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

[via llvm-commits]

goldsteinn wrote:

> > Any luck with the remaining regressions?
> 
> Most of the regressions are fixed with #84360, the only outstanding one is when we later simplify to `xmm` w/ demandedelts in which case we have converted from int -> float without any benefit. For example: https://github.com/llvm/llvm-project/pull/84360/files#diff-ec93308ebc948e632bea97e65a5aeedc4342aab5fa49d2c4a43be983131fba47R17
> 
> Thats tough to get around, as float -> int is much more difficult to pull off.

Im trying to refactor so the code arbitrarily works to cvt int <-> fp compares. Then we can undo as needed.

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