[llvm] [AArch64][GlobalISel] FP Info implementation for AArch64. (PR #177158)

[Ryan Cowan via llvm-commits]

HolyMolyCowMan wrote:

Those MachineVerifier changes make sense to me!

---

I imagine we want to stick as close to the semantics of the IR as possible, it's less for the developer to hold in their head and reason about as you suggest. I can see where a speedup may come from in option 2, would you imagine it being a large speedup?

You also suggested that option 2 may lead to fewer changes in other backends, I would imagine that changing the semantics of the floating point MIR instructions would likely be a larger change than inserting bitcasts right?

I can understand the idea behind treating integer as a superclass of float, I've seen a few example of this in backends  where integer stands in for representing the underlying memory. I wonder if that makes more sense as being a generic scalar, the semantics are a little more separate that way?

https://github.com/llvm/llvm-project/pull/177158