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

[via llvm-commits]

gonzalobg wrote:

> In OpenCL and CUDA, it is undefined behavior for an atomic instruction
to modify an object in thread private memory.

This is not true for CUDA, an LLVM IR atomic instruction on thread private memory is fine.
(The LLVM NVPTX backend may have a bug that generates wrong PTX, but that's a backend bug).

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