iterator_facade_body.rst 8.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195
  1. .. Distributed under the Boost
  2. .. Software License, Version 1.0. (See accompanying
  3. .. file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
  4. .. Version 1.1 of this ReStructuredText document corresponds to
  5. n1530_, the paper accepted by the LWG for TR1.
  6. .. Copyright David Abrahams, Jeremy Siek, and Thomas Witt 2003.
  7. While the iterator interface is rich, there is a core subset of the
  8. interface that is necessary for all the functionality. We have
  9. identified the following core behaviors for iterators:
  10. * dereferencing
  11. * incrementing
  12. * decrementing
  13. * equality comparison
  14. * random-access motion
  15. * distance measurement
  16. In addition to the behaviors listed above, the core interface elements
  17. include the associated types exposed through iterator traits:
  18. ``value_type``, ``reference``, ``difference_type``, and
  19. ``iterator_category``.
  20. Iterator facade uses the Curiously Recurring Template
  21. Pattern (CRTP) [Cop95]_ so that the user can specify the behavior
  22. of ``iterator_facade`` in a derived class. Former designs used
  23. policy objects to specify the behavior, but that approach was
  24. discarded for several reasons:
  25. 1. the creation and eventual copying of the policy object may create
  26. overhead that can be avoided with the current approach.
  27. 2. The policy object approach does not allow for custom constructors
  28. on the created iterator types, an essential feature if
  29. ``iterator_facade`` should be used in other library
  30. implementations.
  31. 3. Without the use of CRTP, the standard requirement that an
  32. iterator's ``operator++`` returns the iterator type itself
  33. would mean that all iterators built with the library would
  34. have to be specializations of ``iterator_facade<...>``, rather
  35. than something more descriptive like
  36. ``indirect_iterator<T*>``. Cumbersome type generator
  37. metafunctions would be needed to build new parameterized
  38. iterators, and a separate ``iterator_adaptor`` layer would be
  39. impossible.
  40. Usage
  41. -----
  42. The user of ``iterator_facade`` derives his iterator class from a
  43. specialization of ``iterator_facade`` and passes the derived
  44. iterator class as ``iterator_facade``\ 's first template parameter.
  45. The order of the other template parameters have been carefully
  46. chosen to take advantage of useful defaults. For example, when
  47. defining a constant lvalue iterator, the user can pass a
  48. const-qualified version of the iterator's ``value_type`` as
  49. ``iterator_facade``\ 's ``Value`` parameter and omit the
  50. ``Reference`` parameter which follows.
  51. The derived iterator class must define member functions implementing
  52. the iterator's core behaviors. The following table describes
  53. expressions which are required to be valid depending on the category
  54. of the derived iterator type. These member functions are described
  55. briefly below and in more detail in the iterator facade
  56. requirements.
  57. +------------------------+-------------------------------+
  58. |Expression |Effects |
  59. +========================+===============================+
  60. |``i.dereference()`` |Access the value referred to |
  61. +------------------------+-------------------------------+
  62. |``i.equal(j)`` |Compare for equality with ``j``|
  63. +------------------------+-------------------------------+
  64. |``i.increment()`` |Advance by one position |
  65. +------------------------+-------------------------------+
  66. |``i.decrement()`` |Retreat by one position |
  67. +------------------------+-------------------------------+
  68. |``i.advance(n)`` |Advance by ``n`` positions |
  69. +------------------------+-------------------------------+
  70. |``i.distance_to(j)`` |Measure the distance to ``j`` |
  71. +------------------------+-------------------------------+
  72. .. Should we add a comment that a zero overhead implementation of iterator_facade
  73. is possible with proper inlining?
  74. In addition to implementing the core interface functions, an iterator
  75. derived from ``iterator_facade`` typically defines several
  76. constructors. To model any of the standard iterator concepts, the
  77. iterator must at least have a copy constructor. Also, if the iterator
  78. type ``X`` is meant to be automatically interoperate with another
  79. iterator type ``Y`` (as with constant and mutable iterators) then
  80. there must be an implicit conversion from ``X`` to ``Y`` or from ``Y``
  81. to ``X`` (but not both), typically implemented as a conversion
  82. constructor. Finally, if the iterator is to model Forward Traversal
  83. Iterator or a more-refined iterator concept, a default constructor is
  84. required.
  85. Iterator Core Access
  86. --------------------
  87. ``iterator_facade`` and the operator implementations need to be able
  88. to access the core member functions in the derived class. Making the
  89. core member functions public would expose an implementation detail to
  90. the user. The design used here ensures that implementation details do
  91. not appear in the public interface of the derived iterator type.
  92. Preventing direct access to the core member functions has two
  93. advantages. First, there is no possibility for the user to accidently
  94. use a member function of the iterator when a member of the value_type
  95. was intended. This has been an issue with smart pointer
  96. implementations in the past. The second and main advantage is that
  97. library implementers can freely exchange a hand-rolled iterator
  98. implementation for one based on ``iterator_facade`` without fear of
  99. breaking code that was accessing the public core member functions
  100. directly.
  101. In a naive implementation, keeping the derived class' core member
  102. functions private would require it to grant friendship to
  103. ``iterator_facade`` and each of the seven operators. In order to
  104. reduce the burden of limiting access, ``iterator_core_access`` is
  105. provided, a class that acts as a gateway to the core member functions
  106. in the derived iterator class. The author of the derived class only
  107. needs to grant friendship to ``iterator_core_access`` to make his core
  108. member functions available to the library.
  109. .. This is no long uptodate -thw
  110. .. Yes it is; I made sure of it! -DWA
  111. ``iterator_core_access`` will be typically implemented as an empty
  112. class containing only private static member functions which invoke the
  113. iterator core member functions. There is, however, no need to
  114. standardize the gateway protocol. Note that even if
  115. ``iterator_core_access`` used public member functions it would not
  116. open a safety loophole, as every core member function preserves the
  117. invariants of the iterator.
  118. ``operator[]``
  119. --------------
  120. The indexing operator for a generalized iterator presents special
  121. challenges. A random access iterator's ``operator[]`` is only
  122. required to return something convertible to its ``value_type``.
  123. Requiring that it return an lvalue would rule out currently-legal
  124. random-access iterators which hold the referenced value in a data
  125. member (e.g. |counting|_), because ``*(p+n)`` is a reference
  126. into the temporary iterator ``p+n``, which is destroyed when
  127. ``operator[]`` returns.
  128. .. |counting| replace:: ``counting_iterator``
  129. Writable iterators built with ``iterator_facade`` implement the
  130. semantics required by the preferred resolution to `issue 299`_ and
  131. adopted by proposal n1550_: the result of ``p[n]`` is an object
  132. convertible to the iterator's ``value_type``, and ``p[n] = x`` is
  133. equivalent to ``*(p + n) = x`` (Note: This result object may be
  134. implemented as a proxy containing a copy of ``p+n``). This approach
  135. will work properly for any random-access iterator regardless of the
  136. other details of its implementation. A user who knows more about
  137. the implementation of her iterator is free to implement an
  138. ``operator[]`` that returns an lvalue in the derived iterator
  139. class; it will hide the one supplied by ``iterator_facade`` from
  140. clients of her iterator.
  141. .. _n1550: http://www.open-std.org/JTC1/SC22/WG21/docs/papers/2003/n1550.htm
  142. .. _`issue 299`: http://www.open-std.org/jtc1/sc22/wg21/docs/lwg-active.html#299
  143. .. _`operator arrow`:
  144. ``operator->``
  145. --------------
  146. The ``reference`` type of a readable iterator (and today's input
  147. iterator) need not in fact be a reference, so long as it is
  148. convertible to the iterator's ``value_type``. When the ``value_type``
  149. is a class, however, it must still be possible to access members
  150. through ``operator->``. Therefore, an iterator whose ``reference``
  151. type is not in fact a reference must return a proxy containing a copy
  152. of the referenced value from its ``operator->``.
  153. The return types for ``iterator_facade``\ 's ``operator->`` and
  154. ``operator[]`` are not explicitly specified. Instead, those types
  155. are described in terms of a set of requirements, which must be
  156. satisfied by the ``iterator_facade`` implementation.
  157. .. [Cop95] [Coplien, 1995] Coplien, J., Curiously Recurring Template
  158. Patterns, C++ Report, February 1995, pp. 24-27.