Re: [Boost-users] C++ Template Metaprogramming Book

From: David Abrahams (dave_at_[hidden])
Date: 2005-08-04 19:55:57

Eduardo Bezerra <edubez_at_[hidden]> writes:

> I'm reading the C++ Template Metaprogramming book and I'm not
> understanding the following code from page 65-66:
>
> template <class T>
> struct param_type
>   : mpl::eval_if <
>       typename boost::is_scalar<T>::type
>     , mpl::indentity<T>
>     , boost::add_reference<T const>
>   >
> {};
>
> template <class T>
> struct param_type
>   : mpl::eval_if <
>       boost::is_scalar<T>
> -->  ^^^^^^^^^^^^^^^^^^^^^^^ <--
>     , mpl::indentity<T>
>     , boost::add_reference<T const>
>   >
> {};
>
> After reading again and again the explanation I still don't get why we
> don't need to use typename ....::type anymore. The text explains that
> all Boost's integral metafunctions suply a nested ::value, but how it is
> used in this context ?

Well,

    typename boost::is_scalar<T>::type

is some

   mpl::bool_<X>

which happens to be a class that fulfils the requirements that eval_if
places on its first argument.

   http://www.boost.org/libs/mpl/doc/refmanual/eval-if.html

shows that the requirement is that the argument be a model of Integral
Constant.  If you click through to the requirements page
           
  http://www.boost.org/libs/mpl/doc/refmanual/integral-constant.html

for Integral Constant you can see what those are; (**) they include the
requirement that the class have a nested ::value.  However, we can drop
typename...::type because

  boost::is_scalar<T>

also happens to be a model of Integral constant with the same
properties as its nested ::type.  That might be achieved this way:

  template <class T>
  struct is_scalar_impl
  {  
    typedef mpl::bool_< ... > type; 
  };

  template <class T>
  struct is_scalar
   : is_scalar_impl<T>::type  // derive from mpl::bool_< ... >
  {
     typedef typename is_scalar_impl<T>::type type;
  };

It turns out that any class derived from mpl::bool_<X> is also a model
of Integral Constant with the same properties as mpl::bool_<X> has.

I hope this explains it.

(**) Aleksey, I think Integral Constant is much too strong a
requirement for the first argument to eval_if.  Most of those things
requiring a bool_ or similar really just need a nested ::value.

-- 
Dave Abrahams
Boost Consulting
www.boost-consulting.com