diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -13076,6 +13076,33 @@
   return ScaledMask;
 }
 
+// Helper to attempt to widen/narrow blend masks.
+static bool adjustBlendMask(uint64_t OldMask, uint64_t OldWidth,
+                            uint64_t NewWidth, uint64_t *pNewMask = nullptr) {
+  assert(((OldWidth % NewWidth) == 0 || (NewWidth % OldWidth) == 0) &&
+         "Illegal blend mask scale");
+  uint64_t NewMask = 0;
+
+  if ((OldWidth % NewWidth) == 0) {
+    uint64_t Scale = OldWidth / NewWidth;
+    uint64_t SubMask = (1u << Scale) - 1;
+    for (uint64_t i = 0; i != NewWidth; ++i) {
+      uint64_t Sub = (OldMask >> (i * Scale)) & SubMask;
+      if (Sub == SubMask)
+        NewMask |= (1u << i);
+      else if (Sub != 0x0)
+        return false;
+    }
+  } else {
+      NewMask = scaleVectorShuffleBlendMask(OldMask, OldWidth, NewWidth / OldWidth);
+  }
+
+  if (pNewMask)
+    *pNewMask = NewMask;
+  return true;
+}
+
+
 /// Try to emit a blend instruction for a shuffle.
 ///
 /// This doesn't do any checks for the availability of instructions for blending
@@ -13102,6 +13129,35 @@
 
   unsigned NumElts = VT.getVectorNumElements();
 
+  auto Geti16Blend = [&](MVT TmpVT, unsigned NewWidth) {
+    uint64_t NewBlendMask = 0;
+    if (adjustBlendMask(BlendMask, NewWidth * 2, NewWidth, &NewBlendMask)) {
+      return DAG.getBitcast(
+          VT, DAG.getNode(X86ISD::BLENDI, DL, TmpVT, DAG.getBitcast(TmpVT, V1),
+                          DAG.getBitcast(TmpVT, V2),
+                          DAG.getTargetConstant(NewBlendMask, DL, MVT::i8)));
+    }
+    return SDValue();
+  };
+
+  // blendw is slower than blendps/blendd on many targets. If we can scale
+  // blendw -> blendps do so. NB: We go for blendps because anytime blendw is
+  // available for a given width, blendps is also available.
+  // X86ExecutionDomainFix will convert back to blendw (or to blendd) if its
+  // needed.
+  switch (VT.SimpleTy) {
+  case MVT::v8i16:
+    if (SDValue R = Geti16Blend(MVT::v4f32, 4))
+      return R;
+    break;
+  case MVT::v16i16:
+    if (SDValue R = Geti16Blend(MVT::v8f32, 8))
+      return R;
+    break;
+  default:
+    break;
+  }
+
   switch (VT.SimpleTy) {
   case MVT::v4i64:
   case MVT::v8i32:
diff --git a/llvm/test/CodeGen/X86/combine-sdiv.ll b/llvm/test/CodeGen/X86/combine-sdiv.ll
--- a/llvm/test/CodeGen/X86/combine-sdiv.ll
+++ b/llvm/test/CodeGen/X86/combine-sdiv.ll
@@ -2889,26 +2889,12 @@
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5,6,7]
 ; AVX1-NEXT:    retq
 ;
-; AVX2-LABEL: combine_vec_sdiv_nonuniform7:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpsubw %xmm0, %xmm1, %xmm1
-; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
-; AVX2-NEXT:    retq
-;
-; AVX512F-LABEL: combine_vec_sdiv_nonuniform7:
-; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512F-NEXT:    vpsubw %xmm0, %xmm1, %xmm1
-; AVX512F-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
-; AVX512F-NEXT:    retq
-;
-; AVX512BW-LABEL: combine_vec_sdiv_nonuniform7:
-; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpsubw %xmm0, %xmm1, %xmm1
-; AVX512BW-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5,6,7]
-; AVX512BW-NEXT:    retq
+; AVX2ORLATER-LABEL: combine_vec_sdiv_nonuniform7:
+; AVX2ORLATER:       # %bb.0:
+; AVX2ORLATER-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX2ORLATER-NEXT:    vpsubw %xmm0, %xmm1, %xmm1
+; AVX2ORLATER-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
+; AVX2ORLATER-NEXT:    retq
 ;
 ; XOP-LABEL: combine_vec_sdiv_nonuniform7:
 ; XOP:       # %bb.0:
diff --git a/llvm/test/CodeGen/X86/shuffle-blendw.ll b/llvm/test/CodeGen/X86/shuffle-blendw.ll
--- a/llvm/test/CodeGen/X86/shuffle-blendw.ll
+++ b/llvm/test/CodeGen/X86/shuffle-blendw.ll
@@ -66,13 +66,13 @@
 ; X86-AVX512-LABEL: blendw_to_blendd_32:
 ; X86-AVX512:       # %bb.0:
 ; X86-AVX512-NEXT:    vpaddw %ymm2, %ymm0, %ymm0
-; X86-AVX512-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7],ymm0[8,9],ymm1[10,11],ymm0[12,13],ymm1[14,15]
+; X86-AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
 ; X86-AVX512-NEXT:    retl
 ;
 ; X64-AVX512-LABEL: blendw_to_blendd_32:
 ; X64-AVX512:       # %bb.0:
 ; X64-AVX512-NEXT:    vpaddw %ymm2, %ymm0, %ymm0
-; X64-AVX512-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7],ymm0[8,9],ymm1[10,11],ymm0[12,13],ymm1[14,15]
+; X64-AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
 ; X64-AVX512-NEXT:    retq
   %x1 = add <16 x i16> %x, %z
   %shuffle = shufflevector <16 x i16> %x1, <16 x i16> %y, <16 x i32> <i32 0, i32 1, i32 18, i32 19, i32 4, i32 5, i32 22, i32 23, i32 8, i32 9, i32 26, i32 27, i32 12, i32 13, i32 30, i32 31>
@@ -119,13 +119,13 @@
 ; X86-AVX512-LABEL: blendw_to_blendd_16:
 ; X86-AVX512:       # %bb.0:
 ; X86-AVX512-NEXT:    vpaddw %xmm2, %xmm0, %xmm0
-; X86-AVX512-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7]
+; X86-AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
 ; X86-AVX512-NEXT:    retl
 ;
 ; X64-AVX512-LABEL: blendw_to_blendd_16:
 ; X64-AVX512:       # %bb.0:
 ; X64-AVX512-NEXT:    vpaddw %xmm2, %xmm0, %xmm0
-; X64-AVX512-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7]
+; X64-AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
 ; X64-AVX512-NEXT:    retq
   %x1 = add <8 x i16> %x, %z
   %shuffle = shufflevector <8 x i16> %x1, <8 x i16> %y, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 12, i32 13, i32 14, i32 15>