Revert "[X86][SSE] Generalize X86ISD::BLENDI support to more value ty…

…pes"

This reverts commit r353610.
It causes a miscompile visible in macro expansion in a bootstrapped clang.

http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20190211/626590.html

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@353699 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/X86/X86ISelLowering.cpp

@@ -10409,24 +10409,45 @@ static SDValue lowerShuffleAsBlend(const SDLoc &DL, MVT VT, SDValue V1,
     V2 = getZeroVector(VT, Subtarget, DAG, DL);
 
   switch (VT.SimpleTy) {
+  case MVT::v2f64:
+  case MVT::v4f32:
+  case MVT::v4f64:
+  case MVT::v8f32:
+    return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
+                       DAG.getConstant(BlendMask, DL, MVT::i8));
   case MVT::v4i64:
   case MVT::v8i32:
     assert(Subtarget.hasAVX2() && "256-bit integer blends require AVX2!");
     LLVM_FALLTHROUGH;
-  case MVT::v4f64:
-  case MVT::v8f32:
-    assert(Subtarget.hasAVX() && "256-bit float blends require AVX!");
-    LLVM_FALLTHROUGH;
-  case MVT::v2f64:
   case MVT::v2i64:
-  case MVT::v4f32:
   case MVT::v4i32:
-  case MVT::v8i16:
-    assert(Subtarget.hasSSE41() && "128-bit blends require SSE41!");
-    return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
-                       DAG.getConstant(BlendMask, DL, MVT::i8));
+    // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
+    // that instruction.
+    if (Subtarget.hasAVX2()) {
+      // Scale the blend by the number of 32-bit dwords per element.
+      int Scale =  VT.getScalarSizeInBits() / 32;
+      BlendMask = scaleVectorShuffleBlendMask(BlendMask, Mask.size(), Scale);
+      MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
+      V1 = DAG.getBitcast(BlendVT, V1);
+      V2 = DAG.getBitcast(BlendVT, V2);
+      return DAG.getBitcast(
+          VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
+                          DAG.getConstant(BlendMask, DL, MVT::i8)));
+    }
+    LLVM_FALLTHROUGH;
+  case MVT::v8i16: {
+    // For integer shuffles we need to expand the mask and cast the inputs to
+    // v8i16s prior to blending.
+    int Scale = 8 / VT.getVectorNumElements();
+    BlendMask = scaleVectorShuffleBlendMask(BlendMask, Mask.size(), Scale);
+    V1 = DAG.getBitcast(MVT::v8i16, V1);
+    V2 = DAG.getBitcast(MVT::v8i16, V2);
+    return DAG.getBitcast(VT,
+                          DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
+                                      DAG.getConstant(BlendMask, DL, MVT::i8)));
+  }
   case MVT::v16i16: {
-    assert(Subtarget.hasAVX2() && "v16i16 blends require AVX2!");
+    assert(Subtarget.hasAVX2() && "256-bit integer blends require AVX2!");
     SmallVector<int, 8> RepeatedMask;
     if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
       // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
@@ -10454,11 +10475,10 @@ static SDValue lowerShuffleAsBlend(const SDLoc &DL, MVT VT, SDValue V1,
     }
     LLVM_FALLTHROUGH;
   }
-  case MVT::v32i8:
-    assert(Subtarget.hasAVX2() && "256-bit byte-blends require AVX2!");
-    LLVM_FALLTHROUGH;
-  case MVT::v16i8: {
-    assert(Subtarget.hasSSE41() && "128-bit byte-blends require SSE41!");
+  case MVT::v16i8:
+  case MVT::v32i8: {
+    assert((VT.is128BitVector() || Subtarget.hasAVX2()) &&
+           "256-bit byte-blends require AVX2 support!");
 
     // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
     if (SDValue Masked = lowerShuffleAsBitMask(DL, VT, V1, V2, Mask, Zeroable,
@@ -30954,11 +30974,34 @@ static bool matchBinaryPermuteShuffle(
           return true;
         }
       } else {
+        // Determine a type compatible with X86ISD::BLENDI.
+        ShuffleVT = MaskVT;
+        if (Subtarget.hasAVX2()) {
+          if (ShuffleVT == MVT::v4i64)
+            ShuffleVT = MVT::v8i32;
+          else if (ShuffleVT == MVT::v2i64)
+            ShuffleVT = MVT::v4i32;
+        } else {
+          if (ShuffleVT == MVT::v2i64 || ShuffleVT == MVT::v4i32)
+            ShuffleVT = MVT::v8i16;
+          else if (ShuffleVT == MVT::v4i64)
+            ShuffleVT = MVT::v4f64;
+          else if (ShuffleVT == MVT::v8i32)
+            ShuffleVT = MVT::v8f32;
+        }
+
+        if (!ShuffleVT.isFloatingPoint()) {
+          int Scale = EltSizeInBits / ShuffleVT.getScalarSizeInBits();
+          BlendMask =
+              scaleVectorShuffleBlendMask(BlendMask, NumMaskElts, Scale);
+          ShuffleVT = MVT::getIntegerVT(EltSizeInBits / Scale);
+          ShuffleVT = MVT::getVectorVT(ShuffleVT, NumMaskElts * Scale);
+        }
+
         V1 = ForceV1Zero ? getZeroVector(MaskVT, Subtarget, DAG, DL) : V1;
         V2 = ForceV2Zero ? getZeroVector(MaskVT, Subtarget, DAG, DL) : V2;
         PermuteImm = (unsigned)BlendMask;
         Shuffle = X86ISD::BLENDI;
-        ShuffleVT = MaskVT;
         return true;
       }
     }
@@ -32123,29 +32166,6 @@ static SDValue combineTargetShuffle(SDValue N, SelectionDAG &DAG,
 
     return SDValue();
   }
-  case X86ISD::BLENDI: {
-    SDValue N0 = N.getOperand(0);
-    SDValue N1 = N.getOperand(1);
-
-    // blend(bitcast(x),bitcast(y)) -> bitcast(blend(x,y)) to narrower types.
-    // TODO: Handle MVT::v16i16 repeated blend mask.
-    if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
-        N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
-      MVT SrcVT = N0.getOperand(0).getSimpleValueType();
-      if ((VT.getScalarSizeInBits() % SrcVT.getScalarSizeInBits()) == 0 &&
-          SrcVT.getScalarSizeInBits() >= 32) {
-        unsigned Mask = N.getConstantOperandVal(2);
-        unsigned Size = VT.getVectorNumElements();
-        unsigned Scale = VT.getScalarSizeInBits() / SrcVT.getScalarSizeInBits();
-        unsigned ScaleMask = scaleVectorShuffleBlendMask(Mask, Size, Scale);
-        return DAG.getBitcast(
-            VT, DAG.getNode(X86ISD::BLENDI, DL, SrcVT, N0.getOperand(0),
-                            N1.getOperand(0),
-                            DAG.getConstant(ScaleMask, DL, MVT::i8)));
-      }
-    }
-    return SDValue();
-  }
   case X86ISD::PSHUFD:
   case X86ISD::PSHUFLW:
   case X86ISD::PSHUFHW:

lib/Target/X86/X86InstrSSE.td

@@ -6507,22 +6507,6 @@ let Predicates = [HasAVX2] in {
                                    VEX_4V, VEX_L, VEX_WIG;
 }
 
-// Emulate vXi32/vXi64 blends with vXf32/vXf64.
-// ExecutionDomainFixPass will cleanup domains later on.
-let Predicates = [HasAVX] in {
-def : Pat<(X86Blendi (v4i64 VR256:$src1), (v4i64 VR256:$src2), (iPTR imm:$src3)),
-          (VBLENDPDYrri VR256:$src1, VR256:$src2, imm:$src3)>;
-def : Pat<(X86Blendi (v2i64 VR128:$src1), (v2i64 VR128:$src2), (iPTR imm:$src3)),
-          (VBLENDPDrri VR128:$src1, VR128:$src2, imm:$src3)>;
-}
-
-let Predicates = [HasAVX1Only] in {
-def : Pat<(X86Blendi (v8i32 VR256:$src1), (v8i32 VR256:$src2), (iPTR imm:$src3)),
-          (VBLENDPSYrri VR256:$src1, VR256:$src2, imm:$src3)>;
-def : Pat<(X86Blendi (v4i32 VR128:$src1), (v4i32 VR128:$src2), (iPTR imm:$src3)),
-          (VBLENDPSrri VR128:$src1, VR128:$src2, imm:$src3)>;
-}
-
 defm BLENDPS : SS41I_blend_rmi<0x0C, "blendps", X86Blendi, v4f32,
                                VR128, memop, f128mem, 1, SSEPackedSingle,
                                SchedWriteFBlend.XMM, BlendCommuteImm4>;
@@ -6533,13 +6517,6 @@ defm PBLENDW : SS41I_blend_rmi<0x0E, "pblendw", X86Blendi, v8i16,
                                VR128, memop, i128mem, 1, SSEPackedInt,
                                SchedWriteBlend.XMM, BlendCommuteImm8>;
 
-let Predicates = [UseSSE41] in {
-def : Pat<(X86Blendi (v2i64 VR128:$src1), (v2i64 VR128:$src2), (iPTR imm:$src3)),
-          (BLENDPDrri VR128:$src1, VR128:$src2, imm:$src3)>;
-def : Pat<(X86Blendi (v4i32 VR128:$src1), (v4i32 VR128:$src2), (iPTR imm:$src3)),
-          (BLENDPSrri VR128:$src1, VR128:$src2, imm:$src3)>;
-}
-
 // For insertion into the zero index (low half) of a 256-bit vector, it is
 // more efficient to generate a blend with immediate instead of an insert*128.
 let Predicates = [HasAVX] in {

test/CodeGen/X86/avx512-shuffles/partial_permute.ll

@@ -1912,8 +1912,8 @@ define <2 x i64> @test_masked_z_4xi64_to_2xi64_perm_mem_mask0(<4 x i64>* %vp, <2
 define <2 x i64> @test_masked_4xi64_to_2xi64_perm_mem_mask1(<4 x i64>* %vp, <2 x i64> %vec2, <2 x i64> %mask) {
 ; CHECK-LABEL: test_masked_4xi64_to_2xi64_perm_mem_mask1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vmovdqa (%rdi), %xmm2
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm2 = mem[0,1],xmm2[2,3]
+; CHECK-NEXT:    vmovdqa 16(%rdi), %xmm2
+; CHECK-NEXT:    vpblendd {{.*#+}} xmm2 = xmm2[0,1],mem[2,3]
 ; CHECK-NEXT:    vptestnmq %xmm1, %xmm1, %k1
 ; CHECK-NEXT:    vmovdqa64 %xmm2, %xmm0 {%k1}
 ; CHECK-NEXT:    retq
@@ -1927,8 +1927,8 @@ define <2 x i64> @test_masked_4xi64_to_2xi64_perm_mem_mask1(<4 x i64>* %vp, <2 x
 define <2 x i64> @test_masked_z_4xi64_to_2xi64_perm_mem_mask1(<4 x i64>* %vp, <2 x i64> %mask) {
 ; CHECK-LABEL: test_masked_z_4xi64_to_2xi64_perm_mem_mask1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vmovdqa (%rdi), %xmm1
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm1 = mem[0,1],xmm1[2,3]
+; CHECK-NEXT:    vmovdqa 16(%rdi), %xmm1
+; CHECK-NEXT:    vpblendd {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]
 ; CHECK-NEXT:    vptestnmq %xmm0, %xmm0, %k1
 ; CHECK-NEXT:    vmovdqa64 %xmm1, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    retq
@@ -2553,8 +2553,9 @@ define <4 x i64> @test_masked_z_8xi64_to_4xi64_perm_mem_mask7(<8 x i64>* %vp, <4
 define <2 x i64> @test_8xi64_to_2xi64_perm_mem_mask0(<8 x i64>* %vp) {
 ; CHECK-LABEL: test_8xi64_to_2xi64_perm_mem_mask0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vmovaps (%rdi), %xmm0
-; CHECK-NEXT:    vblendps {{.*#+}} xmm0 = mem[0,1],xmm0[2,3]
+; CHECK-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    vmovaps 32(%rdi), %xmm1
+; CHECK-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
 ; CHECK-NEXT:    retq
   %vec = load <8 x i64>, <8 x i64>* %vp
   %res = shufflevector <8 x i64> %vec, <8 x i64> undef, <2 x i32> <i32 4, i32 1>
@@ -2563,10 +2564,10 @@ define <2 x i64> @test_8xi64_to_2xi64_perm_mem_mask0(<8 x i64>* %vp) {
 define <2 x i64> @test_masked_8xi64_to_2xi64_perm_mem_mask0(<8 x i64>* %vp, <2 x i64> %vec2, <2 x i64> %mask) {
 ; CHECK-LABEL: test_masked_8xi64_to_2xi64_perm_mem_mask0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vmovdqa (%rdi), %xmm2
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm2 = mem[0,1],xmm2[2,3]
+; CHECK-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
+; CHECK-NEXT:    vmovdqa 32(%rdi), %xmm3
 ; CHECK-NEXT:    vptestnmq %xmm1, %xmm1, %k1
-; CHECK-NEXT:    vmovdqa64 %xmm2, %xmm0 {%k1}
+; CHECK-NEXT:    vpunpcklqdq {{.*#+}} xmm0 {%k1} = xmm3[0],xmm2[0]
 ; CHECK-NEXT:    retq
   %vec = load <8 x i64>, <8 x i64>* %vp
   %shuf = shufflevector <8 x i64> %vec, <8 x i64> undef, <2 x i32> <i32 4, i32 1>
@@ -2578,10 +2579,10 @@ define <2 x i64> @test_masked_8xi64_to_2xi64_perm_mem_mask0(<8 x i64>* %vp, <2 x
 define <2 x i64> @test_masked_z_8xi64_to_2xi64_perm_mem_mask0(<8 x i64>* %vp, <2 x i64> %mask) {
 ; CHECK-LABEL: test_masked_z_8xi64_to_2xi64_perm_mem_mask0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vmovdqa (%rdi), %xmm1
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm1 = mem[0,1],xmm1[2,3]
+; CHECK-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
+; CHECK-NEXT:    vmovdqa 32(%rdi), %xmm2
 ; CHECK-NEXT:    vptestnmq %xmm0, %xmm0, %k1
-; CHECK-NEXT:    vmovdqa64 %xmm1, %xmm0 {%k1} {z}
+; CHECK-NEXT:    vpunpcklqdq {{.*#+}} xmm0 {%k1} {z} = xmm2[0],xmm1[0]
 ; CHECK-NEXT:    retq
   %vec = load <8 x i64>, <8 x i64>* %vp
   %shuf = shufflevector <8 x i64> %vec, <8 x i64> undef, <2 x i32> <i32 4, i32 1>

test/CodeGen/X86/combine-sdiv.ll

@@ -1701,7 +1701,8 @@ define <4 x i64> @combine_vec_sdiv_by_pow2b_v4i64(<4 x i64> %x) {
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm2
 ; AVX1-NEXT:    vpsrlq $62, %xmm2, %xmm2
 ; AVX1-NEXT:    vpaddq %xmm2, %xmm0, %xmm2
-; AVX1-NEXT:    vpsrlq $2, %xmm2, %xmm2
+; AVX1-NEXT:    vpsrlq $2, %xmm2, %xmm3
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1,2,3],xmm3[4,5,6,7]
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [9223372036854775808,2305843009213693952]
 ; AVX1-NEXT:    vpxor %xmm3, %xmm2, %xmm2
 ; AVX1-NEXT:    vpsubq %xmm3, %xmm2, %xmm2
@@ -1889,7 +1890,8 @@ define <8 x i64> @combine_vec_sdiv_by_pow2b_v8i64(<8 x i64> %x) {
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm2, %xmm5
 ; AVX1-NEXT:    vpsrlq $62, %xmm5, %xmm5
 ; AVX1-NEXT:    vpaddq %xmm5, %xmm0, %xmm5
-; AVX1-NEXT:    vpsrlq $2, %xmm5, %xmm5
+; AVX1-NEXT:    vpsrlq $2, %xmm5, %xmm6
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm5 = xmm5[0,1,2,3],xmm6[4,5,6,7]
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm6 = [9223372036854775808,2305843009213693952]
 ; AVX1-NEXT:    vpxor %xmm6, %xmm5, %xmm5
 ; AVX1-NEXT:    vpsubq %xmm6, %xmm5, %xmm5
@@ -1909,7 +1911,8 @@ define <8 x i64> @combine_vec_sdiv_by_pow2b_v8i64(<8 x i64> %x) {
 ; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm2, %xmm2
 ; AVX1-NEXT:    vpsrlq $62, %xmm2, %xmm2
 ; AVX1-NEXT:    vpaddq %xmm2, %xmm1, %xmm2
-; AVX1-NEXT:    vpsrlq $2, %xmm2, %xmm2
+; AVX1-NEXT:    vpsrlq $2, %xmm2, %xmm4
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1,2,3],xmm4[4,5,6,7]
 ; AVX1-NEXT:    vpxor %xmm6, %xmm2, %xmm2
 ; AVX1-NEXT:    vpsubq %xmm6, %xmm2, %xmm2
 ; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2

test/CodeGen/X86/insertelement-ones.ll

@@ -291,15 +291,21 @@ define <16 x i16> @insert_v16i16_x12345x789ABCDEx(<16 x i16> %a) {
 ; AVX2-LABEL: insert_v16i16_x12345x789ABCDEx:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
-; AVX2-NEXT:    vpblendw {{.*#+}} ymm2 = ymm1[0],ymm0[1,2,3,4,5],ymm1[6],ymm0[7],ymm1[8],ymm0[9,10,11,12,13],ymm1[14],ymm0[15]
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm2 = ymm1[0],ymm0[1,2,3,4,5,6,7],ymm1[8],ymm0[9,10,11,12,13,14,15]
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm1[6],ymm2[7,8,9,10,11,12,13],ymm1[14],ymm2[15]
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
 ; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm1[7],ymm0[8,9,10,11,12,13,14],ymm1[15]
 ; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: insert_v16i16_x12345x789ABCDEx:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
-; AVX512-NEXT:    vpblendw {{.*#+}} ymm2 = ymm1[0],ymm0[1,2,3,4,5],ymm1[6],ymm0[7],ymm1[8],ymm0[9,10,11,12,13],ymm1[14],ymm0[15]
+; AVX512-NEXT:    vpblendw {{.*#+}} ymm2 = ymm1[0],ymm0[1,2,3,4,5,6,7],ymm1[8],ymm0[9,10,11,12,13,14,15]
+; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm0[4,5,6,7]
+; AVX512-NEXT:    vpblendw {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm1[6],ymm2[7,8,9,10,11,12,13],ymm1[14],ymm2[15]
+; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
 ; AVX512-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm1[7],ymm0[8,9,10,11,12,13,14],ymm1[15]
 ; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
 ; AVX512-NEXT:    retq

test/CodeGen/X86/known-signbits-vector.ll

@@ -89,7 +89,7 @@ define float @signbits_ashr_extract_sitofp_1(<2 x i64> %a0) nounwind {
 ; X32:       # %bb.0:
 ; X32-NEXT:    pushl %eax
 ; X32-NEXT:    vpsrlq $32, %xmm0, %xmm0
-; X32-NEXT:    vmovdqa {{.*#+}} xmm1 = [2147483648,0,1,0]
+; X32-NEXT:    vmovdqa {{.*#+}} xmm1 = [0,32768,0,0,1,0,0,0]
 ; X32-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X32-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
 ; X32-NEXT:    vcvtdq2ps %xmm0, %xmm0
@@ -115,7 +115,7 @@ define float @signbits_ashr_shl_extract_sitofp(<2 x i64> %a0) nounwind {
 ; X32:       # %bb.0:
 ; X32-NEXT:    pushl %eax
 ; X32-NEXT:    vpsrlq $61, %xmm0, %xmm0
-; X32-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,0,8,0]
+; X32-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,0,0,0,8,0,0,0]
 ; X32-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X32-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
 ; X32-NEXT:    vpsllq $20, %xmm0, %xmm0
@@ -231,7 +231,7 @@ define float @signbits_ashr_sext_sextinreg_and_extract_sitofp(<2 x i64> %a0, <2
 ; X32:       # %bb.0:
 ; X32-NEXT:    pushl %eax
 ; X32-NEXT:    vpsrlq $61, %xmm0, %xmm0
-; X32-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,0,8,0]
+; X32-NEXT:    vmovdqa {{.*#+}} xmm1 = [4,0,0,0,8,0,0,0]
 ; X32-NEXT:    vpxor %xmm1, %xmm0, %xmm0
 ; X32-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
 ; X32-NEXT:    vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
@@ -272,7 +272,7 @@ define float @signbits_ashr_sextvecinreg_bitops_extract_sitofp(<2 x i64> %a0, <4
 ; X32-NEXT:    vpsrlq $60, %xmm0, %xmm2
 ; X32-NEXT:    vpsrlq $61, %xmm0, %xmm0
 ; X32-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4,5,6,7]
-; X32-NEXT:    vmovdqa {{.*#+}} xmm2 = [4,0,8,0]
+; X32-NEXT:    vmovdqa {{.*#+}} xmm2 = [4,0,0,0,8,0,0,0]
 ; X32-NEXT:    vpxor %xmm2, %xmm0, %xmm0
 ; X32-NEXT:    vpsubq %xmm2, %xmm0, %xmm0
 ; X32-NEXT:    vpmovsxdq %xmm1, %xmm1
@@ -322,7 +322,7 @@ define <4 x float> @signbits_ashr_sext_select_shuffle_sitofp(<4 x i64> %a0, <4 x
 ; X32-NEXT:    vpmovsxdq 8(%ebp), %xmm4
 ; X32-NEXT:    vextractf128 $1, %ymm2, %xmm5
 ; X32-NEXT:    vpsrlq $33, %xmm5, %xmm5
-; X32-NEXT:    vmovdqa {{.*#+}} xmm6 = [1073741824,0,1,0]
+; X32-NEXT:    vmovdqa {{.*#+}} xmm6 = [0,16384,0,0,1,0,0,0]
 ; X32-NEXT:    vpxor %xmm6, %xmm5, %xmm5
 ; X32-NEXT:    vpsubq %xmm6, %xmm5, %xmm5
 ; X32-NEXT:    vpsrlq $33, %xmm2, %xmm2