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

[via llvm-commits]

goldsteinn wrote:

> > > > I recall I read someone commented with concern when DAZ is set, fcmp will always return equal for denormal values, but I cannot find the comment now. Is this a real concern? Did we fix it?
> > > 
> > > 
> > > Looks like that would be the result of DAZ, but didn't test: https://stackoverflow.com/questions/54047415/do-denormal-flags-like-denormals-are-zero-daz-affect-comparisons-for-equality
> > > Ill update with `sub` flag.
> > 
> > 
> > Actually, I don't think this is an issue since we don't handle `eq`/`ne`
> 
> It affects all predicates, not only `eq`/`ne`, see https://godbolt.org/z/jxs5s5qde

ah.

We don't naturally have `subnormal` fpclass in DAG, so think we just need to disable the simplifications involving `Op1C == 0` if `-mdaz-ftz`? Or would we need to disable it regardless unless we can prove `Op0` is normal?

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