Use different sets of functions based on template parameters (C++ traits?)

Question

I have defined a class in C++ which holds an array of scalars of type T for which I want to define operators like sin, cos, etc. For defining the meaning of sin applied on an object of this class I need to know the meaning of sin applied on the single scalar type T. This means I need to use appropriate math libraries (corresponding to the scalar type T) within the class. Here's the code as it is now:

template<class T>
class MyType<T>
{
    private:
        std::vector<T> list;

    // ...

        template<class U> friend const UTP<U> sin(const UTP<U>& a);
        template<class U> friend const UTP<U> cos(const UTP<U>& a);
        template<class U> friend const UTP<U> tan(const UTP<U>& a);

    //...
};

template<class T> const UTP<T> sin(const UTP<T>& a)
{
   // use the sin(..) appropriate for type T here 
   // if T were double I want to use double std::sin(double)
   // if T were BigNum I want to use BigNum somelib::bigtype::sin(BigNum)
}

Currently, I have code that exposes the appropriate math library (using namespace std;) and then use ::sin(a) inside the sin function for my class MyType. While this works, it seems like a major hack.

I see that C++ traits can be used to store instance specific information (like which set of math functions to use when T is double, when T is BigNum, etc..)

I want to do something like this: (I know this doesn't compile but I hope this conveys what I want to do)

template<T>
struct MyType_traits {
};

template<>
struct MyType_traits<double> {
    namespace math = std;
};

template<>
struct MyType_traits<BigNum> {
    namespace math = somelib::bigtype;
};

and then in redefine my MyType class as:

template<T, traits = MyType_traits<T> >
class MyType
{
// ...
}

and then use traits::math::sin in my friend function. Is there a way in which I can obtain the correct namespace (parameterized by T) containing the math functions?

Answer 1

Isn't argument-dependent look-up good enough?

#include <cmath>
#include <iostream>

namespace xxx {
class X
{
};

X sin(X) { return X(); }
} //xxx

std::ostream& operator<< (std::ostream& os, xxx::X)
{
    return os << "X";
}

template <class T>
void use_sin(T t)
{
    using std::sin; //primitive types are not in a namespace,
                    //and with some implementation sin(double) etc might not be available
                    //in global namespace
    std::cout << sin(t) << '\n';
}

int main()
{
    use_sin(1.0);
    use_sin(xxx::X());
}

This would work for X, because sin(X) is defined in the same namespace as X. If you expect that not to be so, this probably won't help...

Answer 2

It's not the specific answer you're looking for, but wouldn't using template specialization be a simpler option?

As in...

template <typename T> T sin(T& t)
{
    // does nothing
}

template <> float sin(float& t)
{
    ...
}

template <> double sin(double& t)
{
    ...
}

And so on?

Answer 3

I'm including this answer because I finally managed to get what I want (with help from very nice folk on ##c++ at irc.freenode.net). This method enables both ADL as well as a static set of places (xxx::math) to look in for the definition of the math functions.

This way, if the type parameter T of the class Test is such that:

1. if T defines math functions as members then we can use ADL and not touch (add to) the namespace xxx::math.
2. if T doesn't define math functions as members but uses functions from a specific namespace then we can add to the namespace xxx::math like in the example below.

Library looks like this:

#include <vector>
#include <cmath>

namespace xxx {

// include the usual math
namespace math {
    using std::asin;
}

template <class T>
class Test
{
    std::vector<T> array;

    public:
    Test(const typename std::vector<T>::size_type length)
    {
        assert(length >= 1);
        array.assign(length, T(0.0));
    }
    friend std::ostream& operator<<(std::ostream& out, const Test<T>& a)
    {
        out << "(";
        std::copy(a.array.begin(), a.array.end(), std::ostream_iterator<T>(out, ", "));
        out << "\b\b)";
        return out;
    }

    template<class U> friend const Test<U> asin(const Test<U>& a);
};

template<class U> const Test<U> asin(const Test<U>& a)
{
    using math::asin;

    Test<U> ret(a.array.size());
    for (typename std::vector<U>::size_type i = 0; i < a.array.size(); ++i) 
        ret.array[i] = asin(a.array[i]);

    // note how we use have a using math::asin; and then a call to asin(..) here,
    // instead of a math::asin(..). This allows for ADL.

    return ret;
}

} // xxx

Client looks like this:

#include <iostream>
#include <boost/math/complex.hpp>

// client, with some foresight, includes complex math
namespace xxx { namespace math {
    using boost::math::asin;
} }

#include "test.h"

// demo
int main(int argc, char **argv)
{
    using std::cout; using std::endl;

    xxx::Test<double> atest(3);
    cout << "atest: " <<  atest << endl;
    cout << "asin(atest): " <<  asin(atest) << endl;
    cout << endl;

    xxx::Test<std::complex<double> > btest(3);
    cout << "btest: " <<  btest << endl;
    cout << "asin(btest): " <<  asin(btest) << endl;
    cout << endl;

    return 0;
}