[llvm] [GlobalOpt] Look through non-PointerType constant expressions. (PR #125205)

[Florian Hahn via llvm-commits]

fhahn wrote:

> My primary concern here would be that the result of the inttoptr might point to a different global. That isn't relevant in the context of an icmp (icmp ignores provenance anyway), but could be relevant in other contexts.

Are you thinking about something like below? 

GlobalStatus determines how a particular global is used. If starting from `@G1` we reach an `inttoptr` that may actually refer to a different global `@G2`, then we may have marked `@G1` as used/stored unnecessarily, but that should be conservatively correct I think? It is also what we are doing at the moment and I don't think with the new code we could accidentally return the status of a different global than we were starting from.


```
  store ptr inttoptr (i64 add (i64 ptrtoint (ptr @G) to i64, i64 ptrtoint(ptr @G2) to i64)) to ptr, ptr %dst
```` 



https://github.com/llvm/llvm-project/pull/125205