[llvm] [InstCombine] Fix regressions from canonicalizing `(icmp eq/ne (X, Y), Y` -> `(icmp eq/ne (~X, Y), 0)` (PR #93656)

[via llvm-commits]

goldsteinn wrote:

> > @dtcxzyw can you test this to see if the regression actually matters?
> 
> Done. I don't think the regression exists.

Thanks for checking

https://github.com/llvm/llvm-project/pull/93656