[llvm] AMDGPU: Refactor lowering of s_barrier to split barriers (PR #154648)

[Nicolai Hähnle via llvm-commits]

nhaehnle wrote:

I'm not a fan of doing ever more stuff in MachineIR. The tooling around IR is better, the compiler code working on it tends to be cleaner, and it's generally better understood by more people.

https://github.com/llvm/llvm-project/pull/154648