fusion/doc/extension.qbk

[section Extension]

The Fusion library is designed to be extensible, new sequences types can easily
be added. In fact, the library support for `std::pair`, `boost::array` and __mpl__ 
sequences is entirely provided using the extension mechanism.

The process for adding a new sequence type to Fusion is:

# Enable the __tag_dispatching__ mechanism used by Fusion for your sequence type
# Design an iterator type for the sequence
# Provide specialized behaviour for the intrinsic operations of the new Fusion sequence

[heading Our example]

In order to illustrate enabling a new sequence type for use with Fusion, we
are going to use the type:

    namespace example
    {
        struct example_struct
        {
            std::string name;
            int age;
            example_struct(
                const std::string& n,
                int a)
                : name(n), age(a)
            {}
        };
    }
    
We are going to pretend that this type has been provided by a 3rd party
library, and therefore cannot be modified. We shall work through all the
necessary steps to enable `example_struct` to serve as an __associative_sequence__
as described in the __quick_start__ guide.

[heading Enabling Tag Dispatching]

The Fusion extensibility mechanism uses __tag_dispatching__ to call the
correct code for a given sequence type. In order to exploit the tag
dispatching mechanism we must first declare a new tag type for the
mechanism to use. For example:

    namespace example {
        struct example_sequence_tag; // Only definition needed
    }

Next we need to enable the `traits::tag_of` metafunction to return our newly chosen
tag type for operations involving our sequence. This is done by specializing
`traits::tag_of` for our sequence type.

    #include <boost/fusion/support/tag_of_fwd.hpp>

    namespace boost { namespace fusion { namespace traits {        
        template<>
        struct tag_of<example_struct>
        {
            typedef example::example_sequence_tag type;
        };
    }}}

`traits::tag_of` also has a second template argument, 
that can be used in conjuction with `boost::enable_if` to provide tag 
support for groups of related types. This feature is not necessary
for our sequence, but for an example see the code in:

    #include <boost/fusion/sequence/adapted/mpl/tag_of.hpp>

[heading Designing a suitable iterator]

We need an iterator to describe positions, and provide access to
the data within our sequence. As it is straightforward to do,
we are going to provide a random access iterator in our example.

We will use a simple design, in which the 2 members of
`example_struct` are given numbered indices, 0 for `name` and
1 for `age` respectively.

    template<typename Struct, int Pos>
    struct example_struct_iterator
        : boost::fusion::iterator_base<example_struct_iterator<Struct, Pos> >
    {
        BOOST_STATIC_ASSERT(Pos >=0 && Pos < 3);
        typedef Struct struct_type;
        typedef boost::mpl::int_<Pos> index;
        typedef boost::fusion::random_access_traversal_tag category;

        example_struct_iterator(Struct& str)
            : struct_(str) {}

        Struct& struct_;
    };

A quick summary of the details of our iterator:

# The iterator is parameterized by the type it is iterating over, and the index of the current element.
# The typedefs `struct_type` and `index` provide convenient access to information we will need later in
  the implementation.
# The typedef `category` allows the `traits::__category_of__` metafunction to establish 
  the traversal category of the iterator.
# The constructor stores a reference to the `example_struct` being iterated over.

We also need to enable __tag_dispatching__ for our iterator type, with another specialization of
`traits::tag_of`.

In isolation, the iterator implementation is pretty dry. Things should become clearer as we
add features to our implementation.

[heading A first couple of instructive features]

To start with, we will get the __result_of_value_of__ metafunction working. To
do this, we provide a specialization of the `boost::fusion::extension::value_of_impl` template for 
our iterator's tag type.

    template<>
    struct value_of_impl<example::example_struct_iterator_tag>
    {
        template<typename Iterator>
        struct apply;

        template<typename Struct>
        struct apply<example::example_struct_iterator<Struct, 0> >
        {
            typedef std::string type;
        };

        template<typename Struct>
        struct apply<example::example_struct_iterator<Struct, 1> >
        {
            typedef int type;
        };
    };

The implementation itself is pretty simple, it just uses 2 partial specializations to
provide the type of the 2 different members of `example_struct`, based on the index of the iterator.

To understand how `value_of_impl` is used by the library we will look at the implementation of __value_of__:

    template <typename Iterator>
    struct __value_of__
        : extension::value_of_impl<typename detail::tag_of<Iterator>::type>::
            template apply<Iterator>
    {};

So __value_of__ uses __tag_dispatching__ to select an __mpl_metafunction_class__
to provide its functionality. You will notice this pattern throughout the
implementation of Fusion.

Ok, lets enable dereferencing of our iterator. In this case we must provide a suitable
specialization of `deref_impl`.

    template<>
    struct deref_impl<example::example_struct_iterator_tag>
    {
        template<typename Iterator>
        struct apply;

        template<typename Struct>
        struct apply<example::example_struct_iterator<Struct, 0> >
        {
            typedef typename mpl::if_<
                is_const<Struct>, std::string const&, std::string&>::type type;

            static type
            call(example::example_struct_iterator<Struct, 0> const& it)
            {
                return it.struct_.name;
            }
        };

        template<typename Struct>
        struct apply<example::example_struct_iterator<Struct, 1> >
        {
            typedef typename mpl::if_<
                is_const<Struct>, int const&, int&>::type type;

            static type
            call(example::example_struct_iterator<Struct, 1> const& it)
            {
                    return it.struct_.age;
                }
            };
        };
    }

The use of `deref_impl` is very similar to that of `value_of_impl`, but it also
provides some runtime functionality this time via the `call` static member function. 
To see how `deref_impl` is used, lets have a look at the implementation of __deref__:

    namespace result_of
    {
        template <typename Iterator>
        struct __deref__
            : extension::deref_impl<typename detail::tag_of<Iterator>::type>::
                template apply<Iterator>
        {};
    }

    template <typename Iterator>
    typename result_of::deref<Iterator>::type
    __deref__(Iterator const& i)
    {
        typedef result_of::deref<Iterator> deref_meta;
        return deref_meta::call(i);
    }

So again __result_of_deref__ uses __tag_dispatching__ in exactly the
same way as the __value_of__ implementation. The runtime functionality used
by __deref__ is provided by the `call` static function of the selected
__mpl_metafunction_class__.

The actual implementation of `deref_impl` is slightly more complex than that of  `value_of_impl`. 
We also need to implement the `call` function, which returns a reference
to the appropriate member of the underlying sequence. We also require a little
bit of metaprogramming to return `const` references if the underlying sequence
is const.

[note Although there is a fair amount of left to do to produce a fully fledged
Fusion sequence, __value_of__ and __deref__ illustrate all the signficant concepts
required. The remainder of the process is very repetitive, simply requiring
implementation of a suitable `xxxx_impl` for each feature `xxxx`.
]

[heading Implementing the remaining iterator functionality]

Ok, now we have seen the way __value_of__ and __deref__ work, everything else will work
 in pretty much the same way. Lets start with forward iteration,
by providing a `next_impl`:

    template<>
    struct next_impl<example::example_struct_iterator_tag>
    {
        template<typename Iterator>
        struct apply
        {
            typedef typename Iterator::struct_type struct_type;
            typedef typename Iterator::index index;
            typedef example::example_struct_iterator<struct_type, index::value + 1> type;

            static type
            call(Iterator const& i)
            {
                 return type(i.struct_);
            }
        };
    };

This should be very familiar from our `deref_impl` implementation, we will be
using this approach again and again now. Our design is simply to increment
the `index` counter to move on to the next element. The various other iterator
manipulations we need to perform will all just involve simple calculations
with the `index` variables.

We also need to provide a suitable `equal_to_impl` so that iterators can be
correctly compared. A __bidirectional_iterator__ will also need an implementation of `prior_impl`. For a
__random_access_iterator__ `distance_impl` and `advance_impl` also need to be provided
in order to satisfy the necessary complexity guarantees. As our iterator is
a __random_access_iterator__ we will have to implement all of these functions.

Full implementations of `prior_impl`, `advance_impl`, `distance_impl` and `equal_to_impl` are 
provided in the example code.

[heading Implementing the intrinsic functions of the sequence]

In order that Fusion can correctly identify our sequence as a Fusion sequence, we
need to enable `is_sequence` for our sequence type. As usual we just create
an `impl` type specialized for our sequence tag:

     template<>
     struct is_sequence_impl<example::example_sequence_tag>
     {
         template<typename T>
         struct apply : mpl::true_ {};
     };

We've some similar formalities to complete, providing `category_of_impl` so Fusion
can correctly identify our sequence type, and `is_view_impl` so Fusion can correctly
identify our sequence as not being a __view__ type. Implementations are 
provide in the example code.

Now we've completed some formalities, on to more interesting features. Lets get
__begin__ working so that we can get an iterator to start accessing the data in
our sequence.

    template<>
    struct begin_impl<example::example_sequence_tag>
    {
        template<typename Sequence>
        struct apply
        {
            typedef example::example_struct_iterator<Sequence, 0> type;

            static type
            call(Sequence& seq)
            {
                return type(seq);
            }
        };
    };

The implementation uses the same ideas we have applied throughout, in this case
we are just creating one of the iterators we developed earlier, pointing to the
first element in the sequence. The implementation of __end__ is very similar, and 
is provided in the example code.

For our __random_access_sequence__ we will also need to implement `size_impl`,
`value_at_impl` and `at_impl`.

[heading Enabling our type as an associative container]

In order for `example_struct` to serve as an associative container, 
we need to enable 3 lookup features, __at_key__, __value_at_key__ and __has_key__.
We also need to provide an implementation of the `is_associative` trait
so that our sequence can be correctly identified as an associative container. 

To implement `at_key_impl` we need to associate the `fields::age` and `fields::age`
types described in the __quick_start__ guide with the appropriate members of `example_struct`.
Our implementation is as follows:

    template<>
    struct at_key_impl<example::example_sequence_tag>
    {
        template<typename Sequence, typename Key>
        struct apply;

        template<typename Sequence>
        struct apply<Sequence, fields::name>
        {
            typedef typename mpl::if_<
                is_const<Sequence>,
                std::string const&,
                std::string&>::type type;

            static type
            call(Sequence& seq)
            {
                return seq.name;
            };
        };

        template<typename Sequence>
        struct apply<Sequence, fields::age>
        {
            typedef typename mpl::if_<
                is_const<Sequence>,
                int const&,
                int&>::type type;

            static type
            call(Sequence& seq)
            {
                return seq.age;
            };
        };
    };

Its all very similar to the implementations we've seen previously,
such as `deref_impl` and `value_of_impl`. Instead of identifying
the members by index or position, we are now selecting them using
the types `fields::name` and `fields::age`. The implementations of
`value_at_key_impl` and `has_key_impl` are equally straightforward,
and are provided in the example code, along with an implementation
of `is_associative_impl`.

[heading Summary]

We've now worked through the entire process for adding a new random
access sequence and we've also enabled our type to serve as an associative
container. The implementation was slightly longwinded, but followed
a simple repeating pattern.

The support for `std::pair`, __mpl__ sequences, and `boost::array` all
use the same approach, and provide additional examples of the approach
for a variety of types.

[endsect]