Proposed resolution:

This wording is relative to N4964.

1. Modify [view.interface], class template view_interface synopsis, as indicated:

   namespace std::ranges {
     template<class D>
       requires is_class_v<D> && same_as<D, remove_cv_t<D>>
     class view_interface {
       […]
     public:
       […]
       constexpr decltype(auto) back() requires (bidirectional_range<D> && common_range<D>) ||
                                                (random_access_range<D> && sized_range<D>);
       constexpr decltype(auto) back() const
         requires (bidirectional_range<const D> && common_range<const D>) ||
                  (random_access_range<const D> && sized_range<const D>);
       […]
     };
   }
   

2. Modify [view.interface.members] as indicated:

   constexpr decltype(auto) back() requires (bidirectional_range<D> && common_range<D>) ||
                                            (random_access_range<D> && sized_range<D>);
   constexpr decltype(auto) back() const
     requires (bidirectional_range<const D> && common_range<const D>) ||
              (random_access_range<const D> && sized_range<const D>);
   

   -3- Preconditions: !empty() is true.

   -4- Effects: Equivalent to:

   auto common-arg-end = []<class R>(R& r) {
     if constexpr (common_range<R>) {
       return ranges::end(r);
     } else {
       return ranges::begin(r) + ranges::distance(r);
     }
   };
   return *ranges::prev(common-arg-endranges::end(derived()));
   

[ 2023-11-07; Reflector poll ]

NAD. "During the concat discussion LEWG decided not to support the corner case of random-access sized but not-common ranges." "If we did want to address such ranges, would be better to enforce commonness for random-access sized ranges by having ranges::end return ranges::begin(r) + ranges::size(r)."

Currently, view_interface only provides the back member when the derived class satisfies both bidirectional_range and common_range, which ensures that ranges::prev can act its sentinel.

However, requiring common_range seems to be too strict because when the derived class satisfies both random_access_range and sized_range, its end iterator can still be calculated in constant time, which is what some range adaptors currently do to greedily become common ranges.

I think we should follow similar rules to eliminate this inconsistency (demo):

#include <ranges>

constexpr auto r = std::ranges::subrange(std::views::iota(0), 5);
constexpr auto z = std::views::zip(r);
static_assert(r.back() == 4); // ill-formed
static_assert(std::get<0>(z.back()) == 4); // ok