SFINAE check for operator[] is more confused than I am?

Question

I wrote simple check for operator[], but has_subscript_op struct template instantiation chooses wrong overload:

#include <iostream>
#include <type_traits>
#include <string>
#include <map>

template<class, class, class = void>
struct has_subscript_op : std::false_type
{ };

template<class T, class S>
struct has_subscript_op<T, S, std::void_t<decltype(&std::declval<T>()[S()])>> : std::true_type
{ };

int main()
{
    //true, nice
    std::cout << "int[][int]: " << has_subscript_op<int[], int>::value << std::endl;
    //false, nice
    std::cout << "int[][float]: " << has_subscript_op<int[], float>::value << std::endl;
    //true, nice
    std::cout << "std::string[int]: " << has_subscript_op<std::string, int>::value << std::endl;
    //true, WAT?
    std::cout << "std::map<std::string, std::string>[int]: " << has_subscript_op<std::map<std::string, std::string>, int>::value << std::endl;
}

I'm using GCC 6.2.0

Coliru

Is this GCC bug, general bug, or I made an obvious mistake somewhere?

Answer 1

Just drop the & and use declval for the key too:

template<class T, class S>
struct has_subscript_op<T, S, std::void_t<decltype(std::declval<T>()[std::declval<S>()])>> : std::true_type {};

Live example at coliru

Why did the check with S() gave the wrong result? Because in GCC, it's considered as it was 0. A std::string can be constructed with a pointer, and 0 happens to be a null pointer constant.

Other compilers should not treat S() as it was 0 in C++14.

You can try for yourself:

std::map<std::string, std::string> test;

// compile fine, segfault at runtime
auto a = test[0];

// compile error!
auto b = test[2]

The check works better with std::declval because it's not a 0, neither a 2 but a plain int. Bonus, with declval, your check won't require the key to be default constructible.