[llvm] [AMDGPU][ASAN] Handle special GVs lowering in amdgpu-sw-lower-lds (PR #161827)

[Jay Foad via llvm-commits]

jayfoad wrote:

> > On platform without hardware support, we can emulate named-barrier using LDS atomics.
> 
> Why? What's the software need to do that on older hardware?

There is no software need, it's just part of the rationale for making them look like LDS allocations. As I understand it:
- they are allocated per-workgroup just like LDS allocations
- you could in theory have a software implementation that actually allocated them in LDS
- with that in place, the path to allocate them to hardware barrier IDs would be "just an optimizastion"
- but there's no use case for that fallback path, so we have only implemented the optimized path

https://github.com/llvm/llvm-project/pull/161827