1 [/==============================================================================
2 Copyright (C) 2001-2011 Hartmut Kaiser
3 Copyright (C) 2001-2011 Joel de Guzman
4 Copyright (C) 2001-2002 Daniel C. Nuffer
6 Distributed under the Boost Software License, Version 1.0. (See accompanying
7 file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
8 ===============================================================================/]
10 [section:multi_pass The multi pass iterator]
12 Backtracking in __qi__ requires the use of the following types of iterator:
13 forward, bidirectional, or random access. Because of backtracking, input
14 iterators cannot be used. Therefore, the standard library classes
15 `std::istreambuf_iterator` and `std::istream_iterator`, that fall under the
16 category of input iterators, cannot be used. Another input iterator that is of
17 interest is one that wraps a lexer, such as LEX.
19 [note In general, __qi__ generates recursive descent parser which require
20 backtracking parsers by design. For this reason we need to provide at
21 least forward iterators to any of __qi__'s API functions. This is not an
22 absolute requirement though. In the future, we shall see more
23 deterministic parsers that require no more than 1 character (token) of
24 lookahead. Such parsers allow us to use input iterators such as the
25 `std::istream_iterator` as is. ]
27 Backtracking can be implemented only if we are allowed to save an iterator
28 position, i.e. making a copy of the current iterator. Unfortunately, with an
29 input iterator, there is no way to do so, and thus input iterators will not
30 work with backtracking in __qi__. One solution to this problem is to simply
31 load all the data to be parsed into a container, such as a vector or deque,
32 and then pass the begin and end of the container to __qi__. This method can be
33 too memory intensive for certain applications, which is why the `multi_pass`
36 [heading Using the multi_pass]
38 The `multi_pass` iterator will convert any input iterator into a forward
39 iterator suitable for use with __qi__. `multi_pass` will buffer data when
40 needed and will discard the buffer when its contents is not needed anymore.
41 This happens either if only one copy of the iterator exists or if no
42 backtracking can occur.
44 A grammar must be designed with care if the `multi_pass` iterator is used.
45 Any rule that may need to backtrack, such as one that contains an alternative,
46 will cause data to be buffered. The rules that are optimal to use are
47 repetition constructs (as kleene and plus).
49 Sequences of the form `a >> b` will buffer data as well. This is different from
50 the behavior of __classic__ but for a good reason. Sequences need to reset the
51 current iterator to its initial state if one of the components of a sequence
52 fails to match. To compensate for this behavior we added functionality to
53 the `expect` parsers (i.e. constructs like `a > b`). Expectation points introduce
54 deterministic points into the grammar ensuring no backtracking can occur if
55 they match. For this reason we clear the buffers of any multi_pass iterator
56 on each expectation point, ensuring minimal buffer content even for large
59 [important If you use an error handler in conjunction with the `expect` parser
60 while utilizing a `multi_pass` iterator and you intend to use
61 the error handler to force a `retry` or a `fail` (see the
62 description of error handlers - __fixme__: insert link), then
63 you need to instantiate the error handler using `retry` or `fail`,
66 rule r<iterator_type> r;
67 on_error<retry>(r, std::cout << phoenix::val("Error!"));
69 If you fail to do so the resulting code will trigger an assert
70 statement at runtime.]
72 Any rule that repeats, such as kleene_star (`*a`) or positive such as (`+a`),
73 will only buffer the data for the current repetition.
75 In typical grammars, ambiguity and therefore lookahead is often localized. In
76 fact, many well designed languages are fully deterministic and require no
77 lookahead at all. Peeking at the first character from the input will
78 immediately determine the alternative branch to take. Yet, even with highly
79 ambiguous grammars, alternatives are often of the form `*(a | b | c | d)`.
80 The input iterator moves on and is never stuck at the beginning. Let's look at
81 a Pascal snippet for example:
84 programHeading >> block >> '.'
88 *( labelDeclarationPart
89 | constantDefinitionPart
91 | variableDeclarationPart
92 | procedureAndFunctionDeclarationPart
97 Notice the alternatives inside the Kleene star in the rule block . The rule
98 gobbles the input in a linear manner and throws away the past history with each
99 iteration. As this is fully deterministic LL(1) grammar, each failed
100 alternative only has to peek 1 character (token). The alternative that consumes
101 more than 1 character (token) is definitely a winner. After which, the Kleene
102 star moves on to the next.
104 Now, after the lecture on the features to be careful with when using
105 `multi_pass`, you may think that `multi_pass` is way too restrictive to use.
106 That's not the case. If your grammar is deterministic, you can make use of
107 the `flush_multi_pass` pseudo parser in your grammar to ensure that data is not
108 buffered when unnecessary (`flush_multi_pass` is available from the __qi__
111 Here we present a minimal example showing a minimal use case. The `multi_pass`
112 iterator is highly configurable, but the default policies have been chosen so
113 that its easily usable with input iterators such as `std::istreambuf_iterator`.
114 For the complete source code of this example please refer to
115 [@../../example/support/multi_pass.cpp multi_pass.cpp].
117 [import ../../example/support/multi_pass.cpp]
118 [tutorial_multi_pass]
120 [heading Using the flush_multi_pass parser]
122 The __spirit__ __repo__ contains the `flush_multi_pass` parser component.
123 This is usable in conjunction with the `multi_pass` iterator to minimize the
124 buffering. It allows to insert explicit synchronization points into your
125 grammar where it is safe to clear any stored input as it is ensured that no
126 backtracking can occur at this point anymore.
128 When the `flush_multi_pass` parser is used with `multi_pass`, it will call
129 `multi_pass::clear_queue()`. This will cause any buffered data to be erased.
130 This also will invalidate all other copies of multi_pass and they should not
131 be used. If they are, an `boost::illegal_backtracking` exception will be
134 [heading The multi_pass Policies]
136 The `multi_pass` iterator is a templated class configurable using policies.
137 The description of `multi_pass` above is how it was originally implemented
138 (before it used policies), and is the default configuration now. But,
139 `multi_pass` is capable of much more. Because of the open-ended nature of
140 policies, you can write your own policy to make `multi_pass` behave in a way
141 that we never before imagined.
143 The multi_pass class has two template parameters:
145 [variablelist The multi_pass template parameters
146 [[Input] [The type multi_pass uses to acquire it's input. This is
147 typically an input iterator, or functor.]]
148 [[Policies] [The combined policies to use to create an instance of a
149 multi_pass iterator. This combined policy type is described
153 It is possible to implement all of the required functionality of the combined
154 policy in a single class. But it has shown to be more convenient to split this
155 into four different groups of functions, i.e. four separate, but well
156 coordinated policies. For this reason the `multi_pass` library
157 implements a template `iterator_policies::default_policy` allowing to combine
158 several different policies, each implementing one of the functionality groups:
160 [table Policies needed for default_policy template
161 [[Template Parameter] [Description]]
162 [[`OwnershipPolicy`] [This policy determines how `multi_pass` deals with
163 it's shared components.]]
164 [[`CheckingPolicy`] [This policy determines how checking for invalid
166 [[`InputPolicy`] [A class that defines how `multi_pass` acquires its
167 input. The `InputPolicy` is parameterized by the
168 `Input` template parameter to the `multi_pass`.]]
169 [[`StoragePolicy`] [The buffering scheme used by `multi_pass` is
170 determined and managed by the StoragePolicy.]]
173 The `multi_pass` library contains several predefined policy implementations
174 for each of the policy types as described above. First we will describe those
175 predefined types. Afterwards we will give some guidelines how you can write
176 your own policy implementations.
178 [heading Predefined policies]
180 All predefined `multi_pass` policies are defined in the namespace
181 `boost::spirit::iterator_policies`.
183 [table Predefined policy classes
184 [[Class name] [Description]]
185 [[*InputPolicy* classes]]
186 [[`input_iterator`] [This policy directs `multi_pass` to read from an
187 input iterator of type `Input`.]]
188 [[`buffering_input_iterator`] [This policy directs `multi_pass` to read from an
189 input iterator of type `Input`. Additionally it buffers
190 the last character received from the underlying iterator.
191 This allows to wrap iterators not buffering the last
192 character on their own (as `std::istreambuf_iterator`).]]
193 [[`istream`] [This policy directs `multi_pass` to read from an
194 input stream of type `Input` (usually a
195 `std::basic_istream`).]]
196 [[`lex_input`] [This policy obtains it's input by calling yylex(),
197 which would typically be provided by a scanner
198 generated by __flex__. If you use this policy your code
199 must link against a __flex__ generated scanner.]]
200 [[`functor_input`] [This input policy obtains it's data by calling a
201 functor of type `Input`. The functor must meet
202 certain requirements. It must have a typedef called
203 `result_type` which should be the type returned
204 from `operator()`. Also, since an input policy needs
205 a way to determine when the end of input has been
206 reached, the functor must contain a static variable
207 named `eof` which is comparable to a variable of
209 [[`split_functor_input`][This is essentially the same as the `functor_input`
210 policy except that the (user supplied) function
211 object exposes separate `unique` and `shared` sub
212 classes, allowing to integrate the functors /unique/
213 data members with the `multi_pass` data items held
214 by each instance and its /shared/ data members will
215 be integrated with the `multi_pass` members shared
218 [[*OwnershipPolicy* classes]]
219 [[`ref_counted`] [This class uses a reference counting scheme.
220 The `multi_pass` will delete it's shared components
221 when the count reaches zero.]]
222 [[`first_owner`] [When this policy is used, the first `multi_pass`
223 created will be the one that deletes the shared data.
224 Each copy will not take ownership of the shared data.
225 This works well for __spirit__, since no dynamic
226 allocation of iterators is done. All copies are made
227 on the stack, so the original iterator has the
230 [[*CheckingPolicy* classes]]
231 [[`no_check`] [This policy does no checking at all.]]
232 [[`buf_id_check`] [This policy keeps around a buffer id, or a buffer
233 age. Every time `clear_queue()` is called on a
234 `multi_pass` iterator, it is possible that all other
235 iterators become invalid. When `clear_queue()` is
236 called, `buf_id_check` increments the buffer id.
237 When an iterator is dereferenced, this policy checks
238 that the buffer id of the iterator matches the shared
239 buffer id. This policy is most effective when used
240 together with the `split_std_deque` StoragePolicy.
241 It should not be used with the `fixed_size_queue`
242 StoragePolicy, because it will not detect iterator
243 dereferences that are out of range.]]
244 [[full_check] [This policy has not been implemented yet. When it
245 is, it will keep track of all iterators and make
246 sure that they are all valid. This will be mostly
247 useful for debugging purposes as it will incur
248 significant overhead.]]
250 [[*StoragePolicy* classes]]
251 [[`split_std_deque`] [Despite its name this policy keeps all buffered data
252 in a `std::vector`. All data is stored as long as
253 there is more than one iterator. Once the iterator
254 count goes down to one, and the queue is no longer
255 needed, it is cleared, freeing up memory. The queue
256 can also be forcibly cleared by calling
257 `multi_pass::clear_queue()`.]]
258 [[`fixed_size_queue<N>`][This policy keeps a circular buffer that is size
259 `N+1` and stores `N` elements. `fixed_size_queue`
260 is a template with a `std::size_t` parameter that
261 specified the queue size. It is your responsibility
262 to ensure that `N` is big enough for your parser.
263 Whenever the foremost iterator is incremented, the
264 last character of the buffer is automatically
265 erased. Currently there is no way to tell if an
266 iterator is trailing too far behind and has become
267 invalid. No dynamic allocation is done by this
268 policy during normal iterator operation, only on
269 initial construction. The memory usage of this
270 `StoragePolicy` is set at `N+1` bytes, unlike
271 `split_std_deque`, which is unbounded.]]
274 [heading Combinations: How to specify your own custom multi_pass]
276 The beauty of policy based designs is that you can mix and match policies to
277 create your own custom iterator by selecting the policies you want. Here's an
278 example of how to specify a custom `multi_pass` that wraps an
279 `std::istream_iterator<char>`, and is slightly more efficient than the default
280 `multi_pass` (as generated by the `make_default_multi_pass()` API function)
281 because it uses the `iterator_policies::first_owner` OwnershipPolicy and the
282 `iterator_policies::no_check` CheckingPolicy:
285 std::istream_iterator<char>
286 , iterator_policies::default_policy<
287 iterator_policies::first_owner
288 , iterator_policies::no_check
289 , iterator_policies::buffering_input_iterator
290 , iterator_policies::split_std_deque
292 > first_owner_multi_pass_type;
294 The default template parameters for `iterator_policies::default_policy` are:
296 * `iterator_policies::ref_counted` OwnershipPolicy
297 * `iterator_policies::no_check` CheckingPolicy, if `BOOST_SPIRIT_DEBUG` is
298 defined: `iterator_policies::buf_id_check` CheckingPolicy
299 * `iterator_policies::buffering_input_iterator` InputPolicy, and
300 * `iterator_policies::split_std_deque` StoragePolicy.
302 So if you use `multi_pass<std::istream_iterator<char> >` you will get those
303 pre-defined behaviors while wrapping an `std::istream_iterator<char>`.
305 [heading Dealing with constant look ahead]
307 There is one other pre-defined class called `look_ahead`. The class
308 `look_ahead` is another predefine `multi_pass` iterator type. It has two
309 template parameters: `Input`, the type of the input iterator to wrap, and a
310 `std::size_t N`, which specifies the size of the buffer to the
311 `fixed_size_queue` policy. While the default multi_pass configuration is
312 designed for safety, `look_ahead` is designed for speed. `look_ahead` is derived
313 from a multi_pass with the following policies: `input_iterator` InputPolicy,
314 `first_owner` OwnershipPolicy, `no_check` CheckingPolicy, and
315 `fixed_size_queue<N>` StoragePolicy.
317 This iterator is defined by including the files:
319 // forwards to <boost/spirit/home/support/look_ahead.hpp>
320 #include <boost/spirit/include/support_look_ahead.hpp>
322 Also, see __include_structure__.
324 [heading Reading from standard input streams]
326 Yet another predefined iterator for wrapping standard input streams (usually a
327 `std::basic_istream<>`) is called `basic_istream_iterator<Char, Traits>`. This
328 class is usable as a drop in replacement for `std::istream_iterator<Char, Traits>`.
329 Its only difference is that it is a forward iterator (instead of the
330 `std::istream_iterator`, which is an input iterator). `basic_istream_iterator`
331 is derived from a multi_pass with the following policies: `istream` InputPolicy,
332 `ref_counted` OwnershipPolicy, `no_check` CheckingPolicy, and
333 `split_std_deque` StoragePolicy.
335 There exists an additional predefined typedef:
337 typedef basic_istream_iterator<char, std::char_traits<char> > istream_iterator;
339 This iterator is defined by including the files:
341 // forwards to <boost/spirit/home/support/istream_iterator.hpp>
342 #include <boost/spirit/include/support_istream_iterator.hpp>
344 Also, see __include_structure__.
346 [heading How to write a functor for use with the `functor_input` InputPolicy]
348 If you want to use the `functor_input` InputPolicy, you can write your own
349 function object that will supply the input to `multi_pass`. The function object
350 must satisfy several requirements. It must have a typedef `result_type` which
351 specifies the return type of its `operator()`. This is standard practice in the
352 STL. Also, it must supply a static variable called eof which is compared against
353 to know whether the input has reached the end. Last but not least the function
354 object must be default constructible. Here is an example:
357 #include <boost/spirit/home/qi.hpp>
358 #include <boost/spirit/home/support.hpp>
359 #include <boost/spirit/home/support/multi_pass.hpp>
360 #include <boost/spirit/home/support/iterators/detail/functor_input_policy.hpp>
362 // define the function object
366 typedef char result_type;
368 iterate_a2m() : c_('A') {}
369 iterate_a2m(char c) : c_(c) {}
371 result_type operator()()
378 static result_type eof;
384 iterate_a2m::result_type iterate_a2m::eof = iterate_a2m::result_type('M');
386 using namespace boost::spirit;
388 // create two iterators using the define function object, one of which is
390 typedef multi_pass<iterate_a2m
391 , iterator_policies::first_owner
392 , iterator_policies::no_check
393 , iterator_policies::functor_input
394 , iterator_policies::split_std_deque>
395 functor_multi_pass_type;
399 functor_multi_pass_type first = functor_multi_pass_type(iterate_a2m());
400 functor_multi_pass_type last;
402 // use the iterators: this will print "ABCDEFGHIJKL"
403 while (first != last) {
407 std::cout << std::endl;
411 [heading How to write policies for use with multi_pass]
413 All policies to be used with the `default_policy` template need to have two
414 embedded classes: `unique` and `shared`. The `unique` class needs to implement
415 all required functions for a particular policy type. In addition it may hold
416 all member data items being /unique/ for a particular instance of a `multi_pass`
417 (hence the name). The `shared` class does not expose any member functions
418 (except sometimes a constructor), but it may hold all member data items to be
419 /shared/ between all copies of a particular `multi_pass`.
421 [heading InputPolicy]
423 An `InputPolicy` must have the following interface:
427 // Input is the same type used as the first template parameter
428 // while instantiating the multi_pass
429 template <typename Input>
432 // these typedef's will be exposed as the multi_pass iterator
434 typedef __unspecified_type__ value_type;
435 typedef __unspecified_type__ difference_type;
436 typedef __unspecified_type__ distance_type;
437 typedef __unspecified_type__ pointer;
438 typedef __unspecified_type__ reference;
441 explicit unique(Input) {}
443 // destroy is called whenever the last copy of a multi_pass is
444 // destructed (ownership_policy::release() returned true)
446 // mp: is a reference to the whole multi_pass instance
447 template <typename MultiPass>
448 static void destroy(MultiPass& mp);
450 // swap is called by multi_pass::swap()
453 // get_input is called whenever the next input character/token
454 // should be fetched.
456 // mp: is a reference to the whole multi_pass instance
458 // This method is expected to return a reference to the next
460 template <typename MultiPass>
461 static typename MultiPass::reference get_input(MultiPass& mp);
463 // advance_input is called whenever the underlying input stream
464 // should be advanced so that the next call to get_input will be
465 // able to return the next input character/token
467 // mp: is a reference to the whole multi_pass instance
468 template <typename MultiPass>
469 static void advance_input(MultiPass& mp);
471 // input_at_eof is called to test whether this instance is a
472 // end of input iterator.
474 // mp: is a reference to the whole multi_pass instance
476 // This method is expected to return true if the end of input is
477 // reached. It is often used in the implementation of the function
478 // storage_policy::is_eof.
479 template <typename MultiPass>
480 static bool input_at_eof(MultiPass const& mp);
482 // input_is_valid is called to verify if the parameter t represents
483 // a valid input character/token
485 // mp: is a reference to the whole multi_pass instance
486 // t: is the character/token to test for validity
488 // This method is expected to return true if the parameter t
489 // represents a valid character/token.
490 template <typename MultiPass>
491 static bool input_is_valid(MultiPass const& mp, value_type const& t);
494 // Input is the same type used as the first template parameter passed
495 // while instantiating the multi_pass
496 template <typename Input>
499 explicit shared(Input) {}
503 It is possible to derive the struct `unique` from the type
504 `boost::spirit::detail::default_input_policy`. This type implements a minimal
505 sufficient interface for some of the required functions, simplifying the task
506 of writing a new input policy.
508 This class may implement a function `destroy()` being called during destruction
509 of the last copy of a `multi_pass`. This function should be used to free any of
510 the shared data items the policy might have allocated during construction of
511 its `shared` part. Because of the way `multi_pass` is implemented any allocated
512 data members in `shared` should _not_ be deep copied in a copy constructor of
515 [heading OwnershipPolicy]
517 The `OwnershipPolicy` must have the following interface:
519 struct ownership_policy
523 // destroy is called whenever the last copy of a multi_pass is
524 // destructed (ownership_policy::release() returned true)
526 // mp: is a reference to the whole multi_pass instance
527 template <typename MultiPass>
528 static void destroy(MultiPass& mp);
530 // swap is called by multi_pass::swap()
533 // clone is called whenever a multi_pass is copied
535 // mp: is a reference to the whole multi_pass instance
536 template <typename MultiPass>
537 static void clone(MultiPass& mp);
539 // release is called whenever a multi_pass is destroyed
541 // mp: is a reference to the whole multi_pass instance
543 // The method is expected to return true if the destructed
544 // instance is the last copy of a particular multi_pass.
545 template <typename MultiPass>
546 static bool release(MultiPass& mp);
548 // is_unique is called to test whether this instance is the only
549 // existing copy of a particular multi_pass
551 // mp: is a reference to the whole multi_pass instance
553 // The method is expected to return true if this instance is unique
554 // (no other copies of this multi_pass exist).
555 template <typename MultiPass>
556 static bool is_unique(MultiPass const& mp);
562 It is possible to derive the struct `unique` from the type
563 `boost::spirit::detail::default_ownership_policy`. This type implements a
564 minimal sufficient interface for some of the required functions, simplifying
565 the task of writing a new ownership policy.
567 This class may implement a function `destroy()` being called during destruction
568 of the last copy of a `multi_pass`. This function should be used to free any of
569 the shared data items the policy might have allocated during construction of
570 its `shared` part. Because of the way `multi_pass` is implemented any allocated
571 data members in `shared` should _not_ be deep copied in a copy constructor of
574 [heading CheckingPolicy]
576 The `CheckingPolicy` must have the following interface:
578 struct checking_policy
582 // swap is called by multi_pass::swap()
585 // destroy is called whenever the last copy of a multi_pass is
586 // destructed (ownership_policy::release() returned true)
588 // mp: is a reference to the whole multi_pass instance
589 template <typename MultiPass>
590 static void destroy(MultiPass& mp);
592 // docheck is called before the multi_pass is dereferenced or
595 // mp: is a reference to the whole multi_pass instance
597 // This method is expected to make sure the multi_pass instance is
598 // still valid. If it is invalid an exception should be thrown.
599 template <typename MultiPass>
600 static void docheck(MultiPass const& mp);
602 // clear_queue is called whenever the function
603 // multi_pass::clear_queue is called on this instance
605 // mp: is a reference to the whole multi_pass instance
606 template <typename MultiPass>
607 static void clear_queue(MultiPass& mp);
613 It is possible to derive the struct `unique` from the type
614 `boost::spirit::detail::default_checking_policy`. This type implements a
615 minimal sufficient interface for some of the required functions, simplifying
616 the task of writing a new checking policy.
618 This class may implement a function `destroy()` being called during destruction
619 of the last copy of a `multi_pass`. This function should be used to free any of
620 the shared data items the policy might have allocated during construction of
621 its `shared` part. Because of the way `multi_pass` is implemented any allocated
622 data members in `shared` should _not_ be deep copied in a copy constructor of
625 [heading StoragePolicy]
627 A `StoragePolicy` must have the following interface:
629 struct storage_policy
631 // Value is the same type as typename MultiPass::value_type
632 template <typename Value>
635 // destroy is called whenever the last copy of a multi_pass is
636 // destructed (ownership_policy::release() returned true)
638 // mp: is a reference to the whole multi_pass instance
639 template <typename MultiPass>
640 static void destroy(MultiPass& mp);
642 // swap is called by multi_pass::swap()
645 // dereference is called whenever multi_pass::operator*() is invoked
647 // mp: is a reference to the whole multi_pass instance
649 // This function is expected to return a reference to the current
651 template <typename MultiPass>
652 static typename MultiPass::reference dereference(MultiPass const& mp);
654 // increment is called whenever multi_pass::operator++ is invoked
656 // mp: is a reference to the whole multi_pass instance
657 template <typename MultiPass>
658 static void increment(MultiPass& mp);
661 // mp: is a reference to the whole multi_pass instance
662 template <typename MultiPass>
663 static void clear_queue(MultiPass& mp);
665 // is_eof is called to test whether this instance is a end of input
668 // mp: is a reference to the whole multi_pass instance
670 // This method is expected to return true if the end of input is
672 template <typename MultiPass>
673 static bool is_eof(MultiPass const& mp);
675 // less_than is called whenever multi_pass::operator==() is invoked
677 // mp: is a reference to the whole multi_pass instance
678 // rhs: is the multi_pass reference this instance is compared
681 // This function is expected to return true if the current instance
682 // is equal to the right hand side multi_pass instance
683 template <typename MultiPass>
684 static bool equal_to(MultiPass const& mp, MultiPass const& rhs);
686 // less_than is called whenever multi_pass::operator<() is invoked
688 // mp: is a reference to the whole multi_pass instance
689 // rhs: is the multi_pass reference this instance is compared
692 // This function is expected to return true if the current instance
693 // is less than the right hand side multi_pass instance
694 template <typename MultiPass>
695 static bool less_than(MultiPass const& mp, MultiPass const& rhs);
698 // Value is the same type as typename MultiPass::value_type
699 template <typename Value>
703 It is possible to derive the struct `unique` from the type
704 `boost::spirit::detail::default_storage_policy`. This type implements a
705 minimal sufficient interface for some of the required functions, simplifying
706 the task of writing a new storage policy.
708 This class may implement a function `destroy()` being called during destruction
709 of the last copy of a `multi_pass`. This function should be used to free any of
710 the shared data items the policy might have allocated during construction of
711 its `shared` part. Because of the way `multi_pass` is implemented any allocated
712 data members in `shared` should _not_ be deep copied in a copy constructor of
715 Generally, a `StoragePolicy` is the trickiest policy to implement. You should
716 study and understand the existing `StoragePolicy` classes before you try and