One limitation at the moment is the pass only reasons about inbounds GEPs, if we can prove that the index is positive, to ensure the GEP doesn’t wrap in the unsigned sense.
This means it won’t perform simplifications on GEPs with size_t indices, unless there’s additional info that the index is positive. I posted Retain fact that index/offsets for GEP are non-negative for `size_t/uint64_t` offsets? to discuss options to improve the situation.
An example from the issue tracker is LLVM doesn't eliminate bounds check inside loop · Issue #51358 · llvm/llvm-project · GitHub