Michael Marcin wrote:
IIUC what that forward declared partial specialization is doing is akin to saying: - I know the default implementation of tag_of can't be correct for boost::array<T,N> - There should be an implementation of tag_of that will make it work for boost::array<T,N> - I don't know what that implementation is, you or someone else better provide it
Yes. It's done to prevent (different kinds of) ODR violations. Example 1: template<typename T> struct X { }; template<typename T> struct Y { }; // [...] <--- 'X< Y<void> >' gets instantiated template<typename T> struct X< Y<void> > { // !!! ERROR, already defined }; Example 2: // first translation unit includes template<typename T> struct is_xyz : mpl::false_ { }; template<> struct is_xyz<void> : mpl::true_ { }; // now is_xyz<xyz>::value == true ---- // second translation unit misses to include the specialization template<typename T> struct is_xyz : mpl::false_ { }; // now is_xyz<xyz>::value == false Note there are compilers that happily compile and link example 2.
Is it then true that when proper definition of the partial specialization is provided this partial specialization of the declaration has no effect whatsoever?
Yes, once the type is completed it's complete ;-). It just prevents the type to be completed prematurely by instantiating the primary template.
If that is true and this is a compiler bug then perpahs we can use a configuration macro like BOOST_NO_DECLARTION_PARTIAL_SPECIALIZATION to simply omit these these forward declarations on compilers that choke on them (like CW 9.4 apparerently does).
Are you sure it only applies to /partial/ specializations? How about the full ones? My suggestions for the names would be: BOOST_NO_FORWARD_DECLARED_TEMPLATE_SPECIALIZATION or BOOST_NO_FORWARD_DECLARED_TEMPLATE_PARTIAL_SPECIALIZATION Regards, Tobias