[llvm] [X86] Try Folding icmp of v8i32 -> fcmp of v8f32 on AVX (PR #82290)

[via llvm-commits]

goldsteinn wrote:

> It'd be good to get #82242 fixed - have you any thoughts on how we can reduce the scope of this patch to avoid having to chase regressions so much?

I think most of the scope ends up being there too fix regressions. The fundemental issue is that going backwards is a lot more difficult and we are able to do some simplifications on the integer versions that we can't on the float ones.

Not really sure how to address them TBH.

https://github.com/llvm/llvm-project/pull/82290