Is it possible to detect if a class exists in C++ using SFINAE? If possible then how?
Suppose we have a class that is provided only by some versions of a library. I'd like to know if it is possible to use SFINAE to detect whether the class exists or not. The result of detection is arbitrary, say an enum constant which is 1 if it exists, 0 otherwise.
If we ask the compiler to tell us anything about a class type T that has not
even been declared we are bound to get a compilation error. There is no way
around that. Therefore if we want to know whether class T "exists", where T
might not even have been declared yet, we must declare T first.
But that is OK, because merely declaring T will not make it "exist", since
what we must mean by T exists is T is defined. And if, having declared T,
you can then determine whether it is already defined, you need not be in
any confusion.
So the problem is to determine whether T is a defined class type.
sizeof(T) is no help here. If T is undefined then it will give an
incomplete type T error. Likewise typeid(T). Nor is it any good
crafting an SFINAE probe on the type T *, because T * is a defined type
as long as T has been declared, even if T isn't. And since we are
obliged to have a declaration of class T, std::is_class<T> is not the
answer either, because that declaration will suffice for it to say "Yes".
C++11 provides std::is_constructible<T ...Args> in <type_traits>. Can
this offer an off-the-peg solution? - given that if T is defined, then it must have at least one constructor.
I'm afraid not. If you know the signature of at least one public
constructor of T then GCC's <type_traits> (as of 4.6.3) will indeed do
the business. Say that one known public constructor is T::T(int). Then:
std::is_constructible<T,int>::value
will be true if T is defined and false if T is merely declared.
But this isn't portable. <type_traits> in VC++ 2010 doesn't yet provide std::is_constructible and even its std::has_trivial_constructor<T> will barf if T is not defined: most likely when std::is_constructible does arrive it will follow suit. Furthermore, in the eventuality that only private constructors of T exist for offering to std::is_constructible then even GCC
will barf (which is eyebrow raising).
If T is defined, it must have a destructor, and only one destructor. And that destructor is likelier to be public than any other possible member of T. In that light, the simplest and strongest play we can make is to craft an SFINAE probe for the existence of T::~T.
This SFINAE probe cannot be crafted in the routine way for determining
whether T has an ordinary member function mf - making the "Yes overload"
of the SFINAE probe function takes an argument that is defined in terms
of the type of &T::mf. Because we're not allowed to take the address of a
destructor (or constructor).
Nevertheless, if T is defined, then T::~T has a type DT- which must be
yielded by decltype(dt) whenever dt is an expression that evaluates to an
invocation of T::~T; and therefore DT * will be a type also, that can in
principle be given as the argument type of a function overload. Therefore we
can write the probe like this (GCC 4.6.3):
#ifndef HAS_DESTRUCTOR_H
#define HAS_DESTRUCTOR_H
#include <type_traits>
/*! The template `has_destructor<T>` exports a
boolean constant `value that is true iff `T` has
a public destructor.
N.B. A compile error will occur if T has non-public destructor.
*/
template< typename T>
struct has_destructor
{
/* Has destructor :) */
template <typename A>
static std::true_type test(decltype(std::declval<A>().~A()) *) {
return std::true_type();
}
/* Has no destructor :( */
template<typename A>
static std::false_type test(...) {
return std::false_type();
}
/* This will be either `std::true_type` or `std::false_type` */
typedef decltype(test<T>(0)) type;
static const bool value = type::value; /* Which is it? */
};
#endif // EOF
with only the restriction that T must have a public destructor to be legally invoked in the argument expression of decltype(std::declval<A>().~A()). (has_destructor<T> is a simplified adaptation of the method-introspecting template I contributed here.)
The meaning of that argument expression std::declval<A>().~A() may be obscure to some, specifically std::declval<A>(). The function template std::declval<T>() is defined in <type_traits> and returns a T&& (rvalue-reference to T) - although it may only be invoked in unevaluated contexts, such as the argument of decltype. So the meaning of std::declval<A>().~A() is a call to ~A() upon some given A. std::declval<A>() serves us well here by obviating the need for there to be any public constructor of T, or for us to know about it.
Accordingly, the argument type of the SFINAE probe for the "Yes overload" is: pointer to the type of the destructor of A, and test<T>(0) will match that overload just in case there is such a type as destructor of A, for A = T.
With has_destructor<T> in hand - and its limitation to publicly destructible values of T firmly in mind - you can test whether a class T is defined at some point in your code by ensuring that you declare it before asking the question. Here is a test program.
#include "has_destructor.h"
#include <iostream>
class bar {}; // Defined
template<
class CharT,
class Traits
> class basic_iostream; //Defined
template<typename T>
struct vector; //Undefined
class foo; // Undefined
int main()
{
std::cout << has_destructor<bar>::value << std::endl;
std::cout << has_destructor<std::basic_iostream<char>>::value
<< std::endl;
std::cout << has_destructor<foo>::value << std::endl;
std::cout << has_destructor<vector<int>>::value << std::endl;
std::cout << has_destructor<int>::value << std::endl;
std::count << std::has_trivial_destructor<int>::value << std::endl;
return 0;
}
Built with GCC 4.6.3, this will tell you that the 2 // Defined classes
have destructors and the 2 // Undefined classes do not. The fifth
line of output will say that int is destructible, and the final
line will show that std::has_trivial_destructor<int> agrees. If we want
to narrow the field to class types, std::is_class<T> can be applied after
we determine that T is destructible.
Visual C++ 2010 does not provide std::declval(). To support that compiler
you can add the following at the top of has_destructor.h:
#ifdef _MSC_VER
namespace std {
template <typename T>
typename add_rvalue_reference<T>::type declval();
}
#endif
delctype(…), i.e., the return type of the destructor, is void &. And it's not possible to create a pointer to it. Here's what my GCC tells me: cannot declare pointer to 'using DestructorReturnType = void&' {aka 'void&'}. –
Pinstripe Still did not find a satisfying answer in this post...
Mike Kinghan started the answer right and told a smart thing:
So the problem is to determine whether T is a defined class type.
But
sizeof(T) is no help here
is not correct...
Here is how you can do it with sizeof(T):
template <class T, class Enable = void>
struct is_defined
{
static constexpr bool value = false;
};
template <class T>
struct is_defined<T, std::enable_if_t<(sizeof(T) > 0)>>
{
static constexpr bool value = true;
};
is_complete but class-wise it is just is_defined. BTW, after a year I've just realized, that sizeof(T) > 0 can be replaced with simply sizeof(T). –
Teleutospore std::false_type and std::true_type, so that they automatically provide operator()? –
Cladding With SFINAE, no. I think name lookup tricks are the way to get this done. If you aren't afraid to inject a name into the library's namespace:
namespace lib {
#if DEFINE_A
class A;
#endif
}
namespace {
struct local_tag;
using A = local_tag;
}
namespace lib {
template <typename T = void>
A is_a_defined();
}
constexpr bool A_is_defined =
!std::is_same<local_tag, decltype(lib::is_a_defined())>::value;
If A is declared in the global namespace:
#if DEFINE_A
class A;
#endif
namespace {
struct local_tag;
using A = local_tag;
}
namespace foo {
template <typename T = void>
::A is_a_defined();
}
constexpr bool A_is_defined =
!std::is_same<local_tag, decltype(foo::is_a_defined())>::value;
Alright, I think I found a way to do this, though there may be better ways. Suppose we have class A that is included in some instances of the library and not in others. The trick is to define a special private conversion constructor in A and then use SFINAE to detect the conversion constructor. When A is included, the detection succeeds; when it's not, detection fails.
Here's a concrete example. First the header for the detection template, class_defined.hpp:
struct class_defined_helper { };
template< typename T >
struct class_defined {
typedef char yes;
typedef long no;
static yes test( T const & );
static no test( ... );
enum { value = sizeof( test( class_defined_helper( )) == sizeof( yes ) };
};
#define CLASS_DEFINED_CHECK( type ) \
type( class_defined_helper const & ); \
\
friend struct class_defined< type >;
Now a header that contains a class definition, blah.hpp:
#include "class_defined.hpp"
#ifdef INCLUDE_BLAH
class blah {
CLASS_DEFINED_CHECK( blah );
};
#else
class blah;
#endif
Now the source file, main.cpp:
#include "blah.hpp"
int main( ) {
std::cout << class_defined< blah >::value << std::endl;
}
Compiled with BLAH_INCLUDED defined this prints 1. Without BLAH_INCLUDED defined it prints 0. Unfortunately this still requires a forward declaration of the class to compile in both cases. I don't see a way to avoid that.
#if CLASS_DEFINED_MACRO and then the dependent code. –
Cilo © 2022 - 2024 — McMap. All rights reserved.
#ifdefs? – Antistrophe