[llvm] [InstCombine] Relax the same-underlying-object constraint for the GEP canonicalization (PR #76583)

[Nikita Popov via llvm-commits]

nikic wrote:

> Do you actually need the ptrtoint case? The icmp one is definitely correct, the ptrtoint one may or may not be depending on what ptrtoint semantics we decide on (e.g. it would be incorrect if it has "expose" semantics).

Nevermind, this should be fine for ptrtoint as well, regardless of the specific semantics, because we already have ptrtoints for both x and y in the first place, so the amount of exposed provenance remains the same.

https://github.com/llvm/llvm-project/pull/76583