The following two things are not necessarily the same:
- Atomic operations on a C++ level (std::atomic)
- Assembler CPU instructions that are guaranteed to be atomic
For example, a modern Intel processor guarantees that 64-bit reads/writes to a 64-bit aligned address is atomic. It also guarantees that an instruction with a LOCK prefix is executed atomically.
In C++, on the other hand, std::atomic guarantees that an operation on an object is executed atomically. A good compiler will use a single atomic CPU instructions for this, if possible. However, for larger objects, this atomicity cannot be accomplished with a single atomic CPU instruction. Therefore, atomicity can only be guaranteed by using more complex synchronization methods, such as locks. Such complex synchronization methods require several CPU instructions and can therefore always be interrupted.
You can determine whether std:atomic is using locks or not by calling std::atomic::is_lock_free().