Index: llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -7530,6 +7530,16 @@
       return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
   }
 
+  // fold (sext_in_reg (*_extend_vector_inreg x)) -> (sext_vector_in_reg x)
+  if ((N0.getOpcode() == ISD::ANY_EXTEND_VECTOR_INREG ||
+       N0.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG ||
+       N0.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG) &&
+      N0.getOperand(0).getScalarValueSizeInBits() == EVTBits) {
+    if (!LegalOperations ||
+        TLI.isOperationLegal(ISD::SIGN_EXTEND_VECTOR_INREG, VT))
+      return DAG.getSignExtendVectorInReg(N0.getOperand(0), SDLoc(N), VT);
+  }
+
   // fold (sext_in_reg (zext x)) -> (sext x)
   // iff we are extending the source sign bit.
   if (N0.getOpcode() == ISD::ZERO_EXTEND) {
@@ -14194,6 +14204,113 @@
   return DAG.getBuildVector(VT, SDLoc(SVN), Ops);
 }
 
+// Match shuffles that can be converted to any_vector_extend_in_reg.
+// This is often generated during legalization.
+// e.g. v4i32 <0,u,1,u> -> (v2i64 any_vector_extend_in_reg(v4i32 src))
+// TODO Add support for ZERO_EXTEND_VECTOR_INREG when we have a test case.
+SDValue combineShuffleToVectorExtend(ShuffleVectorSDNode *SVN,
+                                     SelectionDAG &DAG,
+                                     const TargetLowering &TLI,
+                                     bool LegalOperations) {
+  EVT VT = SVN->getValueType(0);
+  bool IsBigEndian = DAG.getDataLayout().isBigEndian();
+
+  // TODO Add support for big-endian when we have a test case.
+  if (!VT.isInteger() || IsBigEndian)
+    return SDValue();
+
+  unsigned NumElts = VT.getVectorNumElements();
+  unsigned EltSizeInBits = VT.getScalarSizeInBits();
+  ArrayRef<int> Mask = SVN->getMask();
+  SDValue N0 = SVN->getOperand(0);
+
+  // shuffle<0,-1,1,-1> == (v2i64 anyextend_vector_inreg(v4i32))
+  auto isAnyExtend = [&Mask, &NumElts](unsigned Scale) {
+    for (unsigned i = 0; i != NumElts; ++i) {
+      if (Mask[i] < 0)
+        continue;
+      if ((i % Scale) == 0 && Mask[i] == (i / Scale))
+        continue;
+      return false;
+    }
+    return true;
+  };
+
+  // Attempt to match a '*_extend_vector_inreg' shuffle, we just search for
+  // power-of-2 extensions as they are the most likely.
+  for (unsigned Scale = 2; Scale < NumElts; Scale *= 2) {
+    if (!isAnyExtend(Scale))
+      continue;
+
+    EVT OutSVT = EVT::getIntegerVT(*DAG.getContext(), EltSizeInBits * Scale);
+    EVT OutVT = EVT::getVectorVT(*DAG.getContext(), OutSVT, NumElts / Scale);
+    if (!LegalOperations ||
+        TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND_VECTOR_INREG, OutVT))
+      return DAG.getBitcast(VT,
+                            DAG.getAnyExtendVectorInReg(N0, SDLoc(SVN), OutVT));
+  }
+
+  return SDValue();
+}
+
+// Detect 'truncate_vector_inreg' style shuffles that pack the lower parts of
+// each source element of a large type into the lowest elements of a smaller
+// destination type. This is often generated during legalization.
+// If the source node itself was a '*_extend_vector_inreg' node then we should
+// then be able to remove it.
+SDValue combineTruncationShuffle(ShuffleVectorSDNode *SVN, SelectionDAG &DAG) {
+  EVT VT = SVN->getValueType(0);
+  bool IsBigEndian = DAG.getDataLayout().isBigEndian();
+
+  // TODO Add support for big-endian when we have a test case.
+  if (!VT.isInteger() || IsBigEndian)
+    return SDValue();
+
+  SDValue N0 = SVN->getOperand(0);
+  while (N0.getOpcode() == ISD::BITCAST)
+    N0 = N0.getOperand(0);
+
+  unsigned Opcode = N0.getOpcode();
+  if (Opcode != ISD::ANY_EXTEND_VECTOR_INREG &&
+      Opcode != ISD::SIGN_EXTEND_VECTOR_INREG &&
+      Opcode != ISD::ZERO_EXTEND_VECTOR_INREG)
+    return SDValue();
+
+  SDValue N00 = N0.getOperand(0);
+  ArrayRef<int> Mask = SVN->getMask();
+  unsigned NumElts = VT.getVectorNumElements();
+  unsigned EltSizeInBits = VT.getScalarSizeInBits();
+  unsigned ExtSrcSizeInBits = N00.getScalarValueSizeInBits();
+
+  // (v4i32 truncate_vector_inreg(v2i64)) == shuffle<0,2-1,-1>
+  // (v8i16 truncate_vector_inreg(v4i32)) == shuffle<0,2,4,6,-1,-1,-1,-1>
+  // (v8i16 truncate_vector_inreg(v2i64)) == shuffle<0,4,-1,-1,-1,-1,-1,-1>
+  auto isTruncate = [&Mask, &NumElts](unsigned Scale) {
+    for (unsigned i = 0; i != NumElts; ++i) {
+      if (Mask[i] < 0)
+        continue;
+      if ((i * Scale) < NumElts && Mask[i] == (i * Scale))
+        continue;
+      return false;
+    }
+    return true;
+  };
+
+  // At the moment we just handle the case where we've truncated back to the
+  // same size as before the extension.
+  // TODO: handle more extension/truncation cases as cases arise.
+  if (EltSizeInBits != ExtSrcSizeInBits)
+    return SDValue();
+
+  // Attempt to match a 'truncate_vector_inreg' shuffle, we just search for
+  // power-of-2 truncations as they are the most likely.
+  for (unsigned Scale = 2; Scale < NumElts; Scale *= 2)
+    if (isTruncate(Scale))
+      return DAG.getBitcast(VT, N00);
+
+  return SDValue();
+}
+
 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
   EVT VT = N->getValueType(0);
   unsigned NumElts = VT.getVectorNumElements();
@@ -14298,6 +14415,14 @@
   if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
     return S;
 
+  // Match shuffles that can be converted to any_vector_extend_in_reg.
+  if (SDValue V = combineShuffleToVectorExtend(SVN, DAG, TLI, LegalOperations))
+    return V;
+
+  // Combine "truncate_vector_in_reg" style shuffles.
+  if (SDValue V = combineTruncationShuffle(SVN, DAG))
+    return V;
+
   if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
       Level < AfterLegalizeVectorOps &&
       (N1.isUndef() ||
Index: llvm/trunk/test/CodeGen/X86/2011-10-21-widen-cmp.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/2011-10-21-widen-cmp.ll
+++ llvm/trunk/test/CodeGen/X86/2011-10-21-widen-cmp.ll
@@ -9,7 +9,8 @@
 ; CHECK:       # BB#0: # %entry
 ; CHECK-NEXT:    movaps %xmm0, %xmm2
 ; CHECK-NEXT:    cmpordps %xmm0, %xmm0
-; CHECK-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1],zero,xmm0[1]
+; CHECK-NEXT:    pmovsxdq %xmm0, %xmm0
+; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
 ; CHECK-NEXT:    pslld $31, %xmm0
 ; CHECK-NEXT:    blendvps %xmm0, %xmm2, %xmm1
 ; CHECK-NEXT:    movlps %xmm1, (%rax)
Index: llvm/trunk/test/CodeGen/X86/avx512-cvt.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/avx512-cvt.ll
+++ llvm/trunk/test/CodeGen/X86/avx512-cvt.ll
@@ -1075,7 +1075,6 @@
 ; NOVL:       ## BB#0:
 ; NOVL-NEXT:    vxorps %xmm1, %xmm1, %xmm1
 ; NOVL-NEXT:    vcmpltps %xmm0, %xmm1, %xmm0
-; NOVL-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1],zero,xmm0[1]
 ; NOVL-NEXT:    vcvtdq2ps %xmm0, %xmm0
 ; NOVL-NEXT:    retq
 ;
Index: llvm/trunk/test/CodeGen/X86/avx512-vec-cmp.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/avx512-vec-cmp.ll
+++ llvm/trunk/test/CodeGen/X86/avx512-vec-cmp.ll
@@ -1226,11 +1226,7 @@
 ; KNL-LABEL: test46:
 ; KNL:       ## BB#0:
 ; KNL-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm0
-; KNL-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; KNL-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; KNL-NEXT:    vpsrad $31, %xmm0, %xmm1
-; KNL-NEXT:    vmovshdup {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; KNL-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3]
+; KNL-NEXT:    vpmovsxdq %xmm0, %xmm0
 ; KNL-NEXT:    vpand {{.*}}(%rip), %xmm0, %xmm0
 ; KNL-NEXT:    retq
 ;
Index: llvm/trunk/test/CodeGen/X86/sse-minmax.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/sse-minmax.ll
+++ llvm/trunk/test/CodeGen/X86/sse-minmax.ll
@@ -1277,8 +1277,6 @@
 ; STRICT-NEXT:    movaps %xmm0, %xmm2
 ; STRICT-NEXT:    movaps %xmm1, %xmm0
 ; STRICT-NEXT:    cmpleps %xmm2, %xmm0
-; STRICT-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1],zero,xmm0[1]
-; STRICT-NEXT:    pslld $31, %xmm0
 ; STRICT-NEXT:    blendvps %xmm0, %xmm2, %xmm1
 ; STRICT-NEXT:    movaps %xmm1, %xmm0
 ; STRICT-NEXT:    retq
@@ -1297,8 +1295,6 @@
 ; STRICT:       # BB#0:
 ; STRICT-NEXT:    movaps %xmm0, %xmm2
 ; STRICT-NEXT:    cmpleps %xmm1, %xmm0
-; STRICT-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1],zero,xmm0[1]
-; STRICT-NEXT:    pslld $31, %xmm0
 ; STRICT-NEXT:    blendvps %xmm0, %xmm2, %xmm1
 ; STRICT-NEXT:    movaps %xmm1, %xmm0
 ; STRICT-NEXT:    retq