"Daniel James" <daniel@calamity.org.uk> wrote
Using mpl views might be better. Currently the implementation works by passing a list/vector to encode and adding to it. Instead encode could just return the sequence for it's sub-type, which can then be combined using mpl::joint_view, or similar.
Note that push_back<mpl::vector<...> > requires just one template instantiation. I am not sure why views are better. And views definitely don't provide constant-time lookup.
For example, in the current implementation:
template<class V, class T> struct encode_type_impl<V, const T> { typedef typename encode_type< typename detail::push_back< V , mpl::int_<CONST_ID> >::type , T>::type type; };
Would become something like:
template<class T> struct encode_type_impl<const T> { typedef boost::mpl::joint_view< boost::mpl::single_view<mpl::int_<CONST_ID> >, encode_type<T> >; };
The big advantage here is that if T has already been encoded (even as part of a completely different type) it can be reused. For example if the following were encoded:
std::pair<int, int> std::pair<std::pair<int, int>, std::pair<int, int> >
The existing versions would encode std::pair<int, int> 3 times, this would do it once.
Agreed, but this should be weighted against not having direct access. Regards, Arkadiy