[X86] Call SimplifyDemandedBits on conditions of X86ISD::SHRUNKBLEND

[craig.topper]

Summary

There can be additional simplifications after VSELECT is converted to SHRUNKBLEND. So we should call SimplifyDemandedBits

This should help some of the regressions from D56387

[RKSimon]

lib/Target/X86/X86ISelLowering.cpp

34318 ↗

(On Diff #180588)

Would it be any better to add X86ISD::SHRUNKBLEND handling to X86TargetLowering::SimplifyDemandedBitsForTargetNode directly?

[RKSimon]

Also, should we remove the similar SimplifyDemandedBits call in combineVSelectToShrunkBlend?

[craig.topper]

In D56421#1349790, @RKSimon wrote:

Also, should we remove the similar SimplifyDemandedBits call in combineVSelectToShrunkBlend?

The one that decides when to turn VSELECT into SHRUNKBLEND? I don't think we can remove that. We might be able to merge this with it and turn SHRUNKBLEND into SHRUNKBLEND again if another simplification is made. That would probably help with cases where the condition is used multiple times.

[craig.topper]

Reuse combineVSelectToShrunkBlend. This reduces the code and should allow better optimization when the select condition is used by multiple selects.

[RKSimon]

LGTM - thanks

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

static SDValue combineSelectOfTwoConstants(SDNode *N, SelectionDAG &DAG) {
   }
 
   return SDValue();
 }
 
 /// If this is a *dynamic* select (non-constant condition) and we can match
 /// this node with one of the variable blend instructions, restructure the
 /// condition so that blends can use the high (sign) bit of each element.
+/// This function will also call SimplfiyDemandedBits on already created
+/// SHRUNKBLENDS to perform additional simplifications.
 static SDValue combineVSelectToShrunkBlend(SDNode *N, SelectionDAG &DAG,
                                            TargetLowering::DAGCombinerInfo &DCI,
                                            const X86Subtarget &Subtarget) {
   SDValue Cond = N->getOperand(0);
-  if (N->getOpcode() != ISD::VSELECT ||
+  if ((N->getOpcode() != ISD::VSELECT &&
+       N->getOpcode() != X86ISD::SHRUNKBLEND) ||
       ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
     return SDValue();
 
   // Don't optimize before the condition has been transformed to a legal type
   // and don't ever optimize vector selects that map to AVX512 mask-registers.
   unsigned BitWidth = Cond.getScalarValueSizeInBits();
   if (BitWidth < 8 || BitWidth > 64)
     return SDValue();
if (VT == MVT::v32i8 && !Subtarget.hasAVX2())
     return SDValue();
   // There are no 512-bit blend instructions that use sign bits.
   if (VT.is512BitVector())
     return SDValue();
 
   // TODO: Add other opcodes eventually lowered into BLEND.
   for (SDNode::use_iterator UI = Cond->use_begin(), UE = Cond->use_end();
        UI != UE; ++UI)
-    if (UI->getOpcode() != ISD::VSELECT || UI.getOperandNo() != 0)
+    if ((UI->getOpcode() != ISD::VSELECT &&
+         UI->getOpcode() != X86ISD::SHRUNKBLEND) ||
+        UI.getOperandNo() != 0)
       return SDValue();
 
   APInt DemandedMask(APInt::getSignMask(BitWidth));
   KnownBits Known;
   TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
                                         !DCI.isBeforeLegalizeOps());
   if (!TLI.SimplifyDemandedBits(Cond, DemandedMask, Known, TLO, 0, true))
     return SDValue();
 
   // If we changed the computation somewhere in the DAG, this change will
   // affect all users of Cond. Update all the nodes so that we do not use
   // the generic VSELECT anymore. Otherwise, we may perform wrong
   // optimizations as we messed with the actual expectation for the vector
   // boolean values.
   for (SDNode *U : Cond->uses()) {
+    if (U->getOpcode() == X86ISD::SHRUNKBLEND)
+      continue;
+
     SDValue SB = DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(U), U->getValueType(0),
                              Cond, U->getOperand(1), U->getOperand(2));
     DAG.ReplaceAllUsesOfValueWith(SDValue(U, 0), SB);
     DCI.AddToWorklist(U);
   }
   DCI.CommitTargetLoweringOpt(TLO);
   return SDValue(N, 0);
 }

llvm/trunk/test/CodeGen/X86/avx2-logic.ll

entry:
   %x = xor <4 x i64> %a2, %b
   ret <4 x i64> %x
 }
 
 define <32 x i8> @vpblendvb(<32 x i1> %cond, <32 x i8> %x, <32 x i8> %y) {
 ; X32-LABEL: vpblendvb:
 ; X32:       # %bb.0:
 ; X32-NEXT:    vpsllw $7, %ymm0, %ymm0
-; X32-NEXT:    vpand {{\.LCPI.*}}, %ymm0, %ymm0
 ; X32-NEXT:    vpblendvb %ymm0, %ymm1, %ymm2, %ymm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: vpblendvb:
 ; X64:       # %bb.0:
 ; X64-NEXT:    vpsllw $7, %ymm0, %ymm0
-; X64-NEXT:    vpand {{.*}}(%rip), %ymm0, %ymm0
 ; X64-NEXT:    vpblendvb %ymm0, %ymm1, %ymm2, %ymm0
 ; X64-NEXT:    retq
   %min = select <32 x i1> %cond, <32 x i8> %x, <32 x i8> %y
   ret <32 x i8> %min
 }
 
 define <8 x i32> @allOnes() nounwind {
 ; X32-LABEL: allOnes:

llvm/trunk/test/CodeGen/X86/vector-reduce-smax.ll

 ; SSE41-NEXT:    movdqa %xmm0, %xmm1
 ; SSE41-NEXT:    psllq $32, %xmm1
 ; SSE41-NEXT:    psrad $31, %xmm1
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,2,3,3]
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,2,2,3]
 ; SSE41-NEXT:    psrad $31, %xmm3
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm2[0,1],xmm3[2,3],xmm2[4,5],xmm3[6,7]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
-; SSE41-NEXT:    movdqa %xmm3, %xmm2
-; SSE41-NEXT:    pxor %xmm0, %xmm2
-; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm2, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm2, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    pxor %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm1, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    por %xmm2, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm3
 ; SSE41-NEXT:    movd %xmm3, %eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: test_v2i32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpsllq $32, %xmm0, %xmm1
 ; AVX1-NEXT:    vpsrad $31, %xmm1, %xmm1

 ; SSE41-NEXT:    psrad $16, %xmm1
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,3],xmm1[4,5],xmm0[6,7]
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm2
 ; SSE41-NEXT:    pxor %xmm0, %xmm2
 ; SSE41-NEXT:    pxor %xmm1, %xmm0
 ; SSE41-NEXT:    movdqa %xmm2, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm2, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    por %xmm2, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm1
 ; SSE41-NEXT:    movd %xmm1, %eax
 ; SSE41-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: test_v2i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpsllq $48, %xmm0, %xmm1

 ; SSE41-NEXT:    psrad $24, %xmm1
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,3],xmm1[4,5],xmm0[6,7]
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
 ; SSE41-NEXT:    movdqa %xmm3, %xmm2
 ; SSE41-NEXT:    pxor %xmm0, %xmm2
 ; SSE41-NEXT:    pxor %xmm1, %xmm0
 ; SSE41-NEXT:    movdqa %xmm2, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm2, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    por %xmm2, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm1
 ; SSE41-NEXT:    pextrb $0, %xmm1, %eax
 ; SSE41-NEXT:    # kill: def $al killed $al killed $eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: test_v2i8:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpsllq $56, %xmm0, %xmm1

llvm/trunk/test/CodeGen/X86/vector-reduce-smin.ll

 ; SSE41-NEXT:    movdqa %xmm0, %xmm1
 ; SSE41-NEXT:    psllq $32, %xmm1
 ; SSE41-NEXT:    psrad $31, %xmm1
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,2,3,3]
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[0,2,2,3]
 ; SSE41-NEXT:    psrad $31, %xmm3
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm2[0,1],xmm3[2,3],xmm2[4,5],xmm3[6,7]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
-; SSE41-NEXT:    movdqa %xmm1, %xmm2
-; SSE41-NEXT:    pxor %xmm0, %xmm2
-; SSE41-NEXT:    pxor %xmm3, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm2, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm2, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pxor %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm3, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    por %xmm2, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm3
 ; SSE41-NEXT:    movd %xmm3, %eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: test_v2i32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpsllq $32, %xmm0, %xmm1
 ; AVX1-NEXT:    vpsrad $31, %xmm1, %xmm1

 ; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,3],xmm3[4,5],xmm2[6,7]
 ; SSE41-NEXT:    psllq $48, %xmm1
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
 ; SSE41-NEXT:    psrad $31, %xmm0
 ; SSE41-NEXT:    psrad $16, %xmm1
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,3],xmm1[4,5],xmm0[6,7]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
-; SSE41-NEXT:    movdqa %xmm3, %xmm2
-; SSE41-NEXT:    pxor %xmm0, %xmm2
-; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm2, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm2, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    pxor %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm1, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    por %xmm2, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm1
 ; SSE41-NEXT:    movd %xmm1, %eax
 ; SSE41-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: test_v2i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpsllq $48, %xmm0, %xmm1

 ; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,3],xmm3[4,5],xmm2[6,7]
 ; SSE41-NEXT:    psllq $56, %xmm1
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
 ; SSE41-NEXT:    psrad $31, %xmm0
 ; SSE41-NEXT:    psrad $24, %xmm1
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,3],xmm1[4,5],xmm0[6,7]
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
-; SSE41-NEXT:    movdqa %xmm3, %xmm2
-; SSE41-NEXT:    pxor %xmm0, %xmm2
-; SSE41-NEXT:    pxor %xmm1, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm2, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm2, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm3, %xmm0
+; SSE41-NEXT:    pxor %xmm2, %xmm0
+; SSE41-NEXT:    pxor %xmm1, %xmm2
+; SSE41-NEXT:    movdqa %xmm2, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    por %xmm2, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm3, %xmm1
 ; SSE41-NEXT:    pextrb $0, %xmm1, %eax
 ; SSE41-NEXT:    # kill: def $al killed $al killed $eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: test_v2i8:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpsllq $56, %xmm0, %xmm1

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

 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 ; SSE41-NEXT:    pmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
 ; SSE41-NEXT:    movdqa %xmm1, %xmm3
 ; SSE41-NEXT:    pxor %xmm0, %xmm3
 ; SSE41-NEXT:    pxor %xmm2, %xmm0
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    por %xmm3, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
 ; SSE41-NEXT:    movd %xmm2, %eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: test_v2i32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]

 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3],xmm0[4],xmm1[5,6,7]
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 ; SSE41-NEXT:    pmovzxwq {{.*#+}} xmm2 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
 ; SSE41-NEXT:    movdqa %xmm1, %xmm3
 ; SSE41-NEXT:    por %xmm0, %xmm3
 ; SSE41-NEXT:    por %xmm2, %xmm0
 ; SSE41-NEXT:    movdqa %xmm3, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm0, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    por %xmm3, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
 ; SSE41-NEXT:    movd %xmm2, %eax
 ; SSE41-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: test_v2i16:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vpxor %xmm1, %xmm1, %xmm1

 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 ; SSE41-NEXT:    pand {{.*}}(%rip), %xmm1
 ; SSE41-NEXT:    pmovzxbq {{.*#+}} xmm2 = xmm0[0],zero,zero,zero,zero,zero,zero,zero,xmm0[1],zero,zero,zero,zero,zero,zero,zero
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
 ; SSE41-NEXT:    movdqa %xmm1, %xmm4
 ; SSE41-NEXT:    por %xmm3, %xmm4
 ; SSE41-NEXT:    por %xmm2, %xmm3
 ; SSE41-NEXT:    movdqa %xmm4, %xmm0
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm4, %xmm3
-; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
-; SSE41-NEXT:    pand %xmm5, %xmm3
-; SSE41-NEXT:    por %xmm3, %xmm0
+; SSE41-NEXT:    pcmpeqd %xmm3, %xmm0
+; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
+; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[0,0,2,2]
+; SSE41-NEXT:    pand %xmm3, %xmm0
+; SSE41-NEXT:    por %xmm4, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
 ; SSE41-NEXT:    pextrb $0, %xmm2, %eax
 ; SSE41-NEXT:    # kill: def $al killed $al killed $eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: test_v2i8:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vpand {{.*}}(%rip), %xmm0, %xmm1

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

 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: test_v2i32:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    pxor %xmm1, %xmm1
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 ; SSE41-NEXT:    pmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
-; SSE41-NEXT:    movdqa %xmm1, %xmm3
-; SSE41-NEXT:    pxor %xmm0, %xmm3
-; SSE41-NEXT:    pxor %xmm2, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    pxor %xmm3, %xmm0
+; SSE41-NEXT:    pxor %xmm2, %xmm3
+; SSE41-NEXT:    movdqa %xmm3, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    por %xmm3, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
 ; SSE41-NEXT:    movd %xmm2, %eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: test_v2i32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]

 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: test_v2i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    pxor %xmm1, %xmm1
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3],xmm0[4],xmm1[5,6,7]
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 ; SSE41-NEXT:    pmovzxwq {{.*#+}} xmm2 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
-; SSE41-NEXT:    movdqa %xmm1, %xmm3
-; SSE41-NEXT:    por %xmm0, %xmm3
-; SSE41-NEXT:    por %xmm2, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    por %xmm3, %xmm0
+; SSE41-NEXT:    por %xmm2, %xmm3
+; SSE41-NEXT:    movdqa %xmm3, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    por %xmm3, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
 ; SSE41-NEXT:    movd %xmm2, %eax
 ; SSE41-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: test_v2i16:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vpxor %xmm1, %xmm1, %xmm1

 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: test_v2i8:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm1
 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 ; SSE41-NEXT:    pand {{.*}}(%rip), %xmm1
 ; SSE41-NEXT:    pmovzxbq {{.*#+}} xmm2 = xmm0[0],zero,zero,zero,zero,zero,zero,zero,xmm0[1],zero,zero,zero,zero,zero,zero,zero
-; SSE41-NEXT:    movdqa {{.*#+}} xmm0 = [2147483648,2147483648]
-; SSE41-NEXT:    movdqa %xmm1, %xmm3
-; SSE41-NEXT:    por %xmm0, %xmm3
-; SSE41-NEXT:    por %xmm2, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm4
-; SSE41-NEXT:    pcmpgtd %xmm3, %xmm4
-; SSE41-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
-; SSE41-NEXT:    pcmpeqd %xmm3, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT:    pand %xmm5, %xmm0
-; SSE41-NEXT:    por %xmm4, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
+; SSE41-NEXT:    movdqa %xmm1, %xmm0
+; SSE41-NEXT:    por %xmm3, %xmm0
+; SSE41-NEXT:    por %xmm2, %xmm3
+; SSE41-NEXT:    movdqa %xmm3, %xmm4
+; SSE41-NEXT:    pcmpeqd %xmm0, %xmm4
+; SSE41-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
+; SSE41-NEXT:    pand %xmm4, %xmm0
+; SSE41-NEXT:    por %xmm3, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
 ; SSE41-NEXT:    pextrb $0, %xmm2, %eax
 ; SSE41-NEXT:    # kill: def $al killed $al killed $eax
 ; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: test_v2i8:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vpand {{.*}}(%rip), %xmm0, %xmm1