[PATCH] D46282: [AA] cfl-anders-aa with field sensitivity

[Daniel Berlin via Phabricator via llvm-commits]

dberlin added a comment.

So, the loop detection thing you've done is just broken. It's really unclear what you are trying to do.
The normal way to handle such loops would be to do SCC detection and collapse the SCCs properly (which, for field sensitivity, involves resolving what cycle does, usually by marking all things in the relevant type/variable as accessed/written to/whatever)

You appear to just be trying to see if the next list is the same size, which doesn't make any sense.
It's entirely possible for the size to stay the same with their not being a cycle.

The offset handling is definitely broken. Because LLVM's type system is too weak, you really need offset, size pairs that you can resolve through determining overlap, because the size depends on what happens at the access, and not in the type itself (unlike most other IR's). 
IE  size portion can only be determined at load/store points.

Anything with unknown size or unknown offset conflicts with any other pair that exists.


Repository:
  rL LLVM

https://reviews.llvm.org/D46282