[X86][SSE] Add support for PACKSS/PACKUS constant folding

Pulled out of D38472

llvm-svn: 314776

4 files changed, 133 insertions, 68 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 0981d39fe5c..3dd4d74ca40 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -5350,6 +5350,13 @@ static bool getTargetConstantBitsFromNode(SDValue Op, unsigned EltSizeInBits,
     return false;
   };
 
+  // Handle UNDEFs.
+  if (Op.isUndef()) {
+    APInt UndefSrcElts = APInt::getAllOnesValue(NumElts);
+    SmallVector<APInt, 64> SrcEltBits(NumElts, APInt(EltSizeInBits, 0));
+    return CastBitData(UndefSrcElts, SrcEltBits);
+  }
+
   // Extract constant bits from build vector.
   if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
     unsigned SrcEltSizeInBits = VT.getScalarSizeInBits();
@@ -31838,6 +31845,82 @@ static SDValue combineShift(SDNode* N, SelectionDAG &DAG,
   return SDValue();
 }
 
+static SDValue combineVectorPack(SDNode *N, SelectionDAG &DAG,
+                                 TargetLowering::DAGCombinerInfo &DCI,
+                                 const X86Subtarget &Subtarget) {
+  unsigned Opcode = N->getOpcode();
+  assert((X86ISD::PACKSS == Opcode || X86ISD::PACKUS == Opcode) &&
+         "Unexpected shift opcode");
+
+  EVT VT = N->getValueType(0);
+  EVT SVT = VT.getScalarType();
+  SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
+  unsigned DstBitsPerElt = VT.getScalarSizeInBits();
+  unsigned SrcBitsPerElt = 2 * DstBitsPerElt;
+  assert(N0.getScalarValueSizeInBits() == SrcBitsPerElt &&
+         N1.getScalarValueSizeInBits() == SrcBitsPerElt &&
+         "Unexpected PACKSS/PACKUS input type");
+
+  // Constant Folding.
+  APInt UndefElts0, UndefElts1;
+  SmallVector<APInt, 32> EltBits0, EltBits1;
+  if ((N0->isUndef() || N->isOnlyUserOf(N0.getNode())) &&
+      (N1->isUndef() || N->isOnlyUserOf(N1.getNode())) &&
+      getTargetConstantBitsFromNode(N0, SrcBitsPerElt, UndefElts0, EltBits0) &&
+      getTargetConstantBitsFromNode(N1, SrcBitsPerElt, UndefElts1, EltBits1)) {
+    unsigned NumLanes = VT.getSizeInBits() / 128;
+    unsigned NumDstElts = VT.getVectorNumElements();
+    unsigned NumSrcElts = NumDstElts / 2;
+    unsigned NumDstEltsPerLane = NumDstElts / NumLanes;
+    unsigned NumSrcEltsPerLane = NumSrcElts / NumLanes;
+    bool IsSigned = (X86ISD::PACKSS == Opcode);
+
+    APInt Undefs(NumDstElts, 0);
+    SmallVector<APInt, 32> Bits(NumDstElts, APInt::getNullValue(DstBitsPerElt));
+    for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
+      for (unsigned Elt = 0; Elt != NumDstEltsPerLane; ++Elt) {
+        unsigned SrcIdx = Lane * NumSrcEltsPerLane + Elt % NumSrcEltsPerLane;
+        auto &UndefElts = (Elt >= NumSrcEltsPerLane ? UndefElts1 : UndefElts0);
+        auto &EltBits = (Elt >= NumSrcEltsPerLane ? EltBits1 : EltBits0);
+
+        if (UndefElts[SrcIdx]) {
+          Undefs.setBit(Lane * NumDstEltsPerLane + Elt);
+          continue;
+        }
+
+        APInt &Val = EltBits[SrcIdx];
+        if (IsSigned) {
+          // PACKSS: Truncate signed value with signed saturation.
+          // Source values less than dst minint are saturated to minint.
+          // Source values greater than dst maxint are saturated to maxint.
+          if (Val.isSignedIntN(DstBitsPerElt))
+            Val = Val.trunc(DstBitsPerElt);
+          else if (Val.isNegative())
+            Val = APInt::getSignedMinValue(DstBitsPerElt);
+          else
+            Val = APInt::getSignedMaxValue(DstBitsPerElt);
+        } else {
+          // PACKUS: Truncate signed value with unsigned saturation.
+          // Source values less than zero are saturated to zero.
+          // Source values greater than dst maxuint are saturated to maxuint.
+          if (Val.isIntN(DstBitsPerElt))
+            Val = Val.trunc(DstBitsPerElt);
+          else if (Val.isNegative())
+            Val = APInt::getNullValue(DstBitsPerElt);
+          else
+            Val = APInt::getAllOnesValue(DstBitsPerElt);
+        }
+        Bits[Lane * NumDstEltsPerLane + Elt] = Val;
+      }
+    }
+
+    return getConstVector(Bits, Undefs, VT.getSimpleVT(), DAG, SDLoc(N));
+  }
+
+  return SDValue();
+}
+
 static SDValue combineVectorShiftImm(SDNode *N, SelectionDAG &DAG,
                                      TargetLowering::DAGCombinerInfo &DCI,
                                      const X86Subtarget &Subtarget) {
@@ -36069,6 +36152,8 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::SETCC:          return combineSetCC(N, DAG, Subtarget);
   case X86ISD::SETCC:       return combineX86SetCC(N, DAG, Subtarget);
   case X86ISD::BRCOND:      return combineBrCond(N, DAG, Subtarget);
+  case X86ISD::PACKSS:
+  case X86ISD::PACKUS:      return combineVectorPack(N, DAG, DCI, Subtarget);
   case X86ISD::VSHLI:
   case X86ISD::VSRAI:
   case X86ISD::VSRLI:
diff --git a/llvm/test/CodeGen/X86/avx2-intrinsics-x86.ll b/llvm/test/CodeGen/X86/avx2-intrinsics-x86.ll
index 7c9874e9a48..1329c243924 100644
--- a/llvm/test/CodeGen/X86/avx2-intrinsics-x86.ll
+++ b/llvm/test/CodeGen/X86/avx2-intrinsics-x86.ll
@@ -21,15 +21,15 @@ declare <16 x i16> @llvm.x86.avx2.packssdw(<8 x i32>, <8 x i32>) nounwind readno
 define <16 x i16> @test_x86_avx2_packssdw_fold() {
 ; AVX2-LABEL: test_x86_avx2_packssdw_fold:
 ; AVX2:       ## BB#0:
-; AVX2-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0xef,0xc0]
-; AVX2-NEXT:    vpackssdw LCPI1_0, %ymm0, %ymm0 ## encoding: [0xc5,0xfd,0x6b,0x05,A,A,A,A]
+; AVX2-NEXT:    vmovaps {{.*#+}} ymm0 = [0,0,0,0,255,32767,32767,65535,0,0,0,0,32769,32768,0,65280]
+; AVX2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x05,A,A,A,A]
 ; AVX2-NEXT:    ## fixup A - offset: 4, value: LCPI1_0, kind: FK_Data_4
 ; AVX2-NEXT:    retl ## encoding: [0xc3]
 ;
 ; AVX512VL-LABEL: test_x86_avx2_packssdw_fold:
 ; AVX512VL:       ## BB#0:
-; AVX512VL-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
-; AVX512VL-NEXT:    vpackssdw LCPI1_0, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6b,0x05,A,A,A,A]
+; AVX512VL-NEXT:    vmovaps LCPI1_0, %ymm0 ## EVEX TO VEX Compression ymm0 = [0,0,0,0,255,32767,32767,65535,0,0,0,0,32769,32768,0,65280]
+; AVX512VL-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x05,A,A,A,A]
 ; AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI1_0, kind: FK_Data_4
 ; AVX512VL-NEXT:    retl ## encoding: [0xc3]
   %res = call <16 x i16> @llvm.x86.avx2.packssdw(<8 x i32> zeroinitializer, <8 x i32> <i32 255, i32 32767, i32 65535, i32 -1, i32 -32767, i32 -65535, i32 0, i32 -256>)
@@ -56,20 +56,16 @@ declare <32 x i8> @llvm.x86.avx2.packsswb(<16 x i16>, <16 x i16>) nounwind readn
 define <32 x i8> @test_x86_avx2_packsswb_fold() {
 ; AVX2-LABEL: test_x86_avx2_packsswb_fold:
 ; AVX2:       ## BB#0:
-; AVX2-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0xef,0xc0]
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [0,255,256,65535,65535,65281,65280,32858,0,255,256,65535,65535,65281,65280,32858]
-; AVX2-NEXT:    ## encoding: [0xc5,0xfd,0x6f,0x0d,A,A,A,A]
+; AVX2-NEXT:    vmovaps {{.*#+}} ymm0 = [0,127,127,255,255,128,128,128,0,0,0,0,0,0,0,0,0,127,127,255,255,128,128,128,0,0,0,0,0,0,0,0]
+; AVX2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x05,A,A,A,A]
 ; AVX2-NEXT:    ## fixup A - offset: 4, value: LCPI3_0, kind: FK_Data_4
-; AVX2-NEXT:    vpacksswb %ymm0, %ymm1, %ymm0 ## encoding: [0xc5,0xf5,0x63,0xc0]
 ; AVX2-NEXT:    retl ## encoding: [0xc3]
 ;
 ; AVX512VL-LABEL: test_x86_avx2_packsswb_fold:
 ; AVX512VL:       ## BB#0:
-; AVX512VL-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
-; AVX512VL-NEXT:    vmovdqa LCPI3_0, %ymm1 ## EVEX TO VEX Compression ymm1 = [0,255,256,65535,65535,65281,65280,32858,0,255,256,65535,65535,65281,65280,32858]
-; AVX512VL-NEXT:    ## encoding: [0xc5,0xfd,0x6f,0x0d,A,A,A,A]
+; AVX512VL-NEXT:    vmovaps LCPI3_0, %ymm0 ## EVEX TO VEX Compression ymm0 = [0,127,127,255,255,128,128,128,0,0,0,0,0,0,0,0,0,127,127,255,255,128,128,128,0,0,0,0,0,0,0,0]
+; AVX512VL-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x05,A,A,A,A]
 ; AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI3_0, kind: FK_Data_4
-; AVX512VL-NEXT:    vpacksswb %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0x63,0xc0]
 ; AVX512VL-NEXT:    retl ## encoding: [0xc3]
   %res = call <32 x i8> @llvm.x86.avx2.packsswb(<16 x i16> <i16 0, i16 255, i16 256, i16 65535, i16 -1, i16 -255, i16 -256, i16 -32678, i16 0, i16 255, i16 256, i16 65535, i16 -1, i16 -255, i16 -256, i16 -32678>, <16 x i16> zeroinitializer)
   ret <32 x i8> %res
@@ -95,20 +91,16 @@ declare <32 x i8> @llvm.x86.avx2.packuswb(<16 x i16>, <16 x i16>) nounwind readn
 define <32 x i8> @test_x86_avx2_packuswb_fold() {
 ; AVX2-LABEL: test_x86_avx2_packuswb_fold:
 ; AVX2:       ## BB#0:
-; AVX2-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0xef,0xc0]
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [0,255,256,65535,65535,65281,65280,32858,0,255,256,65535,65535,65281,65280,32858]
-; AVX2-NEXT:    ## encoding: [0xc5,0xfd,0x6f,0x0d,A,A,A,A]
+; AVX2-NEXT:    vmovaps {{.*#+}} ymm0 = [0,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0,0,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0]
+; AVX2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x05,A,A,A,A]
 ; AVX2-NEXT:    ## fixup A - offset: 4, value: LCPI5_0, kind: FK_Data_4
-; AVX2-NEXT:    vpackuswb %ymm0, %ymm1, %ymm0 ## encoding: [0xc5,0xf5,0x67,0xc0]
 ; AVX2-NEXT:    retl ## encoding: [0xc3]
 ;
 ; AVX512VL-LABEL: test_x86_avx2_packuswb_fold:
 ; AVX512VL:       ## BB#0:
-; AVX512VL-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
-; AVX512VL-NEXT:    vmovdqa LCPI5_0, %ymm1 ## EVEX TO VEX Compression ymm1 = [0,255,256,65535,65535,65281,65280,32858,0,255,256,65535,65535,65281,65280,32858]
-; AVX512VL-NEXT:    ## encoding: [0xc5,0xfd,0x6f,0x0d,A,A,A,A]
+; AVX512VL-NEXT:    vmovaps LCPI5_0, %ymm0 ## EVEX TO VEX Compression ymm0 = [0,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0,0,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0]
+; AVX512VL-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x05,A,A,A,A]
 ; AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI5_0, kind: FK_Data_4
-; AVX512VL-NEXT:    vpackuswb %ymm0, %ymm1, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf5,0x67,0xc0]
 ; AVX512VL-NEXT:    retl ## encoding: [0xc3]
   %res = call <32 x i8> @llvm.x86.avx2.packuswb(<16 x i16> <i16 0, i16 255, i16 256, i16 65535, i16 -1, i16 -255, i16 -256, i16 -32678, i16 0, i16 255, i16 256, i16 65535, i16 -1, i16 -255, i16 -256, i16 -32678>, <16 x i16> zeroinitializer)
   ret <32 x i8> %res
@@ -850,16 +842,16 @@ declare <16 x i16> @llvm.x86.avx2.packusdw(<8 x i32>, <8 x i32>) nounwind readno
 define <16 x i16> @test_x86_avx2_packusdw_fold() {
 ; AVX2-LABEL: test_x86_avx2_packusdw_fold:
 ; AVX2:       ## BB#0:
-; AVX2-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0xef,0xc0]
-; AVX2-NEXT:    vpackusdw LCPI55_0, %ymm0, %ymm0 ## encoding: [0xc4,0xe2,0x7d,0x2b,0x05,A,A,A,A]
-; AVX2-NEXT:    ## fixup A - offset: 5, value: LCPI55_0, kind: FK_Data_4
+; AVX2-NEXT:    vmovaps {{.*#+}} ymm0 = [0,0,0,0,255,32767,65535,0,0,0,0,0,0,0,0,0]
+; AVX2-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x05,A,A,A,A]
+; AVX2-NEXT:    ## fixup A - offset: 4, value: LCPI55_0, kind: FK_Data_4
 ; AVX2-NEXT:    retl ## encoding: [0xc3]
 ;
 ; AVX512VL-LABEL: test_x86_avx2_packusdw_fold:
 ; AVX512VL:       ## BB#0:
-; AVX512VL-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
-; AVX512VL-NEXT:    vpackusdw LCPI55_0, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x2b,0x05,A,A,A,A]
-; AVX512VL-NEXT:    ## fixup A - offset: 5, value: LCPI55_0, kind: FK_Data_4
+; AVX512VL-NEXT:    vmovaps LCPI55_0, %ymm0 ## EVEX TO VEX Compression ymm0 = [0,0,0,0,255,32767,65535,0,0,0,0,0,0,0,0,0]
+; AVX512VL-NEXT:    ## encoding: [0xc5,0xfc,0x28,0x05,A,A,A,A]
+; AVX512VL-NEXT:    ## fixup A - offset: 4, value: LCPI55_0, kind: FK_Data_4
 ; AVX512VL-NEXT:    retl ## encoding: [0xc3]
   %res = call <16 x i16> @llvm.x86.avx2.packusdw(<8 x i32> zeroinitializer, <8 x i32> <i32 255, i32 32767, i32 65535, i32 -1, i32 -32767, i32 -65535, i32 0, i32 -256>)
   ret <16 x i16> %res
diff --git a/llvm/test/CodeGen/X86/sse2-intrinsics-x86.ll b/llvm/test/CodeGen/X86/sse2-intrinsics-x86.ll
index e6d3057fc5d..7324e855088 100644
--- a/llvm/test/CodeGen/X86/sse2-intrinsics-x86.ll
+++ b/llvm/test/CodeGen/X86/sse2-intrinsics-x86.ll
@@ -684,22 +684,22 @@ declare <8 x i16> @llvm.x86.sse2.packssdw.128(<4 x i32>, <4 x i32>) nounwind rea
 define <8 x i16> @test_x86_sse2_packssdw_128_fold() {
 ; SSE-LABEL: test_x86_sse2_packssdw_128_fold:
 ; SSE:       ## BB#0:
-; SSE-NEXT:    pxor %xmm0, %xmm0 ## encoding: [0x66,0x0f,0xef,0xc0]
-; SSE-NEXT:    packssdw LCPI32_0, %xmm0 ## encoding: [0x66,0x0f,0x6b,0x05,A,A,A,A]
-; SSE-NEXT:    ## fixup A - offset: 4, value: LCPI32_0, kind: FK_Data_4
+; SSE-NEXT:    movaps {{.*#+}} xmm0 = [0,0,0,0,32767,32767,65535,32768]
+; SSE-NEXT:    ## encoding: [0x0f,0x28,0x05,A,A,A,A]
+; SSE-NEXT:    ## fixup A - offset: 3, value: LCPI32_0, kind: FK_Data_4
 ; SSE-NEXT:    retl ## encoding: [0xc3]
 ;
 ; AVX2-LABEL: test_x86_sse2_packssdw_128_fold:
 ; AVX2:       ## BB#0:
-; AVX2-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0xef,0xc0]
-; AVX2-NEXT:    vpackssdw LCPI32_0, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0x6b,0x05,A,A,A,A]
+; AVX2-NEXT:    vmovaps {{.*#+}} xmm0 = [0,0,0,0,32767,32767,65535,32768]
+; AVX2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x05,A,A,A,A]
 ; AVX2-NEXT:    ## fixup A - offset: 4, value: LCPI32_0, kind: FK_Data_4
 ; AVX2-NEXT:    retl ## encoding: [0xc3]
 ;
 ; SKX-LABEL: test_x86_sse2_packssdw_128_fold:
 ; SKX:       ## BB#0:
-; SKX-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
-; SKX-NEXT:    vpackssdw LCPI32_0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6b,0x05,A,A,A,A]
+; SKX-NEXT:    vmovaps LCPI32_0, %xmm0 ## EVEX TO VEX Compression xmm0 = [0,0,0,0,32767,32767,65535,32768]
+; SKX-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x05,A,A,A,A]
 ; SKX-NEXT:    ## fixup A - offset: 4, value: LCPI32_0, kind: FK_Data_4
 ; SKX-NEXT:    retl ## encoding: [0xc3]
   %res = call <8 x i16> @llvm.x86.sse2.packssdw.128(<4 x i32> zeroinitializer, <4 x i32> <i32 65535, i32 65536, i32 -1, i32 -131072>)
@@ -731,29 +731,23 @@ declare <16 x i8> @llvm.x86.sse2.packsswb.128(<8 x i16>, <8 x i16>) nounwind rea
 define <16 x i8> @test_x86_sse2_packsswb_128_fold() {
 ; SSE-LABEL: test_x86_sse2_packsswb_128_fold:
 ; SSE:       ## BB#0:
-; SSE-NEXT:    pxor %xmm1, %xmm1 ## encoding: [0x66,0x0f,0xef,0xc9]
-; SSE-NEXT:    movdqa {{.*#+}} xmm0 = [0,255,256,65535,65535,65281,65280,32858]
-; SSE-NEXT:    ## encoding: [0x66,0x0f,0x6f,0x05,A,A,A,A]
-; SSE-NEXT:    ## fixup A - offset: 4, value: LCPI34_0, kind: FK_Data_4
-; SSE-NEXT:    packsswb %xmm1, %xmm0 ## encoding: [0x66,0x0f,0x63,0xc1]
+; SSE-NEXT:    movaps {{.*#+}} xmm0 = [0,127,127,255,255,128,128,128,0,0,0,0,0,0,0,0]
+; SSE-NEXT:    ## encoding: [0x0f,0x28,0x05,A,A,A,A]
+; SSE-NEXT:    ## fixup A - offset: 3, value: LCPI34_0, kind: FK_Data_4
 ; SSE-NEXT:    retl ## encoding: [0xc3]
 ;
 ; AVX2-LABEL: test_x86_sse2_packsswb_128_fold:
 ; AVX2:       ## BB#0:
-; AVX2-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0xef,0xc0]
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,255,256,65535,65535,65281,65280,32858]
-; AVX2-NEXT:    ## encoding: [0xc5,0xf9,0x6f,0x0d,A,A,A,A]
+; AVX2-NEXT:    vmovaps {{.*#+}} xmm0 = [0,127,127,255,255,128,128,128,0,0,0,0,0,0,0,0]
+; AVX2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x05,A,A,A,A]
 ; AVX2-NEXT:    ## fixup A - offset: 4, value: LCPI34_0, kind: FK_Data_4
-; AVX2-NEXT:    vpacksswb %xmm0, %xmm1, %xmm0 ## encoding: [0xc5,0xf1,0x63,0xc0]
 ; AVX2-NEXT:    retl ## encoding: [0xc3]
 ;
 ; SKX-LABEL: test_x86_sse2_packsswb_128_fold:
 ; SKX:       ## BB#0:
-; SKX-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
-; SKX-NEXT:    vmovdqa LCPI34_0, %xmm1 ## EVEX TO VEX Compression xmm1 = [0,255,256,65535,65535,65281,65280,32858]
-; SKX-NEXT:    ## encoding: [0xc5,0xf9,0x6f,0x0d,A,A,A,A]
+; SKX-NEXT:    vmovaps LCPI34_0, %xmm0 ## EVEX TO VEX Compression xmm0 = [0,127,127,255,255,128,128,128,0,0,0,0,0,0,0,0]
+; SKX-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x05,A,A,A,A]
 ; SKX-NEXT:    ## fixup A - offset: 4, value: LCPI34_0, kind: FK_Data_4
-; SKX-NEXT:    vpacksswb %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0x63,0xc0]
 ; SKX-NEXT:    retl ## encoding: [0xc3]
   %res = call <16 x i8> @llvm.x86.sse2.packsswb.128(<8 x i16> <i16 0, i16 255, i16 256, i16 65535, i16 -1, i16 -255, i16 -256, i16 -32678>, <8 x i16> zeroinitializer)
   ret <16 x i8> %res
@@ -784,29 +778,23 @@ declare <16 x i8> @llvm.x86.sse2.packuswb.128(<8 x i16>, <8 x i16>) nounwind rea
 define <16 x i8> @test_x86_sse2_packuswb_128_fold() {
 ; SSE-LABEL: test_x86_sse2_packuswb_128_fold:
 ; SSE:       ## BB#0:
-; SSE-NEXT:    pxor %xmm1, %xmm1 ## encoding: [0x66,0x0f,0xef,0xc9]
-; SSE-NEXT:    movdqa {{.*#+}} xmm0 = [0,255,256,65535,65535,65281,65280,32858]
-; SSE-NEXT:    ## encoding: [0x66,0x0f,0x6f,0x05,A,A,A,A]
-; SSE-NEXT:    ## fixup A - offset: 4, value: LCPI36_0, kind: FK_Data_4
-; SSE-NEXT:    packuswb %xmm1, %xmm0 ## encoding: [0x66,0x0f,0x67,0xc1]
+; SSE-NEXT:    movaps {{.*#+}} xmm0 = [0,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0]
+; SSE-NEXT:    ## encoding: [0x0f,0x28,0x05,A,A,A,A]
+; SSE-NEXT:    ## fixup A - offset: 3, value: LCPI36_0, kind: FK_Data_4
 ; SSE-NEXT:    retl ## encoding: [0xc3]
 ;
 ; AVX2-LABEL: test_x86_sse2_packuswb_128_fold:
 ; AVX2:       ## BB#0:
-; AVX2-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0xef,0xc0]
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,255,256,65535,65535,65281,65280,32858]
-; AVX2-NEXT:    ## encoding: [0xc5,0xf9,0x6f,0x0d,A,A,A,A]
+; AVX2-NEXT:    vmovaps {{.*#+}} xmm0 = [0,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0]
+; AVX2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x05,A,A,A,A]
 ; AVX2-NEXT:    ## fixup A - offset: 4, value: LCPI36_0, kind: FK_Data_4
-; AVX2-NEXT:    vpackuswb %xmm0, %xmm1, %xmm0 ## encoding: [0xc5,0xf1,0x67,0xc0]
 ; AVX2-NEXT:    retl ## encoding: [0xc3]
 ;
 ; SKX-LABEL: test_x86_sse2_packuswb_128_fold:
 ; SKX:       ## BB#0:
-; SKX-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
-; SKX-NEXT:    vmovdqa LCPI36_0, %xmm1 ## EVEX TO VEX Compression xmm1 = [0,255,256,65535,65535,65281,65280,32858]
-; SKX-NEXT:    ## encoding: [0xc5,0xf9,0x6f,0x0d,A,A,A,A]
+; SKX-NEXT:    vmovaps LCPI36_0, %xmm0 ## EVEX TO VEX Compression xmm0 = [0,255,255,0,0,0,0,0,0,0,0,0,0,0,0,0]
+; SKX-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x05,A,A,A,A]
 ; SKX-NEXT:    ## fixup A - offset: 4, value: LCPI36_0, kind: FK_Data_4
-; SKX-NEXT:    vpackuswb %xmm0, %xmm1, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf1,0x67,0xc0]
 ; SKX-NEXT:    retl ## encoding: [0xc3]
   %res = call <16 x i8> @llvm.x86.sse2.packuswb.128(<8 x i16> <i16 0, i16 255, i16 256, i16 65535, i16 -1, i16 -255, i16 -256, i16 -32678>, <8 x i16> zeroinitializer)
   ret <16 x i8> %res
diff --git a/llvm/test/CodeGen/X86/sse41-intrinsics-x86.ll b/llvm/test/CodeGen/X86/sse41-intrinsics-x86.ll
index 39acde0b5cd..98300a526a9 100644
--- a/llvm/test/CodeGen/X86/sse41-intrinsics-x86.ll
+++ b/llvm/test/CodeGen/X86/sse41-intrinsics-x86.ll
@@ -138,23 +138,23 @@ declare <8 x i16> @llvm.x86.sse41.packusdw(<4 x i32>, <4 x i32>) nounwind readno
 define <8 x i16> @test_x86_sse41_packusdw_fold() {
 ; SSE41-LABEL: test_x86_sse41_packusdw_fold:
 ; SSE41:       ## BB#0:
-; SSE41-NEXT:    pxor %xmm0, %xmm0 ## encoding: [0x66,0x0f,0xef,0xc0]
-; SSE41-NEXT:    packusdw LCPI7_0, %xmm0 ## encoding: [0x66,0x0f,0x38,0x2b,0x05,A,A,A,A]
-; SSE41-NEXT:    ## fixup A - offset: 5, value: LCPI7_0, kind: FK_Data_4
+; SSE41-NEXT:    movaps {{.*#+}} xmm0 = [0,0,0,0,65535,65535,0,0]
+; SSE41-NEXT:    ## encoding: [0x0f,0x28,0x05,A,A,A,A]
+; SSE41-NEXT:    ## fixup A - offset: 3, value: LCPI7_0, kind: FK_Data_4
 ; SSE41-NEXT:    retl ## encoding: [0xc3]
 ;
 ; AVX2-LABEL: test_x86_sse41_packusdw_fold:
 ; AVX2:       ## BB#0:
-; AVX2-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0xef,0xc0]
-; AVX2-NEXT:    vpackusdw LCPI7_0, %xmm0, %xmm0 ## encoding: [0xc4,0xe2,0x79,0x2b,0x05,A,A,A,A]
-; AVX2-NEXT:    ## fixup A - offset: 5, value: LCPI7_0, kind: FK_Data_4
+; AVX2-NEXT:    vmovaps {{.*#+}} xmm0 = [0,0,0,0,65535,65535,0,0]
+; AVX2-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x05,A,A,A,A]
+; AVX2-NEXT:    ## fixup A - offset: 4, value: LCPI7_0, kind: FK_Data_4
 ; AVX2-NEXT:    retl ## encoding: [0xc3]
 ;
 ; SKX-LABEL: test_x86_sse41_packusdw_fold:
 ; SKX:       ## BB#0:
-; SKX-NEXT:    vpxor %xmm0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xef,0xc0]
-; SKX-NEXT:    vpackusdw LCPI7_0, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x2b,0x05,A,A,A,A]
-; SKX-NEXT:    ## fixup A - offset: 5, value: LCPI7_0, kind: FK_Data_4
+; SKX-NEXT:    vmovaps LCPI7_0, %xmm0 ## EVEX TO VEX Compression xmm0 = [0,0,0,0,65535,65535,0,0]
+; SKX-NEXT:    ## encoding: [0xc5,0xf8,0x28,0x05,A,A,A,A]
+; SKX-NEXT:    ## fixup A - offset: 4, value: LCPI7_0, kind: FK_Data_4
 ; SKX-NEXT:    retl ## encoding: [0xc3]
   %res = call <8 x i16> @llvm.x86.sse41.packusdw(<4 x i32> zeroinitializer, <4 x i32> <i32 65535, i32 65536, i32 -1, i32 -131072>)
   ret <8 x i16> %res