Optimization for the following SIGN_EXTEND pairs:
v8i8  -> v8i64, 
v8i8  -> v8i32, 
v4i8  -> v4i64, 
v4i16 -> v4i64 
for AVX and AVX2.

Bug 14865.

llvm-svn: 172708
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index a82410a..3e5a446 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -4298,11 +4298,19 @@
   if (isa<ConstantSDNode>(N0))
     return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N0);
 
-  // fold (sext (sext x)) -> (sext x)
-  // fold (sext (aext x)) -> (sext x)
-  if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
-    return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
-                       N0.getOperand(0));
+  // Folding (sext (sext x)) is obvious, but we do it only after the type 
+  // legalization phase. When the sequence is like {(T1->T2), (T2->T3)} and 
+  // T1 or T3 (or the both) are illegal types, the TypeLegalizer may not 
+  // give a good sequence for the (T1->T3) pair.
+  // So we give a chance to target specific combiner to optimize T1->T2 and T2->T3
+  // separately and may be fold them in a preceding of subsequent instruction.
+  if (Level >= AfterLegalizeTypes) {
+    // fold (sext (sext x)) -> (sext x)
+    // fold (sext (aext x)) -> (sext x)
+    if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
+      return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
+                         N0.getOperand(0));
+  }
 
   if (N0.getOpcode() == ISD::TRUNCATE) {
     // fold (sext (truncate (load x))) -> (sext (smaller load x))
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 344d144..91491bf 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -2554,9 +2554,7 @@
             VT.getVectorNumElements() ==
             Operand.getValueType().getVectorNumElements()) &&
            "Vector element count mismatch!");
-    if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND)
-      return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0));
-    else if (OpOpcode == ISD::UNDEF)
+    if (OpOpcode == ISD::UNDEF)
       // sext(undef) = 0, because the top bits will all be the same.
       return getConstant(0, VT);
     break;
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index f42884d..a8294b6 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -16970,14 +16970,37 @@
 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
                                   TargetLowering::DAGCombinerInfo &DCI,
                                   const X86Subtarget *Subtarget) {
+  EVT VT = N->getValueType(0);
+  
+  if (!VT.isVector())
+    return SDValue();
+
+  SDValue In = N->getOperand(0);
+  EVT InVT = In.getValueType();
+  DebugLoc dl = N->getDebugLoc();
+  unsigned ExtenedEltSize = VT.getVectorElementType().getSizeInBits(); 
+
+  // Split SIGN_EXTEND operation to use vmovsx instruction when possible
+  if (InVT == MVT::v8i8) {
+    if (ExtenedEltSize > 16 && !Subtarget->hasInt256())
+      In = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, In);
+    if (ExtenedEltSize > 32)
+      In = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i32, In);
+    return DAG.getNode(ISD::SIGN_EXTEND, dl, VT, In);
+  }
+
+  if ((InVT == MVT::v4i8 || InVT == MVT::v4i16) &&
+      ExtenedEltSize > 32 && !Subtarget->hasInt256()) {
+    In = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, In);
+    return DAG.getNode(ISD::SIGN_EXTEND, dl, VT, In);
+  }
   if (!DCI.isBeforeLegalizeOps())
     return SDValue();
 
   if (!Subtarget->hasFp256())
     return SDValue();
 
-  EVT VT = N->getValueType(0);
-  if (VT.isVector() && VT.getSizeInBits() == 256) {
+  if (VT.is256BitVector()) {
     SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
     if (R.getNode())
       return R;