[X86][SSE] Use ISD::MULHS for constant vXi16 ISD::SRA lowering (PR38151)

[RKSimon]

Summary

Similar to the existing ISD::SRL constant vector shifts from D49562, this patch adds ISD::SRA support with ISD::MULHS.

As we're dealing with signed values, we have to handle shift by zero and shift by one special cases, so XOP+AVX2/AVX512 splitting/extension is still a better solution - really we should still use ISD::MULHS if one of the special cases are used but for now I've just left a TODO and filtered by isKnownNeverZero.

[RKSimon]

ping?

[craig.topper]

LGTM

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

if (Opc == ISD::SRL && ConstantAmt &&
     if (SDValue Scale = convertShiftLeftToScale(RAmt, dl, Subtarget, DAG)) {
       SDValue Zero = DAG.getConstant(0, dl, VT);
       SDValue ZAmt = DAG.getSetCC(dl, VT, Amt, Zero, ISD::SETEQ);
       SDValue Res = DAG.getNode(ISD::MULHU, dl, VT, R, Scale);
       return DAG.getSelect(dl, VT, ZAmt, R, Res);
     }
   }
 
+  // Constant ISD::SRA can be performed efficiently on vXi16 vectors as we
+  // can replace with ISD::MULHS, creating scale factor from (NumEltBits - Amt).
+  // TODO: Special case handling for shift by 0/1, really we can afford either
+  // of these cases in pre-SSE41/XOP/AVX512 but not both.
+  if (Opc == ISD::SRA && ConstantAmt &&
+      (VT == MVT::v8i16 || (VT == MVT::v16i16 && Subtarget.hasInt256())) &&
+      ((Subtarget.hasSSE41() && !Subtarget.hasXOP() &&
+        !Subtarget.hasAVX512()) ||
+       DAG.isKnownNeverZero(Amt))) {
+    SDValue EltBits = DAG.getConstant(EltSizeInBits, dl, VT);
+    SDValue RAmt = DAG.getNode(ISD::SUB, dl, VT, EltBits, Amt);
+    if (SDValue Scale = convertShiftLeftToScale(RAmt, dl, Subtarget, DAG)) {
+      SDValue Amt0 =
+          DAG.getSetCC(dl, VT, Amt, DAG.getConstant(0, dl, VT), ISD::SETEQ);
+      SDValue Amt1 =
+          DAG.getSetCC(dl, VT, Amt, DAG.getConstant(1, dl, VT), ISD::SETEQ);
+      SDValue Sra1 =
+          getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, 1, DAG);
+      SDValue Res = DAG.getNode(ISD::MULHS, dl, VT, R, Scale);
+      Res = DAG.getSelect(dl, VT, Amt0, R, Res);
+      return DAG.getSelect(dl, VT, Amt1, Sra1, Res);
+    }
+  }
+
   // v4i32 Non Uniform Shifts.
   // If the shift amount is constant we can shift each lane using the SSE2
   // immediate shifts, else we need to zero-extend each lane to the lower i64
   // and shift using the SSE2 variable shifts.
   // The separate results can then be blended together.
   if (VT == MVT::v4i32) {
     SDValue Amt0, Amt1, Amt2, Amt3;
     if (ConstantAmt) {

llvm/trunk/test/CodeGen/X86/combine-sdiv.ll

 ;
 ; SSE41-LABEL: combine_vec_sdiv_by_pow2b_v8i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm1
 ; SSE41-NEXT:    psraw $15, %xmm1
 ; SSE41-NEXT:    pmulhuw {{.*}}(%rip), %xmm1
 ; SSE41-NEXT:    paddw %xmm0, %xmm1
 ; SSE41-NEXT:    movdqa %xmm1, %xmm2
-; SSE41-NEXT:    psraw $4, %xmm2
-; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm1[0,1,2],xmm2[3],xmm1[4],xmm2[5,6],xmm1[7]
-; SSE41-NEXT:    movdqa %xmm2, %xmm3
-; SSE41-NEXT:    psraw $2, %xmm3
-; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm2[0],xmm3[1],xmm2[2,3],xmm3[4],xmm2[5],xmm3[6],xmm2[7]
-; SSE41-NEXT:    movdqa %xmm3, %xmm1
-; SSE41-NEXT:    psraw $1, %xmm1
-; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm3[0,1],xmm1[2],xmm3[3],xmm1[4,5],xmm3[6],xmm1[7]
+; SSE41-NEXT:    psraw $1, %xmm2
+; SSE41-NEXT:    pmulhw {{.*}}(%rip), %xmm1
+; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3,4,5,6],xmm2[7]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3,4,5,6,7]
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: combine_vec_sdiv_by_pow2b_v8i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpsraw $15, %xmm0, %xmm1
 ; AVX1-NEXT:    vpmulhuw {{.*}}(%rip), %xmm1, %xmm1
 ; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm1
-; AVX1-NEXT:    vpsraw $4, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm2[3],xmm1[4],xmm2[5,6],xmm1[7]
-; AVX1-NEXT:    vpsraw $2, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3],xmm2[4],xmm1[5],xmm2[6],xmm1[7]
 ; AVX1-NEXT:    vpsraw $1, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3],xmm2[4,5],xmm1[6],xmm2[7]
+; AVX1-NEXT:    vpmulhw {{.*}}(%rip), %xmm1, %xmm1
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3,4,5,6],xmm2[7]
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: combine_vec_sdiv_by_pow2b_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpsraw $15, %xmm0, %xmm1
 ; AVX2-NEXT:    vpmulhuw {{.*}}(%rip), %xmm1, %xmm1
 ; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm1
-; AVX2-NEXT:    vpmovsxwd %xmm1, %ymm1
-; AVX2-NEXT:    vpsravd {{.*}}(%rip), %ymm1, %ymm1
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
+; AVX2-NEXT:    vpsraw $1, %xmm1, %xmm2
+; AVX2-NEXT:    vpmulhw {{.*}}(%rip), %xmm1, %xmm1
+; AVX2-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3,4,5,6],xmm2[7]
 ; AVX2-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
-; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: combine_vec_sdiv_by_pow2b_v8i16:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vpsraw $15, %xmm0, %xmm1
 ; AVX512F-NEXT:    vpmulhuw {{.*}}(%rip), %xmm1, %xmm1
 ; AVX512F-NEXT:    vpaddw %xmm1, %xmm0, %xmm1
 ; AVX512F-NEXT:    vpmovsxwd %xmm1, %ymm1

 ; SSE2-NEXT:    por %xmm7, %xmm2
 ; SSE2-NEXT:    movdqa %xmm2, %xmm1
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: combine_vec_sdiv_by_pow2b_v16i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    psraw $15, %xmm2
-; SSE41-NEXT:    movdqa {{.*#+}} xmm3 = [1,4,2,16,8,32,64,2]
-; SSE41-NEXT:    pmulhuw %xmm3, %xmm2
+; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [1,4,2,16,8,32,64,2]
+; SSE41-NEXT:    pmulhuw %xmm4, %xmm2
 ; SSE41-NEXT:    paddw %xmm0, %xmm2
-; SSE41-NEXT:    movdqa %xmm2, %xmm4
-; SSE41-NEXT:    psraw $4, %xmm4
-; SSE41-NEXT:    pblendw {{.*#+}} xmm4 = xmm2[0,1,2],xmm4[3],xmm2[4],xmm4[5,6],xmm2[7]
-; SSE41-NEXT:    movdqa %xmm4, %xmm5
-; SSE41-NEXT:    psraw $2, %xmm5
-; SSE41-NEXT:    pblendw {{.*#+}} xmm5 = xmm4[0],xmm5[1],xmm4[2,3],xmm5[4],xmm4[5],xmm5[6],xmm4[7]
-; SSE41-NEXT:    movdqa %xmm5, %xmm2
+; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = <u,16384,32768,4096,8192,2048,1024,32768>
+; SSE41-NEXT:    movdqa %xmm2, %xmm3
+; SSE41-NEXT:    pmulhw %xmm5, %xmm3
 ; SSE41-NEXT:    psraw $1, %xmm2
-; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm5[0,1],xmm2[2],xmm5[3],xmm2[4,5],xmm5[6],xmm2[7]
+; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2],xmm3[3,4,5,6],xmm2[7]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3,4,5,6,7]
-; SSE41-NEXT:    movdqa %xmm1, %xmm0
-; SSE41-NEXT:    psraw $15, %xmm0
-; SSE41-NEXT:    pmulhuw %xmm3, %xmm0
-; SSE41-NEXT:    paddw %xmm1, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm3
-; SSE41-NEXT:    psraw $4, %xmm3
-; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm0[0,1,2],xmm3[3],xmm0[4],xmm3[5,6],xmm0[7]
-; SSE41-NEXT:    movdqa %xmm3, %xmm0
-; SSE41-NEXT:    psraw $2, %xmm0
-; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm3[0],xmm0[1],xmm3[2,3],xmm0[4],xmm3[5],xmm0[6],xmm3[7]
-; SSE41-NEXT:    movdqa %xmm0, %xmm3
+; SSE41-NEXT:    movdqa %xmm1, %xmm3
+; SSE41-NEXT:    psraw $15, %xmm3
+; SSE41-NEXT:    pmulhuw %xmm4, %xmm3
+; SSE41-NEXT:    paddw %xmm1, %xmm3
+; SSE41-NEXT:    pmulhw %xmm3, %xmm5
 ; SSE41-NEXT:    psraw $1, %xmm3
-; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm0[0,1],xmm3[2],xmm0[3],xmm3[4,5],xmm0[6],xmm3[7]
+; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm5[0,1],xmm3[2],xmm5[3,4,5,6],xmm3[7]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm1[0],xmm3[1,2,3,4,5,6,7]
 ; SSE41-NEXT:    movdqa %xmm2, %xmm0
 ; SSE41-NEXT:    movdqa %xmm3, %xmm1
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: combine_vec_sdiv_by_pow2b_v16i16:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; AVX1-NEXT:    vpsraw $15, %xmm1, %xmm2
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [1,4,2,16,8,32,64,2]
 ; AVX1-NEXT:    vpmulhuw %xmm3, %xmm2, %xmm2
 ; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vpsraw $4, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm2[3],xmm1[4],xmm2[5,6],xmm1[7]
-; AVX1-NEXT:    vpsraw $2, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3],xmm2[4],xmm1[5],xmm2[6],xmm1[7]
-; AVX1-NEXT:    vpsraw $1, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3],xmm2[4,5],xmm1[6],xmm2[7]
-; AVX1-NEXT:    vpsraw $15, %xmm0, %xmm2
-; AVX1-NEXT:    vpmulhuw %xmm3, %xmm2, %xmm2
-; AVX1-NEXT:    vpaddw %xmm2, %xmm0, %xmm2
-; AVX1-NEXT:    vpsraw $4, %xmm2, %xmm3
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1,2],xmm3[3],xmm2[4],xmm3[5,6],xmm2[7]
-; AVX1-NEXT:    vpsraw $2, %xmm2, %xmm3
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0],xmm3[1],xmm2[2,3],xmm3[4],xmm2[5],xmm3[6],xmm2[7]
-; AVX1-NEXT:    vpsraw $1, %xmm2, %xmm3
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm3[2],xmm2[3],xmm3[4,5],xmm2[6],xmm3[7]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = <u,16384,32768,4096,8192,2048,1024,32768>
+; AVX1-NEXT:    vpmulhw %xmm2, %xmm1, %xmm4
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm4 = xmm1[0],xmm4[1,2,3,4,5,6,7]
+; AVX1-NEXT:    vpsraw $1, %xmm1, %xmm1
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm4[0,1],xmm1[2],xmm4[3,4,5,6],xmm1[7]
+; AVX1-NEXT:    vpsraw $15, %xmm0, %xmm4
+; AVX1-NEXT:    vpmulhuw %xmm3, %xmm4, %xmm3
+; AVX1-NEXT:    vpaddw %xmm3, %xmm0, %xmm3
+; AVX1-NEXT:    vpmulhw %xmm2, %xmm3, %xmm2
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0],xmm2[1,2,3,4,5,6,7]
+; AVX1-NEXT:    vpsraw $1, %xmm3, %xmm3
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm3[2],xmm2[3,4,5,6],xmm3[7]
 ; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm2, %ymm1
 ; AVX1-NEXT:    vmovaps {{.*#+}} ymm2 = [0,65535,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,65535]
 ; AVX1-NEXT:    vandps %ymm2, %ymm1, %ymm1
 ; AVX1-NEXT:    vandnps %ymm0, %ymm2, %ymm0
 ; AVX1-NEXT:    vorps %ymm0, %ymm1, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: combine_vec_sdiv_by_pow2b_v16i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,2,1,4,3,5,6,1,0,2,1,4,3,5,6,1]
-; AVX2-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
-; AVX2-NEXT:    vpsraw $15, %ymm0, %ymm4
-; AVX2-NEXT:    vpmulhuw {{.*}}(%rip), %ymm4, %ymm4
-; AVX2-NEXT:    vpaddw %ymm4, %ymm0, %ymm4
-; AVX2-NEXT:    vpunpckhwd {{.*#+}} ymm5 = ymm1[4],ymm4[4],ymm1[5],ymm4[5],ymm1[6],ymm4[6],ymm1[7],ymm4[7],ymm1[12],ymm4[12],ymm1[13],ymm4[13],ymm1[14],ymm4[14],ymm1[15],ymm4[15]
-; AVX2-NEXT:    vpsravd %ymm3, %ymm5, %ymm3
-; AVX2-NEXT:    vpsrld $16, %ymm3, %ymm3
-; AVX2-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; AVX2-NEXT:    vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm4[0],ymm1[1],ymm4[1],ymm1[2],ymm4[2],ymm1[3],ymm4[3],ymm1[8],ymm4[8],ymm1[9],ymm4[9],ymm1[10],ymm4[10],ymm1[11],ymm4[11]
-; AVX2-NEXT:    vpsravd %ymm2, %ymm1, %ymm1
-; AVX2-NEXT:    vpsrld $16, %ymm1, %ymm1
-; AVX2-NEXT:    vpackusdw %ymm3, %ymm1, %ymm1
+; AVX2-NEXT:    vpsraw $15, %ymm0, %ymm1
+; AVX2-NEXT:    vpmulhuw {{.*}}(%rip), %ymm1, %ymm1
+; AVX2-NEXT:    vpaddw %ymm1, %ymm0, %ymm1
+; AVX2-NEXT:    vpsraw $1, %ymm1, %ymm2
+; AVX2-NEXT:    vpmulhw {{.*}}(%rip), %ymm1, %ymm1
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6],ymm2[7],ymm1[8,9],ymm2[10],ymm1[11,12,13,14],ymm2[15]
 ; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3,4,5,6,7],ymm0[8],ymm1[9,10,11,12,13,14,15]
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: combine_vec_sdiv_by_pow2b_v16i16:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vpsraw $15, %ymm0, %ymm1
 ; AVX512F-NEXT:    vpmulhuw {{.*}}(%rip), %ymm1, %ymm1
 ; AVX512F-NEXT:    vpaddw %ymm1, %ymm0, %ymm1

 ; SSE2-NEXT:    movdqa %xmm12, %xmm3
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: combine_vec_sdiv_by_pow2b_v32i16:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa %xmm1, %xmm4
 ; SSE41-NEXT:    movdqa %xmm0, %xmm1
 ; SSE41-NEXT:    psraw $15, %xmm0
-; SSE41-NEXT:    movdqa {{.*#+}} xmm5 = [1,4,2,16,8,32,64,2]
-; SSE41-NEXT:    pmulhuw %xmm5, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm7 = [1,4,2,16,8,32,64,2]
+; SSE41-NEXT:    pmulhuw %xmm7, %xmm0
 ; SSE41-NEXT:    paddw %xmm1, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm6
-; SSE41-NEXT:    psraw $4, %xmm6
-; SSE41-NEXT:    pblendw {{.*#+}} xmm6 = xmm0[0,1,2],xmm6[3],xmm0[4],xmm6[5,6],xmm0[7]
-; SSE41-NEXT:    movdqa %xmm6, %xmm7
-; SSE41-NEXT:    psraw $2, %xmm7
-; SSE41-NEXT:    pblendw {{.*#+}} xmm7 = xmm6[0],xmm7[1],xmm6[2,3],xmm7[4],xmm6[5],xmm7[6],xmm6[7]
-; SSE41-NEXT:    movdqa %xmm7, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm6 = <u,16384,32768,4096,8192,2048,1024,32768>
+; SSE41-NEXT:    movdqa %xmm0, %xmm5
+; SSE41-NEXT:    pmulhw %xmm6, %xmm5
 ; SSE41-NEXT:    psraw $1, %xmm0
-; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm7[0,1],xmm0[2],xmm7[3],xmm0[4,5],xmm7[6],xmm0[7]
+; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm5[0,1],xmm0[2],xmm5[3,4,5,6],xmm0[7]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3,4,5,6,7]
 ; SSE41-NEXT:    movdqa %xmm4, %xmm1
 ; SSE41-NEXT:    psraw $15, %xmm1
-; SSE41-NEXT:    pmulhuw %xmm5, %xmm1
+; SSE41-NEXT:    pmulhuw %xmm7, %xmm1
 ; SSE41-NEXT:    paddw %xmm4, %xmm1
-; SSE41-NEXT:    movdqa %xmm1, %xmm6
-; SSE41-NEXT:    psraw $4, %xmm6
-; SSE41-NEXT:    pblendw {{.*#+}} xmm6 = xmm1[0,1,2],xmm6[3],xmm1[4],xmm6[5,6],xmm1[7]
-; SSE41-NEXT:    movdqa %xmm6, %xmm7
-; SSE41-NEXT:    psraw $2, %xmm7
-; SSE41-NEXT:    pblendw {{.*#+}} xmm7 = xmm6[0],xmm7[1],xmm6[2,3],xmm7[4],xmm6[5],xmm7[6],xmm6[7]
-; SSE41-NEXT:    movdqa %xmm7, %xmm1
+; SSE41-NEXT:    movdqa %xmm1, %xmm5
+; SSE41-NEXT:    pmulhw %xmm6, %xmm5
 ; SSE41-NEXT:    psraw $1, %xmm1
-; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm7[0,1],xmm1[2],xmm7[3],xmm1[4,5],xmm7[6],xmm1[7]
+; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm5[0,1],xmm1[2],xmm5[3,4,5,6],xmm1[7]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm4[0],xmm1[1,2,3,4,5,6,7]
 ; SSE41-NEXT:    movdqa %xmm2, %xmm4
 ; SSE41-NEXT:    psraw $15, %xmm4
-; SSE41-NEXT:    pmulhuw %xmm5, %xmm4
+; SSE41-NEXT:    pmulhuw %xmm7, %xmm4
 ; SSE41-NEXT:    paddw %xmm2, %xmm4
-; SSE41-NEXT:    movdqa %xmm4, %xmm6
-; SSE41-NEXT:    psraw $4, %xmm6
-; SSE41-NEXT:    pblendw {{.*#+}} xmm6 = xmm4[0,1,2],xmm6[3],xmm4[4],xmm6[5,6],xmm4[7]
-; SSE41-NEXT:    movdqa %xmm6, %xmm7
-; SSE41-NEXT:    psraw $2, %xmm7
-; SSE41-NEXT:    pblendw {{.*#+}} xmm7 = xmm6[0],xmm7[1],xmm6[2,3],xmm7[4],xmm6[5],xmm7[6],xmm6[7]
-; SSE41-NEXT:    movdqa %xmm7, %xmm4
+; SSE41-NEXT:    movdqa %xmm4, %xmm5
+; SSE41-NEXT:    pmulhw %xmm6, %xmm5
 ; SSE41-NEXT:    psraw $1, %xmm4
-; SSE41-NEXT:    pblendw {{.*#+}} xmm4 = xmm7[0,1],xmm4[2],xmm7[3],xmm4[4,5],xmm7[6],xmm4[7]
+; SSE41-NEXT:    pblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2],xmm5[3,4,5,6],xmm4[7]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm4 = xmm2[0],xmm4[1,2,3,4,5,6,7]
-; SSE41-NEXT:    movdqa %xmm3, %xmm2
-; SSE41-NEXT:    psraw $15, %xmm2
-; SSE41-NEXT:    pmulhuw %xmm5, %xmm2
-; SSE41-NEXT:    paddw %xmm3, %xmm2
-; SSE41-NEXT:    movdqa %xmm2, %xmm5
-; SSE41-NEXT:    psraw $4, %xmm5
-; SSE41-NEXT:    pblendw {{.*#+}} xmm5 = xmm2[0,1,2],xmm5[3],xmm2[4],xmm5[5,6],xmm2[7]
-; SSE41-NEXT:    movdqa %xmm5, %xmm2
-; SSE41-NEXT:    psraw $2, %xmm2
-; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm5[0],xmm2[1],xmm5[2,3],xmm2[4],xmm5[5],xmm2[6],xmm5[7]
-; SSE41-NEXT:    movdqa %xmm2, %xmm5
+; SSE41-NEXT:    movdqa %xmm3, %xmm5
+; SSE41-NEXT:    psraw $15, %xmm5
+; SSE41-NEXT:    pmulhuw %xmm7, %xmm5
+; SSE41-NEXT:    paddw %xmm3, %xmm5
+; SSE41-NEXT:    pmulhw %xmm5, %xmm6
 ; SSE41-NEXT:    psraw $1, %xmm5
-; SSE41-NEXT:    pblendw {{.*#+}} xmm5 = xmm2[0,1],xmm5[2],xmm2[3],xmm5[4,5],xmm2[6],xmm5[7]
+; SSE41-NEXT:    pblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2],xmm6[3,4,5,6],xmm5[7]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm5 = xmm3[0],xmm5[1,2,3,4,5,6,7]
 ; SSE41-NEXT:    movdqa %xmm4, %xmm2
 ; SSE41-NEXT:    movdqa %xmm5, %xmm3
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: combine_vec_sdiv_by_pow2b_v32i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
-; AVX1-NEXT:    vpsraw $15, %xmm3, %xmm4
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [1,4,2,16,8,32,64,2]
-; AVX1-NEXT:    vpmulhuw %xmm2, %xmm4, %xmm4
-; AVX1-NEXT:    vpaddw %xmm4, %xmm3, %xmm3
-; AVX1-NEXT:    vpsraw $4, %xmm3, %xmm4
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1,2],xmm4[3],xmm3[4],xmm4[5,6],xmm3[7]
-; AVX1-NEXT:    vpsraw $2, %xmm3, %xmm4
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0],xmm4[1],xmm3[2,3],xmm4[4],xmm3[5],xmm4[6],xmm3[7]
-; AVX1-NEXT:    vpsraw $1, %xmm3, %xmm4
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm4[2],xmm3[3],xmm4[4,5],xmm3[6],xmm4[7]
-; AVX1-NEXT:    vpsraw $15, %xmm0, %xmm4
-; AVX1-NEXT:    vpmulhuw %xmm2, %xmm4, %xmm4
-; AVX1-NEXT:    vpaddw %xmm4, %xmm0, %xmm4
-; AVX1-NEXT:    vpsraw $4, %xmm4, %xmm5
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm4 = xmm4[0,1,2],xmm5[3],xmm4[4],xmm5[5,6],xmm4[7]
-; AVX1-NEXT:    vpsraw $2, %xmm4, %xmm5
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm4 = xmm4[0],xmm5[1],xmm4[2,3],xmm5[4],xmm4[5],xmm5[6],xmm4[7]
-; AVX1-NEXT:    vpsraw $1, %xmm4, %xmm5
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm4 = xmm4[0,1],xmm5[2],xmm4[3],xmm5[4,5],xmm4[6],xmm5[7]
-; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm4, %ymm3
-; AVX1-NEXT:    vmovaps {{.*#+}} ymm4 = [0,65535,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,65535]
-; AVX1-NEXT:    vandps %ymm4, %ymm3, %ymm3
-; AVX1-NEXT:    vandnps %ymm0, %ymm4, %ymm0
-; AVX1-NEXT:    vorps %ymm0, %ymm3, %ymm0
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3
-; AVX1-NEXT:    vpsraw $15, %xmm3, %xmm5
-; AVX1-NEXT:    vpmulhuw %xmm2, %xmm5, %xmm5
-; AVX1-NEXT:    vpaddw %xmm5, %xmm3, %xmm3
-; AVX1-NEXT:    vpsraw $4, %xmm3, %xmm5
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1,2],xmm5[3],xmm3[4],xmm5[5,6],xmm3[7]
-; AVX1-NEXT:    vpsraw $2, %xmm3, %xmm5
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0],xmm5[1],xmm3[2,3],xmm5[4],xmm3[5],xmm5[6],xmm3[7]
-; AVX1-NEXT:    vpsraw $1, %xmm3, %xmm5
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm5[2],xmm3[3],xmm5[4,5],xmm3[6],xmm5[7]
-; AVX1-NEXT:    vpsraw $15, %xmm1, %xmm5
-; AVX1-NEXT:    vpmulhuw %xmm2, %xmm5, %xmm2
-; AVX1-NEXT:    vpaddw %xmm2, %xmm1, %xmm2
-; AVX1-NEXT:    vpsraw $4, %xmm2, %xmm5
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1,2],xmm5[3],xmm2[4],xmm5[5,6],xmm2[7]
-; AVX1-NEXT:    vpsraw $2, %xmm2, %xmm5
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0],xmm5[1],xmm2[2,3],xmm5[4],xmm2[5],xmm5[6],xmm2[7]
-; AVX1-NEXT:    vpsraw $1, %xmm2, %xmm5
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm5[2],xmm2[3],xmm5[4,5],xmm2[6],xmm5[7]
-; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
-; AVX1-NEXT:    vandps %ymm4, %ymm2, %ymm2
-; AVX1-NEXT:    vandnps %ymm1, %ymm4, %ymm1
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT:    vpsraw $15, %xmm2, %xmm3
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [1,4,2,16,8,32,64,2]
+; AVX1-NEXT:    vpmulhuw %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vpaddw %xmm3, %xmm2, %xmm2
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = <u,16384,32768,4096,8192,2048,1024,32768>
+; AVX1-NEXT:    vpmulhw %xmm3, %xmm2, %xmm5
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm5 = xmm2[0],xmm5[1,2,3,4,5,6,7]
+; AVX1-NEXT:    vpsraw $1, %xmm2, %xmm2
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm2 = xmm5[0,1],xmm2[2],xmm5[3,4,5,6],xmm2[7]
+; AVX1-NEXT:    vpsraw $15, %xmm0, %xmm5
+; AVX1-NEXT:    vpmulhuw %xmm4, %xmm5, %xmm5
+; AVX1-NEXT:    vpaddw %xmm5, %xmm0, %xmm5
+; AVX1-NEXT:    vpmulhw %xmm3, %xmm5, %xmm6
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm6 = xmm5[0],xmm6[1,2,3,4,5,6,7]
+; AVX1-NEXT:    vpsraw $1, %xmm5, %xmm5
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2],xmm6[3,4,5,6],xmm5[7]
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm5, %ymm2
+; AVX1-NEXT:    vmovaps {{.*#+}} ymm5 = [0,65535,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,65535]
+; AVX1-NEXT:    vandps %ymm5, %ymm2, %ymm2
+; AVX1-NEXT:    vandnps %ymm0, %ymm5, %ymm0
+; AVX1-NEXT:    vorps %ymm0, %ymm2, %ymm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpsraw $15, %xmm2, %xmm6
+; AVX1-NEXT:    vpmulhuw %xmm4, %xmm6, %xmm6
+; AVX1-NEXT:    vpaddw %xmm6, %xmm2, %xmm2
+; AVX1-NEXT:    vpmulhw %xmm3, %xmm2, %xmm6
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm6 = xmm2[0],xmm6[1,2,3,4,5,6,7]
+; AVX1-NEXT:    vpsraw $1, %xmm2, %xmm2
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm2 = xmm6[0,1],xmm2[2],xmm6[3,4,5,6],xmm2[7]
+; AVX1-NEXT:    vpsraw $15, %xmm1, %xmm6
+; AVX1-NEXT:    vpmulhuw %xmm4, %xmm6, %xmm4
+; AVX1-NEXT:    vpaddw %xmm4, %xmm1, %xmm4
+; AVX1-NEXT:    vpmulhw %xmm3, %xmm4, %xmm3
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm3 = xmm4[0],xmm3[1,2,3,4,5,6,7]
+; AVX1-NEXT:    vpsraw $1, %xmm4, %xmm4
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm4[2],xmm3[3,4,5,6],xmm4[7]
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
+; AVX1-NEXT:    vandps %ymm5, %ymm2, %ymm2
+; AVX1-NEXT:    vandnps %ymm1, %ymm5, %ymm1
 ; AVX1-NEXT:    vorps %ymm1, %ymm2, %ymm1
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: combine_vec_sdiv_by_pow2b_v32i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [0,2,1,4,3,5,6,1,0,2,1,4,3,5,6,1]
-; AVX2-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
-; AVX2-NEXT:    vpsraw $15, %ymm0, %ymm5
-; AVX2-NEXT:    vbroadcasti128 {{.*#+}} ymm6 = [1,4,2,16,8,32,64,2,1,4,2,16,8,32,64,2]
-; AVX2-NEXT:    # ymm6 = mem[0,1,0,1]
-; AVX2-NEXT:    vpmulhuw %ymm6, %ymm5, %ymm5
-; AVX2-NEXT:    vpaddw %ymm5, %ymm0, %ymm5
-; AVX2-NEXT:    vpunpckhwd {{.*#+}} ymm7 = ymm2[4],ymm5[4],ymm2[5],ymm5[5],ymm2[6],ymm5[6],ymm2[7],ymm5[7],ymm2[12],ymm5[12],ymm2[13],ymm5[13],ymm2[14],ymm5[14],ymm2[15],ymm5[15]
-; AVX2-NEXT:    vpsravd %ymm4, %ymm7, %ymm7
-; AVX2-NEXT:    vpsrld $16, %ymm7, %ymm7
-; AVX2-NEXT:    vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX2-NEXT:    vpunpcklwd {{.*#+}} ymm5 = ymm2[0],ymm5[0],ymm2[1],ymm5[1],ymm2[2],ymm5[2],ymm2[3],ymm5[3],ymm2[8],ymm5[8],ymm2[9],ymm5[9],ymm2[10],ymm5[10],ymm2[11],ymm5[11]
-; AVX2-NEXT:    vpsravd %ymm3, %ymm5, %ymm5
-; AVX2-NEXT:    vpsrld $16, %ymm5, %ymm5
-; AVX2-NEXT:    vpackusdw %ymm7, %ymm5, %ymm5
-; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm5[1,2,3,4,5,6,7],ymm0[8],ymm5[9,10,11,12,13,14,15]
-; AVX2-NEXT:    vpsraw $15, %ymm1, %ymm5
-; AVX2-NEXT:    vpmulhuw %ymm6, %ymm5, %ymm5
-; AVX2-NEXT:    vpaddw %ymm5, %ymm1, %ymm5
-; AVX2-NEXT:    vpunpckhwd {{.*#+}} ymm6 = ymm2[4],ymm5[4],ymm2[5],ymm5[5],ymm2[6],ymm5[6],ymm2[7],ymm5[7],ymm2[12],ymm5[12],ymm2[13],ymm5[13],ymm2[14],ymm5[14],ymm2[15],ymm5[15]
-; AVX2-NEXT:    vpsravd %ymm4, %ymm6, %ymm4
-; AVX2-NEXT:    vpsrld $16, %ymm4, %ymm4
-; AVX2-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm5[0],ymm2[1],ymm5[1],ymm2[2],ymm5[2],ymm2[3],ymm5[3],ymm2[8],ymm5[8],ymm2[9],ymm5[9],ymm2[10],ymm5[10],ymm2[11],ymm5[11]
-; AVX2-NEXT:    vpsravd %ymm3, %ymm2, %ymm2
-; AVX2-NEXT:    vpsrld $16, %ymm2, %ymm2
-; AVX2-NEXT:    vpackusdw %ymm4, %ymm2, %ymm2
+; AVX2-NEXT:    vpsraw $15, %ymm0, %ymm2
+; AVX2-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [1,4,2,16,8,32,64,2,1,4,2,16,8,32,64,2]
+; AVX2-NEXT:    # ymm3 = mem[0,1,0,1]
+; AVX2-NEXT:    vpmulhuw %ymm3, %ymm2, %ymm2
+; AVX2-NEXT:    vpaddw %ymm2, %ymm0, %ymm2
+; AVX2-NEXT:    vbroadcasti128 {{.*#+}} ymm4 = [0,16384,32768,4096,8192,2048,1024,32768,0,16384,32768,4096,8192,2048,1024,32768]
+; AVX2-NEXT:    # ymm4 = mem[0,1,0,1]
+; AVX2-NEXT:    vpmulhw %ymm4, %ymm2, %ymm5
+; AVX2-NEXT:    vpsraw $1, %ymm2, %ymm2
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2],ymm5[3,4,5,6],ymm2[7],ymm5[8,9],ymm2[10],ymm5[11,12,13,14],ymm2[15]
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1,2,3,4,5,6,7],ymm0[8],ymm2[9,10,11,12,13,14,15]
+; AVX2-NEXT:    vpsraw $15, %ymm1, %ymm2
+; AVX2-NEXT:    vpmulhuw %ymm3, %ymm2, %ymm2
+; AVX2-NEXT:    vpaddw %ymm2, %ymm1, %ymm2
+; AVX2-NEXT:    vpmulhw %ymm4, %ymm2, %ymm3
+; AVX2-NEXT:    vpsraw $1, %ymm2, %ymm2
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4,5,6],ymm2[7],ymm3[8,9],ymm2[10],ymm3[11,12,13,14],ymm2[15]
 ; AVX2-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3,4,5,6,7],ymm1[8],ymm2[9,10,11,12,13,14,15]
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: combine_vec_sdiv_by_pow2b_v32i16:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vpsraw $15, %ymm0, %ymm2
 ; AVX512F-NEXT:    vbroadcasti128 {{.*#+}} ymm3 = [1,4,2,16,8,32,64,2,1,4,2,16,8,32,64,2]
 ; AVX512F-NEXT:    # ymm3 = mem[0,1,0,1]

 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: combine_vec_sdiv_nonuniform5:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [65535,0,65535,0,0,0,1,1]
 ; SSE41-NEXT:    pmullw %xmm0, %xmm1
 ; SSE41-NEXT:    pmulhw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    paddw %xmm1, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm1
-; SSE41-NEXT:    psraw $8, %xmm1
-; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm1[0],xmm0[1,2,3,4,5,6],xmm1[7]
-; SSE41-NEXT:    movdqa %xmm1, %xmm2
-; SSE41-NEXT:    psraw $4, %xmm2
-; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm1[0,1],xmm2[2],xmm1[3,4,5],xmm2[6],xmm1[7]
-; SSE41-NEXT:    movdqa %xmm2, %xmm1
-; SSE41-NEXT:    psraw $2, %xmm1
-; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1],xmm2[2,3,4,5],xmm1[6],xmm2[7]
-; SSE41-NEXT:    movdqa %xmm1, %xmm2
+; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = <256,16384,4096,u,u,32768,512,256>
+; SSE41-NEXT:    pmulhw %xmm0, %xmm1
+; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm0[3,4],xmm1[5,6,7]
+; SSE41-NEXT:    movdqa %xmm0, %xmm2
 ; SSE41-NEXT:    psraw $1, %xmm2
-; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm1[0,1,2,3,4],xmm2[5,6],xmm1[7]
+; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm1[0,1,2,3,4],xmm2[5],xmm1[6,7]
 ; SSE41-NEXT:    psrlw $15, %xmm0
 ; SSE41-NEXT:    paddw %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: combine_vec_sdiv_nonuniform5:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpmullw {{.*}}(%rip), %xmm0, %xmm1
 ; AVX1-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpsraw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm0[1,2,3,4,5,6],xmm1[7]
-; AVX1-NEXT:    vpsraw $4, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3,4,5],xmm2[6],xmm1[7]
-; AVX1-NEXT:    vpsraw $2, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3,4,5],xmm2[6],xmm1[7]
-; AVX1-NEXT:    vpsraw $1, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4],xmm2[5,6],xmm1[7]
+; AVX1-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm1
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm0[3,4],xmm1[5,6,7]
+; AVX1-NEXT:    vpsraw $1, %xmm0, %xmm2
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4],xmm2[5],xmm1[6,7]
 ; AVX1-NEXT:    vpsrlw $15, %xmm0, %xmm0
 ; AVX1-NEXT:    vpaddw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: combine_vec_sdiv_nonuniform5:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpmullw {{.*}}(%rip), %xmm0, %xmm1
 ; AVX2-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpmovsxwd %xmm0, %ymm1
-; AVX2-NEXT:    vpsravd {{.*}}(%rip), %ymm1, %ymm1
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
+; AVX2-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm1
+; AVX2-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm0[3,4],xmm1[5,6,7]
+; AVX2-NEXT:    vpsraw $1, %xmm0, %xmm2
+; AVX2-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4],xmm2[5],xmm1[6,7]
 ; AVX2-NEXT:    vpsrlw $15, %xmm0, %xmm0
 ; AVX2-NEXT:    vpaddw %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: combine_vec_sdiv_nonuniform5:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vpmullw {{.*}}(%rip), %xmm0, %xmm1
 ; AVX512F-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX512F-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
 ; AVX512F-NEXT:    vpsrlw $15, %xmm0, %xmm1

 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: combine_vec_sdiv_nonuniform6:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535,1,1,1,0]
 ; SSE41-NEXT:    pmullw %xmm0, %xmm1
 ; SSE41-NEXT:    pmulhw {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    paddw %xmm1, %xmm0
-; SSE41-NEXT:    movdqa %xmm0, %xmm1
-; SSE41-NEXT:    psraw $8, %xmm1
-; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm0[3,4,5],xmm1[6,7]
-; SSE41-NEXT:    movdqa %xmm1, %xmm2
-; SSE41-NEXT:    psraw $4, %xmm2
-; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm2[0],xmm1[1,2,3,4],xmm2[5],xmm1[6],xmm2[7]
-; SSE41-NEXT:    movdqa %xmm2, %xmm3
-; SSE41-NEXT:    psraw $2, %xmm3
-; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm3[0],xmm2[1,2,3,4],xmm3[5],xmm2[6,7]
-; SSE41-NEXT:    movdqa %xmm3, %xmm1
-; SSE41-NEXT:    psraw $1, %xmm1
-; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm3[0,1,2,3,4],xmm1[5],xmm3[6],xmm1[7]
+; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = <4,256,256,u,u,512,256,8>
+; SSE41-NEXT:    pmulhw %xmm0, %xmm2
+; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm2[0,1,2],xmm0[3,4],xmm2[5,6,7]
 ; SSE41-NEXT:    psrlw $15, %xmm0
-; SSE41-NEXT:    pxor %xmm2, %xmm2
-; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm0[0,1,2],xmm2[3,4],xmm0[5,6,7]
+; SSE41-NEXT:    pxor %xmm1, %xmm1
+; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0,1,2],xmm1[3,4],xmm0[5,6,7]
 ; SSE41-NEXT:    paddw %xmm2, %xmm1
 ; SSE41-NEXT:    movdqa %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: combine_vec_sdiv_nonuniform6:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpmullw {{.*}}(%rip), %xmm0, %xmm1
 ; AVX1-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpsraw $8, %xmm0, %xmm1
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm0[3,4,5],xmm1[6,7]
-; AVX1-NEXT:    vpsraw $4, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1,2,3,4],xmm2[5],xmm1[6],xmm2[7]
-; AVX1-NEXT:    vpsraw $2, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1,2,3,4],xmm2[5],xmm1[6,7]
-; AVX1-NEXT:    vpsraw $1, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4],xmm2[5],xmm1[6],xmm2[7]
+; AVX1-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm1
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm0[3,4],xmm1[5,6,7]
 ; AVX1-NEXT:    vpsrlw $15, %xmm0, %xmm0
 ; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm2[3,4],xmm0[5,6,7]
 ; AVX1-NEXT:    vpaddw %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: combine_vec_sdiv_nonuniform6:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpmullw {{.*}}(%rip), %xmm0, %xmm1
 ; AVX2-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpmovsxwd %xmm0, %ymm1
-; AVX2-NEXT:    vpsravd {{.*}}(%rip), %ymm1, %ymm1
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
+; AVX2-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm1
+; AVX2-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2],xmm0[3,4],xmm1[5,6,7]
 ; AVX2-NEXT:    vpsrlw $15, %xmm0, %xmm0
 ; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
 ; AVX2-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2],xmm2[3,4],xmm0[5,6,7]
 ; AVX2-NEXT:    vpaddw %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: combine_vec_sdiv_nonuniform6:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vpmullw {{.*}}(%rip), %xmm0, %xmm1
 ; AVX512F-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm0
 ; AVX512F-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
 ; AVX512F-NEXT:    vpsrlw $15, %xmm0, %xmm1

llvm/trunk/test/CodeGen/X86/vector-shift-ashr-128.ll

 ; SSE2-NEXT:    andps %xmm1, %xmm0
 ; SSE2-NEXT:    psraw $1, %xmm2
 ; SSE2-NEXT:    andnps %xmm2, %xmm1
 ; SSE2-NEXT:    orps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: constant_shift_v8i16:
 ; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa %xmm0, %xmm1
-; SSE41-NEXT:    psraw $4, %xmm1
-; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0,1,2,3],xmm1[4,5,6,7]
-; SSE41-NEXT:    movdqa %xmm1, %xmm2
-; SSE41-NEXT:    psraw $2, %xmm2
-; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm1[0,1],xmm2[2,3],xmm1[4,5],xmm2[6,7]
-; SSE41-NEXT:    movdqa %xmm2, %xmm0
+; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = <u,32768,16384,8192,4096,2048,1024,512>
+; SSE41-NEXT:    pmulhw %xmm0, %xmm1
+; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3,4,5,6,7]
 ; SSE41-NEXT:    psraw $1, %xmm0
-; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm2[0],xmm0[1],xmm2[2],xmm0[3],xmm2[4],xmm0[5],xmm2[6],xmm0[7]
+; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3,4,5,6,7]
 ; SSE41-NEXT:    retq
 ;
-; AVX1-LABEL: constant_shift_v8i16:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vpsraw $4, %xmm0, %xmm1
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7]
-; AVX1-NEXT:    vpsraw $2, %xmm0, %xmm1
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
-; AVX1-NEXT:    vpsraw $1, %xmm0, %xmm1
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7]
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: constant_shift_v8i16:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpmovsxwd %xmm0, %ymm0
-; AVX2-NEXT:    vpsravd {{.*}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT:    vpackssdw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vzeroupper
-; AVX2-NEXT:    retq
+; AVX-LABEL: constant_shift_v8i16:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm1
+; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3,4,5,6,7]
+; AVX-NEXT:    vpsraw $1, %xmm0, %xmm0
+; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3,4,5,6,7]
+; AVX-NEXT:    retq
 ;
 ; XOP-LABEL: constant_shift_v8i16:
 ; XOP:       # %bb.0:
 ; XOP-NEXT:    vpshaw {{.*}}(%rip), %xmm0, %xmm0
 ; XOP-NEXT:    retq
 ;
 ; AVX512DQ-LABEL: constant_shift_v8i16:
 ; AVX512DQ:       # %bb.0:

llvm/trunk/test/CodeGen/X86/vector-shift-ashr-256.ll

 ; X32-AVX2-NEXT:    retl
   %shift = ashr <8 x i32> %a, <i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 8, i32 7>
   ret <8 x i32> %shift
 }
 
 define <16 x i16> @constant_shift_v16i16(<16 x i16> %a) nounwind {
 ; AVX1-LABEL: constant_shift_v16i16:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT:    vpsraw $8, %xmm1, %xmm1
-; AVX1-NEXT:    vpsraw $4, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3],xmm2[4,5,6,7]
-; AVX1-NEXT:    vpsraw $2, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3],xmm1[4,5],xmm2[6,7]
-; AVX1-NEXT:    vpsraw $1, %xmm1, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2],xmm2[3],xmm1[4],xmm2[5],xmm1[6],xmm2[7]
-; AVX1-NEXT:    vpsraw $4, %xmm0, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4,5,6,7]
-; AVX1-NEXT:    vpsraw $2, %xmm0, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
+; AVX1-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm1
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3,4,5,6,7]
 ; AVX1-NEXT:    vpsraw $1, %xmm0, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3],xmm0[4],xmm2[5],xmm0[6],xmm2[7]
-; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3,4,5,6,7]
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vpmulhw {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: constant_shift_v16i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15]
-; AVX2-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
-; AVX2-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX2-NEXT:    vpsravd %ymm3, %ymm4, %ymm3
-; AVX2-NEXT:    vpsrld $16, %ymm3, %ymm3
-; AVX2-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; AVX2-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX2-NEXT:    vpsravd %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpsrld $16, %ymm0, %ymm0
-; AVX2-NEXT:    vpackusdw %ymm3, %ymm0, %ymm0
+; AVX2-NEXT:    vpmulhw {{.*}}(%rip), %ymm0, %ymm1
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm2 = ymm0[0],ymm1[1,2,3,4,5,6,7],ymm0[8],ymm1[9,10,11,12,13,14,15]
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm1[4,5,6,7]
+; AVX2-NEXT:    vpsraw $1, %ymm0, %ymm0
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3,4,5,6,7,8],ymm0[9],ymm2[10,11,12,13,14,15]
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
 ; AVX2-NEXT:    retq
 ;
 ; XOPAVX1-LABEL: constant_shift_v16i16:
 ; XOPAVX1:       # %bb.0:
 ; XOPAVX1-NEXT:    vpshaw {{.*}}(%rip), %xmm0, %xmm1
 ; XOPAVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; XOPAVX1-NEXT:    vpshaw {{.*}}(%rip), %xmm0, %xmm0
 ; XOPAVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0

 ;
 ; AVX512BWVL-LABEL: constant_shift_v16i16:
 ; AVX512BWVL:       # %bb.0:
 ; AVX512BWVL-NEXT:    vpsravw {{.*}}(%rip), %ymm0, %ymm0
 ; AVX512BWVL-NEXT:    retq
 ;
 ; X32-AVX1-LABEL: constant_shift_v16i16:
 ; X32-AVX1:       # %bb.0:
-; X32-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; X32-AVX1-NEXT:    vpsraw $8, %xmm1, %xmm1
-; X32-AVX1-NEXT:    vpsraw $4, %xmm1, %xmm2
-; X32-AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3],xmm2[4,5,6,7]
-; X32-AVX1-NEXT:    vpsraw $2, %xmm1, %xmm2
-; X32-AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3],xmm1[4,5],xmm2[6,7]
-; X32-AVX1-NEXT:    vpsraw $1, %xmm1, %xmm2
-; X32-AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2],xmm2[3],xmm1[4],xmm2[5],xmm1[6],xmm2[7]
-; X32-AVX1-NEXT:    vpsraw $4, %xmm0, %xmm2
-; X32-AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4,5,6,7]
-; X32-AVX1-NEXT:    vpsraw $2, %xmm0, %xmm2
-; X32-AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
+; X32-AVX1-NEXT:    vpmulhw {{\.LCPI.*}}, %xmm0, %xmm1
+; X32-AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3,4,5,6,7]
 ; X32-AVX1-NEXT:    vpsraw $1, %xmm0, %xmm2
-; X32-AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3],xmm0[4],xmm2[5],xmm0[6],xmm2[7]
-; X32-AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; X32-AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3,4,5,6,7]
+; X32-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; X32-AVX1-NEXT:    vpmulhw {{\.LCPI.*}}, %xmm0, %xmm0
+; X32-AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
 ; X32-AVX1-NEXT:    retl
 ;
 ; X32-AVX2-LABEL: constant_shift_v16i16:
 ; X32-AVX2:       # %bb.0:
-; X32-AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; X32-AVX2-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15]
-; X32-AVX2-NEXT:    vpunpckhwd {{.*#+}} ymm3 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
-; X32-AVX2-NEXT:    vpunpckhwd {{.*#+}} ymm4 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; X32-AVX2-NEXT:    vpsravd %ymm3, %ymm4, %ymm3
-; X32-AVX2-NEXT:    vpsrld $16, %ymm3, %ymm3
-; X32-AVX2-NEXT:    vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; X32-AVX2-NEXT:    vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; X32-AVX2-NEXT:    vpsravd %ymm2, %ymm0, %ymm0
-; X32-AVX2-NEXT:    vpsrld $16, %ymm0, %ymm0
-; X32-AVX2-NEXT:    vpackusdw %ymm3, %ymm0, %ymm0
+; X32-AVX2-NEXT:    vpmulhw {{\.LCPI.*}}, %ymm0, %ymm1
+; X32-AVX2-NEXT:    vpblendw {{.*#+}} ymm2 = ymm0[0],ymm1[1,2,3,4,5,6,7],ymm0[8],ymm1[9,10,11,12,13,14,15]
+; X32-AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm1[4,5,6,7]
+; X32-AVX2-NEXT:    vpsraw $1, %ymm0, %ymm0
+; X32-AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3,4,5,6,7,8],ymm0[9],ymm2[10,11,12,13,14,15]
+; X32-AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
 ; X32-AVX2-NEXT:    retl
   %shift = ashr <16 x i16> %a, <i16 0, i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
   ret <16 x i16> %shift
 }
 
 define <32 x i8> @constant_shift_v32i8(<32 x i8> %a) nounwind {
 ; AVX1-LABEL: constant_shift_v32i8:
 ; AVX1:       # %bb.0: