Overload resolution with ref-qualifiers

While working with ref-qualified function overloads, I'm getting different results from GCC (4.8.1) and Clang (2.9 and trunk). Consider the following code:

#include <iostream>
#include <utility>

struct foo
{
    int& bar() &
    {
        std::cout << "non-const lvalue" << std::endl;
        return _bar;
    }
    //~ int&& bar() &&
    //~ {
    //~     std::cout << "non-const rvalue" << std::endl;
    //~     return std::move(_bar);
    //~ }
    int const& bar() const &
    {
        std::cout << "const lvalue" << std::endl;
        return _bar;
    }
    int const&& bar() const &&
    {
        std::cout << "const rvalue" << std::endl;
        return std::move(_bar);
    }

    int _bar;
};

int main(int argc, char** argv)
{
    foo().bar();
}

Clang compiles it and outputs "const rvalue", while GCC thinks this is an ambiguous call with the two const-qualified functions both being best viable candidates. If I provide all 4 overloads, then both compilers output "non-const rvalue".

I would like to know which compiler --if any-- is doing the right thing, and what are the relevant standard pieces in play.

Note: The reason this actually matters is that the real code declares both const-qualified functions as constexpr. Of course, there is no output to std::cout and static_cast is used instead of std::move, so that they are valid constexpr definitions. And since in C++11 constexpr still implies const, the overload commented out in the sample code cannot be provided as it would redefine the const-qualified rvalue overload.

Solution 1:

Firstly, the implicit object parameter is treated as a normal parameter as per 13.3.1.4:

For non-static member functions, the type of the implicit object parameter is

— “lvalue reference to cv X” for functions declared without a ref-qualifier or with the & ref-qualifier

— “rvalue reference to cv X” for functions declared with the && ref-qualifier

where X is the class of which the function is a member and cv is the cv-qualification on the member function declaration.

So what you are asking is equivalent to the following:

void bar(foo&);
void bar(foo&&);
void bar(const foo&);
void bar(const foo&&);

int main()
{
    bar(foo());
}

The expression foo() is a class prvalue.

Secondly, the non-const lvalue reference version is not viable, as a prvalue cannot bind to it.

This leaves us with three viable functions for overload resolution.

Each has a single implicit object parameter (const foo&, foo&& or const foo&&), so we must rank these three to determine the best match.

In all three case it is a directly bound reference binding. This is described in declarators/initialization (8.5.3).

The ranking of the three possible bindings (const foo&, foo&& and const foo&&) is described in 13.3.3.2.3:

Standard conversion sequence S1 is a better conversion sequence than standard conversion sequence S2 if

• S1 and S2 are reference bindings and neither refers to an implicit object parameter of a non-static member function declared without a ref-qualifier [this exception doesn't apply here, they all have ref-qualifiers], and S1 binds an rvalue reference to an rvalue [a class prvalue is an rvalue] and S2 binds an lvalue reference.

This means that both foo&& and const foo&& are better then const foo&.

• S1 and S2 are reference bindings, and the types to which the references refer are the same type except for top-level cv-qualifiers, and the type to which the reference initialized by S2 refers is more cv-qualified than the type to which the reference initialized by S1 refers.

This means that foo&& is better than const foo&&.

So Clang is right, and it is a bug in GCC. The overload ranking for foo().bar() is as follows:

struct foo
{
    int&& bar() &&;             // VIABLE - BEST  (1)
    int const&& bar() const &&; // VIABLE -       (2)
    int const& bar() const &;   // VIABLE - WORST (3)
    int& bar() &;               // NOT VIABLE

    int _bar;
};

The bug in GCC seems to apply purely to implicit object parameters (with ref-qualifiers), for a normal parameter it seems to get the ranking correct, at least in 4.7.2.