[PATCH] [x86] Teach the target-specific combining how to aggressively foldhalf-shuffles, even looking through intervening instructions in a chain.

[Chandler Carruth]

Generally, will apply fixes before submitting. One note though.

================
Comment at: test/CodeGen/X86/vector-shuffle-combining.ll:11
@@ +10,3 @@
+; CHECK-SSE2-LABEL: @combine_pshuflw1
+; CHECK-SSE2:       # BB#0:
+; CHECK-SSE2-NEXT:    retq
----------------
Jim Grosbach wrote:
> I prefer not to look for these sorts of comments in test lines if at all possible and to just run llc with -asm-verbose=false. That enables explicit CHECK-NEXT for instructions right after the function label to make sure there's not extraneous stuff showing up, which looks like what you're trying to do here.
I understand your hesitance, but with shuffles, I think we have to use comment-based checking. It is essential to write the shuffle instruction check lines using the comments that decode the operands rather than having opaque "$-27" and company in check lines.

http://reviews.llvm.org/D4291