I don’t know of any statements or guarantees in CUDA C++ that a __shared__ address will numerically fall within the PTX shared logical space. In fact, a simple experiment can be devised to prove, at least in some cases, it does not. Since all pointers in CUDA C++ for current systems/setups are 64-bit, and the shared state space in PTX land evidently involves a 32-bit space, we have another datapoint that says that the __shared__ pointer in CUDA C++ cannot be considered absolutely identical to a shared state space address in PTX. This question may also be of interest.