[llvm] [RFC] IR: Define noalias.addrspace metadata (PR #102461)

[via llvm-commits]

gonzalobg wrote:

> > This is not true for CUDA, an LLVM IR atomic instruction on thread private memory is fine.
> 
> Are you sure?

Yes, it is only true for the pre sm_70 CUDA `__atomicXXX` builtins, but not true for the post sm_70 and newer CUDA atomics APIs (`cuda::atomic`, `cuda::atomic_ref`).

https://github.com/llvm/llvm-project/pull/102461