[X86][SSE] Use (V)PHMINPOSUW for vXi16 SMAX/SMIN/UMAX/UMIN horizontal reductions (PR32841)

[RKSimon]

Summary

(V)PHMINPOSUW determines the UMIN element in an v8i16 input, with suitable bit flipping it can also be used for SMAX/SMIN/UMAX cases as well.

This patch matches vXi16 SMAX/SMIN/UMAX/UMIN horizontal reductions and reduces the input down to a v8i16 vector before calling (V)PHMINPOSUW.

I can split the commits into X86ISD::PHMINPOS, UMIN reduction and SMAX/SMIN/UMAX reduction stages if it would make things clearer.

A later patch will use this for v16i8 reductions as well (PR32841).

Depending on when D39726 lands I might need to update this to use that as well.

[RKSimon]

Updated now that D39726 has landed

[RKSimon]

ping?

[zvi]

LGTM

lib/Target/X86/X86ISelLowering.h

enum NodeType : unsigned {
       // Vector packed double/float comparison.
       CMPP,
 
       // Vector integer comparisons.
       PCMPEQ, PCMPGT,
       // Vector integer comparisons, the result is in a mask vector.
       PCMPEQM, PCMPGTM,
 
+      // v8i16 Horizontal minimum and position.
+      PHMINPOS,
+
       MULTISHIFT,
 
       /// Vector comparison generating mask bits for fp and
       /// integer signed and unsigned data types.
       CMPM,
       CMPMU,
       // Vector comparison with rounding mode for FP values
       CMPM_RND,

lib/Target/X86/X86ISelLowering.cpp

const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
   case X86ISD::VROTLI:             return "X86ISD::VROTLI";
   case X86ISD::VROTRI:             return "X86ISD::VROTRI";
   case X86ISD::VPPERM:             return "X86ISD::VPPERM";
   case X86ISD::CMPP:               return "X86ISD::CMPP";
   case X86ISD::PCMPEQ:             return "X86ISD::PCMPEQ";
   case X86ISD::PCMPGT:             return "X86ISD::PCMPGT";
   case X86ISD::PCMPEQM:            return "X86ISD::PCMPEQM";
   case X86ISD::PCMPGTM:            return "X86ISD::PCMPGTM";
+  case X86ISD::PHMINPOS:           return "X86ISD::PHMINPOS";
   case X86ISD::ADD:                return "X86ISD::ADD";
   case X86ISD::SUB:                return "X86ISD::SUB";
   case X86ISD::ADC:                return "X86ISD::ADC";
   case X86ISD::SBB:                return "X86ISD::SBB";
   case X86ISD::SMUL:               return "X86ISD::SMUL";
   case X86ISD::UMUL:               return "X86ISD::UMUL";
   case X86ISD::SMUL8:              return "X86ISD::SMUL8";
   case X86ISD::UMUL8:              return "X86ISD::UMUL8";
static SDValue createPSADBW(SelectionDAG &DAG, const SDValue &Zext0,
   Ops[0] = Zext1.getOperand(0);
   SDValue SadOp1 = DAG.getNode(ISD::CONCAT_VECTORS, DL, ExtendedVT, Ops);
 
   // Actually build the SAD
   MVT SadVT = MVT::getVectorVT(MVT::i64, RegSize / 64);
   return DAG.getNode(X86ISD::PSADBW, DL, SadVT, SadOp0, SadOp1);
 }
 
+// Attempt to replace an min/max v8i16 horizontal reduction with PHMINPOSUW.
+static SDValue combineHorizontalMinMaxResult(SDNode *Extract,
+                                             SelectionDAG &DAG,
+                                             const X86Subtarget &Subtarget) {
+  // Bail without SSE41.
+  if (!Subtarget.hasSSE41())
+    return SDValue();
+
+  EVT ExtractVT = Extract->getValueType(0);
+  if (ExtractVT != MVT::i16)
+    return SDValue();
+
+  // Check for SMAX/SMIN/UMAX/UMIN horizontal reduction patterns.
+  for (ISD::NodeType Op : {ISD::SMAX, ISD::SMIN, ISD::UMAX, ISD::UMIN}) {
+    SDValue Src = matchBinOpReduction(Extract, Op);
+    if (!Src)
+      continue;
+
+    EVT SrcVT = Src.getValueType();
+    if (SrcVT.getScalarType() != MVT::i16 || (SrcVT.getSizeInBits() % 128) != 0)
+      return SDValue();
+
+    SDLoc DL(Extract);
+    SDValue MinPos = Src;
+
+    // First, reduce the source down to 128-bit, applying the op to lo/hi.
+    while (SrcVT.getSizeInBits() > 128) {
+      unsigned NumElts = SrcVT.getVectorNumElements();
+      unsigned NumSubElts = NumElts / 2;
+      SrcVT = EVT::getVectorVT(*DAG.getContext(), MVT::i16, NumSubElts);
+      unsigned SubSizeInBits = SrcVT.getSizeInBits();
+      SDValue Lo = extractSubVector(MinPos, 0, DAG, DL, SubSizeInBits);
+      SDValue Hi = extractSubVector(MinPos, NumSubElts, DAG, DL, SubSizeInBits);
+      MinPos = DAG.getNode(Op, DL, SrcVT, Lo, Hi);
+    }
+    assert(SrcVT == MVT::v8i16 && "Unexpected value type");
+
+    // PHMINPOSUW applies to UMIN(v8i16), for SMIN/SMAX/UMAX we must apply a mask
+    // to flip the value accordingly.
+    SDValue Mask;
+    if (Op == ISD::SMAX)
+      Mask = DAG.getConstant(APInt::getSignedMaxValue(16), DL, MVT::v8i16);
+    else if (Op == ISD::SMIN)
+      Mask = DAG.getConstant(APInt::getSignedMinValue(16), DL, MVT::v8i16);
+    else if (Op == ISD::UMAX)
+      Mask = DAG.getConstant(APInt::getAllOnesValue(16), DL, MVT::v8i16);
+
+    if (Mask)
+      MinPos = DAG.getNode(ISD::XOR, DL, MVT::v8i16, Mask, MinPos);
+
+    MinPos = DAG.getNode(X86ISD::PHMINPOS, DL, MVT::v8i16, MinPos);
+
+    if (Mask)
+      MinPos = DAG.getNode(ISD::XOR, DL, MVT::v8i16, Mask, MinPos);
+
+    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i16, MinPos,
+                       DAG.getIntPtrConstant(0, DL));
+  }
+
+  return SDValue();
+}
+
 // Attempt to replace an all_of/any_of style horizontal reduction with a MOVMSK.
 static SDValue combineHorizontalPredicateResult(SDNode *Extract,
                                                 SelectionDAG &DAG,
                                                 const X86Subtarget &Subtarget) {
   // Bail without SSE2 or with AVX512VL (which uses predicate registers).
   if (!Subtarget.hasSSE2() || Subtarget.hasVLX())
     return SDValue();
 
static SDValue combineExtractVectorElt(SDNode *N, SelectionDAG &DAG,
   // pre-legalization,
   if (SDValue SAD = combineBasicSADPattern(N, DAG, Subtarget))
     return SAD;
 
   // Attempt to replace an all_of/any_of horizontal reduction with a MOVMSK.
   if (SDValue Cmp = combineHorizontalPredicateResult(N, DAG, Subtarget))
     return Cmp;
 
+  // Attempt to replace min/max v8i16 reductions with PHMINPOSUW.
+  if (SDValue MinMax = combineHorizontalMinMaxResult(N, DAG, Subtarget))
+    return MinMax;
+
   // Only operate on vectors of 4 elements, where the alternative shuffling
   // gets to be more expensive.
   if (SrcVT != MVT::v4i32)
     return SDValue();
 
   // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
   // single use which is a sign-extend or zero-extend, and all elements are
   // used.

lib/Target/X86/X86InstrFragmentsSIMD.td

def X86CmpMaskCCScalarRound :
                            SDTCisVT<3, i8>, SDTCisVT<4, i32>]>;
 
 def X86cmpm     : SDNode<"X86ISD::CMPM",     X86CmpMaskCC>;
 def X86cmpmRnd  : SDNode<"X86ISD::CMPM_RND", X86CmpMaskCCRound>;
 def X86cmpmu    : SDNode<"X86ISD::CMPMU",    X86CmpMaskCC>;
 def X86cmpms    : SDNode<"X86ISD::FSETCCM",   X86CmpMaskCCScalar>;
 def X86cmpmsRnd : SDNode<"X86ISD::FSETCCM_RND",   X86CmpMaskCCScalarRound>;
 
+def X86phminpos: SDNode<"X86ISD::PHMINPOS", 
+                 SDTypeProfile<1, 1, [SDTCisVT<0, v8i16>, SDTCisVT<1, v8i16>]>>;
+
 def X86vshiftuniform : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
                                             SDTCisVec<2>, SDTCisInt<0>,
                                             SDTCisInt<1>]>;
 
 def X86vshl    : SDNode<"X86ISD::VSHL", X86vshiftuniform>;
 def X86vsrl    : SDNode<"X86ISD::VSRL", X86vshiftuniform>;
 def X86vsra    : SDNode<"X86ISD::VSRA", X86vshiftuniform>;
 

lib/Target/X86/X86InstrSSE.td

def POPCNT64rr : RI<0xB8, MRMSrcReg, (outs GR64:$dst), (ins GR64:$src),
                       IIC_SSE_POPCNT_RR>, Sched<[WriteFAdd]>, XS;
   def POPCNT64rm : RI<0xB8, MRMSrcMem, (outs GR64:$dst), (ins i64mem:$src),
                       "popcnt{q}\t{$src, $dst|$dst, $src}",
                       [(set GR64:$dst, (ctpop (loadi64 addr:$src))),
                        (implicit EFLAGS)], IIC_SSE_POPCNT_RM>,
                        Sched<[WriteFAddLd]>, XS;
 }
 
-
-
 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
-                                 Intrinsic IntId128, PatFrag ld_frag,
+                                 SDNode OpNode, PatFrag ld_frag,
                                  X86FoldableSchedWrite Sched> {
   def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
                     (ins VR128:$src),
                     !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
-                    [(set VR128:$dst, (IntId128 VR128:$src))]>,
+                    [(set VR128:$dst, (v8i16 (OpNode (v8i16 VR128:$src))))]>,
                     Sched<[Sched]>;
   def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
                      (ins i128mem:$src),
                      !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
                      [(set VR128:$dst,
-                       (IntId128 (bitconvert (ld_frag addr:$src))))]>,
+                       (v8i16 (OpNode (v8i16 (bitconvert (ld_frag addr:$src))))))]>,
                     Sched<[Sched.Folded]>;
 }
 
 // PHMIN has the same profile as PSAD, thus we use the same scheduling
 // model, although the naming is misleading.
 let Predicates = [HasAVX] in
 defm VPHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "vphminposuw",
-                                         int_x86_sse41_phminposuw, loadv2i64,
+                                         X86phminpos, loadv2i64,
                                          WriteVecIMul>, VEX, VEX_WIG;
 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
-                                         int_x86_sse41_phminposuw, memopv2i64,
+                                         X86phminpos, memopv2i64,
                                          WriteVecIMul>;
 
 /// SS48I_binop_rm - Simple SSE41 binary operator.
 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
                           ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
                           X86MemOperand x86memop, bit Is2Addr = 1,
                           OpndItins itins = SSE_INTALU_ITINS_P> {
   let isCommutable = 1 in

lib/Target/X86/X86IntrinsicsInfo.h

   X86_INTRINSIC_DATA(sse3_addsub_pd,    INTR_TYPE_2OP, X86ISD::ADDSUB, 0),
   X86_INTRINSIC_DATA(sse3_addsub_ps,    INTR_TYPE_2OP, X86ISD::ADDSUB, 0),
   X86_INTRINSIC_DATA(sse3_hadd_pd,      INTR_TYPE_2OP, X86ISD::FHADD, 0),
   X86_INTRINSIC_DATA(sse3_hadd_ps,      INTR_TYPE_2OP, X86ISD::FHADD, 0),
   X86_INTRINSIC_DATA(sse3_hsub_pd,      INTR_TYPE_2OP, X86ISD::FHSUB, 0),
   X86_INTRINSIC_DATA(sse3_hsub_ps,      INTR_TYPE_2OP, X86ISD::FHSUB, 0),
   X86_INTRINSIC_DATA(sse41_insertps,    INTR_TYPE_3OP, X86ISD::INSERTPS, 0),
   X86_INTRINSIC_DATA(sse41_packusdw,    INTR_TYPE_2OP, X86ISD::PACKUS, 0),
+  X86_INTRINSIC_DATA(sse41_phminposuw,  INTR_TYPE_1OP, X86ISD::PHMINPOS, 0),
   X86_INTRINSIC_DATA(sse41_pmuldq,      INTR_TYPE_2OP, X86ISD::PMULDQ, 0),
   X86_INTRINSIC_DATA(sse4a_extrqi,      INTR_TYPE_3OP, X86ISD::EXTRQI, 0),
   X86_INTRINSIC_DATA(sse4a_insertqi,    INTR_TYPE_4OP, X86ISD::INSERTQI, 0),
   X86_INTRINSIC_DATA(ssse3_phadd_d_128, INTR_TYPE_2OP, X86ISD::HADD, 0),
   X86_INTRINSIC_DATA(ssse3_phadd_w_128, INTR_TYPE_2OP, X86ISD::HADD, 0),
   X86_INTRINSIC_DATA(ssse3_phsub_d_128, INTR_TYPE_2OP, X86ISD::HSUB, 0),
   X86_INTRINSIC_DATA(ssse3_phsub_w_128, INTR_TYPE_2OP, X86ISD::HSUB, 0),
   X86_INTRINSIC_DATA(ssse3_pmadd_ub_sw_128, INTR_TYPE_2OP, X86ISD::VPMADDUBSW, 0),

test/CodeGen/X86/horizontal-reduce-smax.ll

; X64-AVX-NEXT:    retq
   %4 = shufflevector <4 x i32> %3, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
   %5 = icmp sgt <4 x i32> %3, %4
   %6 = select <4 x i1> %5, <4 x i32> %3, <4 x i32> %4
   %7 = extractelement <4 x i32> %6, i32 0
   ret i32 %7
 }
 
 define i16 @test_reduce_v8i16(<8 x i16> %a0) {
-; X86-SSE-LABEL: test_reduce_v8i16:
-; X86-SSE:       ## BB#0:
-; X86-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-SSE-NEXT:    pmaxsw %xmm0, %xmm1
-; X86-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X86-SSE-NEXT:    pmaxsw %xmm1, %xmm0
-; X86-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE-NEXT:    psrld $16, %xmm1
-; X86-SSE-NEXT:    pmaxsw %xmm0, %xmm1
-; X86-SSE-NEXT:    movd %xmm1, %eax
-; X86-SSE-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
-; X86-SSE-NEXT:    retl
+; X86-SSE2-LABEL: test_reduce_v8i16:
+; X86-SSE2:       ## BB#0:
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
+; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
+; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT:    psrld $16, %xmm1
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-SSE2-NEXT:    retl
+;
+; X86-SSE42-LABEL: test_reduce_v8i16:
+; X86-SSE42:       ## BB#0:
+; X86-SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    movd %xmm0, %eax
+; X86-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX-LABEL: test_reduce_v8i16:
 ; X86-AVX:       ## BB#0:
-; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X86-AVX-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
+; X86-AVX-NEXT:    vmovdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X86-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX-NEXT:    retl
 ;
-; X64-SSE-LABEL: test_reduce_v8i16:
-; X64-SSE:       ## BB#0:
-; X64-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-SSE-NEXT:    pmaxsw %xmm0, %xmm1
-; X64-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X64-SSE-NEXT:    pmaxsw %xmm1, %xmm0
-; X64-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X64-SSE-NEXT:    psrld $16, %xmm1
-; X64-SSE-NEXT:    pmaxsw %xmm0, %xmm1
-; X64-SSE-NEXT:    movd %xmm1, %eax
-; X64-SSE-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
-; X64-SSE-NEXT:    retq
+; X64-SSE2-LABEL: test_reduce_v8i16:
+; X64-SSE2:       ## BB#0:
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
+; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
+; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X64-SSE2-NEXT:    psrld $16, %xmm1
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-SSE2-NEXT:    retq
+;
+; X64-SSE42-LABEL: test_reduce_v8i16:
+; X64-SSE42:       ## BB#0:
+; X64-SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    movd %xmm0, %eax
+; X64-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-SSE42-NEXT:    retq
 ;
 ; X64-AVX-LABEL: test_reduce_v8i16:
 ; X64-AVX:       ## BB#0:
-; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X64-AVX-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
+; X64-AVX-NEXT:    vmovdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X64-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX-NEXT:    retq
   %1  = shufflevector <8 x i16> %a0, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
   %2  = icmp sgt <8 x i16> %a0, %1
   %3  = select <8 x i1> %2, <8 x i16> %a0, <8 x i16> %1
   %4  = shufflevector <8 x i16> %3, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %5  = icmp sgt <8 x i16> %3, %4
; X64-AVX512-NEXT:    retq
   %7  = shufflevector <8 x i32> %6, <8 x i32> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %8  = icmp sgt <8 x i32> %6, %7
   %9  = select <8 x i1> %8, <8 x i32> %6, <8 x i32> %7
   %10 = extractelement <8 x i32> %9, i32 0
   ret i32 %10
 }
 
 define i16 @test_reduce_v16i16(<16 x i16> %a0) {
-; X86-SSE-LABEL: test_reduce_v16i16:
-; X86-SSE:       ## BB#0:
-; X86-SSE-NEXT:    pmaxsw %xmm1, %xmm0
-; X86-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-SSE-NEXT:    pmaxsw %xmm0, %xmm1
-; X86-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X86-SSE-NEXT:    pmaxsw %xmm1, %xmm0
-; X86-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE-NEXT:    psrld $16, %xmm1
-; X86-SSE-NEXT:    pmaxsw %xmm0, %xmm1
-; X86-SSE-NEXT:    movd %xmm1, %eax
-; X86-SSE-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
-; X86-SSE-NEXT:    retl
+; X86-SSE2-LABEL: test_reduce_v16i16:
+; X86-SSE2:       ## BB#0:
+; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
+; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
+; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT:    psrld $16, %xmm1
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-SSE2-NEXT:    retl
+;
+; X86-SSE42-LABEL: test_reduce_v16i16:
+; X86-SSE42:       ## BB#0:
+; X86-SSE42-NEXT:    pmaxsw %xmm1, %xmm0
+; X86-SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    movd %xmm0, %eax
+; X86-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX1-LABEL: test_reduce_v16i16:
 ; X86-AVX1:       ## BB#0:
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; X86-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX1-NEXT:    vzeroupper
 ; X86-AVX1-NEXT:    retl
 ;
 ; X86-AVX2-LABEL: test_reduce_v16i16:
 ; X86-AVX2:       ## BB#0:
 ; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X86-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX2-NEXT:    vzeroupper
 ; X86-AVX2-NEXT:    retl
 ;
-; X64-SSE-LABEL: test_reduce_v16i16:
-; X64-SSE:       ## BB#0:
-; X64-SSE-NEXT:    pmaxsw %xmm1, %xmm0
-; X64-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-SSE-NEXT:    pmaxsw %xmm0, %xmm1
-; X64-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X64-SSE-NEXT:    pmaxsw %xmm1, %xmm0
-; X64-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X64-SSE-NEXT:    psrld $16, %xmm1
-; X64-SSE-NEXT:    pmaxsw %xmm0, %xmm1
-; X64-SSE-NEXT:    movd %xmm1, %eax
-; X64-SSE-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
-; X64-SSE-NEXT:    retq
+; X64-SSE2-LABEL: test_reduce_v16i16:
+; X64-SSE2:       ## BB#0:
+; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
+; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
+; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X64-SSE2-NEXT:    psrld $16, %xmm1
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-SSE2-NEXT:    retq
+;
+; X64-SSE42-LABEL: test_reduce_v16i16:
+; X64-SSE42:       ## BB#0:
+; X64-SSE42-NEXT:    pmaxsw %xmm1, %xmm0
+; X64-SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    movd %xmm0, %eax
+; X64-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-SSE42-NEXT:    retq
 ;
 ; X64-AVX1-LABEL: test_reduce_v16i16:
 ; X64-AVX1:       ## BB#0:
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; X64-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX1-NEXT:    vzeroupper
 ; X64-AVX1-NEXT:    retq
 ;
 ; X64-AVX2-LABEL: test_reduce_v16i16:
 ; X64-AVX2:       ## BB#0:
 ; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX2-NEXT:    vzeroupper
 ; X64-AVX2-NEXT:    retq
 ;
 ; X64-AVX512-LABEL: test_reduce_v16i16:
 ; X64-AVX512:       ## BB#0:
 ; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX512-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX512-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX512-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX512-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
+; X64-AVX512-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vmovdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X64-AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX512-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX512-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX512-NEXT:    vzeroupper
 ; X64-AVX512-NEXT:    retq
   %1  = shufflevector <16 x i16> %a0, <16 x i16> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %2  = icmp sgt <16 x i16> %a0, %1
   %3  = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %1
   %4  = shufflevector <16 x i16> %3, <16 x i16> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
; X64-AVX512-NEXT:    retq
   %10 = shufflevector <16 x i32> %9, <16 x i32> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %11 = icmp sgt <16 x i32> %9, %10
   %12 = select <16 x i1> %11, <16 x i32> %9, <16 x i32> %10
   %13 = extractelement <16 x i32> %12, i32 0
   ret i32 %13
 }
 
 define i16 @test_reduce_v32i16(<32 x i16> %a0) {
-; X86-SSE-LABEL: test_reduce_v32i16:
-; X86-SSE:       ## BB#0:
-; X86-SSE-NEXT:    pmaxsw %xmm3, %xmm1
-; X86-SSE-NEXT:    pmaxsw %xmm2, %xmm0
-; X86-SSE-NEXT:    pmaxsw %xmm1, %xmm0
-; X86-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-SSE-NEXT:    pmaxsw %xmm0, %xmm1
-; X86-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X86-SSE-NEXT:    pmaxsw %xmm1, %xmm0
-; X86-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE-NEXT:    psrld $16, %xmm1
-; X86-SSE-NEXT:    pmaxsw %xmm0, %xmm1
-; X86-SSE-NEXT:    movd %xmm1, %eax
-; X86-SSE-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
-; X86-SSE-NEXT:    retl
+; X86-SSE2-LABEL: test_reduce_v32i16:
+; X86-SSE2:       ## BB#0:
+; X86-SSE2-NEXT:    pmaxsw %xmm3, %xmm1
+; X86-SSE2-NEXT:    pmaxsw %xmm2, %xmm0
+; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
+; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
+; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT:    psrld $16, %xmm1
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-SSE2-NEXT:    retl
+;
+; X86-SSE42-LABEL: test_reduce_v32i16:
+; X86-SSE42:       ## BB#0:
+; X86-SSE42-NEXT:    pmaxsw %xmm3, %xmm1
+; X86-SSE42-NEXT:    pmaxsw %xmm2, %xmm0
+; X86-SSE42-NEXT:    pmaxsw %xmm1, %xmm0
+; X86-SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    movd %xmm0, %eax
+; X86-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX1-LABEL: test_reduce_v32i16:
 ; X86-AVX1:       ## BB#0:
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; X86-AVX1-NEXT:    vpmaxsw %xmm2, %xmm3, %xmm2
 ; X86-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vpmaxsw %xmm2, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX1-NEXT:    vzeroupper
 ; X86-AVX1-NEXT:    retl
 ;
 ; X86-AVX2-LABEL: test_reduce_v32i16:
 ; X86-AVX2:       ## BB#0:
 ; X86-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
 ; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X86-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX2-NEXT:    vzeroupper
 ; X86-AVX2-NEXT:    retl
 ;
-; X64-SSE-LABEL: test_reduce_v32i16:
-; X64-SSE:       ## BB#0:
-; X64-SSE-NEXT:    pmaxsw %xmm3, %xmm1
-; X64-SSE-NEXT:    pmaxsw %xmm2, %xmm0
-; X64-SSE-NEXT:    pmaxsw %xmm1, %xmm0
-; X64-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-SSE-NEXT:    pmaxsw %xmm0, %xmm1
-; X64-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X64-SSE-NEXT:    pmaxsw %xmm1, %xmm0
-; X64-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X64-SSE-NEXT:    psrld $16, %xmm1
-; X64-SSE-NEXT:    pmaxsw %xmm0, %xmm1
-; X64-SSE-NEXT:    movd %xmm1, %eax
-; X64-SSE-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
-; X64-SSE-NEXT:    retq
+; X64-SSE2-LABEL: test_reduce_v32i16:
+; X64-SSE2:       ## BB#0:
+; X64-SSE2-NEXT:    pmaxsw %xmm3, %xmm1
+; X64-SSE2-NEXT:    pmaxsw %xmm2, %xmm0
+; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
+; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
+; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X64-SSE2-NEXT:    psrld $16, %xmm1
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-SSE2-NEXT:    retq
+;
+; X64-SSE42-LABEL: test_reduce_v32i16:
+; X64-SSE42:       ## BB#0:
+; X64-SSE42-NEXT:    pmaxsw %xmm3, %xmm1
+; X64-SSE42-NEXT:    pmaxsw %xmm2, %xmm0
+; X64-SSE42-NEXT:    pmaxsw %xmm1, %xmm0
+; X64-SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    movd %xmm0, %eax
+; X64-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-SSE42-NEXT:    retq
 ;
 ; X64-AVX1-LABEL: test_reduce_v32i16:
 ; X64-AVX1:       ## BB#0:
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; X64-AVX1-NEXT:    vpmaxsw %xmm2, %xmm3, %xmm2
 ; X64-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
 ; X64-AVX1-NEXT:    vpmaxsw %xmm2, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX1-NEXT:    vzeroupper
 ; X64-AVX1-NEXT:    retq
 ;
 ; X64-AVX2-LABEL: test_reduce_v32i16:
 ; X64-AVX2:       ## BB#0:
 ; X64-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
 ; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX2-NEXT:    vzeroupper
 ; X64-AVX2-NEXT:    retq
 ;
 ; X64-AVX512-LABEL: test_reduce_v32i16:
 ; X64-AVX512:       ## BB#0:
 ; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; X64-AVX512-NEXT:    vpmaxsw %zmm1, %zmm0, %zmm0
+; X64-AVX512-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
 ; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX512-NEXT:    vpmaxsw %zmm1, %zmm0, %zmm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX512-NEXT:    vpmaxsw %zmm1, %zmm0, %zmm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX512-NEXT:    vpmaxsw %zmm1, %zmm0, %zmm0
-; X64-AVX512-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX512-NEXT:    vpmaxsw %zmm1, %zmm0, %zmm0
+; X64-AVX512-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vmovdqa {{.*#+}} xmm1 = [32767,32767,32767,32767,32767,32767,32767,32767]
+; X64-AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX512-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX512-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX512-NEXT:    vzeroupper
 ; X64-AVX512-NEXT:    retq
   %1  = shufflevector <32 x i16> %a0, <32 x i16> undef, <32 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %2  = icmp sgt <32 x i16> %a0, %1
   %3  = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %1
   %4  = shufflevector <32 x i16> %3, <32 x i16> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>

test/CodeGen/X86/horizontal-reduce-smin.ll

; X64-AVX-NEXT:    retq
   %4 = shufflevector <4 x i32> %3, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
   %5 = icmp slt <4 x i32> %3, %4
   %6 = select <4 x i1> %5, <4 x i32> %3, <4 x i32> %4
   %7 = extractelement <4 x i32> %6, i32 0
   ret i32 %7
 }
 
 define i16 @test_reduce_v8i16(<8 x i16> %a0) {
-; X86-SSE-LABEL: test_reduce_v8i16:
-; X86-SSE:       ## BB#0:
-; X86-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-SSE-NEXT:    pminsw %xmm0, %xmm1
-; X86-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X86-SSE-NEXT:    pminsw %xmm1, %xmm0
-; X86-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE-NEXT:    psrld $16, %xmm1
-; X86-SSE-NEXT:    pminsw %xmm0, %xmm1
-; X86-SSE-NEXT:    movd %xmm1, %eax
-; X86-SSE-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
-; X86-SSE-NEXT:    retl
+; X86-SSE2-LABEL: test_reduce_v8i16:
+; X86-SSE2:       ## BB#0:
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
+; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0
+; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT:    psrld $16, %xmm1
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-SSE2-NEXT:    retl
+;
+; X86-SSE42-LABEL: test_reduce_v8i16:
+; X86-SSE42:       ## BB#0:
+; X86-SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    movd %xmm0, %eax
+; X86-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX-LABEL: test_reduce_v8i16:
 ; X86-AVX:       ## BB#0:
-; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X86-AVX-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
+; X86-AVX-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X86-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX-NEXT:    retl
 ;
-; X64-SSE-LABEL: test_reduce_v8i16:
-; X64-SSE:       ## BB#0:
-; X64-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-SSE-NEXT:    pminsw %xmm0, %xmm1
-; X64-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X64-SSE-NEXT:    pminsw %xmm1, %xmm0
-; X64-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X64-SSE-NEXT:    psrld $16, %xmm1
-; X64-SSE-NEXT:    pminsw %xmm0, %xmm1
-; X64-SSE-NEXT:    movd %xmm1, %eax
-; X64-SSE-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
-; X64-SSE-NEXT:    retq
+; X64-SSE2-LABEL: test_reduce_v8i16:
+; X64-SSE2:       ## BB#0:
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
+; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0
+; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X64-SSE2-NEXT:    psrld $16, %xmm1
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-SSE2-NEXT:    retq
+;
+; X64-SSE42-LABEL: test_reduce_v8i16:
+; X64-SSE42:       ## BB#0:
+; X64-SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    movd %xmm0, %eax
+; X64-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-SSE42-NEXT:    retq
 ;
 ; X64-AVX-LABEL: test_reduce_v8i16:
 ; X64-AVX:       ## BB#0:
-; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X64-AVX-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
+; X64-AVX-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X64-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX-NEXT:    retq
   %1  = shufflevector <8 x i16> %a0, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
   %2  = icmp slt <8 x i16> %a0, %1
   %3  = select <8 x i1> %2, <8 x i16> %a0, <8 x i16> %1
   %4  = shufflevector <8 x i16> %3, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %5  = icmp slt <8 x i16> %3, %4
; X64-AVX512-NEXT:    retq
   %7  = shufflevector <8 x i32> %6, <8 x i32> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %8  = icmp slt <8 x i32> %6, %7
   %9  = select <8 x i1> %8, <8 x i32> %6, <8 x i32> %7
   %10 = extractelement <8 x i32> %9, i32 0
   ret i32 %10
 }
 
 define i16 @test_reduce_v16i16(<16 x i16> %a0) {
-; X86-SSE-LABEL: test_reduce_v16i16:
-; X86-SSE:       ## BB#0:
-; X86-SSE-NEXT:    pminsw %xmm1, %xmm0
-; X86-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-SSE-NEXT:    pminsw %xmm0, %xmm1
-; X86-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X86-SSE-NEXT:    pminsw %xmm1, %xmm0
-; X86-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE-NEXT:    psrld $16, %xmm1
-; X86-SSE-NEXT:    pminsw %xmm0, %xmm1
-; X86-SSE-NEXT:    movd %xmm1, %eax
-; X86-SSE-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
-; X86-SSE-NEXT:    retl
+; X86-SSE2-LABEL: test_reduce_v16i16:
+; X86-SSE2:       ## BB#0:
+; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
+; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0
+; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT:    psrld $16, %xmm1
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-SSE2-NEXT:    retl
+;
+; X86-SSE42-LABEL: test_reduce_v16i16:
+; X86-SSE42:       ## BB#0:
+; X86-SSE42-NEXT:    pminsw %xmm1, %xmm0
+; X86-SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    movd %xmm0, %eax
+; X86-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX1-LABEL: test_reduce_v16i16:
 ; X86-AVX1:       ## BB#0:
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; X86-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX1-NEXT:    vzeroupper
 ; X86-AVX1-NEXT:    retl
 ;
 ; X86-AVX2-LABEL: test_reduce_v16i16:
 ; X86-AVX2:       ## BB#0:
 ; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X86-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX2-NEXT:    vzeroupper
 ; X86-AVX2-NEXT:    retl
 ;
-; X64-SSE-LABEL: test_reduce_v16i16:
-; X64-SSE:       ## BB#0:
-; X64-SSE-NEXT:    pminsw %xmm1, %xmm0
-; X64-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-SSE-NEXT:    pminsw %xmm0, %xmm1
-; X64-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X64-SSE-NEXT:    pminsw %xmm1, %xmm0
-; X64-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X64-SSE-NEXT:    psrld $16, %xmm1
-; X64-SSE-NEXT:    pminsw %xmm0, %xmm1
-; X64-SSE-NEXT:    movd %xmm1, %eax
-; X64-SSE-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
-; X64-SSE-NEXT:    retq
+; X64-SSE2-LABEL: test_reduce_v16i16:
+; X64-SSE2:       ## BB#0:
+; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
+; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0
+; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X64-SSE2-NEXT:    psrld $16, %xmm1
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-SSE2-NEXT:    retq
+;
+; X64-SSE42-LABEL: test_reduce_v16i16:
+; X64-SSE42:       ## BB#0:
+; X64-SSE42-NEXT:    pminsw %xmm1, %xmm0
+; X64-SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    movd %xmm0, %eax
+; X64-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-SSE42-NEXT:    retq
 ;
 ; X64-AVX1-LABEL: test_reduce_v16i16:
 ; X64-AVX1:       ## BB#0:
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; X64-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX1-NEXT:    vzeroupper
 ; X64-AVX1-NEXT:    retq
 ;
 ; X64-AVX2-LABEL: test_reduce_v16i16:
 ; X64-AVX2:       ## BB#0:
 ; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX2-NEXT:    vzeroupper
 ; X64-AVX2-NEXT:    retq
 ;
 ; X64-AVX512-LABEL: test_reduce_v16i16:
 ; X64-AVX512:       ## BB#0:
 ; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX512-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX512-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX512-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX512-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
+; X64-AVX512-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X64-AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX512-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX512-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX512-NEXT:    vzeroupper
 ; X64-AVX512-NEXT:    retq
   %1  = shufflevector <16 x i16> %a0, <16 x i16> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %2  = icmp slt <16 x i16> %a0, %1
   %3  = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %1
   %4  = shufflevector <16 x i16> %3, <16 x i16> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
; X64-AVX512-NEXT:    retq
   %10 = shufflevector <16 x i32> %9, <16 x i32> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %11 = icmp slt <16 x i32> %9, %10
   %12 = select <16 x i1> %11, <16 x i32> %9, <16 x i32> %10
   %13 = extractelement <16 x i32> %12, i32 0
   ret i32 %13
 }
 
 define i16 @test_reduce_v32i16(<32 x i16> %a0) {
-; X86-SSE-LABEL: test_reduce_v32i16:
-; X86-SSE:       ## BB#0:
-; X86-SSE-NEXT:    pminsw %xmm3, %xmm1
-; X86-SSE-NEXT:    pminsw %xmm2, %xmm0
-; X86-SSE-NEXT:    pminsw %xmm1, %xmm0
-; X86-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-SSE-NEXT:    pminsw %xmm0, %xmm1
-; X86-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X86-SSE-NEXT:    pminsw %xmm1, %xmm0
-; X86-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE-NEXT:    psrld $16, %xmm1
-; X86-SSE-NEXT:    pminsw %xmm0, %xmm1
-; X86-SSE-NEXT:    movd %xmm1, %eax
-; X86-SSE-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
-; X86-SSE-NEXT:    retl
+; X86-SSE2-LABEL: test_reduce_v32i16:
+; X86-SSE2:       ## BB#0:
+; X86-SSE2-NEXT:    pminsw %xmm3, %xmm1
+; X86-SSE2-NEXT:    pminsw %xmm2, %xmm0
+; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
+; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0
+; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT:    psrld $16, %xmm1
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-SSE2-NEXT:    retl
+;
+; X86-SSE42-LABEL: test_reduce_v32i16:
+; X86-SSE42:       ## BB#0:
+; X86-SSE42-NEXT:    pminsw %xmm3, %xmm1
+; X86-SSE42-NEXT:    pminsw %xmm2, %xmm0
+; X86-SSE42-NEXT:    pminsw %xmm1, %xmm0
+; X86-SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    movd %xmm0, %eax
+; X86-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX1-LABEL: test_reduce_v32i16:
 ; X86-AVX1:       ## BB#0:
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; X86-AVX1-NEXT:    vpminsw %xmm2, %xmm3, %xmm2
 ; X86-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vpminsw %xmm2, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX1-NEXT:    vzeroupper
 ; X86-AVX1-NEXT:    retl
 ;
 ; X86-AVX2-LABEL: test_reduce_v32i16:
 ; X86-AVX2:       ## BB#0:
 ; X86-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
 ; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X86-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX2-NEXT:    vzeroupper
 ; X86-AVX2-NEXT:    retl
 ;
-; X64-SSE-LABEL: test_reduce_v32i16:
-; X64-SSE:       ## BB#0:
-; X64-SSE-NEXT:    pminsw %xmm3, %xmm1
-; X64-SSE-NEXT:    pminsw %xmm2, %xmm0
-; X64-SSE-NEXT:    pminsw %xmm1, %xmm0
-; X64-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-SSE-NEXT:    pminsw %xmm0, %xmm1
-; X64-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X64-SSE-NEXT:    pminsw %xmm1, %xmm0
-; X64-SSE-NEXT:    movdqa %xmm0, %xmm1
-; X64-SSE-NEXT:    psrld $16, %xmm1
-; X64-SSE-NEXT:    pminsw %xmm0, %xmm1
-; X64-SSE-NEXT:    movd %xmm1, %eax
-; X64-SSE-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
-; X64-SSE-NEXT:    retq
+; X64-SSE2-LABEL: test_reduce_v32i16:
+; X64-SSE2:       ## BB#0:
+; X64-SSE2-NEXT:    pminsw %xmm3, %xmm1
+; X64-SSE2-NEXT:    pminsw %xmm2, %xmm0
+; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
+; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0
+; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X64-SSE2-NEXT:    psrld $16, %xmm1
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-SSE2-NEXT:    retq
+;
+; X64-SSE42-LABEL: test_reduce_v32i16:
+; X64-SSE42:       ## BB#0:
+; X64-SSE42-NEXT:    pminsw %xmm3, %xmm1
+; X64-SSE42-NEXT:    pminsw %xmm2, %xmm0
+; X64-SSE42-NEXT:    pminsw %xmm1, %xmm0
+; X64-SSE42-NEXT:    movdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    movd %xmm0, %eax
+; X64-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
+; X64-SSE42-NEXT:    retq
 ;
 ; X64-AVX1-LABEL: test_reduce_v32i16:
 ; X64-AVX1:       ## BB#0:
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; X64-AVX1-NEXT:    vpminsw %xmm2, %xmm3, %xmm2
 ; X64-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
 ; X64-AVX1-NEXT:    vpminsw %xmm2, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX1-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX1-NEXT:    vzeroupper
 ; X64-AVX1-NEXT:    retq
 ;
 ; X64-AVX2-LABEL: test_reduce_v32i16:
 ; X64-AVX2:       ## BB#0:
 ; X64-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
 ; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX2-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX2-NEXT:    vzeroupper
 ; X64-AVX2-NEXT:    retq
 ;
 ; X64-AVX512-LABEL: test_reduce_v32i16:
 ; X64-AVX512:       ## BB#0:
 ; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; X64-AVX512-NEXT:    vpminsw %zmm1, %zmm0, %zmm0
+; X64-AVX512-NEXT:    vpminsw %ymm1, %ymm0, %ymm0
 ; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX512-NEXT:    vpminsw %zmm1, %zmm0, %zmm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX512-NEXT:    vpminsw %zmm1, %zmm0, %zmm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX512-NEXT:    vpminsw %zmm1, %zmm0, %zmm0
-; X64-AVX512-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX512-NEXT:    vpminsw %zmm1, %zmm0, %zmm0
+; X64-AVX512-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
+; X64-AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX512-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX512-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX512-NEXT:    vzeroupper
 ; X64-AVX512-NEXT:    retq
   %1  = shufflevector <32 x i16> %a0, <32 x i16> undef, <32 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %2  = icmp slt <32 x i16> %a0, %1
   %3  = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %1
   %4  = shufflevector <32 x i16> %3, <32 x i16> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>

test/CodeGen/X86/horizontal-reduce-umax.ll

 ; X86-SSE2-NEXT:    pandn %xmm0, %xmm3
 ; X86-SSE2-NEXT:    por %xmm2, %xmm3
 ; X86-SSE2-NEXT:    movd %xmm3, %eax
 ; X86-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v8i16:
 ; X86-SSE42:       ## BB#0:
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-SSE42-NEXT:    pmaxuw %xmm0, %xmm1
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X86-SSE42-NEXT:    pmaxuw %xmm1, %xmm0
-; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE42-NEXT:    psrld $16, %xmm1
-; X86-SSE42-NEXT:    pmaxuw %xmm0, %xmm1
-; X86-SSE42-NEXT:    movd %xmm1, %eax
+; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    movd %xmm0, %eax
 ; X86-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX-LABEL: test_reduce_v8i16:
 ; X86-AVX:       ## BB#0:
-; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X86-AVX-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
+; X86-AVX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X86-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v8i16:
 ; X64-SSE2:       ## BB#0:
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
 ; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]

 ; X64-SSE2-NEXT:    pandn %xmm0, %xmm3
 ; X64-SSE2-NEXT:    por %xmm2, %xmm3
 ; X64-SSE2-NEXT:    movd %xmm3, %eax
 ; X64-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v8i16:
 ; X64-SSE42:       ## BB#0:
-; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-SSE42-NEXT:    pmaxuw %xmm0, %xmm1
-; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X64-SSE42-NEXT:    pmaxuw %xmm1, %xmm0
-; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1
-; X64-SSE42-NEXT:    psrld $16, %xmm1
-; X64-SSE42-NEXT:    pmaxuw %xmm0, %xmm1
-; X64-SSE42-NEXT:    movd %xmm1, %eax
+; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    movd %xmm0, %eax
 ; X64-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-SSE42-NEXT:    retq
 ;
 ; X64-AVX-LABEL: test_reduce_v8i16:
 ; X64-AVX:       ## BB#0:
-; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X64-AVX-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
+; X64-AVX-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X64-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX-NEXT:    retq
   %1  = shufflevector <8 x i16> %a0, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
   %2  = icmp ugt <8 x i16> %a0, %1
   %3  = select <8 x i1> %2, <8 x i16> %a0, <8 x i16> %1
   %4  = shufflevector <8 x i16> %3, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %5  = icmp ugt <8 x i16> %3, %4

 ; X86-SSE2-NEXT:    por %xmm3, %xmm1
 ; X86-SSE2-NEXT:    movd %xmm1, %eax
 ; X86-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v16i16:
 ; X86-SSE42:       ## BB#0:
 ; X86-SSE42-NEXT:    pmaxuw %xmm1, %xmm0
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-SSE42-NEXT:    pmaxuw %xmm0, %xmm1
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X86-SSE42-NEXT:    pmaxuw %xmm1, %xmm0
-; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE42-NEXT:    psrld $16, %xmm1
-; X86-SSE42-NEXT:    pmaxuw %xmm0, %xmm1
-; X86-SSE42-NEXT:    movd %xmm1, %eax
+; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    movd %xmm0, %eax
 ; X86-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX1-LABEL: test_reduce_v16i16:
 ; X86-AVX1:       ## BB#0:
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; X86-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX1-NEXT:    vzeroupper
 ; X86-AVX1-NEXT:    retl
 ;
 ; X86-AVX2-LABEL: test_reduce_v16i16:
 ; X86-AVX2:       ## BB#0:
 ; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X86-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX2-NEXT:    vzeroupper
 ; X86-AVX2-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v16i16:
 ; X64-SSE2:       ## BB#0:
 ; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]

 ; X64-SSE2-NEXT:    por %xmm3, %xmm1
 ; X64-SSE2-NEXT:    movd %xmm1, %eax
 ; X64-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v16i16:
 ; X64-SSE42:       ## BB#0:
 ; X64-SSE42-NEXT:    pmaxuw %xmm1, %xmm0
-; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-SSE42-NEXT:    pmaxuw %xmm0, %xmm1
-; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X64-SSE42-NEXT:    pmaxuw %xmm1, %xmm0
-; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1
-; X64-SSE42-NEXT:    psrld $16, %xmm1
-; X64-SSE42-NEXT:    pmaxuw %xmm0, %xmm1
-; X64-SSE42-NEXT:    movd %xmm1, %eax
+; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    movd %xmm0, %eax
 ; X64-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-SSE42-NEXT:    retq
 ;
 ; X64-AVX1-LABEL: test_reduce_v16i16:
 ; X64-AVX1:       ## BB#0:
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; X64-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX1-NEXT:    vzeroupper
 ; X64-AVX1-NEXT:    retq
 ;
 ; X64-AVX2-LABEL: test_reduce_v16i16:
 ; X64-AVX2:       ## BB#0:
 ; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX2-NEXT:    vzeroupper
 ; X64-AVX2-NEXT:    retq
 ;
 ; X64-AVX512-LABEL: test_reduce_v16i16:
 ; X64-AVX512:       ## BB#0:
 ; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX512-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX512-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX512-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX512-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
+; X64-AVX512-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X64-AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX512-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX512-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX512-NEXT:    vzeroupper
 ; X64-AVX512-NEXT:    retq
   %1  = shufflevector <16 x i16> %a0, <16 x i16> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %2  = icmp ugt <16 x i16> %a0, %1
   %3  = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %1
   %4  = shufflevector <16 x i16> %3, <16 x i16> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>

 ; X86-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v32i16:
 ; X86-SSE42:       ## BB#0:
 ; X86-SSE42-NEXT:    pmaxuw %xmm3, %xmm1
 ; X86-SSE42-NEXT:    pmaxuw %xmm2, %xmm0
 ; X86-SSE42-NEXT:    pmaxuw %xmm1, %xmm0
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-SSE42-NEXT:    pmaxuw %xmm0, %xmm1
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X86-SSE42-NEXT:    pmaxuw %xmm1, %xmm0
-; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE42-NEXT:    psrld $16, %xmm1
-; X86-SSE42-NEXT:    pmaxuw %xmm0, %xmm1
-; X86-SSE42-NEXT:    movd %xmm1, %eax
+; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X86-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X86-SSE42-NEXT:    movd %xmm0, %eax
 ; X86-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX1-LABEL: test_reduce_v32i16:
 ; X86-AVX1:       ## BB#0:
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; X86-AVX1-NEXT:    vpmaxuw %xmm2, %xmm3, %xmm2
 ; X86-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vpmaxuw %xmm2, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX1-NEXT:    vzeroupper
 ; X86-AVX1-NEXT:    retl
 ;
 ; X86-AVX2-LABEL: test_reduce_v32i16:
 ; X86-AVX2:       ## BB#0:
 ; X86-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
 ; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X86-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
+; X86-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X86-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX2-NEXT:    vzeroupper
 ; X86-AVX2-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v32i16:
 ; X64-SSE2:       ## BB#0:
 ; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [32768,32768,32768,32768,32768,32768,32768,32768]

 ; X64-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v32i16:
 ; X64-SSE42:       ## BB#0:
 ; X64-SSE42-NEXT:    pmaxuw %xmm3, %xmm1
 ; X64-SSE42-NEXT:    pmaxuw %xmm2, %xmm0
 ; X64-SSE42-NEXT:    pmaxuw %xmm1, %xmm0
-; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-SSE42-NEXT:    pmaxuw %xmm0, %xmm1
-; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X64-SSE42-NEXT:    pmaxuw %xmm1, %xmm0
-; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1
-; X64-SSE42-NEXT:    psrld $16, %xmm1
-; X64-SSE42-NEXT:    pmaxuw %xmm0, %xmm1
-; X64-SSE42-NEXT:    movd %xmm1, %eax
+; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X64-SSE42-NEXT:    pxor %xmm1, %xmm0
+; X64-SSE42-NEXT:    movd %xmm0, %eax
 ; X64-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-SSE42-NEXT:    retq
 ;
 ; X64-AVX1-LABEL: test_reduce_v32i16:
 ; X64-AVX1:       ## BB#0:
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; X64-AVX1-NEXT:    vpmaxuw %xmm2, %xmm3, %xmm2
 ; X64-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
 ; X64-AVX1-NEXT:    vpmaxuw %xmm2, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX1-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX1-NEXT:    vzeroupper
 ; X64-AVX1-NEXT:    retq
 ;
 ; X64-AVX2-LABEL: test_reduce_v32i16:
 ; X64-AVX2:       ## BB#0:
 ; X64-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
 ; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX2-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX2-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX2-NEXT:    vzeroupper
 ; X64-AVX2-NEXT:    retq
 ;
 ; X64-AVX512-LABEL: test_reduce_v32i16:
 ; X64-AVX512:       ## BB#0:
 ; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; X64-AVX512-NEXT:    vpmaxuw %zmm1, %zmm0, %zmm0
+; X64-AVX512-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm0
 ; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX512-NEXT:    vpmaxuw %zmm1, %zmm0, %zmm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX512-NEXT:    vpmaxuw %zmm1, %zmm0, %zmm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX512-NEXT:    vpmaxuw %zmm1, %zmm0, %zmm0
-; X64-AVX512-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX512-NEXT:    vpmaxuw %zmm1, %zmm0, %zmm0
+; X64-AVX512-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X64-AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
+; X64-AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X64-AVX512-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX512-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX512-NEXT:    vzeroupper
 ; X64-AVX512-NEXT:    retq
   %1  = shufflevector <32 x i16> %a0, <32 x i16> undef, <32 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %2  = icmp ugt <32 x i16> %a0, %1
   %3  = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %1
   %4  = shufflevector <32 x i16> %3, <32 x i16> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>

test/CodeGen/X86/horizontal-reduce-umin.ll

 ; X86-SSE2-NEXT:    pandn %xmm0, %xmm1
 ; X86-SSE2-NEXT:    por %xmm3, %xmm1
 ; X86-SSE2-NEXT:    movd %xmm1, %eax
 ; X86-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v8i16:
 ; X86-SSE42:       ## BB#0:
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-SSE42-NEXT:    pminuw %xmm0, %xmm1
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X86-SSE42-NEXT:    pminuw %xmm1, %xmm0
-; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE42-NEXT:    psrld $16, %xmm1
-; X86-SSE42-NEXT:    pminuw %xmm0, %xmm1
-; X86-SSE42-NEXT:    movd %xmm1, %eax
+; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X86-SSE42-NEXT:    movd %xmm0, %eax
 ; X86-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX-LABEL: test_reduce_v8i16:
 ; X86-AVX:       ## BB#0:
-; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X86-AVX-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
+; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
 ; X86-AVX-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v8i16:
 ; X64-SSE2:       ## BB#0:
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,0,1]
 ; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]

 ; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
 ; X64-SSE2-NEXT:    por %xmm3, %xmm1
 ; X64-SSE2-NEXT:    movd %xmm1, %eax
 ; X64-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v8i16:
 ; X64-SSE42:       ## BB#0:
-; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-SSE42-NEXT:    pminuw %xmm0, %xmm1
-; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X64-SSE42-NEXT:    pminuw %xmm1, %xmm0
-; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1
-; X64-SSE42-NEXT:    psrld $16, %xmm1
-; X64-SSE42-NEXT:    pminuw %xmm0, %xmm1
-; X64-SSE42-NEXT:    movd %xmm1, %eax
+; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X64-SSE42-NEXT:    movd %xmm0, %eax
 ; X64-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-SSE42-NEXT:    retq
 ;
 ; X64-AVX-LABEL: test_reduce_v8i16:
 ; X64-AVX:       ## BB#0:
-; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X64-AVX-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
+; X64-AVX-NEXT:    vphminposuw %xmm0, %xmm0
 ; X64-AVX-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX-NEXT:    retq
   %1  = shufflevector <8 x i16> %a0, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
   %2  = icmp ult <8 x i16> %a0, %1
   %3  = select <8 x i1> %2, <8 x i16> %a0, <8 x i16> %1
   %4  = shufflevector <8 x i16> %3, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %5  = icmp ult <8 x i16> %3, %4

 ; X86-SSE2-NEXT:    por %xmm4, %xmm2
 ; X86-SSE2-NEXT:    movd %xmm2, %eax
 ; X86-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v16i16:
 ; X86-SSE42:       ## BB#0:
 ; X86-SSE42-NEXT:    pminuw %xmm1, %xmm0
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-SSE42-NEXT:    pminuw %xmm0, %xmm1
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X86-SSE42-NEXT:    pminuw %xmm1, %xmm0
-; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE42-NEXT:    psrld $16, %xmm1
-; X86-SSE42-NEXT:    pminuw %xmm0, %xmm1
-; X86-SSE42-NEXT:    movd %xmm1, %eax
+; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X86-SSE42-NEXT:    movd %xmm0, %eax
 ; X86-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX1-LABEL: test_reduce_v16i16:
 ; X86-AVX1:       ## BB#0:
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; X86-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX1-NEXT:    vzeroupper
 ; X86-AVX1-NEXT:    retl
 ;
 ; X86-AVX2-LABEL: test_reduce_v16i16:
 ; X86-AVX2:       ## BB#0:
 ; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X86-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
 ; X86-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX2-NEXT:    vzeroupper
 ; X86-AVX2-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v16i16:
 ; X64-SSE2:       ## BB#0:
 ; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]

 ; X64-SSE2-NEXT:    por %xmm4, %xmm2
 ; X64-SSE2-NEXT:    movd %xmm2, %eax
 ; X64-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v16i16:
 ; X64-SSE42:       ## BB#0:
 ; X64-SSE42-NEXT:    pminuw %xmm1, %xmm0
-; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-SSE42-NEXT:    pminuw %xmm0, %xmm1
-; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X64-SSE42-NEXT:    pminuw %xmm1, %xmm0
-; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1
-; X64-SSE42-NEXT:    psrld $16, %xmm1
-; X64-SSE42-NEXT:    pminuw %xmm0, %xmm1
-; X64-SSE42-NEXT:    movd %xmm1, %eax
+; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X64-SSE42-NEXT:    movd %xmm0, %eax
 ; X64-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-SSE42-NEXT:    retq
 ;
 ; X64-AVX1-LABEL: test_reduce_v16i16:
 ; X64-AVX1:       ## BB#0:
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; X64-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
 ; X64-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX1-NEXT:    vzeroupper
 ; X64-AVX1-NEXT:    retq
 ;
 ; X64-AVX2-LABEL: test_reduce_v16i16:
 ; X64-AVX2:       ## BB#0:
 ; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
 ; X64-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX2-NEXT:    vzeroupper
 ; X64-AVX2-NEXT:    retq
 ;
 ; X64-AVX512-LABEL: test_reduce_v16i16:
 ; X64-AVX512:       ## BB#0:
 ; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX512-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX512-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX512-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX512-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
+; X64-AVX512-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
 ; X64-AVX512-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX512-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX512-NEXT:    vzeroupper
 ; X64-AVX512-NEXT:    retq
   %1  = shufflevector <16 x i16> %a0, <16 x i16> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %2  = icmp ult <16 x i16> %a0, %1
   %3  = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %1
   %4  = shufflevector <16 x i16> %3, <16 x i16> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>

 ; X86-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v32i16:
 ; X86-SSE42:       ## BB#0:
 ; X86-SSE42-NEXT:    pminuw %xmm3, %xmm1
 ; X86-SSE42-NEXT:    pminuw %xmm2, %xmm0
 ; X86-SSE42-NEXT:    pminuw %xmm1, %xmm0
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-SSE42-NEXT:    pminuw %xmm0, %xmm1
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X86-SSE42-NEXT:    pminuw %xmm1, %xmm0
-; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE42-NEXT:    psrld $16, %xmm1
-; X86-SSE42-NEXT:    pminuw %xmm0, %xmm1
-; X86-SSE42-NEXT:    movd %xmm1, %eax
+; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X86-SSE42-NEXT:    movd %xmm0, %eax
 ; X86-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX1-LABEL: test_reduce_v32i16:
 ; X86-AVX1:       ## BB#0:
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; X86-AVX1-NEXT:    vpminuw %xmm2, %xmm3, %xmm2
 ; X86-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vpminuw %xmm2, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X86-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX1-NEXT:    vzeroupper
 ; X86-AVX1-NEXT:    retl
 ;
 ; X86-AVX2-LABEL: test_reduce_v32i16:
 ; X86-AVX2:       ## BB#0:
 ; X86-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
 ; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X86-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X86-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X86-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X86-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
 ; X86-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X86-AVX2-NEXT:    vzeroupper
 ; X86-AVX2-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v32i16:
 ; X64-SSE2:       ## BB#0:
 ; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [32768,32768,32768,32768,32768,32768,32768,32768]

 ; X64-SSE2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v32i16:
 ; X64-SSE42:       ## BB#0:
 ; X64-SSE42-NEXT:    pminuw %xmm3, %xmm1
 ; X64-SSE42-NEXT:    pminuw %xmm2, %xmm0
 ; X64-SSE42-NEXT:    pminuw %xmm1, %xmm0
-; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-SSE42-NEXT:    pminuw %xmm0, %xmm1
-; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; X64-SSE42-NEXT:    pminuw %xmm1, %xmm0
-; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1
-; X64-SSE42-NEXT:    psrld $16, %xmm1
-; X64-SSE42-NEXT:    pminuw %xmm0, %xmm1
-; X64-SSE42-NEXT:    movd %xmm1, %eax
+; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
+; X64-SSE42-NEXT:    movd %xmm0, %eax
 ; X64-SSE42-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-SSE42-NEXT:    retq
 ;
 ; X64-AVX1-LABEL: test_reduce_v32i16:
 ; X64-AVX1:       ## BB#0:
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; X64-AVX1-NEXT:    vpminuw %xmm2, %xmm3, %xmm2
 ; X64-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
 ; X64-AVX1-NEXT:    vpminuw %xmm2, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
-; X64-AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX1-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
 ; X64-AVX1-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX1-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX1-NEXT:    vzeroupper
 ; X64-AVX1-NEXT:    retq
 ;
 ; X64-AVX2-LABEL: test_reduce_v32i16:
 ; X64-AVX2:       ## BB#0:
 ; X64-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
 ; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
-; X64-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX2-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
 ; X64-AVX2-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX2-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX2-NEXT:    vzeroupper
 ; X64-AVX2-NEXT:    retq
 ;
 ; X64-AVX512-LABEL: test_reduce_v32i16:
 ; X64-AVX512:       ## BB#0:
 ; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; X64-AVX512-NEXT:    vpminuw %zmm1, %zmm0, %zmm0
+; X64-AVX512-NEXT:    vpminuw %ymm1, %ymm0, %ymm0
 ; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; X64-AVX512-NEXT:    vpminuw %zmm1, %zmm0, %zmm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; X64-AVX512-NEXT:    vpminuw %zmm1, %zmm0, %zmm0
-; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; X64-AVX512-NEXT:    vpminuw %zmm1, %zmm0, %zmm0
-; X64-AVX512-NEXT:    vpsrld $16, %xmm0, %xmm1
-; X64-AVX512-NEXT:    vpminuw %zmm1, %zmm0, %zmm0
+; X64-AVX512-NEXT:    vpminuw %xmm1, %xmm0, %xmm0
+; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
 ; X64-AVX512-NEXT:    vmovd %xmm0, %eax
 ; X64-AVX512-NEXT:    ## kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-AVX512-NEXT:    vzeroupper
 ; X64-AVX512-NEXT:    retq
   %1  = shufflevector <32 x i16> %a0, <32 x i16> undef, <32 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %2  = icmp ult <32 x i16> %a0, %1
   %3  = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %1
   %4  = shufflevector <32 x i16> %3, <32 x i16> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>