"David Abrahams" <dave@boost-consulting.com> wrote
That's the basic idea behind all typeof() emulations that I've ever seen. They all require some type registration to associate a compile-time integral constant with a type. The innovation that Steve brought, IIUC, was that with the registration of only class types, the rest of the transformations possible in the type system could be encoded by the library automatically. For example, if you register class Foo, then the library can generate encodings for Foo*, Foo[2], Foo**, int (Foo), Foo(*)(Foo,Foo), etc., so you don't have to register those types. The problem is that those numbers can get very large...
Right. What I am even more interested in, is the ability to generate encodings of complicated templates. Say I registered int, std::vector, and mpl::vector2, and the library should be able to work with std::vector<int>, mpl::vector2<std::vector<int>, mpl::vector2<int, int> >, etc. It looks like this can be solved pretty easily. I imagine that Boost.Lambda, for example might also benefit from the ability to encode/decode templates (and maybe Spirit, but I am not sure).
My guess is that manipulating array types in that way has a lower cost at compile-time than using a vector_c, though using mpl probably yields much nicer library code and vector_c is more portable.
I don't suspect my solution is very compile-time efficient. But I don't have enough knowledge of compilers internals to estimate this. I believe the real advantage is in is simplicity. Anyway, here it is... Let's recall that to achieve our goal we have to: 1) Pass our expression to a function template, like: template<class T> SOMETHING foo(const T&); 2) inside foo() encode our type into a compile-time list of integers; 3) somehow manage to pass this list through the function boundaries; 4) decode the type. The most chalanging is 3. Let's define the following template: template<class N> struct sizer { char_[N::value]; }; The whole purpose of this is to create classes with a given sizeof (to be completely accurate we have to then divide this by sizeof(char)). Next template is what it is all about: template<class IntList> struct multi_int { typedef IntList int_list; template<class T> sizer<typename mpl::apply_if< typename mpl::less<T, typename mpl::size<int_list>::type>, mpl::at<IntList, T>, mpl::identity<mpl::int_<1> > >::type> get(const T&); }; So, if we have an object of this class, x, then sizeof(x.get(mpl::int_<n>())) defines the nth element of the list, and out-of-range values produce 1. The list can be reproduced like, say: #define BOOST_RTL_TYPEITEM(Expr, N)\ mpl::int_<sizeof(Expr.get(mpl::int_<N>()))> mpl::vector<\ BOOST_RTL_TYPEITEM(x, 0) , BOOST_RTL_TYPEITEM(x, 1) ... , BOOST_RTL_TYPEITEM(x, N)
Where it can be generated by the preprocessor for a fairly large N. The reproduced list is larger than the original one, and is 1-padded. The encoding scheme I use does not require to know where the list ends, so this doesn't matter. This is basically it. The return type of the function foo above (sorry for such name) is going to be multi_int<encode<T>::type>, where T is the original type, and encode template works accordingly to some encoding scheme. The encoding scheme to choose is a separate issue. In my example I encode templates by adding the encodings of its parameters to the code of template itself. So, if int is 1, std::vector is 2, and mpl::vector2 is 3, the above example: mpl::vector2<std::vector<int>, mpl::vector2<int, int> > becomes 3 2 1 3 1 1 Also note that all templates with one parameter behave similarly, so only one registration macro is required. Another macro is required to register templates with two parameters, etc.
For what it's worth, I'm pretty sure I know how to solve the problem that most typeof() implementations strip top-level references. See boost/iterator/is_lvalue_iterator.hpp for a demonstration.
Cool. I imagine pointers are pretty easy, but references surely represent a big problem... I will follow up with the working sketch... Regards, Arkadiy