Robert Ramey wrote:
BOOST_CLASS_VERSION should be in the header. This might be overkill, but its a natural place for it, it's bonehead proof and doesn't have any side effects. BOOST_CLASS_EXPORT has the side effect of instantiatine code so it's "special"
My original intention was that all these "serialization traits" be part of the class header description. BOOST_CLASS_EXPORT sort of broke this idea but it still holds for the other ones.
Right, the version macro just creates a version trait specialization-majig. Here's our case... maybe there's no good solution, if so, no big deal. To keep compile times down we try to keep all the serialization stuff out of our headers except for the (dependencyless, happily) declaration of the serialize function struct C : Base { template <class Archive> void serialize(Archive&, const unsigned); }; and we instantiate the serialize function for several different archive types in the .cpp, via macro #include <all_the_heavy_serialization_headers> template <class Archive> void C::serialize(Archive& ar, const unsigned version) { ... } SERIALIZABLE(C) // instantiates and exports I've mentioned this before. Nearly all of these hundred-or-so classes are serialized in one place via pointer-to-base. So it'd be nice to have the headers completely clean of serialization #includes. If not, no big deal, but: I don't really get why, if the library that does the serialization for C can't see C's header (it goes via shared_ptr<Base>, and has loaded C's code from a shared library), does the BOOST_CLASS_VERSION need to be in the header? Shouldn't the version-checking business get taken care of at the point of instantiation of the serialization method for C? -t