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

[Timothy Herchen via llvm-commits]

anematode wrote:

Yeah, but only the bitcast optimization and only if we can't prove the operands are normal. I never looked at `simplifySExtOfDecomposedSetCC` – sorry for the confusion. The issue is that in general I don't think we can rely on DAZ not being set, just because we didn't compile with it enabled. For example, if you link a library that's compiled with `-ffast-math` it sometimes will enable DAZ without you noticing. (The curse of global state!) Then it would be doubly unexpected to have what looks like integer math give spurious results.

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