34.4. Ranges

We've seen that there are two issues in the enumeration of ranges and the use of functors. First, we have a desire to avoid the general, though not universal, repeated stipulation of the begin() and end() methods for sequence types. Second, we have a need to avoid writing overly specific "write-once, use nowhere else" functors.

In a discussion with my friend John Torjo,^[7] we discovered that we've had similar disenchantments with these issues, and wanted a better solution. We came up with the Range concept. Naturally, he and I see the issue slightly differently;^[8] the concept definition and component implementations presented here primarily reflect my view.

^[7] John was also a one of the reviewers for this book.

^[8] You can fit more angels on the head of a pin than you can find software engineers who wouldn't argue over a blade of grass. Notwithstanding that, John and I are in a reasonable state of agreement. You can see the differences in our interpretation of the concept, and understand the different design principles and implementation mechanisms in our implementations, because John's expansive and impressive Boost-compatible implementation of the Range concept is included on the CD along with my own STLSoft version.

34.4.1 The Range Concept

The range concept is quite simple.

Definition: A Range represents a bounded collection of elements, which may be accessed in an incremental fashion. It encapsulates a logical range—that is, a beginning and end point, along with rules for how to walk through it (move from the beginning to the end point)—and embodies a single entity with which client code may access the values contained within the range.

Sounds almost too simple to be meaningful, doesn't it? Well, the salient part of the definition is that it's a single entity. The canonical form of use is:

for(R r = . . .; r; ++r)


{


  f(*r);


}

or, if you don't go for all that operator overloading:

for(R r = . . .; r.is_open(); r.advance())


{


  f(r.current());


}

There's noticeable brevity of form, which is part the raison d'être of ranges. A range has the characteristics shown in Table 34.2.

Table 34.2. Characteristics of a notional range.

Name	Expressions	Semantics	Precondition	Postcondition
Dereference	*r or r.current()	Returns the value of represented at the current position	r has not reached its end condition	r is unchanged
Advance	++r or r.advance()	Advances r's current position	r has not reached its end condition	r is dereferenceable or has reached its end condition
State	r or r.is_open()	Evaluates to true if r has reached its end condition, false otherwise	-	r is unchanged

Let's look at a couple of ranges in action:

// A range over a sequence


glob_sequence gs("/usr/include/", "impcpp*");


for(sequence_range<glob_sequence> r(gs); r; ++r)


{


  puts(*r); // Print the matched file entry to stdout


}





// A "notional range"


for(integral_range<int> r(10, 20, 2); r; ++r)


{


  cout << *r << endl; // Print int value at current point


}

There's no way any pointer can fit the range syntax^[9] so we don't have to worry about catering for any fundamental types; this provides us with a lot of extra flexibility in the implementation. In other words, we can rely on all range instances being of class types, and therefore we can rely on the presence or absence of class type characteristics in order to specialize behavior for the algorithms that operate on ranges (see section 34.4.4).

^[9] Except if we were to point it high into the top of memory and then iterate until it reached 0, but this would result in an access violation on many platforms, so the issue is moot.

The way this works is similar to the implementation of algorithms that manipulate Iterators [Aust1999], only simpler. My implementation of ranges simply relies on the specific range types deriving from the appropriate range type tag structure:

struct simple_range_tag


{};





struct iterable_range_tag


  : public simple_range_tag


{};

We'll see how these are used in the next sections.

34.4.2 Notional Range

Sometimes you don't have a concrete range, that is, one defined by two iterators; rather you have a notional range. It might be the odd numbers between 1 and 99. In this simple case you know that the number of odd numbers is 49 (1–97 inclusive), and so what you might do in classic STL is as shown in Listing 34.8:

Listing 34.8.

struct next_odd_number


{


  next_odd_number(int first)


    : m_current(first - 2)


  {}


  int operator ()()


  {


    return ++++m_current; // Nasty, but it fits on one line ;/


  }


private:


  int m_current;


};





std::vector<int>  ints(49);


std::generate(ints.begin(), ints.end(), next_odd_number(1));

Books on STL tend to show such things as perfectly reasonable examples of STL best practice. That may indeed be the case, but to me it's a whole lot of pain. First, we have the creation of a functor that will probably not be used elsewhere.

Second, and more significantly, the technique works by treating the functor as a passive producer whose actions are directed by the standard library algorithm generate(), whose bounds are, in turn, defined by the vector. This is very important, because it means that we need to know the number of results that will be created beforehand, in order to create spaces for them in the vector. Hence, we need a deep appreciation of how the producer—the functor—works. In this simple case, that is relatively straightforward—although I must confess I did get it wrong when preparing the test program! But imagine if we were computing the members of the Fibonacci series up to a user-supplied maximum value. It would be impossible to anticipate the number of steps in such a range other than by enumerating it to completion, which would be a drag, not to mention inefficient.

What we want in such cases is for the producer to drive the process, in accordance with the criteria stipulated by the code's author. Time to crank out my favorite example range, the integral_range template, whose simplified definition is shown in Listing 34.9:

Listing 34.9.

template <typename T>


class integral_range


  : public simple_range_tag


{


public:


  typedef simple_range_tag  range_tag_type;


  typedef T                 value_type;


// Construction


public:


  integral_range( value_type first, value_type last


                , value_type increment = +1)


  ~integral_range()


  {


    // Ensure that the parameterising type is integral


    STATIC_ASSERT(0 != is_integral_type<T>::value);


  }


// Enumeration


public:


  operator "safe bool"() const; // See Chapter 24


  value_type operator *() const;


  class_type &operator ++();


// Members


  . . .


};

The integral_range performs enumeration over its logical range via member variables of the given integral type, T.

We can now rewrite the instantiation of the vector with odd numbers as follows:

std::vector<int>  ints;


ints.reserve(49); // No problem if this is wrong.


r_copy(integral_range<int>(1, 99, 2), std::back_inserter(ints));

Note that the call to reserve() is an optimization and can be omitted without causing any change to the correctness of the code. r_copy() is a range algorithm (see section 34.4.3) that has the same semantics as the standard library algorithm copy() [Muss2001]. Now the producer, the instance of integral_range<int>, drives the process, which is as it should be. We could easily substitute a Fibonacci_range here, and the code would work correctly and efficiently, which cannot be said for the STL version.

34.4.3 Iterable Range

The Iterable range is an extension of the Notional range, with the additional provision of begin(), end() and range() methods, allowing it to be fully compatible with standard STL algorithms and usage (see section 34.4.4).

Iterable ranges usually maintain internal iterators reflecting the current and past-the-end condition positions, and return these via the begin() and end() methods. The range() method is provided to allow a given range to be cloned in its current enumeration state, although this is subject to the restrictions on copying iterators if the range's underlying iterators are of the Input or Output Iterator concepts [Aust1999, Muss2001]: only the Forward and "higher" models [Aust1999, Muss2001] support the ability to pass through a given range more than once.

It is possible to write iterable range classes, but the common idiom is to use an adaptor. In my implementation of the Range concept, I have two adaptors, the sequence_range and iterator_range template classes. Obviously the sequence_range adapts STL Sequences, and the iterator_range adapts STL Iterators.

Given a sequence type, we can adapt it to a range by passing the sequence instance to the constructor of an appropriate instantiation of sequence_range, as in:

std::deque<std::string> d = . . . ;





sequence_range< std::deque<std::string> > r(d);


for(; r; ++r)


{


  . . . // Use *r


}

Similarly, an iterator_range is constructed from a pair of iterators, as in:

vector<int> v = . . . ;





for(iterator_range<vector<int>::iterator> r(v.begin(), v.end());


    r; ++r)


{


  . . . // Use *r


}

Iterable ranges at first glance appear to be nothing more than a syntactic convenience for reducing the clutter of unwanted variables when processing iterator ranges in unrolled loops. I must confess that this in itself represents a big attraction to draw for me, but the concept offers much more as we'll see in the next couple of sections.

34.4.4 Range Algorithms and Tags

All the examples so far have shown ranges used in unrolled loops. Ranges may also be used with algorithms. That's where the separation of Notional and Iterable range concepts comes in.

Consider the standard library algorithm distance().

template <typename I>


size_t distance(I first, I last);

This algorithm returns the number of elements represented by the range [first, last). For all iterator types other than Random Access [Aust1999, Muss2001], the number is calculated by iterating first until it equals last. But the template is implemented to simply and efficiently calculate (last – first) for random access iterators. We certainly don't want to lose such efficiency gains when using algorithms with ranges.

The answer is very simple: we implement the range algorithms in terms of the standard library algorithms where possible. The range algorithm r_distance() is defined as shown in Listing 34.10:

Listing 34.10.

template <typename R>


ptrdiff_t r_distance_1(R r, iterable_range_tag const &)


{


  return std::distance(r.begin(), r.end());


}


template <typename R>


ptrdiff_t r_distance_1(R r, simple_range_tag const &)


{


  ptrdiff_t d = 0;


  for(; r; ++r, ++d)


  {}


  return d;


}


template <typename R>


inline ptrdiff_t r_distance(R r)


{


  return r_distance_1(r, r);


}

r_distance() is implemented in terms of r_distance_1(),^[10] which has two definitions: one for the iterable ranges that defers to a standard library algorithm, and one for notional ranges that carries out the enumeration manually. The two overloads of r_distance_1() are distinguished by their second parameter, which discriminate on whether the ranges as simple or iterable.

^[10] This is a general implementation naming policy, which helps to avoid clashes that can occur when the "outer" function and the implementing "inner" functions have the same name, and there are multiple overloads of the "outer" function. Different compilers have different aptitudes for resolving such things, so the simplest solution is to avoid it, and name the implementing "inner" functions unambiguously.

We use run time polymorphism (inheritance) to select the compile-time polymorphism (template type resolution), so we need to pass the range to r_distance_1() in two parameters, to preserve the actual type in the first parameter while at the same type discriminating the overload in the second. Since compilers can make mincemeat of such simple things, there're no concerns regarding inefficiency, and the mechanism ably suffices. We saw in section 34.4.2 that the integral_range template inherited from simple_range_tag. Iterable ranges inherit from iterable_range_tag. Hence the implementations of all range algorithms unambiguously select the most suitable implementation.

We can now come back and address the algorithm name problem (see section 34.2.2). Since we're clearly distinguishing between iterator-pair ranges and instances of a range class, we're never going to need the same name to facilitate generalized programming because the former ranges come as two parameters and the latter as one. Therefore, we can avoid all the namespace tribulations and just use algorithms with an r_ prefix.

34.4.5 Filters

Another powerful aspect of the Range concept abstraction is that of filters. This aspect of ranges is only just developing, but is already promising great benefits. I'll illustrate with a simple filter divisible, which we'll apply to our integral_range.

Listing 34.11.

template <typename R>


struct divisible


       : public R::range_tag_type


{


public:


  typedef R                       filtered_range_type;


  typedef typename R::value_type  value_type;


public:


  divisible(filtered_range_type r, value_type div)


    : m_r(r)


    , m_div(div)


  {


    assert(div > 0);


    for(; m_r && 0 != (*m_r % m_div); ++m_r)


    {}


  }


public:


  operator "safe bool"() const; // See Chapter 24


  {


    . . . // implemented in terms of m_r's op safe bool


  }


  value_type operator *() const


  {


    return *m_r;


  }


  class_type &operator ++()


  {


    for(; m_r && 0 != (*++m_r % m_div); )


    {}


    return *this;


  }


private:


  filtered_range_type m_r;


  value_type          m_div;


};

When combined with integral_range, we can filter out all the odd numbers in a given range to only those that are wholly divisible by three, as in:

std::vector<int>    ints;


integral_range<int> ir(1, 99, 2);


r_copy( divisible<integral_range<int> >(ir, 3)


      , std::back_inserter(ints));

Naturally range filters can be a lot more sophisticated than this in the real world.

34.4.6 Hypocrisy?

As someone who is against the abuse of operator overloading,^[11] I am certainly open to accusations of hypocrisy when it comes to the operators that ranges support. Clearly it can be argued, if one wished to be terribly uncharitable, that the unrealistic combination of operators provided by ranges is an abuse.

^[11] Although you'll have noted that I managed to overcome my squeamishness in Chapter 25.

I can't really defend the concept in any purist sense, so I can fall back on the Imperfect Practitioner's Philosophy, and state that the fact that it works so well, and so simply, is worth the pricked conscience.

If you like the range concept, but don't want to go for the operator overloading, it is a simple matter to replace some or all of the operators with methods. The semantics and efficiency will be the same; only the syntax will be slightly heavier.

for(R r = . . .; r.is_open(); r.advance())


{


  . . . = r.current();


}