Index: llvm/trunk/lib/Target/X86/X86ISelLowering.h
===================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.h
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.h
@@ -335,6 +335,9 @@
       // 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
Index: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
@@ -25073,6 +25073,7 @@
   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";
@@ -30326,6 +30327,66 @@
   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.
+  unsigned BinOp;
+  SDValue Src = matchBinOpReduction(
+      Extract, BinOp, {ISD::SMAX, ISD::SMIN, ISD::UMAX, ISD::UMIN});
+  if (!Src)
+    return SDValue();
+
+  EVT SrcVT = Src.getValueType();
+  EVT SrcSVT = SrcVT.getScalarType();
+  if (SrcSVT != MVT::i16 || (SrcVT.getSizeInBits() % 128) != 0)
+    return SDValue();
+
+  SDLoc DL(Extract);
+  SDValue MinPos = Src;
+
+  // First, reduce the source down to 128-bit, applying BinOp to lo/hi.
+  while (SrcVT.getSizeInBits() > 128) {
+    unsigned NumElts = SrcVT.getVectorNumElements();
+    unsigned NumSubElts = NumElts / 2;
+    SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcSVT, NumSubElts);
+    unsigned SubSizeInBits = SrcVT.getSizeInBits();
+    SDValue Lo = extractSubVector(MinPos, 0, DAG, DL, SubSizeInBits);
+    SDValue Hi = extractSubVector(MinPos, NumSubElts, DAG, DL, SubSizeInBits);
+    MinPos = DAG.getNode(BinOp, 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 (BinOp == ISD::SMAX)
+    Mask = DAG.getConstant(APInt::getSignedMaxValue(16), DL, SrcVT);
+  else if (BinOp == ISD::SMIN)
+    Mask = DAG.getConstant(APInt::getSignedMinValue(16), DL, SrcVT);
+  else if (BinOp == ISD::UMAX)
+    Mask = DAG.getConstant(APInt::getAllOnesValue(16), DL, SrcVT);
+
+  if (Mask)
+    MinPos = DAG.getNode(ISD::XOR, DL, SrcVT, Mask, MinPos);
+
+  MinPos = DAG.getNode(X86ISD::PHMINPOS, DL, SrcVT, MinPos);
+
+  if (Mask)
+    MinPos = DAG.getNode(ISD::XOR, DL, SrcVT, Mask, MinPos);
+
+  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtractVT, MinPos,
+                     DAG.getIntPtrConstant(0, DL));
+}
+
 // Attempt to replace an all_of/any_of style horizontal reduction with a MOVMSK.
 static SDValue combineHorizontalPredicateResult(SDNode *Extract,
                                                 SelectionDAG &DAG,
@@ -30633,6 +30694,10 @@
   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)
Index: llvm/trunk/lib/Target/X86/X86InstrFragmentsSIMD.td
===================================================================
--- llvm/trunk/lib/Target/X86/X86InstrFragmentsSIMD.td
+++ llvm/trunk/lib/Target/X86/X86InstrFragmentsSIMD.td
@@ -182,6 +182,9 @@
 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>]>;
Index: llvm/trunk/lib/Target/X86/X86InstrSSE.td
===================================================================
--- llvm/trunk/lib/Target/X86/X86InstrSSE.td
+++ llvm/trunk/lib/Target/X86/X86InstrSSE.td
@@ -6188,22 +6188,20 @@
                        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]>;
 }
 
@@ -6211,10 +6209,10 @@
 // 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.
Index: llvm/trunk/lib/Target/X86/X86IntrinsicsInfo.h
===================================================================
--- llvm/trunk/lib/Target/X86/X86IntrinsicsInfo.h
+++ llvm/trunk/lib/Target/X86/X86IntrinsicsInfo.h
@@ -1679,6 +1679,7 @@
   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(sse41_round_pd,    ROUNDP, X86ISD::VRNDSCALE, 0),
   X86_INTRINSIC_DATA(sse41_round_ps,    ROUNDP, X86ISD::VRNDSCALE, 0),
Index: llvm/trunk/test/CodeGen/X86/horizontal-reduce-smax.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/horizontal-reduce-smax.ll
+++ llvm/trunk/test/CodeGen/X86/horizontal-reduce-smax.ll
@@ -196,52 +196,68 @@
 }
 
 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
@@ -719,30 +735,39 @@
 }
 
 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
@@ -751,42 +776,49 @@
 ; 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
@@ -795,13 +827,11 @@
 ; 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
@@ -810,13 +840,11 @@
 ; 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
@@ -1511,21 +1539,34 @@
 }
 
 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:
@@ -1534,12 +1575,10 @@
 ; 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
@@ -1549,33 +1588,44 @@
 ; 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:
@@ -1584,12 +1634,10 @@
 ; 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
@@ -1599,13 +1647,11 @@
 ; 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
@@ -1614,15 +1660,13 @@
 ; 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
Index: llvm/trunk/test/CodeGen/X86/horizontal-reduce-smin.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/horizontal-reduce-smin.ll
+++ llvm/trunk/test/CodeGen/X86/horizontal-reduce-smin.ll
@@ -198,52 +198,68 @@
 }
 
 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
@@ -723,30 +739,39 @@
 }
 
 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
@@ -755,42 +780,49 @@
 ; 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
@@ -799,13 +831,11 @@
 ; 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
@@ -814,13 +844,11 @@
 ; 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
@@ -1513,21 +1541,34 @@
 }
 
 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:
@@ -1536,12 +1577,10 @@
 ; 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
@@ -1551,33 +1590,44 @@
 ; 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:
@@ -1586,12 +1636,10 @@
 ; 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
@@ -1601,13 +1649,11 @@
 ; 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
@@ -1616,15 +1662,13 @@
 ; 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
Index: llvm/trunk/test/CodeGen/X86/horizontal-reduce-umax.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/horizontal-reduce-umax.ll
+++ llvm/trunk/test/CodeGen/X86/horizontal-reduce-umax.ll
@@ -259,25 +259,20 @@
 ;
 ; 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
@@ -318,25 +313,20 @@
 ;
 ; 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
@@ -879,14 +869,11 @@
 ; 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
 ;
@@ -894,12 +881,10 @@
 ; 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
@@ -908,13 +893,11 @@
 ; 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
@@ -965,14 +948,11 @@
 ; 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
 ;
@@ -980,12 +960,10 @@
 ; 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
@@ -994,13 +972,11 @@
 ; 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
@@ -1009,13 +985,11 @@
 ; 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
@@ -1821,14 +1795,11 @@
 ; 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
 ;
@@ -1839,12 +1810,10 @@
 ; 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
@@ -1854,13 +1823,11 @@
 ; 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
@@ -1929,14 +1896,11 @@
 ; 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
 ;
@@ -1947,12 +1911,10 @@
 ; 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
@@ -1962,13 +1924,11 @@
 ; 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
@@ -1977,15 +1937,13 @@
 ; 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
Index: llvm/trunk/test/CodeGen/X86/horizontal-reduce-umin.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/horizontal-reduce-umin.ll
+++ llvm/trunk/test/CodeGen/X86/horizontal-reduce-umin.ll
@@ -261,25 +261,14 @@
 ;
 ; 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
@@ -320,25 +309,14 @@
 ;
 ; 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
@@ -885,14 +863,8 @@
 ; 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
 ;
@@ -900,12 +872,7 @@
 ; 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
@@ -914,13 +881,8 @@
 ; 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
@@ -971,14 +933,8 @@
 ; 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
 ;
@@ -986,12 +942,7 @@
 ; 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
@@ -1000,13 +951,8 @@
 ; 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
@@ -1015,13 +961,8 @@
 ; 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
@@ -1825,14 +1766,8 @@
 ; 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
 ;
@@ -1843,12 +1778,7 @@
 ; 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
@@ -1858,13 +1788,8 @@
 ; 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
@@ -1933,14 +1858,8 @@
 ; 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
 ;
@@ -1951,12 +1870,7 @@
 ; 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
@@ -1966,13 +1880,8 @@
 ; 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
@@ -1981,15 +1890,10 @@
 ; 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