[llvm] [Analysis] Track offsets in PtrUseVisitor to revisit when required (PR #179726)

[Jameson Nash via llvm-commits]

vtjnash wrote:

That makes some sense, as before I only enabled it for PHI node loops and not for all instructions (dead code loops). With the more optimal implementation, it made more sense to enable it completely. Any idea where or how to dissect where the extra time is from? (e.g. is it the hashing? the APInt copies? the revisit logic? the maintenance of the InstsInPath stack?). There are probably a number of knobs to turn to test out the different possibilities, unless profiling is revealing of anything. The easiest way to preserve the old performance is perhaps to add a flag for whether the user (most likely SROA) cares if we are revisiting offsets, and then dynamically selecting the old or new code. It perhaps could also check `use_empty()` before re-enqueuing itself, which may save some effort?

https://github.com/llvm/llvm-project/pull/179726