Template programming: specialization and enable_if

Question

I'm working with libffi and I've made a class with a similar template to std::function (i.e class Func<Ret (Args...)> { /* ... */};. I want to convert the return type (Ret) and each argument type (Args) to their corresponding libffi type (see this for reference). So far I've come up with this:

// Member function of 'Func' class
Prepare(void)
{
 // This vector holds all the type structures
 std::vector<ffi_type*> argumentTypes{ GetFFIType<Args>()... };
 ffi_type * returnType = GetFFIType<Ret>();

 // Rest of the code below
 // ....
}

Where the GetFFIType function is implemented as the following:

template <typename T>
ffi_type * GetFFIType(void)
{
    // We will check for any kind of pointer types
    if(std::is_pointer<T>::value || std::is_array<T>::value ||
       std::is_reference<T>::value || std::is_function<T>::value)
        return &ffi_type_pointer;

    if(std::is_enum<T>::value)
        //return GetFFIType<std::underlying_type<T>::type>();
    {
        // Since the size of the enum may vary, we will identify the size
        if(sizeof(T) == ffi_type_schar.size)    return std::is_unsigned<T>::value ? &ffi_type_uchar : &ffi_type_schar;
        if(sizeof(T) == ffi_type_sshort.size)   return std::is_unsigned<T>::value ? &ffi_type_ushort : &ffi_type_sshort;
        if(sizeof(T) == ffi_type_sint.size) return std::is_unsigned<T>::value ? &ffi_type_uint : &ffi_type_sint;
        if(sizeof(T) == ffi_type_slong.size)    return std::is_unsigned<T>::value ? &ffi_type_ulong : &ffi_type_slong;
    }

    assert(false && "cannot identify type");
}

// These are all of our specializations
template <> ffi_type * GetFFIType<void>(void)       { return &ffi_type_void; }
template <> ffi_type * GetFFIType<byte>(void)       { return &ffi_type_uchar; }
template <> ffi_type * GetFFIType<char>(void)       { return &ffi_type_schar; }
template <> ffi_type * GetFFIType<ushort>(void)     { return &ffi_type_ushort; }
template <> ffi_type * GetFFIType<short>(void)      { return &ffi_type_sshort; }
template <> ffi_type * GetFFIType<uint>(void)       { return &ffi_type_uint; }
template <> ffi_type * GetFFIType<int>(void)        { return &ffi_type_sint; }
template <> ffi_type * GetFFIType<ulong>(void)      { return &ffi_type_ulong; }
template <> ffi_type * GetFFIType<long>(void)       { return &ffi_type_slong; }
template <> ffi_type * GetFFIType<float>(void)      { return &ffi_type_float; }
template <> ffi_type * GetFFIType<double>(void)     { return &ffi_type_double; }
template <> ffi_type * GetFFIType<long double>(void)    { return &ffi_type_longdouble; }

This works, but obviously there is some room for improvements. If the type is invalid (i.e a class or a structure) it is not identified at compile-time (a runtime-error occurs instead using assert). How would I avoid this, and make this function determine whether a type is valid (a primitive type) or not during compilation?

I also dislike the way I am identifying the underlying type in case of enums. I would prefer using std::underlying_type<T> instead (commented out in the code) but it issues compile-errors if the type is for example a void pointer (type_traits:1762:38: error: ‘void*’ is not an enumeration type)

I tried to achieve this behavior using std::enable_if but without success... Do tell if I should explain something in case it sounded a bit fuzzy!

Summary: I want to get the GetFFIType function to determine everything during compilation and the function should only support primitive types (see this for a more extensive reference)

EDIT: Sorry for the title, nothing better came to mind :(

Answer 1

It's easier and usually better to overload function templates, rather than specialize them. I'll add a version of the function with a pointer argument so it can be called without a template parameter list:

inline ffi_type * GetFFITypeHelper( void* ) { return &ffi_type_void; }
inline ffi_type * GetFFITypeHelper( byte* ) { return &ffi_type_uchar; }
// ...

Then you can use enable_if for the more generalized cases you want to cover.

template<typename T> auto GetFFITypeHelper( T* ) ->
    std::enable_if< std::is_function<T>::value, ffi_type* >::type
{ return &ffi_type_pointer; }
template<typename T> auto GetFFITypeHelper( T* ) ->
    std::enable_if< std::is_enum<T>::value, ffi_type* >::type
{ return GetFFITypeHelper( static_cast<std::underlying_type<T>::type*>(nullptr) ); }

AFTER all those overloads are declared, the version you wanted is:

template<typename T> ffi_type * GetFFIType()
{ return GetFFITypeHelper( static_cast<T*>(nullptr) ); }

Answer 2

Putting the logic inside a class template rather than a function template will allow for partial specializations, which we can also take advantage of for SFINAE tricks:

// Second parameter is an implementation detail
template<typename T, typename Sfinae = std::true_type>
struct ToFFIType;

// Front-end
template<typename T>
ffi_type* GetFFIType()
{ return ToFFIType<T>::make(); }

// Primary template where we end up if we don't know what to do with the type
template<typename T, typename = std::true_type>
struct ToFFIType {
    static_assert( dependent_false_type<T>::value,
                   "Write your clever error message to explain why we ended up here" );

    static ffi_type* make() = delete;
};

// Trait-like to match what we want with ffi_type_pointer
template<typename T>
struct treat_as_pointer: or_<
    std::is_pointer<T>
    , std::is_array<T>
    , std::is_reference<T>
    , std::is_function<T>
> {};

template<typename T>
struct ToFFIType<T, typename treat_as_pointer<T>::type> {
    static ffi_type* make()
    { return &fii_type_pointer; }
};

// Matches enumeration types
template<typename T>
struct ToFFIType<T, typename std::is_enum<T>::type> {
    static ffi_type* make()
    {
        return ToFFIType<typename std::underlying_type<T>::type>::make();
    }
};

The total specializations are straightforward to write so I won't show them. Although note that you can choose to instead match e.g. std::is_integral and switch on sizeof(T) if you want, similar to what you did to work around std::underlying_type.

Finally here are two suggested implementations of the two utilities which are assumed in the above code; obviously you don't need to use them verbatim as long as you write something else along in the same vein.

// Same functionality as std::false_type but useful
// for static_assert in templates
template<typename Dummy>
struct dependent_false_type: std::false_type {};

// Disjunction of boolean TMP integral constants
// Take care to inherit from std::true_type/std::false_type so
// the previous SFINAE trick works
template<typename... T>
struct or_: std::false_type {};

// There likely are better implementations
template<typename Head, typename... Tail>
struct or_<Head, Tail...>: std::conditional<
    Head::value
    , std::true_type              // short circuit to desired base
    , typename or_<Tail...>::type // or inherit from recursive base
>::type {}; // Note: std::conditional is NOT the base