http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53646
--- Comment #5 from Michael Matz <matz at gcc dot gnu.org> 2012-06-12 15:36:01 UTC --- (In reply to comment #2) > N.B. std::pair is not a POD in c++98 or c++11, so I don't know what libstdc++ > could have done to cause the FE to change how it returns a std::pair. I don't know if it's PODness (but I believe that it is), but the calling convention is changed by the existence of this c++11 construct in std::pair: pair(pair&& __p) noexcept(__and_<is_nothrow_move_constructible<_T1>, is_nothrow_move_constructible<_T2>>::value) : first(std::forward<first_type>(__p.first)), second(std::forward<second_type>(__p.second)) { } It's some sort of copy-ctor, right? In any case when it's there (and it's only there when compiling/preprocessing in c++11 mode), then the frontend makes this type be TREE_ADDRESSABLE, aggregate_value_p will return true, and that makes the ABI use an invisible reference.