[PATCH] [x86] Fold extract_vector_elt of a load into the load's address computation.

Jim Grosbach grosbach at apple.com
Tue Apr 15 18:03:14 PDT 2014


  This looks like a really nice improvement, and you're right, probably not just for x86.

  I suggest looking into making this a target-independent DAG combine that can be predicated on queries to TargetLowering about what types and operations are legal to see when it's profitable. I suspect this is going to be the right thing to do for pretty much all CPU targets.

http://reviews.llvm.org/D3375






More information about the llvm-commits mailing list