<utility>: Side effects in self-swaps of pair are skipped

[frederick-vs-ja]

Describe the bug

Currently, self-swaps of pair are no-op:

STL/stl/inc/utility

Lines 441 to 460 in 9aca224

	_CONSTEXPR20 void swap(pair& _Right) noexcept(
	_Is_nothrow_swappable<_Ty1>::value && _Is_nothrow_swappable<_Ty2>::value) {
	using _STD swap;
	if (this != _STD addressof(_Right)) {
	swap(first, _Right.first); // intentional ADL
	swap(second, _Right.second); // intentional ADL
	}
	}
	#if _HAS_CXX23
	template <int = 0> // see GH-3013
	constexpr void swap(const pair& _Right) const
	noexcept(is_nothrow_swappable_v<const _Ty1> && is_nothrow_swappable_v<const _Ty2>) {
	using _STD swap;
	if (this != _STD addressof(_Right)) {
	swap(first, _Right.first); // intentional ADL
	swap(second, _Right.second); // intentional ADL
	}
	}
	#endif // _HAS_CXX23

However, the Cpp17Swappable requirements ([swappable.requirements]) doesn't seem to exclude self-swaps or side effects in them, and no special case is specified for self-swaps of pair.

Perhaps we should just unconditionally swap members.

Command-line test case

Godbolt link.

#include <utility>

struct swap_counter {
    unsigned int* pcnt_ = nullptr;

    friend constexpr void swap(swap_counter& lhs, swap_counter& rhs) noexcept
    {
        std::swap(lhs.pcnt_, rhs.pcnt_);
        if (lhs.pcnt_ != nullptr)
            ++*lhs.pcnt_;
        if (rhs.pcnt_ != nullptr)
            ++*lhs.pcnt_;
    }
};

static_assert([]
{
    unsigned int cnt{};
    std::pair<swap_counter, int> pr{swap_counter{&cnt}, 0};
    pr.swap(pr);
    return cnt == 2u;
}());

Expected behavior

The code snippet compiles.

[StephanTLavavej]

We looked at this during the weekly maintainer meeting and agree that the Standardese says we should be directly swapping both members without a self-address check.

@CaseyCarter remembered that this is vaguely related to #1485 but is distinct.