[llvm] [SCEV] Fix incorrect nw-inference in zext-addrec (PR #217785)

[Ramkumar Ramachandra via llvm-commits]

artagnon wrote:

> I think you're confusing pre-inc and post-inc addrecs in your analysis?
> 
> The two values for i are -1 and 127 (or 255 and 127 after zext). i.inc does take -1 on the last iteration, but i.inc is not i. i itself does not self-wrap -- right?

Yes, I think you're right -- I confused pre-inc and post-inc values :(

https://github.com/llvm/llvm-project/pull/217785