Closed ranges may be a problem; breaking counted_iterator is not the solution

Document #: P2799R0
Date: 2023-02-14
Project: Programming Language C++
Audience: SG9
LEWG
Reply-to: Tim Song
<>

1 Abstract

The US 46-107/[P2406R5] complaint boils down to a desire to adapt a closed range - something foreign to the C++ iterator model - into a half-open range, and making breaking changes to counted_iterator is the wrong way to solve that problem.

2 C++ ranges are half-open

C++ ranges are half-open.

This is fundamental to the C++ iterator model. As 25.3.1 [iterator.requirements.general] p7 explains:

Just as a regular pointer to an array guarantees that there is a pointer value pointing past the last element of the array, so for any iterator type there is an iterator value that points past the last element of a corresponding sequence. Such a value is called a past-the-end value.

A range of N elements has N + 1 unique iterator values. N values corresponding to the N elements, and the past-the-end value. This is iterator 101.

C++20 introduced the notion of counted ranges, but the same principle still applies. 25.3.1 [iterator.requirements.general] p11:

A counted range i + [0, n) is empty if n == 0; otherwise, i + [0, n) refers to the n elements in the data structure starting with the element pointed to by i and up to but not including the element, if any, pointed to by the result of n applications of ++i. A counted range i + [0, n) is valid if and only if n == 0; or n is positive, i is dereferenceable, and ++i + [0, --n) is valid.

3 What does counted_iterator do?

It attaches an integer count to an iterator:

template <input_or_output_iterator It>
struct counted_iterator {
  It it;
  int count;
};

Nothing less, nothing more. count changes as it moves, allowing the distance between two iterators to the same range to be easily computed.

Counting down instead of up (i.e., increment is ++it; --count; and decrementing is --it; ++count;) allows the past-the-end test for a simple counted range to be expressed as count == 0 regardless of what the number of elements in the range is, but that’s just an implementation detail to allow for the convenience of a stateless sentinel. It could have incremented count instead and have the past-the-end check be count == N, at the cost of a stateful sentinel that stores N. take_view, of course, adds its own sentinel in many cases, because the count we are given is the maximum number of elements to be taken, not the exact number: its test for past-the-end is it == end or count == 0.

4 Closed ranges

As it turns out, people occasionally find the requirement of a past-the-end value too constraining. They want to represent a range of N elements with just N iterator values - one for each element. In other words, they want a closed range, not a half-open one.

Consider the two examples given in [P2406R5]:

    auto iss = std::istringstream("0 1 2");
    for (auto i : rn::istream_view<int>(iss)
                  | rv::take(1))
        std::cout << i << '\n';
    auto i = 0;
    iss >> i;
    assert(i == 1); // FAILS, i == 2

Here, the actual problem is that istream_view expects all access to the stream to happen through its iterators. Nothing is actually lost if its iterator is consistently used. In the example above, we can extract the istream_view iterator out of the past-the-end counted_iterator using the allegedly useless base() function:

    auto iss = std::istringstream("0 1 2");
    auto tv = rn::istream_view<int>(iss)
                  | rv::take(1);
    auto it = tv.begin();
    for(; it != tv.end(); ++it)
        std::cout << *it << '\n';
    auto next = std::move(it).base();
    auto i = *next;
    assert(i == 1); // OK

But since the code wants to break istream_view’s expectation and rely on what is basically an implementation detail2, we are left again with a range of N elements that needs to represented using N iterator values.3

Of course there are other cases that could be aided by recognizing a notion of closed ranges. For example, iota_view can’t in general represent a range that contains the maximum of its value type, because there’s no valid past-the-end value for its iterator in that case.

But this notion of closed ranges is, as discussed above, foreign to the current C++ iterator model, and it is entirely unsurprising that counted_iterator isn’t designed for it.

5 Adapting a closed range into a half-open range

But let’s say we have a closed range of N elements for whatever reason and want to adapt it to a half-open range that everything else in C++ expects. What do we need to do?

Well, we have N iterator values to start with and have to come up with N + 1 ones for the half-open range. In other words, we need to invent a new iterator value out of thin air. The adapted iterator has a single type that needs to represent both the original positions and this past-the-end value everything else expects - it is effectively a variant<OriginalIterator, PastTheEndSentinel> - though perhaps not literally that.

It should be unsurprising that such an adaptation necessarily adds overhead and limits functionality - every increment needs to check “did I become the sentinel”; every decrement - if supported at all - would need to check “am I the sentinel”. And supporting decrement at all requires the PastTheEndSentinel to contain an OriginalIterator value. Returning non-void from postfix increment on input or output iterators cannot be supported - those can return a proxy, and we can’t invent one out of thin air. And since output iterators require *o++ to work, we lose support for those iterators entirely.

It’s worth emphasizing that this problem has nothing to do with having a count. Adapting a closed range to a half-open range can be done without one. Presumably the reason this first arose in the counted_iterator context is that having a decrementing count alongside the iterator happens to provide a natural representation for the PastTheEndSentinel state. On the other hand, having a count also leads to a confusing iterator construction API if decrement is to be supported - especially if the adaptor can be used with normal C++ iterators where support for past-the-end values is the norm.

As a concrete example, consider range-v3’s closed_iota, which represents an integer range [first, last]. Its iterator is basically

struct iterator {
  I current;
  I last;
  bool past_the_end;
};

where operator++ sets past_the_end to true (and does not increment current) when current == last.

6 Whatever the right approach for adapting a closed range into a half-open range is, counted_iterator is not it

Closed ranges are not a common occurrence in C++. How can they be, when types with no past-the-end values aren’t even iterator types and every algorithm in the standard expects half-open ranges?

There are certainly interesting uses of closed ranges, and relaxing the requirement for a past-the-end value naturally means that you can express more things. It may well be worthwhile to add a closed-to-half-open adaptation mechanism so that such ranges can be used with other things in the standard library.

But counted_iterator should not be that mechanism. It does the job it is designed for well.4 Making a breaking change that adds performance overhead, inhibits compiler optimizations and limits functionality - at the last minute before we ship C++23 no less - is simply the wrong way to support closed ranges.

7 Conclusion

There’s nothing wrong with the design of counted_iterator for the use cases it was designed for.

If we want a facility to adapt a fully-closed range to a half-open range, we can add one in C++26. That’s not counted_iterator’s job.

8 References

[LWG2471] Jonathan Wakely. copy_n’s number of InputIterator increments unspecified.
https://wg21.link/lwg2471

[P2406R5] Yehezkel Bernat, Yehuda Bernat. 2023. Add lazy_counted_iterator.
https://wg21.link/P2406R5


  1. “range” is in scary quotes because this combination does not actually meet the semantic requirements (due to the lack of a past-the-end value), despite meeting the syntactic requirements.↩︎

  2. range-v3 adds a workaround for this issue by providing a cached() member in its istream_view, so that the user can use tv.base().cached().↩︎

  3. It’s worth noting - as explained eight years ago in [LWG2471] - that anything that “fixes” istream_iterator (and istream_view - the two are indistinguishable for this purpose) - will likely break comparable uses of istreambuf_iterator instead.↩︎

  4. It should be very clear from the discussion in this paper that the author strongly disagrees with any suggestion that counted_iterator “has the wrong default” or that the proposed lazy_take/lazy_counted_iterator should be used by default.↩︎