[llvm] [AMDGPU] Add backward compatibility layer for kernarg preloading (PR #119167)

[Austin Kerbow via llvm-commits]

kerbowa wrote:

RFC. I'm looking for a better way to do the padding here, which right now is just filled out with s_branch instructions. Ideally we could do some kind of fill that is less verbose, I was thinking maybe a pseudo instruction that expands with the proper padding, but I need something that also wont upset the verifier if we have instructions after the branch. Curious if anyone had any thoughts on this?

https://github.com/llvm/llvm-project/pull/119167