[X86] Create PSUBUS from (add (umax X, C), -C)

[craig.topper]

Summary

InstCombine seems to canonicalize or PSUB patter into a max with the cosntant and an add with an inverse of the constant.

This patch recognizes this pattern and turns it into PSUBUS.

Fixes some of PR40053

[spatel]

1. Preliminary: add the tests with baseline trunk checks.
2. We should have test coverage for v8i16, v16i16, and v32i8.
3. This doesn't solve the addusb part of PR40053, right? That needs some enhancement around MatchADDUS I think.

[craig.topper]

Rebase after committing baseline tests for all types.

[andreadb]

Hi Craig,

Your patch addresses the issue with the subs example from PR40053. However, as soon as we change the code to something like this, then your rule would not trigger:

unsigned long long test_sub_2(__m128i x) {
    __m128i c = _mm_set1_epi8(70);
    return _mm_subs_epu8(x, c)[0];
}

This is similar to the example from PR40053, with the difference that only element zero is effectively used.
If I build this with clang, the optimizer performs a simplify demanded elts-like optimization that propagates undefs to the constant vector.
That undef propagation later on breaks the UMAX pattern. So, it won't appear in the "Initial Selection DAG".
As a consequence, your new pattern would not work, and we end up with this codegen (avx):

vpmaxub .LCPI1_0(%rip), %xmm0, %xmm1
vpcmpeqb        %xmm1, %xmm0, %xmm1
vpaddb  .LCPI1_1(%rip), %xmm0, %xmm0
vpand   %xmm0, %xmm1, %xmm0
vmovq   %xmm0, %rax

Your patch improves the matching of SUBUS, but more work would need to be done in this area.

There is also another problem caused by undef being aggressively propagated by the logic that simplifies demanded vector elements. See my comment below.

lib/Target/X86/X86ISelLowering.cpp

40594

My understanding is that matchBinaryPredicate() bails out if not all elements from the input vectors are ConstantSDNode. In the presence of constant build vectors with some undefs, that logic would not be able to match a SUBUS. However, the presence of undef values shouldn't really matter in this particular case. It should be safe to ignore them.

Recently simplify demanded vector elts has become more effective (and a bit more aggressive). Being able to propagate undef to unused lanes of an input vector is great. However, in one particular case (D55600) it caused a regression.
The root cause of that regression was the the inability of matchBinaryPredicate() to handle undefs.

I am worried that, the more we improve that simplification logic, the more likely it is to end up propagating undef values to constant vectors.

I think that we need a version of matchBinaryPredicate that knows how to skip undefs. Alternatively, `matchBinaryPredicate could accept an extra (optional) bool argument named IgnoreUndefs.

[RKSimon]

lib/Target/X86/X86ISelLowering.cpp

40594

I'm looking at this now - D55819 does the same for matchUnaryPredicate and I should have the equivalent matchBinaryPredicate patch available soon

[spatel]

This change handles psubus cases with no undefs, so LGTM.
Depending on how the patches land, we could add the undef capability and tests for it directly here or make that an enhancement.
But as I mentioned, I don't think this patch alone is enough to close PR40053 because that's at least 2 independent bugs in 1 report.

[RKSimon]

lib/Target/X86/X86ISelLowering.cpp

40594

D55822 is for matchBinaryPredicate undef handling - I have no objections for this patch to go in first and I can add support for it to D55822

[andreadb]

In D55780#1334383, @spatel wrote:

This change handles psubus cases with no undefs, so LGTM.
Depending on how the patches land, we could add the undef capability and tests for it directly here or make that an enhancement.
But as I mentioned, I don't think this patch alone is enough to close PR40053 because that's at least 2 independent bugs in 1 report.

Don't get me wrong: I am not suggesting that we shouldn't commit this patch.

I just wanted to point out that it is not true that all the psubus cases are fixed as you wrote. To fully fix psubus we need to make sure that we match UMAX even in the presence of undefs.
As I wrote in my previous comment, if you slightly change the example from the bugzilla, then we no longer get a UMAX in the DAG, and we lose this optimization (see below):

unsigned long long test_sub_2(__m128i x) {
    __m128i c = _mm_set1_epi8(70);
    return _mm_subs_epu8(x, c)[0];
}

[spatel]

In D55780#1334563, @andreadb wrote:

In D55780#1334383, @spatel wrote:

This change handles psubus cases with no undefs, so LGTM.
Depending on how the patches land, we could add the undef capability and tests for it directly here or make that an enhancement.
But as I mentioned, I don't think this patch alone is enough to close PR40053 because that's at least 2 independent bugs in 1 report.

Don't get me wrong: I am not suggesting that we shouldn't commit this patch.

I just wanted to point out that it is not true that all the psubus cases are fixed as you wrote. To fully fix psubus we need to make sure that we match UMAX even in the presence of undefs.
As I wrote in my previous comment, if you slightly change the example from the bugzilla, then we no longer get a UMAX in the DAG, and we lose this optimization (see below):

unsigned long long test_sub_2(__m128i x) {
    __m128i c = _mm_set1_epi8(70);
    return _mm_subs_epu8(x, c)[0];
}

Ah, sorry I forgot to address that example. I filed it here:
https://bugs.llvm.org/show_bug.cgi?id=40083

And I agree, we're missing many saturating math patterns. I started looking into that with:
https://bugs.llvm.org/show_bug.cgi?id=14613
...but now that we have IR intrinsics and DAG nodes for saturating math, I think it needs to be revisited so we use those ops.

[andreadb]

In D55780#1334596, @spatel wrote:

In D55780#1334563, @andreadb wrote:

In D55780#1334383, @spatel wrote:

This change handles psubus cases with no undefs, so LGTM.
Depending on how the patches land, we could add the undef capability and tests for it directly here or make that an enhancement.
But as I mentioned, I don't think this patch alone is enough to close PR40053 because that's at least 2 independent bugs in 1 report.

Don't get me wrong: I am not suggesting that we shouldn't commit this patch.

I just wanted to point out that it is not true that all the psubus cases are fixed as you wrote. To fully fix psubus we need to make sure that we match UMAX even in the presence of undefs.
As I wrote in my previous comment, if you slightly change the example from the bugzilla, then we no longer get a UMAX in the DAG, and we lose this optimization (see below):

unsigned long long test_sub_2(__m128i x) {
    __m128i c = _mm_set1_epi8(70);
    return _mm_subs_epu8(x, c)[0];
}

Ah, sorry I forgot to address that example. I filed it here:
https://bugs.llvm.org/show_bug.cgi?id=40083

No problem. Thanks for raising that bug!

And I agree, we're missing many saturating math patterns. I started looking into that with:
https://bugs.llvm.org/show_bug.cgi?id=14613
...but now that we have IR intrinsics and DAG nodes for saturating math, I think it needs to be revisited so we use those ops.

As Simon wrote, we could wait for D55822, so that we get support for undefs too. That being said, I am okay even if this patch is committed first.

-Andrea

lib/Target/X86/X86ISelLowering.cpp

return DAG.getNode(X86ISD::VPMADDWD, DL, ResVT,
                        DAG.getNode(ISD::TRUNCATE, DL, TruncVT, Ops[0]),
                        DAG.getNode(ISD::TRUNCATE, DL, TruncVT, Ops[1]));
   };
   return SplitOpsAndApply(DAG, Subtarget, DL, VT,
                           { Mul.getOperand(0), Mul.getOperand(1) },
                           PMADDBuilder);
 }
 
+// Try to turn (add (umax X, C), -C) into (psubus X, C)
+static SDValue combineAddToSUBUS(SDNode *N, SelectionDAG &DAG,
+                                 const X86Subtarget &Subtarget) {
+  if (!Subtarget.hasSSE2())
+    return SDValue();
+
+  EVT VT = N->getValueType(0);
+
+  // psubus is available in SSE2 for i8 and i16 vectors.
+  if (!VT.isVector() || VT.getVectorNumElements() < 2 ||
+      !isPowerOf2_32(VT.getVectorNumElements()) ||
+      !(VT.getVectorElementType() == MVT::i8 ||
+        VT.getVectorElementType() == MVT::i16))
+    return SDValue();
+
+  SDValue Op0 = N->getOperand(0);
+  SDValue Op1 = N->getOperand(1);
+  if (Op0.getOpcode() != ISD::UMAX)
+    return SDValue();
+
+  // The add should have a constant that is the negative of the max.
+  auto MatchSUBUS = [](ConstantSDNode *Max, ConstantSDNode *Op) {
+    return Max->getAPIntValue() == (-Op->getAPIntValue());
+  };
+  if (!ISD::matchBinaryPredicate(Op0.getOperand(1), Op1, MatchSUBUS))
+    return SDValue();
+
+  auto SUBUSBuilder = [](SelectionDAG &DAG, const SDLoc &DL,
+                         ArrayRef<SDValue> Ops) {
+    return DAG.getNode(X86ISD::SUBUS, DL, Ops[0].getValueType(), Ops);
+  };
+
+  // Take both operands from the umax node.
+  SDLoc DL(N);
+  return SplitOpsAndApply(DAG, Subtarget, DL, VT,
+                          { Op0.getOperand(0), Op0.getOperand(1) },
+                          SUBUSBuilder);
+}
+
 // Attempt to turn this pattern into PMADDWD.
 // (mul (add (zext (build_vector)), (zext (build_vector))),
 //      (add (zext (build_vector)), (zext (build_vector)))
 static SDValue matchPMADDWD_2(SelectionDAG &DAG, SDValue N0, SDValue N1,
                               const SDLoc &DL, EVT VT,
                               const X86Subtarget &Subtarget) {
   if (!Subtarget.hasSSE2())
     return SDValue();
if ((VT == MVT::v8i16 || VT == MVT::v4i32 || VT == MVT::v16i16 ||
     };
     return SplitOpsAndApply(DAG, Subtarget, SDLoc(N), VT, {Op0, Op1},
                             HADDBuilder);
   }
 
   if (SDValue V = combineIncDecVector(N, DAG))
     return V;
 
+  if (SDValue V = combineAddToSUBUS(N, DAG, Subtarget))
+    return V;
+
   return combineAddOrSubToADCOrSBB(N, DAG);
 }
 
 static SDValue combineSubToSubus(SDNode *N, SelectionDAG &DAG,
                                  const X86Subtarget &Subtarget) {
   SDValue Op0 = N->getOperand(0);
   SDValue Op1 = N->getOperand(1);
   EVT VT = N->getValueType(0);

test/CodeGen/X86/psubus.ll

; AVX512-NEXT:    retq
   %sh3 = select <2 x i1> %2, <2 x i16> %1, <2 x i16> zeroinitializer
   store <2 x i16> %sh3, <2 x i16>* %p1, align 8
   ret void
 }
 
 define <16 x i8> @test19(<16 x i8> %x) {
 ; SSE-LABEL: test19:
 ; SSE:       # %bb.0: # %entry
-; SSE-NEXT:    pmaxub {{.*}}(%rip), %xmm0
-; SSE-NEXT:    paddb {{.*}}(%rip), %xmm0
+; SSE-NEXT:    psubusb {{.*}}(%rip), %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: test19:
 ; AVX:       # %bb.0: # %entry
-; AVX-NEXT:    vpmaxub {{.*}}(%rip), %xmm0, %xmm0
-; AVX-NEXT:    vpaddb {{.*}}(%rip), %xmm0, %xmm0
+; AVX-NEXT:    vpsubusb {{.*}}(%rip), %xmm0, %xmm0
 ; AVX-NEXT:    retq
 entry:
   %0 = icmp ugt <16 x i8> %x, <i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70>
   %1 = select <16 x i1> %0, <16 x i8> %x, <16 x i8> <i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70>
   %2 = add <16 x i8> %1, <i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70>
   ret <16 x i8> %2
 }
 
 define <16 x i8> @test20(<16 x i8> %x) {
 ; SSE-LABEL: test20:
 ; SSE:       # %bb.0: # %entry
-; SSE-NEXT:    pmaxub {{.*}}(%rip), %xmm0
-; SSE-NEXT:    paddb {{.*}}(%rip), %xmm0
+; SSE-NEXT:    psubusb {{.*}}(%rip), %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: test20:
 ; AVX:       # %bb.0: # %entry
-; AVX-NEXT:    vpmaxub {{.*}}(%rip), %xmm0, %xmm0
-; AVX-NEXT:    vpaddb {{.*}}(%rip), %xmm0, %xmm0
+; AVX-NEXT:    vpsubusb {{.*}}(%rip), %xmm0, %xmm0
 ; AVX-NEXT:    retq
 entry:
   %0 = icmp ugt <16 x i8> %x, <i8 1, i8 -22, i8 -50, i8 -114, i8 -77, i8 -70, i8 123, i8 98, i8 63, i8 19, i8 -22, i8 100, i8 25, i8 34, i8 55, i8 70>
   %1 = select <16 x i1> %0, <16 x i8> %x, <16 x i8> <i8 1, i8 -22, i8 -50, i8 -114, i8 -77, i8 -70, i8 123, i8 98, i8 63, i8 19, i8 -22, i8 100, i8 25, i8 34, i8 55, i8 70>
   %2 = add <16 x i8> %1, <i8 -1, i8 22, i8 50, i8 114, i8 77, i8 70, i8 -123, i8 -98, i8 -63, i8 -19, i8 22, i8 -100, i8 -25, i8 -34, i8 -55, i8 -70>
   ret <16 x i8> %2
 }
 
 define <8 x i16> @test21(<8 x i16> %x) {
-; SSE2-LABEL: test21:
-; SSE2:       # %bb.0: # %entry
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
-; SSE2-NEXT:    pxor %xmm1, %xmm0
-; SSE2-NEXT:    pmaxsw {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    pxor %xmm1, %xmm0
-; SSE2-NEXT:    paddw {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    retq
-;
-; SSSE3-LABEL: test21:
-; SSSE3:       # %bb.0: # %entry
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
-; SSSE3-NEXT:    pxor %xmm1, %xmm0
-; SSSE3-NEXT:    pmaxsw {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    pxor %xmm1, %xmm0
-; SSSE3-NEXT:    paddw {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: test21:
-; SSE41:       # %bb.0: # %entry
-; SSE41-NEXT:    pmaxuw {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    paddw {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: test21:
+; SSE:       # %bb.0: # %entry
+; SSE-NEXT:    psubusw {{.*}}(%rip), %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: test21:
 ; AVX:       # %bb.0: # %entry
-; AVX-NEXT:    vpmaxuw {{.*}}(%rip), %xmm0, %xmm0
-; AVX-NEXT:    vpaddw {{.*}}(%rip), %xmm0, %xmm0
+; AVX-NEXT:    vpsubusw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX-NEXT:    retq
 entry:
   %0 = icmp ugt <8 x i16> %x, <i16 700, i16 700, i16 700, i16 700, i16 700, i16 700, i16 700, i16 700>
   %1 = select <8 x i1> %0, <8 x i16> %x, <8 x i16> <i16 700, i16 700, i16 700, i16 700, i16 700, i16 700, i16 700, i16 700>
   %2 = add <8 x i16> %1, <i16 -700, i16 -700, i16 -700, i16 -700, i16 -700, i16 -700, i16 -700, i16 -700>
   ret <8 x i16> %2
 }
 
 define <8 x i16> @test22(<8 x i16> %x) {
-; SSE2-LABEL: test22:
-; SSE2:       # %bb.0: # %entry
-; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
-; SSE2-NEXT:    pxor %xmm1, %xmm0
-; SSE2-NEXT:    pmaxsw {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    pxor %xmm1, %xmm0
-; SSE2-NEXT:    paddw {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    retq
-;
-; SSSE3-LABEL: test22:
-; SSSE3:       # %bb.0: # %entry
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
-; SSSE3-NEXT:    pxor %xmm1, %xmm0
-; SSSE3-NEXT:    pmaxsw {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    pxor %xmm1, %xmm0
-; SSSE3-NEXT:    paddw {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: test22:
-; SSE41:       # %bb.0: # %entry
-; SSE41-NEXT:    pmaxuw {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    paddw {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    retq
+; SSE-LABEL: test22:
+; SSE:       # %bb.0: # %entry
+; SSE-NEXT:    psubusw {{.*}}(%rip), %xmm0
+; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: test22:
 ; AVX:       # %bb.0: # %entry
-; AVX-NEXT:    vpmaxuw {{.*}}(%rip), %xmm0, %xmm0
-; AVX-NEXT:    vpaddw {{.*}}(%rip), %xmm0, %xmm0
+; AVX-NEXT:    vpsubusw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX-NEXT:    retq
 entry:
   %0 = icmp ugt <8 x i16> %x, <i16 1, i16 -22000, i16 -770, i16 98, i16 19, i16 1000, i16 3456, i16 70>
   %1 = select <8 x i1> %0, <8 x i16> %x, <8 x i16> <i16 1, i16 -22000, i16 -770, i16 98, i16 19, i16 1000, i16 3456, i16 70>
   %2 = add <8 x i16> %1, <i16 -1, i16 22000, i16 770, i16 -98, i16 -19, i16 -1000, i16 -3456, i16 -70>
   ret <8 x i16> %2
 }
 
 define <32 x i8> @test23(<32 x i8> %x) {
 ; SSE-LABEL: test23:
 ; SSE:       # %bb.0: # %entry
 ; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [70,70,70,70,70,70,70,70,70,70,70,70,70,70,70,70]
-; SSE-NEXT:    pmaxub %xmm2, %xmm1
-; SSE-NEXT:    pmaxub %xmm2, %xmm0
-; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [186,186,186,186,186,186,186,186,186,186,186,186,186,186,186,186]
-; SSE-NEXT:    paddb %xmm2, %xmm0
-; SSE-NEXT:    paddb %xmm2, %xmm1
+; SSE-NEXT:    psubusb %xmm2, %xmm0
+; SSE-NEXT:    psubusb %xmm2, %xmm1
 ; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: test23:
 ; AVX1:       # %bb.0: # %entry
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [70,70,70,70,70,70,70,70,70,70,70,70,70,70,70,70]
-; AVX1-NEXT:    vpmaxub %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpmaxub %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [186,186,186,186,186,186,186,186,186,186,186,186,186,186,186,186]
-; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpaddb %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [70,70,70,70,70,70,70,70,70,70,70,70,70,70,70,70]
+; AVX1-NEXT:    vpsubusb %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpsubusb %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: test23:
 ; AVX2:       # %bb.0: # %entry
-; AVX2-NEXT:    vpmaxub {{.*}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT:    vpaddb {{.*}}(%rip), %ymm0, %ymm0
+; AVX2-NEXT:    vpsubusb {{.*}}(%rip), %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: test23:
 ; AVX512:       # %bb.0: # %entry
-; AVX512-NEXT:    vpmaxub {{.*}}(%rip), %ymm0, %ymm0
-; AVX512-NEXT:    vpaddb {{.*}}(%rip), %ymm0, %ymm0
+; AVX512-NEXT:    vpsubusb {{.*}}(%rip), %ymm0, %ymm0
 ; AVX512-NEXT:    retq
 entry:
   %0 = icmp ugt <32 x i8> %x, <i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70>
   %1 = select <32 x i1> %0, <32 x i8> %x, <32 x i8> <i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70, i8 70>
   %2 = add <32 x i8> %1, <i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70, i8 -70>
   ret <32 x i8> %2
 }
 
 define <32 x i8> @test24(<32 x i8> %x) {
 ; SSE-LABEL: test24:
 ; SSE:       # %bb.0: # %entry
-; SSE-NEXT:    pmaxub {{.*}}(%rip), %xmm1
-; SSE-NEXT:    pmaxub {{.*}}(%rip), %xmm0
-; SSE-NEXT:    paddb {{.*}}(%rip), %xmm0
-; SSE-NEXT:    paddb {{.*}}(%rip), %xmm1
+; SSE-NEXT:    psubusb {{.*}}(%rip), %xmm0
+; SSE-NEXT:    psubusb {{.*}}(%rip), %xmm1
 ; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: test24:
 ; AVX1:       # %bb.0: # %entry
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT:    vpmaxub {{.*}}(%rip), %xmm1, %xmm1
-; AVX1-NEXT:    vpmaxub {{.*}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT:    vpaddb {{.*}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT:    vpaddb {{.*}}(%rip), %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT:    vpsubusb {{.*}}(%rip), %xmm0, %xmm1
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vpsubusb {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: test24:
 ; AVX2:       # %bb.0: # %entry
-; AVX2-NEXT:    vpmaxub {{.*}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT:    vpaddb {{.*}}(%rip), %ymm0, %ymm0
+; AVX2-NEXT:    vpsubusb {{.*}}(%rip), %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: test24:
 ; AVX512:       # %bb.0: # %entry
-; AVX512-NEXT:    vpmaxub {{.*}}(%rip), %ymm0, %ymm0
-; AVX512-NEXT:    vpaddb {{.*}}(%rip), %ymm0, %ymm0
+; AVX512-NEXT:    vpsubusb {{.*}}(%rip), %ymm0, %ymm0
 ; AVX512-NEXT:    retq
 entry:
   %0 = icmp ugt <32 x i8> %x, <i8 1, i8 -22, i8 -50, i8 -114, i8 -77, i8 -70, i8 123, i8 98, i8 63, i8 19, i8 -22, i8 100, i8 25, i8 34, i8 55, i8 70, i8 2, i8 -23, i8 -49, i8 -114, i8 -77, i8 -70, i8 123, i8 98, i8 63, i8 19, i8 -22, i8 110, i8 25, i8 34, i8 55, i8 70>
   %1 = select <32 x i1> %0, <32 x i8> %x, <32 x i8> <i8 1, i8 -22, i8 -50, i8 -114, i8 -77, i8 -70, i8 123, i8 98, i8 63, i8 19, i8 -22, i8 100, i8 25, i8 34, i8 55, i8 70, i8 2, i8 -23, i8 -49, i8 -114, i8 -77, i8 -70, i8 123, i8 98, i8 63, i8 19, i8 -22, i8 110, i8 25, i8 34, i8 55, i8 70>
   %2 = add <32 x i8> %1, <i8 -1, i8 22, i8 50, i8 114, i8 77, i8 70, i8 -123, i8 -98, i8 -63, i8 -19, i8 22, i8 -100, i8 -25, i8 -34, i8 -55, i8 -70, i8 -2, i8 23, i8 49, i8 114, i8 77, i8 70, i8 -123, i8 -98, i8 -63, i8 -19, i8 22, i8 -110, i8 -25, i8 -34, i8 -55, i8 -70>
   ret <32 x i8> %2
 }
 
 define <16 x i16> @test25(<16 x i16> %x) {
-; SSE2-LABEL: test25:
-; SSE2:       # %bb.0: # %entry
-; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
-; SSE2-NEXT:    pxor %xmm2, %xmm1
-; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [37768,37768,37768,37768,37768,37768,37768,37768]
-; SSE2-NEXT:    pmaxsw %xmm3, %xmm1
-; SSE2-NEXT:    pxor %xmm2, %xmm1
-; SSE2-NEXT:    pxor %xmm2, %xmm0
-; SSE2-NEXT:    pmaxsw %xmm3, %xmm0
-; SSE2-NEXT:    pxor %xmm2, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [60536,60536,60536,60536,60536,60536,60536,60536]
-; SSE2-NEXT:    paddw %xmm2, %xmm0
-; SSE2-NEXT:    paddw %xmm2, %xmm1
-; SSE2-NEXT:    retq
-;
-; SSSE3-LABEL: test25:
-; SSSE3:       # %bb.0: # %entry
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
-; SSSE3-NEXT:    pxor %xmm2, %xmm1
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [37768,37768,37768,37768,37768,37768,37768,37768]
-; SSSE3-NEXT:    pmaxsw %xmm3, %xmm1
-; SSSE3-NEXT:    pxor %xmm2, %xmm1
-; SSSE3-NEXT:    pxor %xmm2, %xmm0
-; SSSE3-NEXT:    pmaxsw %xmm3, %xmm0
-; SSSE3-NEXT:    pxor %xmm2, %xmm0
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [60536,60536,60536,60536,60536,60536,60536,60536]
-; SSSE3-NEXT:    paddw %xmm2, %xmm0
-; SSSE3-NEXT:    paddw %xmm2, %xmm1
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: test25:
-; SSE41:       # %bb.0: # %entry
-; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [5000,5000,5000,5000,5000,5000,5000,5000]
-; SSE41-NEXT:    pmaxuw %xmm2, %xmm1
-; SSE41-NEXT:    pmaxuw %xmm2, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [60536,60536,60536,60536,60536,60536,60536,60536]
-; SSE41-NEXT:    paddw %xmm2, %xmm0
-; SSE41-NEXT:    paddw %xmm2, %xmm1
-; SSE41-NEXT:    retq
+; SSE-LABEL: test25:
+; SSE:       # %bb.0: # %entry
+; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [5000,5000,5000,5000,5000,5000,5000,5000]
+; SSE-NEXT:    psubusw %xmm2, %xmm0
+; SSE-NEXT:    psubusw %xmm2, %xmm1
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: test25:
 ; AVX1:       # %bb.0: # %entry
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [5000,5000,5000,5000,5000,5000,5000,5000]
-; AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm1 = [60536,60536,60536,60536,60536,60536,60536,60536]
-; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpaddw %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [5000,5000,5000,5000,5000,5000,5000,5000]
+; AVX1-NEXT:    vpsubusw %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vpsubusw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: test25:
 ; AVX2:       # %bb.0: # %entry
-; AVX2-NEXT:    vpmaxuw {{.*}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT:    vpaddw {{.*}}(%rip), %ymm0, %ymm0
+; AVX2-NEXT:    vpsubusw {{.*}}(%rip), %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: test25:
 ; AVX512:       # %bb.0: # %entry
-; AVX512-NEXT:    vpmaxuw {{.*}}(%rip), %ymm0, %ymm0
-; AVX512-NEXT:    vpaddw {{.*}}(%rip), %ymm0, %ymm0
+; AVX512-NEXT:    vpsubusw {{.*}}(%rip), %ymm0, %ymm0
 ; AVX512-NEXT:    retq
 entry:
   %0 = icmp ugt <16 x i16> %x, <i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000>
   %1 = select <16 x i1> %0, <16 x i16> %x, <16 x i16> <i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000, i16 5000>
   %2 = add <16 x i16> %1, <i16 -5000, i16 -5000, i16 -5000, i16 -5000, i16 -5000, i16 -5000, i16 -5000, i16 -5000, i16 -5000, i16 -5000, i16 -5000, i16 -5000, i16 -5000, i16 -5000, i16 -5000, i16 -5000>
   ret <16 x i16> %2
 }
 
 define <16 x i16> @test26(<16 x i16> %x) {
-; SSE2-LABEL: test26:
-; SSE2:       # %bb.0: # %entry
-; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
-; SSE2-NEXT:    pxor %xmm2, %xmm1
-; SSE2-NEXT:    pmaxsw {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm2, %xmm1
-; SSE2-NEXT:    pxor %xmm2, %xmm0
-; SSE2-NEXT:    pmaxsw {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    pxor %xmm2, %xmm0
-; SSE2-NEXT:    paddw {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    paddw {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    retq
-;
-; SSSE3-LABEL: test26:
-; SSSE3:       # %bb.0: # %entry
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
-; SSSE3-NEXT:    pxor %xmm2, %xmm1
-; SSSE3-NEXT:    pmaxsw {{.*}}(%rip), %xmm1
-; SSSE3-NEXT:    pxor %xmm2, %xmm1
-; SSSE3-NEXT:    pxor %xmm2, %xmm0
-; SSSE3-NEXT:    pmaxsw {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    pxor %xmm2, %xmm0
-; SSSE3-NEXT:    paddw {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    paddw {{.*}}(%rip), %xmm1
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: test26:
-; SSE41:       # %bb.0: # %entry
-; SSE41-NEXT:    pmaxuw {{.*}}(%rip), %xmm1
-; SSE41-NEXT:    pmaxuw {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    paddw {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    paddw {{.*}}(%rip), %xmm1
-; SSE41-NEXT:    retq
+; SSE-LABEL: test26:
+; SSE:       # %bb.0: # %entry
+; SSE-NEXT:    psubusw {{.*}}(%rip), %xmm0
+; SSE-NEXT:    psubusw {{.*}}(%rip), %xmm1
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: test26:
 ; AVX1:       # %bb.0: # %entry
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT:    vpmaxuw {{.*}}(%rip), %xmm1, %xmm1
-; AVX1-NEXT:    vpmaxuw {{.*}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT:    vpaddw {{.*}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT:    vpaddw {{.*}}(%rip), %xmm1, %xmm1
-; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT:    vpsubusw {{.*}}(%rip), %xmm0, %xmm1
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vpsubusw {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: test26:
 ; AVX2:       # %bb.0: # %entry
-; AVX2-NEXT:    vpmaxuw {{.*}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT:    vpaddw {{.*}}(%rip), %ymm0, %ymm0
+; AVX2-NEXT:    vpsubusw {{.*}}(%rip), %ymm0, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: test26:
 ; AVX512:       # %bb.0: # %entry
-; AVX512-NEXT:    vpmaxuw {{.*}}(%rip), %ymm0, %ymm0
-; AVX512-NEXT:    vpaddw {{.*}}(%rip), %ymm0, %ymm0
+; AVX512-NEXT:    vpsubusw {{.*}}(%rip), %ymm0, %ymm0
 ; AVX512-NEXT:    retq
 entry:
   %0 = icmp ugt <16 x i16> %x, <i16 1, i16 -2200, i16 -50, i16 -114, i16 -77, i16 -70, i16 123, i16 9800, i16 635, i16 19567, i16 -22, i16 100, i16 2534, i16 34, i16 55, i16 70>
   %1 = select <16 x i1> %0, <16 x i16> %x, <16 x i16> <i16 1, i16 -2200, i16 -50, i16 -114, i16 -77, i16 -70, i16 123, i16 9800, i16 635, i16 19567, i16 -22, i16 100, i16 2534, i16 34, i16 55, i16 70>
   %2 = add <16 x i16> %1, <i16 -1, i16 2200, i16 50, i16 114, i16 77, i16 70, i16 -123, i16 -9800, i16 -635, i16 -19567, i16 22, i16 -100, i16 -2534, i16 -34, i16 -55, i16 -70>
   ret <16 x i16> %2
 }
 
 define <64 x i8> @test27(<64 x i8> %x) {
 ; SSE-LABEL: test27:
 ; SSE:       # %bb.0: # %entry
 ; SSE-NEXT:    movdqa {{.*#+}} xmm4 = [154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154]
-; SSE-NEXT:    pmaxub %xmm4, %xmm3
-; SSE-NEXT:    pmaxub %xmm4, %xmm2
-; SSE-NEXT:    pmaxub %xmm4, %xmm1
-; SSE-NEXT:    pmaxub %xmm4, %xmm0
-; SSE-NEXT:    movdqa {{.*#+}} xmm4 = [102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102]
-; SSE-NEXT:    paddb %xmm4, %xmm0
-; SSE-NEXT:    paddb %xmm4, %xmm1
-; SSE-NEXT:    paddb %xmm4, %xmm2
-; SSE-NEXT:    paddb %xmm4, %xmm3
+; SSE-NEXT:    psubusb %xmm4, %xmm0
+; SSE-NEXT:    psubusb %xmm4, %xmm1
+; SSE-NEXT:    psubusb %xmm4, %xmm2
+; SSE-NEXT:    psubusb %xmm4, %xmm3
 ; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: test27:
 ; AVX1:       # %bb.0: # %entry
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154]
-; AVX1-NEXT:    vpmaxub %xmm2, %xmm1, %xmm3
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vpmaxub %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vpmaxub %xmm2, %xmm0, %xmm4
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpmaxub %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102]
-; AVX1-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm4
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm4, %ymm0
-; AVX1-NEXT:    vpaddb %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vpaddb %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm2, %ymm1
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154]
+; AVX1-NEXT:    vpsubusb %xmm3, %xmm2, %xmm2
+; AVX1-NEXT:    vpsubusb %xmm3, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpsubusb %xmm3, %xmm2, %xmm2
+; AVX1-NEXT:    vpsubusb %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: test27:
 ; AVX2:       # %bb.0: # %entry
 ; AVX2-NEXT:    vmovdqa {{.*#+}} ymm2 = [154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154,154]
-; AVX2-NEXT:    vpmaxub %ymm2, %ymm1, %ymm1
-; AVX2-NEXT:    vpmaxub %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm2 = [102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102,102]
-; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpaddb %ymm2, %ymm1, %ymm1
+; AVX2-NEXT:    vpsubusb %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpsubusb %ymm2, %ymm1, %ymm1
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: test27:
 ; AVX512:       # %bb.0: # %entry
-; AVX512-NEXT:    vpmaxub {{.*}}(%rip), %zmm0, %zmm0
-; AVX512-NEXT:    vpaddb {{.*}}(%rip), %zmm0, %zmm0
+; AVX512-NEXT:    vpsubusb {{.*}}(%rip), %zmm0, %zmm0
 ; AVX512-NEXT:    retq
 entry:
   %0 = icmp ugt <64 x i8> %x, <i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154>
   %1 = select <64 x i1> %0, <64 x i8> %x, <64 x i8> <i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154, i8 154>
   %2 = add <64 x i8> %1, <i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154, i8 -154>
   ret <64 x i8> %2
 }
 
 define <64 x i8> @test28(<64 x i8> %x) {
 ; SSE-LABEL: test28:
 ; SSE:       # %bb.0: # %entry
 ; SSE-NEXT:    movdqa {{.*#+}} xmm4 = [1,234,206,142,179,186,123,98,63,19,234,100,25,34,55,70]
-; SSE-NEXT:    pmaxub %xmm4, %xmm2
-; SSE-NEXT:    pmaxub %xmm4, %xmm0
-; SSE-NEXT:    pmaxub {{.*}}(%rip), %xmm3
-; SSE-NEXT:    pmaxub {{.*}}(%rip), %xmm1
-; SSE-NEXT:    movdqa {{.*#+}} xmm4 = [255,22,50,114,77,70,133,158,193,237,22,156,231,222,201,186]
-; SSE-NEXT:    paddb %xmm4, %xmm0
-; SSE-NEXT:    paddb %xmm4, %xmm2
-; SSE-NEXT:    paddb {{.*}}(%rip), %xmm1
-; SSE-NEXT:    paddb {{.*}}(%rip), %xmm3
+; SSE-NEXT:    psubusb %xmm4, %xmm0
+; SSE-NEXT:    psubusb %xmm4, %xmm2
+; SSE-NEXT:    psubusb {{.*}}(%rip), %xmm1
+; SSE-NEXT:    psubusb {{.*}}(%rip), %xmm3
 ; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: test28:
 ; AVX1:       # %bb.0: # %entry
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [1,234,206,142,179,186,123,98,63,19,234,100,25,34,55,70]
-; AVX1-NEXT:    vpmaxub %xmm2, %xmm1, %xmm3
-; AVX1-NEXT:    vpmaxub %xmm2, %xmm0, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vpmaxub {{.*}}(%rip), %xmm1, %xmm1
+; AVX1-NEXT:    vpsubusb %xmm2, %xmm0, %xmm3
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpmaxub {{.*}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [255,22,50,114,77,70,133,158,193,237,22,156,231,222,201,186]
-; AVX1-NEXT:    vpaddb %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpaddb {{.*}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
-; AVX1-NEXT:    vpaddb %xmm4, %xmm3, %xmm2
-; AVX1-NEXT:    vpaddb {{.*}}(%rip), %xmm1, %xmm1
+; AVX1-NEXT:    vpsubusb {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm3, %ymm0
+; AVX1-NEXT:    vpsubusb %xmm2, %xmm1, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; AVX1-NEXT:    vpsubusb {{.*}}(%rip), %xmm1, %xmm1
 ; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm2, %ymm1
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: test28:
 ; AVX2:       # %bb.0: # %entry
-; AVX2-NEXT:    vpmaxub {{.*}}(%rip), %ymm1, %ymm1
-; AVX2-NEXT:    vpmaxub {{.*}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT:    vpaddb {{.*}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT:    vpaddb {{.*}}(%rip), %ymm1, %ymm1
+; AVX2-NEXT:    vpsubusb {{.*}}(%rip), %ymm0, %ymm0
+; AVX2-NEXT:    vpsubusb {{.*}}(%rip), %ymm1, %ymm1
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: test28:
 ; AVX512:       # %bb.0: # %entry
-; AVX512-NEXT:    vpmaxub {{.*}}(%rip), %zmm0, %zmm0
-; AVX512-NEXT:    vpaddb {{.*}}(%rip), %zmm0, %zmm0
+; AVX512-NEXT:    vpsubusb {{.*}}(%rip), %zmm0, %zmm0
 ; AVX512-NEXT:    retq
 entry:
   %0 = icmp ugt <64 x i8> %x, <i8 1, i8 -22, i8 -50, i8 -114, i8 -77, i8 -70, i8 123, i8 98, i8 63, i8 19, i8 -22, i8 100, i8 25, i8 34, i8 55, i8 70, i8 2, i8 -23, i8 -49, i8 -114, i8 -77, i8 -70, i8 123, i8 98, i8 63, i8 19, i8 -22, i8 110, i8 25, i8 34, i8 55, i8 70, i8 1, i8 -22, i8 -50, i8 -114, i8 -77, i8 -70, i8 123, i8 98, i8 63, i8 19, i8 -22, i8 100, i8 25, i8 34, i8 55, i8 70, i8 2, i8 -23, i8 -49, i8 -116, i8 -77, i8 -70, i8 123, i8 98, i8 67, i8 19, i8 -22, i8 110, i8 25, i8 34, i8 55, i8 70>
   %1 = select <64 x i1> %0, <64 x i8> %x, <64 x i8> <i8 1, i8 -22, i8 -50, i8 -114, i8 -77, i8 -70, i8 123, i8 98, i8 63, i8 19, i8 -22, i8 100, i8 25, i8 34, i8 55, i8 70, i8 2, i8 -23, i8 -49, i8 -114, i8 -77, i8 -70, i8 123, i8 98, i8 63, i8 19, i8 -22, i8 110, i8 25, i8 34, i8 55, i8 70, i8 1, i8 -22, i8 -50, i8 -114, i8 -77, i8 -70, i8 123, i8 98, i8 63, i8 19, i8 -22, i8 100, i8 25, i8 34, i8 55, i8 70, i8 2, i8 -23, i8 -49, i8 -116, i8 -77, i8 -70, i8 123, i8 98, i8 67, i8 19, i8 -22, i8 110, i8 25, i8 34, i8 55, i8 70>
   %2 = add <64 x i8> %1, <i8 -1, i8 22, i8 50, i8 114, i8 77, i8 70, i8 -123, i8 -98, i8 -63, i8 -19, i8 22, i8 -100, i8 -25, i8 -34, i8 -55, i8 -70, i8 -2, i8 23, i8 49, i8 114, i8 77, i8 70, i8 -123, i8 -98, i8 -63, i8 -19, i8 22, i8 -110, i8 -25, i8 -34, i8 -55, i8 -70, i8 -1, i8 22, i8 50, i8 114, i8 77, i8 70, i8 -123, i8 -98, i8 -63, i8 -19, i8 22, i8 -100, i8 -25, i8 -34, i8 -55, i8 -70, i8 -2, i8 23, i8 49, i8 116, i8 77, i8 70, i8 -123, i8 -98, i8 -67, i8 -19, i8 22, i8 -110, i8 -25, i8 -34, i8 -55, i8 -70>
   ret <64 x i8> %2
 }
 
 define <32 x i16> @test29(<32 x i16> %x) {
-; SSE2-LABEL: test29:
-; SSE2:       # %bb.0: # %entry
-; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [32768,32768,32768,32768,32768,32768,32768,32768]
-; SSE2-NEXT:    pxor %xmm4, %xmm3
-; SSE2-NEXT:    pmaxsw {{.*}}(%rip), %xmm3
-; SSE2-NEXT:    pxor %xmm4, %xmm3
-; SSE2-NEXT:    pxor %xmm4, %xmm2
-; SSE2-NEXT:    pmaxsw {{.*}}(%rip), %xmm2
-; SSE2-NEXT:    pxor %xmm4, %xmm2
-; SSE2-NEXT:    pxor %xmm4, %xmm1
-; SSE2-NEXT:    pmaxsw {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    pxor %xmm4, %xmm1
-; SSE2-NEXT:    pxor %xmm4, %xmm0
-; SSE2-NEXT:    pmaxsw {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    pxor %xmm4, %xmm0
-; SSE2-NEXT:    paddw {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    paddw {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    paddw {{.*}}(%rip), %xmm2
-; SSE2-NEXT:    paddw {{.*}}(%rip), %xmm3
-; SSE2-NEXT:    retq
-;
-; SSSE3-LABEL: test29:
-; SSSE3:       # %bb.0: # %entry
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm4 = [32768,32768,32768,32768,32768,32768,32768,32768]
-; SSSE3-NEXT:    pxor %xmm4, %xmm3
-; SSSE3-NEXT:    pmaxsw {{.*}}(%rip), %xmm3
-; SSSE3-NEXT:    pxor %xmm4, %xmm3
-; SSSE3-NEXT:    pxor %xmm4, %xmm2
-; SSSE3-NEXT:    pmaxsw {{.*}}(%rip), %xmm2
-; SSSE3-NEXT:    pxor %xmm4, %xmm2
-; SSSE3-NEXT:    pxor %xmm4, %xmm1
-; SSSE3-NEXT:    pmaxsw {{.*}}(%rip), %xmm1
-; SSSE3-NEXT:    pxor %xmm4, %xmm1
-; SSSE3-NEXT:    pxor %xmm4, %xmm0
-; SSSE3-NEXT:    pmaxsw {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    pxor %xmm4, %xmm0
-; SSSE3-NEXT:    paddw {{.*}}(%rip), %xmm0
-; SSSE3-NEXT:    paddw {{.*}}(%rip), %xmm1
-; SSSE3-NEXT:    paddw {{.*}}(%rip), %xmm2
-; SSSE3-NEXT:    paddw {{.*}}(%rip), %xmm3
-; SSSE3-NEXT:    retq
-;
-; SSE41-LABEL: test29:
-; SSE41:       # %bb.0: # %entry
-; SSE41-NEXT:    pmaxuw {{.*}}(%rip), %xmm3
-; SSE41-NEXT:    pmaxuw {{.*}}(%rip), %xmm2
-; SSE41-NEXT:    pmaxuw {{.*}}(%rip), %xmm1
-; SSE41-NEXT:    pmaxuw {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    paddw {{.*}}(%rip), %xmm0
-; SSE41-NEXT:    paddw {{.*}}(%rip), %xmm1
-; SSE41-NEXT:    paddw {{.*}}(%rip), %xmm2
-; SSE41-NEXT:    paddw {{.*}}(%rip), %xmm3
-; SSE41-NEXT:    retq
+; SSE-LABEL: test29:
+; SSE:       # %bb.0: # %entry
+; SSE-NEXT:    psubusw {{.*}}(%rip), %xmm0
+; SSE-NEXT:    psubusw {{.*}}(%rip), %xmm1
+; SSE-NEXT:    psubusw {{.*}}(%rip), %xmm2
+; SSE-NEXT:    psubusw {{.*}}(%rip), %xmm3
+; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: test29:
 ; AVX1:       # %bb.0: # %entry
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT:    vpmaxuw {{.*}}(%rip), %xmm2, %xmm2
-; AVX1-NEXT:    vpmaxuw {{.*}}(%rip), %xmm1, %xmm1
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
-; AVX1-NEXT:    vpmaxuw {{.*}}(%rip), %xmm3, %xmm3
-; AVX1-NEXT:    vpmaxuw {{.*}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT:    vpaddw {{.*}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT:    vpaddw {{.*}}(%rip), %xmm3, %xmm3
-; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm0
-; AVX1-NEXT:    vpaddw {{.*}}(%rip), %xmm1, %xmm1
-; AVX1-NEXT:    vpaddw {{.*}}(%rip), %xmm2, %xmm2
-; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
+; AVX1-NEXT:    vpsubusw {{.*}}(%rip), %xmm0, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vpsubusw {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
+; AVX1-NEXT:    vpsubusw {{.*}}(%rip), %xmm1, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; AVX1-NEXT:    vpsubusw {{.*}}(%rip), %xmm1, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm2, %ymm1
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: test29:
 ; AVX2:       # %bb.0: # %entry
-; AVX2-NEXT:    vpmaxuw {{.*}}(%rip), %ymm1, %ymm1
-; AVX2-NEXT:    vpmaxuw {{.*}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT:    vpaddw {{.*}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT:    vpaddw {{.*}}(%rip), %ymm1, %ymm1
+; AVX2-NEXT:    vpsubusw {{.*}}(%rip), %ymm0, %ymm0
+; AVX2-NEXT:    vpsubusw {{.*}}(%rip), %ymm1, %ymm1
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: test29:
 ; AVX512:       # %bb.0: # %entry
-; AVX512-NEXT:    vpmaxuw {{.*}}(%rip), %zmm0, %zmm0
-; AVX512-NEXT:    vpaddw {{.*}}(%rip), %zmm0, %zmm0
+; AVX512-NEXT:    vpsubusw {{.*}}(%rip), %zmm0, %zmm0
 ; AVX512-NEXT:    retq
 entry:
   %0 = icmp ugt <32 x i16> %x, <i16 1, i16 -2200, i16 -50, i16 -114, i16 -77, i16 -70, i16 123, i16 9800, i16 635, i16 19567, i16 -22, i16 100, i16 2534, i16 34, i16 55, i16 70, i16 1, i16 -2200, i16 -50, i16 -114, i16 -77, i16 -70, i16 123, i16 9805, i16 635, i16 19567, i16 -22, i16 100, i16 2534, i16 346, i16 55, i16 70>
   %1 = select <32 x i1> %0, <32 x i16> %x, <32 x i16> <i16 1, i16 -2200, i16 -50, i16 -114, i16 -77, i16 -70, i16 123, i16 9800, i16 635, i16 19567, i16 -22, i16 100, i16 2534, i16 34, i16 55, i16 70, i16 1, i16 -2200, i16 -50, i16 -114, i16 -77, i16 -70, i16 123, i16 9805, i16 635, i16 19567, i16 -22, i16 100, i16 2534, i16 346, i16 55, i16 70>
   %2 = add <32 x i16> %1, <i16 -1, i16 2200, i16 50, i16 114, i16 77, i16 70, i16 -123, i16 -9800, i16 -635, i16 -19567, i16 22, i16 -100, i16 -2534, i16 -34, i16 -55, i16 -70, i16 -1, i16 2200, i16 50, i16 114, i16 77, i16 70, i16 -123, i16 -9805, i16 -635, i16 -19567, i16 22, i16 -100, i16 -2534, i16 -346, i16 -55, i16 -70>
   ret <32 x i16> %2
 }