[X86][SSE2] Follow up patch for D87236 to avoid the regression regarding horizontal unsigned 16 bit minimums and maximums.

[RKSimon]

Summary

This is a follow up patch for D87236 based on the comments about the test regressions.

The new code is even slightly better than it was originally because the first sign bit flip is pulled before the first shuffle. Maybe this should also be done for lowering for 8 bit signed minimums and maximums.

An entirely alternative approach to fix the regression might be to switch to the "sign bit flip" method depending on some heuristic based on the number of chained UMIN/UMAX instructions. For example when calculating the median of 3 integers via max(min(x, y), min(max(x, y), z)), we also want to use the old way, I think. However this is more complex to implement.

[RKSimon]

llvm/lib/Target/X86/X86ISelLowering.cpp

39249

Is it worth pulling this in as well?

39282

clang-format this

[RKSimon]

We might be better off waiting until D88787 lands then we can decide if we want to use shouldExpandReduction() to allow us to handle these reduction opcodes in dag directly

[RKSimon]

@TomHender Are you able to continue with this, otherwise would you mind if I comandeered this please?

llvm/lib/Target/X86/X86ISelLowering.cpp

static SDValue createPSADBW(SelectionDAG &DAG, const SDValue &Zext0,
   return SplitOpsAndApply(DAG, Subtarget, DL, SadVT, { SadOp0, SadOp1 },
                           PSADBWBuilder);
 }
 
 // Attempt to replace an min/max v8i16/v16i8 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 && ExtractVT != MVT::i8)
     return SDValue();
 
   // Check for SMAX/SMIN/UMAX/UMIN horizontal reduction patterns.
   ISD::NodeType BinOp;
   SDValue Src = DAG.matchBinOpReduction(
       Extract, BinOp, {ISD::SMAX, ISD::SMIN, ISD::UMAX, ISD::UMIN}, true);
   if (!Src)
     return SDValue();
 
   EVT SrcVT = Src.getValueType();
   EVT SrcSVT = SrcVT.getScalarType();
   if (SrcSVT != ExtractVT || (SrcVT.getSizeInBits() % 128) != 0)
     return SDValue();
 
   SDLoc DL(Extract);
-  SDValue MinPos = Src;
+  SDValue Result = Src;
 
   // First, reduce the source down to 128-bit, applying BinOp to lo/hi.
   while (SrcVT.getSizeInBits() > 128) {
     SDValue Lo, Hi;
-    std::tie(Lo, Hi) = splitVector(MinPos, DAG, DL);
+    std::tie(Lo, Hi) = splitVector(Result, DAG, DL);
     SrcVT = Lo.getValueType();
-    MinPos = DAG.getNode(BinOp, DL, SrcVT, Lo, Hi);
+    Result = DAG.getNode(BinOp, DL, SrcVT, Lo, Hi);
   }
   assert(((SrcVT == MVT::v8i16 && ExtractVT == MVT::i16) ||
           (SrcVT == MVT::v16i8 && ExtractVT == MVT::i8)) &&
          "Unexpected value type");
 
+  // PHMINPOSUW is only available when SSE41 is supported. If it is not supported
+  // we just rewrite UMIN/UMAX of v8i16 into SMIN/SMAX to avoid the inefficient
+  // lowering into pairs of PSUBUSW and PADDW/PSUBW.
+  if (!Subtarget.hasSSE41())
+  {
+    if (SrcVT != MVT::v8i16 || (BinOp != ISD::UMAX && BinOp != ISD::UMIN))
+      return SDValue();
+    SDValue Mask = DAG.getConstant(
+        APInt::getSignedMinValue(ExtractVT.getSizeInBits()), DL, SrcVT);
+    Result = DAG.getNode(ISD::XOR, DL, SrcVT, Mask, Result);
+    
+    // Emit a reduction tree with SMIN/SMAX operations.
+    ISD::NodeType SignedBinOp = BinOp == ISD::UMAX ? ISD::SMAX : ISD::SMIN;
+    Result = DAG.getNode(SignedBinOp, DL, SrcVT, 
+        DAG.getVectorShuffle(SrcVT, DL, Result, DAG.getUNDEF(SrcVT),
+            {4, 5, 6, 7, -1, -1, -1, -1}), Result);
+    Result = DAG.getNode(SignedBinOp, DL, SrcVT, 
+        DAG.getVectorShuffle(SrcVT, DL, Result, DAG.getUNDEF(SrcVT),
+            {2, 3, -1, -1, -1, -1, -1, -1}), Result);
+    Result = DAG.getNode(SignedBinOp, DL, SrcVT, 
+        DAG.getVectorShuffle(SrcVT, DL, Result, DAG.getUNDEF(SrcVT),
+            {1, -1, -1, -1, -1, -1, -1, -1}), Result);
+
+    // Flip the sign bit back and return.
+    Result = DAG.getNode(ISD::XOR, DL, SrcVT, Mask, Result);
+    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtractVT, Result,
+        DAG.getIntPtrConstant(0, DL));
+  }
+
   // PHMINPOSUW applies to UMIN(v8i16), for SMIN/SMAX/UMAX we must apply a mask
   // to flip the value accordingly.
   SDValue Mask;
   unsigned MaskEltsBits = ExtractVT.getSizeInBits();
   if (BinOp == ISD::SMAX)
     Mask = DAG.getConstant(APInt::getSignedMaxValue(MaskEltsBits), DL, SrcVT);
   else if (BinOp == ISD::SMIN)
     Mask = DAG.getConstant(APInt::getSignedMinValue(MaskEltsBits), DL, SrcVT);
   else if (BinOp == ISD::UMAX)
     Mask = DAG.getConstant(APInt::getAllOnesValue(MaskEltsBits), DL, SrcVT);
 
   if (Mask)
-    MinPos = DAG.getNode(ISD::XOR, DL, SrcVT, Mask, MinPos);
+    Result = DAG.getNode(ISD::XOR, DL, SrcVT, Mask, Result);
 
   // For v16i8 cases we need to perform UMIN on pairs of byte elements,
   // shuffling each upper element down and insert zeros. This means that the
   // v16i8 UMIN will leave the upper element as zero, performing zero-extension
   // ready for the PHMINPOS.
   if (ExtractVT == MVT::i8) {
     SDValue Upper = DAG.getVectorShuffle(
-        SrcVT, DL, MinPos, DAG.getConstant(0, DL, MVT::v16i8),
+        SrcVT, DL, Result, DAG.getConstant(0, DL, MVT::v16i8),
         {1, 16, 3, 16, 5, 16, 7, 16, 9, 16, 11, 16, 13, 16, 15, 16});
-    MinPos = DAG.getNode(ISD::UMIN, DL, SrcVT, MinPos, Upper);
+    Result = DAG.getNode(ISD::UMIN, DL, SrcVT, Result, Upper);
   }
 
   // Perform the PHMINPOS on a v8i16 vector,
-  MinPos = DAG.getBitcast(MVT::v8i16, MinPos);
-  MinPos = DAG.getNode(X86ISD::PHMINPOS, DL, MVT::v8i16, MinPos);
-  MinPos = DAG.getBitcast(SrcVT, MinPos);
+  Result = DAG.getBitcast(MVT::v8i16, Result);
+  Result = DAG.getNode(X86ISD::PHMINPOS, DL, MVT::v8i16, Result);
+  Result = DAG.getBitcast(SrcVT, Result);
 
   if (Mask)
-    MinPos = DAG.getNode(ISD::XOR, DL, SrcVT, Mask, MinPos);
+    Result = DAG.getNode(ISD::XOR, DL, SrcVT, Mask, Result);
 
-  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtractVT, MinPos,
+  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtractVT, Result,
                      DAG.getIntPtrConstant(0, DL));
 }
 
 // Attempt to replace an all_of/any_of/parity style horizontal reduction with a MOVMSK.
 static SDValue combineHorizontalPredicateResult(SDNode *Extract,
                                                 SelectionDAG &DAG,
                                                 const X86Subtarget &Subtarget) {
   // Bail without SSE2.

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

; X64-AVX-NEXT:    retq
   %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-SSE2-LABEL: test_reduce_v8i16:
 ; X86-SSE2:       ## %bb.0:
+; X86-SSE2-NEXT:    pxor LCPI2_0, %xmm0
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X86-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm0
-; X86-SSE2-NEXT:    paddw %xmm1, %xmm0
+; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X86-SSE2-NEXT:    psrld $16, %xmm1
-; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X86-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v8i16:
 ; X86-SSE42:       ## %bb.0:
 ; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
 ; X86-SSE42-NEXT:    pxor %xmm0, %xmm1
 ; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0

 ; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
 ; X86-AVX-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX-NEXT:    notl %eax
 ; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-AVX-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v8i16:
 ; X64-SSE2:       ## %bb.0:
+; X64-SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X64-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm0
-; X64-SSE2-NEXT:    paddw %xmm1, %xmm0
+; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X64-SSE2-NEXT:    psrld $16, %xmm1
-; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X64-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v8i16:
 ; X64-SSE42:       ## %bb.0:
 ; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
 ; X64-SSE42-NEXT:    pxor %xmm0, %xmm1
 ; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0
; X64-AVX512-NEXT:    retq
   ret i32 %10
 }
 
 define i16 @test_reduce_v16i16(<16 x i16> %a0) {
 ; X86-SSE2-LABEL: test_reduce_v16i16:
 ; X86-SSE2:       ## %bb.0:
 ; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1
 ; X86-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pxor LCPI6_0, %xmm1
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm0
-; X86-SSE2-NEXT:    paddw %xmm1, %xmm0
+; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X86-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0
 ; X86-SSE2-NEXT:    psrld $16, %xmm0
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm0
-; X86-SSE2-NEXT:    paddw %xmm1, %xmm0
+; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; X86-SSE2-NEXT:    movd %xmm0, %eax
+; X86-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v16i16:
 ; X86-SSE42:       ## %bb.0:
 ; X86-SSE42-NEXT:    pmaxuw %xmm1, %xmm0
 ; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
 ; X86-SSE42-NEXT:    pxor %xmm0, %xmm1

 ; X86-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-AVX2-NEXT:    vzeroupper
 ; X86-AVX2-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v16i16:
 ; X64-SSE2:       ## %bb.0:
 ; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1
 ; X64-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pxor {{.*}}(%rip), %xmm1
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm0
-; X64-SSE2-NEXT:    paddw %xmm1, %xmm0
+; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X64-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0
 ; X64-SSE2-NEXT:    psrld $16, %xmm0
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm0
-; X64-SSE2-NEXT:    paddw %xmm1, %xmm0
+; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; X64-SSE2-NEXT:    movd %xmm0, %eax
+; X64-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v16i16:
 ; X64-SSE42:       ## %bb.0:
 ; X64-SSE42-NEXT:    pmaxuw %xmm1, %xmm0
 ; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
 ; X64-SSE42-NEXT:    pxor %xmm0, %xmm1

 ; X86-SSE2-LABEL: test_reduce_v32i16:
 ; X86-SSE2:       ## %bb.0:
 ; X86-SSE2-NEXT:    psubusw %xmm0, %xmm2
 ; X86-SSE2-NEXT:    paddw %xmm0, %xmm2
 ; X86-SSE2-NEXT:    psubusw %xmm1, %xmm3
 ; X86-SSE2-NEXT:    paddw %xmm1, %xmm3
 ; X86-SSE2-NEXT:    psubusw %xmm2, %xmm3
 ; X86-SSE2-NEXT:    paddw %xmm2, %xmm3
+; X86-SSE2-NEXT:    pxor LCPI10_0, %xmm3
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
-; X86-SSE2-NEXT:    psubusw %xmm3, %xmm0
-; X86-SSE2-NEXT:    paddw %xmm3, %xmm0
+; X86-SSE2-NEXT:    pmaxsw %xmm3, %xmm0
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X86-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0
 ; X86-SSE2-NEXT:    psrld $16, %xmm0
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm0
-; X86-SSE2-NEXT:    paddw %xmm1, %xmm0
+; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; X86-SSE2-NEXT:    movd %xmm0, %eax
+; X86-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; 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, %xmm1
 ; X86-SSE42-NEXT:    pmaxuw %xmm0, %xmm1

 ; X64-SSE2-LABEL: test_reduce_v32i16:
 ; X64-SSE2:       ## %bb.0:
 ; X64-SSE2-NEXT:    psubusw %xmm0, %xmm2
 ; X64-SSE2-NEXT:    paddw %xmm0, %xmm2
 ; X64-SSE2-NEXT:    psubusw %xmm1, %xmm3
 ; X64-SSE2-NEXT:    paddw %xmm1, %xmm3
 ; X64-SSE2-NEXT:    psubusw %xmm2, %xmm3
 ; X64-SSE2-NEXT:    paddw %xmm2, %xmm3
+; X64-SSE2-NEXT:    pxor {{.*}}(%rip), %xmm3
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
-; X64-SSE2-NEXT:    psubusw %xmm3, %xmm0
-; X64-SSE2-NEXT:    paddw %xmm3, %xmm0
+; X64-SSE2-NEXT:    pmaxsw %xmm3, %xmm0
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X64-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0
 ; X64-SSE2-NEXT:    psrld $16, %xmm0
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm0
-; X64-SSE2-NEXT:    paddw %xmm1, %xmm0
+; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; X64-SSE2-NEXT:    movd %xmm0, %eax
+; X64-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; 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, %xmm1
 ; X64-SSE42-NEXT:    pmaxuw %xmm0, %xmm1

 
 ;
 ; Partial Vector Reductions
 ;
 
 define i16 @test_reduce_v16i16_v8i16(<16 x i16> %a0) {
 ; X86-SSE2-LABEL: test_reduce_v16i16_v8i16:
 ; X86-SSE2:       ## %bb.0:
+; X86-SSE2-NEXT:    pxor LCPI12_0, %xmm0
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X86-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm0
-; X86-SSE2-NEXT:    paddw %xmm1, %xmm0
+; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X86-SSE2-NEXT:    psrld $16, %xmm1
-; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X86-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v16i16_v8i16:
 ; X86-SSE42:       ## %bb.0:
 ; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
 ; X86-SSE42-NEXT:    pxor %xmm0, %xmm1
 ; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0

 ; X86-AVX-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX-NEXT:    notl %eax
 ; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-AVX-NEXT:    vzeroupper
 ; X86-AVX-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v16i16_v8i16:
 ; X64-SSE2:       ## %bb.0:
+; X64-SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X64-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm0
-; X64-SSE2-NEXT:    paddw %xmm1, %xmm0
+; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X64-SSE2-NEXT:    psrld $16, %xmm1
-; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X64-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v16i16_v8i16:
 ; X64-SSE42:       ## %bb.0:
 ; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
 ; X64-SSE42-NEXT:    pxor %xmm0, %xmm1
 ; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0
; X64-AVX512-NEXT:    retq
   %9  = select <16 x i1> %8, <16 x i16> %6, <16 x i16> %7
   %10 = extractelement <16 x i16> %9, i32 0
   ret i16 %10
 }
 
 define i16 @test_reduce_v32i16_v8i16(<32 x i16> %a0) {
 ; X86-SSE2-LABEL: test_reduce_v32i16_v8i16:
 ; X86-SSE2:       ## %bb.0:
+; X86-SSE2-NEXT:    pxor LCPI13_0, %xmm0
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X86-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm0
-; X86-SSE2-NEXT:    paddw %xmm1, %xmm0
+; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X86-SSE2-NEXT:    psrld $16, %xmm1
-; X86-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X86-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v32i16_v8i16:
 ; X86-SSE42:       ## %bb.0:
 ; X86-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
 ; X86-SSE42-NEXT:    pxor %xmm0, %xmm1
 ; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0

 ; X86-AVX-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX-NEXT:    notl %eax
 ; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-AVX-NEXT:    vzeroupper
 ; X86-AVX-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v32i16_v8i16:
 ; X64-SSE2:       ## %bb.0:
+; X64-SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X64-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm0
-; X64-SSE2-NEXT:    paddw %xmm1, %xmm0
+; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X64-SSE2-NEXT:    psrld $16, %xmm1
-; X64-SSE2-NEXT:    psubusw %xmm0, %xmm1
-; X64-SSE2-NEXT:    paddw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v32i16_v8i16:
 ; X64-SSE42:       ## %bb.0:
 ; X64-SSE42-NEXT:    pcmpeqd %xmm1, %xmm1
 ; X64-SSE42-NEXT:    pxor %xmm0, %xmm1
 ; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0

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

; X64-AVX-NEXT:    retq
   %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-SSE2-LABEL: test_reduce_v8i16:
 ; X86-SSE2:       ## %bb.0:
+; X86-SSE2-NEXT:    pxor LCPI2_0, %xmm0
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X86-SSE2-NEXT:    psrld $16, %xmm1
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X86-SSE2-NEXT:    movd %xmm0, %eax
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v8i16:
 ; X86-SSE42:       ## %bb.0:
 ; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
 ; X86-SSE42-NEXT:    movd %xmm0, %eax
 ; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX-LABEL: test_reduce_v8i16:
 ; X86-AVX:       ## %bb.0:
 ; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
 ; X86-AVX-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-AVX-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v8i16:
 ; X64-SSE2:       ## %bb.0:
+; X64-SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X64-SSE2-NEXT:    psrld $16, %xmm1
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X64-SSE2-NEXT:    movd %xmm0, %eax
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v8i16:
 ; X64-SSE42:       ## %bb.0:
 ; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
 ; X64-SSE42-NEXT:    movd %xmm0, %eax
 ; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax

 }
 
 define i16 @test_reduce_v16i16(<16 x i16> %a0) {
 ; X86-SSE2-LABEL: test_reduce_v16i16:
 ; X86-SSE2:       ## %bb.0:
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
 ; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
 ; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X86-SSE2-NEXT:    pxor LCPI6_0, %xmm0
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X86-SSE2-NEXT:    psrld $16, %xmm1
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X86-SSE2-NEXT:    movd %xmm0, %eax
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v16i16:
 ; X86-SSE42:       ## %bb.0:
 ; X86-SSE42-NEXT:    pminuw %xmm1, %xmm0
 ; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
 ; X86-SSE42-NEXT:    movd %xmm0, %eax

 ; X86-AVX2-NEXT:    vzeroupper
 ; X86-AVX2-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v16i16:
 ; X64-SSE2:       ## %bb.0:
 ; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
 ; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
 ; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X64-SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X64-SSE2-NEXT:    psrld $16, %xmm1
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X64-SSE2-NEXT:    movd %xmm0, %eax
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v16i16:
 ; X64-SSE42:       ## %bb.0:
 ; X64-SSE42-NEXT:    pminuw %xmm1, %xmm0
 ; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
 ; X64-SSE42-NEXT:    movd %xmm0, %eax

 ; X86-SSE2-NEXT:    psubusw %xmm3, %xmm4
 ; X86-SSE2-NEXT:    psubw %xmm4, %xmm1
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm3
 ; X86-SSE2-NEXT:    psubusw %xmm2, %xmm3
 ; X86-SSE2-NEXT:    psubw %xmm3, %xmm0
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
 ; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
 ; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X86-SSE2-NEXT:    pxor LCPI10_0, %xmm0
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X86-SSE2-NEXT:    psrld $16, %xmm1
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X86-SSE2-NEXT:    movd %xmm0, %eax
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; 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, %xmm1
 ; X86-SSE42-NEXT:    pminuw %xmm0, %xmm1

 ; X64-SSE2-NEXT:    psubusw %xmm3, %xmm4
 ; X64-SSE2-NEXT:    psubw %xmm4, %xmm1
 ; X64-SSE2-NEXT:    movdqa %xmm0, %xmm3
 ; X64-SSE2-NEXT:    psubusw %xmm2, %xmm3
 ; X64-SSE2-NEXT:    psubw %xmm3, %xmm0
 ; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
 ; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
 ; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X64-SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X64-SSE2-NEXT:    psrld $16, %xmm1
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X64-SSE2-NEXT:    movd %xmm0, %eax
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; 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, %xmm1
 ; X64-SSE42-NEXT:    pminuw %xmm0, %xmm1

 
 ;
 ; Partial Vector Reductions
 ;
 
 define i16 @test_reduce_v16i16_v8i16(<16 x i16> %a0) {
 ; X86-SSE2-LABEL: test_reduce_v16i16_v8i16:
 ; X86-SSE2:       ## %bb.0:
+; X86-SSE2-NEXT:    pxor LCPI12_0, %xmm0
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X86-SSE2-NEXT:    psrld $16, %xmm1
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X86-SSE2-NEXT:    movd %xmm0, %eax
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v16i16_v8i16:
 ; X86-SSE42:       ## %bb.0:
 ; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
 ; X86-SSE42-NEXT:    movd %xmm0, %eax
 ; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX-LABEL: test_reduce_v16i16_v8i16:
 ; X86-AVX:       ## %bb.0:
 ; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
 ; X86-AVX-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-AVX-NEXT:    vzeroupper
 ; X86-AVX-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v16i16_v8i16:
 ; X64-SSE2:       ## %bb.0:
+; X64-SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X64-SSE2-NEXT:    psrld $16, %xmm1
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X64-SSE2-NEXT:    movd %xmm0, %eax
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v16i16_v8i16:
 ; X64-SSE42:       ## %bb.0:
 ; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
 ; X64-SSE42-NEXT:    movd %xmm0, %eax
 ; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
; X64-AVX-NEXT:    retq
   %9  = select <16 x i1> %8, <16 x i16> %6, <16 x i16> %7
   %10 = extractelement <16 x i16> %9, i32 0
   ret i16 %10
 }
 
 define i16 @test_reduce_v32i16_v8i16(<32 x i16> %a0) {
 ; X86-SSE2-LABEL: test_reduce_v32i16_v8i16:
 ; X86-SSE2:       ## %bb.0:
+; X86-SSE2-NEXT:    pxor LCPI13_0, %xmm0
 ; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X86-SSE2-NEXT:    psrld $16, %xmm1
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X86-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X86-SSE2-NEXT:    movd %xmm0, %eax
+; X86-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X86-SSE2-NEXT:    movd %xmm1, %eax
+; X86-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-SSE2-NEXT:    retl
 ;
 ; X86-SSE42-LABEL: test_reduce_v32i16_v8i16:
 ; X86-SSE42:       ## %bb.0:
 ; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
 ; X86-SSE42-NEXT:    movd %xmm0, %eax
 ; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-SSE42-NEXT:    retl
 ;
 ; X86-AVX-LABEL: test_reduce_v32i16_v8i16:
 ; X86-AVX:       ## %bb.0:
 ; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
 ; X86-AVX-NEXT:    vmovd %xmm0, %eax
 ; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X86-AVX-NEXT:    vzeroupper
 ; X86-AVX-NEXT:    retl
 ;
 ; X64-SSE2-LABEL: test_reduce_v32i16_v8i16:
 ; X64-SSE2:       ## %bb.0:
+; X64-SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X64-SSE2-NEXT:    psrld $16, %xmm1
-; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X64-SSE2-NEXT:    psubusw %xmm1, %xmm2
-; X64-SSE2-NEXT:    psubw %xmm2, %xmm0
-; X64-SSE2-NEXT:    movd %xmm0, %eax
+; X64-SSE2-NEXT:    pminsw %xmm0, %xmm1
+; X64-SSE2-NEXT:    movd %xmm1, %eax
+; X64-SSE2-NEXT:    xorl $32768, %eax ## imm = 0x8000
 ; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
 ; X64-SSE2-NEXT:    retq
 ;
 ; X64-SSE42-LABEL: test_reduce_v32i16_v8i16:
 ; X64-SSE42:       ## %bb.0:
 ; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
 ; X64-SSE42-NEXT:    movd %xmm0, %eax
 ; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax

llvm/test/CodeGen/X86/vector-reduce-umax.ll

 ; AVX512-NEXT:    retq
   %1 = call i16 @llvm.experimental.vector.reduce.umax.v4i16(<4 x i16> %a0)
   ret i16 %1
 }
 
 define i16 @test_v8i16(<8 x i16> %a0) {
 ; SSE2-LABEL: test_v8i16:
 ; SSE2:       # %bb.0:
+; SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT:    psubusw %xmm0, %xmm1
-; SSE2-NEXT:    paddw %xmm0, %xmm1
+; SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE2-NEXT:    psubusw %xmm1, %xmm0
-; SSE2-NEXT:    paddw %xmm1, %xmm0
+; SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrld $16, %xmm1
-; SSE2-NEXT:    psubusw %xmm0, %xmm1
-; SSE2-NEXT:    paddw %xmm0, %xmm1
+; SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; SSE2-NEXT:    movd %xmm1, %eax
+; SSE2-NEXT:    xorl $32768, %eax # imm = 0x8000
 ; SSE2-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: test_v8i16:
 ; SSE4:       # %bb.0:
 ; SSE4-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE4-NEXT:    pxor %xmm0, %xmm1
 ; SSE4-NEXT:    phminposuw %xmm1, %xmm0
; AVX512VL-NEXT:    retq
   ret i16 %1
 }
 
 define i16 @test_v16i16(<16 x i16> %a0) {
 ; SSE2-LABEL: test_v16i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    psubusw %xmm0, %xmm1
 ; SSE2-NEXT:    paddw %xmm0, %xmm1
+; SSE2-NEXT:    pxor {{.*}}(%rip), %xmm1
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE2-NEXT:    psubusw %xmm1, %xmm0
-; SSE2-NEXT:    paddw %xmm1, %xmm0
+; SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT:    psubusw %xmm0, %xmm1
-; SSE2-NEXT:    paddw %xmm0, %xmm1
+; SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; SSE2-NEXT:    movdqa %xmm1, %xmm0
 ; SSE2-NEXT:    psrld $16, %xmm0
-; SSE2-NEXT:    psubusw %xmm1, %xmm0
-; SSE2-NEXT:    paddw %xmm1, %xmm0
+; SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; SSE2-NEXT:    movd %xmm0, %eax
+; SSE2-NEXT:    xorl $32768, %eax # imm = 0x8000
 ; SSE2-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: test_v16i16:
 ; SSE4:       # %bb.0:
 ; SSE4-NEXT:    pmaxuw %xmm1, %xmm0
 ; SSE4-NEXT:    pcmpeqd %xmm1, %xmm1
 ; SSE4-NEXT:    pxor %xmm0, %xmm1

 ; SSE2-LABEL: test_v32i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    psubusw %xmm0, %xmm2
 ; SSE2-NEXT:    paddw %xmm0, %xmm2
 ; SSE2-NEXT:    psubusw %xmm1, %xmm3
 ; SSE2-NEXT:    paddw %xmm1, %xmm3
 ; SSE2-NEXT:    psubusw %xmm2, %xmm3
 ; SSE2-NEXT:    paddw %xmm2, %xmm3
+; SSE2-NEXT:    pxor {{.*}}(%rip), %xmm3
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
-; SSE2-NEXT:    psubusw %xmm3, %xmm0
-; SSE2-NEXT:    paddw %xmm3, %xmm0
+; SSE2-NEXT:    pmaxsw %xmm3, %xmm0
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT:    psubusw %xmm0, %xmm1
-; SSE2-NEXT:    paddw %xmm0, %xmm1
+; SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; SSE2-NEXT:    movdqa %xmm1, %xmm0
 ; SSE2-NEXT:    psrld $16, %xmm0
-; SSE2-NEXT:    psubusw %xmm1, %xmm0
-; SSE2-NEXT:    paddw %xmm1, %xmm0
+; SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; SSE2-NEXT:    movd %xmm0, %eax
+; SSE2-NEXT:    xorl $32768, %eax # imm = 0x8000
 ; SSE2-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: test_v32i16:
 ; SSE4:       # %bb.0:
 ; SSE4-NEXT:    pmaxuw %xmm3, %xmm1
 ; SSE4-NEXT:    pmaxuw %xmm2, %xmm1
 ; SSE4-NEXT:    pmaxuw %xmm0, %xmm1

 ; AVX512VL-NEXT:    retq
   %1 = call i16 @llvm.experimental.vector.reduce.umax.v32i16(<32 x i16> %a0)
   ret i16 %1
 }
 
 define i16 @test_v64i16(<64 x i16> %a0) {
 ; SSE2-LABEL: test_v64i16:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    psubusw %xmm1, %xmm5
-; SSE2-NEXT:    paddw %xmm1, %xmm5
-; SSE2-NEXT:    psubusw %xmm3, %xmm7
-; SSE2-NEXT:    paddw %xmm3, %xmm7
 ; SSE2-NEXT:    psubusw %xmm0, %xmm4
 ; SSE2-NEXT:    paddw %xmm0, %xmm4
 ; SSE2-NEXT:    psubusw %xmm2, %xmm6
 ; SSE2-NEXT:    paddw %xmm2, %xmm6
 ; SSE2-NEXT:    psubusw %xmm4, %xmm6
 ; SSE2-NEXT:    paddw %xmm4, %xmm6
+; SSE2-NEXT:    psubusw %xmm1, %xmm5
+; SSE2-NEXT:    paddw %xmm1, %xmm5
+; SSE2-NEXT:    psubusw %xmm3, %xmm7
+; SSE2-NEXT:    paddw %xmm3, %xmm7
 ; SSE2-NEXT:    psubusw %xmm5, %xmm7
 ; SSE2-NEXT:    paddw %xmm5, %xmm7
 ; SSE2-NEXT:    psubusw %xmm6, %xmm7
 ; SSE2-NEXT:    paddw %xmm6, %xmm7
+; SSE2-NEXT:    pxor {{.*}}(%rip), %xmm7
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[2,3,2,3]
-; SSE2-NEXT:    psubusw %xmm7, %xmm0
-; SSE2-NEXT:    paddw %xmm7, %xmm0
+; SSE2-NEXT:    pmaxsw %xmm7, %xmm0
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT:    psubusw %xmm0, %xmm1
-; SSE2-NEXT:    paddw %xmm0, %xmm1
+; SSE2-NEXT:    pmaxsw %xmm0, %xmm1
 ; SSE2-NEXT:    movdqa %xmm1, %xmm0
 ; SSE2-NEXT:    psrld $16, %xmm0
-; SSE2-NEXT:    psubusw %xmm1, %xmm0
-; SSE2-NEXT:    paddw %xmm1, %xmm0
+; SSE2-NEXT:    pmaxsw %xmm1, %xmm0
 ; SSE2-NEXT:    movd %xmm0, %eax
+; SSE2-NEXT:    xorl $32768, %eax # imm = 0x8000
 ; SSE2-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: test_v64i16:
 ; SSE4:       # %bb.0:
 ; SSE4-NEXT:    pmaxuw %xmm7, %xmm3
 ; SSE4-NEXT:    pmaxuw %xmm5, %xmm3
 ; SSE4-NEXT:    pmaxuw %xmm1, %xmm3

llvm/test/CodeGen/X86/vector-reduce-umin.ll

 ; AVX512-NEXT:    retq
   %1 = call i16 @llvm.experimental.vector.reduce.umin.v4i16(<4 x i16> %a0)
   ret i16 %1
 }
 
 define i16 @test_v8i16(<8 x i16> %a0) {
 ; SSE2-LABEL: test_v8i16:
 ; SSE2:       # %bb.0:
+; SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psubusw %xmm1, %xmm2
-; SSE2-NEXT:    psubw %xmm2, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psubusw %xmm1, %xmm2
-; SSE2-NEXT:    psubw %xmm2, %xmm0
+; SSE2-NEXT:    pminsw %xmm0, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrld $16, %xmm1
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psubusw %xmm1, %xmm2
-; SSE2-NEXT:    psubw %xmm2, %xmm0
-; SSE2-NEXT:    movd %xmm0, %eax
+; SSE2-NEXT:    pminsw %xmm0, %xmm1
+; SSE2-NEXT:    movd %xmm1, %eax
+; SSE2-NEXT:    xorl $32768, %eax # imm = 0x8000
 ; SSE2-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: test_v8i16:
 ; SSE4:       # %bb.0:
 ; SSE4-NEXT:    phminposuw %xmm0, %xmm0
 ; SSE4-NEXT:    movd %xmm0, %eax
 ; SSE4-NEXT:    # kill: def $ax killed $ax killed $eax

 }
 
 define i16 @test_v16i16(<16 x i16> %a0) {
 ; SSE2-LABEL: test_v16i16:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa %xmm0, %xmm2
 ; SSE2-NEXT:    psubusw %xmm1, %xmm2
 ; SSE2-NEXT:    psubw %xmm2, %xmm0
+; SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psubusw %xmm1, %xmm2
-; SSE2-NEXT:    psubw %xmm2, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psubusw %xmm1, %xmm2
-; SSE2-NEXT:    psubw %xmm2, %xmm0
+; SSE2-NEXT:    pminsw %xmm0, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrld $16, %xmm1
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psubusw %xmm1, %xmm2
-; SSE2-NEXT:    psubw %xmm2, %xmm0
-; SSE2-NEXT:    movd %xmm0, %eax
+; SSE2-NEXT:    pminsw %xmm0, %xmm1
+; SSE2-NEXT:    movd %xmm1, %eax
+; SSE2-NEXT:    xorl $32768, %eax # imm = 0x8000
 ; SSE2-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: test_v16i16:
 ; SSE4:       # %bb.0:
 ; SSE4-NEXT:    pminuw %xmm1, %xmm0
 ; SSE4-NEXT:    phminposuw %xmm0, %xmm0
 ; SSE4-NEXT:    movd %xmm0, %eax

 ; SSE2-NEXT:    psubusw %xmm3, %xmm4
 ; SSE2-NEXT:    psubw %xmm4, %xmm1
 ; SSE2-NEXT:    movdqa %xmm0, %xmm3
 ; SSE2-NEXT:    psubusw %xmm2, %xmm3
 ; SSE2-NEXT:    psubw %xmm3, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm2
 ; SSE2-NEXT:    psubusw %xmm1, %xmm2
 ; SSE2-NEXT:    psubw %xmm2, %xmm0
+; SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psubusw %xmm1, %xmm2
-; SSE2-NEXT:    psubw %xmm2, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psubusw %xmm1, %xmm2
-; SSE2-NEXT:    psubw %xmm2, %xmm0
+; SSE2-NEXT:    pminsw %xmm0, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrld $16, %xmm1
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psubusw %xmm1, %xmm2
-; SSE2-NEXT:    psubw %xmm2, %xmm0
-; SSE2-NEXT:    movd %xmm0, %eax
+; SSE2-NEXT:    pminsw %xmm0, %xmm1
+; SSE2-NEXT:    movd %xmm1, %eax
+; SSE2-NEXT:    xorl $32768, %eax # imm = 0x8000
 ; SSE2-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: test_v32i16:
 ; SSE4:       # %bb.0:
 ; SSE4-NEXT:    pminuw %xmm3, %xmm1
 ; SSE4-NEXT:    pminuw %xmm2, %xmm1
 ; SSE4-NEXT:    pminuw %xmm0, %xmm1

 ; AVX512-NEXT:    retq
   %1 = call i16 @llvm.experimental.vector.reduce.umin.v32i16(<32 x i16> %a0)
   ret i16 %1
 }
 
 define i16 @test_v64i16(<64 x i16> %a0) {
 ; SSE2-LABEL: test_v64i16:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm2, %xmm8
-; SSE2-NEXT:    psubusw %xmm6, %xmm8
-; SSE2-NEXT:    psubw %xmm8, %xmm2
-; SSE2-NEXT:    movdqa %xmm0, %xmm6
-; SSE2-NEXT:    psubusw %xmm4, %xmm6
-; SSE2-NEXT:    psubw %xmm6, %xmm0
-; SSE2-NEXT:    movdqa %xmm3, %xmm4
-; SSE2-NEXT:    psubusw %xmm7, %xmm4
-; SSE2-NEXT:    psubw %xmm4, %xmm3
-; SSE2-NEXT:    movdqa %xmm1, %xmm4
-; SSE2-NEXT:    psubusw %xmm5, %xmm4
-; SSE2-NEXT:    psubw %xmm4, %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm4
-; SSE2-NEXT:    psubusw %xmm3, %xmm4
-; SSE2-NEXT:    psubw %xmm4, %xmm1
+; SSE2-NEXT:    movdqa %xmm3, %xmm8
+; SSE2-NEXT:    psubusw %xmm7, %xmm8
+; SSE2-NEXT:    psubw %xmm8, %xmm3
+; SSE2-NEXT:    movdqa %xmm1, %xmm7
+; SSE2-NEXT:    psubusw %xmm5, %xmm7
+; SSE2-NEXT:    psubw %xmm7, %xmm1
+; SSE2-NEXT:    movdqa %xmm1, %xmm5
+; SSE2-NEXT:    psubusw %xmm3, %xmm5
+; SSE2-NEXT:    psubw %xmm5, %xmm1
+; SSE2-NEXT:    movdqa %xmm2, %xmm3
+; SSE2-NEXT:    psubusw %xmm6, %xmm3
+; SSE2-NEXT:    psubw %xmm3, %xmm2
+; SSE2-NEXT:    movdqa %xmm0, %xmm3
+; SSE2-NEXT:    psubusw %xmm4, %xmm3
+; SSE2-NEXT:    psubw %xmm3, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm3
 ; SSE2-NEXT:    psubusw %xmm2, %xmm3
 ; SSE2-NEXT:    psubw %xmm3, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm2
 ; SSE2-NEXT:    psubusw %xmm1, %xmm2
 ; SSE2-NEXT:    psubw %xmm2, %xmm0
+; SSE2-NEXT:    pxor {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psubusw %xmm1, %xmm2
-; SSE2-NEXT:    psubw %xmm2, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psubusw %xmm1, %xmm2
-; SSE2-NEXT:    psubw %xmm2, %xmm0
+; SSE2-NEXT:    pminsw %xmm0, %xmm1
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE2-NEXT:    pminsw %xmm1, %xmm0
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; SSE2-NEXT:    psrld $16, %xmm1
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psubusw %xmm1, %xmm2
-; SSE2-NEXT:    psubw %xmm2, %xmm0
-; SSE2-NEXT:    movd %xmm0, %eax
+; SSE2-NEXT:    pminsw %xmm0, %xmm1
+; SSE2-NEXT:    movd %xmm1, %eax
+; SSE2-NEXT:    xorl $32768, %eax # imm = 0x8000
 ; SSE2-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE2-NEXT:    retq
 ;
 ; SSE4-LABEL: test_v64i16:
 ; SSE4:       # %bb.0:
 ; SSE4-NEXT:    pminuw %xmm7, %xmm3
 ; SSE4-NEXT:    pminuw %xmm5, %xmm3
 ; SSE4-NEXT:    pminuw %xmm1, %xmm3