[llvm] [InstCombine] Fold `trunc nuw/nsw (x xor y) to i1` to `x != y` (PR #90408)

[via llvm-commits]

alexfh wrote:

I suppose, this particular function is only involved in what _might_ be a miscompile, when it's inlined into another function, but disabling the optimization just on it using `#pragma clang optimize off/on` is enough to `fix` the test we see failing. The problem is that the whole setup leading to the problem is quite convoluted, and there's no isolated reproducer so far.

We're also actively investigating other possibilities like a preexisting UB in the code.

https://github.com/llvm/llvm-project/pull/90408