Understanding declval optimized implementation

Looking at libstdc++ source code, I found the following declval implementation:

template<typename _Tp, typename _Up = _Tp&&>
_Up __declval(int);  // (1)

template<typename _Tp>
_Tp __declval(long); // (2)

template<typename _Tp>
auto declval() noexcept -> decltype(__declval<_Tp>(0));

This implementation was proposed by Eric Niebler as a compile time optimization: he explains that overload resolution is faster than template instantiation.

However, I can't understand how it works. Specifically:

1. In (1), why is using _Up better than just returning _Tp&& ?
2. It seems that the overload (2) is never used. Why is it needed?

How all this prevents template instantiations, as opposed to the most naive implementation:

template<typename T>
T&& declval() noexcept;