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

[Jun Wang via llvm-commits]

jwanggit86 wrote:

> For the primary use, these constraints are really only used by flat_load_ and flat_store_ instructions.
> 
> So in inline asm, you'd really only want to use "RF" if you literally have a flat_load or flat_store in the asm string -- perhaps as part of a larger block of inline code, say some hand-optimised sequence or something.
> 
So what does GCC do if (1) Something other than 'RF', say, 'm' or 'RB', is used in flat_load/store_, OR (2) 'RF' is used in other load/store instructions, eg., global_load? Thanks!



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