Here is a possible definition of std::swap:
template
void swap(T& a, T& b) {
T tmp(std::move(a));
a = std::move(b);
[res.on.arguments] is a statement about how the client should use the std::lib. When the client sends an xvalue to a std::lib function, the client has to be willing to pretend that the xvalue is really a prvalue, and expect the std::lib to take advantage of that.
However when the client calls std::swap(x, x), the client isn't sending an xvalue to a std::lib function. It is the implementation that is doing so instead. And so the onus is on the implementation to make std::swap(x, x) work.
That being said, the std has given the implementor a guarantee: X shall satisfy MoveAssignable. Even if in a moved-from state, the client must ensure that X is MoveAssignable. Furthermore, the implementation of std::swap doesn't really care what self-move-assignment does, as long as it is not undefined behavior for X. I.e. as long as it doesn't crash.
a = std::move(b);
When &a == &b, both the source and target of this assignment have an unspecified (moved-from) value. This can be a no-op, or it can do something else. As long as it doesn't crash, std::swap will work correctly. This is because in the next line:
b = std::move(tmp);
Whatever value went into a from the previous line is going to be given a new value from tmp. And tmp has the original value of a. So besides burning up a lot of cpu cycles, swap(a, a) is a no-op.
Update
The latest working draft, N4618 has been modified to clearly state that in the MoveAssignable requirements the expression:
t = rv
(where rv is an rvalue), t need only be the equivalent value of rv prior to the assignment if t and rv do not reference the same object. And regardless, rv's state is unspecified after the assignment. There is an additional note for further clarification:
rvmust still meet the requirements of the library component that is using it, whether or nottandrvrefer to the same object.