Fast C++ Delegate: Boost.Function 'drop-in' replacement and multicast

JaeWook Choi

4.86/5 (51 votes)

1 Jun 200733 min read

1.9K

An article on the implementation of a fast C++ delegate with many advanced features.

Introduction

There have been several C++ delegates which declared themselves as a 'fast' or 'fastest' delegate, while Boost.Function and its siblings, Boost.Bind and Boost.Mem_fn, were adopted as part of C++ Standards Committee's Library Technical Report (TR1). So, what are those called 'fast' or 'fastest' delegates and how much 'faster' are they than Boost.Function?

The prefix 'fast' in the term 'fast(est) delegate' means either 'fast' invocation or 'fast' copy, or both. But, I believe, what is really an issue between the two when using the 'non-fast' Boost.Function is more likely its awful copy performance. This is due to the expensive heap memory allocation that is required to store the member function and the bound object on which member function call is made. So, 'fast' delegate often refers to a delegate that does not require heap memory allocation for storing the member function and the bound object. In C++, as an object oriented programming paradigm, use of delegate or closure for the member function and the bound object is one of the most frequently occurring practices. Thus, a 'fast' delegate can 'boost' the performance by far in some situations.

The following four graphs are the result of the invocation speed comparison among three fast delegates and Boost.Function in the various function call scenarios. See the '%FD_ROOT%/benchmark' for details.

Don's Fastest Delegate
Sergey's Fast Delegate
Jae's Fast Delegate
<a href="http://www.boost.org/doc/html/function.html">Boost.Function</a>

The following two graphs are the result of the copy speed comparison among three fast delegates and Boost.Function. For a bound member function call, it was found that Boost.Function can take 150 times longer than the fastest. The result may vary based on the benchmark platform and environment, but it is obvious that the copy performance of Boost.Function is not acceptable in certain cases.

In spite of the prominent speed boost of the fast delegates in specific cases, it is not comfortable for many programmers to switch and start using the fast delegates. This is because their features are not as rich as those which Boost.Function and its siblings provide, and we are already accustomed to using Boosts. These fast delegates support very limited types of callable entities to store and mostly do not support the storing of a function object, which is another frequently occurring practice in C++.

I had implemented a fast delegate some time ago, but it was not as fast as other fast delegates nor as C++ Standard compliant as I thought it was. I actually patched it to be C++ Standard compliant later. This is the second version, but it is completely re-implemented from the scratch. The old version is obsolete. It is another 'fast' delegate, but it is also a Boost.Function 'drop-in' replacement and more. I say 'more' because it supports the multicast feature which is missing in the most of C++ delegates currently available. It is not like an ancillary class to support multicast, but one class instance acts as single cast and multicast on demand, without any runtime performance penalty. FD.Delegate can be thought of as an aggregation of Boost.Function and its siblings (Boost.Bind and Boost.Mem_fn) plus some features from Boost.Signals. See the 'Delegates Comparison Chart' at the end of the article for particulars.

Using the code

As stated previously, FD.Delegate is a Boost.Function 'drop-in' replacement. So it is reasonable to refer to the online documentation of Boost.Function and especially Boost.Function tutorial for features of FD.Delegate. Just make sure to add '%FD_ROOT%/include' as a system include directory.

- Example #1 from Boost.Function.

#include <iostream>
#include <fd/delegate.hpp>

struct int_div
{
    float operator()(int x, int y) const { return ((float)x)/y; };
};

int main()
{
    fd::delegate<float (int, int)> f;
    f = int_div();

    std::cout << f(5, 3) << std::endl; // 1.66667

    return 0;
}

- Example #2 from Boost.Function.

#include <iostream>
#include <fd/delegate.hpp>

void do_sum_avg(int values[], int n, int& sum, float& avg)
{
    sum = 0;
    for (int i = 0; i < n; i++)
        sum += values[i];
    avg = (float)sum / n;
}

int main()
{
    // The second parameter should be int[], but some compilers (e.g., GCC)
    // complain about this
    fd::delegate<void (int*, int, int&, float&)> sum_avg;

    sum_avg = &do_sum_avg;

    int values[5] = { 1, 1, 2, 3, 5 };
    int sum;
    float avg;
    sum_avg(values, 5, sum, avg);

    std::cout << "sum = " << sum << std::endl;
    std::cout << "avg = " << avg << std::endl;
    return 0;
}

FD.Delegate supports multicast and uses C#'s multicast syntax, operator += and operator -=.

#include <iostream>
#include <fd/delegate/delegate2.hpp>

struct print_sum
{
    void operator()(int x, int y) const { std::cout << x+y << std::endl; }
};

struct print_product
{
    void operator()(int x, int y) const { std::cout << x*y << std::endl; }
};

int main()
{
    fd::delegate2<void, int, int> dg;

    dg += print_sum();
    dg += print_product();

    dg(3, 5); // prints 8 and 15

    return 0;
}

While a function pointer is equality comparable, a function object is not quite determinant at compile-time, whether equality comparable or not. This fact makes operator -= pretty much useless for removing a function object from multicast. FD.Delegate has add() and remove() member function pairs to remedy the issue. add() returns an instance of fd::multicast::token which can be used to remove the added delegate(s).

#include <iostream>
#include <fd/delegate.hpp>
#include <cassert>

struct print_sum
{
    void operator()(int x, int y) const { std::cout << x+y << std::endl; }
};

struct print_product
{
    void operator()(int x, int y) const { std::cout << x*y << std::endl; }
};

struct print_difference
{
    void operator()(int x, int y) const { std::cout << x-y << std::endl; }
};

struct print_quotient
{
    void operator()(int x, int y) const { std::cout << x/-y << std::endl; }
};

int main()
{
    fd::delegate2<void, int, int> dg;

    dg += print_sum();
    dg += print_product();

    dg(3, 5);

    fd::multicast::token print_diff_tok = dg.add(print_difference());

    // print_diff_tok is still connected to dg
    assert(print_diff_tok.valid());

    dg(5, 3); // prints 8, 15, and 2

    print_diff_tok.remove(); // remove the print_difference delegate

    dg(5, 3);  // now prints 8 and 15, but not the difference

    assert(!print_diff_tok.valid()); // not connected anymore
    {
        fd::multicast::scoped_token t = dg.add(print_quotient());
        dg(5, 3); // prints 8, 15, and 1
    } // t falls out of scope, so print_quotient is not a member of dg

    dg(5, 3); // prints 8 and 15

    return 0;
}

It has been one of the main concerns for a multicast delegate how to manage multiple return values. Combiner interface of Boost.Signals has been adopted, but has slightly different usage and syntax. The type of the combiner interface is not a part of FD.Delegate type, although it is for Boost.Signals, as the form of the template parameter when declaring a signal variable. Instead, FD.Delegate has a special function call operator which takes the instance of the combiner interface as the last function call argument.

#include <algorithm>
#include <iostream>
#include <fd/delegate.hpp>

template<typename T>
struct maximum
{
    typedef T result_type;

    template<typename InputIterator>
    T operator()(InputIterator first, InputIterator last) const
    {
        if(first == last)
            throw std::runtime_error("Cannot compute maximum of zero elements!");
        return *std::max_element(first, last);
  }
};

template<typename Container>
struct aggregate_values
{
    typedef Container result_type;

    template<typename InputIterator>
    Container operator()(InputIterator first, InputIterator last) const
    {
        return Container(first, last);
    }
};

int main()
{
    fd::delegate2<int, int, int> dg_max;
    dg_max += std::plus<int>();
    dg_max += std::multiplies<int>();
    dg_max += std::minus<int>();
    dg_max += std::divides<int>();

    std::cout << dg_max(5, 3, maximum<int>()) << std::endl; // prints 15

    std::vector<int> vec_result = dg_max(5, 3, 
        aggregate_values<std::vector<int> >());
    assert(vec_result.size() == 4);

    std::cout << vec_result[0] << std::endl; // prints 8
    std::cout << vec_result[1] << std::endl; // prints 15
    std::cout << vec_result[2] << std::endl; // prints 2
    std::cout << vec_result[3] << std::endl; // prints 0

    return 0;
}

Under the hood

Part A: storing a function pointer for later invocation without requiring heap memory allocation.

According to C++ standards, a function pointer -- both free function pointer and member function pointer -- cannot be converted or stored into a void *. A function pointer may be converted into a function pointer of a different type signature, however, the result of such conversion cannot be used; it can only be converted back. The size of the member function varies over the different platforms from 4 bytes to 16 bytes. To avoid heap allocation to store the member function, some well-known template meta programming techniques have been adapted. These permit a member function pointer whose size is less than or equal to the size of the predefined generic member function pointer to be stored without heap memory allocation. The stored generic member function pointer is restored back to its original member function type before use.

typedef void generic_fxn();

class alignment_dummy_base1 { };
class alignment_dummy_base2 { };

class alignment_dummy_s : alignment_dummy_base1 { };                         
    // single inheritance.
class alignment_dummy_m : alignment_dummy_base1, alignment_dummy_base2 { };  
    // multiple inheritance.
class alignment_dummy_v : virtual alignment_dummy_base1 { };                 
    // virtual inheritance.
class alignment_dummy_u;                                                     
    // unknown (incomplete).

typedef void (alignment_dummy_s::*mfn_ptr_s)();  
    // member function pointer of single inheritance class.
typedef void (alignment_dummy_m::*mfn_ptr_m)();  
    // member function pointer of multiple inheritance class.
typedef void (alignment_dummy_v::*mfn_ptr_v)();  
    // member function pointer of virtual inheritance class.
typedef void (alignment_dummy_u::*mfn_ptr_u)();  
    // member function pointer of unknown (incomplete) class.

typedef void (alignment_dummy_m::*generic_mfn_ptr)();

union max_align_for_funtion_pointer
{
    void const * dummy_vp;
    generic_fxn * dummy_fp;
    boost::ct_if<( sizeof( generic_mfn_ptr ) < sizeof( mfn_ptr_s ) ),
      generic_mfn_ptr, mfn_ptr_s>::type dummy_mfp1;
    boost::ct_if<( sizeof( generic_mfn_ptr ) < sizeof( mfn_ptr_m ) ),
      generic_mfn_ptr, mfn_ptr_m>::type dummy_mfp2;
    boost::ct_if<( sizeof( generic_mfn_ptr ) < sizeof( mfn_ptr_v ) ),
      generic_mfn_ptr, mfn_ptr_v>::type dummy_mfp3;
    boost::ct_if<( sizeof( generic_mfn_ptr ) < sizeof( mfn_ptr_u ) ),
      generic_mfn_ptr, mfn_ptr_u>::type dummy_mfp4;
};

BOOST_STATIC_CONSTANT( unsigned, 
    any_fxn_size = sizeof( max_align_for_funtion_pointer ) );

union any_fxn_pointer
{
    void const * obj_ptr;
    generic_fxn * fxn_ptr;
    generic_mfn_ptr mfn_ptr;
    max_align_for_funtion_pointer m_;
};

A member function pointer whose size is less than or equal to any_fxn_size is stored into any_fxn_pointer. any_fxn_pointer is implemented to make it able to store one out of three different pointer types -- a void data pointer, a function pointer, or a member function pointer -- whose size is less than a function pointer to the member of a multiple inherited class. Only one pointer type is stored at one specific time. Care has been taken regarding the misalignment issue, which may cause undefined behavior according to the C++ standard, by applying the specialized version of the well-known max alignment union trickery.

void hello(int, float) { }
typedef void (*MyFxn)(int, float);

struct foobar
{
    void foo(int, float) { }
};
typedef void (foobar::*MyMfn)(int, float);


void test1(any_fxn_pointer any)
{
    ( *reinterpret_cast<MyFxn>( any.fxn_ptr ) )( 1, 1.0f );
}

void test2(any_fxn_pointer any, foobar * pfb)
{
    ( pfb->*reinterpret_cast<MyMfn>( any.mfn_ptr ) )( 1, 1.0f );
}

void main()
{
    any_fxn_pointer any;
    any.fxn_ptr = reinterpret_cast<generic_fxn *>( &hello );

    test1( any );

    foobar fb;
    any.mfn_ptr = reinterpret_cast<generic_mfn_ptr>( &foobar::foo );
    test2( any, &fb );
}

When the size of a member function pointer is greater than any_fxn_size, takes for an example when storing a member function pointer to virtual inherited class in MSVC, it is stored by allocating heap memory in the same way that Boost.Function does, as a non-fast delegate. However, in real world practices, virtual inheritance is rarely used.

template<typename U, typename T>
void bind(UR (U::*fxn)(int, float), T t)
{
    struct select_stub
    {
        typedef void (U::*TFxn)(int, float);
        typedef typename boost::ct_if<( sizeof( TFxn ) <= any_fxn_size ),
        typename impl_class::fast_mfn_delegate,
        typename impl_class::normal_mfn_delegate
        >::type type;
    };

    select_stub::type::bind( *this, fxn, t, );
}

Part B: using any_fxn_pointer

any_fxn_pointer is used for storing an arbitrary function pointer with the class template. This is done in order to erase the type while storing the function and to restore the original type safely when required later. Sergey Ryazanov demonstrated in his article that a C++ standard compliant fast delegate for member function can be implemented using the class template with a non-type member function template parameter. The sample below shows a rough idea of how it had been implemented.

class delegate
{
    typedef void (*invoke_stub)(void const *, int);

    void const * obj_ptr_;
    invoke_stub stub_ptr_;

    template<typename T, void (T::*Fxn)(int)>
    struct mem_fn_stub
    {
        static void invoke(void const * obj_ptr, int a0)
        {
            T * obj = static_cast<T *>( const_cast<void *>( obj_ptr ) );
            (obj->*Fxn)( a0 );
        }
    };

    template<typename T, void (T::*Fxn)(int) const>
    struct mem_fn_const_stub
    {
        static void invoke(void const * obj_ptr, int a0)
        {
            T const * obj = static_cast<T const *>( obj_ptr );
            (obj->*Fxn)( a0 );
        }
    };

    template<void (*Fxn)(int)>
    struct function_stub
    {
        static void invoke(void const *, int a0)
        {
            (*Fxn)( a0 );
        }
    };

public:
    delegate() : obj_ptr_( 0 ), stub_ptr_( 0 ) { }

    template<typename T, void (T::*Fxn)(int)>
    void from_function(T * obj)
    {
        obj_ptr_ = const_cast<T const *>( obj );
        stub_ptr_ = &mem_fn_stub<T, Fxn>::invoke;
    }

    template<typename T, void (T::*Fxn)(int) const>
    void from_function(T const * obj)
    {
        obj_ptr_ = obj;
        stub_ptr_ = &mem_fn_const_stub<T, Fxn>::invoke;
    }

    template<void (*Fxn)(int)>
    void from_function()
    {
        obj_ptr_ = 0;
        stub_ptr_ = &function_stub<Fxn>::invoke;
    }

    void operator ()(int a0) const
    {
        ( *stub_ptr_ )( obj_ptr_, a0 );
    }
};

Even though passing a member function as a non-type template parameter is a legitimate C++ feature -- somewhat outdated, but still widely used -- compilers do not support it. The real problem with Sergey's implementation of using the member function as non-type template parameter is not the lack of support from those outdated compilers, but rather its awful syntax. This is due to the fact that non-type template parameters cannot participate in template argument deduction from the function arguments provided. Therefore it must be explicitly specified all the time.

struct foobar
{
    void foo(int) { }
    void bar(int) const { }
};

void hello(int) { }

void main()
{
    foobar fb;
    foobar * pfb = &fb;

    delegate dg;

    dg.from_function<foobar, &foobar::foo>( pfb );
    dg( 1 ); // (pfb->*&foobar::foo)( 1 );

    dg.from_function<foobar const, &foobar::bar>( pfb );
    dg( 1 ); // (pfb->*&foobar::bar)( 1 );

    dg.from_function<&hello>();
    dg( 1 ); // hello( 1 );
}

It is a really bad idea in practice that you should provide template arguments explicitly every time. Using any_fxn_pointer introduced previously, we can significantly improve the syntax of the usage and, in turn, the degree of convenience. We just need to add an any_fxn_pointer as a member of the delegate class to store the address of the function of interest.

class delegate
{
    typedef void (*invoke_stub)(void const *, any_fxn_pointer, int);

    void const * obj_ptr_;
    any_fxn_pointer any_;
    invoke_stub stub_ptr_;

    template<typename T, typename TFxn>
    struct mem_fn_stub
    {
        static void invoke(void const * obj_ptr, any_fxn_pointer any, int a0)
        {
            T * obj = static_cast<T *>( const_cast<void *>( obj_ptr ) );
            (obj->*reinterpret_cast<TFxn>( any.mfn_ptr ) )( a0 );
        }
    };

    template<typename TFxn>
    struct function_stub
    {
        static void invoke(void const *, any_fxn_pointer any, int a0)
        {
            (*reinterpret_cast<TFxn>( any.fxn_ptr ) )( a0 );
        }
    };

public:
    delegate() : obj_ptr_( 0 ), any(), stub_ptr_( 0 ) { }

    template<typename T>
    void from_function(void (T::*fxn)(int ), T * obj)
    {
        typedef void (T::*TFxn)(int);

        obj_ptr_ = const_cast<T const *>( obj );
        any_.mfn_ptr = reinterpret_cast<generic_mfn_ptr>( fxn );
        stub_ptr_ = &mem_fn_stub<T, TFxn>::invoke;
    }

    template<typename T>
    void from_function(void (T::*fxn)(int) const, T const * obj)
    {
        typedef void (T::*TFxn)(int) const;

        obj_ptr_ = obj;
        any_.mfn_ptr = reinterpret_cast<generic_mfn_ptr>( fxn );
        stub_ptr_ = &mem_fn_stub<T const, TFxn>::invoke;
    }

    void from_function(void (*fxn)(int))
    {
        typedef void (*TFxn)(int);

        obj_ptr_ = 0;
        any_.fxn_ptr = reinterpret_cast<generic_fxn *>( fxn );
        stub_ptr_ = &function_stub<TFxn>::invoke;
    }

    void operator ()(int a0) const
    {
        ( *stub_ptr_ )( obj_ptr_, any_, a0 );
    }
}; // delegate

Not even this works for those outdated compilers that do not support member functions as non-type template parameters. It becomes more intuitive syntax and thus easier to use, since function overloading and automatic template argument deduction is now applicable.

struct foobar
{
    void foo(int) { }
    void bar(int) const { }
};

void hello(int) { }

void main()
{
    foobar fb;
    foobar * pfb = &fb;

    delegate dg;

    dg.from_function( &foobar::foo, pfb );
    dg( 1 ); // (pfb->*&foobar::foo)( 1 );

    dg.from_function( &foobar::bar, pfb );
    dg( 1 ); // (pfb->*&foobar::bar)( 1 );

    dg.from_funcion( &hello );
    dg( 1 ); // hello( 1 );
}

Part C: using function reference tables to support rich features

One of interesting features of Boost.Function is its ability to store a member function along with the bound object in various forms: a) a pointer or a reference to the bound object of the type on which the member function call is made, or b) an instance of an arbitrary smart pointer to the bound object. As a matter of fact, it is precise to say that this feature is of Boost.Mem_fn and Boost.Bind. Boost.Function has a 'manager' member and defines enumerate tags of so-called 'functor_manager_operation_type'.

enum functor_manager_operation_type
{
    clone_functor_tag,
    destroy_functor_tag,
    check_functor_type_tag
};

It also defines several tags to distinguish among different types of functions.

struct function_ptr_tag {};
struct function_obj_tag {};
struct member_ptr_tag {};
struct function_obj_ref_tag {};
struct stateless_function_obj_tag {};

This is a traditional tag dispatching technique that uses function overloading to dispatch based on properties of type. This sort of tag dispatching works quite reasonably, but with slightly over complicated construct detail. This is because an implementation for a specific function type is usually scattered here and there. If you have ever tried to look into the Boost.Function detail, you know what I mean. However, there exists one very neat and elegant alternative. It was initially introduced by Chris Diggings and Jonathan Turkanis in their series of articles about BIL (Boost.Interface.Libarary). We can start off by declaring a function reference table that contains a set of function pointers. These function pointers determine the behavioral requirement or concepts of the delegate that we want to design.

class delegate
{
    typedef void generic_fxn();
    // Function reference table.
    struct fxn_table
    {
        void invoke(void const *, any_fxn_pointer, int);
        void copy_obj(void cons **, void const *);
        void delete_obj(void const *);
        bool is_empty();
    };
 
    void const * obj_ptr_;
    any_fxn_pointer any_;
    fxn_table const * tbl_ptr_;

Then we define class templates that implement all of the entries of function reference table respectively.

template<typename T, typename TFxn>
struct mem_fn_stub
{
    static void init(
        void const ** obj_pptr, any_fxn_pointer & any, T * obj, TFxn fxn)
    {
        *obj_pptr = obj;
        any.mfn_ptr = reinterpret_cast<generic_mfn_ptr>( fxn );
    }

    static void invoke(void const * obj_ptr, any_fxn_pointer any, int a0)
    {
        T * obj = static_cast<T *>( const_cast<void *>( obj_ptr ) );
        (obj->*reinterpret_cast<TFxn>( any.mfn_ptr ) )( a0 );
    }

    static void copy_obj(void const ** obj_pptr_dest, void const * obj_ptr_src)
    {
        *obj_pptr_dest = obj_ptr_src; // Simply copies between pointers
    }

    static void delete_obj(void const *)
    {
        // Do nothing.
    }

    static bool is_empty()
    {
        return false;
    }

    static fxn_table const * get_table()
    {
        static fxn_table const static_table = { &invoke, &copy_obj,
            &delete_obj, &is_empty, };
        return &static_table;
    }
};

template<typename T, typename TFxn>
struct mem_fn_obj_stub
{
    static void init(
        void const ** obj_pptr, any_fxn_pointer & any, T * obj, TFxn fxn)
    {
        *obj_pptr = new T( *obj );
        any.mfn_ptr = reinterpret_cast<generic_mfn_ptr>( fxn );
    }

    static void invoke(void const * obj_ptr, any_fxn_pointer any, int a0)
    {
        T * obj = static_cast<T *>( const_cast<void *>( obj_ptr ) );
        ( get_pointer(*obj)->*reinterpret_cast<TFxn>( any.mfn_ptr ) )( a0 );
    }

    static void copy_obj(void const ** obj_pptr_dest, void const * obj_ptr_src)
    {
        // Clones the pointed object.
        *obj_pptr_dest = new T( *static_cast<T const *>( obj_ptr_src ) );
    }

    static void delete_obj(void const * obj_ptr)
    {
        // Deletes the pointed object.
        delete static_cast<T const *>( obj_ptr );
    }

    static bool is_empty()
    {
        return false;
    }

    static fxn_table const * get_table()
    {
        static fxn_table const static_table = { &invoke, &copy_obj,
            &delete_obj, &is_empty, };
        return &static_table;
    }
};

template<typename TFxn>
struct function_stub
{
    static void init(void const ** obj_pptr, any_fxn_pointer & any, TFxn fxn)
    {
        *obj_pptr = 0;
        any.fxn_ptr = reinterpret_cast<generic_fxn *>( fxn );
    }

    static void invoke(void const *, any_fxn_pointer any, int a0)
    {
        (*reinterpret_cast<TFxn>( any.fxn_ptr ) )( a0 );
    }

    static void copy_obj(void const ** obj_pptr_dest, void const * obj_ptr_src)
    {
        *obj_pptr_dest = obj_ptr_src; // Simply copies between pointers
    }

    static void delete_obj(void const *)
    {
        // Do nothing.
    }

    static bool is_empty()
    {
        return false;
    }

    static fxn_table const * get_table()
    {
        static fxn_table const static_table = { &invoke, &copy_obj,
            &delete_obj, &is_empty, };
        return &static_table;
    }
};

We have defined three class templates above. Actually, a class template is not necessary if no extra template parameter is required to implement the behavioral requirement of the specific function type. A normal class will be just fine in such case.

struct null_stub
{
    static void invoke(void const *, any_fxn_pointer, int)
    {
        throw bad_function_call();
    }

    static void copy_obj(void const ** obj_pptr_dest, void const *)
    {
        *obj_pptr_dest = 0;
    }

    static void delete_obj(void const *)
    {
        // Do nothing.
    }

    static bool is_empty()
    {
        return true;
    }

    static fxn_table const * get_table()
    {
        static fxn_table const static_table = { &invoke, &copy_obj,
            &delete_obj, &is_empty, };
        return &static_table;
    }
};

We can now implement delegate class. We only use entries of the function reference table we declared in the first place through the pointer to the function reference table, tbl_ptr_, to implement member functions of a 'generic' delegate itself.

public:
    delegate()
        : obj_ptr_( 0 ), any_(), tbl_ptr_( null_stub::get_table() ) { }

    delegate(delegate const & other)
        : obj_ptr_( other.obj_ptr_ ), any_( other.any_ ), 
        tbl_ptr_( other.tbl_ptr_ )
    {
        if( other.tbl_ptr_ )
            other.tbl_ptr_->copy_obj( &obj_ptr_, other.obj_ptr_ );
    }

    ~delegate()
    {
        if( tbl_ptr_ )
            tbl_ptr_->delete_obj( obj_ptr_ );
    }

    void swap(delegate & other)
    {
        std::swap( obj_ptr_, other.obj_ptr_ );
        std::swap( any_, other.any_ );
        std::swap( tbl_ptr_, other.tbl_ptr_ );
    }

    bool empty() const
    {
        return tbl_ptr_->is_empty();
    }

    delegate & operator =(delegate const & other)
    {
        if( this != &other )
            delegate( other ).swap( *this );
        return *this;
    }

    void reset()
    {
        delegate().swap( *this );
    }

    void operator ()(int a0) const
    {
        tbl_ptr->invoke( obj_ptr_, any_, a0 );
    }

Defining a class to represent 'null' has several benefits over initializing the pointer to the function reference table, tbl_ptr_, to a zero value. It is not necessary to check nullness of delegate at runtime. If the pointer to the function reference table were to be assigned to zero to represent null delegate instead, the function call operator should have contained an if-clause to check the nullness. Most of all, it should also have had an exception statement to throw when the call had been made on the empty delegate, which seemed quite an unreasonable penalty.

Since nullness of delegate is determined at compile time by introducing 'null' class, null_stub, we can improve the performance by not checking the nullness at runtime and by not having throw statement if it is not a 'null' delegate. Finally, we can complete our new delegate class by adding interface member functions to support storing from various function types.

  template<typename T, typename U>
  void from_function(void (U::*fxn)(int), T obj)
  {
      reset();

      typedef void (U::*TFxn)(int);

      struct select_stub
      {
          typedef typename boost::ct_if<
              (::boost::is_pointer<T>::value),
              mem_fn_stub<typename boost::remove_pointer<T>::type, TFxn>,
              mem_fn_obj_stub<T, TFxn>
              > type;
      };

      select_stub::type::init( &obj_ptr_, any_, obj, fxn );
      tbl_ptr_ = select_stub::type::get_table();
  }

  template<typename T, typename U>
  void from_function(void (U::*fxn)(int) const, T obj)
  {
      reset();

      typedef void (U::*TFxn)(int) const;

      struct select_stub
      {
          typedef typename boost::ct_if<
              (::boost::is_pointer<T>::value),
              mem_fn_stub<typename boost::remove_pointer<T>::type, TFxn>,
              mem_fn_obj_stub<T, TFxn>
              > type;
      };

      select_stub::type::init( &obj_ptr_, any_, obj, fxn );
      tbl_ptr_ = select_stub::type::get_table();
  }

  void from_function(void (*fxn)(int))
  {
      reset();

      typedef void (*TFxn)(int);

      function_stub<TFxn>::init( &obj_ptr_, any_, fxn );
      tbl_ptr_ = function_stub<TFxn>::get_table();
  }

}; // class delegate

We just added smart pointer support into our delegate. We can now use any kind of smart pointer to bind the target object on which the member function call is made, as long as the get_pointer() overload for the smart pointer is provided in the visible namespace scope.

struct foobar
{
    void foo(int) { }
    void bar(int) const { }
};

template<typename T>
T * get_pointer(boost::shared_ptr<T> const & t)
{
    return t.get();
}

void main()
{
    boost::shared_ptr<foobar> spfb(new foobar);

    delegate dg;

    dg.from_function( &foobar::foo, spfb );
    dg( 1 ); // (get_pointer(spfb)->*&foobar::foo)( 1 );
}

As demonstrated above, it is quite simple to add new function type support into our delegate. First, it determines all the behavioral requirements according to the characteristic of the function type that we are interested in storing. Then it adds a class template and starts defining all the entries of function reference table for the new function type per requirements. Finally, it adds an interface function overload -- from_function() member function in the example above -- to support the newly added function type.

If we want to add more behavioral requirements or concepts into the delegate, we can add new entries into the function reference table to fulfill the purpose. Of course, all of the existing class templates already defined for some other function types must implement newly added entries accordingly. Otherwise, it will assert a compile error. Be aware that even if these newly added entries might not apply for some of class templates already defined for certain function types, they still have to define empty entries, which may do nothing and thus possibly be optimized away by the compiler.

As easily seen from the example above, all entries in a function reference table are declared in a type neutral form. This means that it may be determined by the function call type signature, but nothing else. Also, we can see that void pointers are passed over as their arguments, and the specific type information is given as the template parameters of the class template that implements all of the entries of the function reference table. Those void pointers passed in and out come alive and restore themselves back to their original identities by the aid of casting to one of the template parameters of class template. However, it will assuredly erase the type information only during its storage life span and convert back to the right identity when required. This is the key point of how the different set of behavioral requirements of a delegate for a specific function type is implemented in the type-safe manner, but with type-neutral persistency.

Implementation details are to be defined in one location, in a class template for a specific function type, so it has better readability and manageability than the tag dispatching technique. It is like a 'strategy pattern' to make it possible to encapsulate the interchangeable algorithm of delegate for various function types. Actually, a function reference table is an interface and the technique can be thought of as a static typing or static binding version of the virtual table feature in C++.

Every individual class template that implements all of the entries of a function reference table according to their own requirements for the specific function type are orthogonal to each other. This means that we don't need to concern ourselves with how a function object is stored into the delegate while designing a class template to store a member function, and vice versa. So what does this mean? I can now implement and support a multicast delegate much easier in one single delegate class definition. There will be absolutely no degradation in runtime performance nor will there be a size increase of non-multicast delegate. This is because they -- that is, the multicast delegate and non-multicast delegate -- are treated as completely different beasts of the same name.

Therefore it becomes very important how robustly we declare entries of a function reference table to generalize the behavioral requirements of the generic delegate that we are interested in storing from the given function type, as well as how to categorize these function types based on their own unique traits and requirements.

Selecting a proper function reference table for the specific function type takes place in the implementation of the interface function overloads of delegate class. It usually involves template meta programming along with type traits to determine or to transform from a function type. Unfortunately, some type traits I used while implementing the selection structure in the interface function overloads only work in places based on the template partial specialization feature of C++ (boost::remove_pointer<> for example). This is why FD.Delegate does not work on some outdated compilers that have no support or broken support for this feature. If I were to support only a limited feature set, like Don's fastest delegate or Sergey's fast delegate, I could have supported many outdated compilers as well. However, I wanted it to be a 'drop-in' replacement for Boost.Function, so this was not an option for me.

D. Categorizing function types.

There are three dominant callable entities in the C++ world.

Free function
Member function
Function object

These basic callable entities are strained off more refined species depending on how their bound object or function object is stored and managed internally. See the table below.

Function type	Category	Description

FT01.	1.	Free function (including static member function)

FT02.	2-a.	Member function on a pointer to an object of the type that the member function belongs to.
FT03.	2-b.	Member function on a reference to an object of the type that the member function belongs to.
FT04.	2-c.	Member function on a copy of an object of either the type that the member function belongs to or any type that supports `get_pointer()` overload (i.e. smart pointers).

FT05.	2-d.	Member function bound with a pointer or a reference to an object of the type that the member function belongs to.
FT06.	2-e.	Member function bound with a copy of an object of any type that supports `get_pointer()` overload (i.e. smart pointers).

FT07.	3-a.	A pointer or a reference to a function object.
FT08.	3-b.	A copy of a function object.
FT09.	3-c.	Stateless function object.

FT10.		Empty delegate.

FT11.		Multicast delegate.

In the previous delegate example, we added supports for function type FT01, FT05, FT06 and FT10. To make it easier to understand what these function types are, see the example as illustrated below.

struct foobar
{
    int id_;

    void foo(int) { }
    static void bar(int) { }

    void operator ()(int) const { }
};

void hello(int) { }

struct stateless
{
    void operator ()(int) const { }
};

void main()
{
    delegate<void (int)> dg1;

    foobar fb;
    foobar * pfb = &fb; 
    boost::shared_ptr<foobar> spfb( new foobar );

    dg1 = &hello;                               // FT01
    dg1( 1 );                                   // hello( 1 );

    dg1 = &foobar::bar;                         // FT01
    dg1( 1 );                                   // foobar::bar( 1 );

    delegate<void (foobar *, int)> dg2;
    dg2 = &foobar::foo;                         // FT02
    dg2( pfb, 1 );                              // (pfb->*&foobar::foo)( 1 );

    delegate<void (foobar &, int)> dg3;
    dg3 = &foobar::foo;                         // FT03
    dg3( fb, 1 );                               // (fb.*&foobar::foo)( 1 );

    delegate<void (foobar, int)> dg4;
    dg4 = &foobar::foo;                         // FT04
    dg4( fb, 1 );                       // ((copy of fb).*&foobar::foo)( 1 );

    delegate<void (boost::shared_ptr<foobar>, int)> dg5;

    dg5 = &foobar::foo;                         // FT04
    dg5( spfb, 1 );                 // (get_pointer(spfb)->*&foobar::foo)( 1 );

    dg1.bind( &foobar::foo, pfb );              // FT05
    dg1( 1 );                                   // (pfb->*&foobar::foo)( 1 );

    dg1.bind( &foobar::foo, boost::ref( fb ) ); // FT05
    dg1( 1 );                                   // (fb.*&foobar::foo)( 1 );

    dg1.bind( &foobar::foo, spfb );             // FT06
    dg1( 1 );                       // (get_pointer(spfb)->*&foobar::foo)( 1 );

    dg1 = pfb;                                  // FT07
    dg1( 1 );                                   // (*pfb)( 1 );

    dg1 = boost::ref( fb );                     // FT07
    dg1( 1 );                                   // fb( 1 );

    dg1 = fb;                                   // FT08
    dg1( 1 );                                   // (copy of fb)( 1 );

    dg1 = stateless();                          // FT09
    dg1( 1 );                                   // stateless()( 1 );

    dg1 = 0;                                    // FT10
    try { dg1( 1 ); }                           // throw bad_function_call();
    catch(bad_function_call) { }

    dg1 += &hello;                              // FT11
    dg1 += delegate<void (int)>( &foobar::foo, spfb );
    dg1 += fb;

    dg1( 1 );                                   // hello( 1 );
                                  // (get_pointer(spfb)->*&foobar::foo)( 1 );
                                              // (copy of fb)( 1 );
}

Part E: list of entries of function reference table

As emphasized previously, it is very important to declare common and representative entries of the function reference table for a generic delegate. It is the basis of the whole FD.Delegate's design and determines performance, as well as robustness. The following list of function entries are declared in order to generalize the common behavioral requirements of a generic delegate for the various function types that we categorized above.

Invocation

invoke() - Invokes the underlying callable entity.

Object management

copy_obj() - Copies the object to destination from source.
delete_obj() - Deletes the object.

General information inquiry

is_empty() - Determines whether or not the delegate is empty.
size_of_fxn() - Retrieves size of the underlying callable entity.
type_of_fxn() - Retrieves std::type_info of the underlying callable entity.
is_functor() - Determines whether or not the underlying callable entity is a function object.

Comparisons

compare_equal_same_type() - Compares equality of two underlying callable entities of the same type.
memcmp_delegate() - Non contextual memory comparison between underlying callable entities of two arbitrary types.

Multicast

is_multicast() - Determines whether or not the delegate is a multicast.
add_delegates() - Adds one or more delegates into multicast.
remove_delegates() - Removes one or more delegates from multicast.
find_delegate() - Determines whether or not the specified delegate is in multicast.

Part F: generic data structure to store a callable entity

It is required to have at least two void pointers and an any_fxn_pointer to store all the previously categorized function types into a generic delegate. To make it easy to access and manipulate these void pointers, as well as any_fxn_pointer, they are bundled up together into a single structure called 'delegate_holder'.

struct delegate_holder
{
    void const * obj_ptr;
    any_fxn_pointer any;
    void const * tbl_ptr;
};

See the table below to understand how these void pointer members are used effectively in a situation of storing one of the various function types.

Function type

delegate_holder

Fast delegate (Yes / No)

Equality comparison with self type (compare_eq_same_type)

.obj_ptr

.any

.tbl_ptr

FT01

Not used

Function Pointer