It's not a missed optimization. The compiler can know that we are
dealing with a B&, and that B& must be returned from f1(), f2(), f3(),
but this is not helpful.

We are still making virtual calls to do_f1(), do_f2(), do_f3(). If we
could magically devirtualize any call by going through a member
function that has "return *this;" in it, it would totally break
polymorphism.

However, there is something that does look like a missed optimization
here: https://www.godbolt.org/z/3Kc8dnMMj

going through a non-virtual member function seems to prevent
de-virtualization for seemingly no reason.