C ++编译时多态

Question

There two unrelated structures A and B

template <typename T>
struct A {};

template <typename T>
struct B {};

one enum type

typedef enum  { ma, mb} M;

and class C containing function templates

class C
{
public: 
    template <typename T>
    static void f1 ( A <T> &a) {}

    template <typename T>
    static void f2 ( B <T> &b) {}

    template <typename U>
    static void algo (U &u, M m)
    {
        /*Long algorithm here                     
        ....
        */
        if ( m == ma) f1(u);
        else f2(u);
    }
};

Static method algo contains some algorithm, that is quite difficult... It modified some values and results into structure A or B.

I would like to run static method algo with objects A or B depending on M value. But how to say it to my compiler :-)

int main()
{
A <double> a;
C::algo (a, ma); //Error

}

Error   1   error C2784: 'void C::f1(A<T>)' : could not deduce template argument for 'A<T>' from 'B<T>

A] I was thinking about pointer to function, but they are not usable with function templates.

B] Maybe a compile polymorphism could help

template <typename U, M m>
static void algo (U &u, M <m> ) { ...}  //Common for ma

template <typename U, M m>
static void algo (U &u, M <mb> ) { ...} //Spec. for mb

But this solution has one big problem: Both implementations should unnecessarily include almost the same code (why to write the algorithm twice?).

So I need one function algo() processing both types of arguments A and B. Is there any more comfortable solution?

Answer 1

It seems that you are using the enum to convey type information from the user. I would suggest that you don't.

In the simplest case if f1 and f2 are renamed f , then you can remove the if altogether and just call it. The compiler will call the appropriate overload for you.

If you cannot or don't want to rename the function templates, then you can write a helper template that will dispatch for you (basic class template undefined, specialisations for A and B that dispatch to the appropriate static function)

If the enum is used for something else (that the compiler cannot resolve for you), you can still pass it around and rewrite the helper to dispatch on the enum rather than the type of the argument and you will have to rewrite the code to have the enum value as a compile time constant (simplest: pass it as template argument to algo ). In this case ou can write function specialisations instead of classes if you want, as they would be full specialisations. But note that if you can avoid having to pass it you will remove a whole family of errors: passing the wrong enum value.

// Remove the enum and rename the functions to be overloads:
//
struct C {  // If everything is static, you might want to consider using a
            // namespace rather than a class to bind the functions together...
            // it will make life easier

   template <typename T>
   static void f( A<T> & ) { /* implement A version */ }

   template <typename T>
   static void f( B<T> & ) { /* implement B version */ }

   template <typename T> // This T is either A<U> or B<U> for a given type U
   static void algo( T & arg ) {
      // common code
      f( arg ); // compiler will pick up the appropriate template from above
   } 
};

For the other alternatives, it is easier if the enclosing scope is a namespace, but the idea would be the same (just might need to fight the syntax a bit harder:

template <typename T>
struct dispatcher;

template <typename T>
struct dispatcher< A<T> > {
   static void f( A<T>& arg ) {
      C::f1( arg );
   }
};
template <typename T>
struct dispatcher< B<T> > {
   static void f( B<T>& arg ) {
      C::f2( arg );
   }
};

template <typename T>
void C::algo( T & arg ) {
   // common code
   dispatcher<T>::f( arg );
}

Again, getting this to work with a class might be a bit trickier as it will probably need a couple of forward declarations, and I don't have a compiler at hand, but the sketch should lead you in the right direction.

Answer 2

Normal function overloading is sufficient:

template <typename T> 
static void f1 ( A <T> &a) {} 

template <typename T> 
static void f2 ( B <T> &b) {} 

template <typename T> 
static void algo (A<T>& u) {
    f1(u);
} 

template <typename T> 
static void algo (B<T>& u) {
    f2(u);
} 

And then:

A<int> a;
Foo::algo(a);

Although it's not clear what you stand to gain from such an arrangement.

Answer 3

If you realy need to do that in one function, you can use typetraits:

 template<typename T, T Val>
 struct value_type { static const T Value = Val; };

 struct true_type   : public value_type<bool, true>{};
 struct false_type  : public value_type<bool, false>{};


 template<class T>
 struct isClassA : public false_type{};

 template<>
 struct isClassA<A> : public true_type{};


 template < typename T >
 void Algo( T& rcT )
 {
    if ( true == isClassA<T>::Value )
    {
        // Class A algorithm
    }
    else
    {
        // Other algorithm
    }
 };

Answer 4

the value of m parameter is unknown until runtime, so the compiler has to generate code for both if (m == ma) and else branches when it specialize the function. It then complains since it can't understand what he should do if you happen to call C::algo(a,mb) or similar.

As Jon suggested, overloading should fix your case, try using this code:

template<typename U>
static void f12(A<U>&u) { f1(u); }

template<typename U>
static void f12(B<U>&u) { f2(u); }

template<typename U>
static void algo(U& u, M m)
{
    /* long algorithm here
      ...
    */
    //use overloading to switch over U type instead of M value
    f12(u);
}

Also you can use function pointers with template functions, as long as you specify the template parameters:

template<typename U>
static void algo(U& u, M m, void(*)(U&) func)
{
    /* ... */
    (*func)(u);
}

int main()
{
    A <double> a;
    C::algo (a, ma, &C::f1<double> );
}