How do I loop over consecutive pairs in an STL container using range-based loop syntax?

Question

How do I create a custom class to loop over consecutive pairs of items in a STL container using a range-based loop?

This is the syntax and output I want:

std::list<int> number_list;
number_list.push_back(1);
number_list.push_back(2);
number_list.push_back(3);

auto paired_list = Paired(number_list);
for (const auto & pair : paired_list) {
  std::printf("The pair is (%d, %d)\n", *(pair[0]), *(pair[1]));
  // or
  //std::printf("The pair is (%d, %d)\n", *(pair.first), *(pair.second));
}
// output:
// The pair is (1, 2)
// The pair is (2, 3)

I know these (and more) are needed, but I can't figure it out:

template <class T>
class Paired {
  ???
  class iterator {
    ???
  }
  iterator begin() {
    ...
  }
  iterator end() {
    ...
  }
}

Don't worry about const modifiers.

No boost.

Do not modify or copy objects in the container.

Answer 1

Here's what I would do.

#include <iterator>
#include <utility>

template <typename FwdIt> class adjacent_iterator {
public:
    adjacent_iterator(FwdIt first, FwdIt last)
        : m_first(first), m_next(first == last ? first : std::next(first)) { }

    bool operator!=(const adjacent_iterator& other) const {
        return m_next != other.m_next; // NOT m_first!
    }

    adjacent_iterator& operator++() {
        ++m_first;
        ++m_next;
        return *this;
    }

    typedef typename std::iterator_traits<FwdIt>::reference Ref;
    typedef std::pair<Ref, Ref> Pair;

    Pair operator*() const {
        return Pair(*m_first, *m_next); // NOT std::make_pair()!
    }

private:
    FwdIt m_first;
    FwdIt m_next;
};

template <typename FwdIt> class adjacent_range {
public:
    adjacent_range(FwdIt first, FwdIt last)
        : m_first(first), m_last(last) { }

    adjacent_iterator<FwdIt> begin() const {
        return adjacent_iterator<FwdIt>(m_first, m_last);
    }

    adjacent_iterator<FwdIt> end() const {
        return adjacent_iterator<FwdIt>(m_last, m_last);
    }

private:
    FwdIt m_first;
    FwdIt m_last;
};

template <typename C> auto make_adjacent_range(C& c) -> adjacent_range<decltype(c.begin())> {
    return adjacent_range<decltype(c.begin())>(c.begin(), c.end());
}

#include <iostream>
#include <vector>
using namespace std;

void test(const vector<int>& v) {
    cout << "[ ";

    for (const auto& p : make_adjacent_range(v)) {
        cout << p.first << "/" << p.second << " ";
    }

    cout << "]" << endl;
}

int main() {
    test({});
    test({11});
    test({22, 33});
    test({44, 55, 66});
    test({10, 20, 30, 40});
}

This prints:

[ ]
[ ]
[ 22/33 ]
[ 44/55 55/66 ]
[ 10/20 20/30 30/40 ]

Notes:

• I haven't exhaustively tested this, but it respects forward iterators (because it doesn't try to use operations beyond ++, !=, and *).

• range-for has extremely weak requirements; it doesn't require all of the things that forward iterators are supposed to provide. Therefore I have achieved range-for's requirements but no more.

• The "NOT m_first" comment is related to how the end of the range is approached. An adjacent_iterator constructed from an empty range has m_first == m_next which is also == last. An adjacent_iterator constructed from a 1-element range has m_first pointing to the element and m_next == last. An adjacent_iterator constructed from a multi-element range has m_first and m_next pointing to consecutive valid elements. As it is incremented, eventually m_first will point to the final element and m_next will be last. What adjacent_range's end() returns is constructed from (m_last, m_last). For a totally empty range, this is physically identical to begin(). For 1+ element ranges, this is not physically identical to a begin() that has been incremented until we don't have a complete pair - such iterators have m_first pointing to the final element. But if we compare iterators based on their m_next, then we get correct semantics.

• The "NOT std::make_pair()" comment is because make_pair() decays, while we actually want a pair of references. (I could have used decltype, but iterator_traits will tell us the answer too.)

• The major remaining subtleties would revolve around banning rvalues as inputs to make_adjacent_range (such temporaries would not have their lives prolonged; the Committee is studying this issue), and playing an ADL dance to respect non-member begin/end, as well as built-in arrays. These exercises are left to the reader.

Answer 2

edit I was using transform.

Use adjacent_difference.

The second version takes a binary function which transforms the two values into a new (different) value:

string make_message(int first, int second) {
    ostringstream oss;
    oss << "The pair is (" << first << ", " << second << ")";
    return oss.str();
}

We can now transform the adjacent pairs into a third range. We'll use the ostream_iterator to use cout like a range:

list<int> numbers;
//...
adjacent_difference(numbers.begin(), numbers.end(),
                    ostream_iterator<string>(cout, "\n"),
                    make_message);

2nd edit

I found a question on comp.lang.c++.moderated asking why there aren't more 'adjacent' functions in the standard library such as for_each_adjacent. The reply said they were trivial to implement using std::mismatch.

I think this would be a better direction to go than implementing a special adjacent iterator.