What is the best way to create a specialization-on

问题:

Is there a better way to do the following?

#include <iostream>

template <typename T>
T Bar();

template <>
int Bar<int>() { return 3; }

// Potentially other specialisations

int main()
{
    std::cout << Bar<int>() << std::endl; // This should work
    std::cout << Bar<float>() << std::endl; // This should fail
}

The problem with this solution is that it fails at (understandably) link time with "undefined reference to float Bar<float>()" or the like. This can be confusing for other developers as they may suspect an implementation file is not being linked.

I do know another potential solution:

template <typename T>
T Bar() { BOOST_STATIC_ASSERT(sizeof(T) == 0); }

This causes a compiler error when Bar<float>() is requested, exactly what I want. However, I'm concerned that technically a compiler may reject this just as gcc rejects BOOST_STATIC_ASSERT(false) because it knows that it will fail regardless of the template parameter, since sizeof(T) can never be zero.

In summary, I want to know whether:

1. There is another way to do this.
2. I'm mistaken and BOOST_STATIC_ASSERT(sizeof(T)) actually can't fail without instantiation.
3. The only way is to let this be a linker error as above.

回答1:

BOOST_STATIC_ASSERT(sizeof(T) == 0); isn't allowed to fail until the template is instantiated, so I would just do that one. You are correct that BOOST_STATIC_ASSERT(false); triggers each time.

---

The reason for this has to do with two-phase name lookup. This is, essentially, the following: when a template is compiled, it's compiled twice. The first time a compielr sees a template it compiles everything except the expressions dependent on template parameters, and the second compilation happens once the template parameter is known, compiling the instantiation fully.

This is why BOOST_STATIC_ASSERT(false); will fail always: nothing here is dependent and the assert is processed immediately, as if the function weren't a template at all. (Note that MSVC does not implement two-phase look-up, so this fails at instantiation, incorrectly.) Contrarily, because T is dependent (§14.6.2.1), BOOST_STATIC_ASSERT(sizeof(T) == 0); is dependent, and is not allowed to be checked until the template is instantiated. (Where upon it will always fail.)

If a compiler tries to be thoughtful and fail it ahead of time, it would be non-conforming. You're suppose to be able to rely on this stuff. That said, if fear gets the best of you it's trivial to really make it wait:

BOOST_STATIC_ASSERT(sizeof(typename T::please_use_specializations) == 0);

This is both guaranteed to fail, and impossible for a compiler to correctly "smartly" fail ahead of time.

回答2:

This could work:

template <typename T>
T Bar() {
  T::ERROR_invalid_template_argument_;
}

template <>
int Bar<int>() { return 3; }

You could also use the highest size possible if you're afraid of using 0:

  static_assert(sizeof(T) == -1, "No specialization");

回答3:

You could do something like the following:

template <typename T>
T Bar()
{ T::unspecialized_method_called; }

This of course assumes, that T doesn't have a member with the given name, so you would have to choose your "error message" accordingly (e.g. by violating naming conventions).

回答4:

use static_assert with c++0x

template <typename T> 
void bar(){
 static_assert(false, " invalid argument type");
}

this will raise an error when compiling.

For c++ 98/2003, we could try this

template <typename T> 
void bar(){
char invalid_arg_[0];
}

array at lest contains one element. So the compiler would complain. But the error message might be useless to show what happened.

The first wouldn't be a choice since it always fails.

回答5:

There's a little caveat if you're using gcc without -pedantic, in which case it is possible to have sizeof(T) == 0 - when T is a zero-length array.

#include <iostream>

#include "boost/static_assert.hpp"

template <typename T>
void Foo()
{
    BOOST_STATIC_ASSERT(sizeof(T) == 0);
    std::cout << "Actually, it is possible to instantiate this." << std::endl;
}

int main()
{
    Foo<int[0]>();

    return 0;
}

In this case you can work around it by using this instead:

BOOST_STATIC_ASSERT(sizeof(T) == sizeof(T) + 1);

It may be better to encapsulate this trick, which can improve readability because it expresses your intent:

#define NEVER_INSTANTIATE(T) BOOST_STATIC_ASSERT(sizeof(T) == sizeof(T) + 1);

As GMan explained, this can't fail without instantiation just like sizeof(T) == 0. However, the moral of this story should probably instead be always compile with -pedantic instead.