[PATCH] [X86] Try to combine vector anyext + and into a vector zext

[Andrea Di Biagio]

Hi Michael,

Overall the change looks good to me. There are a couple of things that could be improved (see comments below).

In the description, you mentioned that this rule would help simplifying a (type-)legalized dag node sequence obtained from a vector zext. It probably makes sense to guard this new rule against `DCI.isBeforeLegalize()`. If the level is 'BeforeLegalizeTypes' then we probably don't want/need to enable this new rule.

Minor question: what about moving this logic into a separate function? I don't have a strong opinion about it, so feel free to keep the code in 'PerformAndCombine' if you think it makes more sense.


================
Comment at: lib/Target/X86/X86ISelLowering.cpp:24668-24669
@@ +24667,4 @@
+      N0.getOperand(0).getOpcode() == ISD::VECTOR_SHUFFLE &&
+      (N1.getOpcode() == ISD::BUILD_VECTOR ||
+       N1.getOpcode() == ISD::BITCAST)) {
+
----------------
I think you can remove these two checks.
Your algorithm already checks (between line 24682 and 24686) that N1 is a build_vector or a bitcast of a build_vector.

================
Comment at: lib/Target/X86/X86ISelLowering.cpp:24691
@@ +24690,3 @@
+    bool HasAnyUndefs;
+    BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(Splat);
+    if (!Vector->isConstantSplat(SplatValue, SplatUndef,
----------------
You don't need a dyn_cast here. Node Splat can only be an ISD::BUILD_VECTOR at this point.
I think you can safely use a 'cast' here.

http://reviews.llvm.org/D7666

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/