[llvm] [AMDGPU] Enhance verification of amdgcn.cs.chain intrinsic (PR #128162)

[Robert Imschweiler via llvm-commits]

ro-i wrote:

Hm, but in that case, we would first need to implement `callbr` ^^
Because for GlobalISel, it isn't implemented at all, and for SelDAG, the code and [the docs](https://llvm.org/docs/LangRef.html#callbr-instruction) say that it's (currently) only "to implement the 'goto' feature of gcc inline assembly". ("This instruction should only be used to implement the “goto” feature of gcc style inline assembly. Any other usage is an error in the IR verifier.")
Why do you think that `callbr` would be a better option than e.g. enforcing tail call and stopping to emit?
See also the [GCC docs on the goto with labels](https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html#GotoLabels).

https://github.com/llvm/llvm-project/pull/128162