Why is std::atomic_ref not implemented in terms of std::atomic

Question

The referenced implementation of std::atomic_ref from the paper P0019r8 roughly keeps the template type as a member variable ( https://github.com/ORNL/cpp-proposals-pub/blob/master/P0019/atomic_ref.hpp ) and uses the GNU built-ins to implement atomic operations.

The question I have here is - why not reinterpret_cast to std::atomic and use the atomic operations instead? Is there a portability concern or detail I am missing?

Answer 1

重新解释某种东西而不使用它是不确定的行为。

Answer 2

There's no guarantee whatsoever that a std::atomic<T> contains nothing but a T and has the same size and alignment requirements as a T . For example, if sizeof(T) == 3 , an implementation of std::atomic<T> may pad it to 4 bytes to enable the use of intrinsics. For another example, if sizeof(T) is too big for an intrinsic, std::atomic<T> might store a synchronization primitive of some sort to serialize the operation.

It follows that reinterpret_cast to std::atomic is not a viable implementation in the general case even if you ignore the general undefined behavior from the object model violations.