[PATCH] D50585: [WIP][LICM] Just rely on AA rather than special case calls

[Philip Reames via Phabricator via llvm-commits]

reames abandoned this revision.
reames added a comment.

This was too good to be true.  The patch made an important conceptual mistake which is that it assumes an instruction maps to a single alias set.  In general, this is not true.  It may instead map to several distinct alias sets.  For a bit more discussion of this, see https://reviews.llvm.org/D50730.

You could extend this approach to multiple aliasing sets, but once you do, you're basically back to the code I was replacing with an extra layer of abstraction added.  Overall, that's not a net win.


Repository:
  rL LLVM

https://reviews.llvm.org/D50585