Rationale behind member function swap

In the standard library, if a class type has a specialized swap algorithm, then it will have a member function swap and a free function swap that simply forwards to the member function. I don't quite get the rationale behind having both of them (and thus code duplication) but not just the free function one. Note that, when I say the free function, I take to mean the specialized free swap function, not the generic std::swap function template. Such a specialized function may have privileged access to the relevant class by being a friend of it.

I have two points to support just leaving the free function rather than the member function. First, it forms a more generic swap interface to facilitate writing generic algorithms, for non-class types like arrays cannot have member functions. Second, swap is a binary operation involving two operands and demonstrates a sense of symmetry. It's more natural and intuitive to perform swap using a free function that does not have bias on either operand. For this reason, I have always felt somewhat weird when using the member function swap as if I was performing some operation that is based on the invoking object.

The free function cannot see the private guts of the class, so the actual functionality needs to be provided as a class member. However, you want the free function because it's a customization point (using ADL); so a generic algorithm will call the free swap in order to achieve whatever swapping is appropriate for the type at hand.

The member function is also useful for constructions like T().swap(existing_thing), since the free function requires lvalue arguments.

I suppose the alternative would have been to make swap a public friend function declared inside the class, which could be found via ADL just as well.