2-NEXT:    vpextrd $1, %xmm3, %eax
-; AVX2-NEXT:    vpextrd $1, %xmm1, %esi
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %esi, %edx
-; AVX2-NEXT:    addl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %esi
-; AVX2-NEXT:    cmovol %ecx, %esi
-; AVX2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX2-NEXT:    vextracti128 $1, %ymm3, %xmm3
-; AVX2-NEXT:    vpextrd $3, %xmm3, %eax
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; AVX2-NEXT:    vpextrd $3, %xmm1, %r14d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r14d, %edx
-; AVX2-NEXT:    addl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %r14d
-; AVX2-NEXT:    cmovol %ecx, %r14d
-; AVX2-NEXT:    vpextrd $2, %xmm3, %eax
-; AVX2-NEXT:    vpextrd $2, %xmm1, %r15d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r15d, %edx
-; AVX2-NEXT:    addl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %r15d
-; AVX2-NEXT:    cmovol %ecx, %r15d
-; AVX2-NEXT:    vmovd %xmm3, %eax
-; AVX2-NEXT:    vmovd %xmm1, %r12d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r12d, %edx
-; AVX2-NEXT:    addl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %r12d
-; AVX2-NEXT:    cmovol %ecx, %r12d
-; AVX2-NEXT:    vpextrd $1, %xmm3, %eax
-; AVX2-NEXT:    vpextrd $1, %xmm1, %r13d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r13d, %esi
-; AVX2-NEXT:    addl %eax, %esi
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %r13d
-; AVX2-NEXT:    cmovol %ecx, %r13d
-; AVX2-NEXT:    vpextrd $3, %xmm2, %eax
-; AVX2-NEXT:    vpextrd $3, %xmm0, %r11d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r11d, %edi
-; AVX2-NEXT:    addl %eax, %edi
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %r11d
-; AVX2-NEXT:    cmovol %ecx, %r11d
-; AVX2-NEXT:    vpextrd $2, %xmm2, %eax
-; AVX2-NEXT:    vpextrd $2, %xmm0, %ebx
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %ebx, %edi
-; AVX2-NEXT:    addl %eax, %edi
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %ebx
-; AVX2-NEXT:    cmovol %ecx, %ebx
-; AVX2-NEXT:    vmovd %xmm2, %eax
-; AVX2-NEXT:    vmovd %xmm0, %ebp
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %ebp, %edi
-; AVX2-NEXT:    addl %eax, %edi
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %ebp
-; AVX2-NEXT:    cmovol %ecx, %ebp
-; AVX2-NEXT:    vpextrd $1, %xmm2, %eax
-; AVX2-NEXT:    vpextrd $1, %xmm0, %r10d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r10d, %edi
-; AVX2-NEXT:    addl %eax, %edi
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %r10d
-; AVX2-NEXT:    cmovol %ecx, %r10d
-; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm1
-; AVX2-NEXT:    vpextrd $3, %xmm1, %ecx
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; AVX2-NEXT:    vpextrd $3, %xmm0, %r8d
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %r8d, %eax
-; AVX2-NEXT:    addl %ecx, %eax
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %ecx, %r8d
-; AVX2-NEXT:    cmovol %edi, %r8d
-; AVX2-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX2-NEXT:    vpextrd $2, %xmm0, %edx
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %edx, %ecx
-; AVX2-NEXT:    addl %esi, %ecx
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %esi, %edx
-; AVX2-NEXT:    cmovol %edi, %edx
-; AVX2-NEXT:    vmovd %xmm1, %eax
-; AVX2-NEXT:    vmovd %xmm0, %ecx
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %ecx, %esi
-; AVX2-NEXT:    addl %eax, %esi
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %eax, %ecx
-; AVX2-NEXT:    cmovol %edi, %ecx
-; AVX2-NEXT:    vpextrd $1, %xmm1, %r9d
-; AVX2-NEXT:    vpextrd $1, %xmm0, %eax
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %eax, %esi
-; AVX2-NEXT:    addl %r9d, %esi
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    addl %r9d, %eax
-; AVX2-NEXT:    cmovol %edi, %eax
-; AVX2-NEXT:    vmovd %ecx, %xmm0
-; AVX2-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
-; AVX2-NEXT:    vpinsrd $2, %edx, %xmm0, %xmm0
-; AVX2-NEXT:    vpinsrd $3, %r8d, %xmm0, %xmm0
-; AVX2-NEXT:    vmovd %ebp, %xmm1
-; AVX2-NEXT:    vpinsrd $1, %r10d, %xmm1, %xmm1
-; AVX2-NEXT:    vpinsrd $2, %ebx, %xmm1, %xmm1
-; AVX2-NEXT:    vpinsrd $3, %r11d, %xmm1, %xmm1
-; AVX2-NEXT:    vmovd %r12d, %xmm2
-; AVX2-NEXT:    vpinsrd $1, %r13d, %xmm2, %xmm2
-; AVX2-NEXT:    vpinsrd $2, %r15d, %xmm2, %xmm2
-; AVX2-NEXT:    vpinsrd $3, %r14d, %xmm2, %xmm2
-; AVX2-NEXT:    vmovd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 4-byte Folded Reload
-; AVX2-NEXT:    # xmm3 = mem[0],zero,zero,zero
-; AVX2-NEXT:    vpinsrd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX2-NEXT:    vpinsrd $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX2-NEXT:    vpinsrd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm1
-; AVX2-NEXT:    popq %rbx
-; AVX2-NEXT:    popq %r12
-; AVX2-NEXT:    popq %r13
-; AVX2-NEXT:    popq %r14
-; AVX2-NEXT:    popq %r15
-; AVX2-NEXT:    popq %rbp
+; AVX2-NEXT:    vpxor %xmm4, %xmm4, %xmm4
+; AVX2-NEXT:    vpcmpgtd %ymm2, %ymm4, %ymm5
+; AVX2-NEXT:    vpcmpeqd %ymm6, %ymm6, %ymm6
+; AVX2-NEXT:    vpxor %ymm6, %ymm5, %ymm5
+; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm4, %ymm7
+; AVX2-NEXT:    vpxor %ymm6, %ymm7, %ymm7
+; AVX2-NEXT:    vpcmpeqd %ymm5, %ymm7, %ymm5
+; AVX2-NEXT:    vpaddd %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm4, %ymm2
+; AVX2-NEXT:    vpxor %ymm6, %ymm2, %ymm2
+; AVX2-NEXT:    vpcmpeqd %ymm2, %ymm7, %ymm2
+; AVX2-NEXT:    vpandn %ymm5, %ymm2, %ymm2
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm5 = [2147483647,2147483647,2147483647,2147483647,2147483647,2147483647,2147483647,2147483647]
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm7 = [2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648]
+; AVX2-NEXT:    vblendvps %ymm0, %ymm5, %ymm7, %ymm8
+; AVX2-NEXT:    vblendvps %ymm2, %ymm8, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpgtd %ymm3, %ymm4, %ymm2
+; AVX2-NEXT:    vpxor %ymm6, %ymm2, %ymm2
+; AVX2-NEXT:    vpcmpgtd %ymm1, %ymm4, %ymm8
+; AVX2-NEXT:    vpxor %ymm6, %ymm8, %ymm8
+; AVX2-NEXT:    vpcmpeqd %ymm2, %ymm8, %ymm2
+; AVX2-NEXT:    vpaddd %ymm3, %ymm1, %ymm1
+; AVX2-NEXT:    vpcmpgtd %ymm1, %ymm4, %ymm3
+; AVX2-NEXT:    vpxor %ymm6, %ymm3, %ymm3
+; AVX2-NEXT:    vpcmpeqd %ymm3, %ymm8, %ymm3
+; AVX2-NEXT:    vpandn %ymm2, %ymm3, %ymm2
+; AVX2-NEXT:    vblendvps %ymm1, %ymm5, %ymm7, %ymm3
+; AVX2-NEXT:    vblendvps %ymm2, %ymm3, %ymm1, %ymm1
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: v16i32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    pushq %rbp
-; AVX512-NEXT:    pushq %r15
-; AVX512-NEXT:    pushq %r14
-; AVX512-NEXT:    pushq %r13
-; AVX512-NEXT:    pushq %r12
-; AVX512-NEXT:    pushq %rbx
-; AVX512-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $3, %xmm0, %esi
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %esi, %edx
-; AVX512-NEXT:    addl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %esi
-; AVX512-NEXT:    cmovol %ecx, %esi
-; AVX512-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX512-NEXT:    vpextrd $2, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $2, %xmm0, %esi
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %esi, %edx
-; AVX512-NEXT:    addl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %esi
-; AVX512-NEXT:    cmovol %ecx, %esi
-; AVX512-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX512-NEXT:    vmovd %xmm1, %eax
-; AVX512-NEXT:    vmovd %xmm0, %esi
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %esi, %edx
-; AVX512-NEXT:    addl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %esi
-; AVX512-NEXT:    cmovol %ecx, %esi
-; AVX512-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX512-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $1, %xmm0, %esi
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %esi, %edx
-; AVX512-NEXT:    addl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %esi
-; AVX512-NEXT:    cmovol %ecx, %esi
-; AVX512-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX512-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX512-NEXT:    vpextrd $3, %xmm2, %eax
-; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm3
-; AVX512-NEXT:    vpextrd $3, %xmm3, %r14d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r14d, %edx
-; AVX512-NEXT:    addl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %r14d
-; AVX512-NEXT:    cmovol %ecx, %r14d
-; AVX512-NEXT:    vpextrd $2, %xmm2, %eax
-; AVX512-NEXT:    vpextrd $2, %xmm3, %r15d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r15d, %edx
-; AVX512-NEXT:    addl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %r15d
-; AVX512-NEXT:    cmovol %ecx, %r15d
-; AVX512-NEXT:    vmovd %xmm2, %eax
-; AVX512-NEXT:    vmovd %xmm3, %r12d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r12d, %edx
-; AVX512-NEXT:    addl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %r12d
-; AVX512-NEXT:    cmovol %ecx, %r12d
-; AVX512-NEXT:    vpextrd $1, %xmm2, %eax
-; AVX512-NEXT:    vpextrd $1, %xmm3, %r13d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r13d, %esi
-; AVX512-NEXT:    addl %eax, %esi
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %r13d
-; AVX512-NEXT:    cmovol %ecx, %r13d
-; AVX512-NEXT:    vextracti32x4 $2, %zmm1, %xmm2
-; AVX512-NEXT:    vpextrd $3, %xmm2, %eax
-; AVX512-NEXT:    vextracti32x4 $2, %zmm0, %xmm3
-; AVX512-NEXT:    vpextrd $3, %xmm3, %r11d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r11d, %edi
-; AVX512-NEXT:    addl %eax, %edi
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %r11d
-; AVX512-NEXT:    cmovol %ecx, %r11d
-; AVX512-NEXT:    vpextrd $2, %xmm2, %eax
-; AVX512-NEXT:    vpextrd $2, %xmm3, %ebx
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %ebx, %edi
-; AVX512-NEXT:    addl %eax, %edi
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %ebx
-; AVX512-NEXT:    cmovol %ecx, %ebx
-; AVX512-NEXT:    vmovd %xmm2, %eax
-; AVX512-NEXT:    vmovd %xmm3, %ebp
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %ebp, %edi
-; AVX512-NEXT:    addl %eax, %edi
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %ebp
-; AVX512-NEXT:    cmovol %ecx, %ebp
-; AVX512-NEXT:    vpextrd $1, %xmm2, %eax
-; AVX512-NEXT:    vpextrd $1, %xmm3, %r10d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r10d, %edi
-; AVX512-NEXT:    addl %eax, %edi
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %r10d
-; AVX512-NEXT:    cmovol %ecx, %r10d
-; AVX512-NEXT:    vextracti32x4 $3, %zmm1, %xmm1
-; AVX512-NEXT:    vpextrd $3, %xmm1, %ecx
-; AVX512-NEXT:    vextracti32x4 $3, %zmm0, %xmm0
-; AVX512-NEXT:    vpextrd $3, %xmm0, %r8d
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %r8d, %eax
-; AVX512-NEXT:    addl %ecx, %eax
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %ecx, %r8d
-; AVX512-NEXT:    cmovol %edi, %r8d
-; AVX512-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX512-NEXT:    vpextrd $2, %xmm0, %edx
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %edx, %ecx
-; AVX512-NEXT:    addl %esi, %ecx
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %esi, %edx
-; AVX512-NEXT:    cmovol %edi, %edx
-; AVX512-NEXT:    vmovd %xmm1, %eax
-; AVX512-NEXT:    vmovd %xmm0, %ecx
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %ecx, %esi
-; AVX512-NEXT:    addl %eax, %esi
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %eax, %ecx
-; AVX512-NEXT:    cmovol %edi, %ecx
-; AVX512-NEXT:    vpextrd $1, %xmm1, %r9d
-; AVX512-NEXT:    vpextrd $1, %xmm0, %eax
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %eax, %esi
-; AVX512-NEXT:    addl %r9d, %esi
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    addl %r9d, %eax
-; AVX512-NEXT:    cmovol %edi, %eax
-; AVX512-NEXT:    vmovd %ecx, %xmm0
-; AVX512-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
-; AVX512-NEXT:    vpinsrd $2, %edx, %xmm0, %xmm0
-; AVX512-NEXT:    vpinsrd $3, %r8d, %xmm0, %xmm0
-; AVX512-NEXT:    vmovd %ebp, %xmm1
-; AVX512-NEXT:    vpinsrd $1, %r10d, %xmm1, %xmm1
-; AVX512-NEXT:    vpinsrd $2, %ebx, %xmm1, %xmm1
-; AVX512-NEXT:    vpinsrd $3, %r11d, %xmm1, %xmm1
-; AVX512-NEXT:    vmovd %r12d, %xmm2
-; AVX512-NEXT:    vpinsrd $1, %r13d, %xmm2, %xmm2
-; AVX512-NEXT:    vpinsrd $2, %r15d, %xmm2, %xmm2
-; AVX512-NEXT:    vpinsrd $3, %r14d, %xmm2, %xmm2
-; AVX512-NEXT:    vmovd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 4-byte Folded Reload
-; AVX512-NEXT:    # xmm3 = mem[0],zero,zero,zero
-; AVX512-NEXT:    vpinsrd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX512-NEXT:    vpinsrd $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX512-NEXT:    vpinsrd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm1
-; AVX512-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512-NEXT:    popq %rbx
-; AVX512-NEXT:    popq %r12
-; AVX512-NEXT:    popq %r13
-; AVX512-NEXT:    popq %r14
-; AVX512-NEXT:    popq %r15
-; AVX512-NEXT:    popq %rbp
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltd %zmm2, %zmm1, %k0
+; AVX512-NEXT:    vpcmpnltd %zmm2, %zmm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpaddd %zmm1, %zmm0, %zmm0
+; AVX512-NEXT:    vpcmpnltd %zmm2, %zmm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandnw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtd %zmm0, %zmm2, %k2
+; AVX512-NEXT:    vpbroadcastd {{.*#+}} zmm1 = [2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648]
+; AVX512-NEXT:    vpbroadcastd {{.*}}(%rip), %zmm1 {%k2}
+; AVX512-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k1}
 ; AVX512-NEXT:    retq
   %z = call <16 x i32> @llvm.sadd.sat.v16i32(<16 x i32> %x, <16 x i32> %y)
   ret <16 x i32> %z
@@ -2818,111 +1669,213 @@
 define <2 x i64> @v2i64(<2 x i64> %x, <2 x i64> %y) nounwind {
 ; SSE2-LABEL: v2i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movq %xmm2, %rax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movq %xmm2, %rcx
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movq %rcx, %rsi
-; SSE2-NEXT:    addq %rax, %rsi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE2-NEXT:    addq %r8, %rdx
-; SSE2-NEXT:    addq %rax, %rcx
-; SSE2-NEXT:    cmovoq %rdx, %rcx
-; SSE2-NEXT:    movq %xmm1, %rax
-; SSE2-NEXT:    movq %xmm0, %rsi
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %rsi, %rdx
-; SSE2-NEXT:    addq %rax, %rdx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    addq %rax, %rsi
-; SSE2-NEXT:    cmovoq %rdi, %rsi
-; SSE2-NEXT:    movq %rsi, %xmm0
-; SSE2-NEXT:    movq %rcx, %xmm1
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
+; SSE2-NEXT:    movdqa %xmm0, %xmm3
+; SSE2-NEXT:    paddq %xmm1, %xmm0
+; SSE2-NEXT:    pxor %xmm2, %xmm1
+; SSE2-NEXT:    movdqa %xmm2, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm1, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm5
+; SSE2-NEXT:    pxor %xmm5, %xmm4
+; SSE2-NEXT:    pxor %xmm2, %xmm3
+; SSE2-NEXT:    movdqa %xmm2, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2-NEXT:    pand %xmm6, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    por %xmm3, %xmm1
+; SSE2-NEXT:    pxor %xmm5, %xmm1
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm3
+; SSE2-NEXT:    movdqa %xmm0, %xmm4
+; SSE2-NEXT:    pxor %xmm2, %xmm4
+; SSE2-NEXT:    movdqa %xmm2, %xmm6
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pand %xmm7, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
+; SSE2-NEXT:    por %xmm2, %xmm4
+; SSE2-NEXT:    pxor %xmm4, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[1,0,3,2]
+; SSE2-NEXT:    pand %xmm5, %xmm1
+; SSE2-NEXT:    pandn %xmm3, %xmm1
+; SSE2-NEXT:    movdqa %xmm4, %xmm2
+; SSE2-NEXT:    pandn {{.*}}(%rip), %xmm2
+; SSE2-NEXT:    pand {{.*}}(%rip), %xmm4
+; SSE2-NEXT:    por %xmm2, %xmm4
+; SSE2-NEXT:    pand %xmm1, %xmm4
+; SSE2-NEXT:    pandn %xmm0, %xmm1
+; SSE2-NEXT:    por %xmm4, %xmm1
+; SSE2-NEXT:    movdqa %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v2i64:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm2, %rax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm2, %rcx
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movq %rcx, %rsi
-; SSSE3-NEXT:    addq %rax, %rsi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSSE3-NEXT:    addq %r8, %rdx
-; SSSE3-NEXT:    addq %rax, %rcx
-; SSSE3-NEXT:    cmovoq %rdx, %rcx
-; SSSE3-NEXT:    movq %xmm1, %rax
-; SSSE3-NEXT:    movq %xmm0, %rsi
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %rsi, %rdx
-; SSSE3-NEXT:    addq %rax, %rdx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    addq %rax, %rsi
-; SSSE3-NEXT:    cmovoq %rdi, %rsi
-; SSSE3-NEXT:    movq %rsi, %xmm0
-; SSSE3-NEXT:    movq %rcx, %xmm1
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
+; SSSE3-NEXT:    movdqa %xmm0, %xmm3
+; SSSE3-NEXT:    paddq %xmm1, %xmm0
+; SSSE3-NEXT:    pxor %xmm2, %xmm1
+; SSSE3-NEXT:    movdqa %xmm2, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm5, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm1, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm5
+; SSSE3-NEXT:    pxor %xmm5, %xmm4
+; SSSE3-NEXT:    pxor %xmm2, %xmm3
+; SSSE3-NEXT:    movdqa %xmm2, %xmm1
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm6, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    por %xmm3, %xmm1
+; SSSE3-NEXT:    pxor %xmm5, %xmm1
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm3
+; SSSE3-NEXT:    movdqa %xmm0, %xmm4
+; SSSE3-NEXT:    pxor %xmm2, %xmm4
+; SSSE3-NEXT:    movdqa %xmm2, %xmm6
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm7, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    por %xmm2, %xmm4
+; SSSE3-NEXT:    pxor %xmm4, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm5, %xmm1
+; SSSE3-NEXT:    pandn %xmm3, %xmm1
+; SSSE3-NEXT:    movdqa %xmm4, %xmm2
+; SSSE3-NEXT:    pandn {{.*}}(%rip), %xmm2
+; SSSE3-NEXT:    pand {{.*}}(%rip), %xmm4
+; SSSE3-NEXT:    por %xmm2, %xmm4
+; SSSE3-NEXT:    pand %xmm1, %xmm4
+; SSSE3-NEXT:    pandn %xmm0, %xmm1
+; SSSE3-NEXT:    por %xmm4, %xmm1
+; SSSE3-NEXT:    movdqa %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v2i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movq %xmm1, %rax
-; SSE41-NEXT:    movq %xmm0, %rcx
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movq %rcx, %rsi
-; SSE41-NEXT:    addq %rax, %rsi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE41-NEXT:    addq %r8, %rdx
-; SSE41-NEXT:    addq %rax, %rcx
-; SSE41-NEXT:    cmovoq %rdx, %rcx
-; SSE41-NEXT:    pextrq $1, %xmm1, %rax
-; SSE41-NEXT:    pextrq $1, %xmm0, %rsi
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %rsi, %rdx
-; SSE41-NEXT:    addq %rax, %rdx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    addq %rax, %rsi
-; SSE41-NEXT:    cmovoq %rdi, %rsi
-; SSE41-NEXT:    movq %rsi, %xmm1
-; SSE41-NEXT:    movq %rcx, %xmm0
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE41-NEXT:    movdqa %xmm0, %xmm2
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm2, %xmm3
+; SSE41-NEXT:    paddq %xmm1, %xmm2
+; SSE41-NEXT:    pxor %xmm0, %xmm1
+; SSE41-NEXT:    movdqa %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm1
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm1, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE41-NEXT:    pxor %xmm1, %xmm4
+; SSE41-NEXT:    pxor %xmm0, %xmm3
+; SSE41-NEXT:    movdqa %xmm0, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE41-NEXT:    pand %xmm6, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE41-NEXT:    por %xmm3, %xmm5
+; SSE41-NEXT:    pxor %xmm1, %xmm5
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm4
+; SSE41-NEXT:    movdqa %xmm2, %xmm3
+; SSE41-NEXT:    pxor %xmm0, %xmm3
+; SSE41-NEXT:    movdqa %xmm0, %xmm6
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE41-NEXT:    pand %xmm7, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
+; SSE41-NEXT:    por %xmm3, %xmm0
+; SSE41-NEXT:    pxor %xmm0, %xmm1
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm1
+; SSE41-NEXT:    pandn %xmm4, %xmm1
+; SSE41-NEXT:    movapd {{.*#+}} xmm3 = [9223372036854775808,9223372036854775808]
+; SSE41-NEXT:    blendvpd %xmm0, {{.*}}(%rip), %xmm3
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm2
+; SSE41-NEXT:    movapd %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX-LABEL: v2i64:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovq %xmm1, %rax
-; AVX-NEXT:    vmovq %xmm0, %rcx
-; AVX-NEXT:    xorl %edx, %edx
-; AVX-NEXT:    movq %rcx, %rsi
-; AVX-NEXT:    addq %rax, %rsi
-; AVX-NEXT:    setns %dl
-; AVX-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX-NEXT:    addq %r8, %rdx
-; AVX-NEXT:    addq %rax, %rcx
-; AVX-NEXT:    cmovoq %rdx, %rcx
-; AVX-NEXT:    vpextrq $1, %xmm1, %rax
-; AVX-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX-NEXT:    xorl %edi, %edi
-; AVX-NEXT:    movq %rsi, %rdx
-; AVX-NEXT:    addq %rax, %rdx
-; AVX-NEXT:    setns %dil
-; AVX-NEXT:    addq %r8, %rdi
-; AVX-NEXT:    addq %rax, %rsi
-; AVX-NEXT:    cmovoq %rdi, %rsi
-; AVX-NEXT:    vmovq %rsi, %xmm0
-; AVX-NEXT:    vmovq %rcx, %xmm1
-; AVX-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX-NEXT:    retq
+; AVX1-LABEL: v2i64:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm5, %xmm3
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm1
+; AVX1-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqq %xmm1, %xmm5, %xmm1
+; AVX1-NEXT:    vpandn %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vmovapd {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
+; AVX1-NEXT:    vblendvpd %xmm0, {{.*}}(%rip), %xmm2, %xmm2
+; AVX1-NEXT:    vblendvpd %xmm1, %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: v2i64:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm3
+; AVX2-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX2-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX2-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm5
+; AVX2-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX2-NEXT:    vpcmpeqq %xmm3, %xmm5, %xmm3
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm1
+; AVX2-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqq %xmm1, %xmm5, %xmm1
+; AVX2-NEXT:    vpandn %xmm3, %xmm1, %xmm1
+; AVX2-NEXT:    vmovapd {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
+; AVX2-NEXT:    vblendvpd %xmm0, {{.*}}(%rip), %xmm2, %xmm2
+; AVX2-NEXT:    vblendvpd %xmm1, %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    retq
+;
+; AVX512-LABEL: v2i64:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltq %xmm2, %xmm1, %k0
+; AVX512-NEXT:    vpcmpnltq %xmm2, %xmm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpaddq %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vpcmpnltq %xmm2, %xmm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandnw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtq %xmm0, %xmm2, %k2
+; AVX512-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
+; AVX512-NEXT:    vmovdqa64 {{.*}}(%rip), %xmm1 {%k2}
+; AVX512-NEXT:    vmovdqa64 %xmm1, %xmm0 {%k1}
+; AVX512-NEXT:    retq
   %z = call <2 x i64> @llvm.sadd.sat.v2i64(<2 x i64> %x, <2 x i64> %y)
   ret <2 x i64> %z
 }
@@ -2930,301 +1883,371 @@
 define <4 x i64> @v4i64(<4 x i64> %x, <4 x i64> %y) nounwind {
 ; SSE2-LABEL: v4i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[2,3,0,1]
-; SSE2-NEXT:    movq %xmm4, %rcx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movq %xmm4, %r9
-; SSE2-NEXT:    xorl %esi, %esi
-; SSE2-NEXT:    movq %r9, %rdx
-; SSE2-NEXT:    addq %rcx, %rdx
-; SSE2-NEXT:    setns %sil
-; SSE2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE2-NEXT:    addq %r8, %rsi
-; SSE2-NEXT:    addq %rcx, %r9
-; SSE2-NEXT:    cmovoq %rsi, %r9
-; SSE2-NEXT:    movq %xmm3, %rsi
-; SSE2-NEXT:    movq %xmm1, %rdx
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %rdx, %rcx
-; SSE2-NEXT:    addq %rsi, %rcx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    addq %rsi, %rdx
-; SSE2-NEXT:    cmovoq %rdi, %rdx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %rcx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %rax
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %rax, %rsi
-; SSE2-NEXT:    addq %rcx, %rsi
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    addq %rcx, %rax
-; SSE2-NEXT:    cmovoq %rdi, %rax
-; SSE2-NEXT:    movq %xmm2, %r10
-; SSE2-NEXT:    movq %xmm0, %rcx
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %rcx, %rsi
-; SSE2-NEXT:    addq %r10, %rsi
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    addq %r10, %rcx
-; SSE2-NEXT:    cmovoq %rdi, %rcx
-; SSE2-NEXT:    movq %rcx, %xmm0
-; SSE2-NEXT:    movq %rax, %xmm1
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE2-NEXT:    movq %rdx, %xmm1
-; SSE2-NEXT:    movq %r9, %xmm2
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; SSE2-NEXT:    movdqa %xmm0, %xmm9
+; SSE2-NEXT:    movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
+; SSE2-NEXT:    paddq %xmm2, %xmm9
+; SSE2-NEXT:    pxor %xmm10, %xmm2
+; SSE2-NEXT:    movdqa %xmm10, %xmm6
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2-NEXT:    pand %xmm7, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,1,3,3]
+; SSE2-NEXT:    por %xmm2, %xmm7
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm6
+; SSE2-NEXT:    pxor %xmm6, %xmm7
+; SSE2-NEXT:    pxor %xmm10, %xmm0
+; SSE2-NEXT:    movdqa %xmm10, %xmm2
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pand %xmm4, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2-NEXT:    por %xmm0, %xmm2
+; SSE2-NEXT:    pxor %xmm6, %xmm2
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[1,0,3,2]
+; SSE2-NEXT:    pand %xmm7, %xmm4
+; SSE2-NEXT:    movdqa %xmm9, %xmm0
+; SSE2-NEXT:    pxor %xmm10, %xmm0
+; SSE2-NEXT:    movdqa %xmm10, %xmm7
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
+; SSE2-NEXT:    por %xmm0, %xmm5
+; SSE2-NEXT:    movdqa %xmm5, %xmm7
+; SSE2-NEXT:    pxor %xmm6, %xmm7
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[1,0,3,2]
+; SSE2-NEXT:    pand %xmm7, %xmm0
+; SSE2-NEXT:    pandn %xmm4, %xmm0
+; SSE2-NEXT:    movdqa {{.*#+}} xmm11 = [9223372036854775808,9223372036854775808]
+; SSE2-NEXT:    movdqa %xmm5, %xmm2
+; SSE2-NEXT:    pandn %xmm11, %xmm2
+; SSE2-NEXT:    movdqa {{.*#+}} xmm8 = [9223372036854775807,9223372036854775807]
+; SSE2-NEXT:    pand %xmm8, %xmm5
+; SSE2-NEXT:    por %xmm2, %xmm5
+; SSE2-NEXT:    pand %xmm0, %xmm5
+; SSE2-NEXT:    pandn %xmm9, %xmm0
+; SSE2-NEXT:    por %xmm5, %xmm0
+; SSE2-NEXT:    movdqa %xmm1, %xmm2
+; SSE2-NEXT:    paddq %xmm3, %xmm1
+; SSE2-NEXT:    pxor %xmm10, %xmm3
+; SSE2-NEXT:    movdqa %xmm10, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm3, %xmm4
+; SSE2-NEXT:    pxor %xmm6, %xmm4
+; SSE2-NEXT:    pxor %xmm10, %xmm2
+; SSE2-NEXT:    movdqa %xmm10, %xmm3
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2-NEXT:    por %xmm2, %xmm3
+; SSE2-NEXT:    pxor %xmm6, %xmm3
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm5
+; SSE2-NEXT:    movdqa %xmm1, %xmm2
+; SSE2-NEXT:    pxor %xmm10, %xmm2
+; SSE2-NEXT:    movdqa %xmm10, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2-NEXT:    pand %xmm7, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm2, %xmm4
+; SSE2-NEXT:    pxor %xmm4, %xmm6
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,0,3,2]
+; SSE2-NEXT:    pand %xmm6, %xmm2
+; SSE2-NEXT:    pandn %xmm5, %xmm2
+; SSE2-NEXT:    movdqa %xmm4, %xmm3
+; SSE2-NEXT:    pandn %xmm11, %xmm3
+; SSE2-NEXT:    pand %xmm8, %xmm4
+; SSE2-NEXT:    por %xmm3, %xmm4
+; SSE2-NEXT:    pand %xmm2, %xmm4
+; SSE2-NEXT:    pandn %xmm1, %xmm2
+; SSE2-NEXT:    por %xmm4, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm1
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v4i64:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm4, %rcx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm4, %r9
-; SSSE3-NEXT:    xorl %esi, %esi
-; SSSE3-NEXT:    movq %r9, %rdx
-; SSSE3-NEXT:    addq %rcx, %rdx
-; SSSE3-NEXT:    setns %sil
-; SSSE3-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSSE3-NEXT:    addq %r8, %rsi
-; SSSE3-NEXT:    addq %rcx, %r9
-; SSSE3-NEXT:    cmovoq %rsi, %r9
-; SSSE3-NEXT:    movq %xmm3, %rsi
-; SSSE3-NEXT:    movq %xmm1, %rdx
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %rdx, %rcx
-; SSSE3-NEXT:    addq %rsi, %rcx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    addq %rsi, %rdx
-; SSSE3-NEXT:    cmovoq %rdi, %rdx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %rcx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %rax
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %rax, %rsi
-; SSSE3-NEXT:    addq %rcx, %rsi
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    addq %rcx, %rax
-; SSSE3-NEXT:    cmovoq %rdi, %rax
-; SSSE3-NEXT:    movq %xmm2, %r10
-; SSSE3-NEXT:    movq %xmm0, %rcx
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %rcx, %rsi
-; SSSE3-NEXT:    addq %r10, %rsi
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    addq %r10, %rcx
-; SSSE3-NEXT:    cmovoq %rdi, %rcx
-; SSSE3-NEXT:    movq %rcx, %xmm0
-; SSSE3-NEXT:    movq %rax, %xmm1
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSSE3-NEXT:    movq %rdx, %xmm1
-; SSSE3-NEXT:    movq %r9, %xmm2
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; SSSE3-NEXT:    movdqa %xmm0, %xmm9
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
+; SSSE3-NEXT:    paddq %xmm2, %xmm9
+; SSSE3-NEXT:    pxor %xmm10, %xmm2
+; SSSE3-NEXT:    movdqa %xmm10, %xmm6
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm7, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    por %xmm2, %xmm7
+; SSSE3-NEXT:    pcmpeqd %xmm6, %xmm6
+; SSSE3-NEXT:    pxor %xmm6, %xmm7
+; SSSE3-NEXT:    pxor %xmm10, %xmm0
+; SSSE3-NEXT:    movdqa %xmm10, %xmm2
+; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm4, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSSE3-NEXT:    por %xmm0, %xmm2
+; SSSE3-NEXT:    pxor %xmm6, %xmm2
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm7, %xmm4
+; SSSE3-NEXT:    movdqa %xmm9, %xmm0
+; SSSE3-NEXT:    pxor %xmm10, %xmm0
+; SSSE3-NEXT:    movdqa %xmm10, %xmm7
+; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm5, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
+; SSSE3-NEXT:    por %xmm0, %xmm5
+; SSSE3-NEXT:    movdqa %xmm5, %xmm7
+; SSSE3-NEXT:    pxor %xmm6, %xmm7
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm7, %xmm0
+; SSSE3-NEXT:    pandn %xmm4, %xmm0
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm11 = [9223372036854775808,9223372036854775808]
+; SSSE3-NEXT:    movdqa %xmm5, %xmm2
+; SSSE3-NEXT:    pandn %xmm11, %xmm2
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm8 = [9223372036854775807,9223372036854775807]
+; SSSE3-NEXT:    pand %xmm8, %xmm5
+; SSSE3-NEXT:    por %xmm2, %xmm5
+; SSSE3-NEXT:    pand %xmm0, %xmm5
+; SSSE3-NEXT:    pandn %xmm9, %xmm0
+; SSSE3-NEXT:    por %xmm5, %xmm0
+; SSSE3-NEXT:    movdqa %xmm1, %xmm2
+; SSSE3-NEXT:    paddq %xmm3, %xmm1
+; SSSE3-NEXT:    pxor %xmm10, %xmm3
+; SSSE3-NEXT:    movdqa %xmm10, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm5, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm3, %xmm4
+; SSSE3-NEXT:    pxor %xmm6, %xmm4
+; SSSE3-NEXT:    pxor %xmm10, %xmm2
+; SSSE3-NEXT:    movdqa %xmm10, %xmm3
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm5, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSSE3-NEXT:    por %xmm2, %xmm3
+; SSSE3-NEXT:    pxor %xmm6, %xmm3
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm5
+; SSSE3-NEXT:    movdqa %xmm1, %xmm2
+; SSSE3-NEXT:    pxor %xmm10, %xmm2
+; SSSE3-NEXT:    movdqa %xmm10, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm7, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm2, %xmm4
+; SSSE3-NEXT:    pxor %xmm4, %xmm6
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm6, %xmm2
+; SSSE3-NEXT:    pandn %xmm5, %xmm2
+; SSSE3-NEXT:    movdqa %xmm4, %xmm3
+; SSSE3-NEXT:    pandn %xmm11, %xmm3
+; SSSE3-NEXT:    pand %xmm8, %xmm4
+; SSSE3-NEXT:    por %xmm3, %xmm4
+; SSSE3-NEXT:    pand %xmm2, %xmm4
+; SSSE3-NEXT:    pandn %xmm1, %xmm2
+; SSSE3-NEXT:    por %xmm4, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm1
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v4i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movq %xmm3, %rcx
-; SSE41-NEXT:    movq %xmm1, %r9
-; SSE41-NEXT:    xorl %esi, %esi
-; SSE41-NEXT:    movq %r9, %rdx
-; SSE41-NEXT:    addq %rcx, %rdx
-; SSE41-NEXT:    setns %sil
-; SSE41-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE41-NEXT:    addq %r8, %rsi
-; SSE41-NEXT:    addq %rcx, %r9
-; SSE41-NEXT:    cmovoq %rsi, %r9
-; SSE41-NEXT:    pextrq $1, %xmm3, %rsi
-; SSE41-NEXT:    pextrq $1, %xmm1, %rdx
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %rdx, %rcx
-; SSE41-NEXT:    addq %rsi, %rcx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    addq %rsi, %rdx
-; SSE41-NEXT:    cmovoq %rdi, %rdx
-; SSE41-NEXT:    movq %xmm2, %rcx
-; SSE41-NEXT:    movq %xmm0, %rax
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %rax, %rsi
-; SSE41-NEXT:    addq %rcx, %rsi
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    addq %rcx, %rax
-; SSE41-NEXT:    cmovoq %rdi, %rax
-; SSE41-NEXT:    pextrq $1, %xmm2, %r10
-; SSE41-NEXT:    pextrq $1, %xmm0, %rcx
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %rcx, %rsi
-; SSE41-NEXT:    addq %r10, %rsi
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    addq %r10, %rcx
-; SSE41-NEXT:    cmovoq %rdi, %rcx
-; SSE41-NEXT:    movq %rcx, %xmm1
-; SSE41-NEXT:    movq %rax, %xmm0
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE41-NEXT:    movq %rdx, %xmm2
-; SSE41-NEXT:    movq %r9, %xmm1
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; SSE41-NEXT:    movdqa %xmm0, %xmm9
+; SSE41-NEXT:    movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
+; SSE41-NEXT:    paddq %xmm2, %xmm9
+; SSE41-NEXT:    pxor %xmm10, %xmm2
+; SSE41-NEXT:    movdqa %xmm10, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE41-NEXT:    pand %xmm7, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm2, %xmm7
+; SSE41-NEXT:    pcmpeqd %xmm4, %xmm4
+; SSE41-NEXT:    pxor %xmm4, %xmm7
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm10, %xmm2
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[1,1,3,3]
+; SSE41-NEXT:    por %xmm0, %xmm5
+; SSE41-NEXT:    pxor %xmm4, %xmm5
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm7
+; SSE41-NEXT:    movdqa %xmm9, %xmm0
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm10, %xmm2
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm8 = xmm2[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm8, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
+; SSE41-NEXT:    por %xmm6, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm2
+; SSE41-NEXT:    pxor %xmm4, %xmm2
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm2
+; SSE41-NEXT:    pandn %xmm7, %xmm2
+; SSE41-NEXT:    movapd {{.*#+}} xmm8 = [9223372036854775807,9223372036854775807]
+; SSE41-NEXT:    movapd {{.*#+}} xmm11 = [9223372036854775808,9223372036854775808]
+; SSE41-NEXT:    movapd %xmm11, %xmm5
+; SSE41-NEXT:    blendvpd %xmm0, %xmm8, %xmm5
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm5, %xmm9
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    paddq %xmm3, %xmm1
+; SSE41-NEXT:    pxor %xmm10, %xmm3
+; SSE41-NEXT:    movdqa %xmm10, %xmm2
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE41-NEXT:    por %xmm3, %xmm2
+; SSE41-NEXT:    pxor %xmm4, %xmm2
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm10, %xmm3
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE41-NEXT:    por %xmm0, %xmm3
+; SSE41-NEXT:    pxor %xmm4, %xmm3
+; SSE41-NEXT:    pcmpeqq %xmm3, %xmm2
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm10, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm6, %xmm7
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3]
+; SSE41-NEXT:    por %xmm7, %xmm0
+; SSE41-NEXT:    pxor %xmm0, %xmm4
+; SSE41-NEXT:    pcmpeqq %xmm3, %xmm4
+; SSE41-NEXT:    pandn %xmm2, %xmm4
+; SSE41-NEXT:    blendvpd %xmm0, %xmm8, %xmm11
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm11, %xmm1
+; SSE41-NEXT:    movapd %xmm9, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: v4i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovq %xmm1, %rdx
-; AVX1-NEXT:    vmovq %xmm0, %r9
-; AVX1-NEXT:    xorl %esi, %esi
-; AVX1-NEXT:    movq %r9, %rcx
-; AVX1-NEXT:    addq %rdx, %rcx
-; AVX1-NEXT:    setns %sil
-; AVX1-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX1-NEXT:    addq %r8, %rsi
-; AVX1-NEXT:    addq %rdx, %r9
-; AVX1-NEXT:    cmovoq %rsi, %r9
-; AVX1-NEXT:    vpextrq $1, %xmm1, %rsi
-; AVX1-NEXT:    vpextrq $1, %xmm0, %rcx
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movq %rcx, %rdx
-; AVX1-NEXT:    addq %rsi, %rdx
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addq %r8, %rdi
-; AVX1-NEXT:    addq %rsi, %rcx
-; AVX1-NEXT:    cmovoq %rdi, %rcx
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vmovq %xmm1, %rdx
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vmovq %xmm0, %rax
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movq %rax, %rsi
-; AVX1-NEXT:    addq %rdx, %rsi
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addq %r8, %rdi
-; AVX1-NEXT:    addq %rdx, %rax
-; AVX1-NEXT:    cmovoq %rdi, %rax
-; AVX1-NEXT:    vpextrq $1, %xmm1, %r10
-; AVX1-NEXT:    vpextrq $1, %xmm0, %rdx
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movq %rdx, %rsi
-; AVX1-NEXT:    addq %r10, %rsi
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addq %r8, %rdi
-; AVX1-NEXT:    addq %r10, %rdx
-; AVX1-NEXT:    cmovoq %rdi, %rdx
-; AVX1-NEXT:    vmovq %rdx, %xmm0
-; AVX1-NEXT:    vmovq %rax, %xmm1
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX1-NEXT:    vmovq %rcx, %xmm1
-; AVX1-NEXT:    vmovq %r9, %xmm2
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm4
+; AVX1-NEXT:    vpcmpeqd %xmm5, %xmm5, %xmm5
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm4
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm6
+; AVX1-NEXT:    vpcmpgtq %xmm6, %xmm3, %xmm7
+; AVX1-NEXT:    vpxor %xmm5, %xmm7, %xmm7
+; AVX1-NEXT:    vpcmpeqq %xmm4, %xmm7, %xmm8
+; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm3, %xmm4
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm9
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm3, %xmm4
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm4
+; AVX1-NEXT:    vpcmpeqq %xmm9, %xmm4, %xmm9
+; AVX1-NEXT:    vinsertf128 $1, %xmm8, %ymm9, %ymm8
+; AVX1-NEXT:    vpaddq %xmm2, %xmm6, %xmm9
+; AVX1-NEXT:    vpcmpgtq %xmm9, %xmm3, %xmm6
+; AVX1-NEXT:    vpxor %xmm5, %xmm6, %xmm2
+; AVX1-NEXT:    vpcmpeqq %xmm2, %xmm7, %xmm2
+; AVX1-NEXT:    vpxor %xmm5, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm3, %xmm1
+; AVX1-NEXT:    vpxor %xmm5, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqq %xmm1, %xmm4, %xmm1
+; AVX1-NEXT:    vpxor %xmm5, %xmm1, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
+; AVX1-NEXT:    vandpd %ymm1, %ymm8, %ymm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm2
+; AVX1-NEXT:    vmovapd {{.*#+}} ymm3 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
+; AVX1-NEXT:    vblendvpd %ymm2, {{.*}}(%rip), %ymm3, %ymm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm9, %ymm0, %ymm0
+; AVX1-NEXT:    vblendvpd %ymm1, %ymm2, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: v4i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovq %xmm1, %rdx
-; AVX2-NEXT:    vmovq %xmm0, %r9
-; AVX2-NEXT:    xorl %esi, %esi
-; AVX2-NEXT:    movq %r9, %rcx
-; AVX2-NEXT:    addq %rdx, %rcx
-; AVX2-NEXT:    setns %sil
-; AVX2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX2-NEXT:    addq %r8, %rsi
-; AVX2-NEXT:    addq %rdx, %r9
-; AVX2-NEXT:    cmovoq %rsi, %r9
-; AVX2-NEXT:    vpextrq $1, %xmm1, %rsi
-; AVX2-NEXT:    vpextrq $1, %xmm0, %rcx
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movq %rcx, %rdx
-; AVX2-NEXT:    addq %rsi, %rdx
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addq %r8, %rdi
-; AVX2-NEXT:    addq %rsi, %rcx
-; AVX2-NEXT:    cmovoq %rdi, %rcx
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; AVX2-NEXT:    vmovq %xmm1, %rdx
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; AVX2-NEXT:    vmovq %xmm0, %rax
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movq %rax, %rsi
-; AVX2-NEXT:    addq %rdx, %rsi
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addq %r8, %rdi
-; AVX2-NEXT:    addq %rdx, %rax
-; AVX2-NEXT:    cmovoq %rdi, %rax
-; AVX2-NEXT:    vpextrq $1, %xmm1, %r10
-; AVX2-NEXT:    vpextrq $1, %xmm0, %rdx
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movq %rdx, %rsi
-; AVX2-NEXT:    addq %r10, %rsi
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addq %r8, %rdi
-; AVX2-NEXT:    addq %r10, %rdx
-; AVX2-NEXT:    cmovoq %rdi, %rdx
-; AVX2-NEXT:    vmovq %rdx, %xmm0
-; AVX2-NEXT:    vmovq %rax, %xmm1
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX2-NEXT:    vmovq %rcx, %xmm1
-; AVX2-NEXT:    vmovq %r9, %xmm2
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpgtq %ymm1, %ymm2, %ymm3
+; AVX2-NEXT:    vpcmpeqd %ymm4, %ymm4, %ymm4
+; AVX2-NEXT:    vpxor %ymm4, %ymm3, %ymm3
+; AVX2-NEXT:    vpcmpgtq %ymm0, %ymm2, %ymm5
+; AVX2-NEXT:    vpxor %ymm4, %ymm5, %ymm5
+; AVX2-NEXT:    vpcmpeqq %ymm3, %ymm5, %ymm3
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpgtq %ymm0, %ymm2, %ymm1
+; AVX2-NEXT:    vpxor %ymm4, %ymm1, %ymm1
+; AVX2-NEXT:    vpcmpeqq %ymm1, %ymm5, %ymm1
+; AVX2-NEXT:    vpandn %ymm3, %ymm1, %ymm1
+; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm2 = [9223372036854775807,9223372036854775807,9223372036854775807,9223372036854775807]
+; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm3 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
+; AVX2-NEXT:    vblendvpd %ymm0, %ymm2, %ymm3, %ymm2
+; AVX2-NEXT:    vblendvpd %ymm1, %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: v4i64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vmovq %xmm1, %rdx
-; AVX512-NEXT:    vmovq %xmm0, %r9
-; AVX512-NEXT:    xorl %esi, %esi
-; AVX512-NEXT:    movq %r9, %rcx
-; AVX512-NEXT:    addq %rdx, %rcx
-; AVX512-NEXT:    setns %sil
-; AVX512-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX512-NEXT:    addq %r8, %rsi
-; AVX512-NEXT:    addq %rdx, %r9
-; AVX512-NEXT:    cmovoq %rsi, %r9
-; AVX512-NEXT:    vpextrq $1, %xmm1, %rsi
-; AVX512-NEXT:    vpextrq $1, %xmm0, %rcx
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movq %rcx, %rdx
-; AVX512-NEXT:    addq %rsi, %rdx
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addq %r8, %rdi
-; AVX512-NEXT:    addq %rsi, %rcx
-; AVX512-NEXT:    cmovoq %rdi, %rcx
-; AVX512-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; AVX512-NEXT:    vmovq %xmm1, %rdx
-; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; AVX512-NEXT:    vmovq %xmm0, %rax
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movq %rax, %rsi
-; AVX512-NEXT:    addq %rdx, %rsi
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addq %r8, %rdi
-; AVX512-NEXT:    addq %rdx, %rax
-; AVX512-NEXT:    cmovoq %rdi, %rax
-; AVX512-NEXT:    vpextrq $1, %xmm1, %r10
-; AVX512-NEXT:    vpextrq $1, %xmm0, %rdx
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movq %rdx, %rsi
-; AVX512-NEXT:    addq %r10, %rsi
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addq %r8, %rdi
-; AVX512-NEXT:    addq %r10, %rdx
-; AVX512-NEXT:    cmovoq %rdi, %rdx
-; AVX512-NEXT:    vmovq %rdx, %xmm0
-; AVX512-NEXT:    vmovq %rax, %xmm1
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX512-NEXT:    vmovq %rcx, %xmm1
-; AVX512-NEXT:    vmovq %r9, %xmm2
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltq %ymm2, %ymm1, %k0
+; AVX512-NEXT:    vpcmpnltq %ymm2, %ymm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpaddq %ymm1, %ymm0, %ymm0
+; AVX512-NEXT:    vpcmpnltq %ymm2, %ymm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandnw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtq %ymm0, %ymm2, %k2
+; AVX512-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
+; AVX512-NEXT:    vpbroadcastq {{.*}}(%rip), %ymm1 {%k2}
+; AVX512-NEXT:    vmovdqa64 %ymm1, %ymm0 {%k1}
 ; AVX512-NEXT:    retq
   %z = call <4 x i64> @llvm.sadd.sat.v4i64(<4 x i64> %x, <4 x i64> %y)
   ret <4 x i64> %z
@@ -3233,609 +2256,692 @@
 define <8 x i64> @v8i64(<8 x i64> %x, <8 x i64> %y) nounwind {
 ; SSE2-LABEL: v8i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pushq %r15
-; SSE2-NEXT:    pushq %r14
-; SSE2-NEXT:    pushq %rbx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[2,3,0,1]
-; SSE2-NEXT:    movq %xmm8, %rax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movq %xmm8, %r9
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movq %r9, %rdx
-; SSE2-NEXT:    addq %rax, %rdx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE2-NEXT:    addq %r8, %rcx
-; SSE2-NEXT:    addq %rax, %r9
-; SSE2-NEXT:    cmovoq %rcx, %r9
-; SSE2-NEXT:    movq %xmm5, %rax
-; SSE2-NEXT:    movq %xmm1, %r10
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movq %r10, %rsi
-; SSE2-NEXT:    addq %rax, %rsi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    addq %r8, %rdx
-; SSE2-NEXT:    addq %rax, %r10
-; SSE2-NEXT:    cmovoq %rdx, %r10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %rax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %r11
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movq %r11, %rdi
-; SSE2-NEXT:    addq %rax, %rdi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    addq %r8, %rdx
-; SSE2-NEXT:    addq %rax, %r11
-; SSE2-NEXT:    cmovoq %rdx, %r11
-; SSE2-NEXT:    movq %xmm6, %rax
-; SSE2-NEXT:    movq %xmm2, %r14
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movq %r14, %rdi
-; SSE2-NEXT:    addq %rax, %rdi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    addq %r8, %rdx
-; SSE2-NEXT:    addq %rax, %r14
-; SSE2-NEXT:    cmovoq %rdx, %r14
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %rax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %r15
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %r15, %rdx
-; SSE2-NEXT:    addq %rax, %rdx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    addq %rax, %r15
-; SSE2-NEXT:    cmovoq %rdi, %r15
-; SSE2-NEXT:    movq %xmm7, %rdx
-; SSE2-NEXT:    movq %xmm3, %rax
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %rax, %rbx
-; SSE2-NEXT:    addq %rdx, %rbx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    addq %rdx, %rax
-; SSE2-NEXT:    cmovoq %rdi, %rax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %rdx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %rcx
-; SSE2-NEXT:    xorl %ebx, %ebx
-; SSE2-NEXT:    movq %rcx, %rdi
-; SSE2-NEXT:    addq %rdx, %rdi
-; SSE2-NEXT:    setns %bl
-; SSE2-NEXT:    addq %r8, %rbx
-; SSE2-NEXT:    addq %rdx, %rcx
-; SSE2-NEXT:    cmovoq %rbx, %rcx
-; SSE2-NEXT:    movq %xmm4, %rdx
-; SSE2-NEXT:    movq %xmm0, %rsi
-; SSE2-NEXT:    xorl %ebx, %ebx
-; SSE2-NEXT:    movq %rsi, %rdi
-; SSE2-NEXT:    addq %rdx, %rdi
-; SSE2-NEXT:    setns %bl
-; SSE2-NEXT:    addq %r8, %rbx
-; SSE2-NEXT:    addq %rdx, %rsi
-; SSE2-NEXT:    cmovoq %rbx, %rsi
-; SSE2-NEXT:    movq %rsi, %xmm0
-; SSE2-NEXT:    movq %rcx, %xmm1
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE2-NEXT:    movq %r10, %xmm1
-; SSE2-NEXT:    movq %r9, %xmm2
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; SSE2-NEXT:    movq %r14, %xmm2
-; SSE2-NEXT:    movq %r11, %xmm3
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
-; SSE2-NEXT:    movq %rax, %xmm3
-; SSE2-NEXT:    movq %r15, %xmm4
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm4[0]
-; SSE2-NEXT:    popq %rbx
-; SSE2-NEXT:    popq %r14
-; SSE2-NEXT:    popq %r15
+; SSE2-NEXT:    movdqa %xmm1, %xmm8
+; SSE2-NEXT:    movdqa %xmm0, %xmm13
+; SSE2-NEXT:    movdqa {{.*#+}} xmm9 = [2147483648,2147483648]
+; SSE2-NEXT:    paddq %xmm4, %xmm13
+; SSE2-NEXT:    pxor %xmm9, %xmm4
+; SSE2-NEXT:    movdqa %xmm9, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm1[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pand %xmm10, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    por %xmm4, %xmm1
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm10
+; SSE2-NEXT:    pxor %xmm10, %xmm1
+; SSE2-NEXT:    pxor %xmm9, %xmm0
+; SSE2-NEXT:    movdqa %xmm9, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pand %xmm11, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm14 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm0, %xmm14
+; SSE2-NEXT:    pxor %xmm10, %xmm14
+; SSE2-NEXT:    pcmpeqd %xmm14, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm1[1,0,3,2]
+; SSE2-NEXT:    pand %xmm1, %xmm11
+; SSE2-NEXT:    movdqa %xmm13, %xmm0
+; SSE2-NEXT:    pxor %xmm9, %xmm0
+; SSE2-NEXT:    movdqa %xmm9, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm1[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pand %xmm12, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    por %xmm0, %xmm1
+; SSE2-NEXT:    movdqa %xmm1, %xmm4
+; SSE2-NEXT:    pxor %xmm10, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm14, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm0
+; SSE2-NEXT:    pandn %xmm11, %xmm0
+; SSE2-NEXT:    movdqa {{.*#+}} xmm11 = [9223372036854775808,9223372036854775808]
+; SSE2-NEXT:    movdqa %xmm1, %xmm4
+; SSE2-NEXT:    pandn %xmm11, %xmm4
+; SSE2-NEXT:    movdqa {{.*#+}} xmm12 = [9223372036854775807,9223372036854775807]
+; SSE2-NEXT:    pand %xmm12, %xmm1
+; SSE2-NEXT:    por %xmm4, %xmm1
+; SSE2-NEXT:    pand %xmm0, %xmm1
+; SSE2-NEXT:    pandn %xmm13, %xmm0
+; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    movdqa %xmm8, %xmm1
+; SSE2-NEXT:    paddq %xmm5, %xmm8
+; SSE2-NEXT:    pxor %xmm9, %xmm5
+; SSE2-NEXT:    movdqa %xmm9, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE2-NEXT:    pand %xmm13, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm5, %xmm4
+; SSE2-NEXT:    pxor %xmm10, %xmm4
+; SSE2-NEXT:    pxor %xmm9, %xmm1
+; SSE2-NEXT:    movdqa %xmm9, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm5[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pand %xmm13, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm15 = xmm5[1,1,3,3]
+; SSE2-NEXT:    por %xmm1, %xmm15
+; SSE2-NEXT:    pxor %xmm10, %xmm15
+; SSE2-NEXT:    pcmpeqd %xmm15, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm13
+; SSE2-NEXT:    movdqa %xmm8, %xmm1
+; SSE2-NEXT:    pxor %xmm9, %xmm1
+; SSE2-NEXT:    movdqa %xmm9, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm14 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pand %xmm14, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm1, %xmm4
+; SSE2-NEXT:    movdqa %xmm4, %xmm5
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm15, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[1,0,3,2]
+; SSE2-NEXT:    pand %xmm5, %xmm1
+; SSE2-NEXT:    pandn %xmm13, %xmm1
+; SSE2-NEXT:    movdqa %xmm4, %xmm5
+; SSE2-NEXT:    pandn %xmm11, %xmm5
+; SSE2-NEXT:    pand %xmm12, %xmm4
+; SSE2-NEXT:    por %xmm5, %xmm4
+; SSE2-NEXT:    pand %xmm1, %xmm4
+; SSE2-NEXT:    pandn %xmm8, %xmm1
+; SSE2-NEXT:    por %xmm4, %xmm1
+; SSE2-NEXT:    movdqa %xmm2, %xmm4
+; SSE2-NEXT:    paddq %xmm6, %xmm2
+; SSE2-NEXT:    pxor %xmm9, %xmm6
+; SSE2-NEXT:    movdqa %xmm9, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm6, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE2-NEXT:    pand %xmm8, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE2-NEXT:    por %xmm6, %xmm5
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pxor %xmm9, %xmm4
+; SSE2-NEXT:    movdqa %xmm9, %xmm6
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pand %xmm8, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE2-NEXT:    por %xmm4, %xmm6
+; SSE2-NEXT:    pxor %xmm10, %xmm6
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[1,0,3,2]
+; SSE2-NEXT:    pand %xmm5, %xmm8
+; SSE2-NEXT:    movdqa %xmm2, %xmm5
+; SSE2-NEXT:    pxor %xmm9, %xmm5
+; SSE2-NEXT:    movdqa %xmm9, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE2-NEXT:    pand %xmm13, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm5, %xmm4
+; SSE2-NEXT:    movdqa %xmm4, %xmm5
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,0,3,2]
+; SSE2-NEXT:    pand %xmm5, %xmm6
+; SSE2-NEXT:    pandn %xmm8, %xmm6
+; SSE2-NEXT:    movdqa %xmm4, %xmm5
+; SSE2-NEXT:    pandn %xmm11, %xmm5
+; SSE2-NEXT:    pand %xmm12, %xmm4
+; SSE2-NEXT:    por %xmm5, %xmm4
+; SSE2-NEXT:    pand %xmm6, %xmm4
+; SSE2-NEXT:    pandn %xmm2, %xmm6
+; SSE2-NEXT:    por %xmm4, %xmm6
+; SSE2-NEXT:    movdqa %xmm3, %xmm2
+; SSE2-NEXT:    paddq %xmm7, %xmm3
+; SSE2-NEXT:    pxor %xmm9, %xmm7
+; SSE2-NEXT:    movdqa %xmm9, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm7, %xmm4
+; SSE2-NEXT:    pxor %xmm10, %xmm4
+; SSE2-NEXT:    pxor %xmm9, %xmm2
+; SSE2-NEXT:    movdqa %xmm9, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2-NEXT:    pand %xmm7, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE2-NEXT:    por %xmm2, %xmm5
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm2
+; SSE2-NEXT:    movdqa %xmm3, %xmm4
+; SSE2-NEXT:    pxor %xmm9, %xmm4
+; SSE2-NEXT:    movdqa %xmm9, %xmm7
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pand %xmm8, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSE2-NEXT:    por %xmm4, %xmm7
+; SSE2-NEXT:    pxor %xmm7, %xmm10
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm10
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm10[1,0,3,2]
+; SSE2-NEXT:    pand %xmm10, %xmm5
+; SSE2-NEXT:    pandn %xmm2, %xmm5
+; SSE2-NEXT:    movdqa %xmm7, %xmm2
+; SSE2-NEXT:    pandn %xmm11, %xmm2
+; SSE2-NEXT:    pand %xmm12, %xmm7
+; SSE2-NEXT:    por %xmm2, %xmm7
+; SSE2-NEXT:    pand %xmm5, %xmm7
+; SSE2-NEXT:    pandn %xmm3, %xmm5
+; SSE2-NEXT:    por %xmm7, %xmm5
+; SSE2-NEXT:    movdqa %xmm6, %xmm2
+; SSE2-NEXT:    movdqa %xmm5, %xmm3
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v8i64:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pushq %r15
-; SSSE3-NEXT:    pushq %r14
-; SSSE3-NEXT:    pushq %rbx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm8, %rax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm8, %r9
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movq %r9, %rdx
-; SSSE3-NEXT:    addq %rax, %rdx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSSE3-NEXT:    addq %r8, %rcx
-; SSSE3-NEXT:    addq %rax, %r9
-; SSSE3-NEXT:    cmovoq %rcx, %r9
-; SSSE3-NEXT:    movq %xmm5, %rax
-; SSSE3-NEXT:    movq %xmm1, %r10
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movq %r10, %rsi
-; SSSE3-NEXT:    addq %rax, %rsi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    addq %r8, %rdx
-; SSSE3-NEXT:    addq %rax, %r10
-; SSSE3-NEXT:    cmovoq %rdx, %r10
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %rax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %r11
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movq %r11, %rdi
-; SSSE3-NEXT:    addq %rax, %rdi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    addq %r8, %rdx
-; SSSE3-NEXT:    addq %rax, %r11
-; SSSE3-NEXT:    cmovoq %rdx, %r11
-; SSSE3-NEXT:    movq %xmm6, %rax
-; SSSE3-NEXT:    movq %xmm2, %r14
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movq %r14, %rdi
-; SSSE3-NEXT:    addq %rax, %rdi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    addq %r8, %rdx
-; SSSE3-NEXT:    addq %rax, %r14
-; SSSE3-NEXT:    cmovoq %rdx, %r14
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %rax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %r15
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %r15, %rdx
-; SSSE3-NEXT:    addq %rax, %rdx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    addq %rax, %r15
-; SSSE3-NEXT:    cmovoq %rdi, %r15
-; SSSE3-NEXT:    movq %xmm7, %rdx
-; SSSE3-NEXT:    movq %xmm3, %rax
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %rax, %rbx
-; SSSE3-NEXT:    addq %rdx, %rbx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    addq %rdx, %rax
-; SSSE3-NEXT:    cmovoq %rdi, %rax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %rdx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %rcx
-; SSSE3-NEXT:    xorl %ebx, %ebx
-; SSSE3-NEXT:    movq %rcx, %rdi
-; SSSE3-NEXT:    addq %rdx, %rdi
-; SSSE3-NEXT:    setns %bl
-; SSSE3-NEXT:    addq %r8, %rbx
-; SSSE3-NEXT:    addq %rdx, %rcx
-; SSSE3-NEXT:    cmovoq %rbx, %rcx
-; SSSE3-NEXT:    movq %xmm4, %rdx
-; SSSE3-NEXT:    movq %xmm0, %rsi
-; SSSE3-NEXT:    xorl %ebx, %ebx
-; SSSE3-NEXT:    movq %rsi, %rdi
-; SSSE3-NEXT:    addq %rdx, %rdi
-; SSSE3-NEXT:    setns %bl
-; SSSE3-NEXT:    addq %r8, %rbx
-; SSSE3-NEXT:    addq %rdx, %rsi
-; SSSE3-NEXT:    cmovoq %rbx, %rsi
-; SSSE3-NEXT:    movq %rsi, %xmm0
-; SSSE3-NEXT:    movq %rcx, %xmm1
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSSE3-NEXT:    movq %r10, %xmm1
-; SSSE3-NEXT:    movq %r9, %xmm2
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; SSSE3-NEXT:    movq %r14, %xmm2
-; SSSE3-NEXT:    movq %r11, %xmm3
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
-; SSSE3-NEXT:    movq %rax, %xmm3
-; SSSE3-NEXT:    movq %r15, %xmm4
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm4[0]
-; SSSE3-NEXT:    popq %rbx
-; SSSE3-NEXT:    popq %r14
-; SSSE3-NEXT:    popq %r15
+; SSSE3-NEXT:    movdqa %xmm1, %xmm8
+; SSSE3-NEXT:    movdqa %xmm0, %xmm13
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm9 = [2147483648,2147483648]
+; SSSE3-NEXT:    paddq %xmm4, %xmm13
+; SSSE3-NEXT:    pxor %xmm9, %xmm4
+; SSSE3-NEXT:    movdqa %xmm9, %xmm1
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm10, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    por %xmm4, %xmm1
+; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm10
+; SSSE3-NEXT:    pxor %xmm10, %xmm1
+; SSSE3-NEXT:    pxor %xmm9, %xmm0
+; SSSE3-NEXT:    movdqa %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm11, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm14 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm0, %xmm14
+; SSSE3-NEXT:    pxor %xmm10, %xmm14
+; SSSE3-NEXT:    pcmpeqd %xmm14, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm1[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm1, %xmm11
+; SSSE3-NEXT:    movdqa %xmm13, %xmm0
+; SSSE3-NEXT:    pxor %xmm9, %xmm0
+; SSSE3-NEXT:    movdqa %xmm9, %xmm1
+; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm12 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm12, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    por %xmm0, %xmm1
+; SSSE3-NEXT:    movdqa %xmm1, %xmm4
+; SSSE3-NEXT:    pxor %xmm10, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm14, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm0
+; SSSE3-NEXT:    pandn %xmm11, %xmm0
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm11 = [9223372036854775808,9223372036854775808]
+; SSSE3-NEXT:    movdqa %xmm1, %xmm4
+; SSSE3-NEXT:    pandn %xmm11, %xmm4
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm12 = [9223372036854775807,9223372036854775807]
+; SSSE3-NEXT:    pand %xmm12, %xmm1
+; SSSE3-NEXT:    por %xmm4, %xmm1
+; SSSE3-NEXT:    pand %xmm0, %xmm1
+; SSSE3-NEXT:    pandn %xmm13, %xmm0
+; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    movdqa %xmm8, %xmm1
+; SSSE3-NEXT:    paddq %xmm5, %xmm8
+; SSSE3-NEXT:    pxor %xmm9, %xmm5
+; SSSE3-NEXT:    movdqa %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm13, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm5, %xmm4
+; SSSE3-NEXT:    pxor %xmm10, %xmm4
+; SSSE3-NEXT:    pxor %xmm9, %xmm1
+; SSSE3-NEXT:    movdqa %xmm9, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm13 = xmm5[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm13, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm15 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    por %xmm1, %xmm15
+; SSSE3-NEXT:    pxor %xmm10, %xmm15
+; SSSE3-NEXT:    pcmpeqd %xmm15, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm13
+; SSSE3-NEXT:    movdqa %xmm8, %xmm1
+; SSSE3-NEXT:    pxor %xmm9, %xmm1
+; SSSE3-NEXT:    movdqa %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm14 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm14, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm1, %xmm4
+; SSSE3-NEXT:    movdqa %xmm4, %xmm5
+; SSSE3-NEXT:    pxor %xmm10, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm15, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm5, %xmm1
+; SSSE3-NEXT:    pandn %xmm13, %xmm1
+; SSSE3-NEXT:    movdqa %xmm4, %xmm5
+; SSSE3-NEXT:    pandn %xmm11, %xmm5
+; SSSE3-NEXT:    pand %xmm12, %xmm4
+; SSSE3-NEXT:    por %xmm5, %xmm4
+; SSSE3-NEXT:    pand %xmm1, %xmm4
+; SSSE3-NEXT:    pandn %xmm8, %xmm1
+; SSSE3-NEXT:    por %xmm4, %xmm1
+; SSSE3-NEXT:    movdqa %xmm2, %xmm4
+; SSSE3-NEXT:    paddq %xmm6, %xmm2
+; SSSE3-NEXT:    pxor %xmm9, %xmm6
+; SSSE3-NEXT:    movdqa %xmm9, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm6, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm8, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    por %xmm6, %xmm5
+; SSSE3-NEXT:    pxor %xmm10, %xmm5
+; SSSE3-NEXT:    pxor %xmm9, %xmm4
+; SSSE3-NEXT:    movdqa %xmm9, %xmm6
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm8, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    por %xmm4, %xmm6
+; SSSE3-NEXT:    pxor %xmm10, %xmm6
+; SSSE3-NEXT:    pcmpeqd %xmm6, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm5, %xmm8
+; SSSE3-NEXT:    movdqa %xmm2, %xmm5
+; SSSE3-NEXT:    pxor %xmm9, %xmm5
+; SSSE3-NEXT:    movdqa %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm13, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm5, %xmm4
+; SSSE3-NEXT:    movdqa %xmm4, %xmm5
+; SSSE3-NEXT:    pxor %xmm10, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm6, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm5, %xmm6
+; SSSE3-NEXT:    pandn %xmm8, %xmm6
+; SSSE3-NEXT:    movdqa %xmm4, %xmm5
+; SSSE3-NEXT:    pandn %xmm11, %xmm5
+; SSSE3-NEXT:    pand %xmm12, %xmm4
+; SSSE3-NEXT:    por %xmm5, %xmm4
+; SSSE3-NEXT:    pand %xmm6, %xmm4
+; SSSE3-NEXT:    pandn %xmm2, %xmm6
+; SSSE3-NEXT:    por %xmm4, %xmm6
+; SSSE3-NEXT:    movdqa %xmm3, %xmm2
+; SSSE3-NEXT:    paddq %xmm7, %xmm3
+; SSSE3-NEXT:    pxor %xmm9, %xmm7
+; SSSE3-NEXT:    movdqa %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm5, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm7, %xmm4
+; SSSE3-NEXT:    pxor %xmm10, %xmm4
+; SSSE3-NEXT:    pxor %xmm9, %xmm2
+; SSSE3-NEXT:    movdqa %xmm9, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm7, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    por %xmm2, %xmm5
+; SSSE3-NEXT:    pxor %xmm10, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm2
+; SSSE3-NEXT:    movdqa %xmm3, %xmm4
+; SSSE3-NEXT:    pxor %xmm9, %xmm4
+; SSSE3-NEXT:    movdqa %xmm9, %xmm7
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm8, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; SSSE3-NEXT:    por %xmm4, %xmm7
+; SSSE3-NEXT:    pxor %xmm7, %xmm10
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm10
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm10[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm10, %xmm5
+; SSSE3-NEXT:    pandn %xmm2, %xmm5
+; SSSE3-NEXT:    movdqa %xmm7, %xmm2
+; SSSE3-NEXT:    pandn %xmm11, %xmm2
+; SSSE3-NEXT:    pand %xmm12, %xmm7
+; SSSE3-NEXT:    por %xmm2, %xmm7
+; SSSE3-NEXT:    pand %xmm5, %xmm7
+; SSSE3-NEXT:    pandn %xmm3, %xmm5
+; SSSE3-NEXT:    por %xmm7, %xmm5
+; SSSE3-NEXT:    movdqa %xmm6, %xmm2
+; SSSE3-NEXT:    movdqa %xmm5, %xmm3
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v8i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    pushq %r15
-; SSE41-NEXT:    pushq %r14
-; SSE41-NEXT:    pushq %rbx
-; SSE41-NEXT:    movq %xmm5, %rax
-; SSE41-NEXT:    movq %xmm1, %r9
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movq %r9, %rdx
-; SSE41-NEXT:    addq %rax, %rdx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE41-NEXT:    addq %r8, %rcx
-; SSE41-NEXT:    addq %rax, %r9
-; SSE41-NEXT:    cmovoq %rcx, %r9
-; SSE41-NEXT:    pextrq $1, %xmm5, %rax
-; SSE41-NEXT:    pextrq $1, %xmm1, %r10
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movq %r10, %rsi
-; SSE41-NEXT:    addq %rax, %rsi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    addq %r8, %rdx
-; SSE41-NEXT:    addq %rax, %r10
-; SSE41-NEXT:    cmovoq %rdx, %r10
-; SSE41-NEXT:    movq %xmm6, %rax
-; SSE41-NEXT:    movq %xmm2, %r11
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movq %r11, %rdi
-; SSE41-NEXT:    addq %rax, %rdi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    addq %r8, %rdx
-; SSE41-NEXT:    addq %rax, %r11
-; SSE41-NEXT:    cmovoq %rdx, %r11
-; SSE41-NEXT:    pextrq $1, %xmm6, %rax
-; SSE41-NEXT:    pextrq $1, %xmm2, %r14
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movq %r14, %rdi
-; SSE41-NEXT:    addq %rax, %rdi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    addq %r8, %rdx
-; SSE41-NEXT:    addq %rax, %r14
-; SSE41-NEXT:    cmovoq %rdx, %r14
-; SSE41-NEXT:    movq %xmm7, %rax
-; SSE41-NEXT:    movq %xmm3, %r15
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %r15, %rdx
-; SSE41-NEXT:    addq %rax, %rdx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    addq %rax, %r15
-; SSE41-NEXT:    cmovoq %rdi, %r15
-; SSE41-NEXT:    pextrq $1, %xmm7, %rdx
-; SSE41-NEXT:    pextrq $1, %xmm3, %rax
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %rax, %rbx
-; SSE41-NEXT:    addq %rdx, %rbx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    addq %rdx, %rax
-; SSE41-NEXT:    cmovoq %rdi, %rax
-; SSE41-NEXT:    movq %xmm4, %rdx
-; SSE41-NEXT:    movq %xmm0, %rcx
-; SSE41-NEXT:    xorl %ebx, %ebx
-; SSE41-NEXT:    movq %rcx, %rdi
-; SSE41-NEXT:    addq %rdx, %rdi
-; SSE41-NEXT:    setns %bl
-; SSE41-NEXT:    addq %r8, %rbx
-; SSE41-NEXT:    addq %rdx, %rcx
-; SSE41-NEXT:    cmovoq %rbx, %rcx
-; SSE41-NEXT:    pextrq $1, %xmm4, %rdx
-; SSE41-NEXT:    pextrq $1, %xmm0, %rsi
-; SSE41-NEXT:    xorl %ebx, %ebx
-; SSE41-NEXT:    movq %rsi, %rdi
-; SSE41-NEXT:    addq %rdx, %rdi
-; SSE41-NEXT:    setns %bl
-; SSE41-NEXT:    addq %r8, %rbx
-; SSE41-NEXT:    addq %rdx, %rsi
-; SSE41-NEXT:    cmovoq %rbx, %rsi
-; SSE41-NEXT:    movq %rsi, %xmm1
-; SSE41-NEXT:    movq %rcx, %xmm0
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE41-NEXT:    movq %r10, %xmm2
-; SSE41-NEXT:    movq %r9, %xmm1
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; SSE41-NEXT:    movq %r14, %xmm3
-; SSE41-NEXT:    movq %r11, %xmm2
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
-; SSE41-NEXT:    movq %rax, %xmm4
-; SSE41-NEXT:    movq %r15, %xmm3
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm4[0]
-; SSE41-NEXT:    popq %rbx
-; SSE41-NEXT:    popq %r14
-; SSE41-NEXT:    popq %r15
+; SSE41-NEXT:    movdqa %xmm0, %xmm8
+; SSE41-NEXT:    movdqa {{.*#+}} xmm10 = [2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm0, %xmm11
+; SSE41-NEXT:    paddq %xmm4, %xmm8
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    movdqa %xmm10, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm4, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm9 = xmm0[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE41-NEXT:    pand %xmm9, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm12 = xmm0[1,1,3,3]
+; SSE41-NEXT:    por %xmm4, %xmm12
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm9
+; SSE41-NEXT:    pxor %xmm9, %xmm12
+; SSE41-NEXT:    pxor %xmm10, %xmm11
+; SSE41-NEXT:    movdqa %xmm10, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm11, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm0[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm11
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm11[1,1,3,3]
+; SSE41-NEXT:    pand %xmm13, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[1,1,3,3]
+; SSE41-NEXT:    por %xmm4, %xmm11
+; SSE41-NEXT:    pxor %xmm9, %xmm11
+; SSE41-NEXT:    pcmpeqq %xmm11, %xmm12
+; SSE41-NEXT:    movdqa %xmm8, %xmm0
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm10, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm14 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm13, %xmm14
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm14, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm4
+; SSE41-NEXT:    pxor %xmm9, %xmm4
+; SSE41-NEXT:    pcmpeqq %xmm11, %xmm4
+; SSE41-NEXT:    pandn %xmm12, %xmm4
+; SSE41-NEXT:    movapd {{.*#+}} xmm12 = [9223372036854775807,9223372036854775807]
+; SSE41-NEXT:    movapd {{.*#+}} xmm11 = [9223372036854775808,9223372036854775808]
+; SSE41-NEXT:    movapd %xmm11, %xmm13
+; SSE41-NEXT:    blendvpd %xmm0, %xmm12, %xmm13
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm13, %xmm8
+; SSE41-NEXT:    movdqa %xmm1, %xmm14
+; SSE41-NEXT:    paddq %xmm5, %xmm1
+; SSE41-NEXT:    pxor %xmm10, %xmm5
+; SSE41-NEXT:    movdqa %xmm10, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm5, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3]
+; SSE41-NEXT:    pand %xmm13, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm15 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm0, %xmm15
+; SSE41-NEXT:    pxor %xmm9, %xmm15
+; SSE41-NEXT:    pxor %xmm10, %xmm14
+; SSE41-NEXT:    movdqa %xmm10, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm14, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm0[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm14
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm14[1,1,3,3]
+; SSE41-NEXT:    pand %xmm13, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3]
+; SSE41-NEXT:    por %xmm4, %xmm5
+; SSE41-NEXT:    pxor %xmm9, %xmm5
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm15
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm10, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm14 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm13, %xmm14
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm14, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm4
+; SSE41-NEXT:    pxor %xmm9, %xmm4
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm4
+; SSE41-NEXT:    pandn %xmm15, %xmm4
+; SSE41-NEXT:    movapd %xmm11, %xmm5
+; SSE41-NEXT:    blendvpd %xmm0, %xmm12, %xmm5
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm5, %xmm1
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    paddq %xmm6, %xmm2
+; SSE41-NEXT:    pxor %xmm10, %xmm6
+; SSE41-NEXT:    movdqa %xmm10, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm6, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm6, %xmm5
+; SSE41-NEXT:    pxor %xmm9, %xmm5
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm10, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm6, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm0, %xmm6
+; SSE41-NEXT:    pxor %xmm9, %xmm6
+; SSE41-NEXT:    pcmpeqq %xmm6, %xmm5
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm10, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm14 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm13, %xmm14
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm14, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm4
+; SSE41-NEXT:    pxor %xmm9, %xmm4
+; SSE41-NEXT:    pcmpeqq %xmm6, %xmm4
+; SSE41-NEXT:    pandn %xmm5, %xmm4
+; SSE41-NEXT:    movapd %xmm11, %xmm5
+; SSE41-NEXT:    blendvpd %xmm0, %xmm12, %xmm5
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm5, %xmm2
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    paddq %xmm7, %xmm3
+; SSE41-NEXT:    pxor %xmm10, %xmm7
+; SSE41-NEXT:    movdqa %xmm10, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm7, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm7
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm7[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm6, %xmm4
+; SSE41-NEXT:    pxor %xmm9, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm10, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm6, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE41-NEXT:    por %xmm0, %xmm5
+; SSE41-NEXT:    pxor %xmm9, %xmm5
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm4
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    pxor %xmm10, %xmm0
+; SSE41-NEXT:    movdqa %xmm10, %xmm6
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm6[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm13, %xmm7
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
+; SSE41-NEXT:    por %xmm7, %xmm0
+; SSE41-NEXT:    pxor %xmm0, %xmm9
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm9
+; SSE41-NEXT:    pandn %xmm4, %xmm9
+; SSE41-NEXT:    blendvpd %xmm0, %xmm12, %xmm11
+; SSE41-NEXT:    movdqa %xmm9, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm11, %xmm3
+; SSE41-NEXT:    movapd %xmm8, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: v8i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    pushq %r15
-; AVX1-NEXT:    pushq %r14
-; AVX1-NEXT:    pushq %rbx
-; AVX1-NEXT:    vmovq %xmm3, %rax
-; AVX1-NEXT:    vmovq %xmm1, %r9
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movq %r9, %rdx
-; AVX1-NEXT:    addq %rax, %rdx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX1-NEXT:    addq %r8, %rcx
-; AVX1-NEXT:    addq %rax, %r9
-; AVX1-NEXT:    cmovoq %rcx, %r9
-; AVX1-NEXT:    vpextrq $1, %xmm3, %rax
-; AVX1-NEXT:    vpextrq $1, %xmm1, %r10
-; AVX1-NEXT:    xorl %edx, %edx
-; AVX1-NEXT:    movq %r10, %rsi
-; AVX1-NEXT:    addq %rax, %rsi
-; AVX1-NEXT:    setns %dl
-; AVX1-NEXT:    addq %r8, %rdx
-; AVX1-NEXT:    addq %rax, %r10
-; AVX1-NEXT:    cmovoq %rdx, %r10
-; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm3
-; AVX1-NEXT:    vmovq %xmm3, %rax
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vmovq %xmm1, %r11
-; AVX1-NEXT:    xorl %edx, %edx
-; AVX1-NEXT:    movq %r11, %rdi
-; AVX1-NEXT:    addq %rax, %rdi
-; AVX1-NEXT:    setns %dl
-; AVX1-NEXT:    addq %r8, %rdx
-; AVX1-NEXT:    addq %rax, %r11
-; AVX1-NEXT:    cmovoq %rdx, %r11
-; AVX1-NEXT:    vpextrq $1, %xmm3, %rax
-; AVX1-NEXT:    vpextrq $1, %xmm1, %r14
-; AVX1-NEXT:    xorl %edx, %edx
-; AVX1-NEXT:    movq %r14, %rdi
-; AVX1-NEXT:    addq %rax, %rdi
-; AVX1-NEXT:    setns %dl
-; AVX1-NEXT:    addq %r8, %rdx
-; AVX1-NEXT:    addq %rax, %r14
-; AVX1-NEXT:    cmovoq %rdx, %r14
-; AVX1-NEXT:    vmovq %xmm2, %rax
-; AVX1-NEXT:    vmovq %xmm0, %r15
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movq %r15, %rdx
-; AVX1-NEXT:    addq %rax, %rdx
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addq %r8, %rdi
-; AVX1-NEXT:    addq %rax, %r15
-; AVX1-NEXT:    cmovoq %rdi, %r15
-; AVX1-NEXT:    vpextrq $1, %xmm2, %rdx
-; AVX1-NEXT:    vpextrq $1, %xmm0, %rax
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movq %rax, %rbx
-; AVX1-NEXT:    addq %rdx, %rbx
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addq %r8, %rdi
-; AVX1-NEXT:    addq %rdx, %rax
-; AVX1-NEXT:    cmovoq %rdi, %rax
-; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm1
-; AVX1-NEXT:    vmovq %xmm1, %rdx
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vmovq %xmm0, %rcx
-; AVX1-NEXT:    xorl %ebx, %ebx
-; AVX1-NEXT:    movq %rcx, %rdi
-; AVX1-NEXT:    addq %rdx, %rdi
-; AVX1-NEXT:    setns %bl
-; AVX1-NEXT:    addq %r8, %rbx
-; AVX1-NEXT:    addq %rdx, %rcx
-; AVX1-NEXT:    cmovoq %rbx, %rcx
-; AVX1-NEXT:    vpextrq $1, %xmm1, %rdx
-; AVX1-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX1-NEXT:    xorl %ebx, %ebx
-; AVX1-NEXT:    movq %rsi, %rdi
-; AVX1-NEXT:    addq %rdx, %rdi
-; AVX1-NEXT:    setns %bl
-; AVX1-NEXT:    addq %r8, %rbx
-; AVX1-NEXT:    addq %rdx, %rsi
-; AVX1-NEXT:    cmovoq %rbx, %rsi
-; AVX1-NEXT:    vmovq %rsi, %xmm0
-; AVX1-NEXT:    vmovq %rcx, %xmm1
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX1-NEXT:    vmovq %rax, %xmm1
-; AVX1-NEXT:    vmovq %r15, %xmm2
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT:    vmovq %r14, %xmm1
-; AVX1-NEXT:    vmovq %r11, %xmm2
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX1-NEXT:    vmovq %r10, %xmm2
-; AVX1-NEXT:    vmovq %r9, %xmm3
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
-; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm2, %ymm1
-; AVX1-NEXT:    popq %rbx
-; AVX1-NEXT:    popq %r14
-; AVX1-NEXT:    popq %r15
+; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm9
+; AVX1-NEXT:    vpxor %xmm12, %xmm12, %xmm12
+; AVX1-NEXT:    vpcmpgtq %xmm9, %xmm12, %xmm7
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm8
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm7
+; AVX1-NEXT:    vpcmpgtq %xmm7, %xmm12, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm10
+; AVX1-NEXT:    vpcmpeqq %xmm8, %xmm10, %xmm8
+; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm12, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm11
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm12, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm6
+; AVX1-NEXT:    vpcmpeqq %xmm11, %xmm6, %xmm11
+; AVX1-NEXT:    vinsertf128 $1, %xmm8, %ymm11, %ymm8
+; AVX1-NEXT:    vpaddq %xmm9, %xmm7, %xmm9
+; AVX1-NEXT:    vpcmpgtq %xmm9, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm5
+; AVX1-NEXT:    vpcmpeqq %xmm5, %xmm10, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm12, %xmm2
+; AVX1-NEXT:    vpxor %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqq %xmm2, %xmm6, %xmm2
+; AVX1-NEXT:    vpxor %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm5, %ymm2, %ymm2
+; AVX1-NEXT:    vandpd %ymm2, %ymm8, %ymm5
+; AVX1-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm7
+; AVX1-NEXT:    vmovapd {{.*#+}} ymm8 = [9223372036854775807,9223372036854775807,9223372036854775807,9223372036854775807]
+; AVX1-NEXT:    vmovapd {{.*#+}} ymm10 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
+; AVX1-NEXT:    vblendvpd %ymm7, %ymm8, %ymm10, %ymm7
+; AVX1-NEXT:    vinsertf128 $1, %xmm9, %ymm0, %ymm0
+; AVX1-NEXT:    vblendvpd %ymm5, %ymm7, %ymm0, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm5
+; AVX1-NEXT:    vpcmpgtq %xmm5, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm7
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm12, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm6
+; AVX1-NEXT:    vpcmpeqq %xmm7, %xmm6, %xmm9
+; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm11
+; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm7
+; AVX1-NEXT:    vpcmpeqq %xmm11, %xmm7, %xmm11
+; AVX1-NEXT:    vinsertf128 $1, %xmm9, %ymm11, %ymm9
+; AVX1-NEXT:    vpaddq %xmm5, %xmm2, %xmm11
+; AVX1-NEXT:    vpcmpgtq %xmm11, %xmm12, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm2
+; AVX1-NEXT:    vpcmpeqq %xmm2, %xmm6, %xmm2
+; AVX1-NEXT:    vpxor %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddq %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm12, %xmm3
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm7, %xmm3
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
+; AVX1-NEXT:    vandpd %ymm2, %ymm9, %ymm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm5, %ymm1, %ymm3
+; AVX1-NEXT:    vblendvpd %ymm3, %ymm8, %ymm10, %ymm3
+; AVX1-NEXT:    vinsertf128 $1, %xmm11, %ymm1, %ymm1
+; AVX1-NEXT:    vblendvpd %ymm2, %ymm3, %ymm1, %ymm1
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: v8i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    pushq %r15
-; AVX2-NEXT:    pushq %r14
-; AVX2-NEXT:    pushq %rbx
-; AVX2-NEXT:    vmovq %xmm3, %rax
-; AVX2-NEXT:    vmovq %xmm1, %r9
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movq %r9, %rdx
-; AVX2-NEXT:    addq %rax, %rdx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX2-NEXT:    addq %r8, %rcx
-; AVX2-NEXT:    addq %rax, %r9
-; AVX2-NEXT:    cmovoq %rcx, %r9
-; AVX2-NEXT:    vpextrq $1, %xmm3, %rax
-; AVX2-NEXT:    vpextrq $1, %xmm1, %r10
-; AVX2-NEXT:    xorl %edx, %edx
-; AVX2-NEXT:    movq %r10, %rsi
-; AVX2-NEXT:    addq %rax, %rsi
-; AVX2-NEXT:    setns %dl
-; AVX2-NEXT:    addq %r8, %rdx
-; AVX2-NEXT:    addq %rax, %r10
-; AVX2-NEXT:    cmovoq %rdx, %r10
-; AVX2-NEXT:    vextracti128 $1, %ymm3, %xmm3
-; AVX2-NEXT:    vmovq %xmm3, %rax
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; AVX2-NEXT:    vmovq %xmm1, %r11
-; AVX2-NEXT:    xorl %edx, %edx
-; AVX2-NEXT:    movq %r11, %rdi
-; AVX2-NEXT:    addq %rax, %rdi
-; AVX2-NEXT:    setns %dl
-; AVX2-NEXT:    addq %r8, %rdx
-; AVX2-NEXT:    addq %rax, %r11
-; AVX2-NEXT:    cmovoq %rdx, %r11
-; AVX2-NEXT:    vpextrq $1, %xmm3, %rax
-; AVX2-NEXT:    vpextrq $1, %xmm1, %r14
-; AVX2-NEXT:    xorl %edx, %edx
-; AVX2-NEXT:    movq %r14, %rdi
-; AVX2-NEXT:    addq %rax, %rdi
-; AVX2-NEXT:    setns %dl
-; AVX2-NEXT:    addq %r8, %rdx
-; AVX2-NEXT:    addq %rax, %r14
-; AVX2-NEXT:    cmovoq %rdx, %r14
-; AVX2-NEXT:    vmovq %xmm2, %rax
-; AVX2-NEXT:    vmovq %xmm0, %r15
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movq %r15, %rdx
-; AVX2-NEXT:    addq %rax, %rdx
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addq %r8, %rdi
-; AVX2-NEXT:    addq %rax, %r15
-; AVX2-NEXT:    cmovoq %rdi, %r15
-; AVX2-NEXT:    vpextrq $1, %xmm2, %rdx
-; AVX2-NEXT:    vpextrq $1, %xmm0, %rax
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movq %rax, %rbx
-; AVX2-NEXT:    addq %rdx, %rbx
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addq %r8, %rdi
-; AVX2-NEXT:    addq %rdx, %rax
-; AVX2-NEXT:    cmovoq %rdi, %rax
-; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm1
-; AVX2-NEXT:    vmovq %xmm1, %rdx
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; AVX2-NEXT:    vmovq %xmm0, %rcx
-; AVX2-NEXT:    xorl %ebx, %ebx
-; AVX2-NEXT:    movq %rcx, %rdi
-; AVX2-NEXT:    addq %rdx, %rdi
-; AVX2-NEXT:    setns %bl
-; AVX2-NEXT:    addq %r8, %rbx
-; AVX2-NEXT:    addq %rdx, %rcx
-; AVX2-NEXT:    cmovoq %rbx, %rcx
-; AVX2-NEXT:    vpextrq $1, %xmm1, %rdx
-; AVX2-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX2-NEXT:    xorl %ebx, %ebx
-; AVX2-NEXT:    movq %rsi, %rdi
-; AVX2-NEXT:    addq %rdx, %rdi
-; AVX2-NEXT:    setns %bl
-; AVX2-NEXT:    addq %r8, %rbx
-; AVX2-NEXT:    addq %rdx, %rsi
-; AVX2-NEXT:    cmovoq %rbx, %rsi
-; AVX2-NEXT:    vmovq %rsi, %xmm0
-; AVX2-NEXT:    vmovq %rcx, %xmm1
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX2-NEXT:    vmovq %rax, %xmm1
-; AVX2-NEXT:    vmovq %r15, %xmm2
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX2-NEXT:    vmovq %r14, %xmm1
-; AVX2-NEXT:    vmovq %r11, %xmm2
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX2-NEXT:    vmovq %r10, %xmm2
-; AVX2-NEXT:    vmovq %r9, %xmm3
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
-; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
-; AVX2-NEXT:    popq %rbx
-; AVX2-NEXT:    popq %r14
-; AVX2-NEXT:    popq %r15
+; AVX2-NEXT:    vpxor %xmm4, %xmm4, %xmm4
+; AVX2-NEXT:    vpcmpgtq %ymm2, %ymm4, %ymm5
+; AVX2-NEXT:    vpcmpeqd %ymm6, %ymm6, %ymm6
+; AVX2-NEXT:    vpxor %ymm6, %ymm5, %ymm5
+; AVX2-NEXT:    vpcmpgtq %ymm0, %ymm4, %ymm7
+; AVX2-NEXT:    vpxor %ymm6, %ymm7, %ymm7
+; AVX2-NEXT:    vpcmpeqq %ymm5, %ymm7, %ymm5
+; AVX2-NEXT:    vpaddq %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpgtq %ymm0, %ymm4, %ymm2
+; AVX2-NEXT:    vpxor %ymm6, %ymm2, %ymm2
+; AVX2-NEXT:    vpcmpeqq %ymm2, %ymm7, %ymm2
+; AVX2-NEXT:    vpandn %ymm5, %ymm2, %ymm2
+; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm5 = [9223372036854775807,9223372036854775807,9223372036854775807,9223372036854775807]
+; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm7 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
+; AVX2-NEXT:    vblendvpd %ymm0, %ymm5, %ymm7, %ymm8
+; AVX2-NEXT:    vblendvpd %ymm2, %ymm8, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpgtq %ymm3, %ymm4, %ymm2
+; AVX2-NEXT:    vpxor %ymm6, %ymm2, %ymm2
+; AVX2-NEXT:    vpcmpgtq %ymm1, %ymm4, %ymm8
+; AVX2-NEXT:    vpxor %ymm6, %ymm8, %ymm8
+; AVX2-NEXT:    vpcmpeqq %ymm2, %ymm8, %ymm2
+; AVX2-NEXT:    vpaddq %ymm3, %ymm1, %ymm1
+; AVX2-NEXT:    vpcmpgtq %ymm1, %ymm4, %ymm3
+; AVX2-NEXT:    vpxor %ymm6, %ymm3, %ymm3
+; AVX2-NEXT:    vpcmpeqq %ymm3, %ymm8, %ymm3
+; AVX2-NEXT:    vpandn %ymm2, %ymm3, %ymm2
+; AVX2-NEXT:    vblendvpd %ymm1, %ymm5, %ymm7, %ymm3
+; AVX2-NEXT:    vblendvpd %ymm2, %ymm3, %ymm1, %ymm1
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: v8i64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    pushq %r15
-; AVX512-NEXT:    pushq %r14
-; AVX512-NEXT:    pushq %rbx
-; AVX512-NEXT:    vmovq %xmm1, %rax
-; AVX512-NEXT:    vmovq %xmm0, %r9
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movq %r9, %rdx
-; AVX512-NEXT:    addq %rax, %rdx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX512-NEXT:    addq %r8, %rcx
-; AVX512-NEXT:    addq %rax, %r9
-; AVX512-NEXT:    cmovoq %rcx, %r9
-; AVX512-NEXT:    vpextrq $1, %xmm1, %rax
-; AVX512-NEXT:    vpextrq $1, %xmm0, %r10
-; AVX512-NEXT:    xorl %edx, %edx
-; AVX512-NEXT:    movq %r10, %rsi
-; AVX512-NEXT:    addq %rax, %rsi
-; AVX512-NEXT:    setns %dl
-; AVX512-NEXT:    addq %r8, %rdx
-; AVX512-NEXT:    addq %rax, %r10
-; AVX512-NEXT:    cmovoq %rdx, %r10
-; AVX512-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX512-NEXT:    vmovq %xmm2, %rax
-; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm3
-; AVX512-NEXT:    vmovq %xmm3, %r11
-; AVX512-NEXT:    xorl %edx, %edx
-; AVX512-NEXT:    movq %r11, %rdi
-; AVX512-NEXT:    addq %rax, %rdi
-; AVX512-NEXT:    setns %dl
-; AVX512-NEXT:    addq %r8, %rdx
-; AVX512-NEXT:    addq %rax, %r11
-; AVX512-NEXT:    cmovoq %rdx, %r11
-; AVX512-NEXT:    vpextrq $1, %xmm2, %rax
-; AVX512-NEXT:    vpextrq $1, %xmm3, %r14
-; AVX512-NEXT:    xorl %edx, %edx
-; AVX512-NEXT:    movq %r14, %rdi
-; AVX512-NEXT:    addq %rax, %rdi
-; AVX512-NEXT:    setns %dl
-; AVX512-NEXT:    addq %r8, %rdx
-; AVX512-NEXT:    addq %rax, %r14
-; AVX512-NEXT:    cmovoq %rdx, %r14
-; AVX512-NEXT:    vextracti32x4 $2, %zmm1, %xmm2
-; AVX512-NEXT:    vmovq %xmm2, %rax
-; AVX512-NEXT:    vextracti32x4 $2, %zmm0, %xmm3
-; AVX512-NEXT:    vmovq %xmm3, %r15
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movq %r15, %rdx
-; AVX512-NEXT:    addq %rax, %rdx
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addq %r8, %rdi
-; AVX512-NEXT:    addq %rax, %r15
-; AVX512-NEXT:    cmovoq %rdi, %r15
-; AVX512-NEXT:    vpextrq $1, %xmm2, %rdx
-; AVX512-NEXT:    vpextrq $1, %xmm3, %rax
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movq %rax, %rbx
-; AVX512-NEXT:    addq %rdx, %rbx
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addq %r8, %rdi
-; AVX512-NEXT:    addq %rdx, %rax
-; AVX512-NEXT:    cmovoq %rdi, %rax
-; AVX512-NEXT:    vextracti32x4 $3, %zmm1, %xmm1
-; AVX512-NEXT:    vmovq %xmm1, %rdx
-; AVX512-NEXT:    vextracti32x4 $3, %zmm0, %xmm0
-; AVX512-NEXT:    vmovq %xmm0, %rcx
-; AVX512-NEXT:    xorl %ebx, %ebx
-; AVX512-NEXT:    movq %rcx, %rdi
-; AVX512-NEXT:    addq %rdx, %rdi
-; AVX512-NEXT:    setns %bl
-; AVX512-NEXT:    addq %r8, %rbx
-; AVX512-NEXT:    addq %rdx, %rcx
-; AVX512-NEXT:    cmovoq %rbx, %rcx
-; AVX512-NEXT:    vpextrq $1, %xmm1, %rdx
-; AVX512-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX512-NEXT:    xorl %ebx, %ebx
-; AVX512-NEXT:    movq %rsi, %rdi
-; AVX512-NEXT:    addq %rdx, %rdi
-; AVX512-NEXT:    setns %bl
-; AVX512-NEXT:    addq %r8, %rbx
-; AVX512-NEXT:    addq %rdx, %rsi
-; AVX512-NEXT:    cmovoq %rbx, %rsi
-; AVX512-NEXT:    vmovq %rsi, %xmm0
-; AVX512-NEXT:    vmovq %rcx, %xmm1
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX512-NEXT:    vmovq %rax, %xmm1
-; AVX512-NEXT:    vmovq %r15, %xmm2
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX512-NEXT:    vmovq %r14, %xmm1
-; AVX512-NEXT:    vmovq %r11, %xmm2
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX512-NEXT:    vmovq %r10, %xmm2
-; AVX512-NEXT:    vmovq %r9, %xmm3
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
-; AVX512-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
-; AVX512-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512-NEXT:    popq %rbx
-; AVX512-NEXT:    popq %r14
-; AVX512-NEXT:    popq %r15
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltq %zmm2, %zmm1, %k0
+; AVX512-NEXT:    vpcmpnltq %zmm2, %zmm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpaddq %zmm1, %zmm0, %zmm0
+; AVX512-NEXT:    vpcmpnltq %zmm2, %zmm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandnw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtq %zmm0, %zmm2, %k2
+; AVX512-NEXT:    vpbroadcastq {{.*#+}} zmm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
+; AVX512-NEXT:    vpbroadcastq {{.*}}(%rip), %zmm1 {%k2}
+; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
 ; AVX512-NEXT:    retq
   %z = call <8 x i64> @llvm.sadd.sat.v8i64(<8 x i64> %x, <8 x i64> %y)
   ret <8 x i64> %z
diff --git a/test/CodeGen/X86/ssub_sat.ll b/test/CodeGen/X86/ssub_sat.ll
index c78e44e..7fda35a 100644
--- a/test/CodeGen/X86/ssub_sat.ll
+++ b/test/CodeGen/X86/ssub_sat.ll
@@ -185,55 +185,30 @@
 ;
 ; X64-LABEL: vec:
 ; X64:       # %bb.0:
-; X64-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,2,3]
-; X64-NEXT:    movd %xmm2, %ecx
-; X64-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
-; X64-NEXT:    movd %xmm2, %r8d
-; X64-NEXT:    xorl %edx, %edx
-; X64-NEXT:    movl %r8d, %esi
-; X64-NEXT:    subl %ecx, %esi
-; X64-NEXT:    setns %dl
-; X64-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; X64-NEXT:    subl %ecx, %r8d
-; X64-NEXT:    cmovol %edx, %r8d
-; X64-NEXT:    movd %xmm1, %edx
-; X64-NEXT:    movd %xmm0, %ecx
-; X64-NEXT:    xorl %esi, %esi
-; X64-NEXT:    movl %ecx, %edi
-; X64-NEXT:    subl %edx, %edi
-; X64-NEXT:    setns %sil
-; X64-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; X64-NEXT:    subl %edx, %ecx
-; X64-NEXT:    cmovol %esi, %ecx
-; X64-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; X64-NEXT:    movd %xmm2, %edx
-; X64-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; X64-NEXT:    movd %xmm2, %eax
-; X64-NEXT:    xorl %edi, %edi
-; X64-NEXT:    movl %eax, %esi
-; X64-NEXT:    subl %edx, %esi
-; X64-NEXT:    setns %dil
-; X64-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; X64-NEXT:    subl %edx, %eax
-; X64-NEXT:    cmovol %edi, %eax
-; X64-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; X64-NEXT:    movd %xmm1, %r9d
-; X64-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; X64-NEXT:    movd %xmm0, %edx
-; X64-NEXT:    xorl %edi, %edi
-; X64-NEXT:    movl %edx, %esi
-; X64-NEXT:    subl %r9d, %esi
-; X64-NEXT:    setns %dil
-; X64-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; X64-NEXT:    subl %r9d, %edx
-; X64-NEXT:    cmovol %edi, %edx
-; X64-NEXT:    movd %edx, %xmm0
-; X64-NEXT:    movd %eax, %xmm1
-; X64-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; X64-NEXT:    movd %ecx, %xmm0
-; X64-NEXT:    movd %r8d, %xmm2
-; X64-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; X64-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; X64-NEXT:    movdqa %xmm0, %xmm2
+; X64-NEXT:    pxor %xmm3, %xmm3
+; X64-NEXT:    pxor %xmm0, %xmm0
+; X64-NEXT:    pcmpgtd %xmm1, %xmm0
+; X64-NEXT:    pcmpeqd %xmm4, %xmm4
+; X64-NEXT:    pxor %xmm4, %xmm0
+; X64-NEXT:    pxor %xmm5, %xmm5
+; X64-NEXT:    pcmpgtd %xmm2, %xmm5
+; X64-NEXT:    pxor %xmm4, %xmm5
+; X64-NEXT:    pcmpeqd %xmm5, %xmm0
+; X64-NEXT:    psubd %xmm1, %xmm2
+; X64-NEXT:    pcmpgtd %xmm2, %xmm3
+; X64-NEXT:    movdqa %xmm3, %xmm1
+; X64-NEXT:    pxor %xmm4, %xmm1
+; X64-NEXT:    pcmpeqd %xmm5, %xmm1
+; X64-NEXT:    pxor %xmm4, %xmm1
+; X64-NEXT:    pandn %xmm1, %xmm0
+; X64-NEXT:    movdqa %xmm3, %xmm1
+; X64-NEXT:    pandn {{.*}}(%rip), %xmm1
+; X64-NEXT:    psrld $1, %xmm3
+; X64-NEXT:    por %xmm1, %xmm3
+; X64-NEXT:    pand %xmm0, %xmm3
+; X64-NEXT:    pandn %xmm2, %xmm0
+; X64-NEXT:    por %xmm3, %xmm0
 ; X64-NEXT:    retq
   %tmp = call <4 x i32> @llvm.ssub.sat.v4i32(<4 x i32> %x, <4 x i32> %y);
   ret <4 x i32> %tmp;
diff --git a/test/CodeGen/X86/ssub_sat_vec.ll b/test/CodeGen/X86/ssub_sat_vec.ll
index 2e6ec60..5ba7b17 100644
--- a/test/CodeGen/X86/ssub_sat_vec.ll
+++ b/test/CodeGen/X86/ssub_sat_vec.ll
@@ -661,128 +661,226 @@
 ; SSE2-LABEL: v2i32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    psllq $32, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movq %xmm2, %rax
+; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
 ; SSE2-NEXT:    psllq $32, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movq %xmm2, %rcx
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movq %rcx, %rsi
-; SSE2-NEXT:    subq %rax, %rsi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE2-NEXT:    addq %r8, %rdx
-; SSE2-NEXT:    subq %rax, %rcx
-; SSE2-NEXT:    cmovoq %rdx, %rcx
-; SSE2-NEXT:    movq %xmm1, %rax
-; SSE2-NEXT:    movq %xmm0, %rsi
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %rsi, %rdx
-; SSE2-NEXT:    subq %rax, %rdx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    subq %rax, %rsi
-; SSE2-NEXT:    cmovoq %rdi, %rsi
-; SSE2-NEXT:    movq %rsi, %xmm1
-; SSE2-NEXT:    movq %rcx, %xmm0
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,3,2,3]
-; SSE2-NEXT:    psrad $31, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,3,2,3]
+; SSE2-NEXT:    movdqa %xmm0, %xmm3
+; SSE2-NEXT:    psubq %xmm1, %xmm0
+; SSE2-NEXT:    pxor %xmm2, %xmm1
+; SSE2-NEXT:    movdqa %xmm2, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm1, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm5
+; SSE2-NEXT:    pxor %xmm5, %xmm4
+; SSE2-NEXT:    pxor %xmm2, %xmm3
+; SSE2-NEXT:    movdqa %xmm2, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2-NEXT:    pand %xmm6, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[1,1,3,3]
+; SSE2-NEXT:    por %xmm3, %xmm6
+; SSE2-NEXT:    pxor %xmm5, %xmm6
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm1
+; SSE2-NEXT:    movdqa %xmm0, %xmm3
+; SSE2-NEXT:    pxor %xmm2, %xmm3
+; SSE2-NEXT:    movdqa %xmm2, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2-NEXT:    pand %xmm7, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm3, %xmm4
+; SSE2-NEXT:    pxor %xmm5, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm3
+; SSE2-NEXT:    pxor %xmm5, %xmm3
+; SSE2-NEXT:    pandn %xmm3, %xmm1
+; SSE2-NEXT:    movdqa %xmm1, %xmm3
+; SSE2-NEXT:    pandn %xmm0, %xmm3
+; SSE2-NEXT:    por %xmm2, %xmm0
+; SSE2-NEXT:    movdqa %xmm2, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm0, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm0
+; SSE2-NEXT:    pandn {{.*}}(%rip), %xmm0
+; SSE2-NEXT:    pand {{.*}}(%rip), %xmm2
+; SSE2-NEXT:    por %xmm0, %xmm2
+; SSE2-NEXT:    pand %xmm1, %xmm2
+; SSE2-NEXT:    por %xmm3, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,3,2,3]
+; SSE2-NEXT:    psrad $31, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,3,2,3]
 ; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v2i32:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    psllq $32, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm2, %rax
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
 ; SSSE3-NEXT:    psllq $32, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm2, %rcx
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movq %rcx, %rsi
-; SSSE3-NEXT:    subq %rax, %rsi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSSE3-NEXT:    addq %r8, %rdx
-; SSSE3-NEXT:    subq %rax, %rcx
-; SSSE3-NEXT:    cmovoq %rdx, %rcx
-; SSSE3-NEXT:    movq %xmm1, %rax
-; SSSE3-NEXT:    movq %xmm0, %rsi
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %rsi, %rdx
-; SSSE3-NEXT:    subq %rax, %rdx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    subq %rax, %rsi
-; SSSE3-NEXT:    cmovoq %rdi, %rsi
-; SSSE3-NEXT:    movq %rsi, %xmm1
-; SSSE3-NEXT:    movq %rcx, %xmm0
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,3,2,3]
-; SSSE3-NEXT:    psrad $31, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,3,2,3]
+; SSSE3-NEXT:    movdqa %xmm0, %xmm3
+; SSSE3-NEXT:    psubq %xmm1, %xmm0
+; SSSE3-NEXT:    pxor %xmm2, %xmm1
+; SSSE3-NEXT:    movdqa %xmm2, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm5, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm1, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm5
+; SSSE3-NEXT:    pxor %xmm5, %xmm4
+; SSSE3-NEXT:    pxor %xmm2, %xmm3
+; SSSE3-NEXT:    movdqa %xmm2, %xmm1
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm6, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    por %xmm3, %xmm6
+; SSSE3-NEXT:    pxor %xmm5, %xmm6
+; SSSE3-NEXT:    pcmpeqd %xmm6, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm1
+; SSSE3-NEXT:    movdqa %xmm0, %xmm3
+; SSSE3-NEXT:    pxor %xmm2, %xmm3
+; SSSE3-NEXT:    movdqa %xmm2, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm7, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm3, %xmm4
+; SSSE3-NEXT:    pxor %xmm5, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm6, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm3
+; SSSE3-NEXT:    pxor %xmm5, %xmm3
+; SSSE3-NEXT:    pandn %xmm3, %xmm1
+; SSSE3-NEXT:    movdqa %xmm1, %xmm3
+; SSSE3-NEXT:    pandn %xmm0, %xmm3
+; SSSE3-NEXT:    por %xmm2, %xmm0
+; SSSE3-NEXT:    movdqa %xmm2, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm5, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm0, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSSE3-NEXT:    pandn {{.*}}(%rip), %xmm0
+; SSSE3-NEXT:    pand {{.*}}(%rip), %xmm2
+; SSSE3-NEXT:    por %xmm0, %xmm2
+; SSSE3-NEXT:    pand %xmm1, %xmm2
+; SSSE3-NEXT:    por %xmm3, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,3,2,3]
+; SSSE3-NEXT:    psrad $31, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,3,2,3]
 ; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v2i32:
 ; SSE41:       # %bb.0:
+; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    psllq $32, %xmm1
-; SSE41-NEXT:    movq %xmm1, %rax
-; SSE41-NEXT:    psllq $32, %xmm0
-; SSE41-NEXT:    movq %xmm0, %rcx
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movq %rcx, %rsi
-; SSE41-NEXT:    subq %rax, %rsi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE41-NEXT:    addq %r8, %rdx
-; SSE41-NEXT:    subq %rax, %rcx
-; SSE41-NEXT:    cmovoq %rdx, %rcx
-; SSE41-NEXT:    pextrq $1, %xmm1, %rax
-; SSE41-NEXT:    pextrq $1, %xmm0, %rsi
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %rsi, %rdx
-; SSE41-NEXT:    subq %rax, %rdx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    subq %rax, %rsi
-; SSE41-NEXT:    cmovoq %rdi, %rsi
-; SSE41-NEXT:    movq %rsi, %xmm1
-; SSE41-NEXT:    movq %rcx, %xmm0
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE41-NEXT:    psrad $31, %xmm0
-; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5],xmm0[6,7]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
+; SSE41-NEXT:    psllq $32, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm3
+; SSE41-NEXT:    psubq %xmm1, %xmm2
+; SSE41-NEXT:    pxor %xmm0, %xmm1
+; SSE41-NEXT:    movdqa %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm6, %xmm1
+; SSE41-NEXT:    pcmpeqd %xmm4, %xmm4
+; SSE41-NEXT:    pxor %xmm4, %xmm1
+; SSE41-NEXT:    pxor %xmm0, %xmm3
+; SSE41-NEXT:    movdqa %xmm0, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE41-NEXT:    pand %xmm6, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE41-NEXT:    por %xmm3, %xmm5
+; SSE41-NEXT:    pxor %xmm4, %xmm5
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm1
+; SSE41-NEXT:    movdqa %xmm2, %xmm3
+; SSE41-NEXT:    pxor %xmm0, %xmm3
+; SSE41-NEXT:    movdqa %xmm0, %xmm6
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE41-NEXT:    pand %xmm7, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE41-NEXT:    por %xmm3, %xmm6
+; SSE41-NEXT:    pxor %xmm4, %xmm6
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm6
+; SSE41-NEXT:    pxor %xmm4, %xmm6
+; SSE41-NEXT:    pandn %xmm6, %xmm1
+; SSE41-NEXT:    movdqa %xmm2, %xmm3
+; SSE41-NEXT:    por %xmm0, %xmm3
+; SSE41-NEXT:    movdqa %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm0
+; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    movapd {{.*#+}} xmm3 = [9223372036854775808,9223372036854775808]
+; SSE41-NEXT:    blendvpd %xmm0, {{.*}}(%rip), %xmm3
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
+; SSE41-NEXT:    psrad $31, %xmm2
+; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: v2i32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpsllq $32, %xmm1, %xmm1
-; AVX1-NEXT:    vmovq %xmm1, %rax
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
 ; AVX1-NEXT:    vpsllq $32, %xmm0, %xmm0
-; AVX1-NEXT:    vmovq %xmm0, %rcx
-; AVX1-NEXT:    xorl %edx, %edx
-; AVX1-NEXT:    movq %rcx, %rsi
-; AVX1-NEXT:    subq %rax, %rsi
-; AVX1-NEXT:    setns %dl
-; AVX1-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX1-NEXT:    addq %r8, %rdx
-; AVX1-NEXT:    subq %rax, %rcx
-; AVX1-NEXT:    cmovoq %rdx, %rcx
-; AVX1-NEXT:    vpextrq $1, %xmm1, %rax
-; AVX1-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movq %rsi, %rdx
-; AVX1-NEXT:    subq %rax, %rdx
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addq %r8, %rdi
-; AVX1-NEXT:    subq %rax, %rsi
-; AVX1-NEXT:    cmovoq %rdi, %rsi
-; AVX1-NEXT:    vmovq %rsi, %xmm0
-; AVX1-NEXT:    vmovq %rcx, %xmm1
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm5, %xmm3
+; AVX1-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm1
+; AVX1-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqq %xmm1, %xmm5, %xmm1
+; AVX1-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX1-NEXT:    vpandn %xmm1, %xmm3, %xmm1
+; AVX1-NEXT:    vmovapd {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
+; AVX1-NEXT:    vblendvpd %xmm0, {{.*}}(%rip), %xmm2, %xmm2
+; AVX1-NEXT:    vblendvpd %xmm1, %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpsrad $31, %xmm0, %xmm1
 ; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
@@ -791,29 +889,23 @@
 ; AVX2-LABEL: v2i32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpsllq $32, %xmm1, %xmm1
-; AVX2-NEXT:    vmovq %xmm1, %rax
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm3
+; AVX2-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX2-NEXT:    vpxor %xmm4, %xmm3, %xmm3
 ; AVX2-NEXT:    vpsllq $32, %xmm0, %xmm0
-; AVX2-NEXT:    vmovq %xmm0, %rcx
-; AVX2-NEXT:    xorl %edx, %edx
-; AVX2-NEXT:    movq %rcx, %rsi
-; AVX2-NEXT:    subq %rax, %rsi
-; AVX2-NEXT:    setns %dl
-; AVX2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX2-NEXT:    addq %r8, %rdx
-; AVX2-NEXT:    subq %rax, %rcx
-; AVX2-NEXT:    cmovoq %rdx, %rcx
-; AVX2-NEXT:    vpextrq $1, %xmm1, %rax
-; AVX2-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movq %rsi, %rdx
-; AVX2-NEXT:    subq %rax, %rdx
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addq %r8, %rdi
-; AVX2-NEXT:    subq %rax, %rsi
-; AVX2-NEXT:    cmovoq %rdi, %rsi
-; AVX2-NEXT:    vmovq %rsi, %xmm0
-; AVX2-NEXT:    vmovq %rcx, %xmm1
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX2-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm5
+; AVX2-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX2-NEXT:    vpcmpeqq %xmm3, %xmm5, %xmm3
+; AVX2-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm1
+; AVX2-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqq %xmm1, %xmm5, %xmm1
+; AVX2-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX2-NEXT:    vpandn %xmm1, %xmm3, %xmm1
+; AVX2-NEXT:    vmovapd {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
+; AVX2-NEXT:    vblendvpd %xmm0, {{.*}}(%rip), %xmm2, %xmm2
+; AVX2-NEXT:    vblendvpd %xmm1, %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpsrad $31, %xmm0, %xmm1
 ; AVX2-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
 ; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3]
@@ -822,29 +914,19 @@
 ; AVX512-LABEL: v2i32:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vpsllq $32, %xmm1, %xmm1
-; AVX512-NEXT:    vmovq %xmm1, %rax
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltq %xmm2, %xmm1, %k0
 ; AVX512-NEXT:    vpsllq $32, %xmm0, %xmm0
-; AVX512-NEXT:    vmovq %xmm0, %rcx
-; AVX512-NEXT:    xorl %edx, %edx
-; AVX512-NEXT:    movq %rcx, %rsi
-; AVX512-NEXT:    subq %rax, %rsi
-; AVX512-NEXT:    setns %dl
-; AVX512-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX512-NEXT:    addq %r8, %rdx
-; AVX512-NEXT:    subq %rax, %rcx
-; AVX512-NEXT:    cmovoq %rdx, %rcx
-; AVX512-NEXT:    vpextrq $1, %xmm1, %rax
-; AVX512-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movq %rsi, %rdx
-; AVX512-NEXT:    subq %rax, %rdx
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addq %r8, %rdi
-; AVX512-NEXT:    subq %rax, %rsi
-; AVX512-NEXT:    cmovoq %rdi, %rsi
-; AVX512-NEXT:    vmovq %rsi, %xmm0
-; AVX512-NEXT:    vmovq %rcx, %xmm1
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX512-NEXT:    vpcmpnltq %xmm2, %xmm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vpcmpnltq %xmm2, %xmm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtq %xmm0, %xmm2, %k2
+; AVX512-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
+; AVX512-NEXT:    vmovdqa64 {{.*}}(%rip), %xmm1 {%k2}
+; AVX512-NEXT:    vmovdqa64 %xmm1, %xmm0 {%k1}
 ; AVX512-NEXT:    vpsraq $32, %xmm0, %xmm0
 ; AVX512-NEXT:    retq
   %z = call <2 x i32> @llvm.ssub.sat.v2i32(<2 x i32> %x, <2 x i32> %y)
@@ -854,197 +936,142 @@
 define <4 x i32> @v4i32(<4 x i32> %x, <4 x i32> %y) nounwind {
 ; SSE2-LABEL: v4i32:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,2,3]
-; SSE2-NEXT:    movd %xmm2, %ecx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
-; SSE2-NEXT:    movd %xmm2, %r8d
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movl %r8d, %esi
-; SSE2-NEXT:    subl %ecx, %esi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %ecx, %r8d
-; SSE2-NEXT:    cmovol %edx, %r8d
-; SSE2-NEXT:    movd %xmm1, %edx
-; SSE2-NEXT:    movd %xmm0, %ecx
-; SSE2-NEXT:    xorl %esi, %esi
-; SSE2-NEXT:    movl %ecx, %edi
-; SSE2-NEXT:    subl %edx, %edi
-; SSE2-NEXT:    setns %sil
-; SSE2-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %edx, %ecx
-; SSE2-NEXT:    cmovol %esi, %ecx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movd %xmm2, %edx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movd %xmm2, %eax
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %eax, %esi
-; SSE2-NEXT:    subl %edx, %esi
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %edx, %eax
-; SSE2-NEXT:    cmovol %edi, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r9d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSE2-NEXT:    movd %xmm0, %edx
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %edx, %esi
-; SSE2-NEXT:    subl %r9d, %esi
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %r9d, %edx
-; SSE2-NEXT:    cmovol %edi, %edx
-; SSE2-NEXT:    movd %edx, %xmm0
-; SSE2-NEXT:    movd %eax, %xmm1
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE2-NEXT:    movd %ecx, %xmm0
-; SSE2-NEXT:    movd %r8d, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE2-NEXT:    movdqa %xmm0, %xmm2
+; SSE2-NEXT:    pxor %xmm3, %xmm3
+; SSE2-NEXT:    pxor %xmm0, %xmm0
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm4
+; SSE2-NEXT:    pxor %xmm4, %xmm0
+; SSE2-NEXT:    pxor %xmm5, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSE2-NEXT:    pxor %xmm4, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm0
+; SSE2-NEXT:    psubd %xmm1, %xmm2
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm3
+; SSE2-NEXT:    movdqa %xmm3, %xmm1
+; SSE2-NEXT:    pxor %xmm4, %xmm1
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm1
+; SSE2-NEXT:    pxor %xmm4, %xmm1
+; SSE2-NEXT:    pandn %xmm1, %xmm0
+; SSE2-NEXT:    movdqa %xmm3, %xmm1
+; SSE2-NEXT:    pandn {{.*}}(%rip), %xmm1
+; SSE2-NEXT:    psrld $1, %xmm3
+; SSE2-NEXT:    por %xmm1, %xmm3
+; SSE2-NEXT:    pand %xmm0, %xmm3
+; SSE2-NEXT:    pandn %xmm2, %xmm0
+; SSE2-NEXT:    por %xmm3, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v4i32:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm2, %ecx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm2, %r8d
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movl %r8d, %esi
-; SSSE3-NEXT:    subl %ecx, %esi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %ecx, %r8d
-; SSSE3-NEXT:    cmovol %edx, %r8d
-; SSSE3-NEXT:    movd %xmm1, %edx
-; SSSE3-NEXT:    movd %xmm0, %ecx
-; SSSE3-NEXT:    xorl %esi, %esi
-; SSSE3-NEXT:    movl %ecx, %edi
-; SSSE3-NEXT:    subl %edx, %edi
-; SSSE3-NEXT:    setns %sil
-; SSSE3-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %edx, %ecx
-; SSSE3-NEXT:    cmovol %esi, %ecx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm2, %edx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm2, %eax
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %eax, %esi
-; SSSE3-NEXT:    subl %edx, %esi
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %edx, %eax
-; SSSE3-NEXT:    cmovol %edi, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r9d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm0, %edx
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %edx, %esi
-; SSSE3-NEXT:    subl %r9d, %esi
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %r9d, %edx
-; SSSE3-NEXT:    cmovol %edi, %edx
-; SSSE3-NEXT:    movd %edx, %xmm0
-; SSSE3-NEXT:    movd %eax, %xmm1
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSSE3-NEXT:    movd %ecx, %xmm0
-; SSSE3-NEXT:    movd %r8d, %xmm2
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSSE3-NEXT:    movdqa %xmm0, %xmm2
+; SSSE3-NEXT:    pxor %xmm3, %xmm3
+; SSSE3-NEXT:    pxor %xmm0, %xmm0
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm0
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm4
+; SSSE3-NEXT:    pxor %xmm4, %xmm0
+; SSSE3-NEXT:    pxor %xmm5, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSSE3-NEXT:    pxor %xmm4, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm0
+; SSSE3-NEXT:    psubd %xmm1, %xmm2
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm3
+; SSSE3-NEXT:    movdqa %xmm3, %xmm1
+; SSSE3-NEXT:    pxor %xmm4, %xmm1
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm1
+; SSSE3-NEXT:    pxor %xmm4, %xmm1
+; SSSE3-NEXT:    pandn %xmm1, %xmm0
+; SSSE3-NEXT:    movdqa %xmm3, %xmm1
+; SSSE3-NEXT:    pandn {{.*}}(%rip), %xmm1
+; SSSE3-NEXT:    psrld $1, %xmm3
+; SSSE3-NEXT:    por %xmm1, %xmm3
+; SSSE3-NEXT:    pand %xmm0, %xmm3
+; SSSE3-NEXT:    pandn %xmm2, %xmm0
+; SSSE3-NEXT:    por %xmm3, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v4i32:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    pextrd $3, %xmm1, %ecx
-; SSE41-NEXT:    pextrd $3, %xmm0, %r8d
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movl %r8d, %esi
-; SSE41-NEXT:    subl %ecx, %esi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %ecx, %r8d
-; SSE41-NEXT:    cmovol %edx, %r8d
-; SSE41-NEXT:    pextrd $2, %xmm1, %edx
-; SSE41-NEXT:    pextrd $2, %xmm0, %ecx
-; SSE41-NEXT:    xorl %esi, %esi
-; SSE41-NEXT:    movl %ecx, %edi
-; SSE41-NEXT:    subl %edx, %edi
-; SSE41-NEXT:    setns %sil
-; SSE41-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %edx, %ecx
-; SSE41-NEXT:    cmovol %esi, %ecx
-; SSE41-NEXT:    movd %xmm1, %edx
-; SSE41-NEXT:    movd %xmm0, %eax
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %eax, %esi
-; SSE41-NEXT:    subl %edx, %esi
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %edx, %eax
-; SSE41-NEXT:    cmovol %edi, %eax
-; SSE41-NEXT:    pextrd $1, %xmm1, %r9d
-; SSE41-NEXT:    pextrd $1, %xmm0, %edx
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %edx, %esi
-; SSE41-NEXT:    subl %r9d, %esi
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %r9d, %edx
-; SSE41-NEXT:    cmovol %edi, %edx
-; SSE41-NEXT:    movd %eax, %xmm0
-; SSE41-NEXT:    pinsrd $1, %edx, %xmm0
-; SSE41-NEXT:    pinsrd $2, %ecx, %xmm0
-; SSE41-NEXT:    pinsrd $3, %r8d, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm2
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pxor %xmm3, %xmm3
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm3
+; SSE41-NEXT:    pcmpeqd %xmm4, %xmm4
+; SSE41-NEXT:    pxor %xmm4, %xmm3
+; SSE41-NEXT:    pxor %xmm5, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSE41-NEXT:    pxor %xmm4, %xmm5
+; SSE41-NEXT:    pcmpeqd %xmm5, %xmm3
+; SSE41-NEXT:    psubd %xmm1, %xmm2
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm4, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm5, %xmm0
+; SSE41-NEXT:    pxor %xmm4, %xmm0
+; SSE41-NEXT:    pandn %xmm0, %xmm3
+; SSE41-NEXT:    movaps {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, {{.*}}(%rip), %xmm1
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm1, %xmm2
+; SSE41-NEXT:    movaps %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX-LABEL: v4i32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpextrd $3, %xmm1, %ecx
-; AVX-NEXT:    vpextrd $3, %xmm0, %r9d
-; AVX-NEXT:    xorl %edx, %edx
-; AVX-NEXT:    movl %r9d, %esi
-; AVX-NEXT:    subl %ecx, %esi
-; AVX-NEXT:    setns %dl
-; AVX-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; AVX-NEXT:    subl %ecx, %r9d
-; AVX-NEXT:    cmovol %edx, %r9d
-; AVX-NEXT:    vpextrd $2, %xmm1, %edx
-; AVX-NEXT:    vpextrd $2, %xmm0, %ecx
-; AVX-NEXT:    xorl %esi, %esi
-; AVX-NEXT:    movl %ecx, %edi
-; AVX-NEXT:    subl %edx, %edi
-; AVX-NEXT:    setns %sil
-; AVX-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX-NEXT:    subl %edx, %ecx
-; AVX-NEXT:    cmovol %esi, %ecx
-; AVX-NEXT:    vmovd %xmm1, %r8d
-; AVX-NEXT:    vmovd %xmm0, %edx
-; AVX-NEXT:    xorl %edi, %edi
-; AVX-NEXT:    movl %edx, %esi
-; AVX-NEXT:    subl %r8d, %esi
-; AVX-NEXT:    setns %dil
-; AVX-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX-NEXT:    subl %r8d, %edx
-; AVX-NEXT:    cmovol %edi, %edx
-; AVX-NEXT:    vpextrd $1, %xmm1, %r8d
-; AVX-NEXT:    vpextrd $1, %xmm0, %eax
-; AVX-NEXT:    xorl %esi, %esi
-; AVX-NEXT:    movl %eax, %edi
-; AVX-NEXT:    subl %r8d, %edi
-; AVX-NEXT:    setns %sil
-; AVX-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX-NEXT:    subl %r8d, %eax
-; AVX-NEXT:    cmovol %esi, %eax
-; AVX-NEXT:    vmovd %edx, %xmm0
-; AVX-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
-; AVX-NEXT:    vpinsrd $2, %ecx, %xmm0, %xmm0
-; AVX-NEXT:    vpinsrd $3, %r9d, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; AVX1-LABEL: v4i32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm2, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm2, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm5, %xmm3
+; AVX1-NEXT:    vpsubd %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm2, %xmm1
+; AVX1-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm5, %xmm1
+; AVX1-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX1-NEXT:    vpandn %xmm1, %xmm3, %xmm1
+; AVX1-NEXT:    vmovaps {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
+; AVX1-NEXT:    vblendvps %xmm0, {{.*}}(%rip), %xmm2, %xmm2
+; AVX1-NEXT:    vblendvps %xmm1, %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: v4i32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpgtd %xmm1, %xmm2, %xmm3
+; AVX2-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX2-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX2-NEXT:    vpcmpgtd %xmm0, %xmm2, %xmm5
+; AVX2-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX2-NEXT:    vpcmpeqd %xmm3, %xmm5, %xmm3
+; AVX2-NEXT:    vpsubd %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpgtd %xmm0, %xmm2, %xmm1
+; AVX2-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm5, %xmm1
+; AVX2-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX2-NEXT:    vpandn %xmm1, %xmm3, %xmm1
+; AVX2-NEXT:    vbroadcastss {{.*#+}} xmm2 = [2147483647,2147483647,2147483647,2147483647]
+; AVX2-NEXT:    vbroadcastss {{.*#+}} xmm3 = [2147483648,2147483648,2147483648,2147483648]
+; AVX2-NEXT:    vblendvps %xmm0, %xmm2, %xmm3, %xmm2
+; AVX2-NEXT:    vblendvps %xmm1, %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    retq
+;
+; AVX512-LABEL: v4i32:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltd %xmm2, %xmm1, %k0
+; AVX512-NEXT:    vpcmpnltd %xmm2, %xmm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpsubd %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vpcmpnltd %xmm2, %xmm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtd %xmm0, %xmm2, %k2
+; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
+; AVX512-NEXT:    vpbroadcastd {{.*}}(%rip), %xmm1 {%k2}
+; AVX512-NEXT:    vmovdqa32 %xmm1, %xmm0 {%k1}
+; AVX512-NEXT:    retq
   %z = call <4 x i32> @llvm.ssub.sat.v4i32(<4 x i32> %x, <4 x i32> %y)
   ret <4 x i32> %z
 }
@@ -1052,575 +1079,227 @@
 define <8 x i32> @v8i32(<8 x i32> %x, <8 x i32> %y) nounwind {
 ; SSE2-LABEL: v8i32:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pushq %rbp
-; SSE2-NEXT:    pushq %rbx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,2,3]
-; SSE2-NEXT:    movd %xmm4, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[1,1,2,3]
-; SSE2-NEXT:    movd %xmm4, %r8d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r8d, %edx
-; SSE2-NEXT:    subl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %r8d
-; SSE2-NEXT:    cmovol %ecx, %r8d
-; SSE2-NEXT:    movd %xmm3, %eax
-; SSE2-NEXT:    movd %xmm1, %r9d
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movl %r9d, %esi
-; SSE2-NEXT:    subl %eax, %esi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %r9d
-; SSE2-NEXT:    cmovol %edx, %r9d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[2,3,0,1]
-; SSE2-NEXT:    movd %xmm4, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movd %xmm4, %r10d
-; SSE2-NEXT:    xorl %esi, %esi
-; SSE2-NEXT:    movl %r10d, %edi
-; SSE2-NEXT:    subl %eax, %edi
-; SSE2-NEXT:    setns %sil
-; SSE2-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %r10d
-; SSE2-NEXT:    cmovol %esi, %r10d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[3,1,2,3]
-; SSE2-NEXT:    movd %xmm3, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r11d
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %r11d, %esi
-; SSE2-NEXT:    subl %eax, %esi
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %r11d
-; SSE2-NEXT:    cmovol %edi, %r11d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %edx
-; SSE2-NEXT:    xorl %esi, %esi
-; SSE2-NEXT:    movl %edx, %edi
-; SSE2-NEXT:    subl %eax, %edi
-; SSE2-NEXT:    setns %sil
-; SSE2-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %edx
-; SSE2-NEXT:    cmovol %esi, %edx
-; SSE2-NEXT:    movd %xmm2, %esi
-; SSE2-NEXT:    movd %xmm0, %ecx
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %ecx, %eax
-; SSE2-NEXT:    subl %esi, %eax
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %esi, %ecx
-; SSE2-NEXT:    cmovol %edi, %ecx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %esi
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %esi, %ebx
-; SSE2-NEXT:    subl %eax, %ebx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %esi
-; SSE2-NEXT:    cmovol %edi, %esi
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSE2-NEXT:    movd %xmm0, %edi
-; SSE2-NEXT:    xorl %ebx, %ebx
-; SSE2-NEXT:    movl %edi, %ebp
-; SSE2-NEXT:    subl %eax, %ebp
-; SSE2-NEXT:    setns %bl
-; SSE2-NEXT:    addl $2147483647, %ebx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %edi
-; SSE2-NEXT:    cmovol %ebx, %edi
-; SSE2-NEXT:    movd %edi, %xmm0
-; SSE2-NEXT:    movd %esi, %xmm1
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE2-NEXT:    movd %ecx, %xmm0
-; SSE2-NEXT:    movd %edx, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE2-NEXT:    movd %r11d, %xmm1
-; SSE2-NEXT:    movd %r10d, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
-; SSE2-NEXT:    movd %r9d, %xmm1
-; SSE2-NEXT:    movd %r8d, %xmm3
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; SSE2-NEXT:    popq %rbx
-; SSE2-NEXT:    popq %rbp
+; SSE2-NEXT:    movdqa %xmm0, %xmm4
+; SSE2-NEXT:    pxor %xmm5, %xmm5
+; SSE2-NEXT:    pxor %xmm0, %xmm0
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm0
+; SSE2-NEXT:    pcmpeqd %xmm8, %xmm8
+; SSE2-NEXT:    pxor %xmm8, %xmm0
+; SSE2-NEXT:    pxor %xmm7, %xmm7
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm7
+; SSE2-NEXT:    pxor %xmm8, %xmm7
+; SSE2-NEXT:    pcmpeqd %xmm7, %xmm0
+; SSE2-NEXT:    psubd %xmm2, %xmm4
+; SSE2-NEXT:    pxor %xmm6, %xmm6
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSE2-NEXT:    movdqa %xmm6, %xmm2
+; SSE2-NEXT:    pxor %xmm8, %xmm2
+; SSE2-NEXT:    pcmpeqd %xmm7, %xmm2
+; SSE2-NEXT:    pxor %xmm8, %xmm2
+; SSE2-NEXT:    pandn %xmm2, %xmm0
+; SSE2-NEXT:    movdqa {{.*#+}} xmm7 = [2147483648,2147483648,2147483648,2147483648]
+; SSE2-NEXT:    movdqa %xmm6, %xmm2
+; SSE2-NEXT:    pandn %xmm7, %xmm2
+; SSE2-NEXT:    psrld $1, %xmm6
+; SSE2-NEXT:    por %xmm2, %xmm6
+; SSE2-NEXT:    pand %xmm0, %xmm6
+; SSE2-NEXT:    pandn %xmm4, %xmm0
+; SSE2-NEXT:    por %xmm6, %xmm0
+; SSE2-NEXT:    pxor %xmm2, %xmm2
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm2
+; SSE2-NEXT:    pxor %xmm8, %xmm2
+; SSE2-NEXT:    pxor %xmm4, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE2-NEXT:    pxor %xmm8, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm2
+; SSE2-NEXT:    psubd %xmm3, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm5
+; SSE2-NEXT:    movdqa %xmm5, %xmm3
+; SSE2-NEXT:    pxor %xmm8, %xmm3
+; SSE2-NEXT:    pcmpeqd %xmm4, %xmm3
+; SSE2-NEXT:    pxor %xmm8, %xmm3
+; SSE2-NEXT:    pandn %xmm3, %xmm2
+; SSE2-NEXT:    movdqa %xmm5, %xmm3
+; SSE2-NEXT:    pandn %xmm7, %xmm3
+; SSE2-NEXT:    psrld $1, %xmm5
+; SSE2-NEXT:    por %xmm3, %xmm5
+; SSE2-NEXT:    pand %xmm2, %xmm5
+; SSE2-NEXT:    pandn %xmm1, %xmm2
+; SSE2-NEXT:    por %xmm5, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm1
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v8i32:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pushq %rbp
-; SSSE3-NEXT:    pushq %rbx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm4, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm4, %r8d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r8d, %edx
-; SSSE3-NEXT:    subl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %r8d
-; SSSE3-NEXT:    cmovol %ecx, %r8d
-; SSSE3-NEXT:    movd %xmm3, %eax
-; SSSE3-NEXT:    movd %xmm1, %r9d
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movl %r9d, %esi
-; SSSE3-NEXT:    subl %eax, %esi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %r9d
-; SSSE3-NEXT:    cmovol %edx, %r9d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm4, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm4, %r10d
-; SSSE3-NEXT:    xorl %esi, %esi
-; SSSE3-NEXT:    movl %r10d, %edi
-; SSSE3-NEXT:    subl %eax, %edi
-; SSSE3-NEXT:    setns %sil
-; SSSE3-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %r10d
-; SSSE3-NEXT:    cmovol %esi, %r10d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm3, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r11d
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %r11d, %esi
-; SSSE3-NEXT:    subl %eax, %esi
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %r11d
-; SSSE3-NEXT:    cmovol %edi, %r11d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %edx
-; SSSE3-NEXT:    xorl %esi, %esi
-; SSSE3-NEXT:    movl %edx, %edi
-; SSSE3-NEXT:    subl %eax, %edi
-; SSSE3-NEXT:    setns %sil
-; SSSE3-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %edx
-; SSSE3-NEXT:    cmovol %esi, %edx
-; SSSE3-NEXT:    movd %xmm2, %esi
-; SSSE3-NEXT:    movd %xmm0, %ecx
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %ecx, %eax
-; SSSE3-NEXT:    subl %esi, %eax
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %esi, %ecx
-; SSSE3-NEXT:    cmovol %edi, %ecx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %esi
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %esi, %ebx
-; SSSE3-NEXT:    subl %eax, %ebx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %esi
-; SSSE3-NEXT:    cmovol %edi, %esi
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm0, %edi
-; SSSE3-NEXT:    xorl %ebx, %ebx
-; SSSE3-NEXT:    movl %edi, %ebp
-; SSSE3-NEXT:    subl %eax, %ebp
-; SSSE3-NEXT:    setns %bl
-; SSSE3-NEXT:    addl $2147483647, %ebx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %edi
-; SSSE3-NEXT:    cmovol %ebx, %edi
-; SSSE3-NEXT:    movd %edi, %xmm0
-; SSSE3-NEXT:    movd %esi, %xmm1
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSSE3-NEXT:    movd %ecx, %xmm0
-; SSSE3-NEXT:    movd %edx, %xmm2
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSSE3-NEXT:    movd %r11d, %xmm1
-; SSSE3-NEXT:    movd %r10d, %xmm2
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
-; SSSE3-NEXT:    movd %r9d, %xmm1
-; SSSE3-NEXT:    movd %r8d, %xmm3
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; SSSE3-NEXT:    popq %rbx
-; SSSE3-NEXT:    popq %rbp
+; SSSE3-NEXT:    movdqa %xmm0, %xmm4
+; SSSE3-NEXT:    pxor %xmm5, %xmm5
+; SSSE3-NEXT:    pxor %xmm0, %xmm0
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm0
+; SSSE3-NEXT:    pcmpeqd %xmm8, %xmm8
+; SSSE3-NEXT:    pxor %xmm8, %xmm0
+; SSSE3-NEXT:    pxor %xmm7, %xmm7
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm7
+; SSSE3-NEXT:    pxor %xmm8, %xmm7
+; SSSE3-NEXT:    pcmpeqd %xmm7, %xmm0
+; SSSE3-NEXT:    psubd %xmm2, %xmm4
+; SSSE3-NEXT:    pxor %xmm6, %xmm6
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSSE3-NEXT:    movdqa %xmm6, %xmm2
+; SSSE3-NEXT:    pxor %xmm8, %xmm2
+; SSSE3-NEXT:    pcmpeqd %xmm7, %xmm2
+; SSSE3-NEXT:    pxor %xmm8, %xmm2
+; SSSE3-NEXT:    pandn %xmm2, %xmm0
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm7 = [2147483648,2147483648,2147483648,2147483648]
+; SSSE3-NEXT:    movdqa %xmm6, %xmm2
+; SSSE3-NEXT:    pandn %xmm7, %xmm2
+; SSSE3-NEXT:    psrld $1, %xmm6
+; SSSE3-NEXT:    por %xmm2, %xmm6
+; SSSE3-NEXT:    pand %xmm0, %xmm6
+; SSSE3-NEXT:    pandn %xmm4, %xmm0
+; SSSE3-NEXT:    por %xmm6, %xmm0
+; SSSE3-NEXT:    pxor %xmm2, %xmm2
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm2
+; SSSE3-NEXT:    pxor %xmm8, %xmm2
+; SSSE3-NEXT:    pxor %xmm4, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSSE3-NEXT:    pxor %xmm8, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm2
+; SSSE3-NEXT:    psubd %xmm3, %xmm1
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm5
+; SSSE3-NEXT:    movdqa %xmm5, %xmm3
+; SSSE3-NEXT:    pxor %xmm8, %xmm3
+; SSSE3-NEXT:    pcmpeqd %xmm4, %xmm3
+; SSSE3-NEXT:    pxor %xmm8, %xmm3
+; SSSE3-NEXT:    pandn %xmm3, %xmm2
+; SSSE3-NEXT:    movdqa %xmm5, %xmm3
+; SSSE3-NEXT:    pandn %xmm7, %xmm3
+; SSSE3-NEXT:    psrld $1, %xmm5
+; SSSE3-NEXT:    por %xmm3, %xmm5
+; SSSE3-NEXT:    pand %xmm2, %xmm5
+; SSSE3-NEXT:    pandn %xmm1, %xmm2
+; SSSE3-NEXT:    por %xmm5, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm1
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v8i32:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    pushq %rbp
-; SSE41-NEXT:    pushq %rbx
-; SSE41-NEXT:    pextrd $3, %xmm3, %eax
-; SSE41-NEXT:    pextrd $3, %xmm1, %r8d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r8d, %edx
-; SSE41-NEXT:    subl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %r8d
-; SSE41-NEXT:    cmovol %ecx, %r8d
-; SSE41-NEXT:    pextrd $2, %xmm3, %eax
-; SSE41-NEXT:    pextrd $2, %xmm1, %r9d
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movl %r9d, %esi
-; SSE41-NEXT:    subl %eax, %esi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %r9d
-; SSE41-NEXT:    cmovol %edx, %r9d
-; SSE41-NEXT:    movd %xmm3, %eax
-; SSE41-NEXT:    movd %xmm1, %r10d
-; SSE41-NEXT:    xorl %esi, %esi
-; SSE41-NEXT:    movl %r10d, %edi
-; SSE41-NEXT:    subl %eax, %edi
-; SSE41-NEXT:    setns %sil
-; SSE41-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %r10d
-; SSE41-NEXT:    cmovol %esi, %r10d
-; SSE41-NEXT:    pextrd $1, %xmm3, %eax
-; SSE41-NEXT:    pextrd $1, %xmm1, %r11d
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %r11d, %esi
-; SSE41-NEXT:    subl %eax, %esi
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %r11d
-; SSE41-NEXT:    cmovol %edi, %r11d
-; SSE41-NEXT:    pextrd $3, %xmm2, %eax
-; SSE41-NEXT:    pextrd $3, %xmm0, %edx
-; SSE41-NEXT:    xorl %esi, %esi
-; SSE41-NEXT:    movl %edx, %edi
-; SSE41-NEXT:    subl %eax, %edi
-; SSE41-NEXT:    setns %sil
-; SSE41-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %edx
-; SSE41-NEXT:    cmovol %esi, %edx
-; SSE41-NEXT:    pextrd $2, %xmm2, %esi
-; SSE41-NEXT:    pextrd $2, %xmm0, %ecx
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %ecx, %eax
-; SSE41-NEXT:    subl %esi, %eax
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %esi, %ecx
-; SSE41-NEXT:    cmovol %edi, %ecx
-; SSE41-NEXT:    movd %xmm2, %eax
-; SSE41-NEXT:    movd %xmm0, %esi
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %esi, %ebx
-; SSE41-NEXT:    subl %eax, %ebx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %esi
-; SSE41-NEXT:    cmovol %edi, %esi
-; SSE41-NEXT:    pextrd $1, %xmm2, %eax
-; SSE41-NEXT:    pextrd $1, %xmm0, %edi
-; SSE41-NEXT:    xorl %ebx, %ebx
-; SSE41-NEXT:    movl %edi, %ebp
-; SSE41-NEXT:    subl %eax, %ebp
-; SSE41-NEXT:    setns %bl
-; SSE41-NEXT:    addl $2147483647, %ebx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %edi
-; SSE41-NEXT:    cmovol %ebx, %edi
-; SSE41-NEXT:    movd %esi, %xmm0
-; SSE41-NEXT:    pinsrd $1, %edi, %xmm0
-; SSE41-NEXT:    pinsrd $2, %ecx, %xmm0
-; SSE41-NEXT:    pinsrd $3, %edx, %xmm0
-; SSE41-NEXT:    movd %r10d, %xmm1
-; SSE41-NEXT:    pinsrd $1, %r11d, %xmm1
-; SSE41-NEXT:    pinsrd $2, %r9d, %xmm1
-; SSE41-NEXT:    pinsrd $3, %r8d, %xmm1
-; SSE41-NEXT:    popq %rbx
-; SSE41-NEXT:    popq %rbp
+; SSE41-NEXT:    movdqa %xmm0, %xmm4
+; SSE41-NEXT:    pxor %xmm6, %xmm6
+; SSE41-NEXT:    pxor %xmm5, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSE41-NEXT:    pcmpeqd %xmm8, %xmm8
+; SSE41-NEXT:    pxor %xmm8, %xmm5
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm4, %xmm0
+; SSE41-NEXT:    pxor %xmm8, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm5
+; SSE41-NEXT:    psubd %xmm2, %xmm4
+; SSE41-NEXT:    pxor %xmm2, %xmm2
+; SSE41-NEXT:    pcmpgtd %xmm4, %xmm2
+; SSE41-NEXT:    pxor %xmm8, %xmm2
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm2
+; SSE41-NEXT:    pxor %xmm8, %xmm2
+; SSE41-NEXT:    pandn %xmm2, %xmm5
+; SSE41-NEXT:    movaps {{.*#+}} xmm9 = [2147483647,2147483647,2147483647,2147483647]
+; SSE41-NEXT:    movaps {{.*#+}} xmm7 = [2147483648,2147483648,2147483648,2147483648]
+; SSE41-NEXT:    movaps %xmm7, %xmm2
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm9, %xmm2
+; SSE41-NEXT:    movdqa %xmm5, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm2, %xmm4
+; SSE41-NEXT:    xorps %xmm2, %xmm2
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm2
+; SSE41-NEXT:    pxor %xmm8, %xmm2
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
+; SSE41-NEXT:    pxor %xmm8, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm2
+; SSE41-NEXT:    psubd %xmm3, %xmm1
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm6
+; SSE41-NEXT:    pxor %xmm8, %xmm6
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm6
+; SSE41-NEXT:    pxor %xmm8, %xmm6
+; SSE41-NEXT:    pandn %xmm6, %xmm2
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm9, %xmm7
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm7, %xmm1
+; SSE41-NEXT:    movaps %xmm4, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: v8i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    pushq %rbp
-; AVX1-NEXT:    pushq %rbx
-; AVX1-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX1-NEXT:    vpextrd $3, %xmm0, %r8d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r8d, %edx
-; AVX1-NEXT:    subl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %r8d
-; AVX1-NEXT:    cmovol %ecx, %r8d
-; AVX1-NEXT:    vpextrd $2, %xmm1, %eax
-; AVX1-NEXT:    vpextrd $2, %xmm0, %r9d
-; AVX1-NEXT:    xorl %edx, %edx
-; AVX1-NEXT:    movl %r9d, %esi
-; AVX1-NEXT:    subl %eax, %esi
-; AVX1-NEXT:    setns %dl
-; AVX1-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %r9d
-; AVX1-NEXT:    cmovol %edx, %r9d
-; AVX1-NEXT:    vmovd %xmm1, %eax
-; AVX1-NEXT:    vmovd %xmm0, %r10d
-; AVX1-NEXT:    xorl %esi, %esi
-; AVX1-NEXT:    movl %r10d, %edi
-; AVX1-NEXT:    subl %eax, %edi
-; AVX1-NEXT:    setns %sil
-; AVX1-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %r10d
-; AVX1-NEXT:    cmovol %esi, %r10d
-; AVX1-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX1-NEXT:    vpextrd $1, %xmm0, %r11d
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %r11d, %esi
-; AVX1-NEXT:    subl %eax, %esi
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %r11d
-; AVX1-NEXT:    cmovol %edi, %r11d
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpextrd $3, %xmm0, %edx
-; AVX1-NEXT:    xorl %esi, %esi
-; AVX1-NEXT:    movl %edx, %edi
-; AVX1-NEXT:    subl %eax, %edi
-; AVX1-NEXT:    setns %sil
-; AVX1-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %edx
-; AVX1-NEXT:    cmovol %esi, %edx
-; AVX1-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX1-NEXT:    vpextrd $2, %xmm0, %ecx
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %ecx, %eax
-; AVX1-NEXT:    subl %esi, %eax
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %esi, %ecx
-; AVX1-NEXT:    cmovol %edi, %ecx
-; AVX1-NEXT:    vmovd %xmm1, %eax
-; AVX1-NEXT:    vmovd %xmm0, %esi
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %esi, %ebx
-; AVX1-NEXT:    subl %eax, %ebx
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %esi
-; AVX1-NEXT:    cmovol %edi, %esi
-; AVX1-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX1-NEXT:    vpextrd $1, %xmm0, %edi
-; AVX1-NEXT:    xorl %ebx, %ebx
-; AVX1-NEXT:    movl %edi, %ebp
-; AVX1-NEXT:    subl %eax, %ebp
-; AVX1-NEXT:    setns %bl
-; AVX1-NEXT:    addl $2147483647, %ebx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %edi
-; AVX1-NEXT:    cmovol %ebx, %edi
-; AVX1-NEXT:    vmovd %esi, %xmm0
-; AVX1-NEXT:    vpinsrd $1, %edi, %xmm0, %xmm0
-; AVX1-NEXT:    vpinsrd $2, %ecx, %xmm0, %xmm0
-; AVX1-NEXT:    vpinsrd $3, %edx, %xmm0, %xmm0
-; AVX1-NEXT:    vmovd %r10d, %xmm1
-; AVX1-NEXT:    vpinsrd $1, %r11d, %xmm1, %xmm1
-; AVX1-NEXT:    vpinsrd $2, %r9d, %xmm1, %xmm1
-; AVX1-NEXT:    vpinsrd $3, %r8d, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT:    popq %rbx
-; AVX1-NEXT:    popq %rbp
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm9
+; AVX1-NEXT:    vpxor %xmm10, %xmm10, %xmm10
+; AVX1-NEXT:    vpcmpgtd %xmm9, %xmm10, %xmm4
+; AVX1-NEXT:    vpcmpeqd %xmm5, %xmm5, %xmm5
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm4
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm6
+; AVX1-NEXT:    vpcmpgtd %xmm6, %xmm10, %xmm7
+; AVX1-NEXT:    vpxor %xmm5, %xmm7, %xmm7
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm7, %xmm4
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm8
+; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm10, %xmm4
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm4
+; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm10, %xmm2
+; AVX1-NEXT:    vpxor %xmm5, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm2, %xmm4
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm4
+; AVX1-NEXT:    vinsertf128 $1, %xmm8, %ymm4, %ymm8
+; AVX1-NEXT:    vpsubd %xmm9, %xmm6, %xmm6
+; AVX1-NEXT:    vpcmpgtd %xmm6, %xmm10, %xmm4
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm7, %xmm3
+; AVX1-NEXT:    vpxor %xmm5, %xmm3, %xmm3
+; AVX1-NEXT:    vpsubd %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm10, %xmm1
+; AVX1-NEXT:    vpxor %xmm5, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm2, %xmm1
+; AVX1-NEXT:    vpxor %xmm5, %xmm1, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
+; AVX1-NEXT:    vandps %ymm1, %ymm8, %ymm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm2
+; AVX1-NEXT:    vmovaps {{.*#+}} ymm3 = [2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648]
+; AVX1-NEXT:    vblendvps %ymm2, {{.*}}(%rip), %ymm3, %ymm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm0
+; AVX1-NEXT:    vblendvps %ymm1, %ymm2, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: v8i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    pushq %rbp
-; AVX2-NEXT:    pushq %rbx
-; AVX2-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX2-NEXT:    vpextrd $3, %xmm0, %r8d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r8d, %edx
-; AVX2-NEXT:    subl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %r8d
-; AVX2-NEXT:    cmovol %ecx, %r8d
-; AVX2-NEXT:    vpextrd $2, %xmm1, %eax
-; AVX2-NEXT:    vpextrd $2, %xmm0, %r9d
-; AVX2-NEXT:    xorl %edx, %edx
-; AVX2-NEXT:    movl %r9d, %esi
-; AVX2-NEXT:    subl %eax, %esi
-; AVX2-NEXT:    setns %dl
-; AVX2-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %r9d
-; AVX2-NEXT:    cmovol %edx, %r9d
-; AVX2-NEXT:    vmovd %xmm1, %eax
-; AVX2-NEXT:    vmovd %xmm0, %r10d
-; AVX2-NEXT:    xorl %esi, %esi
-; AVX2-NEXT:    movl %r10d, %edi
-; AVX2-NEXT:    subl %eax, %edi
-; AVX2-NEXT:    setns %sil
-; AVX2-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %r10d
-; AVX2-NEXT:    cmovol %esi, %r10d
-; AVX2-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX2-NEXT:    vpextrd $1, %xmm0, %r11d
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %r11d, %esi
-; AVX2-NEXT:    subl %eax, %esi
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %r11d
-; AVX2-NEXT:    cmovol %edi, %r11d
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; AVX2-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; AVX2-NEXT:    vpextrd $3, %xmm0, %edx
-; AVX2-NEXT:    xorl %esi, %esi
-; AVX2-NEXT:    movl %edx, %edi
-; AVX2-NEXT:    subl %eax, %edi
-; AVX2-NEXT:    setns %sil
-; AVX2-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %edx
-; AVX2-NEXT:    cmovol %esi, %edx
-; AVX2-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX2-NEXT:    vpextrd $2, %xmm0, %ecx
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %ecx, %eax
-; AVX2-NEXT:    subl %esi, %eax
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %esi, %ecx
-; AVX2-NEXT:    cmovol %edi, %ecx
-; AVX2-NEXT:    vmovd %xmm1, %eax
-; AVX2-NEXT:    vmovd %xmm0, %esi
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %esi, %ebx
-; AVX2-NEXT:    subl %eax, %ebx
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %esi
-; AVX2-NEXT:    cmovol %edi, %esi
-; AVX2-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX2-NEXT:    vpextrd $1, %xmm0, %edi
-; AVX2-NEXT:    xorl %ebx, %ebx
-; AVX2-NEXT:    movl %edi, %ebp
-; AVX2-NEXT:    subl %eax, %ebp
-; AVX2-NEXT:    setns %bl
-; AVX2-NEXT:    addl $2147483647, %ebx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %edi
-; AVX2-NEXT:    cmovol %ebx, %edi
-; AVX2-NEXT:    vmovd %esi, %xmm0
-; AVX2-NEXT:    vpinsrd $1, %edi, %xmm0, %xmm0
-; AVX2-NEXT:    vpinsrd $2, %ecx, %xmm0, %xmm0
-; AVX2-NEXT:    vpinsrd $3, %edx, %xmm0, %xmm0
-; AVX2-NEXT:    vmovd %r10d, %xmm1
-; AVX2-NEXT:    vpinsrd $1, %r11d, %xmm1, %xmm1
-; AVX2-NEXT:    vpinsrd $2, %r9d, %xmm1, %xmm1
-; AVX2-NEXT:    vpinsrd $3, %r8d, %xmm1, %xmm1
-; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX2-NEXT:    popq %rbx
-; AVX2-NEXT:    popq %rbp
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpgtd %ymm1, %ymm2, %ymm3
+; AVX2-NEXT:    vpcmpeqd %ymm4, %ymm4, %ymm4
+; AVX2-NEXT:    vpxor %ymm4, %ymm3, %ymm3
+; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm2, %ymm5
+; AVX2-NEXT:    vpxor %ymm4, %ymm5, %ymm5
+; AVX2-NEXT:    vpcmpeqd %ymm3, %ymm5, %ymm3
+; AVX2-NEXT:    vpsubd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm2, %ymm1
+; AVX2-NEXT:    vpxor %ymm4, %ymm1, %ymm1
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm5, %ymm1
+; AVX2-NEXT:    vpxor %ymm4, %ymm1, %ymm1
+; AVX2-NEXT:    vpandn %ymm1, %ymm3, %ymm1
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [2147483647,2147483647,2147483647,2147483647,2147483647,2147483647,2147483647,2147483647]
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm3 = [2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648]
+; AVX2-NEXT:    vblendvps %ymm0, %ymm2, %ymm3, %ymm2
+; AVX2-NEXT:    vblendvps %ymm1, %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: v8i32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    pushq %rbp
-; AVX512-NEXT:    pushq %rbx
-; AVX512-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $3, %xmm0, %r8d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r8d, %edx
-; AVX512-NEXT:    subl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %r8d
-; AVX512-NEXT:    cmovol %ecx, %r8d
-; AVX512-NEXT:    vpextrd $2, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $2, %xmm0, %r9d
-; AVX512-NEXT:    xorl %edx, %edx
-; AVX512-NEXT:    movl %r9d, %esi
-; AVX512-NEXT:    subl %eax, %esi
-; AVX512-NEXT:    setns %dl
-; AVX512-NEXT:    addl $2147483647, %edx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %r9d
-; AVX512-NEXT:    cmovol %edx, %r9d
-; AVX512-NEXT:    vmovd %xmm1, %eax
-; AVX512-NEXT:    vmovd %xmm0, %r10d
-; AVX512-NEXT:    xorl %esi, %esi
-; AVX512-NEXT:    movl %r10d, %edi
-; AVX512-NEXT:    subl %eax, %edi
-; AVX512-NEXT:    setns %sil
-; AVX512-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %r10d
-; AVX512-NEXT:    cmovol %esi, %r10d
-; AVX512-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $1, %xmm0, %r11d
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %r11d, %esi
-; AVX512-NEXT:    subl %eax, %esi
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %r11d
-; AVX512-NEXT:    cmovol %edi, %r11d
-; AVX512-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; AVX512-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; AVX512-NEXT:    vpextrd $3, %xmm0, %edx
-; AVX512-NEXT:    xorl %esi, %esi
-; AVX512-NEXT:    movl %edx, %edi
-; AVX512-NEXT:    subl %eax, %edi
-; AVX512-NEXT:    setns %sil
-; AVX512-NEXT:    addl $2147483647, %esi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %edx
-; AVX512-NEXT:    cmovol %esi, %edx
-; AVX512-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX512-NEXT:    vpextrd $2, %xmm0, %ecx
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %ecx, %eax
-; AVX512-NEXT:    subl %esi, %eax
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %esi, %ecx
-; AVX512-NEXT:    cmovol %edi, %ecx
-; AVX512-NEXT:    vmovd %xmm1, %eax
-; AVX512-NEXT:    vmovd %xmm0, %esi
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %esi, %ebx
-; AVX512-NEXT:    subl %eax, %ebx
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %esi
-; AVX512-NEXT:    cmovol %edi, %esi
-; AVX512-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $1, %xmm0, %edi
-; AVX512-NEXT:    xorl %ebx, %ebx
-; AVX512-NEXT:    movl %edi, %ebp
-; AVX512-NEXT:    subl %eax, %ebp
-; AVX512-NEXT:    setns %bl
-; AVX512-NEXT:    addl $2147483647, %ebx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %edi
-; AVX512-NEXT:    cmovol %ebx, %edi
-; AVX512-NEXT:    vmovd %esi, %xmm0
-; AVX512-NEXT:    vpinsrd $1, %edi, %xmm0, %xmm0
-; AVX512-NEXT:    vpinsrd $2, %ecx, %xmm0, %xmm0
-; AVX512-NEXT:    vpinsrd $3, %edx, %xmm0, %xmm0
-; AVX512-NEXT:    vmovd %r10d, %xmm1
-; AVX512-NEXT:    vpinsrd $1, %r11d, %xmm1, %xmm1
-; AVX512-NEXT:    vpinsrd $2, %r9d, %xmm1, %xmm1
-; AVX512-NEXT:    vpinsrd $3, %r8d, %xmm1, %xmm1
-; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX512-NEXT:    popq %rbx
-; AVX512-NEXT:    popq %rbp
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltd %ymm2, %ymm1, %k0
+; AVX512-NEXT:    vpcmpnltd %ymm2, %ymm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpsubd %ymm1, %ymm0, %ymm0
+; AVX512-NEXT:    vpcmpnltd %ymm2, %ymm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtd %ymm0, %ymm2, %k2
+; AVX512-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648]
+; AVX512-NEXT:    vpbroadcastd {{.*}}(%rip), %ymm1 {%k2}
+; AVX512-NEXT:    vmovdqa32 %ymm1, %ymm0 {%k1}
 ; AVX512-NEXT:    retq
   %z = call <8 x i32> @llvm.ssub.sat.v8i32(<8 x i32> %x, <8 x i32> %y)
   ret <8 x i32> %z
@@ -1629,1187 +1308,402 @@
 define <16 x i32> @v16i32(<16 x i32> %x, <16 x i32> %y) nounwind {
 ; SSE2-LABEL: v16i32:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pushq %rbp
-; SSE2-NEXT:    pushq %r15
-; SSE2-NEXT:    pushq %r14
-; SSE2-NEXT:    pushq %r13
-; SSE2-NEXT:    pushq %r12
-; SSE2-NEXT:    pushq %rbx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[1,1,2,3]
-; SSE2-NEXT:    movd %xmm8, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[1,1,2,3]
-; SSE2-NEXT:    movd %xmm8, %esi
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %esi, %edx
-; SSE2-NEXT:    subl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %esi
-; SSE2-NEXT:    cmovol %ecx, %esi
-; SSE2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE2-NEXT:    movd %xmm5, %eax
-; SSE2-NEXT:    movd %xmm1, %esi
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %esi, %edx
-; SSE2-NEXT:    subl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %esi
-; SSE2-NEXT:    cmovol %ecx, %esi
-; SSE2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[2,3,0,1]
-; SSE2-NEXT:    movd %xmm8, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movd %xmm8, %esi
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %esi, %edx
-; SSE2-NEXT:    subl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %esi
-; SSE2-NEXT:    cmovol %ecx, %esi
-; SSE2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[3,1,2,3]
-; SSE2-NEXT:    movd %xmm5, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %esi
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %esi, %edx
-; SSE2-NEXT:    subl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %esi
-; SSE2-NEXT:    cmovol %ecx, %esi
-; SSE2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r14d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r14d, %edx
-; SSE2-NEXT:    subl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %r14d
-; SSE2-NEXT:    cmovol %ecx, %r14d
-; SSE2-NEXT:    movd %xmm6, %eax
-; SSE2-NEXT:    movd %xmm2, %r15d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r15d, %edx
-; SSE2-NEXT:    subl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %r15d
-; SSE2-NEXT:    cmovol %ecx, %r15d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %r12d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r12d, %edx
-; SSE2-NEXT:    subl %eax, %edx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %r12d
-; SSE2-NEXT:    cmovol %ecx, %r12d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r13d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r13d, %esi
-; SSE2-NEXT:    subl %eax, %esi
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %r13d
-; SSE2-NEXT:    cmovol %ecx, %r13d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r11d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r11d, %edi
-; SSE2-NEXT:    subl %eax, %edi
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %r11d
-; SSE2-NEXT:    cmovol %ecx, %r11d
-; SSE2-NEXT:    movd %xmm7, %eax
-; SSE2-NEXT:    movd %xmm3, %ebx
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %ebx, %edi
-; SSE2-NEXT:    subl %eax, %edi
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %ebx
-; SSE2-NEXT:    cmovol %ecx, %ebx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %ebp
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %ebp, %edi
-; SSE2-NEXT:    subl %eax, %edi
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %ebp
-; SSE2-NEXT:    cmovol %ecx, %ebp
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r10d
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movl %r10d, %edi
-; SSE2-NEXT:    subl %eax, %edi
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %r10d
-; SSE2-NEXT:    cmovol %ecx, %r10d
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %ecx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r9d
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %r9d, %eax
-; SSE2-NEXT:    subl %ecx, %eax
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %ecx, %r9d
-; SSE2-NEXT:    cmovol %edi, %r9d
-; SSE2-NEXT:    movd %xmm4, %edx
-; SSE2-NEXT:    movd %xmm0, %esi
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %esi, %ecx
-; SSE2-NEXT:    subl %edx, %ecx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %edx, %esi
-; SSE2-NEXT:    cmovol %edi, %esi
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %eax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movd %xmm1, %ecx
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %ecx, %edx
-; SSE2-NEXT:    subl %eax, %edx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %eax, %ecx
-; SSE2-NEXT:    cmovol %edi, %ecx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[3,1,2,3]
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSE2-NEXT:    movd %xmm1, %r8d
-; SSE2-NEXT:    movd %xmm0, %eax
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movl %eax, %edx
-; SSE2-NEXT:    subl %r8d, %edx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE2-NEXT:    subl %r8d, %eax
-; SSE2-NEXT:    cmovol %edi, %eax
-; SSE2-NEXT:    movd %eax, %xmm8
-; SSE2-NEXT:    movd %ecx, %xmm5
-; SSE2-NEXT:    movd %esi, %xmm0
-; SSE2-NEXT:    movd %r9d, %xmm9
-; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 4-byte Reload
-; SSE2-NEXT:    # xmm10 = mem[0],zero,zero,zero
-; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 4-byte Reload
-; SSE2-NEXT:    # xmm4 = mem[0],zero,zero,zero
-; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
-; SSE2-NEXT:    # xmm1 = mem[0],zero,zero,zero
-; SSE2-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 4-byte Reload
-; SSE2-NEXT:    # xmm11 = mem[0],zero,zero,zero
-; SSE2-NEXT:    movd %r13d, %xmm12
-; SSE2-NEXT:    movd %r12d, %xmm6
-; SSE2-NEXT:    movd %r15d, %xmm2
-; SSE2-NEXT:    movd %r14d, %xmm13
-; SSE2-NEXT:    movd %r10d, %xmm14
-; SSE2-NEXT:    movd %ebp, %xmm7
-; SSE2-NEXT:    movd %ebx, %xmm3
-; SSE2-NEXT:    movd %r11d, %xmm15
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm5 = xmm5[0],xmm8[0],xmm5[1],xmm8[1]
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1]
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm5[0]
-; SSE2-NEXT:    unpcklps {{.*#+}} xmm4 = xmm4[0],xmm10[0],xmm4[1],xmm10[1]
-; SSE2-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm11[0],xmm1[1],xmm11[1]
-; SSE2-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm4[0]
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm6 = xmm6[0],xmm12[0],xmm6[1],xmm12[1]
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm13[0],xmm2[1],xmm13[1]
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm6[0]
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm7 = xmm7[0],xmm14[0],xmm7[1],xmm14[1]
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm15[0],xmm3[1],xmm15[1]
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm7[0]
-; SSE2-NEXT:    popq %rbx
-; SSE2-NEXT:    popq %r12
-; SSE2-NEXT:    popq %r13
-; SSE2-NEXT:    popq %r14
-; SSE2-NEXT:    popq %r15
-; SSE2-NEXT:    popq %rbp
+; SSE2-NEXT:    movdqa %xmm1, %xmm8
+; SSE2-NEXT:    movdqa %xmm0, %xmm12
+; SSE2-NEXT:    pxor %xmm9, %xmm9
+; SSE2-NEXT:    pxor %xmm0, %xmm0
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm0
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm10
+; SSE2-NEXT:    pxor %xmm10, %xmm0
+; SSE2-NEXT:    pxor %xmm11, %xmm11
+; SSE2-NEXT:    pcmpgtd %xmm12, %xmm11
+; SSE2-NEXT:    pxor %xmm10, %xmm11
+; SSE2-NEXT:    pcmpeqd %xmm11, %xmm0
+; SSE2-NEXT:    psubd %xmm4, %xmm12
+; SSE2-NEXT:    pxor %xmm1, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm12, %xmm1
+; SSE2-NEXT:    movdqa %xmm1, %xmm4
+; SSE2-NEXT:    pxor %xmm10, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm11, %xmm4
+; SSE2-NEXT:    pxor %xmm10, %xmm4
+; SSE2-NEXT:    pandn %xmm4, %xmm0
+; SSE2-NEXT:    movdqa {{.*#+}} xmm11 = [2147483648,2147483648,2147483648,2147483648]
+; SSE2-NEXT:    movdqa %xmm1, %xmm4
+; SSE2-NEXT:    pandn %xmm11, %xmm4
+; SSE2-NEXT:    psrld $1, %xmm1
+; SSE2-NEXT:    por %xmm4, %xmm1
+; SSE2-NEXT:    pand %xmm0, %xmm1
+; SSE2-NEXT:    pandn %xmm12, %xmm0
+; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    pxor %xmm1, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm1
+; SSE2-NEXT:    pxor %xmm10, %xmm1
+; SSE2-NEXT:    pxor %xmm12, %xmm12
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm12
+; SSE2-NEXT:    pxor %xmm10, %xmm12
+; SSE2-NEXT:    pcmpeqd %xmm12, %xmm1
+; SSE2-NEXT:    psubd %xmm5, %xmm8
+; SSE2-NEXT:    pxor %xmm4, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm8, %xmm4
+; SSE2-NEXT:    movdqa %xmm4, %xmm5
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm12, %xmm5
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pandn %xmm5, %xmm1
+; SSE2-NEXT:    movdqa %xmm4, %xmm5
+; SSE2-NEXT:    pandn %xmm11, %xmm5
+; SSE2-NEXT:    psrld $1, %xmm4
+; SSE2-NEXT:    por %xmm5, %xmm4
+; SSE2-NEXT:    pand %xmm1, %xmm4
+; SSE2-NEXT:    pandn %xmm8, %xmm1
+; SSE2-NEXT:    por %xmm4, %xmm1
+; SSE2-NEXT:    pxor %xmm8, %xmm8
+; SSE2-NEXT:    pcmpgtd %xmm6, %xmm8
+; SSE2-NEXT:    pxor %xmm10, %xmm8
+; SSE2-NEXT:    pxor %xmm5, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm8
+; SSE2-NEXT:    psubd %xmm6, %xmm2
+; SSE2-NEXT:    pxor %xmm4, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm4
+; SSE2-NEXT:    movdqa %xmm4, %xmm6
+; SSE2-NEXT:    pxor %xmm10, %xmm6
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm6
+; SSE2-NEXT:    pxor %xmm10, %xmm6
+; SSE2-NEXT:    pandn %xmm6, %xmm8
+; SSE2-NEXT:    movdqa %xmm4, %xmm5
+; SSE2-NEXT:    pandn %xmm11, %xmm5
+; SSE2-NEXT:    psrld $1, %xmm4
+; SSE2-NEXT:    por %xmm5, %xmm4
+; SSE2-NEXT:    pand %xmm8, %xmm4
+; SSE2-NEXT:    pandn %xmm2, %xmm8
+; SSE2-NEXT:    por %xmm4, %xmm8
+; SSE2-NEXT:    pxor %xmm5, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm5
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pxor %xmm2, %xmm2
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm2
+; SSE2-NEXT:    pxor %xmm10, %xmm2
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm5
+; SSE2-NEXT:    psubd %xmm7, %xmm3
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm9
+; SSE2-NEXT:    movdqa %xmm9, %xmm4
+; SSE2-NEXT:    pxor %xmm10, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm4
+; SSE2-NEXT:    pxor %xmm10, %xmm4
+; SSE2-NEXT:    pandn %xmm4, %xmm5
+; SSE2-NEXT:    movdqa %xmm9, %xmm2
+; SSE2-NEXT:    pandn %xmm11, %xmm2
+; SSE2-NEXT:    psrld $1, %xmm9
+; SSE2-NEXT:    por %xmm2, %xmm9
+; SSE2-NEXT:    pand %xmm5, %xmm9
+; SSE2-NEXT:    pandn %xmm3, %xmm5
+; SSE2-NEXT:    por %xmm9, %xmm5
+; SSE2-NEXT:    movdqa %xmm8, %xmm2
+; SSE2-NEXT:    movdqa %xmm5, %xmm3
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v16i32:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pushq %rbp
-; SSSE3-NEXT:    pushq %r15
-; SSSE3-NEXT:    pushq %r14
-; SSSE3-NEXT:    pushq %r13
-; SSSE3-NEXT:    pushq %r12
-; SSSE3-NEXT:    pushq %rbx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm8, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm8, %esi
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %esi, %edx
-; SSSE3-NEXT:    subl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %esi
-; SSSE3-NEXT:    cmovol %ecx, %esi
-; SSSE3-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSSE3-NEXT:    movd %xmm5, %eax
-; SSSE3-NEXT:    movd %xmm1, %esi
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %esi, %edx
-; SSSE3-NEXT:    subl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %esi
-; SSSE3-NEXT:    cmovol %ecx, %esi
-; SSSE3-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm8, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm8, %esi
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %esi, %edx
-; SSSE3-NEXT:    subl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %esi
-; SSSE3-NEXT:    cmovol %ecx, %esi
-; SSSE3-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm5, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %esi
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %esi, %edx
-; SSSE3-NEXT:    subl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %esi
-; SSSE3-NEXT:    cmovol %ecx, %esi
-; SSSE3-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r14d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r14d, %edx
-; SSSE3-NEXT:    subl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %r14d
-; SSSE3-NEXT:    cmovol %ecx, %r14d
-; SSSE3-NEXT:    movd %xmm6, %eax
-; SSSE3-NEXT:    movd %xmm2, %r15d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r15d, %edx
-; SSSE3-NEXT:    subl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %r15d
-; SSSE3-NEXT:    cmovol %ecx, %r15d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %r12d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r12d, %edx
-; SSSE3-NEXT:    subl %eax, %edx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %r12d
-; SSSE3-NEXT:    cmovol %ecx, %r12d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r13d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r13d, %esi
-; SSSE3-NEXT:    subl %eax, %esi
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %r13d
-; SSSE3-NEXT:    cmovol %ecx, %r13d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r11d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r11d, %edi
-; SSSE3-NEXT:    subl %eax, %edi
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %r11d
-; SSSE3-NEXT:    cmovol %ecx, %r11d
-; SSSE3-NEXT:    movd %xmm7, %eax
-; SSSE3-NEXT:    movd %xmm3, %ebx
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %ebx, %edi
-; SSSE3-NEXT:    subl %eax, %edi
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %ebx
-; SSSE3-NEXT:    cmovol %ecx, %ebx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %ebp
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %ebp, %edi
-; SSSE3-NEXT:    subl %eax, %edi
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %ebp
-; SSSE3-NEXT:    cmovol %ecx, %ebp
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r10d
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movl %r10d, %edi
-; SSSE3-NEXT:    subl %eax, %edi
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %r10d
-; SSSE3-NEXT:    cmovol %ecx, %r10d
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %ecx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r9d
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %r9d, %eax
-; SSSE3-NEXT:    subl %ecx, %eax
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %ecx, %r9d
-; SSSE3-NEXT:    cmovol %edi, %r9d
-; SSSE3-NEXT:    movd %xmm4, %edx
-; SSSE3-NEXT:    movd %xmm0, %esi
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %esi, %ecx
-; SSSE3-NEXT:    subl %edx, %ecx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %edx, %esi
-; SSSE3-NEXT:    cmovol %edi, %esi
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %eax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movd %xmm1, %ecx
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %ecx, %edx
-; SSSE3-NEXT:    subl %eax, %edx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %eax, %ecx
-; SSSE3-NEXT:    cmovol %edi, %ecx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[3,1,2,3]
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSSE3-NEXT:    movd %xmm1, %r8d
-; SSSE3-NEXT:    movd %xmm0, %eax
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movl %eax, %edx
-; SSSE3-NEXT:    subl %r8d, %edx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSSE3-NEXT:    subl %r8d, %eax
-; SSSE3-NEXT:    cmovol %edi, %eax
-; SSSE3-NEXT:    movd %eax, %xmm8
-; SSSE3-NEXT:    movd %ecx, %xmm5
-; SSSE3-NEXT:    movd %esi, %xmm0
-; SSSE3-NEXT:    movd %r9d, %xmm9
-; SSSE3-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 4-byte Reload
-; SSSE3-NEXT:    # xmm10 = mem[0],zero,zero,zero
-; SSSE3-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 4-byte Reload
-; SSSE3-NEXT:    # xmm4 = mem[0],zero,zero,zero
-; SSSE3-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
-; SSSE3-NEXT:    # xmm1 = mem[0],zero,zero,zero
-; SSSE3-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 4-byte Reload
-; SSSE3-NEXT:    # xmm11 = mem[0],zero,zero,zero
-; SSSE3-NEXT:    movd %r13d, %xmm12
-; SSSE3-NEXT:    movd %r12d, %xmm6
-; SSSE3-NEXT:    movd %r15d, %xmm2
-; SSSE3-NEXT:    movd %r14d, %xmm13
-; SSSE3-NEXT:    movd %r10d, %xmm14
-; SSSE3-NEXT:    movd %ebp, %xmm7
-; SSSE3-NEXT:    movd %ebx, %xmm3
-; SSSE3-NEXT:    movd %r11d, %xmm15
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm5 = xmm5[0],xmm8[0],xmm5[1],xmm8[1]
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1]
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm5[0]
-; SSSE3-NEXT:    unpcklps {{.*#+}} xmm4 = xmm4[0],xmm10[0],xmm4[1],xmm10[1]
-; SSSE3-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm11[0],xmm1[1],xmm11[1]
-; SSSE3-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm4[0]
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm6 = xmm6[0],xmm12[0],xmm6[1],xmm12[1]
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm13[0],xmm2[1],xmm13[1]
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm6[0]
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm7 = xmm7[0],xmm14[0],xmm7[1],xmm14[1]
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm15[0],xmm3[1],xmm15[1]
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm7[0]
-; SSSE3-NEXT:    popq %rbx
-; SSSE3-NEXT:    popq %r12
-; SSSE3-NEXT:    popq %r13
-; SSSE3-NEXT:    popq %r14
-; SSSE3-NEXT:    popq %r15
-; SSSE3-NEXT:    popq %rbp
+; SSSE3-NEXT:    movdqa %xmm1, %xmm8
+; SSSE3-NEXT:    movdqa %xmm0, %xmm12
+; SSSE3-NEXT:    pxor %xmm9, %xmm9
+; SSSE3-NEXT:    pxor %xmm0, %xmm0
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm0
+; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm10
+; SSSE3-NEXT:    pxor %xmm10, %xmm0
+; SSSE3-NEXT:    pxor %xmm11, %xmm11
+; SSSE3-NEXT:    pcmpgtd %xmm12, %xmm11
+; SSSE3-NEXT:    pxor %xmm10, %xmm11
+; SSSE3-NEXT:    pcmpeqd %xmm11, %xmm0
+; SSSE3-NEXT:    psubd %xmm4, %xmm12
+; SSSE3-NEXT:    pxor %xmm1, %xmm1
+; SSSE3-NEXT:    pcmpgtd %xmm12, %xmm1
+; SSSE3-NEXT:    movdqa %xmm1, %xmm4
+; SSSE3-NEXT:    pxor %xmm10, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm11, %xmm4
+; SSSE3-NEXT:    pxor %xmm10, %xmm4
+; SSSE3-NEXT:    pandn %xmm4, %xmm0
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm11 = [2147483648,2147483648,2147483648,2147483648]
+; SSSE3-NEXT:    movdqa %xmm1, %xmm4
+; SSSE3-NEXT:    pandn %xmm11, %xmm4
+; SSSE3-NEXT:    psrld $1, %xmm1
+; SSSE3-NEXT:    por %xmm4, %xmm1
+; SSSE3-NEXT:    pand %xmm0, %xmm1
+; SSSE3-NEXT:    pandn %xmm12, %xmm0
+; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    pxor %xmm1, %xmm1
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm1
+; SSSE3-NEXT:    pxor %xmm10, %xmm1
+; SSSE3-NEXT:    pxor %xmm12, %xmm12
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm12
+; SSSE3-NEXT:    pxor %xmm10, %xmm12
+; SSSE3-NEXT:    pcmpeqd %xmm12, %xmm1
+; SSSE3-NEXT:    psubd %xmm5, %xmm8
+; SSSE3-NEXT:    pxor %xmm4, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm8, %xmm4
+; SSSE3-NEXT:    movdqa %xmm4, %xmm5
+; SSSE3-NEXT:    pxor %xmm10, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm12, %xmm5
+; SSSE3-NEXT:    pxor %xmm10, %xmm5
+; SSSE3-NEXT:    pandn %xmm5, %xmm1
+; SSSE3-NEXT:    movdqa %xmm4, %xmm5
+; SSSE3-NEXT:    pandn %xmm11, %xmm5
+; SSSE3-NEXT:    psrld $1, %xmm4
+; SSSE3-NEXT:    por %xmm5, %xmm4
+; SSSE3-NEXT:    pand %xmm1, %xmm4
+; SSSE3-NEXT:    pandn %xmm8, %xmm1
+; SSSE3-NEXT:    por %xmm4, %xmm1
+; SSSE3-NEXT:    pxor %xmm8, %xmm8
+; SSSE3-NEXT:    pcmpgtd %xmm6, %xmm8
+; SSSE3-NEXT:    pxor %xmm10, %xmm8
+; SSSE3-NEXT:    pxor %xmm5, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSSE3-NEXT:    pxor %xmm10, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm8
+; SSSE3-NEXT:    psubd %xmm6, %xmm2
+; SSSE3-NEXT:    pxor %xmm4, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm4
+; SSSE3-NEXT:    movdqa %xmm4, %xmm6
+; SSSE3-NEXT:    pxor %xmm10, %xmm6
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm6
+; SSSE3-NEXT:    pxor %xmm10, %xmm6
+; SSSE3-NEXT:    pandn %xmm6, %xmm8
+; SSSE3-NEXT:    movdqa %xmm4, %xmm5
+; SSSE3-NEXT:    pandn %xmm11, %xmm5
+; SSSE3-NEXT:    psrld $1, %xmm4
+; SSSE3-NEXT:    por %xmm5, %xmm4
+; SSSE3-NEXT:    pand %xmm8, %xmm4
+; SSSE3-NEXT:    pandn %xmm2, %xmm8
+; SSSE3-NEXT:    por %xmm4, %xmm8
+; SSSE3-NEXT:    pxor %xmm5, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm5
+; SSSE3-NEXT:    pxor %xmm10, %xmm5
+; SSSE3-NEXT:    pxor %xmm2, %xmm2
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm2
+; SSSE3-NEXT:    pxor %xmm10, %xmm2
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm5
+; SSSE3-NEXT:    psubd %xmm7, %xmm3
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm9
+; SSSE3-NEXT:    movdqa %xmm9, %xmm4
+; SSSE3-NEXT:    pxor %xmm10, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm4
+; SSSE3-NEXT:    pxor %xmm10, %xmm4
+; SSSE3-NEXT:    pandn %xmm4, %xmm5
+; SSSE3-NEXT:    movdqa %xmm9, %xmm2
+; SSSE3-NEXT:    pandn %xmm11, %xmm2
+; SSSE3-NEXT:    psrld $1, %xmm9
+; SSSE3-NEXT:    por %xmm2, %xmm9
+; SSSE3-NEXT:    pand %xmm5, %xmm9
+; SSSE3-NEXT:    pandn %xmm3, %xmm5
+; SSSE3-NEXT:    por %xmm9, %xmm5
+; SSSE3-NEXT:    movdqa %xmm8, %xmm2
+; SSSE3-NEXT:    movdqa %xmm5, %xmm3
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v16i32:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    pushq %rbp
-; SSE41-NEXT:    pushq %r15
-; SSE41-NEXT:    pushq %r14
-; SSE41-NEXT:    pushq %r13
-; SSE41-NEXT:    pushq %r12
-; SSE41-NEXT:    pushq %rbx
-; SSE41-NEXT:    pextrd $3, %xmm5, %eax
-; SSE41-NEXT:    pextrd $3, %xmm1, %esi
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %esi, %edx
-; SSE41-NEXT:    subl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %esi
-; SSE41-NEXT:    cmovol %ecx, %esi
-; SSE41-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE41-NEXT:    pextrd $2, %xmm5, %eax
-; SSE41-NEXT:    pextrd $2, %xmm1, %esi
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %esi, %edx
-; SSE41-NEXT:    subl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %esi
-; SSE41-NEXT:    cmovol %ecx, %esi
-; SSE41-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE41-NEXT:    movd %xmm5, %eax
-; SSE41-NEXT:    movd %xmm1, %esi
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %esi, %edx
-; SSE41-NEXT:    subl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %esi
-; SSE41-NEXT:    cmovol %ecx, %esi
-; SSE41-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE41-NEXT:    pextrd $1, %xmm5, %eax
-; SSE41-NEXT:    pextrd $1, %xmm1, %esi
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %esi, %edx
-; SSE41-NEXT:    subl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %esi
-; SSE41-NEXT:    cmovol %ecx, %esi
-; SSE41-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE41-NEXT:    pextrd $3, %xmm6, %eax
-; SSE41-NEXT:    pextrd $3, %xmm2, %r14d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r14d, %edx
-; SSE41-NEXT:    subl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %r14d
-; SSE41-NEXT:    cmovol %ecx, %r14d
-; SSE41-NEXT:    pextrd $2, %xmm6, %eax
-; SSE41-NEXT:    pextrd $2, %xmm2, %r15d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r15d, %edx
-; SSE41-NEXT:    subl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %r15d
-; SSE41-NEXT:    cmovol %ecx, %r15d
-; SSE41-NEXT:    movd %xmm6, %eax
-; SSE41-NEXT:    movd %xmm2, %r12d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r12d, %edx
-; SSE41-NEXT:    subl %eax, %edx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %r12d
-; SSE41-NEXT:    cmovol %ecx, %r12d
-; SSE41-NEXT:    pextrd $1, %xmm6, %eax
-; SSE41-NEXT:    pextrd $1, %xmm2, %r13d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r13d, %esi
-; SSE41-NEXT:    subl %eax, %esi
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %r13d
-; SSE41-NEXT:    cmovol %ecx, %r13d
-; SSE41-NEXT:    pextrd $3, %xmm7, %eax
-; SSE41-NEXT:    pextrd $3, %xmm3, %r11d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r11d, %edi
-; SSE41-NEXT:    subl %eax, %edi
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %r11d
-; SSE41-NEXT:    cmovol %ecx, %r11d
-; SSE41-NEXT:    pextrd $2, %xmm7, %eax
-; SSE41-NEXT:    pextrd $2, %xmm3, %ebx
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %ebx, %edi
-; SSE41-NEXT:    subl %eax, %edi
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %ebx
-; SSE41-NEXT:    cmovol %ecx, %ebx
-; SSE41-NEXT:    movd %xmm7, %eax
-; SSE41-NEXT:    movd %xmm3, %ebp
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %ebp, %edi
-; SSE41-NEXT:    subl %eax, %edi
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %ebp
-; SSE41-NEXT:    cmovol %ecx, %ebp
-; SSE41-NEXT:    pextrd $1, %xmm7, %eax
-; SSE41-NEXT:    pextrd $1, %xmm3, %r10d
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movl %r10d, %edi
-; SSE41-NEXT:    subl %eax, %edi
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %r10d
-; SSE41-NEXT:    cmovol %ecx, %r10d
-; SSE41-NEXT:    pextrd $3, %xmm4, %ecx
-; SSE41-NEXT:    pextrd $3, %xmm0, %r9d
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %r9d, %eax
-; SSE41-NEXT:    subl %ecx, %eax
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %ecx, %r9d
-; SSE41-NEXT:    cmovol %edi, %r9d
-; SSE41-NEXT:    pextrd $2, %xmm4, %edx
-; SSE41-NEXT:    pextrd $2, %xmm0, %esi
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %esi, %ecx
-; SSE41-NEXT:    subl %edx, %ecx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %edx, %esi
-; SSE41-NEXT:    cmovol %edi, %esi
-; SSE41-NEXT:    movd %xmm4, %eax
-; SSE41-NEXT:    movd %xmm0, %ecx
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %ecx, %edx
-; SSE41-NEXT:    subl %eax, %edx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %eax, %ecx
-; SSE41-NEXT:    cmovol %edi, %ecx
-; SSE41-NEXT:    pextrd $1, %xmm4, %r8d
-; SSE41-NEXT:    pextrd $1, %xmm0, %eax
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movl %eax, %edx
-; SSE41-NEXT:    subl %r8d, %edx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; SSE41-NEXT:    subl %r8d, %eax
-; SSE41-NEXT:    cmovol %edi, %eax
-; SSE41-NEXT:    movd %ecx, %xmm0
-; SSE41-NEXT:    pinsrd $1, %eax, %xmm0
-; SSE41-NEXT:    pinsrd $2, %esi, %xmm0
-; SSE41-NEXT:    pinsrd $3, %r9d, %xmm0
-; SSE41-NEXT:    movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Folded Reload
-; SSE41-NEXT:    # xmm1 = mem[0],zero,zero,zero
-; SSE41-NEXT:    pinsrd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Folded Reload
-; SSE41-NEXT:    pinsrd $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Folded Reload
-; SSE41-NEXT:    pinsrd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Folded Reload
-; SSE41-NEXT:    movd %r12d, %xmm2
-; SSE41-NEXT:    pinsrd $1, %r13d, %xmm2
-; SSE41-NEXT:    pinsrd $2, %r15d, %xmm2
-; SSE41-NEXT:    pinsrd $3, %r14d, %xmm2
-; SSE41-NEXT:    movd %ebp, %xmm3
-; SSE41-NEXT:    pinsrd $1, %r10d, %xmm3
-; SSE41-NEXT:    pinsrd $2, %ebx, %xmm3
-; SSE41-NEXT:    pinsrd $3, %r11d, %xmm3
-; SSE41-NEXT:    popq %rbx
-; SSE41-NEXT:    popq %r12
-; SSE41-NEXT:    popq %r13
-; SSE41-NEXT:    popq %r14
-; SSE41-NEXT:    popq %r15
-; SSE41-NEXT:    popq %rbp
+; SSE41-NEXT:    movdqa %xmm0, %xmm8
+; SSE41-NEXT:    pxor %xmm10, %xmm10
+; SSE41-NEXT:    pxor %xmm9, %xmm9
+; SSE41-NEXT:    pcmpgtd %xmm4, %xmm9
+; SSE41-NEXT:    pcmpeqd %xmm11, %xmm11
+; SSE41-NEXT:    pxor %xmm11, %xmm9
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm8, %xmm0
+; SSE41-NEXT:    pxor %xmm11, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm9
+; SSE41-NEXT:    psubd %xmm4, %xmm8
+; SSE41-NEXT:    pxor %xmm4, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm8, %xmm4
+; SSE41-NEXT:    pxor %xmm11, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pxor %xmm11, %xmm4
+; SSE41-NEXT:    pandn %xmm4, %xmm9
+; SSE41-NEXT:    movaps {{.*#+}} xmm13 = [2147483647,2147483647,2147483647,2147483647]
+; SSE41-NEXT:    movaps {{.*#+}} xmm12 = [2147483648,2147483648,2147483648,2147483648]
+; SSE41-NEXT:    movaps %xmm12, %xmm4
+; SSE41-NEXT:    movdqa %xmm8, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm13, %xmm4
+; SSE41-NEXT:    movdqa %xmm9, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm4, %xmm8
+; SSE41-NEXT:    xorps %xmm4, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm5, %xmm4
+; SSE41-NEXT:    pxor %xmm11, %xmm4
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
+; SSE41-NEXT:    pxor %xmm11, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    psubd %xmm5, %xmm1
+; SSE41-NEXT:    pxor %xmm5, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm5
+; SSE41-NEXT:    pxor %xmm11, %xmm5
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm5
+; SSE41-NEXT:    pxor %xmm11, %xmm5
+; SSE41-NEXT:    pandn %xmm5, %xmm4
+; SSE41-NEXT:    movaps %xmm12, %xmm5
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm13, %xmm5
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm5, %xmm1
+; SSE41-NEXT:    pxor %xmm4, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm6, %xmm4
+; SSE41-NEXT:    pxor %xmm11, %xmm4
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm11, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    psubd %xmm6, %xmm2
+; SSE41-NEXT:    xorps %xmm5, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSE41-NEXT:    pxor %xmm11, %xmm5
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm5
+; SSE41-NEXT:    pxor %xmm11, %xmm5
+; SSE41-NEXT:    pandn %xmm5, %xmm4
+; SSE41-NEXT:    movaps %xmm12, %xmm5
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm13, %xmm5
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm5, %xmm2
+; SSE41-NEXT:    pxor %xmm4, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm7, %xmm4
+; SSE41-NEXT:    pxor %xmm11, %xmm4
+; SSE41-NEXT:    pxor %xmm0, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
+; SSE41-NEXT:    pxor %xmm11, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    psubd %xmm7, %xmm3
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm10
+; SSE41-NEXT:    pxor %xmm11, %xmm10
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm10
+; SSE41-NEXT:    pxor %xmm11, %xmm10
+; SSE41-NEXT:    pandn %xmm10, %xmm4
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm13, %xmm12
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvps %xmm0, %xmm12, %xmm3
+; SSE41-NEXT:    movaps %xmm8, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: v16i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    pushq %rbp
-; AVX1-NEXT:    pushq %r15
-; AVX1-NEXT:    pushq %r14
-; AVX1-NEXT:    pushq %r13
-; AVX1-NEXT:    pushq %r12
-; AVX1-NEXT:    pushq %rbx
-; AVX1-NEXT:    vpextrd $3, %xmm3, %eax
-; AVX1-NEXT:    vpextrd $3, %xmm1, %esi
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %esi, %edx
-; AVX1-NEXT:    subl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %esi
-; AVX1-NEXT:    cmovol %ecx, %esi
-; AVX1-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX1-NEXT:    vpextrd $2, %xmm3, %eax
-; AVX1-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %esi, %edx
-; AVX1-NEXT:    subl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %esi
-; AVX1-NEXT:    cmovol %ecx, %esi
-; AVX1-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX1-NEXT:    vmovd %xmm3, %eax
-; AVX1-NEXT:    vmovd %xmm1, %esi
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %esi, %edx
-; AVX1-NEXT:    subl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %esi
-; AVX1-NEXT:    cmovol %ecx, %esi
-; AVX1-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX1-NEXT:    vpextrd $1, %xmm3, %eax
-; AVX1-NEXT:    vpextrd $1, %xmm1, %esi
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %esi, %edx
-; AVX1-NEXT:    subl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %esi
-; AVX1-NEXT:    cmovol %ecx, %esi
-; AVX1-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm3
-; AVX1-NEXT:    vpextrd $3, %xmm3, %eax
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vpextrd $3, %xmm1, %r14d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r14d, %edx
-; AVX1-NEXT:    subl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %r14d
-; AVX1-NEXT:    cmovol %ecx, %r14d
-; AVX1-NEXT:    vpextrd $2, %xmm3, %eax
-; AVX1-NEXT:    vpextrd $2, %xmm1, %r15d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r15d, %edx
-; AVX1-NEXT:    subl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %r15d
-; AVX1-NEXT:    cmovol %ecx, %r15d
-; AVX1-NEXT:    vmovd %xmm3, %eax
-; AVX1-NEXT:    vmovd %xmm1, %r12d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r12d, %edx
-; AVX1-NEXT:    subl %eax, %edx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %r12d
-; AVX1-NEXT:    cmovol %ecx, %r12d
-; AVX1-NEXT:    vpextrd $1, %xmm3, %eax
-; AVX1-NEXT:    vpextrd $1, %xmm1, %r13d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r13d, %esi
-; AVX1-NEXT:    subl %eax, %esi
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %r13d
-; AVX1-NEXT:    cmovol %ecx, %r13d
-; AVX1-NEXT:    vpextrd $3, %xmm2, %eax
-; AVX1-NEXT:    vpextrd $3, %xmm0, %r11d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r11d, %edi
-; AVX1-NEXT:    subl %eax, %edi
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %r11d
-; AVX1-NEXT:    cmovol %ecx, %r11d
-; AVX1-NEXT:    vpextrd $2, %xmm2, %eax
-; AVX1-NEXT:    vpextrd $2, %xmm0, %ebx
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %ebx, %edi
-; AVX1-NEXT:    subl %eax, %edi
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %ebx
-; AVX1-NEXT:    cmovol %ecx, %ebx
-; AVX1-NEXT:    vmovd %xmm2, %eax
-; AVX1-NEXT:    vmovd %xmm0, %ebp
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %ebp, %edi
-; AVX1-NEXT:    subl %eax, %edi
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %ebp
-; AVX1-NEXT:    cmovol %ecx, %ebp
-; AVX1-NEXT:    vpextrd $1, %xmm2, %eax
-; AVX1-NEXT:    vpextrd $1, %xmm0, %r10d
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movl %r10d, %edi
-; AVX1-NEXT:    subl %eax, %edi
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %r10d
-; AVX1-NEXT:    cmovol %ecx, %r10d
-; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm1
-; AVX1-NEXT:    vpextrd $3, %xmm1, %ecx
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpextrd $3, %xmm0, %r8d
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %r8d, %eax
-; AVX1-NEXT:    subl %ecx, %eax
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %ecx, %r8d
-; AVX1-NEXT:    cmovol %edi, %r8d
-; AVX1-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX1-NEXT:    vpextrd $2, %xmm0, %edx
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %edx, %ecx
-; AVX1-NEXT:    subl %esi, %ecx
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %esi, %edx
-; AVX1-NEXT:    cmovol %edi, %edx
-; AVX1-NEXT:    vmovd %xmm1, %eax
-; AVX1-NEXT:    vmovd %xmm0, %ecx
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %ecx, %esi
-; AVX1-NEXT:    subl %eax, %esi
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %eax, %ecx
-; AVX1-NEXT:    cmovol %edi, %ecx
-; AVX1-NEXT:    vpextrd $1, %xmm1, %r9d
-; AVX1-NEXT:    vpextrd $1, %xmm0, %eax
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movl %eax, %esi
-; AVX1-NEXT:    subl %r9d, %esi
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX1-NEXT:    subl %r9d, %eax
-; AVX1-NEXT:    cmovol %edi, %eax
-; AVX1-NEXT:    vmovd %ecx, %xmm0
-; AVX1-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
-; AVX1-NEXT:    vpinsrd $2, %edx, %xmm0, %xmm0
-; AVX1-NEXT:    vpinsrd $3, %r8d, %xmm0, %xmm0
-; AVX1-NEXT:    vmovd %ebp, %xmm1
-; AVX1-NEXT:    vpinsrd $1, %r10d, %xmm1, %xmm1
-; AVX1-NEXT:    vpinsrd $2, %ebx, %xmm1, %xmm1
-; AVX1-NEXT:    vpinsrd $3, %r11d, %xmm1, %xmm1
-; AVX1-NEXT:    vmovd %r12d, %xmm2
-; AVX1-NEXT:    vpinsrd $1, %r13d, %xmm2, %xmm2
-; AVX1-NEXT:    vpinsrd $2, %r15d, %xmm2, %xmm2
-; AVX1-NEXT:    vpinsrd $3, %r14d, %xmm2, %xmm2
-; AVX1-NEXT:    vmovd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 4-byte Folded Reload
-; AVX1-NEXT:    # xmm3 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vpinsrd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX1-NEXT:    vpinsrd $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX1-NEXT:    vpinsrd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm1
-; AVX1-NEXT:    popq %rbx
-; AVX1-NEXT:    popq %r12
-; AVX1-NEXT:    popq %r13
-; AVX1-NEXT:    popq %r14
-; AVX1-NEXT:    popq %r15
-; AVX1-NEXT:    popq %rbp
+; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm8
+; AVX1-NEXT:    vpxor %xmm12, %xmm12, %xmm12
+; AVX1-NEXT:    vpcmpgtd %xmm8, %xmm12, %xmm7
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm9
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm6
+; AVX1-NEXT:    vpcmpgtd %xmm6, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm10
+; AVX1-NEXT:    vpcmpeqd %xmm9, %xmm10, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm9
+; AVX1-NEXT:    vpcmpgtd %xmm2, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm11
+; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm7
+; AVX1-NEXT:    vpcmpeqd %xmm11, %xmm7, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vinsertf128 $1, %xmm9, %ymm5, %ymm9
+; AVX1-NEXT:    vpsubd %xmm8, %xmm6, %xmm8
+; AVX1-NEXT:    vpcmpgtd %xmm8, %xmm12, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm6
+; AVX1-NEXT:    vpcmpeqd %xmm6, %xmm10, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm6
+; AVX1-NEXT:    vpsubd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm12, %xmm2
+; AVX1-NEXT:    vpxor %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm7, %xmm2
+; AVX1-NEXT:    vpxor %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm2, %ymm2
+; AVX1-NEXT:    vandps %ymm2, %ymm9, %ymm6
+; AVX1-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm7
+; AVX1-NEXT:    vmovaps {{.*#+}} ymm9 = [2147483647,2147483647,2147483647,2147483647,2147483647,2147483647,2147483647,2147483647]
+; AVX1-NEXT:    vmovaps {{.*#+}} ymm10 = [2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648]
+; AVX1-NEXT:    vblendvps %ymm7, %ymm9, %ymm10, %ymm7
+; AVX1-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm0
+; AVX1-NEXT:    vblendvps %ymm6, %ymm7, %ymm0, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm13
+; AVX1-NEXT:    vpcmpgtd %xmm13, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm7
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpcmpgtd %xmm2, %xmm12, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vpcmpeqd %xmm7, %xmm5, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm8
+; AVX1-NEXT:    vpcmpgtd %xmm3, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm11
+; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm7
+; AVX1-NEXT:    vpcmpeqd %xmm11, %xmm7, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm6
+; AVX1-NEXT:    vinsertf128 $1, %xmm8, %ymm6, %ymm8
+; AVX1-NEXT:    vpsubd %xmm13, %xmm2, %xmm11
+; AVX1-NEXT:    vpcmpgtd %xmm11, %xmm12, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm5, %xmm2
+; AVX1-NEXT:    vpxor %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vpsubd %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm12, %xmm3
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm7, %xmm3
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
+; AVX1-NEXT:    vandps %ymm2, %ymm8, %ymm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm1, %ymm3
+; AVX1-NEXT:    vblendvps %ymm3, %ymm9, %ymm10, %ymm3
+; AVX1-NEXT:    vinsertf128 $1, %xmm11, %ymm1, %ymm1
+; AVX1-NEXT:    vblendvps %ymm2, %ymm3, %ymm1, %ymm1
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: v16i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    pushq %rbp
-; AVX2-NEXT:    pushq %r15
-; AVX2-NEXT:    pushq %r14
-; AVX2-NEXT:    pushq %r13
-; AVX2-NEXT:    pushq %r12
-; AVX2-NEXT:    pushq %rbx
-; AVX2-NEXT:    vpextrd $3, %xmm3, %eax
-; AVX2-NEXT:    vpextrd $3, %xmm1, %esi
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %esi, %edx
-; AVX2-NEXT:    subl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %esi
-; AVX2-NEXT:    cmovol %ecx, %esi
-; AVX2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX2-NEXT:    vpextrd $2, %xmm3, %eax
-; AVX2-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %esi, %edx
-; AVX2-NEXT:    subl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %esi
-; AVX2-NEXT:    cmovol %ecx, %esi
-; AVX2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX2-NEXT:    vmovd %xmm3, %eax
-; AVX2-NEXT:    vmovd %xmm1, %esi
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %esi, %edx
-; AVX2-NEXT:    subl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %esi
-; AVX2-NEXT:    cmovol %ecx, %esi
-; AVX2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX2-NEXT:    vpextrd $1, %xmm3, %eax
-; AVX2-NEXT:    vpextrd $1, %xmm1, %esi
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %esi, %edx
-; AVX2-NEXT:    subl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %esi
-; AVX2-NEXT:    cmovol %ecx, %esi
-; AVX2-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX2-NEXT:    vextracti128 $1, %ymm3, %xmm3
-; AVX2-NEXT:    vpextrd $3, %xmm3, %eax
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; AVX2-NEXT:    vpextrd $3, %xmm1, %r14d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r14d, %edx
-; AVX2-NEXT:    subl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %r14d
-; AVX2-NEXT:    cmovol %ecx, %r14d
-; AVX2-NEXT:    vpextrd $2, %xmm3, %eax
-; AVX2-NEXT:    vpextrd $2, %xmm1, %r15d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r15d, %edx
-; AVX2-NEXT:    subl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %r15d
-; AVX2-NEXT:    cmovol %ecx, %r15d
-; AVX2-NEXT:    vmovd %xmm3, %eax
-; AVX2-NEXT:    vmovd %xmm1, %r12d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r12d, %edx
-; AVX2-NEXT:    subl %eax, %edx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %r12d
-; AVX2-NEXT:    cmovol %ecx, %r12d
-; AVX2-NEXT:    vpextrd $1, %xmm3, %eax
-; AVX2-NEXT:    vpextrd $1, %xmm1, %r13d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r13d, %esi
-; AVX2-NEXT:    subl %eax, %esi
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %r13d
-; AVX2-NEXT:    cmovol %ecx, %r13d
-; AVX2-NEXT:    vpextrd $3, %xmm2, %eax
-; AVX2-NEXT:    vpextrd $3, %xmm0, %r11d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r11d, %edi
-; AVX2-NEXT:    subl %eax, %edi
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %r11d
-; AVX2-NEXT:    cmovol %ecx, %r11d
-; AVX2-NEXT:    vpextrd $2, %xmm2, %eax
-; AVX2-NEXT:    vpextrd $2, %xmm0, %ebx
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %ebx, %edi
-; AVX2-NEXT:    subl %eax, %edi
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %ebx
-; AVX2-NEXT:    cmovol %ecx, %ebx
-; AVX2-NEXT:    vmovd %xmm2, %eax
-; AVX2-NEXT:    vmovd %xmm0, %ebp
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %ebp, %edi
-; AVX2-NEXT:    subl %eax, %edi
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %ebp
-; AVX2-NEXT:    cmovol %ecx, %ebp
-; AVX2-NEXT:    vpextrd $1, %xmm2, %eax
-; AVX2-NEXT:    vpextrd $1, %xmm0, %r10d
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movl %r10d, %edi
-; AVX2-NEXT:    subl %eax, %edi
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %r10d
-; AVX2-NEXT:    cmovol %ecx, %r10d
-; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm1
-; AVX2-NEXT:    vpextrd $3, %xmm1, %ecx
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; AVX2-NEXT:    vpextrd $3, %xmm0, %r8d
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %r8d, %eax
-; AVX2-NEXT:    subl %ecx, %eax
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %ecx, %r8d
-; AVX2-NEXT:    cmovol %edi, %r8d
-; AVX2-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX2-NEXT:    vpextrd $2, %xmm0, %edx
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %edx, %ecx
-; AVX2-NEXT:    subl %esi, %ecx
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %esi, %edx
-; AVX2-NEXT:    cmovol %edi, %edx
-; AVX2-NEXT:    vmovd %xmm1, %eax
-; AVX2-NEXT:    vmovd %xmm0, %ecx
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %ecx, %esi
-; AVX2-NEXT:    subl %eax, %esi
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %eax, %ecx
-; AVX2-NEXT:    cmovol %edi, %ecx
-; AVX2-NEXT:    vpextrd $1, %xmm1, %r9d
-; AVX2-NEXT:    vpextrd $1, %xmm0, %eax
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movl %eax, %esi
-; AVX2-NEXT:    subl %r9d, %esi
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX2-NEXT:    subl %r9d, %eax
-; AVX2-NEXT:    cmovol %edi, %eax
-; AVX2-NEXT:    vmovd %ecx, %xmm0
-; AVX2-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
-; AVX2-NEXT:    vpinsrd $2, %edx, %xmm0, %xmm0
-; AVX2-NEXT:    vpinsrd $3, %r8d, %xmm0, %xmm0
-; AVX2-NEXT:    vmovd %ebp, %xmm1
-; AVX2-NEXT:    vpinsrd $1, %r10d, %xmm1, %xmm1
-; AVX2-NEXT:    vpinsrd $2, %ebx, %xmm1, %xmm1
-; AVX2-NEXT:    vpinsrd $3, %r11d, %xmm1, %xmm1
-; AVX2-NEXT:    vmovd %r12d, %xmm2
-; AVX2-NEXT:    vpinsrd $1, %r13d, %xmm2, %xmm2
-; AVX2-NEXT:    vpinsrd $2, %r15d, %xmm2, %xmm2
-; AVX2-NEXT:    vpinsrd $3, %r14d, %xmm2, %xmm2
-; AVX2-NEXT:    vmovd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 4-byte Folded Reload
-; AVX2-NEXT:    # xmm3 = mem[0],zero,zero,zero
-; AVX2-NEXT:    vpinsrd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX2-NEXT:    vpinsrd $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX2-NEXT:    vpinsrd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm1
-; AVX2-NEXT:    popq %rbx
-; AVX2-NEXT:    popq %r12
-; AVX2-NEXT:    popq %r13
-; AVX2-NEXT:    popq %r14
-; AVX2-NEXT:    popq %r15
-; AVX2-NEXT:    popq %rbp
+; AVX2-NEXT:    vpxor %xmm4, %xmm4, %xmm4
+; AVX2-NEXT:    vpcmpgtd %ymm2, %ymm4, %ymm5
+; AVX2-NEXT:    vpcmpeqd %ymm6, %ymm6, %ymm6
+; AVX2-NEXT:    vpxor %ymm6, %ymm5, %ymm5
+; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm4, %ymm7
+; AVX2-NEXT:    vpxor %ymm6, %ymm7, %ymm7
+; AVX2-NEXT:    vpcmpeqd %ymm5, %ymm7, %ymm5
+; AVX2-NEXT:    vpsubd %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm4, %ymm2
+; AVX2-NEXT:    vpxor %ymm6, %ymm2, %ymm2
+; AVX2-NEXT:    vpcmpeqd %ymm2, %ymm7, %ymm2
+; AVX2-NEXT:    vpxor %ymm6, %ymm2, %ymm2
+; AVX2-NEXT:    vpandn %ymm2, %ymm5, %ymm2
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm5 = [2147483647,2147483647,2147483647,2147483647,2147483647,2147483647,2147483647,2147483647]
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm7 = [2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648]
+; AVX2-NEXT:    vblendvps %ymm0, %ymm5, %ymm7, %ymm8
+; AVX2-NEXT:    vblendvps %ymm2, %ymm8, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpgtd %ymm3, %ymm4, %ymm2
+; AVX2-NEXT:    vpxor %ymm6, %ymm2, %ymm2
+; AVX2-NEXT:    vpcmpgtd %ymm1, %ymm4, %ymm8
+; AVX2-NEXT:    vpxor %ymm6, %ymm8, %ymm8
+; AVX2-NEXT:    vpcmpeqd %ymm2, %ymm8, %ymm2
+; AVX2-NEXT:    vpsubd %ymm3, %ymm1, %ymm1
+; AVX2-NEXT:    vpcmpgtd %ymm1, %ymm4, %ymm3
+; AVX2-NEXT:    vpxor %ymm6, %ymm3, %ymm3
+; AVX2-NEXT:    vpcmpeqd %ymm3, %ymm8, %ymm3
+; AVX2-NEXT:    vpxor %ymm6, %ymm3, %ymm3
+; AVX2-NEXT:    vpandn %ymm3, %ymm2, %ymm2
+; AVX2-NEXT:    vblendvps %ymm1, %ymm5, %ymm7, %ymm3
+; AVX2-NEXT:    vblendvps %ymm2, %ymm3, %ymm1, %ymm1
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: v16i32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    pushq %rbp
-; AVX512-NEXT:    pushq %r15
-; AVX512-NEXT:    pushq %r14
-; AVX512-NEXT:    pushq %r13
-; AVX512-NEXT:    pushq %r12
-; AVX512-NEXT:    pushq %rbx
-; AVX512-NEXT:    vpextrd $3, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $3, %xmm0, %esi
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %esi, %edx
-; AVX512-NEXT:    subl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %esi
-; AVX512-NEXT:    cmovol %ecx, %esi
-; AVX512-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX512-NEXT:    vpextrd $2, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $2, %xmm0, %esi
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %esi, %edx
-; AVX512-NEXT:    subl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %esi
-; AVX512-NEXT:    cmovol %ecx, %esi
-; AVX512-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX512-NEXT:    vmovd %xmm1, %eax
-; AVX512-NEXT:    vmovd %xmm0, %esi
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %esi, %edx
-; AVX512-NEXT:    subl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %esi
-; AVX512-NEXT:    cmovol %ecx, %esi
-; AVX512-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX512-NEXT:    vpextrd $1, %xmm1, %eax
-; AVX512-NEXT:    vpextrd $1, %xmm0, %esi
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %esi, %edx
-; AVX512-NEXT:    subl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %esi
-; AVX512-NEXT:    cmovol %ecx, %esi
-; AVX512-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; AVX512-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX512-NEXT:    vpextrd $3, %xmm2, %eax
-; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm3
-; AVX512-NEXT:    vpextrd $3, %xmm3, %r14d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r14d, %edx
-; AVX512-NEXT:    subl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %r14d
-; AVX512-NEXT:    cmovol %ecx, %r14d
-; AVX512-NEXT:    vpextrd $2, %xmm2, %eax
-; AVX512-NEXT:    vpextrd $2, %xmm3, %r15d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r15d, %edx
-; AVX512-NEXT:    subl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %r15d
-; AVX512-NEXT:    cmovol %ecx, %r15d
-; AVX512-NEXT:    vmovd %xmm2, %eax
-; AVX512-NEXT:    vmovd %xmm3, %r12d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r12d, %edx
-; AVX512-NEXT:    subl %eax, %edx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %r12d
-; AVX512-NEXT:    cmovol %ecx, %r12d
-; AVX512-NEXT:    vpextrd $1, %xmm2, %eax
-; AVX512-NEXT:    vpextrd $1, %xmm3, %r13d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r13d, %esi
-; AVX512-NEXT:    subl %eax, %esi
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %r13d
-; AVX512-NEXT:    cmovol %ecx, %r13d
-; AVX512-NEXT:    vextracti32x4 $2, %zmm1, %xmm2
-; AVX512-NEXT:    vpextrd $3, %xmm2, %eax
-; AVX512-NEXT:    vextracti32x4 $2, %zmm0, %xmm3
-; AVX512-NEXT:    vpextrd $3, %xmm3, %r11d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r11d, %edi
-; AVX512-NEXT:    subl %eax, %edi
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %r11d
-; AVX512-NEXT:    cmovol %ecx, %r11d
-; AVX512-NEXT:    vpextrd $2, %xmm2, %eax
-; AVX512-NEXT:    vpextrd $2, %xmm3, %ebx
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %ebx, %edi
-; AVX512-NEXT:    subl %eax, %edi
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %ebx
-; AVX512-NEXT:    cmovol %ecx, %ebx
-; AVX512-NEXT:    vmovd %xmm2, %eax
-; AVX512-NEXT:    vmovd %xmm3, %ebp
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %ebp, %edi
-; AVX512-NEXT:    subl %eax, %edi
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %ebp
-; AVX512-NEXT:    cmovol %ecx, %ebp
-; AVX512-NEXT:    vpextrd $1, %xmm2, %eax
-; AVX512-NEXT:    vpextrd $1, %xmm3, %r10d
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movl %r10d, %edi
-; AVX512-NEXT:    subl %eax, %edi
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    addl $2147483647, %ecx # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %r10d
-; AVX512-NEXT:    cmovol %ecx, %r10d
-; AVX512-NEXT:    vextracti32x4 $3, %zmm1, %xmm1
-; AVX512-NEXT:    vpextrd $3, %xmm1, %ecx
-; AVX512-NEXT:    vextracti32x4 $3, %zmm0, %xmm0
-; AVX512-NEXT:    vpextrd $3, %xmm0, %r8d
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %r8d, %eax
-; AVX512-NEXT:    subl %ecx, %eax
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %ecx, %r8d
-; AVX512-NEXT:    cmovol %edi, %r8d
-; AVX512-NEXT:    vpextrd $2, %xmm1, %esi
-; AVX512-NEXT:    vpextrd $2, %xmm0, %edx
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %edx, %ecx
-; AVX512-NEXT:    subl %esi, %ecx
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %esi, %edx
-; AVX512-NEXT:    cmovol %edi, %edx
-; AVX512-NEXT:    vmovd %xmm1, %eax
-; AVX512-NEXT:    vmovd %xmm0, %ecx
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %ecx, %esi
-; AVX512-NEXT:    subl %eax, %esi
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %eax, %ecx
-; AVX512-NEXT:    cmovol %edi, %ecx
-; AVX512-NEXT:    vpextrd $1, %xmm1, %r9d
-; AVX512-NEXT:    vpextrd $1, %xmm0, %eax
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movl %eax, %esi
-; AVX512-NEXT:    subl %r9d, %esi
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addl $2147483647, %edi # imm = 0x7FFFFFFF
-; AVX512-NEXT:    subl %r9d, %eax
-; AVX512-NEXT:    cmovol %edi, %eax
-; AVX512-NEXT:    vmovd %ecx, %xmm0
-; AVX512-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
-; AVX512-NEXT:    vpinsrd $2, %edx, %xmm0, %xmm0
-; AVX512-NEXT:    vpinsrd $3, %r8d, %xmm0, %xmm0
-; AVX512-NEXT:    vmovd %ebp, %xmm1
-; AVX512-NEXT:    vpinsrd $1, %r10d, %xmm1, %xmm1
-; AVX512-NEXT:    vpinsrd $2, %ebx, %xmm1, %xmm1
-; AVX512-NEXT:    vpinsrd $3, %r11d, %xmm1, %xmm1
-; AVX512-NEXT:    vmovd %r12d, %xmm2
-; AVX512-NEXT:    vpinsrd $1, %r13d, %xmm2, %xmm2
-; AVX512-NEXT:    vpinsrd $2, %r15d, %xmm2, %xmm2
-; AVX512-NEXT:    vpinsrd $3, %r14d, %xmm2, %xmm2
-; AVX512-NEXT:    vmovd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 4-byte Folded Reload
-; AVX512-NEXT:    # xmm3 = mem[0],zero,zero,zero
-; AVX512-NEXT:    vpinsrd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX512-NEXT:    vpinsrd $2, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX512-NEXT:    vpinsrd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 4-byte Folded Reload
-; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm1
-; AVX512-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512-NEXT:    popq %rbx
-; AVX512-NEXT:    popq %r12
-; AVX512-NEXT:    popq %r13
-; AVX512-NEXT:    popq %r14
-; AVX512-NEXT:    popq %r15
-; AVX512-NEXT:    popq %rbp
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltd %zmm2, %zmm1, %k0
+; AVX512-NEXT:    vpcmpnltd %zmm2, %zmm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpsubd %zmm1, %zmm0, %zmm0
+; AVX512-NEXT:    vpcmpnltd %zmm2, %zmm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtd %zmm0, %zmm2, %k2
+; AVX512-NEXT:    vpbroadcastd {{.*#+}} zmm1 = [2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648,2147483648]
+; AVX512-NEXT:    vpbroadcastd {{.*}}(%rip), %zmm1 {%k2}
+; AVX512-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k1}
 ; AVX512-NEXT:    retq
   %z = call <16 x i32> @llvm.ssub.sat.v16i32(<16 x i32> %x, <16 x i32> %y)
   ret <16 x i32> %z
@@ -2818,111 +1712,221 @@
 define <2 x i64> @v2i64(<2 x i64> %x, <2 x i64> %y) nounwind {
 ; SSE2-LABEL: v2i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movq %xmm2, %rax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movq %xmm2, %rcx
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movq %rcx, %rsi
-; SSE2-NEXT:    subq %rax, %rsi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE2-NEXT:    addq %r8, %rdx
-; SSE2-NEXT:    subq %rax, %rcx
-; SSE2-NEXT:    cmovoq %rdx, %rcx
-; SSE2-NEXT:    movq %xmm1, %rax
-; SSE2-NEXT:    movq %xmm0, %rsi
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %rsi, %rdx
-; SSE2-NEXT:    subq %rax, %rdx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    subq %rax, %rsi
-; SSE2-NEXT:    cmovoq %rdi, %rsi
-; SSE2-NEXT:    movq %rsi, %xmm0
-; SSE2-NEXT:    movq %rcx, %xmm1
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
+; SSE2-NEXT:    movdqa %xmm0, %xmm3
+; SSE2-NEXT:    psubq %xmm1, %xmm0
+; SSE2-NEXT:    pxor %xmm2, %xmm1
+; SSE2-NEXT:    movdqa %xmm2, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm1, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm5
+; SSE2-NEXT:    pxor %xmm5, %xmm4
+; SSE2-NEXT:    pxor %xmm2, %xmm3
+; SSE2-NEXT:    movdqa %xmm2, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2-NEXT:    pand %xmm6, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[1,1,3,3]
+; SSE2-NEXT:    por %xmm3, %xmm6
+; SSE2-NEXT:    pxor %xmm5, %xmm6
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm1
+; SSE2-NEXT:    movdqa %xmm0, %xmm3
+; SSE2-NEXT:    pxor %xmm2, %xmm3
+; SSE2-NEXT:    movdqa %xmm2, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
+; SSE2-NEXT:    pand %xmm7, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm2, %xmm3
+; SSE2-NEXT:    movdqa %xmm3, %xmm2
+; SSE2-NEXT:    pxor %xmm5, %xmm2
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,0,3,2]
+; SSE2-NEXT:    pand %xmm2, %xmm4
+; SSE2-NEXT:    pxor %xmm5, %xmm4
+; SSE2-NEXT:    pandn %xmm4, %xmm1
+; SSE2-NEXT:    movdqa %xmm3, %xmm2
+; SSE2-NEXT:    pandn {{.*}}(%rip), %xmm2
+; SSE2-NEXT:    pand {{.*}}(%rip), %xmm3
+; SSE2-NEXT:    por %xmm2, %xmm3
+; SSE2-NEXT:    pand %xmm1, %xmm3
+; SSE2-NEXT:    pandn %xmm0, %xmm1
+; SSE2-NEXT:    por %xmm3, %xmm1
+; SSE2-NEXT:    movdqa %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v2i64:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm2, %rax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm2, %rcx
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movq %rcx, %rsi
-; SSSE3-NEXT:    subq %rax, %rsi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSSE3-NEXT:    addq %r8, %rdx
-; SSSE3-NEXT:    subq %rax, %rcx
-; SSSE3-NEXT:    cmovoq %rdx, %rcx
-; SSSE3-NEXT:    movq %xmm1, %rax
-; SSSE3-NEXT:    movq %xmm0, %rsi
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %rsi, %rdx
-; SSSE3-NEXT:    subq %rax, %rdx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    subq %rax, %rsi
-; SSSE3-NEXT:    cmovoq %rdi, %rsi
-; SSSE3-NEXT:    movq %rsi, %xmm0
-; SSSE3-NEXT:    movq %rcx, %xmm1
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
+; SSSE3-NEXT:    movdqa %xmm0, %xmm3
+; SSSE3-NEXT:    psubq %xmm1, %xmm0
+; SSSE3-NEXT:    pxor %xmm2, %xmm1
+; SSSE3-NEXT:    movdqa %xmm2, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm5, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm1, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm5
+; SSSE3-NEXT:    pxor %xmm5, %xmm4
+; SSSE3-NEXT:    pxor %xmm2, %xmm3
+; SSSE3-NEXT:    movdqa %xmm2, %xmm1
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm6, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    por %xmm3, %xmm6
+; SSSE3-NEXT:    pxor %xmm5, %xmm6
+; SSSE3-NEXT:    pcmpeqd %xmm6, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm1
+; SSSE3-NEXT:    movdqa %xmm0, %xmm3
+; SSSE3-NEXT:    pxor %xmm2, %xmm3
+; SSSE3-NEXT:    movdqa %xmm2, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm7, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm2, %xmm3
+; SSSE3-NEXT:    movdqa %xmm3, %xmm2
+; SSSE3-NEXT:    pxor %xmm5, %xmm2
+; SSSE3-NEXT:    pcmpeqd %xmm6, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm2, %xmm4
+; SSSE3-NEXT:    pxor %xmm5, %xmm4
+; SSSE3-NEXT:    pandn %xmm4, %xmm1
+; SSSE3-NEXT:    movdqa %xmm3, %xmm2
+; SSSE3-NEXT:    pandn {{.*}}(%rip), %xmm2
+; SSSE3-NEXT:    pand {{.*}}(%rip), %xmm3
+; SSSE3-NEXT:    por %xmm2, %xmm3
+; SSSE3-NEXT:    pand %xmm1, %xmm3
+; SSSE3-NEXT:    pandn %xmm0, %xmm1
+; SSSE3-NEXT:    por %xmm3, %xmm1
+; SSSE3-NEXT:    movdqa %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v2i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movq %xmm1, %rax
-; SSE41-NEXT:    movq %xmm0, %rcx
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movq %rcx, %rsi
-; SSE41-NEXT:    subq %rax, %rsi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE41-NEXT:    addq %r8, %rdx
-; SSE41-NEXT:    subq %rax, %rcx
-; SSE41-NEXT:    cmovoq %rdx, %rcx
-; SSE41-NEXT:    pextrq $1, %xmm1, %rax
-; SSE41-NEXT:    pextrq $1, %xmm0, %rsi
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %rsi, %rdx
-; SSE41-NEXT:    subq %rax, %rdx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    subq %rax, %rsi
-; SSE41-NEXT:    cmovoq %rdi, %rsi
-; SSE41-NEXT:    movq %rsi, %xmm1
-; SSE41-NEXT:    movq %rcx, %xmm0
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; SSE41-NEXT:    movdqa %xmm0, %xmm2
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm2, %xmm3
+; SSE41-NEXT:    psubq %xmm1, %xmm2
+; SSE41-NEXT:    pxor %xmm0, %xmm1
+; SSE41-NEXT:    movdqa %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm1, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm6, %xmm1
+; SSE41-NEXT:    pcmpeqd %xmm4, %xmm4
+; SSE41-NEXT:    pxor %xmm4, %xmm1
+; SSE41-NEXT:    pxor %xmm0, %xmm3
+; SSE41-NEXT:    movdqa %xmm0, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE41-NEXT:    pand %xmm6, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE41-NEXT:    por %xmm3, %xmm5
+; SSE41-NEXT:    pxor %xmm4, %xmm5
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm1
+; SSE41-NEXT:    movdqa %xmm2, %xmm3
+; SSE41-NEXT:    pxor %xmm0, %xmm3
+; SSE41-NEXT:    movdqa %xmm0, %xmm6
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE41-NEXT:    pand %xmm7, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
+; SSE41-NEXT:    por %xmm3, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm3
+; SSE41-NEXT:    pxor %xmm4, %xmm3
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm3
+; SSE41-NEXT:    pxor %xmm4, %xmm3
+; SSE41-NEXT:    pandn %xmm3, %xmm1
+; SSE41-NEXT:    movapd {{.*#+}} xmm3 = [9223372036854775808,9223372036854775808]
+; SSE41-NEXT:    blendvpd %xmm0, {{.*}}(%rip), %xmm3
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm2
+; SSE41-NEXT:    movapd %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX-LABEL: v2i64:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovq %xmm1, %rax
-; AVX-NEXT:    vmovq %xmm0, %rcx
-; AVX-NEXT:    xorl %edx, %edx
-; AVX-NEXT:    movq %rcx, %rsi
-; AVX-NEXT:    subq %rax, %rsi
-; AVX-NEXT:    setns %dl
-; AVX-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX-NEXT:    addq %r8, %rdx
-; AVX-NEXT:    subq %rax, %rcx
-; AVX-NEXT:    cmovoq %rdx, %rcx
-; AVX-NEXT:    vpextrq $1, %xmm1, %rax
-; AVX-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX-NEXT:    xorl %edi, %edi
-; AVX-NEXT:    movq %rsi, %rdx
-; AVX-NEXT:    subq %rax, %rdx
-; AVX-NEXT:    setns %dil
-; AVX-NEXT:    addq %r8, %rdi
-; AVX-NEXT:    subq %rax, %rsi
-; AVX-NEXT:    cmovoq %rdi, %rsi
-; AVX-NEXT:    vmovq %rsi, %xmm0
-; AVX-NEXT:    vmovq %rcx, %xmm1
-; AVX-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX-NEXT:    retq
+; AVX1-LABEL: v2i64:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm5, %xmm3
+; AVX1-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm1
+; AVX1-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqq %xmm1, %xmm5, %xmm1
+; AVX1-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX1-NEXT:    vpandn %xmm1, %xmm3, %xmm1
+; AVX1-NEXT:    vmovapd {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
+; AVX1-NEXT:    vblendvpd %xmm0, {{.*}}(%rip), %xmm2, %xmm2
+; AVX1-NEXT:    vblendvpd %xmm1, %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: v2i64:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm3
+; AVX2-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX2-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX2-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm5
+; AVX2-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX2-NEXT:    vpcmpeqq %xmm3, %xmm5, %xmm3
+; AVX2-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm1
+; AVX2-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqq %xmm1, %xmm5, %xmm1
+; AVX2-NEXT:    vpxor %xmm4, %xmm1, %xmm1
+; AVX2-NEXT:    vpandn %xmm1, %xmm3, %xmm1
+; AVX2-NEXT:    vmovapd {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
+; AVX2-NEXT:    vblendvpd %xmm0, {{.*}}(%rip), %xmm2, %xmm2
+; AVX2-NEXT:    vblendvpd %xmm1, %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    retq
+;
+; AVX512-LABEL: v2i64:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltq %xmm2, %xmm1, %k0
+; AVX512-NEXT:    vpcmpnltq %xmm2, %xmm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vpcmpnltq %xmm2, %xmm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtq %xmm0, %xmm2, %k2
+; AVX512-NEXT:    vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
+; AVX512-NEXT:    vmovdqa64 {{.*}}(%rip), %xmm1 {%k2}
+; AVX512-NEXT:    vmovdqa64 %xmm1, %xmm0 {%k1}
+; AVX512-NEXT:    retq
   %z = call <2 x i64> @llvm.ssub.sat.v2i64(<2 x i64> %x, <2 x i64> %y)
   ret <2 x i64> %z
 }
@@ -2930,301 +1934,383 @@
 define <4 x i64> @v4i64(<4 x i64> %x, <4 x i64> %y) nounwind {
 ; SSE2-LABEL: v4i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[2,3,0,1]
-; SSE2-NEXT:    movq %xmm4, %rcx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movq %xmm4, %r9
-; SSE2-NEXT:    xorl %esi, %esi
-; SSE2-NEXT:    movq %r9, %rdx
-; SSE2-NEXT:    subq %rcx, %rdx
-; SSE2-NEXT:    setns %sil
-; SSE2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE2-NEXT:    addq %r8, %rsi
-; SSE2-NEXT:    subq %rcx, %r9
-; SSE2-NEXT:    cmovoq %rsi, %r9
-; SSE2-NEXT:    movq %xmm3, %rsi
-; SSE2-NEXT:    movq %xmm1, %rdx
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %rdx, %rcx
-; SSE2-NEXT:    subq %rsi, %rcx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    subq %rsi, %rdx
-; SSE2-NEXT:    cmovoq %rdi, %rdx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %rcx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %rax
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %rax, %rsi
-; SSE2-NEXT:    subq %rcx, %rsi
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    subq %rcx, %rax
-; SSE2-NEXT:    cmovoq %rdi, %rax
-; SSE2-NEXT:    movq %xmm2, %r10
-; SSE2-NEXT:    movq %xmm0, %rcx
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %rcx, %rsi
-; SSE2-NEXT:    subq %r10, %rsi
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    subq %r10, %rcx
-; SSE2-NEXT:    cmovoq %rdi, %rcx
-; SSE2-NEXT:    movq %rcx, %xmm0
-; SSE2-NEXT:    movq %rax, %xmm1
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE2-NEXT:    movq %rdx, %xmm1
-; SSE2-NEXT:    movq %r9, %xmm2
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; SSE2-NEXT:    movdqa %xmm0, %xmm10
+; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [2147483648,2147483648]
+; SSE2-NEXT:    psubq %xmm2, %xmm10
+; SSE2-NEXT:    pxor %xmm5, %xmm2
+; SSE2-NEXT:    movdqa %xmm5, %xmm6
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2-NEXT:    pand %xmm7, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,1,3,3]
+; SSE2-NEXT:    por %xmm2, %xmm7
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm9
+; SSE2-NEXT:    pxor %xmm9, %xmm7
+; SSE2-NEXT:    pxor %xmm5, %xmm0
+; SSE2-NEXT:    movdqa %xmm5, %xmm2
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pand %xmm6, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2-NEXT:    por %xmm0, %xmm2
+; SSE2-NEXT:    pxor %xmm9, %xmm2
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[1,0,3,2]
+; SSE2-NEXT:    pand %xmm7, %xmm0
+; SSE2-NEXT:    movdqa %xmm10, %xmm6
+; SSE2-NEXT:    pxor %xmm5, %xmm6
+; SSE2-NEXT:    movdqa %xmm5, %xmm7
+; SSE2-NEXT:    pcmpgtd %xmm6, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE2-NEXT:    pand %xmm4, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[1,1,3,3]
+; SSE2-NEXT:    por %xmm6, %xmm4
+; SSE2-NEXT:    movdqa %xmm4, %xmm6
+; SSE2-NEXT:    pxor %xmm9, %xmm6
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,0,3,2]
+; SSE2-NEXT:    pand %xmm6, %xmm2
+; SSE2-NEXT:    pxor %xmm9, %xmm2
+; SSE2-NEXT:    pandn %xmm2, %xmm0
+; SSE2-NEXT:    movdqa {{.*#+}} xmm8 = [9223372036854775808,9223372036854775808]
+; SSE2-NEXT:    movdqa %xmm4, %xmm2
+; SSE2-NEXT:    pandn %xmm8, %xmm2
+; SSE2-NEXT:    movdqa {{.*#+}} xmm11 = [9223372036854775807,9223372036854775807]
+; SSE2-NEXT:    pand %xmm11, %xmm4
+; SSE2-NEXT:    por %xmm2, %xmm4
+; SSE2-NEXT:    pand %xmm0, %xmm4
+; SSE2-NEXT:    pandn %xmm10, %xmm0
+; SSE2-NEXT:    por %xmm4, %xmm0
+; SSE2-NEXT:    movdqa %xmm1, %xmm2
+; SSE2-NEXT:    psubq %xmm3, %xmm1
+; SSE2-NEXT:    pxor %xmm5, %xmm3
+; SSE2-NEXT:    movdqa %xmm5, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2-NEXT:    pand %xmm6, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm3, %xmm4
+; SSE2-NEXT:    pxor %xmm9, %xmm4
+; SSE2-NEXT:    pxor %xmm5, %xmm2
+; SSE2-NEXT:    movdqa %xmm5, %xmm3
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2-NEXT:    pand %xmm6, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE2-NEXT:    por %xmm2, %xmm3
+; SSE2-NEXT:    pxor %xmm9, %xmm3
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm2
+; SSE2-NEXT:    movdqa %xmm1, %xmm4
+; SSE2-NEXT:    pxor %xmm5, %xmm4
+; SSE2-NEXT:    movdqa %xmm5, %xmm6
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pand %xmm7, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3]
+; SSE2-NEXT:    por %xmm4, %xmm5
+; SSE2-NEXT:    movdqa %xmm5, %xmm4
+; SSE2-NEXT:    pxor %xmm9, %xmm4
+; SSE2-NEXT:    pcmpeqd %xmm3, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm3
+; SSE2-NEXT:    pxor %xmm9, %xmm3
+; SSE2-NEXT:    pandn %xmm3, %xmm2
+; SSE2-NEXT:    movdqa %xmm5, %xmm3
+; SSE2-NEXT:    pandn %xmm8, %xmm3
+; SSE2-NEXT:    pand %xmm11, %xmm5
+; SSE2-NEXT:    por %xmm3, %xmm5
+; SSE2-NEXT:    pand %xmm2, %xmm5
+; SSE2-NEXT:    pandn %xmm1, %xmm2
+; SSE2-NEXT:    por %xmm5, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm1
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v4i64:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm4, %rcx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm4, %r9
-; SSSE3-NEXT:    xorl %esi, %esi
-; SSSE3-NEXT:    movq %r9, %rdx
-; SSSE3-NEXT:    subq %rcx, %rdx
-; SSSE3-NEXT:    setns %sil
-; SSSE3-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSSE3-NEXT:    addq %r8, %rsi
-; SSSE3-NEXT:    subq %rcx, %r9
-; SSSE3-NEXT:    cmovoq %rsi, %r9
-; SSSE3-NEXT:    movq %xmm3, %rsi
-; SSSE3-NEXT:    movq %xmm1, %rdx
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %rdx, %rcx
-; SSSE3-NEXT:    subq %rsi, %rcx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    subq %rsi, %rdx
-; SSSE3-NEXT:    cmovoq %rdi, %rdx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %rcx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %rax
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %rax, %rsi
-; SSSE3-NEXT:    subq %rcx, %rsi
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    subq %rcx, %rax
-; SSSE3-NEXT:    cmovoq %rdi, %rax
-; SSSE3-NEXT:    movq %xmm2, %r10
-; SSSE3-NEXT:    movq %xmm0, %rcx
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %rcx, %rsi
-; SSSE3-NEXT:    subq %r10, %rsi
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    subq %r10, %rcx
-; SSSE3-NEXT:    cmovoq %rdi, %rcx
-; SSSE3-NEXT:    movq %rcx, %xmm0
-; SSSE3-NEXT:    movq %rax, %xmm1
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSSE3-NEXT:    movq %rdx, %xmm1
-; SSSE3-NEXT:    movq %r9, %xmm2
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; SSSE3-NEXT:    movdqa %xmm0, %xmm10
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm5 = [2147483648,2147483648]
+; SSSE3-NEXT:    psubq %xmm2, %xmm10
+; SSSE3-NEXT:    pxor %xmm5, %xmm2
+; SSSE3-NEXT:    movdqa %xmm5, %xmm6
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm7, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    por %xmm2, %xmm7
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm9
+; SSSE3-NEXT:    pxor %xmm9, %xmm7
+; SSSE3-NEXT:    pxor %xmm5, %xmm0
+; SSSE3-NEXT:    movdqa %xmm5, %xmm2
+; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm6, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSSE3-NEXT:    por %xmm0, %xmm2
+; SSSE3-NEXT:    pxor %xmm9, %xmm2
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm7, %xmm0
+; SSSE3-NEXT:    movdqa %xmm10, %xmm6
+; SSSE3-NEXT:    pxor %xmm5, %xmm6
+; SSSE3-NEXT:    movdqa %xmm5, %xmm7
+; SSSE3-NEXT:    pcmpgtd %xmm6, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm4, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm7[1,1,3,3]
+; SSSE3-NEXT:    por %xmm6, %xmm4
+; SSSE3-NEXT:    movdqa %xmm4, %xmm6
+; SSSE3-NEXT:    pxor %xmm9, %xmm6
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm6, %xmm2
+; SSSE3-NEXT:    pxor %xmm9, %xmm2
+; SSSE3-NEXT:    pandn %xmm2, %xmm0
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm8 = [9223372036854775808,9223372036854775808]
+; SSSE3-NEXT:    movdqa %xmm4, %xmm2
+; SSSE3-NEXT:    pandn %xmm8, %xmm2
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm11 = [9223372036854775807,9223372036854775807]
+; SSSE3-NEXT:    pand %xmm11, %xmm4
+; SSSE3-NEXT:    por %xmm2, %xmm4
+; SSSE3-NEXT:    pand %xmm0, %xmm4
+; SSSE3-NEXT:    pandn %xmm10, %xmm0
+; SSSE3-NEXT:    por %xmm4, %xmm0
+; SSSE3-NEXT:    movdqa %xmm1, %xmm2
+; SSSE3-NEXT:    psubq %xmm3, %xmm1
+; SSSE3-NEXT:    pxor %xmm5, %xmm3
+; SSSE3-NEXT:    movdqa %xmm5, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm6, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm3, %xmm4
+; SSSE3-NEXT:    pxor %xmm9, %xmm4
+; SSSE3-NEXT:    pxor %xmm5, %xmm2
+; SSSE3-NEXT:    movdqa %xmm5, %xmm3
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm6, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSSE3-NEXT:    por %xmm2, %xmm3
+; SSSE3-NEXT:    pxor %xmm9, %xmm3
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm2
+; SSSE3-NEXT:    movdqa %xmm1, %xmm4
+; SSSE3-NEXT:    pxor %xmm5, %xmm4
+; SSSE3-NEXT:    movdqa %xmm5, %xmm6
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm7, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    por %xmm4, %xmm5
+; SSSE3-NEXT:    movdqa %xmm5, %xmm4
+; SSSE3-NEXT:    pxor %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpeqd %xmm3, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm3
+; SSSE3-NEXT:    pxor %xmm9, %xmm3
+; SSSE3-NEXT:    pandn %xmm3, %xmm2
+; SSSE3-NEXT:    movdqa %xmm5, %xmm3
+; SSSE3-NEXT:    pandn %xmm8, %xmm3
+; SSSE3-NEXT:    pand %xmm11, %xmm5
+; SSSE3-NEXT:    por %xmm3, %xmm5
+; SSSE3-NEXT:    pand %xmm2, %xmm5
+; SSSE3-NEXT:    pandn %xmm1, %xmm2
+; SSSE3-NEXT:    por %xmm5, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm1
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v4i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movq %xmm3, %rcx
-; SSE41-NEXT:    movq %xmm1, %r9
-; SSE41-NEXT:    xorl %esi, %esi
-; SSE41-NEXT:    movq %r9, %rdx
-; SSE41-NEXT:    subq %rcx, %rdx
-; SSE41-NEXT:    setns %sil
-; SSE41-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE41-NEXT:    addq %r8, %rsi
-; SSE41-NEXT:    subq %rcx, %r9
-; SSE41-NEXT:    cmovoq %rsi, %r9
-; SSE41-NEXT:    pextrq $1, %xmm3, %rsi
-; SSE41-NEXT:    pextrq $1, %xmm1, %rdx
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %rdx, %rcx
-; SSE41-NEXT:    subq %rsi, %rcx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    subq %rsi, %rdx
-; SSE41-NEXT:    cmovoq %rdi, %rdx
-; SSE41-NEXT:    movq %xmm2, %rcx
-; SSE41-NEXT:    movq %xmm0, %rax
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %rax, %rsi
-; SSE41-NEXT:    subq %rcx, %rsi
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    subq %rcx, %rax
-; SSE41-NEXT:    cmovoq %rdi, %rax
-; SSE41-NEXT:    pextrq $1, %xmm2, %r10
-; SSE41-NEXT:    pextrq $1, %xmm0, %rcx
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %rcx, %rsi
-; SSE41-NEXT:    subq %r10, %rsi
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    subq %r10, %rcx
-; SSE41-NEXT:    cmovoq %rdi, %rcx
-; SSE41-NEXT:    movq %rcx, %xmm1
-; SSE41-NEXT:    movq %rax, %xmm0
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE41-NEXT:    movq %rdx, %xmm2
-; SSE41-NEXT:    movq %r9, %xmm1
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; SSE41-NEXT:    movdqa %xmm0, %xmm9
+; SSE41-NEXT:    movdqa {{.*#+}} xmm11 = [2147483648,2147483648]
+; SSE41-NEXT:    psubq %xmm2, %xmm9
+; SSE41-NEXT:    pxor %xmm11, %xmm2
+; SSE41-NEXT:    movdqa %xmm11, %xmm6
+; SSE41-NEXT:    pcmpgtd %xmm2, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm11, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
+; SSE41-NEXT:    pand %xmm7, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]
+; SSE41-NEXT:    por %xmm4, %xmm2
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm10
+; SSE41-NEXT:    pxor %xmm10, %xmm2
+; SSE41-NEXT:    pxor %xmm11, %xmm0
+; SSE41-NEXT:    movdqa %xmm11, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm11, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm7, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm0, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    pcmpeqq %xmm4, %xmm2
+; SSE41-NEXT:    movdqa %xmm9, %xmm0
+; SSE41-NEXT:    pxor %xmm11, %xmm0
+; SSE41-NEXT:    movdqa %xmm11, %xmm7
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm7
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm7[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm11, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm6, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[1,1,3,3]
+; SSE41-NEXT:    por %xmm5, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm5
+; SSE41-NEXT:    pxor %xmm10, %xmm5
+; SSE41-NEXT:    pcmpeqq %xmm4, %xmm5
+; SSE41-NEXT:    pxor %xmm10, %xmm5
+; SSE41-NEXT:    pandn %xmm5, %xmm2
+; SSE41-NEXT:    movapd {{.*#+}} xmm8 = [9223372036854775807,9223372036854775807]
+; SSE41-NEXT:    movapd {{.*#+}} xmm7 = [9223372036854775808,9223372036854775808]
+; SSE41-NEXT:    movapd %xmm7, %xmm4
+; SSE41-NEXT:    blendvpd %xmm0, %xmm8, %xmm4
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm9
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    psubq %xmm3, %xmm1
+; SSE41-NEXT:    pxor %xmm11, %xmm3
+; SSE41-NEXT:    movdqa %xmm11, %xmm2
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm11, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE41-NEXT:    pand %xmm4, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE41-NEXT:    por %xmm3, %xmm2
+; SSE41-NEXT:    pxor %xmm10, %xmm2
+; SSE41-NEXT:    pxor %xmm11, %xmm0
+; SSE41-NEXT:    movdqa %xmm11, %xmm3
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm11, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; SSE41-NEXT:    por %xmm0, %xmm3
+; SSE41-NEXT:    pxor %xmm10, %xmm3
+; SSE41-NEXT:    pcmpeqq %xmm3, %xmm2
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pxor %xmm11, %xmm0
+; SSE41-NEXT:    movdqa %xmm11, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm11, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm6, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    pcmpeqq %xmm3, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    pandn %xmm4, %xmm2
+; SSE41-NEXT:    blendvpd %xmm0, %xmm8, %xmm7
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm7, %xmm1
+; SSE41-NEXT:    movapd %xmm9, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: v4i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovq %xmm1, %rdx
-; AVX1-NEXT:    vmovq %xmm0, %r9
-; AVX1-NEXT:    xorl %esi, %esi
-; AVX1-NEXT:    movq %r9, %rcx
-; AVX1-NEXT:    subq %rdx, %rcx
-; AVX1-NEXT:    setns %sil
-; AVX1-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX1-NEXT:    addq %r8, %rsi
-; AVX1-NEXT:    subq %rdx, %r9
-; AVX1-NEXT:    cmovoq %rsi, %r9
-; AVX1-NEXT:    vpextrq $1, %xmm1, %rsi
-; AVX1-NEXT:    vpextrq $1, %xmm0, %rcx
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movq %rcx, %rdx
-; AVX1-NEXT:    subq %rsi, %rdx
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addq %r8, %rdi
-; AVX1-NEXT:    subq %rsi, %rcx
-; AVX1-NEXT:    cmovoq %rdi, %rcx
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vmovq %xmm1, %rdx
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vmovq %xmm0, %rax
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movq %rax, %rsi
-; AVX1-NEXT:    subq %rdx, %rsi
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addq %r8, %rdi
-; AVX1-NEXT:    subq %rdx, %rax
-; AVX1-NEXT:    cmovoq %rdi, %rax
-; AVX1-NEXT:    vpextrq $1, %xmm1, %r10
-; AVX1-NEXT:    vpextrq $1, %xmm0, %rdx
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movq %rdx, %rsi
-; AVX1-NEXT:    subq %r10, %rsi
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addq %r8, %rdi
-; AVX1-NEXT:    subq %r10, %rdx
-; AVX1-NEXT:    cmovoq %rdi, %rdx
-; AVX1-NEXT:    vmovq %rdx, %xmm0
-; AVX1-NEXT:    vmovq %rax, %xmm1
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX1-NEXT:    vmovq %rcx, %xmm1
-; AVX1-NEXT:    vmovq %r9, %xmm2
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm9
+; AVX1-NEXT:    vpxor %xmm10, %xmm10, %xmm10
+; AVX1-NEXT:    vpcmpgtq %xmm9, %xmm10, %xmm4
+; AVX1-NEXT:    vpcmpeqd %xmm5, %xmm5, %xmm5
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm4
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm6
+; AVX1-NEXT:    vpcmpgtq %xmm6, %xmm10, %xmm7
+; AVX1-NEXT:    vpxor %xmm5, %xmm7, %xmm7
+; AVX1-NEXT:    vpcmpeqq %xmm4, %xmm7, %xmm4
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm8
+; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm10, %xmm4
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm4
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm10, %xmm2
+; AVX1-NEXT:    vpxor %xmm5, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqq %xmm4, %xmm2, %xmm4
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm4
+; AVX1-NEXT:    vinsertf128 $1, %xmm8, %ymm4, %ymm8
+; AVX1-NEXT:    vpsubq %xmm9, %xmm6, %xmm6
+; AVX1-NEXT:    vpcmpgtq %xmm6, %xmm10, %xmm4
+; AVX1-NEXT:    vpxor %xmm5, %xmm4, %xmm3
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm7, %xmm3
+; AVX1-NEXT:    vpxor %xmm5, %xmm3, %xmm3
+; AVX1-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm10, %xmm1
+; AVX1-NEXT:    vpxor %xmm5, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqq %xmm1, %xmm2, %xmm1
+; AVX1-NEXT:    vpxor %xmm5, %xmm1, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
+; AVX1-NEXT:    vandpd %ymm1, %ymm8, %ymm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm2
+; AVX1-NEXT:    vmovapd {{.*#+}} ymm3 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
+; AVX1-NEXT:    vblendvpd %ymm2, {{.*}}(%rip), %ymm3, %ymm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm0
+; AVX1-NEXT:    vblendvpd %ymm1, %ymm2, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: v4i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovq %xmm1, %rdx
-; AVX2-NEXT:    vmovq %xmm0, %r9
-; AVX2-NEXT:    xorl %esi, %esi
-; AVX2-NEXT:    movq %r9, %rcx
-; AVX2-NEXT:    subq %rdx, %rcx
-; AVX2-NEXT:    setns %sil
-; AVX2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX2-NEXT:    addq %r8, %rsi
-; AVX2-NEXT:    subq %rdx, %r9
-; AVX2-NEXT:    cmovoq %rsi, %r9
-; AVX2-NEXT:    vpextrq $1, %xmm1, %rsi
-; AVX2-NEXT:    vpextrq $1, %xmm0, %rcx
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movq %rcx, %rdx
-; AVX2-NEXT:    subq %rsi, %rdx
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addq %r8, %rdi
-; AVX2-NEXT:    subq %rsi, %rcx
-; AVX2-NEXT:    cmovoq %rdi, %rcx
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; AVX2-NEXT:    vmovq %xmm1, %rdx
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; AVX2-NEXT:    vmovq %xmm0, %rax
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movq %rax, %rsi
-; AVX2-NEXT:    subq %rdx, %rsi
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addq %r8, %rdi
-; AVX2-NEXT:    subq %rdx, %rax
-; AVX2-NEXT:    cmovoq %rdi, %rax
-; AVX2-NEXT:    vpextrq $1, %xmm1, %r10
-; AVX2-NEXT:    vpextrq $1, %xmm0, %rdx
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movq %rdx, %rsi
-; AVX2-NEXT:    subq %r10, %rsi
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addq %r8, %rdi
-; AVX2-NEXT:    subq %r10, %rdx
-; AVX2-NEXT:    cmovoq %rdi, %rdx
-; AVX2-NEXT:    vmovq %rdx, %xmm0
-; AVX2-NEXT:    vmovq %rax, %xmm1
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX2-NEXT:    vmovq %rcx, %xmm1
-; AVX2-NEXT:    vmovq %r9, %xmm2
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpgtq %ymm1, %ymm2, %ymm3
+; AVX2-NEXT:    vpcmpeqd %ymm4, %ymm4, %ymm4
+; AVX2-NEXT:    vpxor %ymm4, %ymm3, %ymm3
+; AVX2-NEXT:    vpcmpgtq %ymm0, %ymm2, %ymm5
+; AVX2-NEXT:    vpxor %ymm4, %ymm5, %ymm5
+; AVX2-NEXT:    vpcmpeqq %ymm3, %ymm5, %ymm3
+; AVX2-NEXT:    vpsubq %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpgtq %ymm0, %ymm2, %ymm1
+; AVX2-NEXT:    vpxor %ymm4, %ymm1, %ymm1
+; AVX2-NEXT:    vpcmpeqq %ymm1, %ymm5, %ymm1
+; AVX2-NEXT:    vpxor %ymm4, %ymm1, %ymm1
+; AVX2-NEXT:    vpandn %ymm1, %ymm3, %ymm1
+; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm2 = [9223372036854775807,9223372036854775807,9223372036854775807,9223372036854775807]
+; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm3 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
+; AVX2-NEXT:    vblendvpd %ymm0, %ymm2, %ymm3, %ymm2
+; AVX2-NEXT:    vblendvpd %ymm1, %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: v4i64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vmovq %xmm1, %rdx
-; AVX512-NEXT:    vmovq %xmm0, %r9
-; AVX512-NEXT:    xorl %esi, %esi
-; AVX512-NEXT:    movq %r9, %rcx
-; AVX512-NEXT:    subq %rdx, %rcx
-; AVX512-NEXT:    setns %sil
-; AVX512-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX512-NEXT:    addq %r8, %rsi
-; AVX512-NEXT:    subq %rdx, %r9
-; AVX512-NEXT:    cmovoq %rsi, %r9
-; AVX512-NEXT:    vpextrq $1, %xmm1, %rsi
-; AVX512-NEXT:    vpextrq $1, %xmm0, %rcx
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movq %rcx, %rdx
-; AVX512-NEXT:    subq %rsi, %rdx
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addq %r8, %rdi
-; AVX512-NEXT:    subq %rsi, %rcx
-; AVX512-NEXT:    cmovoq %rdi, %rcx
-; AVX512-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; AVX512-NEXT:    vmovq %xmm1, %rdx
-; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; AVX512-NEXT:    vmovq %xmm0, %rax
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movq %rax, %rsi
-; AVX512-NEXT:    subq %rdx, %rsi
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addq %r8, %rdi
-; AVX512-NEXT:    subq %rdx, %rax
-; AVX512-NEXT:    cmovoq %rdi, %rax
-; AVX512-NEXT:    vpextrq $1, %xmm1, %r10
-; AVX512-NEXT:    vpextrq $1, %xmm0, %rdx
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movq %rdx, %rsi
-; AVX512-NEXT:    subq %r10, %rsi
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addq %r8, %rdi
-; AVX512-NEXT:    subq %r10, %rdx
-; AVX512-NEXT:    cmovoq %rdi, %rdx
-; AVX512-NEXT:    vmovq %rdx, %xmm0
-; AVX512-NEXT:    vmovq %rax, %xmm1
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX512-NEXT:    vmovq %rcx, %xmm1
-; AVX512-NEXT:    vmovq %r9, %xmm2
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltq %ymm2, %ymm1, %k0
+; AVX512-NEXT:    vpcmpnltq %ymm2, %ymm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpsubq %ymm1, %ymm0, %ymm0
+; AVX512-NEXT:    vpcmpnltq %ymm2, %ymm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtq %ymm0, %ymm2, %k2
+; AVX512-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
+; AVX512-NEXT:    vpbroadcastq {{.*}}(%rip), %ymm1 {%k2}
+; AVX512-NEXT:    vmovdqa64 %ymm1, %ymm0 {%k1}
 ; AVX512-NEXT:    retq
   %z = call <4 x i64> @llvm.ssub.sat.v4i64(<4 x i64> %x, <4 x i64> %y)
   ret <4 x i64> %z
@@ -3233,609 +2319,713 @@
 define <8 x i64> @v8i64(<8 x i64> %x, <8 x i64> %y) nounwind {
 ; SSE2-LABEL: v8i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    pushq %r15
-; SSE2-NEXT:    pushq %r14
-; SSE2-NEXT:    pushq %rbx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[2,3,0,1]
-; SSE2-NEXT:    movq %xmm8, %rax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[2,3,0,1]
-; SSE2-NEXT:    movq %xmm8, %r9
-; SSE2-NEXT:    xorl %ecx, %ecx
-; SSE2-NEXT:    movq %r9, %rdx
-; SSE2-NEXT:    subq %rax, %rdx
-; SSE2-NEXT:    setns %cl
-; SSE2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE2-NEXT:    addq %r8, %rcx
-; SSE2-NEXT:    subq %rax, %r9
-; SSE2-NEXT:    cmovoq %rcx, %r9
-; SSE2-NEXT:    movq %xmm5, %rax
-; SSE2-NEXT:    movq %xmm1, %r10
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movq %r10, %rsi
-; SSE2-NEXT:    subq %rax, %rsi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    addq %r8, %rdx
-; SSE2-NEXT:    subq %rax, %r10
-; SSE2-NEXT:    cmovoq %rdx, %r10
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %rax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %r11
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movq %r11, %rdi
-; SSE2-NEXT:    subq %rax, %rdi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    addq %r8, %rdx
-; SSE2-NEXT:    subq %rax, %r11
-; SSE2-NEXT:    cmovoq %rdx, %r11
-; SSE2-NEXT:    movq %xmm6, %rax
-; SSE2-NEXT:    movq %xmm2, %r14
-; SSE2-NEXT:    xorl %edx, %edx
-; SSE2-NEXT:    movq %r14, %rdi
-; SSE2-NEXT:    subq %rax, %rdi
-; SSE2-NEXT:    setns %dl
-; SSE2-NEXT:    addq %r8, %rdx
-; SSE2-NEXT:    subq %rax, %r14
-; SSE2-NEXT:    cmovoq %rdx, %r14
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %rax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %r15
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %r15, %rdx
-; SSE2-NEXT:    subq %rax, %rdx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    subq %rax, %r15
-; SSE2-NEXT:    cmovoq %rdi, %r15
-; SSE2-NEXT:    movq %xmm7, %rdx
-; SSE2-NEXT:    movq %xmm3, %rax
-; SSE2-NEXT:    xorl %edi, %edi
-; SSE2-NEXT:    movq %rax, %rbx
-; SSE2-NEXT:    subq %rdx, %rbx
-; SSE2-NEXT:    setns %dil
-; SSE2-NEXT:    addq %r8, %rdi
-; SSE2-NEXT:    subq %rdx, %rax
-; SSE2-NEXT:    cmovoq %rdi, %rax
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %rdx
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSE2-NEXT:    movq %xmm1, %rcx
-; SSE2-NEXT:    xorl %ebx, %ebx
-; SSE2-NEXT:    movq %rcx, %rdi
-; SSE2-NEXT:    subq %rdx, %rdi
-; SSE2-NEXT:    setns %bl
-; SSE2-NEXT:    addq %r8, %rbx
-; SSE2-NEXT:    subq %rdx, %rcx
-; SSE2-NEXT:    cmovoq %rbx, %rcx
-; SSE2-NEXT:    movq %xmm4, %rdx
-; SSE2-NEXT:    movq %xmm0, %rsi
-; SSE2-NEXT:    xorl %ebx, %ebx
-; SSE2-NEXT:    movq %rsi, %rdi
-; SSE2-NEXT:    subq %rdx, %rdi
-; SSE2-NEXT:    setns %bl
-; SSE2-NEXT:    addq %r8, %rbx
-; SSE2-NEXT:    subq %rdx, %rsi
-; SSE2-NEXT:    cmovoq %rbx, %rsi
-; SSE2-NEXT:    movq %rsi, %xmm0
-; SSE2-NEXT:    movq %rcx, %xmm1
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE2-NEXT:    movq %r10, %xmm1
-; SSE2-NEXT:    movq %r9, %xmm2
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; SSE2-NEXT:    movq %r14, %xmm2
-; SSE2-NEXT:    movq %r11, %xmm3
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
-; SSE2-NEXT:    movq %rax, %xmm3
-; SSE2-NEXT:    movq %r15, %xmm4
-; SSE2-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm4[0]
-; SSE2-NEXT:    popq %rbx
-; SSE2-NEXT:    popq %r14
-; SSE2-NEXT:    popq %r15
+; SSE2-NEXT:    movdqa %xmm1, %xmm8
+; SSE2-NEXT:    movdqa %xmm0, %xmm13
+; SSE2-NEXT:    movdqa {{.*#+}} xmm9 = [2147483648,2147483648]
+; SSE2-NEXT:    psubq %xmm4, %xmm13
+; SSE2-NEXT:    pxor %xmm9, %xmm4
+; SSE2-NEXT:    movdqa %xmm9, %xmm1
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm1[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pand %xmm10, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    por %xmm4, %xmm1
+; SSE2-NEXT:    pcmpeqd %xmm10, %xmm10
+; SSE2-NEXT:    pxor %xmm10, %xmm1
+; SSE2-NEXT:    pxor %xmm9, %xmm0
+; SSE2-NEXT:    movdqa %xmm9, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pand %xmm11, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm15 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm0, %xmm15
+; SSE2-NEXT:    pxor %xmm10, %xmm15
+; SSE2-NEXT:    pcmpeqd %xmm15, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,0,3,2]
+; SSE2-NEXT:    pand %xmm1, %xmm0
+; SSE2-NEXT:    movdqa %xmm13, %xmm1
+; SSE2-NEXT:    pxor %xmm9, %xmm1
+; SSE2-NEXT:    movdqa %xmm9, %xmm11
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm11
+; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm11[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm14 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pand %xmm12, %xmm14
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm11[1,1,3,3]
+; SSE2-NEXT:    por %xmm14, %xmm1
+; SSE2-NEXT:    movdqa %xmm1, %xmm11
+; SSE2-NEXT:    pxor %xmm10, %xmm11
+; SSE2-NEXT:    pcmpeqd %xmm15, %xmm11
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm11[1,0,3,2]
+; SSE2-NEXT:    pand %xmm11, %xmm4
+; SSE2-NEXT:    pxor %xmm10, %xmm4
+; SSE2-NEXT:    pandn %xmm4, %xmm0
+; SSE2-NEXT:    movdqa {{.*#+}} xmm11 = [9223372036854775808,9223372036854775808]
+; SSE2-NEXT:    movdqa %xmm1, %xmm4
+; SSE2-NEXT:    pandn %xmm11, %xmm4
+; SSE2-NEXT:    movdqa {{.*#+}} xmm12 = [9223372036854775807,9223372036854775807]
+; SSE2-NEXT:    pand %xmm12, %xmm1
+; SSE2-NEXT:    por %xmm4, %xmm1
+; SSE2-NEXT:    pand %xmm0, %xmm1
+; SSE2-NEXT:    pandn %xmm13, %xmm0
+; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    movdqa %xmm8, %xmm1
+; SSE2-NEXT:    psubq %xmm5, %xmm8
+; SSE2-NEXT:    pxor %xmm9, %xmm5
+; SSE2-NEXT:    movdqa %xmm9, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE2-NEXT:    pand %xmm13, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm5, %xmm4
+; SSE2-NEXT:    pxor %xmm10, %xmm4
+; SSE2-NEXT:    pxor %xmm9, %xmm1
+; SSE2-NEXT:    movdqa %xmm9, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm1, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm5[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE2-NEXT:    pand %xmm13, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE2-NEXT:    por %xmm1, %xmm5
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm1
+; SSE2-NEXT:    movdqa %xmm8, %xmm4
+; SSE2-NEXT:    pxor %xmm9, %xmm4
+; SSE2-NEXT:    movdqa %xmm9, %xmm13
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm13
+; SSE2-NEXT:    pshufd {{.*#+}} xmm14 = xmm13[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm15 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pand %xmm14, %xmm15
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm13[1,1,3,3]
+; SSE2-NEXT:    por %xmm15, %xmm4
+; SSE2-NEXT:    movdqa %xmm4, %xmm13
+; SSE2-NEXT:    pxor %xmm10, %xmm13
+; SSE2-NEXT:    pcmpeqd %xmm5, %xmm13
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm13[1,0,3,2]
+; SSE2-NEXT:    pand %xmm13, %xmm5
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pandn %xmm5, %xmm1
+; SSE2-NEXT:    movdqa %xmm4, %xmm5
+; SSE2-NEXT:    pandn %xmm11, %xmm5
+; SSE2-NEXT:    pand %xmm12, %xmm4
+; SSE2-NEXT:    por %xmm5, %xmm4
+; SSE2-NEXT:    pand %xmm1, %xmm4
+; SSE2-NEXT:    pandn %xmm8, %xmm1
+; SSE2-NEXT:    por %xmm4, %xmm1
+; SSE2-NEXT:    movdqa %xmm2, %xmm4
+; SSE2-NEXT:    psubq %xmm6, %xmm2
+; SSE2-NEXT:    pxor %xmm9, %xmm6
+; SSE2-NEXT:    movdqa %xmm9, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm6, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE2-NEXT:    pand %xmm8, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE2-NEXT:    por %xmm6, %xmm5
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pxor %xmm9, %xmm4
+; SSE2-NEXT:    movdqa %xmm9, %xmm6
+; SSE2-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    pand %xmm8, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE2-NEXT:    por %xmm4, %xmm6
+; SSE2-NEXT:    pxor %xmm10, %xmm6
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[1,0,3,2]
+; SSE2-NEXT:    pand %xmm5, %xmm8
+; SSE2-NEXT:    movdqa %xmm2, %xmm5
+; SSE2-NEXT:    pxor %xmm9, %xmm5
+; SSE2-NEXT:    movdqa %xmm9, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm5, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE2-NEXT:    pand %xmm13, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm5, %xmm4
+; SSE2-NEXT:    movdqa %xmm4, %xmm5
+; SSE2-NEXT:    pxor %xmm10, %xmm5
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,0,3,2]
+; SSE2-NEXT:    pand %xmm5, %xmm6
+; SSE2-NEXT:    pxor %xmm10, %xmm6
+; SSE2-NEXT:    pandn %xmm6, %xmm8
+; SSE2-NEXT:    movdqa %xmm4, %xmm5
+; SSE2-NEXT:    pandn %xmm11, %xmm5
+; SSE2-NEXT:    pand %xmm12, %xmm4
+; SSE2-NEXT:    por %xmm5, %xmm4
+; SSE2-NEXT:    pand %xmm8, %xmm4
+; SSE2-NEXT:    pandn %xmm2, %xmm8
+; SSE2-NEXT:    por %xmm4, %xmm8
+; SSE2-NEXT:    movdqa %xmm3, %xmm2
+; SSE2-NEXT:    psubq %xmm7, %xmm3
+; SSE2-NEXT:    pxor %xmm9, %xmm7
+; SSE2-NEXT:    movdqa %xmm9, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm7, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm7
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm7[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm6, %xmm4
+; SSE2-NEXT:    pxor %xmm10, %xmm4
+; SSE2-NEXT:    pxor %xmm9, %xmm2
+; SSE2-NEXT:    movdqa %xmm9, %xmm5
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2-NEXT:    pand %xmm6, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,1,3,3]
+; SSE2-NEXT:    por %xmm2, %xmm6
+; SSE2-NEXT:    pxor %xmm10, %xmm6
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,0,3,2]
+; SSE2-NEXT:    pand %xmm4, %xmm5
+; SSE2-NEXT:    movdqa %xmm3, %xmm2
+; SSE2-NEXT:    pxor %xmm9, %xmm2
+; SSE2-NEXT:    movdqa %xmm9, %xmm4
+; SSE2-NEXT:    pcmpgtd %xmm2, %xmm4
+; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSE2-NEXT:    pcmpeqd %xmm9, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSE2-NEXT:    pand %xmm7, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE2-NEXT:    por %xmm2, %xmm4
+; SSE2-NEXT:    movdqa %xmm4, %xmm2
+; SSE2-NEXT:    pxor %xmm10, %xmm2
+; SSE2-NEXT:    pcmpeqd %xmm6, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[1,0,3,2]
+; SSE2-NEXT:    pand %xmm2, %xmm6
+; SSE2-NEXT:    pxor %xmm10, %xmm6
+; SSE2-NEXT:    pandn %xmm6, %xmm5
+; SSE2-NEXT:    movdqa %xmm4, %xmm2
+; SSE2-NEXT:    pandn %xmm11, %xmm2
+; SSE2-NEXT:    pand %xmm12, %xmm4
+; SSE2-NEXT:    por %xmm2, %xmm4
+; SSE2-NEXT:    pand %xmm5, %xmm4
+; SSE2-NEXT:    pandn %xmm3, %xmm5
+; SSE2-NEXT:    por %xmm4, %xmm5
+; SSE2-NEXT:    movdqa %xmm8, %xmm2
+; SSE2-NEXT:    movdqa %xmm5, %xmm3
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: v8i64:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pushq %r15
-; SSSE3-NEXT:    pushq %r14
-; SSSE3-NEXT:    pushq %rbx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm8, %rax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm1[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm8, %r9
-; SSSE3-NEXT:    xorl %ecx, %ecx
-; SSSE3-NEXT:    movq %r9, %rdx
-; SSSE3-NEXT:    subq %rax, %rdx
-; SSSE3-NEXT:    setns %cl
-; SSSE3-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSSE3-NEXT:    addq %r8, %rcx
-; SSSE3-NEXT:    subq %rax, %r9
-; SSSE3-NEXT:    cmovoq %rcx, %r9
-; SSSE3-NEXT:    movq %xmm5, %rax
-; SSSE3-NEXT:    movq %xmm1, %r10
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movq %r10, %rsi
-; SSSE3-NEXT:    subq %rax, %rsi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    addq %r8, %rdx
-; SSSE3-NEXT:    subq %rax, %r10
-; SSSE3-NEXT:    cmovoq %rdx, %r10
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %rax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %r11
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movq %r11, %rdi
-; SSSE3-NEXT:    subq %rax, %rdi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    addq %r8, %rdx
-; SSSE3-NEXT:    subq %rax, %r11
-; SSSE3-NEXT:    cmovoq %rdx, %r11
-; SSSE3-NEXT:    movq %xmm6, %rax
-; SSSE3-NEXT:    movq %xmm2, %r14
-; SSSE3-NEXT:    xorl %edx, %edx
-; SSSE3-NEXT:    movq %r14, %rdi
-; SSSE3-NEXT:    subq %rax, %rdi
-; SSSE3-NEXT:    setns %dl
-; SSSE3-NEXT:    addq %r8, %rdx
-; SSSE3-NEXT:    subq %rax, %r14
-; SSSE3-NEXT:    cmovoq %rdx, %r14
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %rax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %r15
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %r15, %rdx
-; SSSE3-NEXT:    subq %rax, %rdx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    subq %rax, %r15
-; SSSE3-NEXT:    cmovoq %rdi, %r15
-; SSSE3-NEXT:    movq %xmm7, %rdx
-; SSSE3-NEXT:    movq %xmm3, %rax
-; SSSE3-NEXT:    xorl %edi, %edi
-; SSSE3-NEXT:    movq %rax, %rbx
-; SSSE3-NEXT:    subq %rdx, %rbx
-; SSSE3-NEXT:    setns %dil
-; SSSE3-NEXT:    addq %r8, %rdi
-; SSSE3-NEXT:    subq %rdx, %rax
-; SSSE3-NEXT:    cmovoq %rdi, %rax
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %rdx
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSSE3-NEXT:    movq %xmm1, %rcx
-; SSSE3-NEXT:    xorl %ebx, %ebx
-; SSSE3-NEXT:    movq %rcx, %rdi
-; SSSE3-NEXT:    subq %rdx, %rdi
-; SSSE3-NEXT:    setns %bl
-; SSSE3-NEXT:    addq %r8, %rbx
-; SSSE3-NEXT:    subq %rdx, %rcx
-; SSSE3-NEXT:    cmovoq %rbx, %rcx
-; SSSE3-NEXT:    movq %xmm4, %rdx
-; SSSE3-NEXT:    movq %xmm0, %rsi
-; SSSE3-NEXT:    xorl %ebx, %ebx
-; SSSE3-NEXT:    movq %rsi, %rdi
-; SSSE3-NEXT:    subq %rdx, %rdi
-; SSSE3-NEXT:    setns %bl
-; SSSE3-NEXT:    addq %r8, %rbx
-; SSSE3-NEXT:    subq %rdx, %rsi
-; SSSE3-NEXT:    cmovoq %rbx, %rsi
-; SSSE3-NEXT:    movq %rsi, %xmm0
-; SSSE3-NEXT:    movq %rcx, %xmm1
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSSE3-NEXT:    movq %r10, %xmm1
-; SSSE3-NEXT:    movq %r9, %xmm2
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; SSSE3-NEXT:    movq %r14, %xmm2
-; SSSE3-NEXT:    movq %r11, %xmm3
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
-; SSSE3-NEXT:    movq %rax, %xmm3
-; SSSE3-NEXT:    movq %r15, %xmm4
-; SSSE3-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm4[0]
-; SSSE3-NEXT:    popq %rbx
-; SSSE3-NEXT:    popq %r14
-; SSSE3-NEXT:    popq %r15
+; SSSE3-NEXT:    movdqa %xmm1, %xmm8
+; SSSE3-NEXT:    movdqa %xmm0, %xmm13
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm9 = [2147483648,2147483648]
+; SSSE3-NEXT:    psubq %xmm4, %xmm13
+; SSSE3-NEXT:    pxor %xmm9, %xmm4
+; SSSE3-NEXT:    movdqa %xmm9, %xmm1
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm1[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm10, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    por %xmm4, %xmm1
+; SSSE3-NEXT:    pcmpeqd %xmm10, %xmm10
+; SSSE3-NEXT:    pxor %xmm10, %xmm1
+; SSSE3-NEXT:    pxor %xmm9, %xmm0
+; SSSE3-NEXT:    movdqa %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm11, %xmm0
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm15 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm0, %xmm15
+; SSSE3-NEXT:    pxor %xmm10, %xmm15
+; SSSE3-NEXT:    pcmpeqd %xmm15, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm1, %xmm0
+; SSSE3-NEXT:    movdqa %xmm13, %xmm1
+; SSSE3-NEXT:    pxor %xmm9, %xmm1
+; SSSE3-NEXT:    movdqa %xmm9, %xmm11
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm11
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm12 = xmm11[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm14 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm12, %xmm14
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm11[1,1,3,3]
+; SSSE3-NEXT:    por %xmm14, %xmm1
+; SSSE3-NEXT:    movdqa %xmm1, %xmm11
+; SSSE3-NEXT:    pxor %xmm10, %xmm11
+; SSSE3-NEXT:    pcmpeqd %xmm15, %xmm11
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm11[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm11, %xmm4
+; SSSE3-NEXT:    pxor %xmm10, %xmm4
+; SSSE3-NEXT:    pandn %xmm4, %xmm0
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm11 = [9223372036854775808,9223372036854775808]
+; SSSE3-NEXT:    movdqa %xmm1, %xmm4
+; SSSE3-NEXT:    pandn %xmm11, %xmm4
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm12 = [9223372036854775807,9223372036854775807]
+; SSSE3-NEXT:    pand %xmm12, %xmm1
+; SSSE3-NEXT:    por %xmm4, %xmm1
+; SSSE3-NEXT:    pand %xmm0, %xmm1
+; SSSE3-NEXT:    pandn %xmm13, %xmm0
+; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    movdqa %xmm8, %xmm1
+; SSSE3-NEXT:    psubq %xmm5, %xmm8
+; SSSE3-NEXT:    pxor %xmm9, %xmm5
+; SSSE3-NEXT:    movdqa %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm13, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm5, %xmm4
+; SSSE3-NEXT:    pxor %xmm10, %xmm4
+; SSSE3-NEXT:    pxor %xmm9, %xmm1
+; SSSE3-NEXT:    movdqa %xmm9, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm13 = xmm5[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm13, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    por %xmm1, %xmm5
+; SSSE3-NEXT:    pxor %xmm10, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm1
+; SSSE3-NEXT:    movdqa %xmm8, %xmm4
+; SSSE3-NEXT:    pxor %xmm9, %xmm4
+; SSSE3-NEXT:    movdqa %xmm9, %xmm13
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm13
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm14 = xmm13[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm15 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm14, %xmm15
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm13[1,1,3,3]
+; SSSE3-NEXT:    por %xmm15, %xmm4
+; SSSE3-NEXT:    movdqa %xmm4, %xmm13
+; SSSE3-NEXT:    pxor %xmm10, %xmm13
+; SSSE3-NEXT:    pcmpeqd %xmm5, %xmm13
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm13[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm13, %xmm5
+; SSSE3-NEXT:    pxor %xmm10, %xmm5
+; SSSE3-NEXT:    pandn %xmm5, %xmm1
+; SSSE3-NEXT:    movdqa %xmm4, %xmm5
+; SSSE3-NEXT:    pandn %xmm11, %xmm5
+; SSSE3-NEXT:    pand %xmm12, %xmm4
+; SSSE3-NEXT:    por %xmm5, %xmm4
+; SSSE3-NEXT:    pand %xmm1, %xmm4
+; SSSE3-NEXT:    pandn %xmm8, %xmm1
+; SSSE3-NEXT:    por %xmm4, %xmm1
+; SSSE3-NEXT:    movdqa %xmm2, %xmm4
+; SSSE3-NEXT:    psubq %xmm6, %xmm2
+; SSSE3-NEXT:    pxor %xmm9, %xmm6
+; SSSE3-NEXT:    movdqa %xmm9, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm6, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm8, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    por %xmm6, %xmm5
+; SSSE3-NEXT:    pxor %xmm10, %xmm5
+; SSSE3-NEXT:    pxor %xmm9, %xmm4
+; SSSE3-NEXT:    movdqa %xmm9, %xmm6
+; SSSE3-NEXT:    pcmpgtd %xmm4, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm8, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSSE3-NEXT:    por %xmm4, %xmm6
+; SSSE3-NEXT:    pxor %xmm10, %xmm6
+; SSSE3-NEXT:    pcmpeqd %xmm6, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm5, %xmm8
+; SSSE3-NEXT:    movdqa %xmm2, %xmm5
+; SSSE3-NEXT:    pxor %xmm9, %xmm5
+; SSSE3-NEXT:    movdqa %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm5, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm13, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm5, %xmm4
+; SSSE3-NEXT:    movdqa %xmm4, %xmm5
+; SSSE3-NEXT:    pxor %xmm10, %xmm5
+; SSSE3-NEXT:    pcmpeqd %xmm6, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm5, %xmm6
+; SSSE3-NEXT:    pxor %xmm10, %xmm6
+; SSSE3-NEXT:    pandn %xmm6, %xmm8
+; SSSE3-NEXT:    movdqa %xmm4, %xmm5
+; SSSE3-NEXT:    pandn %xmm11, %xmm5
+; SSSE3-NEXT:    pand %xmm12, %xmm4
+; SSSE3-NEXT:    por %xmm5, %xmm4
+; SSSE3-NEXT:    pand %xmm8, %xmm4
+; SSSE3-NEXT:    pandn %xmm2, %xmm8
+; SSSE3-NEXT:    por %xmm4, %xmm8
+; SSSE3-NEXT:    movdqa %xmm3, %xmm2
+; SSSE3-NEXT:    psubq %xmm7, %xmm3
+; SSSE3-NEXT:    pxor %xmm9, %xmm7
+; SSSE3-NEXT:    movdqa %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm7, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm7
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm7[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm5, %xmm6
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm6, %xmm4
+; SSSE3-NEXT:    pxor %xmm10, %xmm4
+; SSSE3-NEXT:    pxor %xmm9, %xmm2
+; SSSE3-NEXT:    movdqa %xmm9, %xmm5
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm5
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm6, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,1,3,3]
+; SSSE3-NEXT:    por %xmm2, %xmm6
+; SSSE3-NEXT:    pxor %xmm10, %xmm6
+; SSSE3-NEXT:    pcmpeqd %xmm6, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm4, %xmm5
+; SSSE3-NEXT:    movdqa %xmm3, %xmm2
+; SSSE3-NEXT:    pxor %xmm9, %xmm2
+; SSSE3-NEXT:    movdqa %xmm9, %xmm4
+; SSSE3-NEXT:    pcmpgtd %xmm2, %xmm4
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
+; SSSE3-NEXT:    pcmpeqd %xmm9, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; SSSE3-NEXT:    pand %xmm7, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSSE3-NEXT:    por %xmm2, %xmm4
+; SSSE3-NEXT:    movdqa %xmm4, %xmm2
+; SSSE3-NEXT:    pxor %xmm10, %xmm2
+; SSSE3-NEXT:    pcmpeqd %xmm6, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm2, %xmm6
+; SSSE3-NEXT:    pxor %xmm10, %xmm6
+; SSSE3-NEXT:    pandn %xmm6, %xmm5
+; SSSE3-NEXT:    movdqa %xmm4, %xmm2
+; SSSE3-NEXT:    pandn %xmm11, %xmm2
+; SSSE3-NEXT:    pand %xmm12, %xmm4
+; SSSE3-NEXT:    por %xmm2, %xmm4
+; SSSE3-NEXT:    pand %xmm5, %xmm4
+; SSSE3-NEXT:    pandn %xmm3, %xmm5
+; SSSE3-NEXT:    por %xmm4, %xmm5
+; SSSE3-NEXT:    movdqa %xmm8, %xmm2
+; SSSE3-NEXT:    movdqa %xmm5, %xmm3
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: v8i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    pushq %r15
-; SSE41-NEXT:    pushq %r14
-; SSE41-NEXT:    pushq %rbx
-; SSE41-NEXT:    movq %xmm5, %rax
-; SSE41-NEXT:    movq %xmm1, %r9
-; SSE41-NEXT:    xorl %ecx, %ecx
-; SSE41-NEXT:    movq %r9, %rdx
-; SSE41-NEXT:    subq %rax, %rdx
-; SSE41-NEXT:    setns %cl
-; SSE41-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; SSE41-NEXT:    addq %r8, %rcx
-; SSE41-NEXT:    subq %rax, %r9
-; SSE41-NEXT:    cmovoq %rcx, %r9
-; SSE41-NEXT:    pextrq $1, %xmm5, %rax
-; SSE41-NEXT:    pextrq $1, %xmm1, %r10
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movq %r10, %rsi
-; SSE41-NEXT:    subq %rax, %rsi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    addq %r8, %rdx
-; SSE41-NEXT:    subq %rax, %r10
-; SSE41-NEXT:    cmovoq %rdx, %r10
-; SSE41-NEXT:    movq %xmm6, %rax
-; SSE41-NEXT:    movq %xmm2, %r11
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movq %r11, %rdi
-; SSE41-NEXT:    subq %rax, %rdi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    addq %r8, %rdx
-; SSE41-NEXT:    subq %rax, %r11
-; SSE41-NEXT:    cmovoq %rdx, %r11
-; SSE41-NEXT:    pextrq $1, %xmm6, %rax
-; SSE41-NEXT:    pextrq $1, %xmm2, %r14
-; SSE41-NEXT:    xorl %edx, %edx
-; SSE41-NEXT:    movq %r14, %rdi
-; SSE41-NEXT:    subq %rax, %rdi
-; SSE41-NEXT:    setns %dl
-; SSE41-NEXT:    addq %r8, %rdx
-; SSE41-NEXT:    subq %rax, %r14
-; SSE41-NEXT:    cmovoq %rdx, %r14
-; SSE41-NEXT:    movq %xmm7, %rax
-; SSE41-NEXT:    movq %xmm3, %r15
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %r15, %rdx
-; SSE41-NEXT:    subq %rax, %rdx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    subq %rax, %r15
-; SSE41-NEXT:    cmovoq %rdi, %r15
-; SSE41-NEXT:    pextrq $1, %xmm7, %rdx
-; SSE41-NEXT:    pextrq $1, %xmm3, %rax
-; SSE41-NEXT:    xorl %edi, %edi
-; SSE41-NEXT:    movq %rax, %rbx
-; SSE41-NEXT:    subq %rdx, %rbx
-; SSE41-NEXT:    setns %dil
-; SSE41-NEXT:    addq %r8, %rdi
-; SSE41-NEXT:    subq %rdx, %rax
-; SSE41-NEXT:    cmovoq %rdi, %rax
-; SSE41-NEXT:    movq %xmm4, %rdx
-; SSE41-NEXT:    movq %xmm0, %rcx
-; SSE41-NEXT:    xorl %ebx, %ebx
-; SSE41-NEXT:    movq %rcx, %rdi
-; SSE41-NEXT:    subq %rdx, %rdi
-; SSE41-NEXT:    setns %bl
-; SSE41-NEXT:    addq %r8, %rbx
-; SSE41-NEXT:    subq %rdx, %rcx
-; SSE41-NEXT:    cmovoq %rbx, %rcx
-; SSE41-NEXT:    pextrq $1, %xmm4, %rdx
-; SSE41-NEXT:    pextrq $1, %xmm0, %rsi
-; SSE41-NEXT:    xorl %ebx, %ebx
-; SSE41-NEXT:    movq %rsi, %rdi
-; SSE41-NEXT:    subq %rdx, %rdi
-; SSE41-NEXT:    setns %bl
-; SSE41-NEXT:    addq %r8, %rbx
-; SSE41-NEXT:    subq %rdx, %rsi
-; SSE41-NEXT:    cmovoq %rbx, %rsi
-; SSE41-NEXT:    movq %rsi, %xmm1
-; SSE41-NEXT:    movq %rcx, %xmm0
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE41-NEXT:    movq %r10, %xmm2
-; SSE41-NEXT:    movq %r9, %xmm1
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; SSE41-NEXT:    movq %r14, %xmm3
-; SSE41-NEXT:    movq %r11, %xmm2
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
-; SSE41-NEXT:    movq %rax, %xmm4
-; SSE41-NEXT:    movq %r15, %xmm3
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm4[0]
-; SSE41-NEXT:    popq %rbx
-; SSE41-NEXT:    popq %r14
-; SSE41-NEXT:    popq %r15
+; SSE41-NEXT:    movdqa %xmm0, %xmm8
+; SSE41-NEXT:    movdqa {{.*#+}} xmm9 = [2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm0, %xmm11
+; SSE41-NEXT:    psubq %xmm4, %xmm8
+; SSE41-NEXT:    pxor %xmm9, %xmm4
+; SSE41-NEXT:    movdqa %xmm9, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm4, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm10 = xmm0[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm12 = xmm4[1,1,3,3]
+; SSE41-NEXT:    pand %xmm10, %xmm12
+; SSE41-NEXT:    pshufd {{.*#+}} xmm15 = xmm0[1,1,3,3]
+; SSE41-NEXT:    por %xmm12, %xmm15
+; SSE41-NEXT:    pcmpeqd %xmm10, %xmm10
+; SSE41-NEXT:    pxor %xmm10, %xmm15
+; SSE41-NEXT:    pxor %xmm9, %xmm11
+; SSE41-NEXT:    movdqa %xmm9, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm11, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm12 = xmm0[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm11
+; SSE41-NEXT:    pshufd {{.*#+}} xmm11 = xmm11[1,1,3,3]
+; SSE41-NEXT:    pand %xmm12, %xmm11
+; SSE41-NEXT:    pshufd {{.*#+}} xmm12 = xmm0[1,1,3,3]
+; SSE41-NEXT:    por %xmm11, %xmm12
+; SSE41-NEXT:    pxor %xmm10, %xmm12
+; SSE41-NEXT:    pcmpeqq %xmm12, %xmm15
+; SSE41-NEXT:    movdqa %xmm8, %xmm0
+; SSE41-NEXT:    pxor %xmm9, %xmm0
+; SSE41-NEXT:    movdqa %xmm9, %xmm11
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm11
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm11[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm14 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm13, %xmm14
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm11[1,1,3,3]
+; SSE41-NEXT:    por %xmm14, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    pcmpeqq %xmm12, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    pandn %xmm4, %xmm15
+; SSE41-NEXT:    movapd {{.*#+}} xmm12 = [9223372036854775807,9223372036854775807]
+; SSE41-NEXT:    movapd {{.*#+}} xmm11 = [9223372036854775808,9223372036854775808]
+; SSE41-NEXT:    movapd %xmm11, %xmm4
+; SSE41-NEXT:    blendvpd %xmm0, %xmm12, %xmm4
+; SSE41-NEXT:    movdqa %xmm15, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm4, %xmm8
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    psubq %xmm5, %xmm1
+; SSE41-NEXT:    pxor %xmm9, %xmm5
+; SSE41-NEXT:    movdqa %xmm9, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm5, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE41-NEXT:    pand %xmm13, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm5, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    pxor %xmm9, %xmm0
+; SSE41-NEXT:    movdqa %xmm9, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm5[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm13, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm14 = xmm5[1,1,3,3]
+; SSE41-NEXT:    por %xmm0, %xmm14
+; SSE41-NEXT:    pxor %xmm10, %xmm14
+; SSE41-NEXT:    pcmpeqq %xmm14, %xmm4
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pxor %xmm9, %xmm0
+; SSE41-NEXT:    movdqa %xmm9, %xmm13
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm13
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm13[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm15 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm15
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm13[1,1,3,3]
+; SSE41-NEXT:    por %xmm15, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm5
+; SSE41-NEXT:    pxor %xmm10, %xmm5
+; SSE41-NEXT:    pcmpeqq %xmm14, %xmm5
+; SSE41-NEXT:    pxor %xmm10, %xmm5
+; SSE41-NEXT:    pandn %xmm5, %xmm4
+; SSE41-NEXT:    movapd %xmm11, %xmm5
+; SSE41-NEXT:    blendvpd %xmm0, %xmm12, %xmm5
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm5, %xmm1
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    psubq %xmm6, %xmm2
+; SSE41-NEXT:    pxor %xmm9, %xmm6
+; SSE41-NEXT:    movdqa %xmm9, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm6, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm6, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    pxor %xmm9, %xmm0
+; SSE41-NEXT:    movdqa %xmm9, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm6, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE41-NEXT:    por %xmm0, %xmm5
+; SSE41-NEXT:    pxor %xmm10, %xmm5
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm4
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm9, %xmm0
+; SSE41-NEXT:    movdqa %xmm9, %xmm6
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm6[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm14 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm13, %xmm14
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
+; SSE41-NEXT:    por %xmm14, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm6
+; SSE41-NEXT:    pxor %xmm10, %xmm6
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm6
+; SSE41-NEXT:    pxor %xmm10, %xmm6
+; SSE41-NEXT:    pandn %xmm6, %xmm4
+; SSE41-NEXT:    movapd %xmm11, %xmm5
+; SSE41-NEXT:    blendvpd %xmm0, %xmm12, %xmm5
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm5, %xmm2
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    psubq %xmm7, %xmm3
+; SSE41-NEXT:    pxor %xmm9, %xmm7
+; SSE41-NEXT:    movdqa %xmm9, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm7, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm7
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm7[1,1,3,3]
+; SSE41-NEXT:    pand %xmm5, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
+; SSE41-NEXT:    por %xmm6, %xmm4
+; SSE41-NEXT:    pxor %xmm10, %xmm4
+; SSE41-NEXT:    pxor %xmm9, %xmm0
+; SSE41-NEXT:    movdqa %xmm9, %xmm5
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm5
+; SSE41-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm6, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; SSE41-NEXT:    por %xmm0, %xmm5
+; SSE41-NEXT:    pxor %xmm10, %xmm5
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm4
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    pxor %xmm9, %xmm0
+; SSE41-NEXT:    movdqa %xmm9, %xmm6
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm6
+; SSE41-NEXT:    pshufd {{.*#+}} xmm13 = xmm6[0,0,2,2]
+; SSE41-NEXT:    pcmpeqd %xmm9, %xmm0
+; SSE41-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
+; SSE41-NEXT:    pand %xmm13, %xmm7
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
+; SSE41-NEXT:    por %xmm7, %xmm0
+; SSE41-NEXT:    movdqa %xmm0, %xmm6
+; SSE41-NEXT:    pxor %xmm10, %xmm6
+; SSE41-NEXT:    pcmpeqq %xmm5, %xmm6
+; SSE41-NEXT:    pxor %xmm10, %xmm6
+; SSE41-NEXT:    pandn %xmm6, %xmm4
+; SSE41-NEXT:    blendvpd %xmm0, %xmm12, %xmm11
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
+; SSE41-NEXT:    blendvpd %xmm0, %xmm11, %xmm3
+; SSE41-NEXT:    movapd %xmm8, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: v8i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    pushq %r15
-; AVX1-NEXT:    pushq %r14
-; AVX1-NEXT:    pushq %rbx
-; AVX1-NEXT:    vmovq %xmm3, %rax
-; AVX1-NEXT:    vmovq %xmm1, %r9
-; AVX1-NEXT:    xorl %ecx, %ecx
-; AVX1-NEXT:    movq %r9, %rdx
-; AVX1-NEXT:    subq %rax, %rdx
-; AVX1-NEXT:    setns %cl
-; AVX1-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX1-NEXT:    addq %r8, %rcx
-; AVX1-NEXT:    subq %rax, %r9
-; AVX1-NEXT:    cmovoq %rcx, %r9
-; AVX1-NEXT:    vpextrq $1, %xmm3, %rax
-; AVX1-NEXT:    vpextrq $1, %xmm1, %r10
-; AVX1-NEXT:    xorl %edx, %edx
-; AVX1-NEXT:    movq %r10, %rsi
-; AVX1-NEXT:    subq %rax, %rsi
-; AVX1-NEXT:    setns %dl
-; AVX1-NEXT:    addq %r8, %rdx
-; AVX1-NEXT:    subq %rax, %r10
-; AVX1-NEXT:    cmovoq %rdx, %r10
-; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm3
-; AVX1-NEXT:    vmovq %xmm3, %rax
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vmovq %xmm1, %r11
-; AVX1-NEXT:    xorl %edx, %edx
-; AVX1-NEXT:    movq %r11, %rdi
-; AVX1-NEXT:    subq %rax, %rdi
-; AVX1-NEXT:    setns %dl
-; AVX1-NEXT:    addq %r8, %rdx
-; AVX1-NEXT:    subq %rax, %r11
-; AVX1-NEXT:    cmovoq %rdx, %r11
-; AVX1-NEXT:    vpextrq $1, %xmm3, %rax
-; AVX1-NEXT:    vpextrq $1, %xmm1, %r14
-; AVX1-NEXT:    xorl %edx, %edx
-; AVX1-NEXT:    movq %r14, %rdi
-; AVX1-NEXT:    subq %rax, %rdi
-; AVX1-NEXT:    setns %dl
-; AVX1-NEXT:    addq %r8, %rdx
-; AVX1-NEXT:    subq %rax, %r14
-; AVX1-NEXT:    cmovoq %rdx, %r14
-; AVX1-NEXT:    vmovq %xmm2, %rax
-; AVX1-NEXT:    vmovq %xmm0, %r15
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movq %r15, %rdx
-; AVX1-NEXT:    subq %rax, %rdx
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addq %r8, %rdi
-; AVX1-NEXT:    subq %rax, %r15
-; AVX1-NEXT:    cmovoq %rdi, %r15
-; AVX1-NEXT:    vpextrq $1, %xmm2, %rdx
-; AVX1-NEXT:    vpextrq $1, %xmm0, %rax
-; AVX1-NEXT:    xorl %edi, %edi
-; AVX1-NEXT:    movq %rax, %rbx
-; AVX1-NEXT:    subq %rdx, %rbx
-; AVX1-NEXT:    setns %dil
-; AVX1-NEXT:    addq %r8, %rdi
-; AVX1-NEXT:    subq %rdx, %rax
-; AVX1-NEXT:    cmovoq %rdi, %rax
-; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm1
-; AVX1-NEXT:    vmovq %xmm1, %rdx
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vmovq %xmm0, %rcx
-; AVX1-NEXT:    xorl %ebx, %ebx
-; AVX1-NEXT:    movq %rcx, %rdi
-; AVX1-NEXT:    subq %rdx, %rdi
-; AVX1-NEXT:    setns %bl
-; AVX1-NEXT:    addq %r8, %rbx
-; AVX1-NEXT:    subq %rdx, %rcx
-; AVX1-NEXT:    cmovoq %rbx, %rcx
-; AVX1-NEXT:    vpextrq $1, %xmm1, %rdx
-; AVX1-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX1-NEXT:    xorl %ebx, %ebx
-; AVX1-NEXT:    movq %rsi, %rdi
-; AVX1-NEXT:    subq %rdx, %rdi
-; AVX1-NEXT:    setns %bl
-; AVX1-NEXT:    addq %r8, %rbx
-; AVX1-NEXT:    subq %rdx, %rsi
-; AVX1-NEXT:    cmovoq %rbx, %rsi
-; AVX1-NEXT:    vmovq %rsi, %xmm0
-; AVX1-NEXT:    vmovq %rcx, %xmm1
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX1-NEXT:    vmovq %rax, %xmm1
-; AVX1-NEXT:    vmovq %r15, %xmm2
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT:    vmovq %r14, %xmm1
-; AVX1-NEXT:    vmovq %r11, %xmm2
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX1-NEXT:    vmovq %r10, %xmm2
-; AVX1-NEXT:    vmovq %r9, %xmm3
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
-; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm2, %ymm1
-; AVX1-NEXT:    popq %rbx
-; AVX1-NEXT:    popq %r14
-; AVX1-NEXT:    popq %r15
+; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm8
+; AVX1-NEXT:    vpxor %xmm12, %xmm12, %xmm12
+; AVX1-NEXT:    vpcmpgtq %xmm8, %xmm12, %xmm7
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm9
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm6
+; AVX1-NEXT:    vpcmpgtq %xmm6, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm10
+; AVX1-NEXT:    vpcmpeqq %xmm9, %xmm10, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm9
+; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm11
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm7
+; AVX1-NEXT:    vpcmpeqq %xmm11, %xmm7, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vinsertf128 $1, %xmm9, %ymm5, %ymm9
+; AVX1-NEXT:    vpsubq %xmm8, %xmm6, %xmm8
+; AVX1-NEXT:    vpcmpgtq %xmm8, %xmm12, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm6
+; AVX1-NEXT:    vpcmpeqq %xmm6, %xmm10, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm6
+; AVX1-NEXT:    vpsubq %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm12, %xmm2
+; AVX1-NEXT:    vpxor %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqq %xmm2, %xmm7, %xmm2
+; AVX1-NEXT:    vpxor %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm2, %ymm2
+; AVX1-NEXT:    vandpd %ymm2, %ymm9, %ymm6
+; AVX1-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm7
+; AVX1-NEXT:    vmovapd {{.*#+}} ymm9 = [9223372036854775807,9223372036854775807,9223372036854775807,9223372036854775807]
+; AVX1-NEXT:    vmovapd {{.*#+}} ymm10 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
+; AVX1-NEXT:    vblendvpd %ymm7, %ymm9, %ymm10, %ymm7
+; AVX1-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm0
+; AVX1-NEXT:    vblendvpd %ymm6, %ymm7, %ymm0, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm13
+; AVX1-NEXT:    vpcmpgtq %xmm13, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm7
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm12, %xmm5
+; AVX1-NEXT:    vpxor %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vpcmpeqq %xmm7, %xmm5, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm8
+; AVX1-NEXT:    vpcmpgtq %xmm3, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm11
+; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm12, %xmm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm7, %xmm7
+; AVX1-NEXT:    vpcmpeqq %xmm11, %xmm7, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm6
+; AVX1-NEXT:    vinsertf128 $1, %xmm8, %ymm6, %ymm8
+; AVX1-NEXT:    vpsubq %xmm13, %xmm2, %xmm11
+; AVX1-NEXT:    vpcmpgtq %xmm11, %xmm12, %xmm6
+; AVX1-NEXT:    vpxor %xmm4, %xmm6, %xmm2
+; AVX1-NEXT:    vpcmpeqq %xmm2, %xmm5, %xmm2
+; AVX1-NEXT:    vpxor %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vpsubq %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm12, %xmm3
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm7, %xmm3
+; AVX1-NEXT:    vpxor %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
+; AVX1-NEXT:    vandpd %ymm2, %ymm8, %ymm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm1, %ymm3
+; AVX1-NEXT:    vblendvpd %ymm3, %ymm9, %ymm10, %ymm3
+; AVX1-NEXT:    vinsertf128 $1, %xmm11, %ymm1, %ymm1
+; AVX1-NEXT:    vblendvpd %ymm2, %ymm3, %ymm1, %ymm1
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: v8i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    pushq %r15
-; AVX2-NEXT:    pushq %r14
-; AVX2-NEXT:    pushq %rbx
-; AVX2-NEXT:    vmovq %xmm3, %rax
-; AVX2-NEXT:    vmovq %xmm1, %r9
-; AVX2-NEXT:    xorl %ecx, %ecx
-; AVX2-NEXT:    movq %r9, %rdx
-; AVX2-NEXT:    subq %rax, %rdx
-; AVX2-NEXT:    setns %cl
-; AVX2-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX2-NEXT:    addq %r8, %rcx
-; AVX2-NEXT:    subq %rax, %r9
-; AVX2-NEXT:    cmovoq %rcx, %r9
-; AVX2-NEXT:    vpextrq $1, %xmm3, %rax
-; AVX2-NEXT:    vpextrq $1, %xmm1, %r10
-; AVX2-NEXT:    xorl %edx, %edx
-; AVX2-NEXT:    movq %r10, %rsi
-; AVX2-NEXT:    subq %rax, %rsi
-; AVX2-NEXT:    setns %dl
-; AVX2-NEXT:    addq %r8, %rdx
-; AVX2-NEXT:    subq %rax, %r10
-; AVX2-NEXT:    cmovoq %rdx, %r10
-; AVX2-NEXT:    vextracti128 $1, %ymm3, %xmm3
-; AVX2-NEXT:    vmovq %xmm3, %rax
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; AVX2-NEXT:    vmovq %xmm1, %r11
-; AVX2-NEXT:    xorl %edx, %edx
-; AVX2-NEXT:    movq %r11, %rdi
-; AVX2-NEXT:    subq %rax, %rdi
-; AVX2-NEXT:    setns %dl
-; AVX2-NEXT:    addq %r8, %rdx
-; AVX2-NEXT:    subq %rax, %r11
-; AVX2-NEXT:    cmovoq %rdx, %r11
-; AVX2-NEXT:    vpextrq $1, %xmm3, %rax
-; AVX2-NEXT:    vpextrq $1, %xmm1, %r14
-; AVX2-NEXT:    xorl %edx, %edx
-; AVX2-NEXT:    movq %r14, %rdi
-; AVX2-NEXT:    subq %rax, %rdi
-; AVX2-NEXT:    setns %dl
-; AVX2-NEXT:    addq %r8, %rdx
-; AVX2-NEXT:    subq %rax, %r14
-; AVX2-NEXT:    cmovoq %rdx, %r14
-; AVX2-NEXT:    vmovq %xmm2, %rax
-; AVX2-NEXT:    vmovq %xmm0, %r15
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movq %r15, %rdx
-; AVX2-NEXT:    subq %rax, %rdx
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addq %r8, %rdi
-; AVX2-NEXT:    subq %rax, %r15
-; AVX2-NEXT:    cmovoq %rdi, %r15
-; AVX2-NEXT:    vpextrq $1, %xmm2, %rdx
-; AVX2-NEXT:    vpextrq $1, %xmm0, %rax
-; AVX2-NEXT:    xorl %edi, %edi
-; AVX2-NEXT:    movq %rax, %rbx
-; AVX2-NEXT:    subq %rdx, %rbx
-; AVX2-NEXT:    setns %dil
-; AVX2-NEXT:    addq %r8, %rdi
-; AVX2-NEXT:    subq %rdx, %rax
-; AVX2-NEXT:    cmovoq %rdi, %rax
-; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm1
-; AVX2-NEXT:    vmovq %xmm1, %rdx
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; AVX2-NEXT:    vmovq %xmm0, %rcx
-; AVX2-NEXT:    xorl %ebx, %ebx
-; AVX2-NEXT:    movq %rcx, %rdi
-; AVX2-NEXT:    subq %rdx, %rdi
-; AVX2-NEXT:    setns %bl
-; AVX2-NEXT:    addq %r8, %rbx
-; AVX2-NEXT:    subq %rdx, %rcx
-; AVX2-NEXT:    cmovoq %rbx, %rcx
-; AVX2-NEXT:    vpextrq $1, %xmm1, %rdx
-; AVX2-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX2-NEXT:    xorl %ebx, %ebx
-; AVX2-NEXT:    movq %rsi, %rdi
-; AVX2-NEXT:    subq %rdx, %rdi
-; AVX2-NEXT:    setns %bl
-; AVX2-NEXT:    addq %r8, %rbx
-; AVX2-NEXT:    subq %rdx, %rsi
-; AVX2-NEXT:    cmovoq %rbx, %rsi
-; AVX2-NEXT:    vmovq %rsi, %xmm0
-; AVX2-NEXT:    vmovq %rcx, %xmm1
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX2-NEXT:    vmovq %rax, %xmm1
-; AVX2-NEXT:    vmovq %r15, %xmm2
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX2-NEXT:    vmovq %r14, %xmm1
-; AVX2-NEXT:    vmovq %r11, %xmm2
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX2-NEXT:    vmovq %r10, %xmm2
-; AVX2-NEXT:    vmovq %r9, %xmm3
-; AVX2-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
-; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
-; AVX2-NEXT:    popq %rbx
-; AVX2-NEXT:    popq %r14
-; AVX2-NEXT:    popq %r15
+; AVX2-NEXT:    vpxor %xmm4, %xmm4, %xmm4
+; AVX2-NEXT:    vpcmpgtq %ymm2, %ymm4, %ymm5
+; AVX2-NEXT:    vpcmpeqd %ymm6, %ymm6, %ymm6
+; AVX2-NEXT:    vpxor %ymm6, %ymm5, %ymm5
+; AVX2-NEXT:    vpcmpgtq %ymm0, %ymm4, %ymm7
+; AVX2-NEXT:    vpxor %ymm6, %ymm7, %ymm7
+; AVX2-NEXT:    vpcmpeqq %ymm5, %ymm7, %ymm5
+; AVX2-NEXT:    vpsubq %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpgtq %ymm0, %ymm4, %ymm2
+; AVX2-NEXT:    vpxor %ymm6, %ymm2, %ymm2
+; AVX2-NEXT:    vpcmpeqq %ymm2, %ymm7, %ymm2
+; AVX2-NEXT:    vpxor %ymm6, %ymm2, %ymm2
+; AVX2-NEXT:    vpandn %ymm2, %ymm5, %ymm2
+; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm5 = [9223372036854775807,9223372036854775807,9223372036854775807,9223372036854775807]
+; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm7 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
+; AVX2-NEXT:    vblendvpd %ymm0, %ymm5, %ymm7, %ymm8
+; AVX2-NEXT:    vblendvpd %ymm2, %ymm8, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpgtq %ymm3, %ymm4, %ymm2
+; AVX2-NEXT:    vpxor %ymm6, %ymm2, %ymm2
+; AVX2-NEXT:    vpcmpgtq %ymm1, %ymm4, %ymm8
+; AVX2-NEXT:    vpxor %ymm6, %ymm8, %ymm8
+; AVX2-NEXT:    vpcmpeqq %ymm2, %ymm8, %ymm2
+; AVX2-NEXT:    vpsubq %ymm3, %ymm1, %ymm1
+; AVX2-NEXT:    vpcmpgtq %ymm1, %ymm4, %ymm3
+; AVX2-NEXT:    vpxor %ymm6, %ymm3, %ymm3
+; AVX2-NEXT:    vpcmpeqq %ymm3, %ymm8, %ymm3
+; AVX2-NEXT:    vpxor %ymm6, %ymm3, %ymm3
+; AVX2-NEXT:    vpandn %ymm3, %ymm2, %ymm2
+; AVX2-NEXT:    vblendvpd %ymm1, %ymm5, %ymm7, %ymm3
+; AVX2-NEXT:    vblendvpd %ymm2, %ymm3, %ymm1, %ymm1
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: v8i64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    pushq %r15
-; AVX512-NEXT:    pushq %r14
-; AVX512-NEXT:    pushq %rbx
-; AVX512-NEXT:    vmovq %xmm1, %rax
-; AVX512-NEXT:    vmovq %xmm0, %r9
-; AVX512-NEXT:    xorl %ecx, %ecx
-; AVX512-NEXT:    movq %r9, %rdx
-; AVX512-NEXT:    subq %rax, %rdx
-; AVX512-NEXT:    setns %cl
-; AVX512-NEXT:    movabsq $9223372036854775807, %r8 # imm = 0x7FFFFFFFFFFFFFFF
-; AVX512-NEXT:    addq %r8, %rcx
-; AVX512-NEXT:    subq %rax, %r9
-; AVX512-NEXT:    cmovoq %rcx, %r9
-; AVX512-NEXT:    vpextrq $1, %xmm1, %rax
-; AVX512-NEXT:    vpextrq $1, %xmm0, %r10
-; AVX512-NEXT:    xorl %edx, %edx
-; AVX512-NEXT:    movq %r10, %rsi
-; AVX512-NEXT:    subq %rax, %rsi
-; AVX512-NEXT:    setns %dl
-; AVX512-NEXT:    addq %r8, %rdx
-; AVX512-NEXT:    subq %rax, %r10
-; AVX512-NEXT:    cmovoq %rdx, %r10
-; AVX512-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX512-NEXT:    vmovq %xmm2, %rax
-; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm3
-; AVX512-NEXT:    vmovq %xmm3, %r11
-; AVX512-NEXT:    xorl %edx, %edx
-; AVX512-NEXT:    movq %r11, %rdi
-; AVX512-NEXT:    subq %rax, %rdi
-; AVX512-NEXT:    setns %dl
-; AVX512-NEXT:    addq %r8, %rdx
-; AVX512-NEXT:    subq %rax, %r11
-; AVX512-NEXT:    cmovoq %rdx, %r11
-; AVX512-NEXT:    vpextrq $1, %xmm2, %rax
-; AVX512-NEXT:    vpextrq $1, %xmm3, %r14
-; AVX512-NEXT:    xorl %edx, %edx
-; AVX512-NEXT:    movq %r14, %rdi
-; AVX512-NEXT:    subq %rax, %rdi
-; AVX512-NEXT:    setns %dl
-; AVX512-NEXT:    addq %r8, %rdx
-; AVX512-NEXT:    subq %rax, %r14
-; AVX512-NEXT:    cmovoq %rdx, %r14
-; AVX512-NEXT:    vextracti32x4 $2, %zmm1, %xmm2
-; AVX512-NEXT:    vmovq %xmm2, %rax
-; AVX512-NEXT:    vextracti32x4 $2, %zmm0, %xmm3
-; AVX512-NEXT:    vmovq %xmm3, %r15
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movq %r15, %rdx
-; AVX512-NEXT:    subq %rax, %rdx
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addq %r8, %rdi
-; AVX512-NEXT:    subq %rax, %r15
-; AVX512-NEXT:    cmovoq %rdi, %r15
-; AVX512-NEXT:    vpextrq $1, %xmm2, %rdx
-; AVX512-NEXT:    vpextrq $1, %xmm3, %rax
-; AVX512-NEXT:    xorl %edi, %edi
-; AVX512-NEXT:    movq %rax, %rbx
-; AVX512-NEXT:    subq %rdx, %rbx
-; AVX512-NEXT:    setns %dil
-; AVX512-NEXT:    addq %r8, %rdi
-; AVX512-NEXT:    subq %rdx, %rax
-; AVX512-NEXT:    cmovoq %rdi, %rax
-; AVX512-NEXT:    vextracti32x4 $3, %zmm1, %xmm1
-; AVX512-NEXT:    vmovq %xmm1, %rdx
-; AVX512-NEXT:    vextracti32x4 $3, %zmm0, %xmm0
-; AVX512-NEXT:    vmovq %xmm0, %rcx
-; AVX512-NEXT:    xorl %ebx, %ebx
-; AVX512-NEXT:    movq %rcx, %rdi
-; AVX512-NEXT:    subq %rdx, %rdi
-; AVX512-NEXT:    setns %bl
-; AVX512-NEXT:    addq %r8, %rbx
-; AVX512-NEXT:    subq %rdx, %rcx
-; AVX512-NEXT:    cmovoq %rbx, %rcx
-; AVX512-NEXT:    vpextrq $1, %xmm1, %rdx
-; AVX512-NEXT:    vpextrq $1, %xmm0, %rsi
-; AVX512-NEXT:    xorl %ebx, %ebx
-; AVX512-NEXT:    movq %rsi, %rdi
-; AVX512-NEXT:    subq %rdx, %rdi
-; AVX512-NEXT:    setns %bl
-; AVX512-NEXT:    addq %r8, %rbx
-; AVX512-NEXT:    subq %rdx, %rsi
-; AVX512-NEXT:    cmovoq %rbx, %rsi
-; AVX512-NEXT:    vmovq %rsi, %xmm0
-; AVX512-NEXT:    vmovq %rcx, %xmm1
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; AVX512-NEXT:    vmovq %rax, %xmm1
-; AVX512-NEXT:    vmovq %r15, %xmm2
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX512-NEXT:    vmovq %r14, %xmm1
-; AVX512-NEXT:    vmovq %r11, %xmm2
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; AVX512-NEXT:    vmovq %r10, %xmm2
-; AVX512-NEXT:    vmovq %r9, %xmm3
-; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
-; AVX512-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
-; AVX512-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512-NEXT:    popq %rbx
-; AVX512-NEXT:    popq %r14
-; AVX512-NEXT:    popq %r15
+; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512-NEXT:    vpcmpnltq %zmm2, %zmm1, %k0
+; AVX512-NEXT:    vpcmpnltq %zmm2, %zmm0, %k1
+; AVX512-NEXT:    kxorw %k0, %k1, %k0
+; AVX512-NEXT:    vpsubq %zmm1, %zmm0, %zmm0
+; AVX512-NEXT:    vpcmpnltq %zmm2, %zmm0, %k2
+; AVX512-NEXT:    kxorw %k2, %k1, %k1
+; AVX512-NEXT:    kandw %k1, %k0, %k1
+; AVX512-NEXT:    vpcmpgtq %zmm0, %zmm2, %k2
+; AVX512-NEXT:    vpbroadcastq {{.*#+}} zmm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
+; AVX512-NEXT:    vpbroadcastq {{.*}}(%rip), %zmm1 {%k2}
+; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k1}
 ; AVX512-NEXT:    retq
   %z = call <8 x i64> @llvm.ssub.sat.v8i64(<8 x i64> %x, <8 x i64> %y)
   ret <8 x i64> %z