Flattening iterator

Question

Is there any existing iterator implementation (perhaps in boost) which implement some sort of flattening iterator?

For example:

unordered_set<vector<int> > s;

s.insert(vector<int>());
s.insert({1,2,3,4,5});
s.insert({6,7,8});
s.insert({9,10,11,12});

flattening_iterator<unordered_set<vector<int> >::iterator> it( ... ), end( ... );
for(; it != end; ++it)
{
    cout << *it << endl;
}
//would print the numbers 1 through 12

Answer 1

you can make one using iterator facade in boost.

I wrote iterator product which you can use as a template perhaps: http://code.google.com/p/asadchev/source/browse/trunk/work/cxx/iterator/product.hpp

Answer 2

I don't know of any implementation in a major library, but it looked like an interesting problem so I wrote a basic implementation. I've only tested it with the test case I present here, so I don't recommend using it without further testing.

The problem is a bit trickier than it looks because some of the "inner" containers may be empty and you have to skip over them. This means that advancing the flattening_iterator by one position may actually advance the iterator into the "outer" container by more than one position. Because of this, the flattening_iterator needs to know where the end of the outer range is so that it knows when it needs to stop.

This implementation is a forward iterator. A bidirectional iterator would also need to keep track of the beginning of the outer range. The flatten function templates are used to make constructing flattening_iterators a bit easier.

#include <iterator>

// A forward iterator that "flattens" a container of containers.  For example,
// a vector<vector<int>> containing { { 1, 2, 3 }, { 4, 5, 6 } } is iterated as
// a single range, { 1, 2, 3, 4, 5, 6 }.
template <typename OuterIterator>
class flattening_iterator
{
public:

    typedef OuterIterator                                outer_iterator;
    typedef typename OuterIterator::value_type::iterator inner_iterator;

    typedef std::forward_iterator_tag                iterator_category;
    typedef typename inner_iterator::value_type      value_type;
    typedef typename inner_iterator::difference_type difference_type;
    typedef typename inner_iterator::pointer         pointer;
    typedef typename inner_iterator::reference       reference;

    flattening_iterator() { }
    flattening_iterator(outer_iterator it) : outer_it_(it), outer_end_(it) { }
    flattening_iterator(outer_iterator it, outer_iterator end) 
        : outer_it_(it), 
          outer_end_(end)
    { 
        if (outer_it_ == outer_end_) { return; }

        inner_it_ = outer_it_->begin();
        advance_past_empty_inner_containers();
    }

    reference operator*()  const { return *inner_it_;  }
    pointer   operator->() const { return &*inner_it_; }

    flattening_iterator& operator++()
    {
        ++inner_it_;
        if (inner_it_ == outer_it_->end())
            advance_past_empty_inner_containers();
        return *this;
    }

    flattening_iterator operator++(int)
    {
        flattening_iterator it(*this);
        ++*this;
        return it;
    }

    friend bool operator==(const flattening_iterator& a, 
                           const flattening_iterator& b)
    {
        if (a.outer_it_ != b.outer_it_)
            return false;

        if (a.outer_it_ != a.outer_end_ && 
            b.outer_it_ != b.outer_end_ &&
            a.inner_it_ != b.inner_it_)
            return false;

        return true;
    }

    friend bool operator!=(const flattening_iterator& a,
                           const flattening_iterator& b)
    {
        return !(a == b);
    }

private:

    void advance_past_empty_inner_containers()
    {
        while (outer_it_ != outer_end_ && inner_it_ == outer_it_->end())
        {
            ++outer_it_;
            if (outer_it_ != outer_end_) 
                inner_it_ = outer_it_->begin();
        }
    }

    outer_iterator outer_it_;
    outer_iterator outer_end_;
    inner_iterator inner_it_;
};

template <typename Iterator>
flattening_iterator<Iterator> flatten(Iterator it)
{
    return flattening_iterator<Iterator>(it, it);
}

template <typename Iterator>
flattening_iterator<Iterator> flatten(Iterator first, Iterator last)
{
    return flattening_iterator<Iterator>(first, last);
}

The following is a minimal test stub:

#include <algorithm>
#include <iostream>
#include <set>
#include <vector>

int main()
{
    // Generate some test data:  it looks like this:
    // { { 0, 1, 2, 3 }, { 4, 5, 6, 7 }, { 8, 9, 10, 11 } }
    std::vector<std::vector<int>> v(3);
    int i(0);
    for (auto it(v.begin()); it != v.end(); ++it)
    {
        it->push_back(i++); it->push_back(i++);
        it->push_back(i++); it->push_back(i++);
    }

    // Flatten the data and print all the elements:
    for (auto it(flatten(v.begin(), v.end())); it != v.end(); ++it)
    {
        std::cout << *it << ", ";
    }
    std::cout << "\n";

    // Or, since the standard library algorithms are awesome:
    std::copy(flatten(v.begin(), v.end()), flatten(v.end()), 
              std::ostream_iterator<int>(std::cout, ", "));
}

Like I said at the beginning, I haven't tested this thoroughly. Let me know if you find any bugs and I'll be happy to correct them.

Answer 3

I decided to "improve" a bit on the flattening iterator concept, though as noted by James you are stuck using Ranges (except for the inner most container), so I just used ranges through and through and thus obtained a flattened range, with an arbitrary depth.

First I used a building brick:

template <typename C>
struct iterator { using type = typename C::iterator; };

template <typename C>
struct iterator<C const> { using type = typename C::const_iterator; };

And then defined a (very minimal) ForwardRange concept:

template <typename C>
class ForwardRange {
    using Iter = typename iterator<C>::type;
public:
    using pointer = typename std::iterator_traits<Iter>::pointer;
    using reference = typename std::iterator_traits<Iter>::reference;
    using value_type = typename std::iterator_traits<Iter>::value_type;

    ForwardRange(): _begin(), _end() {}

    explicit ForwardRange(C& c): _begin(begin(c)), _end(end(c)) {}

    // Observers
    explicit operator bool() const { return _begin != _end; }

    reference operator*() const { assert(*this); return *_begin; }
    pointer operator->() const { assert(*this); return &*_begin; }

    // Modifiers
    ForwardRange& operator++() { assert(*this); ++_begin; return *this; }
    ForwardRange operator++(int) { ForwardRange tmp(*this); ++*this; return tmp; }

private:
    Iter _begin;
    Iter _end;
}; // class ForwardRange

This is our building brick here, though in fact we could make do with just the rest:

template <typename C, size_t N>
class FlattenedForwardRange {
    using Iter = typename iterator<C>::type;
    using Inner = FlattenedForwardRange<typename std::iterator_traits<Iter>::value_type, N-1>;
public:
    using pointer = typename Inner::pointer;
    using reference = typename Inner::reference;
    using value_type = typename Inner::value_type;

    FlattenedForwardRange(): _outer(), _inner() {}

    explicit FlattenedForwardRange(C& outer): _outer(outer), _inner() {
        if (not _outer) { return; }
        _inner = Inner{*_outer};
        this->advance();
    }

    // Observers
    explicit operator bool() const { return static_cast<bool>(_outer); }

    reference operator*() const { assert(*this); return *_inner; }
    pointer operator->() const { assert(*this); return _inner.operator->(); }

    // Modifiers
    FlattenedForwardRange& operator++() { ++_inner; this->advance(); return *this; }
    FlattenedForwardRange operator++(int) { FlattenedForwardRange tmp(*this); ++*this; return tmp; }

private:
    void advance() {
        if (_inner) { return; }

        for (++_outer; _outer; ++_outer) {
            _inner = Inner{*_outer};
            if (_inner) { return; }
        }
        _inner = Inner{};
    }

    ForwardRange<C> _outer;
    Inner _inner;
}; // class FlattenedForwardRange

template <typename C>
class FlattenedForwardRange<C, 0> {
    using Iter = typename iterator<C>::type;
public:
    using pointer = typename std::iterator_traits<Iter>::pointer;
    using reference = typename std::iterator_traits<Iter>::reference;
    using value_type = typename std::iterator_traits<Iter>::value_type;

    FlattenedForwardRange(): _range() {}

    explicit FlattenedForwardRange(C& c): _range(c) {}

    // Observers
    explicit operator bool() const { return static_cast<bool>(_range); }

    reference operator*() const { return *_range; }
    pointer operator->() const { return _range.operator->(); }

    // Modifiers
    FlattenedForwardRange& operator++() { ++_range; return *this; }
    FlattenedForwardRange operator++(int) { FlattenedForwardRange tmp(*this); ++*this; return tmp; }

private:
    ForwardRange<C> _range;
}; // class FlattenedForwardRange

And apparently, it works

Answer 4

I arrive a little late here, but I have just published a library (multidim) to deal with such problem. The usage is quite simple: to use your example,

#include "multidim.hpp"

// ... create "s" as in your example ...

auto view = multidim::makeFlatView(s);
// view offers now a flattened view on s

// You can now use iterators...
for (auto it = begin(view); it != end(view); ++it) cout << *it << endl;

// or a simple range-for loop
for (auto value : view) cout << value;

The library is header-only and has no dependencies. Requires C++11 though.