Proposed resolution:

This wording is relative to N4885.

1. Modify [range.subrange] as indicated:

   namespace std::ranges {
     template<class From, class To>
       concept uses-nonqualification-pointer-conversion = // exposition only
         is_pointer_v<From> && is_pointer_v<To> &&
         !convertible_to<remove_pointer_t<From>(*)[], remove_pointer_t<To>(*)[]>;
   
     template<class From, class To>
       concept convertible-to-non-slicing = // exposition only
         convertible_to<From, To> &&
         !uses-nonqualification-pointer-conversion<decay_t<From>, decay_t<To>>;
         !(is_pointer_v<decay_t<From>> &&
         is_pointer_v<decay_t<To>> &&
         not-same-as<remove_pointer_t<decay_t<From>>, remove_pointer_t<decay_t<To>>>
         );
     […]
   }
   […]
   

[ 2021-05-19 Tim adds wording ]

The wording below, which has been implemented and tested on top of libstdc++, uses the same technique we use for unique_ptr, shared_ptr, and span. It seems especially appropriate to have feature parity between subrange and span in this respect.

[ 2020-08-21; Reflector prioritization ]

Set priority to 3 after reflector discussions.

Consider

#include <ranges>

int main()
{
  int a[3] = { 1, 2, 3 };
  int* b[3] = { &a[2], &a[0], &a[1] };
  auto c = std::ranges::subrange<const int*const*>(b);
}

The construction of c is ill-formed because convertible-to-non-slicing<int**, const int*const*> is false, although the conversion does not involve object slicing.

I think subrange should allow such qualification conversion, just like unique_ptr<T[]> already does.

(Given that this constraint is useful in more than one context, maybe it deserves a named type trait?)