[llvm] [RFC] IR: Support atomicrmw FP ops with vector types (PR #86796)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> > I have support for fp vector typed xchg, and vector of int/ptr separately implemented
> 
> Does supporting those require substantially different code than you've already modified here? Perhaps should just do it at the same time for consistency?

It greatly expands the testing surface, and hits pretty different paths in the expansions. As far as xchg goes, I think we've accumulated some unnecessary cruft where AtomicExpand is inserting bitcasts when the legalizer can do it just as well which I'd prefer to clean up before touching it. Overall it's not much more code, but it's still separable 

https://github.com/llvm/llvm-project/pull/86796