diff --git a/i/llvm/include/llvm/CodeGen/TargetLowering.h b/llvm/include/llvm/CodeGen/TargetLowering.h
--- a/i/llvm/include/llvm/CodeGen/TargetLowering.h
+++ b/llvm/include/llvm/CodeGen/TargetLowering.h
@@ -605,6 +605,12 @@
     return false;
   }
 
+  /// Return the maximum number of "x & (x - 1)" operations that can be done
+  /// instead of deferring to a custom CTPOP.
+  virtual unsigned getCustomCtpopCost(EVT VT, ISD::CondCode Cond) const {
+    return 1;
+  }
+
   /// Return true if instruction generated for equality comparison is folded
   /// with instruction generated for signed comparison.
   virtual bool isEqualityCmpFoldedWithSignedCmp() const { return true; }
diff --git a/i/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
--- a/i/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -3396,21 +3396,34 @@
   EVT CTVT = CTPOP.getValueType();
   SDValue CTOp = CTPOP.getOperand(0);
 
+  // If this is a vector CTPOP, keep the CTPOP if it is legal.
+  // TODO: Should we check if CTPOP is legal(or custom) for scalars?
+  if (VT.isVector() && TLI.isOperationLegal(ISD::CTPOP, CTVT))
+    return SDValue();
+
+
   // (ctpop x) u< 2 -> (x & x-1) == 0
   // (ctpop x) u> 1 -> (x & x-1) != 0
-  if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)) {
-    // If this is a vector CTPOP, keep the CTPOP if it is legal.
+  if (Cond == ISD::SETULT || Cond == ISD::SETUGT) {
     // This based on X86's custom lowering for vector CTPOP which produces more
     // instructions than the expansion here.
-    // TODO: Should we check if CTPOP is legal(or custom) for scalars?
-    if (VT.isVector() && TLI.isOperationLegal(ISD::CTPOP, CTVT))
+
+    unsigned CostLimit = TLI.getCustomCtpopCost(CTVT, Cond);
+    if (C1.ugt(CostLimit + (Cond == ISD::SETULT)))
       return SDValue();
+    if (C1 == 0 && (Cond == ISD::SETULT))
+      return SDValue(); // This is handled elsewhere.
+
+    unsigned Passes = C1.getLimitedValue() - (Cond == ISD::SETULT);
 
     SDValue NegOne = DAG.getAllOnesConstant(dl, CTVT);
-    SDValue Add = DAG.getNode(ISD::ADD, dl, CTVT, CTOp, NegOne);
-    SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Add);
+    SDValue Result = CTOp;
+    for (unsigned i = 0; i < Passes; i++) {
+      SDValue Add = DAG.getNode(ISD::ADD, dl, CTVT, Result, NegOne);
+      Result = DAG.getNode(ISD::AND, dl, CTVT, Result, Add);
+    }
     ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE;
-    return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, dl, CTVT), CC);
+    return DAG.getSetCC(dl, VT, Result, DAG.getConstant(0, dl, CTVT), CC);
   }
 
   // If ctpop is not supported, expand a power-of-2 comparison based on it.
@@ -3418,11 +3431,8 @@
     // For scalars, keep CTPOP if it is legal or custom.
     if (!VT.isVector() && TLI.isOperationLegalOrCustom(ISD::CTPOP, CTVT))
       return SDValue();
-    // For vectors, keep CTPOP only if it is legal.
     // This is based on X86's custom lowering for CTPOP which produces more
     // instructions than the expansion here.
-    if (VT.isVector() && TLI.isOperationLegal(ISD::CTPOP, CTVT))
-      return SDValue();
 
     // (ctpop x) == 1 --> (x != 0) && ((x & x-1) == 0)
     // (ctpop x) != 1 --> (x == 0) || ((x & x-1) != 0)
diff --git a/i/llvm/lib/Target/X86/X86ISelLowering.h b/llvm/lib/Target/X86/X86ISelLowering.h
--- a/i/llvm/lib/Target/X86/X86ISelLowering.h
+++ b/llvm/lib/Target/X86/X86ISelLowering.h
@@ -983,6 +983,8 @@
 
     bool isCtlzFast() const override;
 
+    unsigned getCustomCtpopCost(EVT VT, ISD::CondCode Cond) const override;
+
     bool hasBitPreservingFPLogic(EVT VT) const override {
       return VT == MVT::f32 || VT == MVT::f64 || VT.isVector();
     }
diff --git a/i/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
--- a/i/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -5331,6 +5331,23 @@
   return Subtarget.hasFastLZCNT();
 }
 
+unsigned X86TargetLowering::getCustomCtpopCost(EVT VT, ISD::CondCode Cond) const {
+  if (!VT.isVector())
+    return Subtarget.hasPOPCNT() ? 1 : 4;
+
+  EVT EltVT = VT.getVectorElementType();
+
+  // Horizontal add makes this easy.
+  if (Subtarget.hasBITALG())
+    return EltVT == MVT::i64 ? 1 : 3;
+  if (Subtarget.hasVPOPCNTDQ())
+    return Subtarget.hasBWI() ? 4 : 1;
+
+  if (VT.getSizeInBits() == 512 && (EltVT == MVT::i64 || EltVT == MVT::i32))
+    return 8;
+  return 4;
+}
+
 bool X86TargetLowering::isMaskAndCmp0FoldingBeneficial(
     const Instruction &AndI) const {
   return true;
diff --git a/i/llvm/test/CodeGen/X86/vector-popcnt-128-ult-ugt.ll b/llvm/test/CodeGen/X86/vector-popcnt-128-ult-ugt.ll
--- a/i/llvm/test/CodeGen/X86/vector-popcnt-128-ult-ugt.ll
+++ b/llvm/test/CodeGen/X86/vector-popcnt-128-ult-ugt.ll
@@ -38,98 +38,16 @@
 }
 
 define <16 x i8> @ugt_0_v16i8(<16 x i8> %0) {
-; SSE2-LABEL: ugt_0_v16i8:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    pcmpeqb %xmm0, %xmm1
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSE2-NEXT:    pxor %xmm0, %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ugt_0_v16i8:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    pcmpeqb %xmm0, %xmm1
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSE3-NEXT:    pxor %xmm0, %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ugt_0_v16i8:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm1, %xmm1
-; SSSE3-NEXT:    pcmpeqb %xmm3, %xmm1
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSSE3-NEXT:    pxor %xmm1, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ugt_0_v16i8:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm1, %xmm1
-; SSE41-NEXT:    pcmpeqb %xmm3, %xmm1
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: ugt_0_v16i8:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqb %xmm1, %xmm0
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_0_v16i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
@@ -138,14 +56,6 @@
 ;
 ; AVX2-LABEL: ugt_0_v16i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
@@ -154,9 +64,6 @@
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_0_v16i8:
 ; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
-; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpmovdb %zmm0, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
@@ -166,13 +73,9 @@
 ;
 ; AVX512VPOPCNTDQVL-LABEL: ugt_0_v16i8:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpmovdb %zmm0, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
 ; BITALG_NOVLX-LABEL: ugt_0_v16i8:
@@ -200,127 +103,17 @@
 }
 
 define <16 x i8> @ult_1_v16i8(<16 x i8> %0) {
-; SSE2-LABEL: ult_1_v16i8:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    pcmpeqb %xmm0, %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ult_1_v16i8:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    pcmpeqb %xmm0, %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ult_1_v16i8:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ult_1_v16i8:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSE41-NEXT:    retq
-;
-; AVX1-LABEL: ult_1_v16i8:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: ult_1_v16i8:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    retq
-;
-; AVX512VPOPCNTDQ-LABEL: ult_1_v16i8:
-; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
-; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpmovdb %zmm0, %xmm0
-; AVX512VPOPCNTDQ-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512VPOPCNTDQ-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
-; AVX512VPOPCNTDQ-NEXT:    vzeroupper
-; AVX512VPOPCNTDQ-NEXT:    retq
+; SSE-LABEL: ult_1_v16i8:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqb %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_1_v16i8:
-; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpmovdb %zmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
-; AVX512VPOPCNTDQVL-NEXT:    retq
+; AVX-LABEL: ult_1_v16i8:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
 ;
 ; BITALG_NOVLX-LABEL: ult_1_v16i8:
 ; BITALG_NOVLX:       # %bb.0:
@@ -459,106 +252,42 @@
 }
 
 define <16 x i8> @ugt_2_v16i8(<16 x i8> %0) {
-; SSE2-LABEL: ugt_2_v16i8:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pcmpgtb {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ugt_2_v16i8:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pcmpgtb {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ugt_2_v16i8:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; SSSE3-NEXT:    pmaxub %xmm3, %xmm0
-; SSSE3-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ugt_2_v16i8:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; SSE41-NEXT:    pmaxub %xmm3, %xmm0
-; SSE41-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: ugt_2_v16i8:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddb %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm0
+; SSE-NEXT:    paddb %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm2, %xmm2
+; SSE-NEXT:    pcmpeqb %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_2_v16i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpmaxub {{.*}}(%rip), %xmm0, %xmm1
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqb %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_2_v16i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpmaxub {{.*}}(%rip), %xmm0, %xmm1
-; AVX2-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqb %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_2_v16i8:
@@ -600,105 +329,38 @@
 }
 
 define <16 x i8> @ult_3_v16i8(<16 x i8> %0) {
-; SSE2-LABEL: ult_3_v16i8:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; SSE2-NEXT:    pcmpgtb %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ult_3_v16i8:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; SSE3-NEXT:    pcmpgtb %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ult_3_v16i8:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2]
-; SSSE3-NEXT:    pminub %xmm3, %xmm0
-; SSSE3-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ult_3_v16i8:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2]
-; SSE41-NEXT:    pminub %xmm3, %xmm0
-; SSE41-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: ult_3_v16i8:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm2, %xmm2
+; SSE-NEXT:    movdqa %xmm0, %xmm1
+; SSE-NEXT:    paddb %xmm2, %xmm1
+; SSE-NEXT:    pand %xmm0, %xmm1
+; SSE-NEXT:    paddb %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm2, %xmm1
+; SSE-NEXT:    pxor %xmm0, %xmm0
+; SSE-NEXT:    pcmpeqb %xmm0, %xmm1
+; SSE-NEXT:    movdqa %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_3_v16i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm1
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpminub {{.*}}(%rip), %xmm0, %xmm1
+; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_3_v16i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm1
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpminub {{.*}}(%rip), %xmm0, %xmm1
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
@@ -744,106 +406,49 @@
 }
 
 define <16 x i8> @ugt_3_v16i8(<16 x i8> %0) {
-; SSE2-LABEL: ugt_3_v16i8:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pcmpgtb {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ugt_3_v16i8:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pcmpgtb {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ugt_3_v16i8:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; SSSE3-NEXT:    pmaxub %xmm3, %xmm0
-; SSSE3-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ugt_3_v16i8:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; SSE41-NEXT:    pmaxub %xmm3, %xmm0
-; SSE41-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: ugt_3_v16i8:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddb %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm3
+; SSE-NEXT:    paddb %xmm1, %xmm3
+; SSE-NEXT:    pand %xmm2, %xmm3
+; SSE-NEXT:    movdqa %xmm3, %xmm0
+; SSE-NEXT:    paddb %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm3, %xmm0
+; SSE-NEXT:    pxor %xmm2, %xmm2
+; SSE-NEXT:    pcmpeqb %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_3_v16i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpmaxub {{.*}}(%rip), %xmm0, %xmm1
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqb %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_3_v16i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpmaxub {{.*}}(%rip), %xmm0, %xmm1
-; AVX2-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqb %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_3_v16i8:
@@ -885,105 +490,44 @@
 }
 
 define <16 x i8> @ult_4_v16i8(<16 x i8> %0) {
-; SSE2-LABEL: ult_4_v16i8:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; SSE2-NEXT:    pcmpgtb %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ult_4_v16i8:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; SSE3-NEXT:    pcmpgtb %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ult_4_v16i8:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; SSSE3-NEXT:    pminub %xmm3, %xmm0
-; SSSE3-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ult_4_v16i8:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; SSE41-NEXT:    pminub %xmm3, %xmm0
-; SSE41-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: ult_4_v16i8:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddb %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm0
+; SSE-NEXT:    paddb %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    paddb %xmm0, %xmm1
+; SSE-NEXT:    pand %xmm1, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqb %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_4_v16i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm1
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpminub {{.*}}(%rip), %xmm0, %xmm1
+; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_4_v16i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm1
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpminub {{.*}}(%rip), %xmm0, %xmm1
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
@@ -1029,106 +573,56 @@
 }
 
 define <16 x i8> @ugt_4_v16i8(<16 x i8> %0) {
-; SSE2-LABEL: ugt_4_v16i8:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pcmpgtb {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ugt_4_v16i8:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pcmpgtb {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ugt_4_v16i8:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5]
-; SSSE3-NEXT:    pmaxub %xmm3, %xmm0
-; SSSE3-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ugt_4_v16i8:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5]
-; SSE41-NEXT:    pmaxub %xmm3, %xmm0
-; SSE41-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: ugt_4_v16i8:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddb %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm0
+; SSE-NEXT:    paddb %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddb %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm0
+; SSE-NEXT:    paddb %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm2, %xmm2
+; SSE-NEXT:    pcmpeqb %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_4_v16i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpmaxub {{.*}}(%rip), %xmm0, %xmm1
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqb %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_4_v16i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpmaxub {{.*}}(%rip), %xmm0, %xmm1
-; AVX2-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqb %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_4_v16i8:
@@ -1170,105 +664,51 @@
 }
 
 define <16 x i8> @ult_5_v16i8(<16 x i8> %0) {
-; SSE2-LABEL: ult_5_v16i8:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5]
-; SSE2-NEXT:    pcmpgtb %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ult_5_v16i8:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5]
-; SSE3-NEXT:    pcmpgtb %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ult_5_v16i8:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; SSSE3-NEXT:    pminub %xmm3, %xmm0
-; SSSE3-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ult_5_v16i8:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; SSE41-NEXT:    pminub %xmm3, %xmm0
-; SSE41-NEXT:    pcmpeqb %xmm3, %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: ult_5_v16i8:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddb %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm3
+; SSE-NEXT:    paddb %xmm1, %xmm3
+; SSE-NEXT:    pand %xmm2, %xmm3
+; SSE-NEXT:    movdqa %xmm3, %xmm0
+; SSE-NEXT:    paddb %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm3, %xmm0
+; SSE-NEXT:    paddb %xmm0, %xmm1
+; SSE-NEXT:    pand %xmm1, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqb %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_5_v16i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm1
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpminub {{.*}}(%rip), %xmm0, %xmm1
+; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_5_v16i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddb %xmm1, %xmm0, %xmm1
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpminub {{.*}}(%rip), %xmm0, %xmm1
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
@@ -2480,115 +1920,16 @@
 }
 
 define <8 x i16> @ugt_0_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_0_v8i16:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    psllw $8, %xmm0
-; SSE2-NEXT:    paddb %xmm1, %xmm0
-; SSE2-NEXT:    psrlw $8, %xmm0
-; SSE2-NEXT:    pxor %xmm1, %xmm1
-; SSE2-NEXT:    pcmpeqw %xmm1, %xmm0
-; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
-; SSE2-NEXT:    pxor %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ugt_0_v8i16:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    psllw $8, %xmm0
-; SSE3-NEXT:    paddb %xmm1, %xmm0
-; SSE3-NEXT:    psrlw $8, %xmm0
-; SSE3-NEXT:    pxor %xmm1, %xmm1
-; SSE3-NEXT:    pcmpeqw %xmm1, %xmm0
-; SSE3-NEXT:    pcmpeqd %xmm1, %xmm1
-; SSE3-NEXT:    pxor %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ugt_0_v8i16:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa %xmm3, %xmm0
-; SSSE3-NEXT:    psllw $8, %xmm0
-; SSSE3-NEXT:    paddb %xmm3, %xmm0
-; SSSE3-NEXT:    psrlw $8, %xmm0
-; SSSE3-NEXT:    pxor %xmm1, %xmm1
-; SSSE3-NEXT:    pcmpeqw %xmm0, %xmm1
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSSE3-NEXT:    pxor %xmm1, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ugt_0_v8i16:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa %xmm3, %xmm0
-; SSE41-NEXT:    psllw $8, %xmm0
-; SSE41-NEXT:    paddb %xmm3, %xmm0
-; SSE41-NEXT:    psrlw $8, %xmm0
-; SSE41-NEXT:    pxor %xmm1, %xmm1
-; SSE41-NEXT:    pcmpeqw %xmm0, %xmm1
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: ugt_0_v8i16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqw %xmm1, %xmm0
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_0_v8i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX1-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
@@ -2597,17 +1938,6 @@
 ;
 ; AVX2-LABEL: ugt_0_v8i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
@@ -2616,9 +1946,6 @@
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_0_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
 ; AVX512VPOPCNTDQ-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
@@ -2628,13 +1955,9 @@
 ;
 ; AVX512VPOPCNTDQVL-LABEL: ugt_0_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
-; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
 ; BITALG_NOVLX-LABEL: ugt_0_v8i16:
@@ -2662,147 +1985,17 @@
 }
 
 define <8 x i16> @ult_1_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_1_v8i16:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    psllw $8, %xmm0
-; SSE2-NEXT:    paddb %xmm1, %xmm0
-; SSE2-NEXT:    psrlw $8, %xmm0
-; SSE2-NEXT:    pxor %xmm1, %xmm1
-; SSE2-NEXT:    pcmpeqw %xmm1, %xmm0
-; SSE2-NEXT:    retq
+; SSE-LABEL: ult_1_v8i16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqw %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
-; SSE3-LABEL: ult_1_v8i16:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    psllw $8, %xmm0
-; SSE3-NEXT:    paddb %xmm1, %xmm0
-; SSE3-NEXT:    psrlw $8, %xmm0
-; SSE3-NEXT:    pxor %xmm1, %xmm1
-; SSE3-NEXT:    pcmpeqw %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ult_1_v8i16:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa %xmm3, %xmm1
-; SSSE3-NEXT:    psllw $8, %xmm1
-; SSSE3-NEXT:    paddb %xmm3, %xmm1
-; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    pcmpeqw %xmm1, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ult_1_v8i16:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa %xmm3, %xmm1
-; SSE41-NEXT:    psllw $8, %xmm1
-; SSE41-NEXT:    paddb %xmm3, %xmm1
-; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    pcmpeqw %xmm1, %xmm0
-; SSE41-NEXT:    retq
-;
-; AVX1-LABEL: ult_1_v8i16:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: ult_1_v8i16:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    retq
-;
-; AVX512VPOPCNTDQ-LABEL: ult_1_v8i16:
-; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512VPOPCNTDQ-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
-; AVX512VPOPCNTDQ-NEXT:    vzeroupper
-; AVX512VPOPCNTDQ-NEXT:    retq
-;
-; AVX512VPOPCNTDQVL-LABEL: ult_1_v8i16:
-; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
-; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
-; AVX512VPOPCNTDQVL-NEXT:    retq
+; AVX-LABEL: ult_1_v8i16:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
 ;
 ; BITALG_NOVLX-LABEL: ult_1_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
@@ -2941,120 +2134,42 @@
 }
 
 define <8 x i16> @ugt_2_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_2_v8i16:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    psllw $8, %xmm0
-; SSE2-NEXT:    paddb %xmm1, %xmm0
-; SSE2-NEXT:    psrlw $8, %xmm0
-; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ugt_2_v8i16:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    psllw $8, %xmm0
-; SSE3-NEXT:    paddb %xmm1, %xmm0
-; SSE3-NEXT:    psrlw $8, %xmm0
-; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ugt_2_v8i16:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa %xmm3, %xmm0
-; SSSE3-NEXT:    psllw $8, %xmm0
-; SSSE3-NEXT:    paddb %xmm3, %xmm0
-; SSSE3-NEXT:    psrlw $8, %xmm0
-; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ugt_2_v8i16:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa %xmm3, %xmm0
-; SSE41-NEXT:    psllw $8, %xmm0
-; SSE41-NEXT:    paddb %xmm3, %xmm0
-; SSE41-NEXT:    psrlw $8, %xmm0
-; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: ugt_2_v8i16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddw %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm0
+; SSE-NEXT:    paddw %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm2, %xmm2
+; SSE-NEXT:    pcmpeqw %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_2_v8i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_2_v8i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqw %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_2_v8i16:
@@ -3096,126 +2211,39 @@
 }
 
 define <8 x i16> @ult_3_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_3_v8i16:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    psllw $8, %xmm2
-; SSE2-NEXT:    paddb %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3,3,3,3,3]
-; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ult_3_v8i16:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    psllw $8, %xmm2
-; SSE3-NEXT:    paddb %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3,3,3,3,3]
-; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ult_3_v8i16:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa %xmm3, %xmm1
-; SSSE3-NEXT:    psllw $8, %xmm1
-; SSSE3-NEXT:    paddb %xmm3, %xmm1
-; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3,3,3,3,3]
-; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ult_3_v8i16:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa %xmm3, %xmm1
-; SSE41-NEXT:    psllw $8, %xmm1
-; SSE41-NEXT:    paddb %xmm3, %xmm1
-; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3,3,3,3,3]
-; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: ult_3_v8i16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm2, %xmm2
+; SSE-NEXT:    movdqa %xmm0, %xmm1
+; SSE-NEXT:    paddw %xmm2, %xmm1
+; SSE-NEXT:    pand %xmm0, %xmm1
+; SSE-NEXT:    paddw %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm2, %xmm1
+; SSE-NEXT:    pxor %xmm0, %xmm0
+; SSE-NEXT:    pcmpeqw %xmm0, %xmm1
+; SSE-NEXT:    movdqa %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_3_v8i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm1
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3,3,3,3,3,3,3]
-; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_3_v8i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm1
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3,3,3,3,3,3,3]
-; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_3_v8i16:
@@ -3260,120 +2288,49 @@
 }
 
 define <8 x i16> @ugt_3_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_3_v8i16:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    psllw $8, %xmm0
-; SSE2-NEXT:    paddb %xmm1, %xmm0
-; SSE2-NEXT:    psrlw $8, %xmm0
-; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ugt_3_v8i16:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    psllw $8, %xmm0
-; SSE3-NEXT:    paddb %xmm1, %xmm0
-; SSE3-NEXT:    psrlw $8, %xmm0
-; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ugt_3_v8i16:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa %xmm3, %xmm0
-; SSSE3-NEXT:    psllw $8, %xmm0
-; SSSE3-NEXT:    paddb %xmm3, %xmm0
-; SSSE3-NEXT:    psrlw $8, %xmm0
-; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ugt_3_v8i16:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa %xmm3, %xmm0
-; SSE41-NEXT:    psllw $8, %xmm0
-; SSE41-NEXT:    paddb %xmm3, %xmm0
-; SSE41-NEXT:    psrlw $8, %xmm0
-; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: ugt_3_v8i16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddw %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm3
+; SSE-NEXT:    paddw %xmm1, %xmm3
+; SSE-NEXT:    pand %xmm2, %xmm3
+; SSE-NEXT:    movdqa %xmm3, %xmm0
+; SSE-NEXT:    paddw %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm3, %xmm0
+; SSE-NEXT:    pxor %xmm2, %xmm2
+; SSE-NEXT:    pcmpeqw %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_3_v8i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_3_v8i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqw %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_3_v8i16:
@@ -3415,126 +2372,45 @@
 }
 
 define <8 x i16> @ult_4_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_4_v8i16:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    psllw $8, %xmm2
-; SSE2-NEXT:    paddb %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4,4,4,4,4]
-; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
-; SSE2-NEXT:    retq
+; SSE-LABEL: ult_4_v8i16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddw %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm0
+; SSE-NEXT:    paddw %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    paddw %xmm0, %xmm1
+; SSE-NEXT:    pand %xmm1, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqw %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
-; SSE3-LABEL: ult_4_v8i16:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    psllw $8, %xmm2
-; SSE3-NEXT:    paddb %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4,4,4,4,4]
-; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ult_4_v8i16:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa %xmm3, %xmm1
-; SSSE3-NEXT:    psllw $8, %xmm1
-; SSSE3-NEXT:    paddb %xmm3, %xmm1
-; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4,4,4,4,4]
-; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ult_4_v8i16:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa %xmm3, %xmm1
-; SSE41-NEXT:    psllw $8, %xmm1
-; SSE41-NEXT:    paddb %xmm3, %xmm1
-; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4,4,4,4,4]
-; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
-; SSE41-NEXT:    retq
-;
-; AVX1-LABEL: ult_4_v8i16:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4,4,4,4,4,4,4]
-; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    retq
+; AVX1-LABEL: ult_4_v8i16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm1
+; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_4_v8i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm1
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4,4,4,4,4,4,4]
-; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_4_v8i16:
@@ -3579,7 +2455,188 @@
 }
 
 define <8 x i16> @ugt_4_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_4_v8i16:
+; SSE-LABEL: ugt_4_v8i16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddw %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm0
+; SSE-NEXT:    paddw %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddw %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm0
+; SSE-NEXT:    paddw %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm2, %xmm2
+; SSE-NEXT:    pcmpeqw %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX1-LABEL: ugt_4_v8i16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: ugt_4_v8i16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqw %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    retq
+;
+; AVX512VPOPCNTDQ-LABEL: ugt_4_v8i16:
+; AVX512VPOPCNTDQ:       # %bb.0:
+; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
+; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
+; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
+; AVX512VPOPCNTDQ-NEXT:    vzeroupper
+; AVX512VPOPCNTDQ-NEXT:    retq
+;
+; AVX512VPOPCNTDQVL-LABEL: ugt_4_v8i16:
+; AVX512VPOPCNTDQVL:       # %bb.0:
+; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
+; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
+; AVX512VPOPCNTDQVL-NEXT:    retq
+;
+; BITALG_NOVLX-LABEL: ugt_4_v8i16:
+; BITALG_NOVLX:       # %bb.0:
+; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
+; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vzeroupper
+; BITALG_NOVLX-NEXT:    retq
+;
+; BITALG-LABEL: ugt_4_v8i16:
+; BITALG:       # %bb.0:
+; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
+; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
+; BITALG-NEXT:    vpmovm2w %k0, %xmm0
+; BITALG-NEXT:    retq
+  %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
+  %3 = icmp ugt <8 x i16> %2, <i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4>
+  %4 = sext <8 x i1> %3 to <8 x i16>
+  ret <8 x i16> %4
+}
+
+define <8 x i16> @ult_5_v8i16(<8 x i16> %0) {
+; SSE-LABEL: ult_5_v8i16:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddw %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm3
+; SSE-NEXT:    paddw %xmm1, %xmm3
+; SSE-NEXT:    pand %xmm2, %xmm3
+; SSE-NEXT:    movdqa %xmm3, %xmm0
+; SSE-NEXT:    paddw %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm3, %xmm0
+; SSE-NEXT:    paddw %xmm0, %xmm1
+; SSE-NEXT:    pand %xmm1, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqw %xmm1, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX1-LABEL: ult_5_v8i16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm1
+; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: ult_5_v8i16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm1
+; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    retq
+;
+; AVX512VPOPCNTDQ-LABEL: ult_5_v8i16:
+; AVX512VPOPCNTDQ:       # %bb.0:
+; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
+; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5,5,5,5,5,5,5]
+; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
+; AVX512VPOPCNTDQ-NEXT:    vzeroupper
+; AVX512VPOPCNTDQ-NEXT:    retq
+;
+; AVX512VPOPCNTDQVL-LABEL: ult_5_v8i16:
+; AVX512VPOPCNTDQVL:       # %bb.0:
+; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
+; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5,5,5,5,5,5,5]
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
+; AVX512VPOPCNTDQVL-NEXT:    retq
+;
+; BITALG_NOVLX-LABEL: ult_5_v8i16:
+; BITALG_NOVLX:       # %bb.0:
+; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
+; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5,5,5,5,5,5,5]
+; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
+; BITALG_NOVLX-NEXT:    vzeroupper
+; BITALG_NOVLX-NEXT:    retq
+;
+; BITALG-LABEL: ult_5_v8i16:
+; BITALG:       # %bb.0:
+; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
+; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
+; BITALG-NEXT:    vpmovm2w %k0, %xmm0
+; BITALG-NEXT:    retq
+  %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
+  %3 = icmp ult <8 x i16> %2, <i16 5, i16 5, i16 5, i16 5, i16 5, i16 5, i16 5, i16 5>
+  %4 = sext <8 x i1> %3 to <8 x i16>
+  ret <8 x i16> %4
+}
+
+define <8 x i16> @ugt_5_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ugt_5_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -3602,7 +2659,7 @@
 ; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_4_v8i16:
+; SSE3-LABEL: ugt_5_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -3625,7 +2682,7 @@
 ; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_4_v8i16:
+; SSSE3-LABEL: ugt_5_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -3644,7 +2701,7 @@
 ; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_4_v8i16:
+; SSE41-LABEL: ugt_5_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -3663,7 +2720,7 @@
 ; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_4_v8i16:
+; AVX1-LABEL: ugt_5_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -3679,7 +2736,7 @@
 ; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_4_v8i16:
+; AVX2-LABEL: ugt_5_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -3695,7 +2752,7 @@
 ; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_4_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ugt_5_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
@@ -3704,7 +2761,7 @@
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_4_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ugt_5_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
@@ -3713,7 +2770,7 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_4_v8i16:
+; BITALG_NOVLX-LABEL: ugt_5_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
@@ -3721,20 +2778,20 @@
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_4_v8i16:
+; BITALG-LABEL: ugt_5_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4>
+  %3 = icmp ugt <8 x i16> %2, <i16 5, i16 5, i16 5, i16 5, i16 5, i16 5, i16 5, i16 5>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ult_5_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_5_v8i16:
+define <8 x i16> @ult_6_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ult_6_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -3754,11 +2811,11 @@
 ; SSE2-NEXT:    psllw $8, %xmm2
 ; SSE2-NEXT:    paddb %xmm1, %xmm2
 ; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [5,5,5,5,5,5,5,5]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [6,6,6,6,6,6,6,6]
 ; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ult_5_v8i16:
+; SSE3-LABEL: ult_6_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -3778,11 +2835,11 @@
 ; SSE3-NEXT:    psllw $8, %xmm2
 ; SSE3-NEXT:    paddb %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [5,5,5,5,5,5,5,5]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [6,6,6,6,6,6,6,6]
 ; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_5_v8i16:
+; SSSE3-LABEL: ult_6_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -3798,11 +2855,11 @@
 ; SSSE3-NEXT:    psllw $8, %xmm1
 ; SSSE3-NEXT:    paddb %xmm3, %xmm1
 ; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [5,5,5,5,5,5,5,5]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [6,6,6,6,6,6,6,6]
 ; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ult_5_v8i16:
+; SSE41-LABEL: ult_6_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -3818,11 +2875,11 @@
 ; SSE41-NEXT:    psllw $8, %xmm1
 ; SSE41-NEXT:    paddb %xmm3, %xmm1
 ; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [5,5,5,5,5,5,5,5]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [6,6,6,6,6,6,6,6]
 ; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ult_5_v8i16:
+; AVX1-LABEL: ult_6_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -3835,11 +2892,11 @@
 ; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5,5,5,5,5,5,5]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [6,6,6,6,6,6,6,6]
 ; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_5_v8i16:
+; AVX2-LABEL: ult_6_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -3852,53 +2909,53 @@
 ; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5,5,5,5,5,5,5]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [6,6,6,6,6,6,6,6]
 ; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_5_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ult_6_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5,5,5,5,5,5,5]
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [6,6,6,6,6,6,6,6]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_5_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ult_6_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5,5,5,5,5,5,5]
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [6,6,6,6,6,6,6,6]
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_5_v8i16:
+; BITALG_NOVLX-LABEL: ult_6_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5,5,5,5,5,5,5]
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [6,6,6,6,6,6,6,6]
 ; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_5_v8i16:
+; BITALG-LABEL: ult_6_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 5, i16 5, i16 5, i16 5, i16 5, i16 5, i16 5, i16 5>
+  %3 = icmp ult <8 x i16> %2, <i16 6, i16 6, i16 6, i16 6, i16 6, i16 6, i16 6, i16 6>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ugt_5_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_5_v8i16:
+define <8 x i16> @ugt_6_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ugt_6_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -3921,7 +2978,7 @@
 ; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_5_v8i16:
+; SSE3-LABEL: ugt_6_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -3944,7 +3001,7 @@
 ; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_5_v8i16:
+; SSSE3-LABEL: ugt_6_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -3963,7 +3020,7 @@
 ; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_5_v8i16:
+; SSE41-LABEL: ugt_6_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -3982,7 +3039,7 @@
 ; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_5_v8i16:
+; AVX1-LABEL: ugt_6_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -3998,7 +3055,7 @@
 ; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_5_v8i16:
+; AVX2-LABEL: ugt_6_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4014,7 +3071,7 @@
 ; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_5_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ugt_6_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
@@ -4023,7 +3080,7 @@
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_5_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ugt_6_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
@@ -4032,7 +3089,7 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_5_v8i16:
+; BITALG_NOVLX-LABEL: ugt_6_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
@@ -4040,20 +3097,20 @@
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_5_v8i16:
+; BITALG-LABEL: ugt_6_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 5, i16 5, i16 5, i16 5, i16 5, i16 5, i16 5, i16 5>
+  %3 = icmp ugt <8 x i16> %2, <i16 6, i16 6, i16 6, i16 6, i16 6, i16 6, i16 6, i16 6>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ult_6_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_6_v8i16:
+define <8 x i16> @ult_7_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ult_7_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -4073,11 +3130,11 @@
 ; SSE2-NEXT:    psllw $8, %xmm2
 ; SSE2-NEXT:    paddb %xmm1, %xmm2
 ; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [6,6,6,6,6,6,6,6]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [7,7,7,7,7,7,7,7]
 ; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ult_6_v8i16:
+; SSE3-LABEL: ult_7_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -4097,11 +3154,11 @@
 ; SSE3-NEXT:    psllw $8, %xmm2
 ; SSE3-NEXT:    paddb %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [6,6,6,6,6,6,6,6]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [7,7,7,7,7,7,7,7]
 ; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_6_v8i16:
+; SSSE3-LABEL: ult_7_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -4117,11 +3174,11 @@
 ; SSSE3-NEXT:    psllw $8, %xmm1
 ; SSSE3-NEXT:    paddb %xmm3, %xmm1
 ; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [6,6,6,6,6,6,6,6]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [7,7,7,7,7,7,7,7]
 ; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ult_6_v8i16:
+; SSE41-LABEL: ult_7_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -4137,11 +3194,11 @@
 ; SSE41-NEXT:    psllw $8, %xmm1
 ; SSE41-NEXT:    paddb %xmm3, %xmm1
 ; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [6,6,6,6,6,6,6,6]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [7,7,7,7,7,7,7,7]
 ; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ult_6_v8i16:
+; AVX1-LABEL: ult_7_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4154,11 +3211,11 @@
 ; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [6,6,6,6,6,6,6,6]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [7,7,7,7,7,7,7,7]
 ; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_6_v8i16:
+; AVX2-LABEL: ult_7_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4171,53 +3228,53 @@
 ; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [6,6,6,6,6,6,6,6]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [7,7,7,7,7,7,7,7]
 ; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_6_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ult_7_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [6,6,6,6,6,6,6,6]
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [7,7,7,7,7,7,7,7]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_6_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ult_7_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [6,6,6,6,6,6,6,6]
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [7,7,7,7,7,7,7,7]
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_6_v8i16:
+; BITALG_NOVLX-LABEL: ult_7_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [6,6,6,6,6,6,6,6]
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [7,7,7,7,7,7,7,7]
 ; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_6_v8i16:
+; BITALG-LABEL: ult_7_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 6, i16 6, i16 6, i16 6, i16 6, i16 6, i16 6, i16 6>
+  %3 = icmp ult <8 x i16> %2, <i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ugt_6_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_6_v8i16:
+define <8 x i16> @ugt_7_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ugt_7_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -4240,7 +3297,7 @@
 ; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_6_v8i16:
+; SSE3-LABEL: ugt_7_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -4263,7 +3320,7 @@
 ; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_6_v8i16:
+; SSSE3-LABEL: ugt_7_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -4282,7 +3339,7 @@
 ; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_6_v8i16:
+; SSE41-LABEL: ugt_7_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -4301,7 +3358,7 @@
 ; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_6_v8i16:
+; AVX1-LABEL: ugt_7_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4317,7 +3374,7 @@
 ; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_6_v8i16:
+; AVX2-LABEL: ugt_7_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4333,7 +3390,7 @@
 ; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_6_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ugt_7_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
@@ -4342,7 +3399,7 @@
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_6_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ugt_7_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
@@ -4351,7 +3408,7 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_6_v8i16:
+; BITALG_NOVLX-LABEL: ugt_7_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
@@ -4359,20 +3416,20 @@
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_6_v8i16:
+; BITALG-LABEL: ugt_7_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 6, i16 6, i16 6, i16 6, i16 6, i16 6, i16 6, i16 6>
+  %3 = icmp ugt <8 x i16> %2, <i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ult_7_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_7_v8i16:
+define <8 x i16> @ult_8_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ult_8_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -4392,11 +3449,11 @@
 ; SSE2-NEXT:    psllw $8, %xmm2
 ; SSE2-NEXT:    paddb %xmm1, %xmm2
 ; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [7,7,7,7,7,7,7,7]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [8,8,8,8,8,8,8,8]
 ; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ult_7_v8i16:
+; SSE3-LABEL: ult_8_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -4416,11 +3473,11 @@
 ; SSE3-NEXT:    psllw $8, %xmm2
 ; SSE3-NEXT:    paddb %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [7,7,7,7,7,7,7,7]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [8,8,8,8,8,8,8,8]
 ; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_7_v8i16:
+; SSSE3-LABEL: ult_8_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -4436,11 +3493,11 @@
 ; SSSE3-NEXT:    psllw $8, %xmm1
 ; SSSE3-NEXT:    paddb %xmm3, %xmm1
 ; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [7,7,7,7,7,7,7,7]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [8,8,8,8,8,8,8,8]
 ; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ult_7_v8i16:
+; SSE41-LABEL: ult_8_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -4456,11 +3513,11 @@
 ; SSE41-NEXT:    psllw $8, %xmm1
 ; SSE41-NEXT:    paddb %xmm3, %xmm1
 ; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [7,7,7,7,7,7,7,7]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [8,8,8,8,8,8,8,8]
 ; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ult_7_v8i16:
+; AVX1-LABEL: ult_8_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4473,11 +3530,11 @@
 ; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [7,7,7,7,7,7,7,7]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [8,8,8,8,8,8,8,8]
 ; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_7_v8i16:
+; AVX2-LABEL: ult_8_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4490,53 +3547,53 @@
 ; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [7,7,7,7,7,7,7,7]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [8,8,8,8,8,8,8,8]
 ; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_7_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ult_8_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [7,7,7,7,7,7,7,7]
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [8,8,8,8,8,8,8,8]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_7_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ult_8_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [7,7,7,7,7,7,7,7]
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [8,8,8,8,8,8,8,8]
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_7_v8i16:
+; BITALG_NOVLX-LABEL: ult_8_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [7,7,7,7,7,7,7,7]
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [8,8,8,8,8,8,8,8]
 ; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_7_v8i16:
+; BITALG-LABEL: ult_8_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7>
+  %3 = icmp ult <8 x i16> %2, <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ugt_7_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_7_v8i16:
+define <8 x i16> @ugt_8_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ugt_8_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -4559,7 +3616,7 @@
 ; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_7_v8i16:
+; SSE3-LABEL: ugt_8_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -4582,7 +3639,7 @@
 ; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_7_v8i16:
+; SSSE3-LABEL: ugt_8_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -4601,7 +3658,7 @@
 ; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_7_v8i16:
+; SSE41-LABEL: ugt_8_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -4620,7 +3677,7 @@
 ; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_7_v8i16:
+; AVX1-LABEL: ugt_8_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4636,7 +3693,7 @@
 ; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_7_v8i16:
+; AVX2-LABEL: ugt_8_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4652,7 +3709,7 @@
 ; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_7_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ugt_8_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
@@ -4661,7 +3718,7 @@
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_7_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ugt_8_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
@@ -4670,7 +3727,7 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_7_v8i16:
+; BITALG_NOVLX-LABEL: ugt_8_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
@@ -4678,20 +3735,20 @@
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_7_v8i16:
+; BITALG-LABEL: ugt_8_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7>
+  %3 = icmp ugt <8 x i16> %2, <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ult_8_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_8_v8i16:
+define <8 x i16> @ult_9_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ult_9_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -4711,11 +3768,11 @@
 ; SSE2-NEXT:    psllw $8, %xmm2
 ; SSE2-NEXT:    paddb %xmm1, %xmm2
 ; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [8,8,8,8,8,8,8,8]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [9,9,9,9,9,9,9,9]
 ; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ult_8_v8i16:
+; SSE3-LABEL: ult_9_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -4735,11 +3792,11 @@
 ; SSE3-NEXT:    psllw $8, %xmm2
 ; SSE3-NEXT:    paddb %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [8,8,8,8,8,8,8,8]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [9,9,9,9,9,9,9,9]
 ; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_8_v8i16:
+; SSSE3-LABEL: ult_9_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -4755,11 +3812,11 @@
 ; SSSE3-NEXT:    psllw $8, %xmm1
 ; SSSE3-NEXT:    paddb %xmm3, %xmm1
 ; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [8,8,8,8,8,8,8,8]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [9,9,9,9,9,9,9,9]
 ; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ult_8_v8i16:
+; SSE41-LABEL: ult_9_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -4775,11 +3832,11 @@
 ; SSE41-NEXT:    psllw $8, %xmm1
 ; SSE41-NEXT:    paddb %xmm3, %xmm1
 ; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [8,8,8,8,8,8,8,8]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [9,9,9,9,9,9,9,9]
 ; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ult_8_v8i16:
+; AVX1-LABEL: ult_9_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4792,11 +3849,11 @@
 ; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [8,8,8,8,8,8,8,8]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [9,9,9,9,9,9,9,9]
 ; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_8_v8i16:
+; AVX2-LABEL: ult_9_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4809,53 +3866,53 @@
 ; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [8,8,8,8,8,8,8,8]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [9,9,9,9,9,9,9,9]
 ; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_8_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ult_9_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [8,8,8,8,8,8,8,8]
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [9,9,9,9,9,9,9,9]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_8_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ult_9_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [8,8,8,8,8,8,8,8]
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [9,9,9,9,9,9,9,9]
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_8_v8i16:
+; BITALG_NOVLX-LABEL: ult_9_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [8,8,8,8,8,8,8,8]
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [9,9,9,9,9,9,9,9]
 ; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_8_v8i16:
+; BITALG-LABEL: ult_9_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
+  %3 = icmp ult <8 x i16> %2, <i16 9, i16 9, i16 9, i16 9, i16 9, i16 9, i16 9, i16 9>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ugt_8_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_8_v8i16:
+define <8 x i16> @ugt_9_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ugt_9_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -4878,7 +3935,7 @@
 ; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_8_v8i16:
+; SSE3-LABEL: ugt_9_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -4901,7 +3958,7 @@
 ; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_8_v8i16:
+; SSSE3-LABEL: ugt_9_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -4920,7 +3977,7 @@
 ; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_8_v8i16:
+; SSE41-LABEL: ugt_9_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -4939,7 +3996,7 @@
 ; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_8_v8i16:
+; AVX1-LABEL: ugt_9_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4955,7 +4012,7 @@
 ; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_8_v8i16:
+; AVX2-LABEL: ugt_9_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -4971,7 +4028,7 @@
 ; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_8_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ugt_9_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
@@ -4980,7 +4037,7 @@
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_8_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ugt_9_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
@@ -4989,7 +4046,7 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_8_v8i16:
+; BITALG_NOVLX-LABEL: ugt_9_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
@@ -4997,20 +4054,20 @@
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_8_v8i16:
+; BITALG-LABEL: ugt_9_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
+  %3 = icmp ugt <8 x i16> %2, <i16 9, i16 9, i16 9, i16 9, i16 9, i16 9, i16 9, i16 9>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ult_9_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_9_v8i16:
+define <8 x i16> @ult_10_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ult_10_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -5030,11 +4087,11 @@
 ; SSE2-NEXT:    psllw $8, %xmm2
 ; SSE2-NEXT:    paddb %xmm1, %xmm2
 ; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [9,9,9,9,9,9,9,9]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [10,10,10,10,10,10,10,10]
 ; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ult_9_v8i16:
+; SSE3-LABEL: ult_10_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -5054,11 +4111,11 @@
 ; SSE3-NEXT:    psllw $8, %xmm2
 ; SSE3-NEXT:    paddb %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [9,9,9,9,9,9,9,9]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [10,10,10,10,10,10,10,10]
 ; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_9_v8i16:
+; SSSE3-LABEL: ult_10_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -5074,11 +4131,11 @@
 ; SSSE3-NEXT:    psllw $8, %xmm1
 ; SSSE3-NEXT:    paddb %xmm3, %xmm1
 ; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [9,9,9,9,9,9,9,9]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [10,10,10,10,10,10,10,10]
 ; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ult_9_v8i16:
+; SSE41-LABEL: ult_10_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -5094,11 +4151,11 @@
 ; SSE41-NEXT:    psllw $8, %xmm1
 ; SSE41-NEXT:    paddb %xmm3, %xmm1
 ; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [9,9,9,9,9,9,9,9]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [10,10,10,10,10,10,10,10]
 ; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ult_9_v8i16:
+; AVX1-LABEL: ult_10_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -5111,11 +4168,11 @@
 ; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [9,9,9,9,9,9,9,9]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [10,10,10,10,10,10,10,10]
 ; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_9_v8i16:
+; AVX2-LABEL: ult_10_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -5128,53 +4185,53 @@
 ; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [9,9,9,9,9,9,9,9]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [10,10,10,10,10,10,10,10]
 ; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_9_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ult_10_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [9,9,9,9,9,9,9,9]
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [10,10,10,10,10,10,10,10]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_9_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ult_10_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [9,9,9,9,9,9,9,9]
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [10,10,10,10,10,10,10,10]
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_9_v8i16:
+; BITALG_NOVLX-LABEL: ult_10_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [9,9,9,9,9,9,9,9]
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [10,10,10,10,10,10,10,10]
 ; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_9_v8i16:
+; BITALG-LABEL: ult_10_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 9, i16 9, i16 9, i16 9, i16 9, i16 9, i16 9, i16 9>
+  %3 = icmp ult <8 x i16> %2, <i16 10, i16 10, i16 10, i16 10, i16 10, i16 10, i16 10, i16 10>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ugt_9_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_9_v8i16:
+define <8 x i16> @ugt_10_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ugt_10_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -5197,7 +4254,7 @@
 ; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_9_v8i16:
+; SSE3-LABEL: ugt_10_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -5220,7 +4277,7 @@
 ; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_9_v8i16:
+; SSSE3-LABEL: ugt_10_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -5239,7 +4296,7 @@
 ; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_9_v8i16:
+; SSE41-LABEL: ugt_10_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -5258,7 +4315,7 @@
 ; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_9_v8i16:
+; AVX1-LABEL: ugt_10_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -5274,7 +4331,7 @@
 ; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_9_v8i16:
+; AVX2-LABEL: ugt_10_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -5290,7 +4347,7 @@
 ; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_9_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ugt_10_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
@@ -5299,7 +4356,7 @@
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_9_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ugt_10_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
@@ -5308,7 +4365,7 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_9_v8i16:
+; BITALG_NOVLX-LABEL: ugt_10_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
@@ -5316,20 +4373,20 @@
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_9_v8i16:
+; BITALG-LABEL: ugt_10_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 9, i16 9, i16 9, i16 9, i16 9, i16 9, i16 9, i16 9>
+  %3 = icmp ugt <8 x i16> %2, <i16 10, i16 10, i16 10, i16 10, i16 10, i16 10, i16 10, i16 10>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ult_10_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_10_v8i16:
+define <8 x i16> @ult_11_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ult_11_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -5349,11 +4406,11 @@
 ; SSE2-NEXT:    psllw $8, %xmm2
 ; SSE2-NEXT:    paddb %xmm1, %xmm2
 ; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [10,10,10,10,10,10,10,10]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [11,11,11,11,11,11,11,11]
 ; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ult_10_v8i16:
+; SSE3-LABEL: ult_11_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -5373,11 +4430,11 @@
 ; SSE3-NEXT:    psllw $8, %xmm2
 ; SSE3-NEXT:    paddb %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [10,10,10,10,10,10,10,10]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [11,11,11,11,11,11,11,11]
 ; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_10_v8i16:
+; SSSE3-LABEL: ult_11_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -5393,11 +4450,11 @@
 ; SSSE3-NEXT:    psllw $8, %xmm1
 ; SSSE3-NEXT:    paddb %xmm3, %xmm1
 ; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [10,10,10,10,10,10,10,10]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [11,11,11,11,11,11,11,11]
 ; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ult_10_v8i16:
+; SSE41-LABEL: ult_11_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -5413,11 +4470,11 @@
 ; SSE41-NEXT:    psllw $8, %xmm1
 ; SSE41-NEXT:    paddb %xmm3, %xmm1
 ; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [10,10,10,10,10,10,10,10]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [11,11,11,11,11,11,11,11]
 ; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ult_10_v8i16:
+; AVX1-LABEL: ult_11_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -5430,11 +4487,11 @@
 ; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [10,10,10,10,10,10,10,10]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [11,11,11,11,11,11,11,11]
 ; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_10_v8i16:
+; AVX2-LABEL: ult_11_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -5447,53 +4504,53 @@
 ; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [10,10,10,10,10,10,10,10]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [11,11,11,11,11,11,11,11]
 ; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_10_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ult_11_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [10,10,10,10,10,10,10,10]
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [11,11,11,11,11,11,11,11]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_10_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ult_11_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [10,10,10,10,10,10,10,10]
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [11,11,11,11,11,11,11,11]
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_10_v8i16:
+; BITALG_NOVLX-LABEL: ult_11_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [10,10,10,10,10,10,10,10]
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [11,11,11,11,11,11,11,11]
 ; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_10_v8i16:
+; BITALG-LABEL: ult_11_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 10, i16 10, i16 10, i16 10, i16 10, i16 10, i16 10, i16 10>
+  %3 = icmp ult <8 x i16> %2, <i16 11, i16 11, i16 11, i16 11, i16 11, i16 11, i16 11, i16 11>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ugt_10_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_10_v8i16:
+define <8 x i16> @ugt_11_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ugt_11_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -5516,7 +4573,7 @@
 ; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_10_v8i16:
+; SSE3-LABEL: ugt_11_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -5539,7 +4596,7 @@
 ; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_10_v8i16:
+; SSSE3-LABEL: ugt_11_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -5558,7 +4615,7 @@
 ; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_10_v8i16:
+; SSE41-LABEL: ugt_11_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -5577,7 +4634,7 @@
 ; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_10_v8i16:
+; AVX1-LABEL: ugt_11_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -5593,7 +4650,7 @@
 ; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_10_v8i16:
+; AVX2-LABEL: ugt_11_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -5609,7 +4666,7 @@
 ; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_10_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ugt_11_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
@@ -5618,7 +4675,7 @@
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_10_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ugt_11_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
@@ -5627,7 +4684,7 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_10_v8i16:
+; BITALG_NOVLX-LABEL: ugt_11_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
@@ -5635,20 +4692,20 @@
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_10_v8i16:
+; BITALG-LABEL: ugt_11_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 10, i16 10, i16 10, i16 10, i16 10, i16 10, i16 10, i16 10>
+  %3 = icmp ugt <8 x i16> %2, <i16 11, i16 11, i16 11, i16 11, i16 11, i16 11, i16 11, i16 11>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ult_11_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_11_v8i16:
+define <8 x i16> @ult_12_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ult_12_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -5668,11 +4725,11 @@
 ; SSE2-NEXT:    psllw $8, %xmm2
 ; SSE2-NEXT:    paddb %xmm1, %xmm2
 ; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [11,11,11,11,11,11,11,11]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [12,12,12,12,12,12,12,12]
 ; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ult_11_v8i16:
+; SSE3-LABEL: ult_12_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -5692,11 +4749,11 @@
 ; SSE3-NEXT:    psllw $8, %xmm2
 ; SSE3-NEXT:    paddb %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [11,11,11,11,11,11,11,11]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [12,12,12,12,12,12,12,12]
 ; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_11_v8i16:
+; SSSE3-LABEL: ult_12_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -5712,11 +4769,11 @@
 ; SSSE3-NEXT:    psllw $8, %xmm1
 ; SSSE3-NEXT:    paddb %xmm3, %xmm1
 ; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [11,11,11,11,11,11,11,11]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [12,12,12,12,12,12,12,12]
 ; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ult_11_v8i16:
+; SSE41-LABEL: ult_12_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -5732,11 +4789,11 @@
 ; SSE41-NEXT:    psllw $8, %xmm1
 ; SSE41-NEXT:    paddb %xmm3, %xmm1
 ; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [11,11,11,11,11,11,11,11]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [12,12,12,12,12,12,12,12]
 ; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ult_11_v8i16:
+; AVX1-LABEL: ult_12_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -5749,11 +4806,11 @@
 ; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [11,11,11,11,11,11,11,11]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [12,12,12,12,12,12,12,12]
 ; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_11_v8i16:
+; AVX2-LABEL: ult_12_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -5766,53 +4823,53 @@
 ; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [11,11,11,11,11,11,11,11]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [12,12,12,12,12,12,12,12]
 ; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_11_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ult_12_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [11,11,11,11,11,11,11,11]
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [12,12,12,12,12,12,12,12]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_11_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ult_12_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [11,11,11,11,11,11,11,11]
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [12,12,12,12,12,12,12,12]
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_11_v8i16:
+; BITALG_NOVLX-LABEL: ult_12_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [11,11,11,11,11,11,11,11]
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [12,12,12,12,12,12,12,12]
 ; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_11_v8i16:
+; BITALG-LABEL: ult_12_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 11, i16 11, i16 11, i16 11, i16 11, i16 11, i16 11, i16 11>
+  %3 = icmp ult <8 x i16> %2, <i16 12, i16 12, i16 12, i16 12, i16 12, i16 12, i16 12, i16 12>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ugt_11_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_11_v8i16:
+define <8 x i16> @ugt_12_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ugt_12_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -5835,7 +4892,7 @@
 ; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_11_v8i16:
+; SSE3-LABEL: ugt_12_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -5858,7 +4915,7 @@
 ; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_11_v8i16:
+; SSSE3-LABEL: ugt_12_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -5877,7 +4934,7 @@
 ; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_11_v8i16:
+; SSE41-LABEL: ugt_12_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -5896,7 +4953,7 @@
 ; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_11_v8i16:
+; AVX1-LABEL: ugt_12_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -5912,7 +4969,7 @@
 ; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_11_v8i16:
+; AVX2-LABEL: ugt_12_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -5928,7 +4985,7 @@
 ; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_11_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ugt_12_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
@@ -5937,7 +4994,7 @@
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_11_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ugt_12_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
@@ -5946,7 +5003,7 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_11_v8i16:
+; BITALG_NOVLX-LABEL: ugt_12_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
@@ -5954,20 +5011,20 @@
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_11_v8i16:
+; BITALG-LABEL: ugt_12_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 11, i16 11, i16 11, i16 11, i16 11, i16 11, i16 11, i16 11>
+  %3 = icmp ugt <8 x i16> %2, <i16 12, i16 12, i16 12, i16 12, i16 12, i16 12, i16 12, i16 12>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ult_12_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_12_v8i16:
+define <8 x i16> @ult_13_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ult_13_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -5987,11 +5044,11 @@
 ; SSE2-NEXT:    psllw $8, %xmm2
 ; SSE2-NEXT:    paddb %xmm1, %xmm2
 ; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [12,12,12,12,12,12,12,12]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [13,13,13,13,13,13,13,13]
 ; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ult_12_v8i16:
+; SSE3-LABEL: ult_13_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -6011,11 +5068,11 @@
 ; SSE3-NEXT:    psllw $8, %xmm2
 ; SSE3-NEXT:    paddb %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [12,12,12,12,12,12,12,12]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [13,13,13,13,13,13,13,13]
 ; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_12_v8i16:
+; SSSE3-LABEL: ult_13_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -6031,11 +5088,11 @@
 ; SSSE3-NEXT:    psllw $8, %xmm1
 ; SSSE3-NEXT:    paddb %xmm3, %xmm1
 ; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [12,12,12,12,12,12,12,12]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [13,13,13,13,13,13,13,13]
 ; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ult_12_v8i16:
+; SSE41-LABEL: ult_13_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -6051,11 +5108,11 @@
 ; SSE41-NEXT:    psllw $8, %xmm1
 ; SSE41-NEXT:    paddb %xmm3, %xmm1
 ; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [12,12,12,12,12,12,12,12]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [13,13,13,13,13,13,13,13]
 ; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ult_12_v8i16:
+; AVX1-LABEL: ult_13_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -6068,11 +5125,11 @@
 ; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [12,12,12,12,12,12,12,12]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [13,13,13,13,13,13,13,13]
 ; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_12_v8i16:
+; AVX2-LABEL: ult_13_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -6085,53 +5142,53 @@
 ; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [12,12,12,12,12,12,12,12]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [13,13,13,13,13,13,13,13]
 ; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_12_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ult_13_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [12,12,12,12,12,12,12,12]
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [13,13,13,13,13,13,13,13]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_12_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ult_13_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [12,12,12,12,12,12,12,12]
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [13,13,13,13,13,13,13,13]
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_12_v8i16:
+; BITALG_NOVLX-LABEL: ult_13_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [12,12,12,12,12,12,12,12]
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [13,13,13,13,13,13,13,13]
 ; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_12_v8i16:
+; BITALG-LABEL: ult_13_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 12, i16 12, i16 12, i16 12, i16 12, i16 12, i16 12, i16 12>
+  %3 = icmp ult <8 x i16> %2, <i16 13, i16 13, i16 13, i16 13, i16 13, i16 13, i16 13, i16 13>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ugt_12_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_12_v8i16:
+define <8 x i16> @ugt_13_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ugt_13_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -6154,7 +5211,7 @@
 ; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_12_v8i16:
+; SSE3-LABEL: ugt_13_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -6177,7 +5234,7 @@
 ; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_12_v8i16:
+; SSSE3-LABEL: ugt_13_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -6196,7 +5253,7 @@
 ; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_12_v8i16:
+; SSE41-LABEL: ugt_13_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -6215,7 +5272,7 @@
 ; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_12_v8i16:
+; AVX1-LABEL: ugt_13_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -6231,7 +5288,7 @@
 ; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_12_v8i16:
+; AVX2-LABEL: ugt_13_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -6247,7 +5304,7 @@
 ; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_12_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ugt_13_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
@@ -6256,7 +5313,7 @@
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_12_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ugt_13_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
@@ -6265,7 +5322,7 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_12_v8i16:
+; BITALG_NOVLX-LABEL: ugt_13_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
@@ -6273,20 +5330,20 @@
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_12_v8i16:
+; BITALG-LABEL: ugt_13_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 12, i16 12, i16 12, i16 12, i16 12, i16 12, i16 12, i16 12>
+  %3 = icmp ugt <8 x i16> %2, <i16 13, i16 13, i16 13, i16 13, i16 13, i16 13, i16 13, i16 13>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ult_13_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_13_v8i16:
+define <8 x i16> @ult_14_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ult_14_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -6306,11 +5363,11 @@
 ; SSE2-NEXT:    psllw $8, %xmm2
 ; SSE2-NEXT:    paddb %xmm1, %xmm2
 ; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [13,13,13,13,13,13,13,13]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [14,14,14,14,14,14,14,14]
 ; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ult_13_v8i16:
+; SSE3-LABEL: ult_14_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -6330,11 +5387,11 @@
 ; SSE3-NEXT:    psllw $8, %xmm2
 ; SSE3-NEXT:    paddb %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [13,13,13,13,13,13,13,13]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [14,14,14,14,14,14,14,14]
 ; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_13_v8i16:
+; SSSE3-LABEL: ult_14_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -6350,11 +5407,11 @@
 ; SSSE3-NEXT:    psllw $8, %xmm1
 ; SSSE3-NEXT:    paddb %xmm3, %xmm1
 ; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [13,13,13,13,13,13,13,13]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [14,14,14,14,14,14,14,14]
 ; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ult_13_v8i16:
+; SSE41-LABEL: ult_14_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -6370,11 +5427,11 @@
 ; SSE41-NEXT:    psllw $8, %xmm1
 ; SSE41-NEXT:    paddb %xmm3, %xmm1
 ; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [13,13,13,13,13,13,13,13]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [14,14,14,14,14,14,14,14]
 ; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ult_13_v8i16:
+; AVX1-LABEL: ult_14_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -6387,11 +5444,11 @@
 ; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [13,13,13,13,13,13,13,13]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [14,14,14,14,14,14,14,14]
 ; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_13_v8i16:
+; AVX2-LABEL: ult_14_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -6404,53 +5461,53 @@
 ; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [13,13,13,13,13,13,13,13]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [14,14,14,14,14,14,14,14]
 ; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_13_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ult_14_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [13,13,13,13,13,13,13,13]
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [14,14,14,14,14,14,14,14]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_13_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ult_14_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [13,13,13,13,13,13,13,13]
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [14,14,14,14,14,14,14,14]
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_13_v8i16:
+; BITALG_NOVLX-LABEL: ult_14_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [13,13,13,13,13,13,13,13]
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [14,14,14,14,14,14,14,14]
 ; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_13_v8i16:
+; BITALG-LABEL: ult_14_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 13, i16 13, i16 13, i16 13, i16 13, i16 13, i16 13, i16 13>
+  %3 = icmp ult <8 x i16> %2, <i16 14, i16 14, i16 14, i16 14, i16 14, i16 14, i16 14, i16 14>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ugt_13_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_13_v8i16:
+define <8 x i16> @ugt_14_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ugt_14_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -6473,7 +5530,7 @@
 ; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_13_v8i16:
+; SSE3-LABEL: ugt_14_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -6496,7 +5553,7 @@
 ; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_13_v8i16:
+; SSSE3-LABEL: ugt_14_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -6515,7 +5572,7 @@
 ; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_13_v8i16:
+; SSE41-LABEL: ugt_14_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -6534,7 +5591,7 @@
 ; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_13_v8i16:
+; AVX1-LABEL: ugt_14_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -6550,7 +5607,7 @@
 ; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_13_v8i16:
+; AVX2-LABEL: ugt_14_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -6566,7 +5623,7 @@
 ; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_13_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ugt_14_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
@@ -6575,7 +5632,7 @@
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_13_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ugt_14_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
@@ -6584,7 +5641,7 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_13_v8i16:
+; BITALG_NOVLX-LABEL: ugt_14_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
@@ -6592,20 +5649,20 @@
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_13_v8i16:
+; BITALG-LABEL: ugt_14_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 13, i16 13, i16 13, i16 13, i16 13, i16 13, i16 13, i16 13>
+  %3 = icmp ugt <8 x i16> %2, <i16 14, i16 14, i16 14, i16 14, i16 14, i16 14, i16 14, i16 14>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ult_14_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_14_v8i16:
+define <8 x i16> @ult_15_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ult_15_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -6625,11 +5682,11 @@
 ; SSE2-NEXT:    psllw $8, %xmm2
 ; SSE2-NEXT:    paddb %xmm1, %xmm2
 ; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [14,14,14,14,14,14,14,14]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [15,15,15,15,15,15,15,15]
 ; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ult_14_v8i16:
+; SSE3-LABEL: ult_15_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -6649,11 +5706,11 @@
 ; SSE3-NEXT:    psllw $8, %xmm2
 ; SSE3-NEXT:    paddb %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [14,14,14,14,14,14,14,14]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [15,15,15,15,15,15,15,15]
 ; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_14_v8i16:
+; SSSE3-LABEL: ult_15_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -6669,11 +5726,11 @@
 ; SSSE3-NEXT:    psllw $8, %xmm1
 ; SSSE3-NEXT:    paddb %xmm3, %xmm1
 ; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [14,14,14,14,14,14,14,14]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ult_14_v8i16:
+; SSE41-LABEL: ult_15_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -6689,11 +5746,11 @@
 ; SSE41-NEXT:    psllw $8, %xmm1
 ; SSE41-NEXT:    paddb %xmm3, %xmm1
 ; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [14,14,14,14,14,14,14,14]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ult_14_v8i16:
+; AVX1-LABEL: ult_15_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -6706,11 +5763,11 @@
 ; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [14,14,14,14,14,14,14,14]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_14_v8i16:
+; AVX2-LABEL: ult_15_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -6723,53 +5780,53 @@
 ; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [14,14,14,14,14,14,14,14]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_14_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ult_15_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [14,14,14,14,14,14,14,14]
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_14_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ult_15_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [14,14,14,14,14,14,14,14]
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15]
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_14_v8i16:
+; BITALG_NOVLX-LABEL: ult_15_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [14,14,14,14,14,14,14,14]
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15]
 ; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_14_v8i16:
+; BITALG-LABEL: ult_15_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 14, i16 14, i16 14, i16 14, i16 14, i16 14, i16 14, i16 14>
+  %3 = icmp ult <8 x i16> %2, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ugt_14_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_14_v8i16:
+define <8 x i16> @ugt_15_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ugt_15_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -6792,7 +5849,7 @@
 ; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_14_v8i16:
+; SSE3-LABEL: ugt_15_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -6815,7 +5872,7 @@
 ; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_14_v8i16:
+; SSSE3-LABEL: ugt_15_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -6834,7 +5891,7 @@
 ; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_14_v8i16:
+; SSE41-LABEL: ugt_15_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -6853,7 +5910,7 @@
 ; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_14_v8i16:
+; AVX1-LABEL: ugt_15_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -6869,7 +5926,7 @@
 ; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_14_v8i16:
+; AVX2-LABEL: ugt_15_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -6885,7 +5942,7 @@
 ; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_14_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ugt_15_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
@@ -6894,7 +5951,7 @@
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_14_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ugt_15_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
@@ -6903,7 +5960,7 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_14_v8i16:
+; BITALG_NOVLX-LABEL: ugt_15_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
@@ -6911,20 +5968,20 @@
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_14_v8i16:
+; BITALG-LABEL: ugt_15_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 14, i16 14, i16 14, i16 14, i16 14, i16 14, i16 14, i16 14>
+  %3 = icmp ugt <8 x i16> %2, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ult_15_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_15_v8i16:
+define <8 x i16> @ult_16_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ult_16_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -6944,11 +6001,11 @@
 ; SSE2-NEXT:    psllw $8, %xmm2
 ; SSE2-NEXT:    paddb %xmm1, %xmm2
 ; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [15,15,15,15,15,15,15,15]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [16,16,16,16,16,16,16,16]
 ; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ult_15_v8i16:
+; SSE3-LABEL: ult_16_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -6968,11 +6025,11 @@
 ; SSE3-NEXT:    psllw $8, %xmm2
 ; SSE3-NEXT:    paddb %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [15,15,15,15,15,15,15,15]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [16,16,16,16,16,16,16,16]
 ; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_15_v8i16:
+; SSSE3-LABEL: ult_16_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -6988,11 +6045,11 @@
 ; SSSE3-NEXT:    psllw $8, %xmm1
 ; SSSE3-NEXT:    paddb %xmm3, %xmm1
 ; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [15,15,15,15,15,15,15,15]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [16,16,16,16,16,16,16,16]
 ; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ult_15_v8i16:
+; SSE41-LABEL: ult_16_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -7008,11 +6065,11 @@
 ; SSE41-NEXT:    psllw $8, %xmm1
 ; SSE41-NEXT:    paddb %xmm3, %xmm1
 ; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [15,15,15,15,15,15,15,15]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [16,16,16,16,16,16,16,16]
 ; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ult_15_v8i16:
+; AVX1-LABEL: ult_16_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -7025,11 +6082,11 @@
 ; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [16,16,16,16,16,16,16,16]
 ; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_15_v8i16:
+; AVX2-LABEL: ult_16_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -7042,53 +6099,53 @@
 ; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [16,16,16,16,16,16,16,16]
 ; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_15_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ult_16_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15]
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [16,16,16,16,16,16,16,16]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_15_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ult_16_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15]
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [16,16,16,16,16,16,16,16]
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_15_v8i16:
+; BITALG_NOVLX-LABEL: ult_16_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15]
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [16,16,16,16,16,16,16,16]
 ; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_15_v8i16:
+; BITALG-LABEL: ult_16_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>
+  %3 = icmp ult <8 x i16> %2, <i16 16, i16 16, i16 16, i16 16, i16 16, i16 16, i16 16, i16 16>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ugt_15_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_15_v8i16:
+define <8 x i16> @ugt_16_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ugt_16_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -7111,7 +6168,7 @@
 ; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_15_v8i16:
+; SSE3-LABEL: ugt_16_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -7134,7 +6191,7 @@
 ; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_15_v8i16:
+; SSSE3-LABEL: ugt_16_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -7153,7 +6210,7 @@
 ; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_15_v8i16:
+; SSE41-LABEL: ugt_16_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -7172,7 +6229,7 @@
 ; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_15_v8i16:
+; AVX1-LABEL: ugt_16_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -7188,7 +6245,7 @@
 ; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_15_v8i16:
+; AVX2-LABEL: ugt_16_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -7204,7 +6261,7 @@
 ; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_15_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ugt_16_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
@@ -7213,7 +6270,7 @@
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_15_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ugt_16_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
@@ -7222,7 +6279,7 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_15_v8i16:
+; BITALG_NOVLX-LABEL: ugt_16_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
@@ -7230,20 +6287,20 @@
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_15_v8i16:
+; BITALG-LABEL: ugt_16_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>
+  %3 = icmp ugt <8 x i16> %2, <i16 16, i16 16, i16 16, i16 16, i16 16, i16 16, i16 16, i16 16>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ult_16_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_16_v8i16:
+define <8 x i16> @ult_17_v8i16(<8 x i16> %0) {
+; SSE2-LABEL: ult_17_v8i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrlw $1, %xmm1
@@ -7263,11 +6320,11 @@
 ; SSE2-NEXT:    psllw $8, %xmm2
 ; SSE2-NEXT:    paddb %xmm1, %xmm2
 ; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [16,16,16,16,16,16,16,16]
+; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [17,17,17,17,17,17,17,17]
 ; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
-; SSE3-LABEL: ult_16_v8i16:
+; SSE3-LABEL: ult_17_v8i16:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movdqa %xmm0, %xmm1
 ; SSE3-NEXT:    psrlw $1, %xmm1
@@ -7287,11 +6344,11 @@
 ; SSE3-NEXT:    psllw $8, %xmm2
 ; SSE3-NEXT:    paddb %xmm1, %xmm2
 ; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [16,16,16,16,16,16,16,16]
+; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [17,17,17,17,17,17,17,17]
 ; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_16_v8i16:
+; SSSE3-LABEL: ult_17_v8i16:
 ; SSSE3:       # %bb.0:
 ; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
@@ -7307,11 +6364,11 @@
 ; SSSE3-NEXT:    psllw $8, %xmm1
 ; SSSE3-NEXT:    paddb %xmm3, %xmm1
 ; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [16,16,16,16,16,16,16,16]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [17,17,17,17,17,17,17,17]
 ; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
-; SSE41-LABEL: ult_16_v8i16:
+; SSE41-LABEL: ult_17_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
@@ -7327,11 +6384,11 @@
 ; SSE41-NEXT:    psllw $8, %xmm1
 ; SSE41-NEXT:    paddb %xmm3, %xmm1
 ; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [16,16,16,16,16,16,16,16]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [17,17,17,17,17,17,17,17]
 ; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: ult_16_v8i16:
+; AVX1-LABEL: ult_17_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -7344,11 +6401,11 @@
 ; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [16,16,16,16,16,16,16,16]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [17,17,17,17,17,17,17,17]
 ; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_16_v8i16:
+; AVX2-LABEL: ult_17_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
@@ -7361,1645 +6418,472 @@
 ; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
 ; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [16,16,16,16,16,16,16,16]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [17,17,17,17,17,17,17,17]
 ; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_16_v8i16:
+; AVX512VPOPCNTDQ-LABEL: ult_17_v8i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [16,16,16,16,16,16,16,16]
+; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [17,17,17,17,17,17,17,17]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_16_v8i16:
+; AVX512VPOPCNTDQVL-LABEL: ult_17_v8i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [16,16,16,16,16,16,16,16]
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [17,17,17,17,17,17,17,17]
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_16_v8i16:
+; BITALG_NOVLX-LABEL: ult_17_v8i16:
 ; BITALG_NOVLX:       # %bb.0:
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [16,16,16,16,16,16,16,16]
+; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [17,17,17,17,17,17,17,17]
 ; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
 ; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_16_v8i16:
+; BITALG-LABEL: ult_17_v8i16:
 ; BITALG:       # %bb.0:
 ; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
 ; BITALG-NEXT:    vpmovm2w %k0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 16, i16 16, i16 16, i16 16, i16 16, i16 16, i16 16, i16 16>
+  %3 = icmp ult <8 x i16> %2, <i16 17, i16 17, i16 17, i16 17, i16 17, i16 17, i16 17, i16 17>
   %4 = sext <8 x i1> %3 to <8 x i16>
   ret <8 x i16> %4
 }
 
-define <8 x i16> @ugt_16_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ugt_16_v8i16:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    psllw $8, %xmm0
-; SSE2-NEXT:    paddb %xmm1, %xmm0
-; SSE2-NEXT:    psrlw $8, %xmm0
-; SSE2-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ugt_16_v8i16:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    psllw $8, %xmm0
-; SSE3-NEXT:    paddb %xmm1, %xmm0
-; SSE3-NEXT:    psrlw $8, %xmm0
-; SSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ugt_16_v8i16:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa %xmm3, %xmm0
-; SSSE3-NEXT:    psllw $8, %xmm0
-; SSSE3-NEXT:    paddb %xmm3, %xmm0
-; SSSE3-NEXT:    psrlw $8, %xmm0
-; SSSE3-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ugt_16_v8i16:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa %xmm3, %xmm0
-; SSE41-NEXT:    psllw $8, %xmm0
-; SSE41-NEXT:    paddb %xmm3, %xmm0
-; SSE41-NEXT:    psrlw $8, %xmm0
-; SSE41-NEXT:    pcmpgtw {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    retq
-;
-; AVX1-LABEL: ugt_16_v8i16:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: ugt_16_v8i16:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
-; AVX2-NEXT:    retq
-;
-; AVX512VPOPCNTDQ-LABEL: ugt_16_v8i16:
-; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
-; AVX512VPOPCNTDQ-NEXT:    vzeroupper
-; AVX512VPOPCNTDQ-NEXT:    retq
-;
-; AVX512VPOPCNTDQVL-LABEL: ugt_16_v8i16:
-; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
-; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
-; AVX512VPOPCNTDQVL-NEXT:    retq
-;
-; BITALG_NOVLX-LABEL: ugt_16_v8i16:
-; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vpcmpgtw {{.*}}(%rip), %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vzeroupper
-; BITALG_NOVLX-NEXT:    retq
-;
-; BITALG-LABEL: ugt_16_v8i16:
-; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
-; BITALG-NEXT:    vpcmpnleuw {{.*}}(%rip), %xmm0, %k0
-; BITALG-NEXT:    vpmovm2w %k0, %xmm0
-; BITALG-NEXT:    retq
-  %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ugt <8 x i16> %2, <i16 16, i16 16, i16 16, i16 16, i16 16, i16 16, i16 16, i16 16>
-  %4 = sext <8 x i1> %3 to <8 x i16>
-  ret <8 x i16> %4
-}
-
-define <8 x i16> @ult_17_v8i16(<8 x i16> %0) {
-; SSE2-LABEL: ult_17_v8i16:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    psllw $8, %xmm2
-; SSE2-NEXT:    paddb %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $8, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [17,17,17,17,17,17,17,17]
-; SSE2-NEXT:    pcmpgtw %xmm2, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ult_17_v8i16:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    psllw $8, %xmm2
-; SSE3-NEXT:    paddb %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $8, %xmm2
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [17,17,17,17,17,17,17,17]
-; SSE3-NEXT:    pcmpgtw %xmm2, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ult_17_v8i16:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    movdqa %xmm3, %xmm1
-; SSSE3-NEXT:    psllw $8, %xmm1
-; SSSE3-NEXT:    paddb %xmm3, %xmm1
-; SSSE3-NEXT:    psrlw $8, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [17,17,17,17,17,17,17,17]
-; SSSE3-NEXT:    pcmpgtw %xmm1, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ult_17_v8i16:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    movdqa %xmm3, %xmm1
-; SSE41-NEXT:    psllw $8, %xmm1
-; SSE41-NEXT:    paddb %xmm3, %xmm1
-; SSE41-NEXT:    psrlw $8, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [17,17,17,17,17,17,17,17]
-; SSE41-NEXT:    pcmpgtw %xmm1, %xmm0
-; SSE41-NEXT:    retq
-;
-; AVX1-LABEL: ult_17_v8i16:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [17,17,17,17,17,17,17,17]
-; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: ult_17_v8i16:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX2-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [17,17,17,17,17,17,17,17]
-; AVX2-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    retq
-;
-; AVX512VPOPCNTDQ-LABEL: ult_17_v8i16:
-; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} xmm1 = [17,17,17,17,17,17,17,17]
-; AVX512VPOPCNTDQ-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
-; AVX512VPOPCNTDQ-NEXT:    vzeroupper
-; AVX512VPOPCNTDQ-NEXT:    retq
-;
-; AVX512VPOPCNTDQVL-LABEL: ult_17_v8i16:
-; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %ymm0, %ymm0
-; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %ymm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} xmm1 = [17,17,17,17,17,17,17,17]
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vzeroupper
-; AVX512VPOPCNTDQVL-NEXT:    retq
-;
-; BITALG_NOVLX-LABEL: ult_17_v8i16:
-; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntw %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vmovdqa {{.*#+}} xmm1 = [17,17,17,17,17,17,17,17]
-; BITALG_NOVLX-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
-; BITALG_NOVLX-NEXT:    vzeroupper
-; BITALG_NOVLX-NEXT:    retq
-;
-; BITALG-LABEL: ult_17_v8i16:
-; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntw %xmm0, %xmm0
-; BITALG-NEXT:    vpcmpltuw {{.*}}(%rip), %xmm0, %k0
-; BITALG-NEXT:    vpmovm2w %k0, %xmm0
-; BITALG-NEXT:    retq
-  %2 = tail call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %0)
-  %3 = icmp ult <8 x i16> %2, <i16 17, i16 17, i16 17, i16 17, i16 17, i16 17, i16 17, i16 17>
-  %4 = sext <8 x i1> %3 to <8 x i16>
-  ret <8 x i16> %4
-}
-
-define <4 x i32> @ult_0_v4i32(<4 x i32> %0) {
-; SSE-LABEL: ult_0_v4i32:
-; SSE:       # %bb.0:
-; SSE-NEXT:    xorps %xmm0, %xmm0
-; SSE-NEXT:    retq
-;
-; AVX-LABEL: ult_0_v4i32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; AVX-NEXT:    retq
-;
-; BITALG_NOVLX-LABEL: ult_0_v4i32:
-; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    retq
-;
-; BITALG-LABEL: ult_0_v4i32:
-; BITALG:       # %bb.0:
-; BITALG-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; BITALG-NEXT:    retq
-  %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
-  %3 = icmp ult <4 x i32> %2, <i32 0, i32 0, i32 0, i32 0>
-  %4 = sext <4 x i1> %3 to <4 x i32>
-  ret <4 x i32> %4
-}
-
-define <4 x i32> @ugt_0_v4i32(<4 x i32> %0) {
-; SSE2-LABEL: ugt_0_v4i32:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE2-NEXT:    psadbw %xmm0, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    packuswb %xmm2, %xmm1
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSE2-NEXT:    pxor %xmm0, %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ugt_0_v4i32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE3-NEXT:    psadbw %xmm0, %xmm2
-; SSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    packuswb %xmm2, %xmm1
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSE3-NEXT:    pxor %xmm0, %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ugt_0_v4i32:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    movdqa %xmm3, %xmm1
-; SSSE3-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    packuswb %xmm1, %xmm3
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm3
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSSE3-NEXT:    pxor %xmm3, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ugt_0_v4i32:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    pmovzxdq {{.*#+}} xmm1 = xmm3[0],zero,xmm3[1],zero
-; SSE41-NEXT:    punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    psadbw %xmm0, %xmm1
-; SSE41-NEXT:    packuswb %xmm3, %xmm1
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    retq
-;
-; AVX1-LABEL: ugt_0_v4i32:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: ugt_0_v4i32:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    retq
-;
-; AVX512VPOPCNTDQ-LABEL: ugt_0_v4i32:
-; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512VPOPCNTDQ-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; AVX512VPOPCNTDQ-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
-; AVX512VPOPCNTDQ-NEXT:    vzeroupper
-; AVX512VPOPCNTDQ-NEXT:    retq
-;
-; AVX512VPOPCNTDQVL-LABEL: ugt_0_v4i32:
-; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    retq
-;
-; BITALG_NOVLX-LABEL: ugt_0_v4i32:
-; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG_NOVLX-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
-; BITALG_NOVLX-NEXT:    vzeroupper
-; BITALG_NOVLX-NEXT:    retq
-;
-; BITALG-LABEL: ugt_0_v4i32:
-; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %xmm0, %xmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; BITALG-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0
-; BITALG-NEXT:    retq
-  %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
-  %3 = icmp ugt <4 x i32> %2, <i32 0, i32 0, i32 0, i32 0>
-  %4 = sext <4 x i1> %3 to <4 x i32>
-  ret <4 x i32> %4
-}
-
-define <4 x i32> @ult_1_v4i32(<4 x i32> %0) {
-; SSE2-LABEL: ult_1_v4i32:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE2-NEXT:    psadbw %xmm0, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    packuswb %xmm2, %xmm1
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ult_1_v4i32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE3-NEXT:    psadbw %xmm0, %xmm2
-; SSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    packuswb %xmm2, %xmm1
-; SSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ult_1_v4i32:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm3
-; SSSE3-NEXT:    pand %xmm2, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm1, %xmm4
-; SSSE3-NEXT:    pshufb %xmm3, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm2, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm1
-; SSSE3-NEXT:    paddb %xmm4, %xmm1
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm2
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSSE3-NEXT:    packuswb %xmm2, %xmm1
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ult_1_v4i32:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm1, %xmm1
-; SSE41-NEXT:    pmovzxdq {{.*#+}} xmm0 = xmm3[0],zero,xmm3[1],zero
-; SSE41-NEXT:    punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]
-; SSE41-NEXT:    psadbw %xmm1, %xmm3
-; SSE41-NEXT:    psadbw %xmm1, %xmm0
-; SSE41-NEXT:    packuswb %xmm3, %xmm0
-; SSE41-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE41-NEXT:    retq
-;
-; AVX1-LABEL: ult_1_v4i32:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: ult_1_v4i32:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    retq
-;
-; AVX512VPOPCNTDQ-LABEL: ult_1_v4i32:
-; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512VPOPCNTDQ-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; AVX512VPOPCNTDQ-NEXT:    vzeroupper
-; AVX512VPOPCNTDQ-NEXT:    retq
-;
-; AVX512VPOPCNTDQVL-LABEL: ult_1_v4i32:
-; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    retq
-;
-; BITALG_NOVLX-LABEL: ult_1_v4i32:
-; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG_NOVLX-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vzeroupper
-; BITALG_NOVLX-NEXT:    retq
-;
-; BITALG-LABEL: ult_1_v4i32:
-; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %xmm0, %xmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; BITALG-NEXT:    retq
-  %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
-  %3 = icmp ult <4 x i32> %2, <i32 1, i32 1, i32 1, i32 1>
-  %4 = sext <4 x i1> %3 to <4 x i32>
-  ret <4 x i32> %4
-}
-
-define <4 x i32> @ugt_1_v4i32(<4 x i32> %0) {
-; SSE-LABEL: ugt_1_v4i32:
-; SSE:       # %bb.0:
-; SSE-NEXT:    pcmpeqd %xmm2, %xmm2
-; SSE-NEXT:    movdqa %xmm0, %xmm1
-; SSE-NEXT:    paddd %xmm2, %xmm1
-; SSE-NEXT:    pand %xmm0, %xmm1
-; SSE-NEXT:    pxor %xmm0, %xmm0
-; SSE-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE-NEXT:    pxor %xmm2, %xmm1
-; SSE-NEXT:    movdqa %xmm1, %xmm0
-; SSE-NEXT:    retq
-;
-; AVX1-LABEL: ugt_1_v4i32:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
-; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: ugt_1_v4i32:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
-; AVX2-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    retq
-;
-; AVX512VPOPCNTDQ-LABEL: ugt_1_v4i32:
-; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [1,1,1,1]
-; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
-; AVX512VPOPCNTDQ-NEXT:    vzeroupper
-; AVX512VPOPCNTDQ-NEXT:    retq
-;
-; AVX512VPOPCNTDQVL-LABEL: ugt_1_v4i32:
-; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpnleud {{.*}}(%rip){1to4}, %xmm0, %k1
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
-; AVX512VPOPCNTDQVL-NEXT:    retq
-;
-; BITALG_NOVLX-LABEL: ugt_1_v4i32:
-; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
-; BITALG_NOVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
-; BITALG_NOVLX-NEXT:    vzeroupper
-; BITALG_NOVLX-NEXT:    retq
-;
-; BITALG-LABEL: ugt_1_v4i32:
-; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
-; BITALG-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; BITALG-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0
-; BITALG-NEXT:    retq
-  %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
-  %3 = icmp ugt <4 x i32> %2, <i32 1, i32 1, i32 1, i32 1>
-  %4 = sext <4 x i1> %3 to <4 x i32>
-  ret <4 x i32> %4
-}
-
-define <4 x i32> @ult_2_v4i32(<4 x i32> %0) {
-; SSE-LABEL: ult_2_v4i32:
+define <4 x i32> @ult_0_v4i32(<4 x i32> %0) {
+; SSE-LABEL: ult_0_v4i32:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
-; SSE-NEXT:    paddd %xmm0, %xmm1
-; SSE-NEXT:    pand %xmm1, %xmm0
-; SSE-NEXT:    pxor %xmm1, %xmm1
-; SSE-NEXT:    pcmpeqd %xmm1, %xmm0
+; SSE-NEXT:    xorps %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: ult_2_v4i32:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: ult_2_v4i32:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    retq
-;
-; AVX512VPOPCNTDQ-LABEL: ult_2_v4i32:
-; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2,2,2,2]
-; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
-; AVX512VPOPCNTDQ-NEXT:    vzeroupper
-; AVX512VPOPCNTDQ-NEXT:    retq
-;
-; AVX512VPOPCNTDQVL-LABEL: ult_2_v4i32:
-; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpltud {{.*}}(%rip){1to4}, %xmm0, %k1
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
-; AVX512VPOPCNTDQVL-NEXT:    retq
+; AVX-LABEL: ult_0_v4i32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_2_v4i32:
+; BITALG_NOVLX-LABEL: ult_0_v4i32:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
-; BITALG_NOVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_2_v4i32:
+; BITALG-LABEL: ult_0_v4i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
-; BITALG-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; BITALG-NEXT:    vxorps %xmm0, %xmm0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
-  %3 = icmp ult <4 x i32> %2, <i32 2, i32 2, i32 2, i32 2>
+  %3 = icmp ult <4 x i32> %2, <i32 0, i32 0, i32 0, i32 0>
   %4 = sext <4 x i1> %3 to <4 x i32>
   ret <4 x i32> %4
 }
 
-define <4 x i32> @ugt_2_v4i32(<4 x i32> %0) {
-; SSE2-LABEL: ugt_2_v4i32:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE2-NEXT:    psadbw %xmm0, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    packuswb %xmm2, %xmm1
-; SSE2-NEXT:    pcmpgtd {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ugt_2_v4i32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE3-NEXT:    psadbw %xmm0, %xmm2
-; SSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    packuswb %xmm2, %xmm1
-; SSE3-NEXT:    pcmpgtd {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ugt_2_v4i32:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm3
-; SSSE3-NEXT:    pand %xmm2, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm1, %xmm4
-; SSSE3-NEXT:    pshufb %xmm3, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm2, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm1
-; SSSE3-NEXT:    paddb %xmm4, %xmm1
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm2
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSSE3-NEXT:    packuswb %xmm2, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{.*}}(%rip), %xmm1
-; SSSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: ugt_2_v4i32:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm1, %xmm1
-; SSE41-NEXT:    pmovzxdq {{.*#+}} xmm0 = xmm3[0],zero,xmm3[1],zero
-; SSE41-NEXT:    punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]
-; SSE41-NEXT:    psadbw %xmm1, %xmm3
-; SSE41-NEXT:    psadbw %xmm1, %xmm0
-; SSE41-NEXT:    packuswb %xmm3, %xmm0
-; SSE41-NEXT:    pcmpgtd {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    retq
+define <4 x i32> @ugt_0_v4i32(<4 x i32> %0) {
+; SSE-LABEL: ugt_0_v4i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm0
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_2_v4i32:
+; AVX1-LABEL: ugt_0_v4i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtd {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_2_v4i32:
+; AVX2-LABEL: ugt_0_v4i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2,2,2,2]
-; AVX2-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_2_v4i32:
+; AVX512VPOPCNTDQ-LABEL: ugt_0_v4i32:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2,2,2,2]
-; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
-; AVX512VPOPCNTDQ-NEXT:    vzeroupper
-; AVX512VPOPCNTDQ-NEXT:    retq
-;
-; AVX512VPOPCNTDQVL-LABEL: ugt_2_v4i32:
-; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpnleud {{.*}}(%rip){1to4}, %xmm0, %k1
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
-; AVX512VPOPCNTDQVL-NEXT:    retq
-;
-; BITALG_NOVLX-LABEL: ugt_2_v4i32:
-; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG_NOVLX-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2,2,2,2]
-; BITALG_NOVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vzeroupper
-; BITALG_NOVLX-NEXT:    retq
-;
-; BITALG-LABEL: ugt_2_v4i32:
-; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %xmm0, %xmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG-NEXT:    vpcmpnleud {{.*}}(%rip){1to4}, %xmm0, %k1
-; BITALG-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
-; BITALG-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
-; BITALG-NEXT:    retq
-  %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
-  %3 = icmp ugt <4 x i32> %2, <i32 2, i32 2, i32 2, i32 2>
-  %4 = sext <4 x i1> %3 to <4 x i32>
-  ret <4 x i32> %4
-}
-
-define <4 x i32> @ult_3_v4i32(<4 x i32> %0) {
-; SSE2-LABEL: ult_3_v4i32:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE2-NEXT:    psadbw %xmm0, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    packuswb %xmm2, %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3]
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ult_3_v4i32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE3-NEXT:    psadbw %xmm0, %xmm2
-; SSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    packuswb %xmm2, %xmm1
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3]
-; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; AVX512VPOPCNTDQ-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX512VPOPCNTDQ-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; AVX512VPOPCNTDQ-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
+; AVX512VPOPCNTDQ-NEXT:    vzeroupper
+; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_3_v4i32:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    movdqa %xmm3, %xmm1
-; SSSE3-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    packuswb %xmm1, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    retq
+; AVX512VPOPCNTDQVL-LABEL: ugt_0_v4i32:
+; AVX512VPOPCNTDQVL:       # %bb.0:
+; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; SSE41-LABEL: ult_3_v4i32:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    pmovzxdq {{.*#+}} xmm1 = xmm3[0],zero,xmm3[1],zero
-; SSE41-NEXT:    punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    psadbw %xmm0, %xmm1
-; SSE41-NEXT:    packuswb %xmm3, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [3,3,3,3]
-; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE41-NEXT:    retq
+; BITALG_NOVLX-LABEL: ugt_0_v4i32:
+; BITALG_NOVLX:       # %bb.0:
+; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
+; BITALG_NOVLX-NEXT:    vzeroupper
+; BITALG_NOVLX-NEXT:    retq
 ;
-; AVX1-LABEL: ult_3_v4i32:
+; BITALG-LABEL: ugt_0_v4i32:
+; BITALG:       # %bb.0:
+; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; BITALG-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0
+; BITALG-NEXT:    retq
+  %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
+  %3 = icmp ugt <4 x i32> %2, <i32 0, i32 0, i32 0, i32 0>
+  %4 = sext <4 x i1> %3 to <4 x i32>
+  ret <4 x i32> %4
+}
+
+define <4 x i32> @ult_1_v4i32(<4 x i32> %0) {
+; SSE-LABEL: ult_1_v4i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX1-LABEL: ult_1_v4i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3,3,3]
-; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_3_v4i32:
+; AVX2-LABEL: ult_1_v4i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [3,3,3,3]
-; AVX2-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_3_v4i32:
+; AVX512VPOPCNTDQ-LABEL: ult_1_v4i32:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [3,3,3,3]
-; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
+; AVX512VPOPCNTDQ-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX512VPOPCNTDQ-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_3_v4i32:
+; AVX512VPOPCNTDQVL-LABEL: ult_1_v4i32:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpltud {{.*}}(%rip){1to4}, %xmm0, %k1
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
+; AVX512VPOPCNTDQVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_3_v4i32:
+; BITALG_NOVLX-LABEL: ult_1_v4i32:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG_NOVLX-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [3,3,3,3]
-; BITALG_NOVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
-; BITALG_NOVLX-NEXT:    vzeroupper
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_3_v4i32:
+; BITALG-LABEL: ult_1_v4i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %xmm0, %xmm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG-NEXT:    vpcmpltud {{.*}}(%rip){1to4}, %xmm0, %k1
-; BITALG-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
-; BITALG-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
-  %3 = icmp ult <4 x i32> %2, <i32 3, i32 3, i32 3, i32 3>
+  %3 = icmp ult <4 x i32> %2, <i32 1, i32 1, i32 1, i32 1>
   %4 = sext <4 x i1> %3 to <4 x i32>
   ret <4 x i32> %4
 }
 
-define <4 x i32> @ugt_3_v4i32(<4 x i32> %0) {
-; SSE2-LABEL: ugt_3_v4i32:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE2-NEXT:    psadbw %xmm0, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    packuswb %xmm2, %xmm1
-; SSE2-NEXT:    pcmpgtd {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    retq
+define <4 x i32> @ugt_1_v4i32(<4 x i32> %0) {
+; SSE-LABEL: ugt_1_v4i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm2, %xmm2
+; SSE-NEXT:    movdqa %xmm0, %xmm1
+; SSE-NEXT:    paddd %xmm2, %xmm1
+; SSE-NEXT:    pand %xmm0, %xmm1
+; SSE-NEXT:    pxor %xmm0, %xmm0
+; SSE-NEXT:    pcmpeqd %xmm0, %xmm1
+; SSE-NEXT:    pxor %xmm2, %xmm1
+; SSE-NEXT:    movdqa %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_3_v4i32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE3-NEXT:    psadbw %xmm0, %xmm2
-; SSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    packuswb %xmm2, %xmm1
-; SSE3-NEXT:    pcmpgtd {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; AVX1-LABEL: ugt_1_v4i32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_3_v4i32:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm3
-; SSSE3-NEXT:    pand %xmm2, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm1, %xmm4
-; SSSE3-NEXT:    pshufb %xmm3, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm2, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm1
-; SSSE3-NEXT:    paddb %xmm4, %xmm1
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm2
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSSE3-NEXT:    packuswb %xmm2, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{.*}}(%rip), %xmm1
-; SSSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSSE3-NEXT:    retq
+; AVX2-LABEL: ugt_1_v4i32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    retq
+;
+; AVX512VPOPCNTDQ-LABEL: ugt_1_v4i32:
+; AVX512VPOPCNTDQ:       # %bb.0:
+; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
+; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
+; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [1,1,1,1]
+; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
+; AVX512VPOPCNTDQ-NEXT:    vzeroupper
+; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_3_v4i32:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm1, %xmm1
-; SSE41-NEXT:    pmovzxdq {{.*#+}} xmm0 = xmm3[0],zero,xmm3[1],zero
-; SSE41-NEXT:    punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]
-; SSE41-NEXT:    psadbw %xmm1, %xmm3
-; SSE41-NEXT:    psadbw %xmm1, %xmm0
-; SSE41-NEXT:    packuswb %xmm3, %xmm0
-; SSE41-NEXT:    pcmpgtd {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    retq
+; AVX512VPOPCNTDQVL-LABEL: ugt_1_v4i32:
+; AVX512VPOPCNTDQVL:       # %bb.0:
+; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpnleud {{.*}}(%rip){1to4}, %xmm0, %k1
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
+; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_3_v4i32:
+; BITALG_NOVLX-LABEL: ugt_1_v4i32:
+; BITALG_NOVLX:       # %bb.0:
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
+; BITALG_NOVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
+; BITALG_NOVLX-NEXT:    vzeroupper
+; BITALG_NOVLX-NEXT:    retq
+;
+; BITALG-LABEL: ugt_1_v4i32:
+; BITALG:       # %bb.0:
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
+; BITALG-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; BITALG-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0
+; BITALG-NEXT:    retq
+  %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
+  %3 = icmp ugt <4 x i32> %2, <i32 1, i32 1, i32 1, i32 1>
+  %4 = sext <4 x i1> %3 to <4 x i32>
+  ret <4 x i32> %4
+}
+
+define <4 x i32> @ult_2_v4i32(<4 x i32> %0) {
+; SSE-LABEL: ult_2_v4i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    paddd %xmm0, %xmm1
+; SSE-NEXT:    pand %xmm1, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX1-LABEL: ult_2_v4i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtd {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_3_v4i32:
+; AVX2-LABEL: ult_2_v4i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [3,3,3,3]
-; AVX2-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_3_v4i32:
+; AVX512VPOPCNTDQ-LABEL: ult_2_v4i32:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [3,3,3,3]
-; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
+; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2,2,2,2]
+; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_3_v4i32:
+; AVX512VPOPCNTDQVL-LABEL: ult_2_v4i32:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpnleud {{.*}}(%rip){1to4}, %xmm0, %k1
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpltud {{.*}}(%rip){1to4}, %xmm0, %k1
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_3_v4i32:
+; BITALG_NOVLX-LABEL: ult_2_v4i32:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
+; BITALG_NOVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG_NOVLX-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [3,3,3,3]
-; BITALG_NOVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vzeroupper
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_3_v4i32:
+; BITALG-LABEL: ult_2_v4i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %xmm0, %xmm0
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
+; BITALG-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG-NEXT:    vpcmpnleud {{.*}}(%rip){1to4}, %xmm0, %k1
-; BITALG-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
-; BITALG-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
-  %3 = icmp ugt <4 x i32> %2, <i32 3, i32 3, i32 3, i32 3>
+  %3 = icmp ult <4 x i32> %2, <i32 2, i32 2, i32 2, i32 2>
   %4 = sext <4 x i1> %3 to <4 x i32>
   ret <4 x i32> %4
 }
 
-define <4 x i32> @ult_4_v4i32(<4 x i32> %0) {
-; SSE2-LABEL: ult_4_v4i32:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE2-NEXT:    psadbw %xmm0, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    packuswb %xmm2, %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4]
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    retq
+define <4 x i32> @ugt_2_v4i32(<4 x i32> %0) {
+; SSE-LABEL: ugt_2_v4i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddd %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm0
+; SSE-NEXT:    paddd %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm2, %xmm2
+; SSE-NEXT:    pcmpeqd %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
-; SSE3-LABEL: ult_4_v4i32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE3-NEXT:    psadbw %xmm0, %xmm2
-; SSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    packuswb %xmm2, %xmm1
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4]
-; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; AVX1-LABEL: ugt_2_v4i32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_4_v4i32:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    movdqa %xmm3, %xmm1
-; SSSE3-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    packuswb %xmm1, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    retq
+; AVX2-LABEL: ugt_2_v4i32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    retq
+;
+; AVX512VPOPCNTDQ-LABEL: ugt_2_v4i32:
+; AVX512VPOPCNTDQ:       # %bb.0:
+; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
+; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
+; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2,2,2,2]
+; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
+; AVX512VPOPCNTDQ-NEXT:    vzeroupper
+; AVX512VPOPCNTDQ-NEXT:    retq
+;
+; AVX512VPOPCNTDQVL-LABEL: ugt_2_v4i32:
+; AVX512VPOPCNTDQVL:       # %bb.0:
+; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpnleud {{.*}}(%rip){1to4}, %xmm0, %k1
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
+; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; SSE41-LABEL: ult_4_v4i32:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    pmovzxdq {{.*#+}} xmm1 = xmm3[0],zero,xmm3[1],zero
-; SSE41-NEXT:    punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    psadbw %xmm0, %xmm1
-; SSE41-NEXT:    packuswb %xmm3, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [4,4,4,4]
-; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE41-NEXT:    retq
+; BITALG_NOVLX-LABEL: ugt_2_v4i32:
+; BITALG_NOVLX:       # %bb.0:
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; BITALG_NOVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
+; BITALG_NOVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
+; BITALG_NOVLX-NEXT:    vzeroupper
+; BITALG_NOVLX-NEXT:    retq
 ;
-; AVX1-LABEL: ult_4_v4i32:
+; BITALG-LABEL: ugt_2_v4i32:
+; BITALG:       # %bb.0:
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; BITALG-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
+; BITALG-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; BITALG-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0
+; BITALG-NEXT:    retq
+  %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
+  %3 = icmp ugt <4 x i32> %2, <i32 2, i32 2, i32 2, i32 2>
+  %4 = sext <4 x i1> %3 to <4 x i32>
+  ret <4 x i32> %4
+}
+
+define <4 x i32> @ult_3_v4i32(<4 x i32> %0) {
+; SSE-LABEL: ult_3_v4i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm2, %xmm2
+; SSE-NEXT:    movdqa %xmm0, %xmm1
+; SSE-NEXT:    paddd %xmm2, %xmm1
+; SSE-NEXT:    pand %xmm0, %xmm1
+; SSE-NEXT:    paddd %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm2, %xmm1
+; SSE-NEXT:    pxor %xmm0, %xmm0
+; SSE-NEXT:    pcmpeqd %xmm0, %xmm1
+; SSE-NEXT:    movdqa %xmm1, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX1-LABEL: ult_3_v4i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4,4,4]
-; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ult_4_v4i32:
+; AVX2-LABEL: ult_3_v4i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [4,4,4,4]
-; AVX2-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ult_4_v4i32:
+; AVX512VPOPCNTDQ-LABEL: ult_3_v4i32:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [4,4,4,4]
+; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [3,3,3,3]
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ult_4_v4i32:
+; AVX512VPOPCNTDQVL-LABEL: ult_3_v4i32:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpltud {{.*}}(%rip){1to4}, %xmm0, %k1
@@ -9007,367 +6891,372 @@
 ; AVX512VPOPCNTDQVL-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ult_4_v4i32:
+; BITALG_NOVLX-LABEL: ult_3_v4i32:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; BITALG_NOVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
+; BITALG_NOVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG_NOVLX-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [4,4,4,4]
-; BITALG_NOVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
-; BITALG_NOVLX-NEXT:    vzeroupper
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ult_4_v4i32:
+; BITALG-LABEL: ult_3_v4i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %xmm0, %xmm0
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; BITALG-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
+; BITALG-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG-NEXT:    vpcmpltud {{.*}}(%rip){1to4}, %xmm0, %k1
-; BITALG-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
-; BITALG-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
-  %3 = icmp ult <4 x i32> %2, <i32 4, i32 4, i32 4, i32 4>
+  %3 = icmp ult <4 x i32> %2, <i32 3, i32 3, i32 3, i32 3>
   %4 = sext <4 x i1> %3 to <4 x i32>
   ret <4 x i32> %4
 }
 
-define <4 x i32> @ugt_4_v4i32(<4 x i32> %0) {
-; SSE2-LABEL: ugt_4_v4i32:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE2-NEXT:    psadbw %xmm0, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    packuswb %xmm2, %xmm1
-; SSE2-NEXT:    pcmpgtd {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    retq
+define <4 x i32> @ugt_3_v4i32(<4 x i32> %0) {
+; SSE-LABEL: ugt_3_v4i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddd %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm3
+; SSE-NEXT:    paddd %xmm1, %xmm3
+; SSE-NEXT:    pand %xmm2, %xmm3
+; SSE-NEXT:    movdqa %xmm3, %xmm0
+; SSE-NEXT:    paddd %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm3, %xmm0
+; SSE-NEXT:    pxor %xmm2, %xmm2
+; SSE-NEXT:    pcmpeqd %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
-; SSE3-LABEL: ugt_4_v4i32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE3-NEXT:    psadbw %xmm0, %xmm2
-; SSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    packuswb %xmm2, %xmm1
-; SSE3-NEXT:    pcmpgtd {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; AVX1-LABEL: ugt_3_v4i32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_4_v4i32:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm3
-; SSSE3-NEXT:    pand %xmm2, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm1, %xmm4
-; SSSE3-NEXT:    pshufb %xmm3, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm2, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm1
-; SSSE3-NEXT:    paddb %xmm4, %xmm1
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm2
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSSE3-NEXT:    packuswb %xmm2, %xmm1
-; SSSE3-NEXT:    pcmpgtd {{.*}}(%rip), %xmm1
-; SSSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSSE3-NEXT:    retq
+; AVX2-LABEL: ugt_3_v4i32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    retq
+;
+; AVX512VPOPCNTDQ-LABEL: ugt_3_v4i32:
+; AVX512VPOPCNTDQ:       # %bb.0:
+; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
+; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
+; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [3,3,3,3]
+; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
+; AVX512VPOPCNTDQ-NEXT:    vzeroupper
+; AVX512VPOPCNTDQ-NEXT:    retq
+;
+; AVX512VPOPCNTDQVL-LABEL: ugt_3_v4i32:
+; AVX512VPOPCNTDQVL:       # %bb.0:
+; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpnleud {{.*}}(%rip){1to4}, %xmm0, %k1
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
+; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; SSE41-LABEL: ugt_4_v4i32:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm1, %xmm1
-; SSE41-NEXT:    pmovzxdq {{.*#+}} xmm0 = xmm3[0],zero,xmm3[1],zero
-; SSE41-NEXT:    punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]
-; SSE41-NEXT:    psadbw %xmm1, %xmm3
-; SSE41-NEXT:    psadbw %xmm1, %xmm0
-; SSE41-NEXT:    packuswb %xmm3, %xmm0
-; SSE41-NEXT:    pcmpgtd {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    retq
+; BITALG_NOVLX-LABEL: ugt_3_v4i32:
+; BITALG_NOVLX:       # %bb.0:
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; BITALG_NOVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; BITALG_NOVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
+; BITALG_NOVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
+; BITALG_NOVLX-NEXT:    vzeroupper
+; BITALG_NOVLX-NEXT:    retq
 ;
-; AVX1-LABEL: ugt_4_v4i32:
+; BITALG-LABEL: ugt_3_v4i32:
+; BITALG:       # %bb.0:
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; BITALG-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; BITALG-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
+; BITALG-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; BITALG-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0
+; BITALG-NEXT:    retq
+  %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
+  %3 = icmp ugt <4 x i32> %2, <i32 3, i32 3, i32 3, i32 3>
+  %4 = sext <4 x i1> %3 to <4 x i32>
+  ret <4 x i32> %4
+}
+
+define <4 x i32> @ult_4_v4i32(<4 x i32> %0) {
+; SSE-LABEL: ult_4_v4i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddd %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm0
+; SSE-NEXT:    paddd %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    paddd %xmm0, %xmm1
+; SSE-NEXT:    pand %xmm1, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX1-LABEL: ult_4_v4i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtd {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: ugt_4_v4i32:
+; AVX2-LABEL: ult_4_v4i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [4,4,4,4]
-; AVX2-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
-; AVX512VPOPCNTDQ-LABEL: ugt_4_v4i32:
+; AVX512VPOPCNTDQ-LABEL: ult_4_v4i32:
 ; AVX512VPOPCNTDQ:       # %bb.0:
 ; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [4,4,4,4]
-; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
+; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; AVX512VPOPCNTDQ-NEXT:    vzeroupper
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; AVX512VPOPCNTDQVL-LABEL: ugt_4_v4i32:
+; AVX512VPOPCNTDQVL-LABEL: ult_4_v4i32:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
 ; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpcmpnleud {{.*}}(%rip){1to4}, %xmm0, %k1
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpltud {{.*}}(%rip){1to4}, %xmm0, %k1
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512VPOPCNTDQVL-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
 ; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; BITALG_NOVLX-LABEL: ugt_4_v4i32:
+; BITALG_NOVLX-LABEL: ult_4_v4i32:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; BITALG_NOVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; BITALG_NOVLX-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
+; BITALG_NOVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG_NOVLX-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [4,4,4,4]
-; BITALG_NOVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vzeroupper
+; BITALG_NOVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; BITALG_NOVLX-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_4_v4i32:
+; BITALG-LABEL: ult_4_v4i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %xmm0, %xmm0
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; BITALG-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; BITALG-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; BITALG-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
+; BITALG-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; BITALG-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; BITALG-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; BITALG-NEXT:    vpcmpnleud {{.*}}(%rip){1to4}, %xmm0, %k1
-; BITALG-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
-; BITALG-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
+; BITALG-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
-  %3 = icmp ugt <4 x i32> %2, <i32 4, i32 4, i32 4, i32 4>
+  %3 = icmp ult <4 x i32> %2, <i32 4, i32 4, i32 4, i32 4>
   %4 = sext <4 x i1> %3 to <4 x i32>
   ret <4 x i32> %4
 }
 
-define <4 x i32> @ult_5_v4i32(<4 x i32> %0) {
-; SSE2-LABEL: ult_5_v4i32:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE2-NEXT:    psadbw %xmm0, %xmm2
-; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
-; SSE2-NEXT:    packuswb %xmm2, %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm0 = [5,5,5,5]
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    retq
+define <4 x i32> @ugt_4_v4i32(<4 x i32> %0) {
+; SSE-LABEL: ugt_4_v4i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddd %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm0
+; SSE-NEXT:    paddd %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddd %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm0
+; SSE-NEXT:    paddd %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm2, %xmm2
+; SSE-NEXT:    pcmpeqd %xmm2, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
-; SSE3-LABEL: ult_5_v4i32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; SSE3-NEXT:    psadbw %xmm0, %xmm2
-; SSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSE3-NEXT:    packuswb %xmm2, %xmm1
-; SSE3-NEXT:    movdqa {{.*#+}} xmm0 = [5,5,5,5]
-; SSE3-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; AVX1-LABEL: ugt_4_v4i32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    retq
 ;
-; SSSE3-LABEL: ult_5_v4i32:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    movdqa %xmm3, %xmm1
-; SSSE3-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm1
-; SSSE3-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    packuswb %xmm1, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [5,5,5,5]
-; SSSE3-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSSE3-NEXT:    retq
+; AVX2-LABEL: ugt_4_v4i32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    retq
+;
+; AVX512VPOPCNTDQ-LABEL: ugt_4_v4i32:
+; AVX512VPOPCNTDQ:       # %bb.0:
+; AVX512VPOPCNTDQ-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
+; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
+; AVX512VPOPCNTDQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [4,4,4,4]
+; AVX512VPOPCNTDQ-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
+; AVX512VPOPCNTDQ-NEXT:    vzeroupper
+; AVX512VPOPCNTDQ-NEXT:    retq
+;
+; AVX512VPOPCNTDQVL-LABEL: ugt_4_v4i32:
+; AVX512VPOPCNTDQVL:       # %bb.0:
+; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %xmm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpnleud {{.*}}(%rip){1to4}, %xmm0, %k1
+; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
+; AVX512VPOPCNTDQVL-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
+; AVX512VPOPCNTDQVL-NEXT:    retq
 ;
-; SSE41-LABEL: ult_5_v4i32:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    pmovzxdq {{.*#+}} xmm1 = xmm3[0],zero,xmm3[1],zero
-; SSE41-NEXT:    punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    psadbw %xmm0, %xmm1
-; SSE41-NEXT:    packuswb %xmm3, %xmm1
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [5,5,5,5]
-; SSE41-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE41-NEXT:    retq
+; BITALG_NOVLX-LABEL: ugt_4_v4i32:
+; BITALG_NOVLX:       # %bb.0:
+; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
+; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
+; BITALG_NOVLX-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [4,4,4,4]
+; BITALG_NOVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
+; BITALG_NOVLX-NEXT:    vzeroupper
+; BITALG_NOVLX-NEXT:    retq
+;
+; BITALG-LABEL: ugt_4_v4i32:
+; BITALG:       # %bb.0:
+; BITALG-NEXT:    vpopcntb %xmm0, %xmm0
+; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; BITALG-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
+; BITALG-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; BITALG-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
+; BITALG-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
+; BITALG-NEXT:    vpcmpnleud {{.*}}(%rip){1to4}, %xmm0, %k1
+; BITALG-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
+; BITALG-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
+; BITALG-NEXT:    retq
+  %2 = tail call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %0)
+  %3 = icmp ugt <4 x i32> %2, <i32 4, i32 4, i32 4, i32 4>
+  %4 = sext <4 x i1> %3 to <4 x i32>
+  ret <4 x i32> %4
+}
+
+define <4 x i32> @ult_5_v4i32(<4 x i32> %0) {
+; SSE-LABEL: ult_5_v4i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    paddd %xmm1, %xmm2
+; SSE-NEXT:    pand %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm2, %xmm3
+; SSE-NEXT:    paddd %xmm1, %xmm3
+; SSE-NEXT:    pand %xmm2, %xmm3
+; SSE-NEXT:    movdqa %xmm3, %xmm0
+; SSE-NEXT:    paddd %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm3, %xmm0
+; SSE-NEXT:    paddd %xmm0, %xmm1
+; SSE-NEXT:    pand %xmm1, %xmm0
+; SSE-NEXT:    pxor %xmm1, %xmm1
+; SSE-NEXT:    pcmpeqd %xmm1, %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_5_v4i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5,5,5]
-; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_5_v4i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm1
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} xmm2 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm2, %xmm2
-; AVX2-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpackuswb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [5,5,5,5]
-; AVX2-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_5_v4i32:
@@ -20090,108 +17979,45 @@
 define <2 x i64> @ugt_0_v2i64(<2 x i64> %0) {
 ; SSE2-LABEL: ugt_0_v2i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    pxor %xmm1, %xmm1
 ; SSE2-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,0,3,2]
-; SSE2-NEXT:    pand %xmm0, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,0,3,2]
+; SSE2-NEXT:    pand %xmm1, %xmm2
 ; SSE2-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSE2-NEXT:    pxor %xmm0, %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
+; SSE2-NEXT:    pxor %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_0_v2i64:
 ; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    pxor %xmm1, %xmm1
 ; SSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,0,3,2]
-; SSE3-NEXT:    pand %xmm0, %xmm1
+; SSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,0,3,2]
+; SSE3-NEXT:    pand %xmm1, %xmm2
 ; SSE3-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSE3-NEXT:    pxor %xmm0, %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
+; SSE3-NEXT:    pxor %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_0_v2i64:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
+; SSSE3-NEXT:    pxor %xmm1, %xmm1
+; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,0,3,2]
 ; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,0,3,2]
-; SSSE3-NEXT:    pand %xmm3, %xmm1
 ; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSSE3-NEXT:    pxor %xmm1, %xmm0
+; SSSE3-NEXT:    pxor %xmm2, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_0_v2i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm3
-; SSE41-NEXT:    pcmpeqq %xmm0, %xmm3
-; SSE41-NEXT:    pcmpeqd %xmm0, %xmm0
-; SSE41-NEXT:    pxor %xmm3, %xmm0
+; SSE41-NEXT:    pxor %xmm1, %xmm1
+; SSE41-NEXT:    pcmpeqq %xmm1, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE41-NEXT:    pxor %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_0_v2i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
 ; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
@@ -20199,16 +18025,7 @@
 ;
 ; AVX2-LABEL: ugt_0_v2i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
@@ -20235,10 +18052,7 @@
 ;
 ; BITALG_NOVLX-LABEL: ugt_0_v2i64:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
 ; BITALG_NOVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; BITALG_NOVLX-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
@@ -20247,9 +18061,7 @@
 ;
 ; BITALG-LABEL: ugt_0_v2i64:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %xmm0, %xmm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; BITALG-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0
 ; BITALG-NEXT:    retq
@@ -20262,116 +18074,43 @@
 define <2 x i64> @ult_1_v2i64(<2 x i64> %0) {
 ; SSE2-LABEL: ult_1_v2i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    pxor %xmm1, %xmm1
 ; SSE2-NEXT:    pcmpeqd %xmm0, %xmm1
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,0,3,2]
-; SSE2-NEXT:    pand %xmm0, %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
+; SSE2-NEXT:    pand %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_1_v2i64:
 ; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    pxor %xmm1, %xmm1
 ; SSE3-NEXT:    pcmpeqd %xmm0, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,0,3,2]
-; SSE3-NEXT:    pand %xmm0, %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    retq
-;
-; SSSE3-LABEL: ult_1_v2i64:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,0,3,2]
+; SSE3-NEXT:    pand %xmm1, %xmm0
+; SSE3-NEXT:    retq
+;
+; SSSE3-LABEL: ult_1_v2i64:
+; SSSE3:       # %bb.0:
+; SSSE3-NEXT:    pxor %xmm1, %xmm1
+; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm1
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,0,3,2]
 ; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm0, %xmm3
-; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm3
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,0,3,2]
-; SSSE3-NEXT:    pand %xmm3, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_1_v2i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm3
-; SSE41-NEXT:    pand %xmm2, %xmm3
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm1, %xmm4
-; SSE41-NEXT:    pshufb %xmm3, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm2, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm1
-; SSE41-NEXT:    paddb %xmm4, %xmm1
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm0, %xmm1
-; SSE41-NEXT:    pcmpeqq %xmm0, %xmm1
-; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pxor %xmm1, %xmm1
+; SSE41-NEXT:    pcmpeqq %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_1_v2i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_1_v2i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
@@ -20393,19 +18132,13 @@
 ;
 ; BITALG_NOVLX-LABEL: ult_1_v2i64:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
 ; BITALG_NOVLX-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
-; BITALG_NOVLX-NEXT:    vzeroupper
 ; BITALG_NOVLX-NEXT:    retq
 ;
 ; BITALG-LABEL: ult_1_v2i64:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %xmm0, %xmm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
 ; BITALG-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; BITALG-NEXT:    retq
   %2 = tail call <2 x i64> @llvm.ctpop.v2i64(<2 x i64> %0)
@@ -20633,144 +18366,88 @@
 define <2 x i64> @ugt_2_v2i64(<2 x i64> %0) {
 ; SSE2-LABEL: ugt_2_v2i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm1, %xmm0
-; SSE2-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [2147483650,2147483650]
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE2-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    paddq %xmm1, %xmm2
+; SSE2-NEXT:    pand %xmm0, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm0
+; SSE2-NEXT:    paddq %xmm1, %xmm0
+; SSE2-NEXT:    pand %xmm2, %xmm0
+; SSE2-NEXT:    pxor %xmm2, %xmm2
+; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSE2-NEXT:    pand %xmm2, %xmm0
+; SSE2-NEXT:    pxor %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_2_v2i64:
 ; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm1, %xmm0
-; SSE3-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [2147483650,2147483650]
+; SSE3-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE3-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    pand %xmm3, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
+; SSE3-NEXT:    paddq %xmm1, %xmm2
+; SSE3-NEXT:    pand %xmm0, %xmm2
+; SSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSE3-NEXT:    paddq %xmm1, %xmm0
+; SSE3-NEXT:    pand %xmm2, %xmm0
+; SSE3-NEXT:    pxor %xmm2, %xmm2
+; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSE3-NEXT:    pand %xmm2, %xmm0
+; SSE3-NEXT:    pxor %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_2_v2i64:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm3, %xmm0
-; SSSE3-NEXT:    por {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [2147483650,2147483650]
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    paddq %xmm1, %xmm2
+; SSSE3-NEXT:    pand %xmm0, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSSE3-NEXT:    paddq %xmm1, %xmm0
+; SSSE3-NEXT:    pand %xmm2, %xmm0
+; SSSE3-NEXT:    pxor %xmm2, %xmm2
+; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm2, %xmm0
+; SSSE3-NEXT:    pxor %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_2_v2i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm3, %xmm0
-; SSE41-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [2147483650,2147483650]
+; SSE41-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm3, %xmm1
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE41-NEXT:    por %xmm1, %xmm0
+; SSE41-NEXT:    paddq %xmm1, %xmm2
+; SSE41-NEXT:    pand %xmm0, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    paddq %xmm1, %xmm0
+; SSE41-NEXT:    pand %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm2, %xmm2
+; SSE41-NEXT:    pcmpeqq %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_2_v2i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtq {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqq %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_2_v2i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpcmpgtq {{.*}}(%rip), %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqq %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_2_v2i64:
@@ -20815,148 +18492,81 @@
 }
 
 define <2 x i64> @ult_3_v2i64(<2 x i64> %0) {
-; SSE2-LABEL: ult_3_v2i64:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm1, %xmm0
-; SSE2-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [2147483651,2147483651]
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE2-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE3-LABEL: ult_3_v2i64:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
-; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm1, %xmm0
-; SSE3-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [2147483651,2147483651]
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE3-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    pand %xmm3, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
+; SSE2-LABEL: ult_3_v2i64:
+; SSE2:       # %bb.0:
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE2-NEXT:    movdqa %xmm0, %xmm2
+; SSE2-NEXT:    paddq %xmm1, %xmm2
+; SSE2-NEXT:    pand %xmm0, %xmm2
+; SSE2-NEXT:    paddq %xmm2, %xmm1
+; SSE2-NEXT:    pand %xmm2, %xmm1
+; SSE2-NEXT:    pxor %xmm2, %xmm2
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSE2-NEXT:    pand %xmm2, %xmm0
+; SSE2-NEXT:    retq
+;
+; SSE3-LABEL: ult_3_v2i64:
+; SSE3:       # %bb.0:
+; SSE3-NEXT:    pcmpeqd %xmm1, %xmm1
+; SSE3-NEXT:    movdqa %xmm0, %xmm2
+; SSE3-NEXT:    paddq %xmm1, %xmm2
+; SSE3-NEXT:    pand %xmm0, %xmm2
+; SSE3-NEXT:    paddq %xmm2, %xmm1
+; SSE3-NEXT:    pand %xmm2, %xmm1
+; SSE3-NEXT:    pxor %xmm2, %xmm2
+; SSE3-NEXT:    pcmpeqd %xmm1, %xmm2
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSE3-NEXT:    pand %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_3_v2i64:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm3, %xmm0
-; SSSE3-NEXT:    por {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [2147483651,2147483651]
-; SSSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    paddq %xmm1, %xmm2
+; SSSE3-NEXT:    pand %xmm0, %xmm2
+; SSSE3-NEXT:    paddq %xmm2, %xmm1
+; SSSE3-NEXT:    pand %xmm2, %xmm1
+; SSSE3-NEXT:    pxor %xmm2, %xmm2
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm2, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_3_v2i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
+; SSE41-NEXT:    pcmpeqd %xmm2, %xmm2
+; SSE41-NEXT:    movdqa %xmm0, %xmm1
+; SSE41-NEXT:    paddq %xmm2, %xmm1
+; SSE41-NEXT:    pand %xmm0, %xmm1
+; SSE41-NEXT:    paddq %xmm1, %xmm2
+; SSE41-NEXT:    pand %xmm2, %xmm1
 ; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm3, %xmm0
-; SSE41-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [2147483651,2147483651]
-; SSE41-NEXT:    movdqa %xmm1, %xmm2
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm3, %xmm1
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE41-NEXT:    por %xmm1, %xmm0
+; SSE41-NEXT:    pcmpeqq %xmm0, %xmm1
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_3_v2i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm1
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3]
-; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_3_v2i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm1
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3]
-; AVX2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
+; AVX2-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_3_v2i64:
@@ -21005,144 +18615,104 @@
 define <2 x i64> @ugt_3_v2i64(<2 x i64> %0) {
 ; SSE2-LABEL: ugt_3_v2i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm1, %xmm0
-; SSE2-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [2147483651,2147483651]
+; SSE2-NEXT:    paddq %xmm1, %xmm2
+; SSE2-NEXT:    pand %xmm0, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm0
+; SSE2-NEXT:    paddq %xmm1, %xmm0
+; SSE2-NEXT:    pand %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE2-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    paddq %xmm1, %xmm2
+; SSE2-NEXT:    pand %xmm0, %xmm2
+; SSE2-NEXT:    pxor %xmm3, %xmm3
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,0,3,2]
+; SSE2-NEXT:    pand %xmm3, %xmm0
+; SSE2-NEXT:    pxor %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_3_v2i64:
 ; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
+; SSE3-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm1, %xmm0
-; SSE3-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [2147483651,2147483651]
+; SSE3-NEXT:    paddq %xmm1, %xmm2
+; SSE3-NEXT:    pand %xmm0, %xmm2
+; SSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSE3-NEXT:    paddq %xmm1, %xmm0
+; SSE3-NEXT:    pand %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE3-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    pand %xmm3, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
+; SSE3-NEXT:    paddq %xmm1, %xmm2
+; SSE3-NEXT:    pand %xmm0, %xmm2
+; SSE3-NEXT:    pxor %xmm3, %xmm3
+; SSE3-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,0,3,2]
+; SSE3-NEXT:    pand %xmm3, %xmm0
+; SSE3-NEXT:    pxor %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ugt_3_v2i64:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm3, %xmm0
-; SSSE3-NEXT:    por {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [2147483651,2147483651]
+; SSSE3-NEXT:    paddq %xmm1, %xmm2
+; SSSE3-NEXT:    pand %xmm0, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSSE3-NEXT:    paddq %xmm1, %xmm0
+; SSSE3-NEXT:    pand %xmm2, %xmm0
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    paddq %xmm1, %xmm2
+; SSSE3-NEXT:    pand %xmm0, %xmm2
+; SSSE3-NEXT:    pxor %xmm3, %xmm3
+; SSSE3-NEXT:    pcmpeqd %xmm2, %xmm3
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm3, %xmm0
+; SSSE3-NEXT:    pxor %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_3_v2i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm3, %xmm0
-; SSE41-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [2147483651,2147483651]
+; SSE41-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm3, %xmm1
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE41-NEXT:    por %xmm1, %xmm0
+; SSE41-NEXT:    paddq %xmm1, %xmm2
+; SSE41-NEXT:    pand %xmm0, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm3
+; SSE41-NEXT:    paddq %xmm1, %xmm3
+; SSE41-NEXT:    pand %xmm2, %xmm3
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    paddq %xmm1, %xmm0
+; SSE41-NEXT:    pand %xmm3, %xmm0
+; SSE41-NEXT:    pxor %xmm2, %xmm2
+; SSE41-NEXT:    pcmpeqq %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_3_v2i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtq {{.*}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: ugt_3_v2i64:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpcmpgtq {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqq %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: ugt_3_v2i64:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqq %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_3_v2i64:
@@ -21189,146 +18759,94 @@
 define <2 x i64> @ult_4_v2i64(<2 x i64> %0) {
 ; SSE2-LABEL: ult_4_v2i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm1, %xmm0
-; SSE2-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [2147483652,2147483652]
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE2-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    paddq %xmm1, %xmm2
+; SSE2-NEXT:    pand %xmm0, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm0
+; SSE2-NEXT:    paddq %xmm1, %xmm0
+; SSE2-NEXT:    pand %xmm2, %xmm0
+; SSE2-NEXT:    paddq %xmm0, %xmm1
+; SSE2-NEXT:    pand %xmm0, %xmm1
+; SSE2-NEXT:    pxor %xmm2, %xmm2
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSE2-NEXT:    pand %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_4_v2i64:
 ; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
+; SSE3-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm1, %xmm0
-; SSE3-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [2147483652,2147483652]
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE3-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    pand %xmm3, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
+; SSE3-NEXT:    paddq %xmm1, %xmm2
+; SSE3-NEXT:    pand %xmm0, %xmm2
+; SSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSE3-NEXT:    paddq %xmm1, %xmm0
+; SSE3-NEXT:    pand %xmm2, %xmm0
+; SSE3-NEXT:    paddq %xmm0, %xmm1
+; SSE3-NEXT:    pand %xmm0, %xmm1
+; SSE3-NEXT:    pxor %xmm2, %xmm2
+; SSE3-NEXT:    pcmpeqd %xmm1, %xmm2
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSE3-NEXT:    pand %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_4_v2i64:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm3, %xmm0
-; SSSE3-NEXT:    por {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [2147483652,2147483652]
-; SSSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    paddq %xmm1, %xmm2
+; SSSE3-NEXT:    pand %xmm0, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSSE3-NEXT:    paddq %xmm1, %xmm0
+; SSSE3-NEXT:    pand %xmm2, %xmm0
+; SSSE3-NEXT:    paddq %xmm0, %xmm1
+; SSSE3-NEXT:    pand %xmm0, %xmm1
+; SSSE3-NEXT:    pxor %xmm2, %xmm2
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm2, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_4_v2i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
+; SSE41-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
+; SSE41-NEXT:    paddq %xmm1, %xmm2
+; SSE41-NEXT:    pand %xmm0, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    paddq %xmm1, %xmm0
+; SSE41-NEXT:    pand %xmm2, %xmm0
+; SSE41-NEXT:    paddq %xmm0, %xmm1
 ; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm3, %xmm0
-; SSE41-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [2147483652,2147483652]
-; SSE41-NEXT:    movdqa %xmm1, %xmm2
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm3, %xmm1
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE41-NEXT:    por %xmm1, %xmm0
+; SSE41-NEXT:    pxor %xmm1, %xmm1
+; SSE41-NEXT:    pcmpeqq %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_4_v2i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm1
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4]
-; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_4_v2i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm1
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4]
-; AVX2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
+; AVX2-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_4_v2i64:
@@ -21377,144 +18895,120 @@
 define <2 x i64> @ugt_4_v2i64(<2 x i64> %0) {
 ; SSE2-LABEL: ugt_4_v2i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm1, %xmm0
-; SSE2-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [2147483652,2147483652]
+; SSE2-NEXT:    paddq %xmm1, %xmm2
+; SSE2-NEXT:    pand %xmm0, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm0
+; SSE2-NEXT:    paddq %xmm1, %xmm0
+; SSE2-NEXT:    pand %xmm2, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE2-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    paddq %xmm1, %xmm2
+; SSE2-NEXT:    pand %xmm0, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm0
+; SSE2-NEXT:    paddq %xmm1, %xmm0
+; SSE2-NEXT:    pand %xmm2, %xmm0
+; SSE2-NEXT:    pxor %xmm2, %xmm2
+; SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSE2-NEXT:    pand %xmm2, %xmm0
+; SSE2-NEXT:    pxor %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ugt_4_v2i64:
 ; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
+; SSE3-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm1, %xmm0
-; SSE3-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [2147483652,2147483652]
+; SSE3-NEXT:    paddq %xmm1, %xmm2
+; SSE3-NEXT:    pand %xmm0, %xmm2
+; SSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSE3-NEXT:    paddq %xmm1, %xmm0
+; SSE3-NEXT:    pand %xmm2, %xmm0
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE3-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    pand %xmm3, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
+; SSE3-NEXT:    paddq %xmm1, %xmm2
+; SSE3-NEXT:    pand %xmm0, %xmm2
+; SSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSE3-NEXT:    paddq %xmm1, %xmm0
+; SSE3-NEXT:    pand %xmm2, %xmm0
+; SSE3-NEXT:    pxor %xmm2, %xmm2
+; SSE3-NEXT:    pcmpeqd %xmm0, %xmm2
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSE3-NEXT:    pand %xmm2, %xmm0
+; SSE3-NEXT:    pxor %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
-; SSSE3-LABEL: ugt_4_v2i64:
-; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm3, %xmm0
-; SSSE3-NEXT:    por {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [2147483652,2147483652]
+; SSSE3-LABEL: ugt_4_v2i64:
+; SSSE3:       # %bb.0:
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    paddq %xmm1, %xmm2
+; SSSE3-NEXT:    pand %xmm0, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSSE3-NEXT:    paddq %xmm1, %xmm0
+; SSSE3-NEXT:    pand %xmm2, %xmm0
+; SSSE3-NEXT:    movdqa %xmm0, %xmm2
+; SSSE3-NEXT:    paddq %xmm1, %xmm2
+; SSSE3-NEXT:    pand %xmm0, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSSE3-NEXT:    paddq %xmm1, %xmm0
+; SSSE3-NEXT:    pand %xmm2, %xmm0
+; SSSE3-NEXT:    pxor %xmm2, %xmm2
+; SSSE3-NEXT:    pcmpeqd %xmm0, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm2, %xmm0
+; SSSE3-NEXT:    pxor %xmm1, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ugt_4_v2i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
+; SSE41-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
-; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm3, %xmm0
-; SSE41-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [2147483652,2147483652]
+; SSE41-NEXT:    paddq %xmm1, %xmm2
+; SSE41-NEXT:    pand %xmm0, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    paddq %xmm1, %xmm0
+; SSE41-NEXT:    pand %xmm2, %xmm0
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pcmpgtd %xmm1, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm3, %xmm1
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE41-NEXT:    por %xmm1, %xmm0
+; SSE41-NEXT:    paddq %xmm1, %xmm2
+; SSE41-NEXT:    pand %xmm0, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    paddq %xmm1, %xmm0
+; SSE41-NEXT:    pand %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm2, %xmm2
+; SSE41-NEXT:    pcmpeqq %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ugt_4_v2i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtq {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqq %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_4_v2i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpcmpgtq {{.*}}(%rip), %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqq %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_4_v2i64:
@@ -21561,146 +19055,110 @@
 define <2 x i64> @ult_5_v2i64(<2 x i64> %0) {
 ; SSE2-LABEL: ult_5_v2i64:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $1, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    psubb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    pand %xmm1, %xmm2
-; SSE2-NEXT:    psrlw $2, %xmm0
-; SSE2-NEXT:    pand %xmm1, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrlw $4, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm0, %xmm0
-; SSE2-NEXT:    psadbw %xmm1, %xmm0
-; SSE2-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [2147483653,2147483653]
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE2-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm1, %xmm0
+; SSE2-NEXT:    paddq %xmm1, %xmm2
+; SSE2-NEXT:    pand %xmm0, %xmm2
+; SSE2-NEXT:    movdqa %xmm2, %xmm0
+; SSE2-NEXT:    paddq %xmm1, %xmm0
+; SSE2-NEXT:    pand %xmm2, %xmm0
+; SSE2-NEXT:    movdqa %xmm0, %xmm2
+; SSE2-NEXT:    paddq %xmm1, %xmm2
+; SSE2-NEXT:    pand %xmm0, %xmm2
+; SSE2-NEXT:    paddq %xmm2, %xmm1
+; SSE2-NEXT:    pand %xmm2, %xmm1
+; SSE2-NEXT:    pxor %xmm2, %xmm2
+; SSE2-NEXT:    pcmpeqd %xmm1, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSE2-NEXT:    pand %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: ult_5_v2i64:
 ; SSE3:       # %bb.0:
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $1, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    psubb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
+; SSE3-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSE3-NEXT:    pand %xmm1, %xmm2
-; SSE3-NEXT:    psrlw $2, %xmm0
-; SSE3-NEXT:    pand %xmm1, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psrlw $4, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
-; SSE3-NEXT:    pxor %xmm0, %xmm0
-; SSE3-NEXT:    psadbw %xmm1, %xmm0
-; SSE3-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm1 = [2147483653,2147483653]
-; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE3-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    pand %xmm3, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE3-NEXT:    por %xmm1, %xmm0
+; SSE3-NEXT:    paddq %xmm1, %xmm2
+; SSE3-NEXT:    pand %xmm0, %xmm2
+; SSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSE3-NEXT:    paddq %xmm1, %xmm0
+; SSE3-NEXT:    pand %xmm2, %xmm0
+; SSE3-NEXT:    movdqa %xmm0, %xmm2
+; SSE3-NEXT:    paddq %xmm1, %xmm2
+; SSE3-NEXT:    pand %xmm0, %xmm2
+; SSE3-NEXT:    paddq %xmm2, %xmm1
+; SSE3-NEXT:    pand %xmm2, %xmm1
+; SSE3-NEXT:    pxor %xmm2, %xmm2
+; SSE3-NEXT:    pcmpeqd %xmm1, %xmm2
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSE3-NEXT:    pand %xmm2, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: ult_5_v2i64:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm2
-; SSSE3-NEXT:    pand %xmm1, %xmm2
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm3, %xmm4
-; SSSE3-NEXT:    pshufb %xmm2, %xmm4
-; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm3
-; SSSE3-NEXT:    paddb %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    psadbw %xmm3, %xmm0
-; SSSE3-NEXT:    por {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [2147483653,2147483653]
-; SSSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSSE3-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    paddq %xmm1, %xmm2
+; SSSE3-NEXT:    pand %xmm0, %xmm2
+; SSSE3-NEXT:    movdqa %xmm2, %xmm0
+; SSSE3-NEXT:    paddq %xmm1, %xmm0
+; SSSE3-NEXT:    pand %xmm2, %xmm0
+; SSSE3-NEXT:    movdqa %xmm0, %xmm2
+; SSSE3-NEXT:    paddq %xmm1, %xmm2
+; SSSE3-NEXT:    pand %xmm0, %xmm2
+; SSSE3-NEXT:    paddq %xmm2, %xmm1
+; SSSE3-NEXT:    pand %xmm2, %xmm1
+; SSSE3-NEXT:    pxor %xmm2, %xmm2
+; SSSE3-NEXT:    pcmpeqd %xmm1, %xmm2
+; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,0,3,2]
+; SSSE3-NEXT:    pand %xmm2, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: ult_5_v2i64:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
+; SSE41-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    pand %xmm1, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pshufb %xmm2, %xmm4
-; SSE41-NEXT:    psrlw $4, %xmm0
+; SSE41-NEXT:    paddq %xmm1, %xmm2
+; SSE41-NEXT:    pand %xmm0, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm3
+; SSE41-NEXT:    paddq %xmm1, %xmm3
+; SSE41-NEXT:    pand %xmm2, %xmm3
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    paddq %xmm1, %xmm0
+; SSE41-NEXT:    pand %xmm3, %xmm0
+; SSE41-NEXT:    paddq %xmm0, %xmm1
 ; SSE41-NEXT:    pand %xmm1, %xmm0
-; SSE41-NEXT:    pshufb %xmm0, %xmm3
-; SSE41-NEXT:    paddb %xmm4, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm0
-; SSE41-NEXT:    psadbw %xmm3, %xmm0
-; SSE41-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [2147483653,2147483653]
-; SSE41-NEXT:    movdqa %xmm1, %xmm2
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm1, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm3, %xmm1
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE41-NEXT:    por %xmm1, %xmm0
+; SSE41-NEXT:    pxor %xmm1, %xmm1
+; SSE41-NEXT:    pcmpeqq %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: ult_5_v2i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm1, %xmm0, %xmm1
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5]
-; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_5_v2i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX2-NEXT:    vpsrlw $4, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm2
+; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vpaddq %xmm1, %xmm0, %xmm1
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX2-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5]
-; AVX2-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
+; AVX2-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_5_v2i64:
diff --git a/i/llvm/test/CodeGen/X86/vector-popcnt-256-ult-ugt.ll b/llvm/test/CodeGen/X86/vector-popcnt-256-ult-ugt.ll
--- a/i/llvm/test/CodeGen/X86/vector-popcnt-256-ult-ugt.ll
+++ b/llvm/test/CodeGen/X86/vector-popcnt-256-ult-ugt.ll
@@ -21,40 +21,18 @@
 define <32 x i8> @ugt_0_v32i8(<32 x i8> %0) {
 ; AVX1-LABEL: ugt_0_v32i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqb %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm3, %xmm3
-; AVX1-NEXT:    vpxor %xmm3, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm2, %xmm1
 ; AVX1-NEXT:    vpxor %xmm3, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vpcmpeqb %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_0_v32i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
@@ -63,14 +41,6 @@
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_0_v32i8:
 ; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512VPOPCNTDQ-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX512VPOPCNTDQ-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
 ; AVX512VPOPCNTDQ-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
@@ -79,14 +49,6 @@
 ;
 ; AVX512VPOPCNTDQVL-LABEL: ugt_0_v32i8:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQVL-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQVL-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512VPOPCNTDQVL-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512VPOPCNTDQVL-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512VPOPCNTDQVL-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512VPOPCNTDQVL-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpternlogq $15, %ymm0, %ymm0, %ymm0
@@ -118,65 +80,27 @@
 define <32 x i8> @ult_1_v32i8(<32 x i8> %0) {
 ; AVX1-LABEL: ult_1_v32i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqb %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqb %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_1_v32i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_1_v32i8:
 ; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512VPOPCNTDQ-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512VPOPCNTDQ-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512VPOPCNTDQ-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX512VPOPCNTDQ-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQVL-LABEL: ult_1_v32i8:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQVL-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512VPOPCNTDQVL-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQVL-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512VPOPCNTDQVL-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512VPOPCNTDQVL-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512VPOPCNTDQVL-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512VPOPCNTDQVL-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    retq
@@ -333,41 +257,34 @@
 define <32 x i8> @ugt_2_v32i8(<32 x i8> %0) {
 ; AVX1-LABEL: ugt_2_v32i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; AVX1-NEXT:    vpmaxub %xmm1, %xmm0, %xmm3
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm0, %xmm0
-; AVX1-NEXT:    vpmaxub %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_2_v32i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpmaxub {{.*}}(%rip), %ymm0, %ymm1
-; AVX2-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqb %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_2_v32i8:
@@ -420,40 +337,30 @@
 define <32 x i8> @ult_3_v32i8(<32 x i8> %0) {
 ; AVX1-LABEL: ult_3_v32i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2]
-; AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm3
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm0, %xmm0
-; AVX1-NEXT:    vpminub %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_3_v32i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpminub {{.*}}(%rip), %ymm0, %ymm1
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
@@ -508,41 +415,40 @@
 define <32 x i8> @ugt_3_v32i8(<32 x i8> %0) {
 ; AVX1-LABEL: ugt_3_v32i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; AVX1-NEXT:    vpmaxub %xmm1, %xmm0, %xmm3
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm0, %xmm0
-; AVX1-NEXT:    vpmaxub %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_3_v32i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpmaxub {{.*}}(%rip), %ymm0, %ymm1
-; AVX2-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqb %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_3_v32i8:
@@ -595,40 +501,36 @@
 define <32 x i8> @ult_4_v32i8(<32 x i8> %0) {
 ; AVX1-LABEL: ult_4_v32i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm3
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm0, %xmm0
-; AVX1-NEXT:    vpminub %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_4_v32i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpminub {{.*}}(%rip), %ymm0, %ymm1
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
@@ -683,41 +585,46 @@
 define <32 x i8> @ugt_4_v32i8(<32 x i8> %0) {
 ; AVX1-LABEL: ugt_4_v32i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5]
-; AVX1-NEXT:    vpmaxub %xmm1, %xmm0, %xmm3
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm0, %xmm0
-; AVX1-NEXT:    vpmaxub %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_4_v32i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpmaxub {{.*}}(%rip), %ymm0, %ymm1
-; AVX2-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqb %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_4_v32i8:
@@ -770,40 +677,42 @@
 define <32 x i8> @ult_5_v32i8(<32 x i8> %0) {
 ; AVX1-LABEL: ult_5_v32i8:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm3
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm0, %xmm0
-; AVX1-NEXT:    vpminub %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_5_v32i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpminub {{.*}}(%rip), %ymm0, %ymm1
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
@@ -1569,49 +1478,18 @@
 define <16 x i16> @ugt_0_v16i16(<16 x i16> %0) {
 ; AVX1-LABEL: ugt_0_v16i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsllw $8, %xmm2, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsrlw $8, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm3, %xmm3
-; AVX1-NEXT:    vpxor %xmm3, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqw %xmm1, %xmm2, %xmm1
 ; AVX1-NEXT:    vpxor %xmm3, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vpcmpeqw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_0_v16i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX2-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX2-NEXT:    vpsrlw $8, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
@@ -1620,9 +1498,6 @@
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_0_v16i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
-; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
 ; AVX512VPOPCNTDQ-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
 ; AVX512VPOPCNTDQ-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
@@ -1631,9 +1506,6 @@
 ;
 ; AVX512VPOPCNTDQVL-LABEL: ugt_0_v16i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %zmm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpternlogq $15, %ymm0, %ymm0, %ymm0
@@ -1665,64 +1537,27 @@
 define <16 x i16> @ult_1_v16i16(<16 x i16> %0) {
 ; AVX1-LABEL: ult_1_v16i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsllw $8, %xmm2, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsrlw $8, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqw %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqw %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_1_v16i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX2-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX2-NEXT:    vpsrlw $8, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX2-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_1_v16i16:
 ; AVX512VPOPCNTDQ:       # %bb.0:
-; AVX512VPOPCNTDQ-NEXT:    vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
-; AVX512VPOPCNTDQ-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NEXT:    vpmovdw %zmm0, %ymm0
 ; AVX512VPOPCNTDQ-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VPOPCNTDQ-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQVL-LABEL: ult_1_v16i16:
 ; AVX512VPOPCNTDQVL:       # %bb.0:
-; AVX512VPOPCNTDQVL-NEXT:    vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
-; AVX512VPOPCNTDQVL-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQVL-NEXT:    vpmovdw %zmm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VPOPCNTDQVL-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
 ; AVX512VPOPCNTDQVL-NEXT:    retq
@@ -1879,47 +1714,34 @@
 define <16 x i16> @ugt_2_v16i16(<16 x i16> %0) {
 ; AVX1-LABEL: ugt_2_v16i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsllw $8, %xmm2, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsrlw $8, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [2,2,2,2,2,2,2,2]
-; AVX1-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtw %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_2_v16i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX2-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX2-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqw %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_2_v16i16:
@@ -1960,48 +1782,31 @@
 define <16 x i16> @ult_3_v16i16(<16 x i16> %0) {
 ; AVX1-LABEL: ult_3_v16i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsllw $8, %xmm2, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsrlw $8, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3,3,3,3,3,3,3]
-; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpcmpgtw %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_3_v16i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX2-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX2-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; AVX2-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_3_v16i16:
@@ -2045,47 +1850,40 @@
 define <16 x i16> @ugt_3_v16i16(<16 x i16> %0) {
 ; AVX1-LABEL: ugt_3_v16i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsllw $8, %xmm2, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsrlw $8, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3,3,3,3,3,3,3]
-; AVX1-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtw %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_3_v16i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX2-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX2-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqw %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_3_v16i16:
@@ -2126,48 +1924,37 @@
 define <16 x i16> @ult_4_v16i16(<16 x i16> %0) {
 ; AVX1-LABEL: ult_4_v16i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsllw $8, %xmm2, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsrlw $8, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4,4,4,4,4,4,4]
-; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpcmpgtw %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_4_v16i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX2-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX2-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; AVX2-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_4_v16i16:
@@ -2211,47 +1998,46 @@
 define <16 x i16> @ugt_4_v16i16(<16 x i16> %0) {
 ; AVX1-LABEL: ugt_4_v16i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsllw $8, %xmm2, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsrlw $8, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4,4,4,4,4,4,4]
-; AVX1-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtw %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-NEXT:    retq
-;
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT:    retq
+;
 ; AVX2-LABEL: ugt_4_v16i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX2-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX2-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX2-NEXT:    vpcmpgtw {{.*}}(%rip), %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqw %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_4_v16i16:
@@ -2292,48 +2078,43 @@
 define <16 x i16> @ult_5_v16i16(<16 x i16> %0) {
 ; AVX1-LABEL: ult_5_v16i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsllw $8, %xmm2, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpsrlw $8, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5,5,5,5,5,5,5]
-; AVX1-NEXT:    vpcmpgtw %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpcmpgtw %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_5_v16i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX2-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX2-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5]
-; AVX2-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_5_v16i16:
@@ -4380,56 +4161,19 @@
 define <8 x i32> @ugt_0_v8i32(<8 x i32> %0) {
 ; AVX1-LABEL: ugt_0_v8i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm5 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm5, %xmm5
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpackuswb %xmm5, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm1 = xmm0[2],xmm4[2],xmm0[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm1, %xmm1
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpxor %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpxor %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_0_v8i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; AVX2-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
 ; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
@@ -4455,14 +4199,7 @@
 ;
 ; BITALG_NOVLX-LABEL: ugt_0_v8i32:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; BITALG_NOVLX-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; BITALG_NOVLX-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; BITALG_NOVLX-NEXT:    # kill: def $ymm0 killed $ymm0 killed $zmm0
@@ -4470,13 +4207,7 @@
 ;
 ; BITALG-LABEL: ugt_0_v8i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %ymm0, %ymm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; BITALG-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; BITALG-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
 ; BITALG-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; BITALG-NEXT:    vpternlogq $15, %ymm0, %ymm0, %ymm0
 ; BITALG-NEXT:    retq
@@ -4489,53 +4220,16 @@
 define <8 x i32> @ult_1_v8i32(<8 x i32> %0) {
 ; AVX1-LABEL: ult_1_v8i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm5 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm5, %xmm5
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpackuswb %xmm5, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm1 = xmm0[2],xmm4[2],xmm0[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm1, %xmm1
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_1_v8i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; AVX2-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
@@ -4556,26 +4250,13 @@
 ;
 ; BITALG_NOVLX-LABEL: ult_1_v8i32:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; BITALG_NOVLX-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; BITALG_NOVLX-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    retq
 ;
 ; BITALG-LABEL: ult_1_v8i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %ymm0, %ymm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; BITALG-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; BITALG-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
 ; BITALG-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i32> @llvm.ctpop.v8i32(<8 x i32> %0)
@@ -4719,56 +4400,34 @@
 define <8 x i32> @ugt_2_v8i32(<8 x i32> %0) {
 ; AVX1-LABEL: ugt_2_v8i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm5 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm5, %xmm5
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpackuswb %xmm5, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm1 = xmm0[2],xmm4[2],xmm0[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm1, %xmm1
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [2,2,2,2]
-; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_2_v8i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; AVX2-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [2,2,2,2,2,2,2,2]
-; AVX2-NEXT:    vpcmpgtd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqd %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_2_v8i32:
@@ -4789,30 +4448,27 @@
 ;
 ; BITALG_NOVLX-LABEL: ugt_2_v8i32:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; BITALG_NOVLX-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; BITALG_NOVLX-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; BITALG_NOVLX-NEXT:    vpaddd %ymm1, %ymm0, %ymm1
+; BITALG_NOVLX-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; BITALG_NOVLX-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; BITALG_NOVLX-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; BITALG_NOVLX-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [2,2,2,2,2,2,2,2]
-; BITALG_NOVLX-NEXT:    vpcmpgtd %ymm1, %ymm0, %ymm0
+; BITALG_NOVLX-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
+; BITALG_NOVLX-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    # kill: def $ymm0 killed $ymm0 killed $zmm0
 ; BITALG_NOVLX-NEXT:    retq
 ;
 ; BITALG-LABEL: ugt_2_v8i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %ymm0, %ymm0
+; BITALG-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; BITALG-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; BITALG-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; BITALG-NEXT:    vpaddd %ymm1, %ymm0, %ymm1
+; BITALG-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; BITALG-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; BITALG-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; BITALG-NEXT:    vpcmpnleud {{.*}}(%rip){1to8}, %ymm0, %k1
-; BITALG-NEXT:    vpcmpeqd %ymm0, %ymm0, %ymm0
-; BITALG-NEXT:    vmovdqa32 %ymm0, %ymm0 {%k1} {z}
+; BITALG-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
+; BITALG-NEXT:    vpternlogq $15, %ymm0, %ymm0, %ymm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i32> @llvm.ctpop.v8i32(<8 x i32> %0)
   %3 = icmp ugt <8 x i32> %2, <i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2>
@@ -4823,56 +4479,31 @@
 define <8 x i32> @ult_3_v8i32(<8 x i32> %0) {
 ; AVX1-LABEL: ult_3_v8i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm5 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm5, %xmm5
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpackuswb %xmm5, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm1 = xmm0[2],xmm4[2],xmm0[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm1, %xmm1
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3,3,3]
-; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpcmpgtd %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_3_v8i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; AVX2-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,3,3,3,3,3,3,3]
-; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_3_v8i32:
@@ -4893,30 +4524,24 @@
 ;
 ; BITALG_NOVLX-LABEL: ult_3_v8i32:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; BITALG_NOVLX-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; BITALG_NOVLX-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; BITALG_NOVLX-NEXT:    vpaddd %ymm1, %ymm0, %ymm1
+; BITALG_NOVLX-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; BITALG_NOVLX-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; BITALG_NOVLX-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; BITALG_NOVLX-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,3,3,3,3,3,3,3]
-; BITALG_NOVLX-NEXT:    vpcmpgtd %ymm0, %ymm1, %ymm0
+; BITALG_NOVLX-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    retq
 ;
 ; BITALG-LABEL: ult_3_v8i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %ymm0, %ymm0
+; BITALG-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; BITALG-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; BITALG-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; BITALG-NEXT:    vpaddd %ymm1, %ymm0, %ymm1
+; BITALG-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; BITALG-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; BITALG-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; BITALG-NEXT:    vpcmpltud {{.*}}(%rip){1to8}, %ymm0, %k1
-; BITALG-NEXT:    vpcmpeqd %ymm0, %ymm0, %ymm0
-; BITALG-NEXT:    vmovdqa32 %ymm0, %ymm0 {%k1} {z}
+; BITALG-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i32> @llvm.ctpop.v8i32(<8 x i32> %0)
   %3 = icmp ult <8 x i32> %2, <i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>
@@ -4927,56 +4552,40 @@
 define <8 x i32> @ugt_3_v8i32(<8 x i32> %0) {
 ; AVX1-LABEL: ugt_3_v8i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm5 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm5, %xmm5
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpackuswb %xmm5, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm1 = xmm0[2],xmm4[2],xmm0[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm1, %xmm1
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3,3,3]
-; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_3_v8i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; AVX2-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,3,3,3,3,3,3,3]
-; AVX2-NEXT:    vpcmpgtd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqd %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_3_v8i32:
@@ -4997,30 +4606,31 @@
 ;
 ; BITALG_NOVLX-LABEL: ugt_3_v8i32:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; BITALG_NOVLX-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; BITALG_NOVLX-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; BITALG_NOVLX-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; BITALG_NOVLX-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; BITALG_NOVLX-NEXT:    vpaddd %ymm1, %ymm0, %ymm1
+; BITALG_NOVLX-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; BITALG_NOVLX-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; BITALG_NOVLX-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; BITALG_NOVLX-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,3,3,3,3,3,3,3]
-; BITALG_NOVLX-NEXT:    vpcmpgtd %ymm1, %ymm0, %ymm0
+; BITALG_NOVLX-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
+; BITALG_NOVLX-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    # kill: def $ymm0 killed $ymm0 killed $zmm0
 ; BITALG_NOVLX-NEXT:    retq
 ;
 ; BITALG-LABEL: ugt_3_v8i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %ymm0, %ymm0
+; BITALG-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; BITALG-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; BITALG-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; BITALG-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; BITALG-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; BITALG-NEXT:    vpaddd %ymm1, %ymm0, %ymm1
+; BITALG-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; BITALG-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; BITALG-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; BITALG-NEXT:    vpcmpnleud {{.*}}(%rip){1to8}, %ymm0, %k1
-; BITALG-NEXT:    vpcmpeqd %ymm0, %ymm0, %ymm0
-; BITALG-NEXT:    vmovdqa32 %ymm0, %ymm0 {%k1} {z}
+; BITALG-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
+; BITALG-NEXT:    vpternlogq $15, %ymm0, %ymm0, %ymm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i32> @llvm.ctpop.v8i32(<8 x i32> %0)
   %3 = icmp ugt <8 x i32> %2, <i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>
@@ -5031,56 +4641,37 @@
 define <8 x i32> @ult_4_v8i32(<8 x i32> %0) {
 ; AVX1-LABEL: ult_4_v8i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm5 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm5, %xmm5
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpackuswb %xmm5, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm1 = xmm0[2],xmm4[2],xmm0[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm1, %xmm1
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4,4,4]
-; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpcmpgtd %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_4_v8i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; AVX2-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4]
-; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_4_v8i32:
@@ -5101,30 +4692,28 @@
 ;
 ; BITALG_NOVLX-LABEL: ult_4_v8i32:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
+; BITALG_NOVLX-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; BITALG_NOVLX-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; BITALG_NOVLX-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; BITALG_NOVLX-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; BITALG_NOVLX-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; BITALG_NOVLX-NEXT:    vpaddd %ymm1, %ymm0, %ymm1
+; BITALG_NOVLX-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; BITALG_NOVLX-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; BITALG_NOVLX-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; BITALG_NOVLX-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4]
-; BITALG_NOVLX-NEXT:    vpcmpgtd %ymm0, %ymm1, %ymm0
+; BITALG_NOVLX-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    retq
 ;
 ; BITALG-LABEL: ult_4_v8i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %ymm0, %ymm0
+; BITALG-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; BITALG-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; BITALG-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; BITALG-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; BITALG-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; BITALG-NEXT:    vpaddd %ymm1, %ymm0, %ymm1
+; BITALG-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; BITALG-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; BITALG-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; BITALG-NEXT:    vpcmpltud {{.*}}(%rip){1to8}, %ymm0, %k1
-; BITALG-NEXT:    vpcmpeqd %ymm0, %ymm0, %ymm0
-; BITALG-NEXT:    vmovdqa32 %ymm0, %ymm0 {%k1} {z}
+; BITALG-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; BITALG-NEXT:    retq
   %2 = tail call <8 x i32> @llvm.ctpop.v8i32(<8 x i32> %0)
   %3 = icmp ult <8 x i32> %2, <i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4>
@@ -5135,56 +4724,46 @@
 define <8 x i32> @ugt_4_v8i32(<8 x i32> %0) {
 ; AVX1-LABEL: ugt_4_v8i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm5 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm5, %xmm5
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpackuswb %xmm5, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm1 = xmm0[2],xmm4[2],xmm0[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm1, %xmm1
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4,4,4]
-; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtd %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_4_v8i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; AVX2-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4]
-; AVX2-NEXT:    vpcmpgtd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqd %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_4_v8i32:
@@ -5239,56 +4818,43 @@
 define <8 x i32> @ult_5_v8i32(<8 x i32> %0) {
 ; AVX1-LABEL: ult_5_v8i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm5 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm5, %xmm5
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpackuswb %xmm5, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpunpckhdq {{.*#+}} xmm1 = xmm0[2],xmm4[2],xmm0[3],xmm4[3]
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm1, %xmm1
-; AVX1-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5,5,5]
-; AVX1-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpcmpgtd %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_5_v8i32:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm2, %ymm2
-; AVX2-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [5,5,5,5,5,5,5,5]
-; AVX2-NEXT:    vpcmpgtd %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_5_v8i32:
@@ -11178,44 +10744,19 @@
 define <4 x i64> @ugt_0_v4i64(<4 x i64> %0) {
 ; AVX1-LABEL: ugt_0_v4i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqq %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqq %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpxor %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqq %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpxor %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqq %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_0_v4i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
 ; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
@@ -11241,10 +10782,7 @@
 ;
 ; BITALG_NOVLX-LABEL: ugt_0_v4i64:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; BITALG_NOVLX-NEXT:    # kill: def $ymm0 killed $ymm0 killed $zmm0
@@ -11252,9 +10790,7 @@
 ;
 ; BITALG-LABEL: ugt_0_v4i64:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %ymm0, %ymm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
 ; BITALG-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
 ; BITALG-NEXT:    vpternlogq $15, %ymm0, %ymm0, %ymm0
 ; BITALG-NEXT:    retq
@@ -11267,41 +10803,16 @@
 define <4 x i64> @ult_1_v4i64(<4 x i64> %0) {
 ; AVX1-LABEL: ult_1_v4i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqq %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpeqq %xmm4, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqq %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqq %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_1_v4i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
@@ -11322,18 +10833,13 @@
 ;
 ; BITALG_NOVLX-LABEL: ult_1_v4i64:
 ; BITALG_NOVLX:       # %bb.0:
-; BITALG_NOVLX-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
-; BITALG_NOVLX-NEXT:    vpopcntb %zmm0, %zmm0
 ; BITALG_NOVLX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG_NOVLX-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
 ; BITALG_NOVLX-NEXT:    retq
 ;
 ; BITALG-LABEL: ult_1_v4i64:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %ymm0, %ymm0
 ; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
 ; BITALG-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
 ; BITALG-NEXT:    retq
   %2 = tail call <4 x i64> @llvm.ctpop.v4i64(<4 x i64> %0)
@@ -11477,44 +10983,34 @@
 define <4 x i64> @ugt_2_v4i64(<4 x i64> %0) {
 ; AVX1-LABEL: ugt_2_v4i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [2,2]
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_2_v4i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [2,2,2,2]
-; AVX2-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqq %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_2_v4i64:
@@ -11561,44 +11057,31 @@
 define <4 x i64> @ult_3_v4i64(<4 x i64> %0) {
 ; AVX1-LABEL: ult_3_v4i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3]
-; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_3_v4i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [3,3,3,3]
-; AVX2-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_3_v4i64:
@@ -11645,44 +11128,40 @@
 define <4 x i64> @ugt_3_v4i64(<4 x i64> %0) {
 ; AVX1-LABEL: ugt_3_v4i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [3,3]
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_3_v4i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [3,3,3,3]
-; AVX2-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqq %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_3_v4i64:
@@ -11729,44 +11208,37 @@
 define <4 x i64> @ult_4_v4i64(<4 x i64> %0) {
 ; AVX1-LABEL: ult_4_v4i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4]
-; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_4_v4i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [4,4,4,4]
-; AVX2-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_4_v4i64:
@@ -11813,44 +11285,46 @@
 define <4 x i64> @ugt_4_v4i64(<4 x i64> %0) {
 ; AVX1-LABEL: ugt_4_v4i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,4]
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ugt_4_v4i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [4,4,4,4]
-; AVX2-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX2-NEXT:    vpcmpeqq %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ugt_4_v4i64:
@@ -11897,44 +11371,43 @@
 define <4 x i64> @ult_5_v4i64(<4 x i64> %0) {
 ; AVX1-LABEL: ult_5_v4i64:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX1-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm4
-; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
-; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [5,5]
-; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm3
+; AVX1-NEXT:    vpand %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm4
+; AVX1-NEXT:    vpand %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm2
+; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: ult_5_v4i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX2-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpsrlw $4, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm1
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [5,5,5,5]
-; AVX2-NEXT:    vpcmpgtq %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    vpcmpeqq %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-LABEL: ult_5_v4i64:
diff --git a/i/llvm/test/CodeGen/X86/vector-popcnt-512-ult-ugt.ll b/llvm/test/CodeGen/X86/vector-popcnt-512-ult-ugt.ll
--- a/i/llvm/test/CodeGen/X86/vector-popcnt-512-ult-ugt.ll
+++ b/llvm/test/CodeGen/X86/vector-popcnt-512-ult-ugt.ll
@@ -20,76 +20,32 @@
 define <64 x i8> @ugt_0_v64i8(<64 x i8> %0) {
 ; AVX512F-LABEL: ugt_0_v64i8:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512F-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpcmpeqb %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512F-NEXT:    vpcmpeqb %ymm2, %ymm1, %ymm1
+; AVX512F-NEXT:    vpcmpeqb %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_0_v64i8:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
 ; AVX512BW-NEXT:    vptestmb %zmm0, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-NOBW-LABEL: ugt_0_v64i8:
 ; AVX512VPOPCNTDQ-NOBW:       # %bb.0:
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqb %ymm1, %ymm2, %ymm1
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqb %ymm2, %ymm1, %ymm1
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqb %ymm2, %ymm0, %ymm0
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NOBW-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NOBW-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ugt_0_v64i8:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vptestmb %zmm0, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
@@ -109,74 +65,30 @@
 define <64 x i8> @ult_1_v64i8(<64 x i8> %0) {
 ; AVX512F-LABEL: ult_1_v64i8:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512F-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpcmpeqb %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512F-NEXT:    vpcmpeqb %ymm2, %ymm1, %ymm1
+; AVX512F-NEXT:    vpcmpeqb %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_1_v64i8:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
 ; AVX512BW-NEXT:    vptestnmb %zmm0, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-NOBW-LABEL: ult_1_v64i8:
 ; AVX512VPOPCNTDQ-NOBW:       # %bb.0:
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqb %ymm1, %ymm2, %ymm1
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqb %ymm2, %ymm1, %ymm1
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqb %ymm2, %ymm0, %ymm0
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NOBW-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ult_1_v64i8:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vptestnmb %zmm0, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
@@ -312,40 +224,30 @@
 define <64 x i8> @ugt_2_v64i8(<64 x i8> %0) {
 ; AVX512F-LABEL: ugt_2_v64i8:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; AVX512F-NEXT:    vpmaxub %ymm1, %ymm0, %ymm3
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512F-NEXT:    vpcmpeqb %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
 ; AVX512F-NEXT:    vpcmpeqb %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpmaxub %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vpcmpeqb %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512F-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_2_v64i8:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleub {{.*}}(%rip), %zmm0, %k0
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmb %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
@@ -376,15 +278,11 @@
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ugt_2_v64i8:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpcmpnleub {{.*}}(%rip), %zmm0, %k0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vptestmb %zmm1, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
 ;
@@ -403,40 +301,29 @@
 define <64 x i8> @ult_3_v64i8(<64 x i8> %0) {
 ; AVX512F-LABEL: ult_3_v64i8:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2]
-; AVX512F-NEXT:    vpminub %ymm1, %ymm0, %ymm3
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512F-NEXT:    vpcmpeqb %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
 ; AVX512F-NEXT:    vpcmpeqb %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpminub %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vpcmpeqb %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_3_v64i8:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltub {{.*}}(%rip), %zmm0, %k0
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmb %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
@@ -467,15 +354,11 @@
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ult_3_v64i8:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpcmpltub {{.*}}(%rip), %zmm0, %k0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vptestnmb %zmm1, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
 ;
@@ -494,40 +377,36 @@
 define <64 x i8> @ugt_3_v64i8(<64 x i8> %0) {
 ; AVX512F-LABEL: ugt_3_v64i8:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; AVX512F-NEXT:    vpmaxub %ymm1, %ymm0, %ymm3
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512F-NEXT:    vpcmpeqb %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
 ; AVX512F-NEXT:    vpcmpeqb %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpmaxub %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vpcmpeqb %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512F-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_3_v64i8:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleub {{.*}}(%rip), %zmm0, %k0
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmb %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
@@ -558,15 +437,13 @@
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ugt_3_v64i8:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpcmpnleub {{.*}}(%rip), %zmm0, %k0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vptestmb %zmm1, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
 ;
@@ -585,40 +462,35 @@
 define <64 x i8> @ult_4_v64i8(<64 x i8> %0) {
 ; AVX512F-LABEL: ult_4_v64i8:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; AVX512F-NEXT:    vpminub %ymm1, %ymm0, %ymm3
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512F-NEXT:    vpcmpeqb %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
 ; AVX512F-NEXT:    vpcmpeqb %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpminub %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vpcmpeqb %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_4_v64i8:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltub {{.*}}(%rip), %zmm0, %k0
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmb %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
@@ -649,15 +521,13 @@
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ult_4_v64i8:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpcmpltub {{.*}}(%rip), %zmm0, %k0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vptestnmb %zmm1, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
 ;
@@ -676,40 +546,42 @@
 define <64 x i8> @ugt_4_v64i8(<64 x i8> %0) {
 ; AVX512F-LABEL: ugt_4_v64i8:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5]
-; AVX512F-NEXT:    vpmaxub %ymm1, %ymm0, %ymm3
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512F-NEXT:    vpcmpeqb %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
 ; AVX512F-NEXT:    vpcmpeqb %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpmaxub %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vpcmpeqb %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512F-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_4_v64i8:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleub {{.*}}(%rip), %zmm0, %k0
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmb %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
@@ -740,15 +612,15 @@
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ugt_4_v64i8:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpcmpnleub {{.*}}(%rip), %zmm0, %k0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vptestmb %zmm1, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
 ;
@@ -767,40 +639,41 @@
 define <64 x i8> @ult_5_v64i8(<64 x i8> %0) {
 ; AVX512F-LABEL: ult_5_v64i8:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; AVX512F-NEXT:    vpminub %ymm1, %ymm0, %ymm3
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512F-NEXT:    vpcmpeqb %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddb %ymm2, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
 ; AVX512F-NEXT:    vpcmpeqb %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpminub %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vpcmpeqb %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_5_v64i8:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltub {{.*}}(%rip), %zmm0, %k0
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmb %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
@@ -831,15 +704,15 @@
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ult_5_v64i8:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpcmpltub {{.*}}(%rip), %zmm0, %k0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vptestnmb %zmm1, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2b %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
 ;
@@ -1597,80 +1470,32 @@
 define <32 x i16> @ugt_0_v32i16(<32 x i16> %0) {
 ; AVX512F-LABEL: ugt_0_v32i16:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsllw $8, %ymm2, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsrlw $8, %ymm2, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX512F-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512F-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpcmpeqw %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512F-NEXT:    vpcmpeqw %ymm2, %ymm1, %ymm1
+; AVX512F-NEXT:    vpcmpeqw %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_0_v32i16:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
 ; AVX512BW-NEXT:    vptestmw %zmm0, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-NOBW-LABEL: ugt_0_v32i16:
 ; AVX512VPOPCNTDQ-NOBW:       # %bb.0:
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpmovzxwd {{.*#+}} zmm1 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpopcntd %zmm1, %zmm1
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpmovdw %zmm1, %ymm1
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpmovdw %zmm0, %ymm0
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
 ; AVX512VPOPCNTDQ-NOBW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqw %ymm2, %ymm0, %ymm0
 ; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqw %ymm2, %ymm1, %ymm1
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqw %ymm2, %ymm0, %ymm0
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NOBW-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NOBW-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ugt_0_v32i16:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vptestmw %zmm0, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
@@ -1690,78 +1515,30 @@
 define <32 x i16> @ult_1_v32i16(<32 x i16> %0) {
 ; AVX512F-LABEL: ult_1_v32i16:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsllw $8, %ymm2, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsrlw $8, %ymm2, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX512F-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512F-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpcmpeqw %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX512F-NEXT:    vpcmpeqw %ymm2, %ymm1, %ymm1
+; AVX512F-NEXT:    vpcmpeqw %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_1_v32i16:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
 ; AVX512BW-NEXT:    vptestnmw %zmm0, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-NOBW-LABEL: ult_1_v32i16:
 ; AVX512VPOPCNTDQ-NOBW:       # %bb.0:
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpmovzxwd {{.*#+}} zmm1 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpopcntd %zmm1, %zmm1
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpmovdw %zmm1, %ymm1
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpopcntd %zmm0, %zmm0
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpmovdw %zmm0, %ymm0
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
 ; AVX512VPOPCNTDQ-NOBW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqw %ymm2, %ymm0, %ymm0
 ; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqw %ymm2, %ymm1, %ymm1
-; AVX512VPOPCNTDQ-NOBW-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vpcmpeqw %ymm2, %ymm0, %ymm0
+; AVX512VPOPCNTDQ-NOBW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-NOBW-NEXT:    retq
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ult_1_v32i16:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vptestnmw %zmm0, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
@@ -1897,47 +1674,30 @@
 define <32 x i16> @ugt_2_v32i16(<32 x i16> %0) {
 ; AVX512F-LABEL: ugt_2_v32i16:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsllw $8, %ymm2, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsrlw $8, %ymm2, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2]
-; AVX512F-NEXT:    vpcmpgtw %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpcmpgtw %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512F-NEXT:    vpcmpeqw %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT:    vpcmpeqw %ymm3, %ymm0, %ymm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512F-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_2_v32i16:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleuw {{.*}}(%rip), %zmm0, %k0
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmw %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
@@ -1958,18 +1718,11 @@
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ugt_2_v32i16:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpcmpnleuw {{.*}}(%rip), %zmm0, %k0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vptestmw %zmm1, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
 ;
@@ -1988,47 +1741,29 @@
 define <32 x i16> @ult_3_v32i16(<32 x i16> %0) {
 ; AVX512F-LABEL: ult_3_v32i16:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsllw $8, %ymm2, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsrlw $8, %ymm2, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; AVX512F-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    vpcmpgtw %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512F-NEXT:    vpcmpeqw %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT:    vpcmpeqw %ymm3, %ymm0, %ymm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_3_v32i16:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltuw {{.*}}(%rip), %zmm0, %k0
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmw %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
@@ -2049,18 +1784,11 @@
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ult_3_v32i16:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpcmpltuw {{.*}}(%rip), %zmm0, %k0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vptestnmw %zmm1, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
 ;
@@ -2079,47 +1807,36 @@
 define <32 x i16> @ugt_3_v32i16(<32 x i16> %0) {
 ; AVX512F-LABEL: ugt_3_v32i16:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsllw $8, %ymm2, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsrlw $8, %ymm2, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
-; AVX512F-NEXT:    vpcmpgtw %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpcmpgtw %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512F-NEXT:    vpcmpeqw %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT:    vpcmpeqw %ymm3, %ymm0, %ymm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512F-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_3_v32i16:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleuw {{.*}}(%rip), %zmm0, %k0
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmw %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
@@ -2140,18 +1857,13 @@
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ugt_3_v32i16:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpcmpnleuw {{.*}}(%rip), %zmm0, %k0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vptestmw %zmm1, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
 ;
@@ -2170,47 +1882,35 @@
 define <32 x i16> @ult_4_v32i16(<32 x i16> %0) {
 ; AVX512F-LABEL: ult_4_v32i16:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsllw $8, %ymm2, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsrlw $8, %ymm2, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; AVX512F-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    vpcmpgtw %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512F-NEXT:    vpcmpeqw %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT:    vpcmpeqw %ymm3, %ymm0, %ymm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_4_v32i16:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltuw {{.*}}(%rip), %zmm0, %k0
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmw %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
@@ -2231,18 +1931,13 @@
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ult_4_v32i16:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpcmpltuw {{.*}}(%rip), %zmm0, %k0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vptestnmw %zmm1, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
 ;
@@ -2261,47 +1956,42 @@
 define <32 x i16> @ugt_4_v32i16(<32 x i16> %0) {
 ; AVX512F-LABEL: ugt_4_v32i16:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsllw $8, %ymm2, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsrlw $8, %ymm2, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4]
-; AVX512F-NEXT:    vpcmpgtw %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpcmpgtw %ymm1, %ymm2, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512F-NEXT:    vpcmpeqw %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT:    vpcmpeqw %ymm3, %ymm0, %ymm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512F-NEXT:    vpternlogq $15, %zmm0, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_4_v32i16:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleuw {{.*}}(%rip), %zmm0, %k0
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmw %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
@@ -2322,18 +2012,15 @@
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ugt_4_v32i16:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpcmpnleuw {{.*}}(%rip), %zmm0, %k0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vptestmw %zmm1, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
 ;
@@ -2352,47 +2039,41 @@
 define <32 x i16> @ult_5_v32i16(<32 x i16> %0) {
 ; AVX512F-LABEL: ult_5_v32i16:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm2
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm4
-; AVX512F-NEXT:    vpand %ymm1, %ymm4, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsllw $8, %ymm2, %ymm4
-; AVX512F-NEXT:    vpaddb %ymm2, %ymm4, %ymm2
-; AVX512F-NEXT:    vpsrlw $8, %ymm2, %ymm2
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm4
-; AVX512F-NEXT:    vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsllw $8, %ymm0, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    vpsrlw $8, %ymm0, %ymm0
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm1 = [5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5]
-; AVX512F-NEXT:    vpcmpgtw %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT:    vpcmpgtw %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm1, %ymm3
+; AVX512F-NEXT:    vpand %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
+; AVX512F-NEXT:    vpcmpeqw %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm4
+; AVX512F-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT:    vpcmpeqw %ymm3, %ymm0, %ymm0
+; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_5_v32i16:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltuw {{.*}}(%rip), %zmm0, %k0
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmw %zmm1, %zmm0, %k0
 ; AVX512BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512BW-NEXT:    retq
 ;
@@ -2413,18 +2094,15 @@
 ;
 ; AVX512VPOPCNTDQ-BW-LABEL: ult_5_v32i16:
 ; AVX512VPOPCNTDQ-BW:       # %bb.0:
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsllw $8, %zmm0, %zmm1
-; AVX512VPOPCNTDQ-BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpsrlw $8, %zmm0, %zmm0
-; AVX512VPOPCNTDQ-BW-NEXT:    vpcmpltuw {{.*}}(%rip), %zmm0, %k0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm2
+; AVX512VPOPCNTDQ-BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512VPOPCNTDQ-BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm1
+; AVX512VPOPCNTDQ-BW-NEXT:    vptestnmw %zmm1, %zmm0, %k0
 ; AVX512VPOPCNTDQ-BW-NEXT:    vpmovm2w %k0, %zmm0
 ; AVX512VPOPCNTDQ-BW-NEXT:    retq
 ;
@@ -4638,53 +4316,12 @@
 define <16 x i32> @ugt_0_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ugt_0_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    vptestmd %zmm0, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_0_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
 ; AVX512BW-NEXT:    vptestmd %zmm0, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
@@ -4698,13 +4335,6 @@
 ;
 ; BITALG-LABEL: ugt_0_v16i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; BITALG-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; BITALG-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
 ; BITALG-NEXT:    vptestmd %zmm0, %zmm0, %k1
 ; BITALG-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; BITALG-NEXT:    retq
@@ -4717,53 +4347,12 @@
 define <16 x i32> @ult_1_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ult_1_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    vptestnmd %zmm0, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_1_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
 ; AVX512BW-NEXT:    vptestnmd %zmm0, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
@@ -4777,13 +4366,6 @@
 ;
 ; BITALG-LABEL: ult_1_v16i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; BITALG-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; BITALG-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
 ; BITALG-NEXT:    vptestnmd %zmm0, %zmm0, %k1
 ; BITALG-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; BITALG-NEXT:    retq
@@ -4870,54 +4452,21 @@
 define <16 x i32> @ugt_2_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ugt_2_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_2_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -4930,14 +4479,11 @@
 ;
 ; BITALG-LABEL: ugt_2_v16i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; BITALG-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; BITALG-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; BITALG-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; BITALG-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; BITALG-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; BITALG-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; BITALG-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; BITALG-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; BITALG-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; BITALG-NEXT:    retq
   %2 = tail call <16 x i32> @llvm.ctpop.v16i32(<16 x i32> %0)
@@ -4949,54 +4495,21 @@
 define <16 x i32> @ult_3_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ult_3_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_3_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -5009,14 +4522,11 @@
 ;
 ; BITALG-LABEL: ult_3_v16i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; BITALG-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; BITALG-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; BITALG-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; BITALG-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; BITALG-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; BITALG-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; BITALG-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; BITALG-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; BITALG-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; BITALG-NEXT:    retq
   %2 = tail call <16 x i32> @llvm.ctpop.v16i32(<16 x i32> %0)
@@ -5028,54 +4538,25 @@
 define <16 x i32> @ugt_3_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ugt_3_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_3_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -5088,14 +4569,13 @@
 ;
 ; BITALG-LABEL: ugt_3_v16i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; BITALG-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; BITALG-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; BITALG-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; BITALG-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; BITALG-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; BITALG-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; BITALG-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; BITALG-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; BITALG-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; BITALG-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; BITALG-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; BITALG-NEXT:    retq
   %2 = tail call <16 x i32> @llvm.ctpop.v16i32(<16 x i32> %0)
@@ -5107,54 +4587,25 @@
 define <16 x i32> @ult_4_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ult_4_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_4_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -5167,14 +4618,13 @@
 ;
 ; BITALG-LABEL: ult_4_v16i32:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; BITALG-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; BITALG-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; BITALG-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; BITALG-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; BITALG-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; BITALG-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; BITALG-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; BITALG-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; BITALG-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; BITALG-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; BITALG-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; BITALG-NEXT:    retq
   %2 = tail call <16 x i32> @llvm.ctpop.v16i32(<16 x i32> %0)
@@ -5186,54 +4636,29 @@
 define <16 x i32> @ugt_4_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ugt_4_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_4_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -5265,54 +4690,29 @@
 define <16 x i32> @ult_5_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ult_5_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_5_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -5344,54 +4744,33 @@
 define <16 x i32> @ugt_5_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ugt_5_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_5_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -5423,54 +4802,33 @@
 define <16 x i32> @ult_6_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ult_6_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_6_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -5502,54 +4860,37 @@
 define <16 x i32> @ugt_6_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ugt_6_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_6_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -5581,54 +4922,37 @@
 define <16 x i32> @ult_7_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ult_7_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_7_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -5660,54 +4984,41 @@
 define <16 x i32> @ugt_7_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ugt_7_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_7_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -5739,54 +5050,41 @@
 define <16 x i32> @ult_8_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ult_8_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_8_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -5818,54 +5116,45 @@
 define <16 x i32> @ugt_8_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ugt_8_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_8_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -5897,54 +5186,45 @@
 define <16 x i32> @ult_9_v16i32(<16 x i32> %0) {
 ; AVX512F-LABEL: ult_9_v16i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm5, %ymm5
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpackuswb %ymm5, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm2, %ymm2
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpackuswb %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_9_v16i32:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm2, %zmm2
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpackuswb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltud {{.*}}(%rip){1to16}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandd %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddd %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmd %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -9779,41 +9059,12 @@
 define <8 x i64> @ugt_0_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ugt_0_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    vptestmq %zmm0, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_0_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
 ; AVX512BW-NEXT:    vptestmq %zmm0, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
@@ -9827,9 +9078,6 @@
 ;
 ; BITALG-LABEL: ugt_0_v8i64:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
 ; BITALG-NEXT:    vptestmq %zmm0, %zmm0, %k1
 ; BITALG-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; BITALG-NEXT:    retq
@@ -9842,41 +9090,12 @@
 define <8 x i64> @ult_1_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ult_1_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
 ; AVX512F-NEXT:    vptestnmq %zmm0, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_1_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
 ; AVX512BW-NEXT:    vptestnmq %zmm0, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
@@ -9890,9 +9109,6 @@
 ;
 ; BITALG-LABEL: ult_1_v8i64:
 ; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
 ; BITALG-NEXT:    vptestnmq %zmm0, %zmm0, %k1
 ; BITALG-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; BITALG-NEXT:    retq
@@ -9979,42 +9195,21 @@
 define <8 x i64> @ugt_2_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ugt_2_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_2_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10025,59 +9220,38 @@
 ; AVX512VPOPCNTDQ-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_2_v8i64:
-; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; BITALG-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
-; BITALG-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
-; BITALG-NEXT:    retq
-  %2 = tail call <8 x i64> @llvm.ctpop.v8i64(<8 x i64> %0)
-  %3 = icmp ugt <8 x i64> %2, <i64 2, i64 2, i64 2, i64 2, i64 2, i64 2, i64 2, i64 2>
-  %4 = sext <8 x i1> %3 to <8 x i64>
-  ret <8 x i64> %4
-}
-
-define <8 x i64> @ult_3_v8i64(<8 x i64> %0) {
-; AVX512F-LABEL: ult_3_v8i64:
-; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; BITALG-LABEL: ugt_2_v8i64:
+; BITALG:       # %bb.0:
+; BITALG-NEXT:    vpopcntb %zmm0, %zmm0
+; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
+; BITALG-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; BITALG-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
+; BITALG-NEXT:    retq
+  %2 = tail call <8 x i64> @llvm.ctpop.v8i64(<8 x i64> %0)
+  %3 = icmp ugt <8 x i64> %2, <i64 2, i64 2, i64 2, i64 2, i64 2, i64 2, i64 2, i64 2>
+  %4 = sext <8 x i1> %3 to <8 x i64>
+  ret <8 x i64> %4
+}
+
+define <8 x i64> @ult_3_v8i64(<8 x i64> %0) {
+; AVX512F-LABEL: ult_3_v8i64:
+; AVX512F:       # %bb.0:
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_3_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10105,42 +9279,25 @@
 define <8 x i64> @ugt_3_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ugt_3_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_3_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10168,42 +9325,25 @@
 define <8 x i64> @ult_4_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ult_4_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_4_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10231,42 +9371,29 @@
 define <8 x i64> @ugt_4_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ugt_4_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_4_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10294,42 +9421,29 @@
 define <8 x i64> @ult_5_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ult_5_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_5_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10357,42 +9471,33 @@
 define <8 x i64> @ugt_5_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ugt_5_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_5_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10403,59 +9508,50 @@
 ; AVX512VPOPCNTDQ-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512VPOPCNTDQ-NEXT:    retq
 ;
-; BITALG-LABEL: ugt_5_v8i64:
-; BITALG:       # %bb.0:
-; BITALG-NEXT:    vpopcntb %zmm0, %zmm0
-; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; BITALG-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; BITALG-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
-; BITALG-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
-; BITALG-NEXT:    retq
-  %2 = tail call <8 x i64> @llvm.ctpop.v8i64(<8 x i64> %0)
-  %3 = icmp ugt <8 x i64> %2, <i64 5, i64 5, i64 5, i64 5, i64 5, i64 5, i64 5, i64 5>
-  %4 = sext <8 x i1> %3 to <8 x i64>
-  ret <8 x i64> %4
-}
-
-define <8 x i64> @ult_6_v8i64(<8 x i64> %0) {
-; AVX512F-LABEL: ult_6_v8i64:
-; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; BITALG-LABEL: ugt_5_v8i64:
+; BITALG:       # %bb.0:
+; BITALG-NEXT:    vpopcntb %zmm0, %zmm0
+; BITALG-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; BITALG-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
+; BITALG-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; BITALG-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
+; BITALG-NEXT:    retq
+  %2 = tail call <8 x i64> @llvm.ctpop.v8i64(<8 x i64> %0)
+  %3 = icmp ugt <8 x i64> %2, <i64 5, i64 5, i64 5, i64 5, i64 5, i64 5, i64 5, i64 5>
+  %4 = sext <8 x i1> %3 to <8 x i64>
+  ret <8 x i64> %4
+}
+
+define <8 x i64> @ult_6_v8i64(<8 x i64> %0) {
+; AVX512F-LABEL: ult_6_v8i64:
+; AVX512F:       # %bb.0:
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_6_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10483,42 +9579,37 @@
 define <8 x i64> @ugt_6_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ugt_6_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_6_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10546,42 +9637,37 @@
 define <8 x i64> @ult_7_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ult_7_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_7_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10609,42 +9695,41 @@
 define <8 x i64> @ugt_7_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ugt_7_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_7_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10672,42 +9757,41 @@
 define <8 x i64> @ult_8_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ult_8_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_8_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10735,42 +9819,45 @@
 define <8 x i64> @ugt_8_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ugt_8_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ugt_8_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpnleuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;
@@ -10798,42 +9885,45 @@
 define <8 x i64> @ult_9_v8i64(<8 x i64> %0) {
 ; AVX512F-LABEL: ult_9_v8i64:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm3
-; AVX512F-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512F-NEXT:    vpshufb %ymm3, %ymm4, %ymm3
-; AVX512F-NEXT:    vpsrlw $4, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpshufb %ymm1, %ymm4, %ymm1
-; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm5
-; AVX512F-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
-; AVX512F-NEXT:    vpsrlw $4, %ymm0, %ymm0
-; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512F-NEXT:    vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsadbw %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512F-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512F-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BW-LABEL: ult_9_v8i64:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm2
-; AVX512BW-NEXT:    vmovdqa64 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; AVX512BW-NEXT:    vpshufb %zmm2, %zmm3, %zmm2
-; AVX512BW-NEXT:    vpsrlw $4, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpandq %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpshufb %zmm0, %zmm3, %zmm0
-; AVX512BW-NEXT:    vpaddb %zmm2, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsadbw %zmm1, %zmm0, %zmm0
-; AVX512BW-NEXT:    vpcmpltuq {{.*}}(%rip){1to8}, %zmm0, %k1
+; AVX512BW-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm2
+; AVX512BW-NEXT:    vpandq %zmm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpaddq %zmm1, %zmm0, %zmm1
+; AVX512BW-NEXT:    vptestnmq %zmm1, %zmm0, %k1
 ; AVX512BW-NEXT:    vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512BW-NEXT:    retq
 ;