[llvm] [WIP][AMDGPU] Support the RF inline asm constraint (PR #195123)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> I don't think it'd be easy to find a case for using the generic "m" constraint. It'd be hard to use safely: the danger would be passing an address to an instruction that it is invalid for, likely causing problems compiling at different optimisation levels, etc.

It would fail to assemble or be user error, that's not different than misusing any operand?

> 
> Another point, stepping back: it might be hard to faithfully reproduce the GCC semantics in LLVM because of the different execution models used by the compilers, in the sense of SIMT vs. SIMD.

This won't change encoding details 



https://github.com/llvm/llvm-project/pull/195123