As I understand it, C volatile and optionally inline asm for memory fence have been used for implementing a device driver on top of memory mapped I/O. Several examples can be found in Linux kernel.
If we forget about the risk of uncaught exceptions (if any,) does it make sense to replace them with C++11 atomics? Or, is it possible at all?
In general, you can replace memory fences with atomics, but not volatile, except where it is used together with a fence exclusively for inter thread communication.
Whith regard to memory mapped I/O the reason atomics don't suffice is that:
volatile guarantees you that all memory accesses to that variable in your program do actuall happen and that they happen (whithin a single thread) exactly in the order you specify. std::atomic only guarantees that your program will behave as if all those memory accesses happen (according to C++'s memory model, which doesn't know about memory mapped I/O) and - depending on the specified memory ordering - as if they happen in the specified order.In practical terms that means, that the compiler can e.g. replace consecutive writes to the same (non-volatile) atomic with a single write (if there is no other synchronization in between) and the same is true for reads. If the result of the read is not used, it could even eliminate the read completely (the compiler might still have to issue a memory barrier though).
On a more theoretical level, if your compiler can prove that all your program does is returning 42, then it is allowed to transform this into a single instruction independently of how many threads and atomics your program uses in the process. If your program uses volatile variables that is not the case.
EDIT: E.g. This paper shows a few posssible (and probably unexpected) optimizations the compiler is allowed to apply to an atomic loop variable.