[llvm] [InstCombine] Try to fold add into GEP x, C (PR #85090)

[David Green via llvm-commits]

davemgreen wrote:

I'd like to try this again if I can. This doesn't seem like an unreasonable patch to take in the sort-term, and as far as I understand is a (little) step towards where we expect codegen to end up. As in - it performs an optimization that will happen if geps are canonicalized to individual offsets, bringing the final codegen closer to what it will later be. And doesn't create larger geps, just transforms existing large geps to be a little different. I've given the patch a rebase. Let me know what you think.

https://github.com/llvm/llvm-project/pull/85090