[D156626] Improve constraints in std::iter_move

[ldionne]

This issue tracks picking up https://reviews.llvm.org/D156626 from the Phabricator archive. It looks like the issue was real, but the commit message was insufficient.

[jwakely]

I've just rediscovered the bug that the phabricator patch fixes:

#include <iterator>

template<typename T>
concept iter_movable = requires { std::ranges::iter_move(std::declval<T>()); };

struct I2
{
    int& operator*() &;
    void operator*() &&;
};

static_assert( iter_movable<I2> );

The __move_deref concept is true, because is_lvalue_reference_v<decltype(*__t)> is true, so the second overload of __iter_move::__fn::operator() is used, but that does:

return std::move(*std::forward<_Ip>(__i));

This fails, because *std::forward<_Ip>(__i) dereferences an rvalue, which uses void I2::operator*() &&.

The fix is (as in the phab patch), for __move_deref to use the correct value category for __i

[jwakely]

The fix is apparently held up because there's no accurate description of the problem, so here you go:

The __move_deref and __just_deref concepts for ranges::iter_move treat the local parameter __t as an lvalue, which does not match what ranges::iter_move actually does. The overloads of __iter_move::__fn::operator() that use those concepts are implemented correctly, using *forward<_Ip>(__i) to preserve the value category of the parameter. The concepts don't do that, so they can select a different overload of operator*, and so they test whether the wrong overload returns an lvalue reference or not.

The concepts should use *forward<_Tp>(__t) to test the properties of the correct value category.