template functions with shared private data

Question

I'm looking for 'best-practice' in the following situation: In general, there are three common ways to share private data between two (or more) non-member functions with differential advantages and disadvantages:

// Example 1: using 'static' class

// HPP
namespace foo {
    class bar
    {
    private:
        static const char* const s_private;
        bar();

    public:
        static void s_method0();
        static void s_method1();
    }; /* class bar */
} /* namespace foo */

// CPP
namespace foo {
    const char* const bar::s_private = "why do I need to be visible in HPP?";

    void bar::s_method0() { std::cout << "s_method0 said: " << s_private << std::endl; }
    void bar::s_method1() { std::cout << "s_method1 said: " << s_private << std::endl; } 
} /* namespace foo */


// Example 2: using unnamed-namespace

// HPP
namespace foo { 
    void bar0(); 
    void bar1();
} /* namespace foo */

// CPP
namespace foo {
    namespace {
        const char* const g_anonymous = "why do I need external linkage?";
    } /* unnamed-namespace */

    void bar0() { std::cout << "bar0 said: " << g_anonymous << std::endl; }
    void bar1() { std::cout << "bar1 said: " << g_anonymous << std::endl; }
} /* namespace foo */


// Example 3: using static keyword in namespace-scope

// HPP
namespace foo { 
    void bar0(); 
    void bar1();
} /* namespace foo */

// CPP
namespace foo {
    static const char* const g_internal = "nobody outside this file can see me and I don't have external linkage";

    void bar0() { std::cout << "bar0 said: " << g_internal << std::endl; }
    void bar1() { std::cout << "bar1 said: " << g_internal << std::endl; }
} /* namespace foo */

I prefer 'Example 3' because it's as close to the intention as it could be. But now I'm running in some problem's using templated functions. 'Example 1' seems to be the only way to solve this:

// HPP
namespace foo {
    class bar
    {
    private:
        static const char* const s_private;
        bar();

    public:
        template<typename T> static void s_method0() { std::cout << "s_method0 said: " << s_private << std::endl; }
        template<typename T> static void s_method1() { std::cout << "s_method1 said: " << s_private << std::endl; }
    }; /* class bar */
} /* namespace foo */

// CPP
namespace foo {
    const char* const bar::s_private = "why do I need to be visible in HPP?";
} /* namespace foo */

That's unsatisfying. Especialy because there are other (in this case methods) non-member function which should be in the same (in this case class-) scope, which don't need to access this private data.

Does anybody know an elegant solution?

Thanks for any help. Best regards.

Answer 1

This is, somewhat unfortunately, an issue that springs quite often with template.

But may I suggest that you are over-engineering here ?

The truth is, whether you look at Loki code (by Andrei Alexandrescu) or Boost code (infamous David Abrahams notably), no-one really bothered to provide a better privacy.

Rather, they simply relied on convention and used a Private namespace (Loki) or a detail namespace (Boost, with sometimes a longer and more descriptive name to prevent clashes).

It's annoying, but there is not much you can do in practice.... though I actually have a solution for your specific problem ;)

// Evil solution!

#ifdef MY_SUPER_MACRO
#  error "MY_SUPER_MACRO is already defined!"
#endif

#define MY_SUPER_MACRO "Some string"

template <typename T> void foo() { std::cout << "foo - " MY_SUPER_MACRO "\n"; }
template <typename T> void bar() { std::cout << "bar - " MY_SUPER_MACRO "\n"; }

#undef MY_SUPER_MACRO

And hop, I achieved locality in a header with an evil macro :)

Answer 2

What about that?

namespace foo {
    namespace detail {
        class shared 
        { 
            template<typename> friend void bar0();
            template<typename> friend void bar1();

            static const char* const m_private; 
        }; /* class shared */
    } /* namespace detail */

    template<typename T> void bar0() { std::cout << "bar0 said: " << detail::shared::m_private << std::endl; }
    template<typename T> void bar1() { std::cout << "bar1 said: " << detail::shared::m_private << std::endl; }
} /* namespace foo */

EDIT 2: OBSOLETE ANSWER. THIS CODE ISN'T WORKING! POSTED AN EXPLANATION BELOW.

EDIT: In 'real code' I would replace namespace detail by an unnamed namespace. This would bring up the possibility to add other shared resources in different header files using the same name-scope:

namespace foo {
    template<typename T> void bar0();
    template<typename T> void bar1();
    template<typename T> void bar2();
    template<typename T> void bar3();


    namespace {
        class shared0 
        { 
            template<typename> friend void foo::bar0();
            template<typename> friend void foo::bar1();

            static const char* const m_private0; 
        }; /* class shared0 */

        class shared1
        { 
            template<typename> friend void foo::bar2();
            template<typename> friend void foo::bar3();

            static const char* const m_private1;    
        }; /* class shared1 */
    } /* unnamed */


    template<typename T> void bar0() { std::cout << "bar0 said: " << shared0::m_private0 << std::endl; }
    template<typename T> void bar1() { std::cout << "bar1 said: " << shared0::m_private0 << std::endl; }

    template<typename T> void bar2() { std::cout << "bar0 said: " << shared1::m_private1 << std::endl; }
    template<typename T> void bar3() { std::cout << "bar1 said: " << shared1::m_private1 << std::endl; }
} /* namespace foo */

Answer 3

I had played around with the different techniques. My idea, using the unnamed namespace in the header file, was to 'mark' the 'shared'-classes as 'header-file-only'. Sure, due to the fact they aren't containing public members you cannot make nasty things anyway with it. But I thought it would be more close to the intention.

But I was wrong! After thinking about this I'm in shame. It's so logically simple. This example is showing what's the problem with it (entire code for clearness):

// header0.hpp
#ifndef HPP_HEADER0_INCLUDED
#define HPP_HEADER0_INCLUDED

#include <iostream>
#include <string>

namespace ns {
    template<typename T> void header0_func0();
    template<typename T> void header0_func1();

    namespace {
        class header0
        {
            template<typename> friend void ns::header0_func0();
            template<typename> friend void ns::header0_func1();

            static std::string s_private;
        }; /* class header0 */
    } /* unnamed */

    template<typename T> void header0_func0() { std::cout << "header0_func0: " << header0::s_private << std::endl; }
    template<typename T> void header0_func1() { std::cout << "header0_func1: " << header0::s_private << std::endl; }
} /* namespace ns */

#endif /* HPP_HEADER0_INCLUDED */ 


// header1.hpp
#ifndef HPP_HEADER1_INCLUDED
#define HPP_HEADER1_INCLUDED

#include <iostream>
#include <string>

namespace ns {
    template<typename T> void header1_func0();
    template<typename T> void header1_func1();

    namespace {
        class header1
        {
            template<typename> friend void ns::header1_func0();
            template<typename> friend void ns::header1_func1();

            static std::string s_private;
        }; /* class header1 */
    } /* unnamed */

    template<typename T> void header1_func0() { std::cout << "header1_func0: " << header1::s_private << std::endl; }
    template<typename T> void header1_func1() { std::cout << "header1_func0: " << header1::s_private << std::endl; }
} /* namespace ns */

#endif /* HPP_HEADER1_INCLUDED */


// source.cpp
#include "header0.hpp"
#include "header1.hpp"

std::string ns::header0::s_private = "header0 private data definition by source.cpp",
            ns::header1::s_private = "header1 private data definition by source.cpp";

namespace {
    // hide private class
    class source
    {
        source();
        ~source();

        static source s_instance;
    };

    source::source() { 
        std::cout << "source.cpp:\n";
        ns::header0_func0<int>();
        ns::header0_func1<int>();
        ns::header1_func0<int>();
        ns::header1_func1<int>();
        std::cout << '\n';
    }

    source::~source() { }
    source source::s_instance;
} /* unnamed */

By now everything seems to be OK. But what happen's if we try to use our headers in other translation units? Let's take a look:

// main.cpp
#include "header0.hpp"
#include "header1.hpp"

int main()
{   
    std::cout << "main.cpp:\n";
    ns::header0_func0<int>();
    ns::header0_func1<int>();
    ns::header1_func0<int>();
    ns::header1_func1<int>();
    std::cout << '\n';

    return 0;
}

What happens is, that we are ending with 2 unresolved externals. So, is the linker just an idiot? No, he is not. Thinking about what are unnamed namespace used for, we know what's going on. An unnamed namespace has an unique identifier in each translation unit. Thus, in our main.cpp, the linker doesn't know the definition of our private data in source.cpp. So, what happens if we define this private data in main.cpp - just to bring matters to a head - ?

// main.cpp
#include "header0.hpp"
#include "header1.hpp"

std::string ns::header0::s_private = "header0 private data definition by main.cpp",
            ns::header1::s_private = "header1 private data definition by main.cpp";

int main()
{   
    std::cout << "main.cpp:\n";
    ns::header0_func0<int>();
    ns::header0_func1<int>();
    ns::header1_func0<int>();
    ns::header1_func1<int>();
    std::cout << '\n';

    return 0;
}

Now, everything is compiling and getting linked 'correctly' - or rather, it seems so. This is the console output of that program:

source.cpp:
header0_func0: header0 private data definition by source.cpp
header0_func1: header0 private data definition by source.cpp
header1_func0: header1 private data definition by source.cpp
header1_func0: header1 private data definition by source.cpp

main.cpp:
header0_func0: header0 private data definition by source.cpp
header0_func1: header0 private data definition by source.cpp
header1_func0: header1 private data definition by source.cpp
header1_func0: header1 private data definition by source.cpp

That means: If undefined behavior is what you are looking for, here you are. In other words: Based on the explanation above: Don't use unnamed namespace in header files to encapsulate private shared data.

And the last question is, "what's the solution?" If you don't want to use 'static' (utility) classes, you should prefer my first posted solution (only changed code):

// header0.hpp
// ...
namespace ns {
    // ...  
    namespace detail { 
        class header0 { /*...*/ };
    } /* namespace detail */

    template<typename T> void header0_func0() { std::cout << "header0_func0: " << detail::header0::s_private << std::endl; }
    template<typename T> void header0_func1() { std::cout << "header0_func1: " << detail::header0::s_private << std::endl; }
} /* namespace ns */
// ...

// header1.hpp
// ...
namespace ns {
    // ...  
    namespace detail { 
        class header1 { /*...*/ };
    } /* namespace detail */

    template<typename T> void header0_func0() { std::cout << "header0_func0: " << detail::header1::s_private << std::endl; }
    template<typename T> void header0_func1() { std::cout << "header0_func1: " << detail::header1::s_private << std::endl; }
} /* namespace ns */
// ...

// source.cpp
// ...
std::string ns::detail::header0::s_private = "header0 private data definition by source.cpp",
            ns::detail::header1::s_private = "header1 private data definition by source.cpp";
// ...

I'm looking forward to any comment. Best regards.

Answer 4

If I understand correctly, your complaint/concern is that unlike with templates, with non-templates, one can define the function bodies inside the CPP, rather than the header, in which case they can access non-class statics, which are "invisible" to the outside world, including member functions defined in the header. This is all true.

However, remember that there's nothing stopping one defining other member functions in the CPP as well, in which case, they'd be equally able to access the static data. So really, the situation is no different. Your complaint is based on a false dichotomy.

If you genuinely want to prevent anything but s_method0<T>() and s_method1<T>() accessing s_private, then you must put them all in a dedicated class. It's as simple as that. This would be the case even for non-templates.