The following C++ 11 program fails to link with GCC (both 4.9 and 5.0) because the primary std::atomic template allocates 3 bytes for both a and aa and calls __atomic_store(&aa.a, &tmp, __ATOMIC_SEQ_CST).  GCC only supports calls to __atomic_store with power of 2 arguments of at most 8.  Assuming the code is well-defined (I couldn't find anything to suggest otherwise -- the type meets the trivially-copyable requirement) it seems that specializations of the template on types of such sizes should add padding up to the nearest greater power of 2 and increase their own alignment so as to make it possible to manipulate them atomically.  Larger specializations will need to make use of a mutex.

$ cat a.cpp && g++ -Wall -Wpedantic -std=c++11 a.cpp
#include <atomic>

int main (void) {
    struct A { char a [3]; } a = { { 'a', 'b', 'c' } };
    std::atomic<A> aa;
    aa.store (a);
}
/tmp/ccGQyZte.o: In function `std::atomic<main::A>::store(main::A, std::memory_order)':
a.cpp:(.text+0xf0): undefined reference to `__atomic_store'
collect2: error: ld returned 1 exit status