Cereal breaks with multiple definitions if you include a header with a registered type in different TUs about cereal

[AzothAmmo]

I have a fix for this incoming that is essentially your solution with a ton of documentation. We're going to have two versions of CEREAL_REGISTER_TYPE, one that does the binding and one that doesn't do it.

[AzothAmmo]

After thinking about this some more I'm not convinced that we need to split things off into a header macro e.g. CEREAL_REGISTER_TYPE_WITHOUT_BINDING (that would avoid calling CEREAL_BIND_TO_ARCHIVES(T)) and a separate definition macro (CEREAL_BIND_TO_ARCHIVES(T)).

In all of the tests I've done we can solve the issue by moving CEREAL_REGISTER_TYPE into a .cpp file. We've been doing this in shared library code and some simple static library code. See https://gist.github.com/AzothAmmo/a56b63b225a0886eddc2 for the example I've been working with.

If you can think of any situations which break this idiom (putting it in the .cpp files) let us know.

[sithhell]

Looks like it is working! Thanks for the resolution.

If that solution turns out to work in our project, you guys are my heros. We are currently dealing with massive compile time and link times because boost.serialization is instantiating an insane amount of code. If we only have to do the registration in the source files means a big win.

[chenghuige]

Below might fix this problem

#define CEREAL_BIND_TO_ARCHIVES(T)
namespace cereal {
namespace detail {
template<>
struct init_binding {
static bind_to_archives const & b()
{
const static bind_to_archives _b = ::cereal::detail::StaticObject<bind_to_archives >::getInstance().bind(); \
return _b;
}
static void unused() { (void)b(); }
}; \
}} /* end namespaces */