In r131488 I misunderstood how VREV works. It splits the vector in half and splits each half. Therefore, the real problem was that we were using a VREV64 for a 4xi16, when we should have been using a VREV32.
Updated test case and reverted change to the PerfectShuffle Table.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@131529 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 18d3008..67cb293 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -4182,7 +4182,15 @@
   switch (OpNum) {
   default: llvm_unreachable("Unknown shuffle opcode!");
   case OP_VREV:
-    return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
+    // VREV divides the vector in half and swaps within the half.
+    if (VT.getVectorElementType() == MVT::i32)
+      return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
+    // vrev <4 x i16> -> VREV32
+    if (VT.getVectorElementType() == MVT::i16)
+      return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
+    // vrev <4 x i8> -> VREV16
+    assert(VT.getVectorElementType() == MVT::i8);
+    return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
   case OP_VDUP0:
   case OP_VDUP1:
   case OP_VDUP2: