Header <boost/call_traits.hpp>

All of the contents of <boost/call_traits.hpp> are defined inside namespace boost.

The template class call_traits<T> encapsulates the "best" method to pass a parameter of some type T to or from a function, and consists of a collection of typedefs defined as in the table below. The purpose of call_traits is to ensure that problems like "references to references" never occur, and that parameters are passed in the most efficient manner possible (see examples). In each case if your existing practice is to use the type defined on the left, then replace it with the call_traits defined type on the right.

Note that for compilers that do not support either partial specialization or member templates, no benefit will occur from using call_traits: the call_traits defined types will always be the same as the existing practice in this case. In addition if only member templates and not partial template specialisation is support by the compiler (for example Visual C++ 6) then call_traits can not be used with array types (although it can be used to solve the reference to reference problem).

Notes:

1. If T is already reference type, then call_traits is defined such that references to references do not occur (requires partial specialization).
2. If T is an array type, then call_traits defines value_type as a "constant pointer to type" rather than an "array of type" (requires partial specialization). Note that if you are using value_type as a stored value then this will result in storing a "constant pointer to an array" rather than the array itself. This may or may not be a good thing depending upon what you actually need (in other words take care!).
3. If T is a small built in type or a pointer, then param_type is defined as T const, instead of T const&. This can improve the ability of the compiler to optimize loops in the body of the function if they depend upon the passed parameter, the semantics of the passed parameter is otherwise unchanged (requires partial specialization).

Copy constructibility

The following table defines which call_traits types can always be copy-constructed from which other types, those entries marked with a '?' are true only if and only if T is copy constructible:

If T is an assignable type the following assignments are possible:

Examples

The following table shows the effect that call_traits has on various types, the table assumes that the compiler supports partial specialization: if it doesn't then all types behave in the same way as the entry for "myclass", and call_traits can not be used with reference or array types.

Example 1:

The following class is a trivial class that stores some type T by value (see the call_traits_test.cpp file), the aim is to illustrate how each of the available call_traits typedefs may be used:

template <class T>
struct contained
{
   // define our typedefs first, arrays are stored by value
   // so value_type is not the same as result_type:
   typedef typename boost::call_traits<T>::param_type       param_type;
   typedef typename boost::call_traits<T>::reference        reference;
   typedef typename boost::call_traits<T>::const_reference  const_reference;
   typedef T                                                value_type;
   typedef typename boost::call_traits<T>::value_type       result_type;

   // stored value:
   value_type v_;
   
   // constructors:
   contained() {}
   contained(param_type p) : v_(p){}
   // return byval:
   result_type value() { return v_; }
   // return by_ref:
   reference get() { return v_; }
   const_reference const_get()const { return v_; }
   // pass value:
   void call(param_type p){}

};

Example 2 (the reference to reference problem):

Consider the definition of std::binder1st:

template <class Operation> 
class binder1st : 
   public unary_function<typename Operation::second_argument_type, typename Operation::result_type> 
{ 
protected: 
   Operation op; 
   typename Operation::first_argument_type value; 
public: 
   binder1st(const Operation& x, const typename Operation::first_argument_type& y); 
   typename Operation::result_type operator()(const typename Operation::second_argument_type& x) const; 
}; 

Now consider what happens in the relatively common case that the functor takes its second argument as a reference, that implies that Operation::second_argument_type is a reference type, operator() will now end up taking a reference to a reference as an argument, and that is not currently legal. The solution here is to modify operator() to use call_traits:

typename Operation::result_type operator()(typename call_traits<typename Operation::second_argument_type>::param_type x) const;

Now in the case that Operation::second_argument_type is a reference type, the argument is passed as a reference, and the no "reference to reference" occurs.

Example 3 (the make_pair problem):

If we pass the name of an array as one (or both) arguments to std::make_pair, then template argument deduction deduces the passed parameter as "const reference to array of T", this also applies to string literals (which are really array literals). Consequently instead of returning a pair of pointers, it tries to return a pair of arrays, and since an array type is not copy-constructible the code fails to compile. One solution is to explicitly cast the arguments to make_pair to pointers, but call_traits provides a better (i.e. automatic) solution (and one that works safely even in generic code where the cast might do the wrong thing):

template <class T1, class T2>
std::pair<
   typename boost::call_traits<T1>::value_type, 
   typename boost::call_traits<T2>::value_type> 
      make_pair(const T1& t1, const T2& t2)
{
   return std::pair<
      typename boost::call_traits<T1>::value_type, 
      typename boost::call_traits<T2>::value_type>(t1, t2);
}

Here, the deduced argument types will be automatically degraded to pointers if the deduced types are arrays, similar situations occur in the standard binders and adapters: in principle in any function that "wraps" a temporary whose type is deduced. Note that the function arguments to make_pair are not expressed in terms of call_traits: doing so would prevent template argument deduction from functioning.

Example 4 (optimising fill):

The call_traits template will "optimize" the passing of a small built-in type as a function parameter, this mainly has an effect when the parameter is used within a loop body. In the following example (see fill_example.cpp), a version of std::fill is optimized in two ways: if the type passed is a single byte built-in type then std::memset is used to effect the fill, otherwise a conventional C++ implemention is used, but with the passed parameter "optimized" using call_traits:

namespace detail{

template <bool opt>
struct filler
{
   template <typename I, typename T>
   static void do_fill(I first, I last, typename boost::call_traits<T>::param_type val)
   {
      while(first != last)
      {
         *first = val;
         ++first;
      }
   }
};

template <>
struct filler<true>
{
   template <typename I, typename T>
   static void do_fill(I first, I last, T val)
   {
      memset(first, val, last-first);
   }
};

}

template <class I, class T>
inline void fill(I first, I last, const T& val)
{
   enum{ can_opt = boost::is_pointer<I>::value
                   && boost::is_arithmetic<T>::value
                   && (sizeof(T) == 1) };
   typedef detail::filler<can_opt> filler_t;
   filler_t::template do_fill<I,T>(first, last, val);
}

Footnote: the reason that this is "optimal" for small built-in types is that with the value passed as "T const" instead of "const T&" the compiler is able to tell both that the value is constant and that it is free of aliases. With this information the compiler is able to cache the passed value in a register, unroll the loop, or use explicitly parallel instructions: if any of these are supported. Exactly how much mileage you will get from this depends upon your compiler - we could really use some accurate benchmarking software as part of boost for cases like this.

Note that the function arguments to fill are not expressed in terms of call_traits: doing so would prevent template argument deduction from functioning. Instead fill acts as a "thin wrapper" that is there to perform template argument deduction, the compiler will optimise away the call to fill all together, replacing it with the call to filler<>::do_fill, which does use call_traits.

Rationale

The following notes are intended to briefly describe the rational behind choices made in call_traits.

All user-defined types follow "existing practice" and need no comment.

Small built-in types (what the standard calls fundamental types [3.9.1]) differ from existing practice only in the param_type typedef. In this case passing "T const" is compatible with existing practice, but may improve performance in some cases (see Example 4), in any case this should never be any worse than existing practice.

Pointers follow the same rational as small built-in types.

For reference types the rational follows Example 2 - references to references are not allowed, so the call_traits members must be defined such that these problems do not occur. There is a proposal to modify the language such that "a reference to a reference is a reference" (issue #106, submitted by Bjarne Stroustrup), call_traits<T>::value_type and call_traits<T>::param_type both provide the same effect as that proposal, without the need for a language change (in other words it's a workaround).

For array types, a function that takes an array as an argument will degrade the array type to a pointer type: this means that the type of the actual parameter is different from its declared type, something that can cause endless problems in template code that relies on the declared type of a parameter. For example:

template <class T>
struct A
{
   void foo(T t);
};

In this case if we instantiate A<int[2]> then the declared type of the parameter passed to member function foo is int[2], but it's actual type is const int*, if we try to use the type T within the function body, then there is a strong likelyhood that our code will not compile:

template <class T>
void A<T>::foo(T t)
{
   T dup(t); // doesn't compile for case that T is an array.
}

By using call_traits the degradation from array to pointer is explicit, and the type of the parameter is the same as it's declared type:

template <class T>
struct A
{
   void foo(typename call_traits<T>::value_type t);
};

template <class T>
void A<T>::foo(typename call_traits<T>::value_type t)
{
   typename call_traits<T>::value_type dup(t); // OK even if T is an array type.
}

For value_type (return by value), again only a pointer may be returned, not a copy of the whole array, and again call_traits makes the degradation explicit. The value_type member is useful whenever an array must be explicitly degraded to a pointer - Example 3 provides the test case (Footnote: the array specialisation for call_traits is the least well understood of all the call_traits specialisations, if the given semantics cause specific problems for you, or don't solve a particular array-related problem, then I would be interested to hear about it. Most people though will probably never need to use this specialisation).

Revised 01 September 2000

Copyright 2000 Steve Cleary, Beman Dawes, Howard Hinnant and John Maddock.
Use, modification and distribution are subject to the Boost Software License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt ).