[flang-commits] [flang] [flang][AliasAnalysis] Cray pointers/pointees might alias with anything (PR #170900)

[Slava Zakharin via flang-commits]

vzakhari wrote:

Thanks for the details, Tom!

It looks like this test contradicts with gfortran's documentation that says that such cases might be optimized in a way that the result would be incorrect.  As Jean also pointed, Flang is allowed to optimize it in such a way.

Your previous TBAA change made an improvement that potentially enabled more optimizations, and this test started failing.  To me it is an expected behavior, and I agree with Jean that maybe we do not need to complicate the implementation because of this case.  On ther other hand, it does not look like a huge complication right now.  So I would be okay with supporting this case with an option that would be off by default.

https://github.com/llvm/llvm-project/pull/170900