[llvm] [SeparateConstOffsetFromGEP] Reorder trivial GEP chains to separate constants (PR #73056)

[Krzysztof Drewniak via llvm-commits]

krzysz00 wrote:

... hold on, from looking at that Alive2, I think you can be a bit more aggressive than just not applying this swap to inbounds GEPs (or stripping the inbounds - do we know which of these is more likely to lead to good codegen?).

Specifically, I claim that, if you can prove that the constant and variable indices are non-negative, you should be able to reassociate the inbounds GEPs soundly. 

See the somewhat ham-fisted https://alive2.llvm.org/ce/z/CRkgEN 

https://github.com/llvm/llvm-project/pull/73056