[llvm] [LangRef] No target-specific size limit for atomics (PR #136864)

[Michael Kruse via llvm-commits]

Meinersbur wrote:

AFAIK no GPU provides `__atomic_load` with arbitrary size directly. Possible could be implemented for CUDA after the Pascal-generation using [`cuda::atomic`](https://nvidia.github.io/cccl/libcudacxx/extended_api/synchronization_primitives/atomic.html#cuda-atomic). Currently, if you would try to use the `__atomic_load` with an unsupported size, you may either get a link error or the compiler [would crash](https://github.com/llvm/llvm-project/blob/8b2d269db22d253ef07a48ef433bc94dc70f2c75/llvm/lib/CodeGen/AtomicExpandPass.cpp#L1738). 

For the purpose of #134455 it doesn't matter where the crash occurs (AtomicExpandPass or emitAtomicLoad).

https://github.com/llvm/llvm-project/pull/136864