xref: /netbsd-src/external/gpl3/gcc/dist/libstdc++-v3/doc/xml/manual/containers.xml (revision b7b7574d3bf8eeb51a1fa3977b59142ec6434a55)
1<chapter xmlns="http://docbook.org/ns/docbook" version="5.0"
2	 xml:id="std.containers" xreflabel="Containers">
3<?dbhtml filename="containers.html"?>
4
5<info><title>
6  Containers
7  <indexterm><primary>Containers</primary></indexterm>
8</title>
9  <keywordset>
10    <keyword>ISO C++</keyword>
11    <keyword>library</keyword>
12  </keywordset>
13</info>
14
15
16
17<!-- Sect1 01 : Sequences -->
18<section xml:id="std.containers.sequences" xreflabel="Sequences"><info><title>Sequences</title></info>
19<?dbhtml filename="sequences.html"?>
20
21
22<section xml:id="containers.sequences.list" xreflabel="list"><info><title>list</title></info>
23<?dbhtml filename="list.html"?>
24
25  <section xml:id="sequences.list.size" xreflabel="list::size() is O(n)"><info><title>list::size() is O(n)</title></info>
26
27   <para>
28     Yes it is, and that's okay.  This is a decision that we preserved
29     when we imported SGI's STL implementation.  The following is
30     quoted from <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://www.sgi.com/tech/stl/FAQ.html">their FAQ</link>:
31   </para>
32   <blockquote>
33     <para>
34       The size() member function, for list and slist, takes time
35       proportional to the number of elements in the list.  This was a
36       deliberate tradeoff.  The only way to get a constant-time
37       size() for linked lists would be to maintain an extra member
38       variable containing the list's size.  This would require taking
39       extra time to update that variable (it would make splice() a
40       linear time operation, for example), and it would also make the
41       list larger.  Many list algorithms don't require that extra
42       word (algorithms that do require it might do better with
43       vectors than with lists), and, when it is necessary to maintain
44       an explicit size count, it's something that users can do
45       themselves.
46     </para>
47     <para>
48       This choice is permitted by the C++ standard. The standard says
49       that size() <quote>should</quote> be constant time, and
50       <quote>should</quote> does not mean the same thing as
51       <quote>shall</quote>.  This is the officially recommended ISO
52       wording for saying that an implementation is supposed to do
53       something unless there is a good reason not to.
54      </para>
55      <para>
56	One implication of linear time size(): you should never write
57      </para>
58	 <programlisting>
59	 if (L.size() == 0)
60	     ...
61	 </programlisting>
62
63	 <para>
64	 Instead, you should write
65	 </para>
66
67	 <programlisting>
68	 if (L.empty())
69	     ...
70	 </programlisting>
71   </blockquote>
72  </section>
73</section>
74
75<section xml:id="containers.sequences.vector" xreflabel="vector"><info><title>vector</title></info>
76<?dbhtml filename="vector.html"?>
77
78  <para>
79  </para>
80  <section xml:id="sequences.vector.management" xreflabel="Space Overhead Management"><info><title>Space Overhead Management</title></info>
81
82   <para>
83     In <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://gcc.gnu.org/ml/libstdc++/2002-04/msg00105.html">this
84     message to the list</link>, Daniel Kostecky announced work on an
85     alternate form of <code>std::vector</code> that would support
86     hints on the number of elements to be over-allocated.  The design
87     was also described, along with possible implementation choices.
88   </para>
89   <para>
90     The first two alpha releases were announced <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://gcc.gnu.org/ml/libstdc++/2002-07/msg00048.html">here</link>
91     and <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://gcc.gnu.org/ml/libstdc++/2002-07/msg00111.html">here</link>.
92   </para>
93
94  </section></section>
95</section>
96
97<!-- Sect1 02 : Associative -->
98<section xml:id="std.containers.associative" xreflabel="Associative"><info><title>Associative</title></info>
99<?dbhtml filename="associative.html"?>
100
101
102  <section xml:id="containers.associative.insert_hints" xreflabel="Insertion Hints"><info><title>Insertion Hints</title></info>
103
104   <para>
105     Section [23.1.2], Table 69, of the C++ standard lists this
106     function for all of the associative containers (map, set, etc):
107   </para>
108   <programlisting>
109      a.insert(p,t);
110   </programlisting>
111   <para>
112     where 'p' is an iterator into the container 'a', and 't' is the
113     item to insert.  The standard says that <quote><code>t</code> is
114     inserted as close as possible to the position just prior to
115     <code>p</code>.</quote> (Library DR #233 addresses this topic,
116     referring to <link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2005/n1780.html">N1780</link>.
117     Since version 4.2 GCC implements the resolution to DR 233, so
118     that insertions happen as close as possible to the hint. For
119     earlier releases the hint was only used as described below.
120   </para>
121   <para>
122     Here we'll describe how the hinting works in the libstdc++
123     implementation, and what you need to do in order to take
124     advantage of it.  (Insertions can change from logarithmic
125     complexity to amortized constant time, if the hint is properly
126     used.)  Also, since the current implementation is based on the
127     SGI STL one, these points may hold true for other library
128     implementations also, since the HP/SGI code is used in a lot of
129     places.
130   </para>
131   <para>
132     In the following text, the phrases <emphasis>greater
133     than</emphasis> and <emphasis>less than</emphasis> refer to the
134     results of the strict weak ordering imposed on the container by
135     its comparison object, which defaults to (basically)
136     <quote>&lt;</quote>.  Using those phrases is semantically sloppy,
137     but I didn't want to get bogged down in syntax.  I assume that if
138     you are intelligent enough to use your own comparison objects,
139     you are also intelligent enough to assign <quote>greater</quote>
140     and <quote>lesser</quote> their new meanings in the next
141     paragraph.  *grin*
142   </para>
143   <para>
144     If the <code>hint</code> parameter ('p' above) is equivalent to:
145   </para>
146     <itemizedlist>
147      <listitem>
148	<para>
149	  <code>begin()</code>, then the item being inserted should
150	  have a key less than all the other keys in the container.
151	  The item will be inserted at the beginning of the container,
152	  becoming the new entry at <code>begin()</code>.
153      </para>
154      </listitem>
155      <listitem>
156	<para>
157	  <code>end()</code>, then the item being inserted should have
158	  a key greater than all the other keys in the container.  The
159	  item will be inserted at the end of the container, becoming
160	  the new entry before <code>end()</code>.
161      </para>
162      </listitem>
163      <listitem>
164	<para>
165	  neither <code>begin()</code> nor <code>end()</code>, then:
166	  Let <code>h</code> be the entry in the container pointed to
167	  by <code>hint</code>, that is, <code>h = *hint</code>.  Then
168	  the item being inserted should have a key less than that of
169	  <code>h</code>, and greater than that of the item preceding
170	  <code>h</code>.  The new item will be inserted between
171	  <code>h</code> and <code>h</code>'s predecessor.
172	  </para>
173      </listitem>
174     </itemizedlist>
175   <para>
176     For <code>multimap</code> and <code>multiset</code>, the
177     restrictions are slightly looser: <quote>greater than</quote>
178     should be replaced by <quote>not less than</quote>and <quote>less
179     than</quote> should be replaced by <quote>not greater
180     than.</quote> (Why not replace greater with
181     greater-than-or-equal-to?  You probably could in your head, but
182     the mathematicians will tell you that it isn't the same thing.)
183   </para>
184   <para>
185     If the conditions are not met, then the hint is not used, and the
186     insertion proceeds as if you had called <code> a.insert(t)
187     </code> instead.  (<emphasis>Note </emphasis> that GCC releases
188     prior to 3.0.2 had a bug in the case with <code>hint ==
189     begin()</code> for the <code>map</code> and <code>set</code>
190     classes.  You should not use a hint argument in those releases.)
191   </para>
192   <para>
193     This behavior goes well with other containers'
194     <code>insert()</code> functions which take an iterator: if used,
195     the new item will be inserted before the iterator passed as an
196     argument, same as the other containers.
197   </para>
198   <para>
199     <emphasis>Note </emphasis> also that the hint in this
200     implementation is a one-shot.  The older insertion-with-hint
201     routines check the immediately surrounding entries to ensure that
202     the new item would in fact belong there.  If the hint does not
203     point to the correct place, then no further local searching is
204     done; the search begins from scratch in logarithmic time.
205   </para>
206  </section>
207
208
209  <section xml:id="containers.associative.bitset" xreflabel="bitset"><info><title>bitset</title></info>
210    <?dbhtml filename="bitset.html"?>
211
212    <section xml:id="associative.bitset.size_variable" xreflabel="Variable"><info><title>Size Variable</title></info>
213
214      <para>
215	No, you cannot write code of the form
216      </para>
217      <!-- Careful, the leading spaces in PRE show up directly. -->
218   <programlisting>
219      #include &lt;bitset&gt;
220
221      void foo (size_t n)
222      {
223	  std::bitset&lt;n&gt;   bits;
224	  ....
225      }
226   </programlisting>
227   <para>
228     because <code>n</code> must be known at compile time.  Your
229     compiler is correct; it is not a bug.  That's the way templates
230     work.  (Yes, it <emphasis>is</emphasis> a feature.)
231   </para>
232   <para>
233     There are a couple of ways to handle this kind of thing.  Please
234     consider all of them before passing judgement.  They include, in
235     no chaptericular order:
236   </para>
237      <itemizedlist>
238	<listitem><para>A very large N in <code>bitset&lt;N&gt;</code>.</para></listitem>
239	<listitem><para>A container&lt;bool&gt;.</para></listitem>
240	<listitem><para>Extremely weird solutions.</para></listitem>
241      </itemizedlist>
242   <para>
243     <emphasis>A very large N in
244     <code>bitset&lt;N&gt;</code>.  </emphasis> It has been
245     pointed out a few times in newsgroups that N bits only takes up
246     (N/8) bytes on most systems, and division by a factor of eight is
247     pretty impressive when speaking of memory.  Half a megabyte given
248     over to a bitset (recall that there is zero space overhead for
249     housekeeping info; it is known at compile time exactly how large
250     the set is) will hold over four million bits.  If you're using
251     those bits as status flags (e.g.,
252     <quote>changed</quote>/<quote>unchanged</quote> flags), that's a
253     <emphasis>lot</emphasis> of state.
254   </para>
255   <para>
256     You can then keep track of the <quote>maximum bit used</quote>
257     during some testing runs on representative data, make note of how
258     many of those bits really need to be there, and then reduce N to
259     a smaller number.  Leave some extra space, of course.  (If you
260     plan to write code like the incorrect example above, where the
261     bitset is a local variable, then you may have to talk your
262     compiler into allowing that much stack space; there may be zero
263     space overhead, but it's all allocated inside the object.)
264   </para>
265   <para>
266     <emphasis>A container&lt;bool&gt;.  </emphasis> The
267     Committee made provision for the space savings possible with that
268     (N/8) usage previously mentioned, so that you don't have to do
269     wasteful things like <code>Container&lt;char&gt;</code> or
270     <code>Container&lt;short int&gt;</code>.  Specifically,
271     <code>vector&lt;bool&gt;</code> is required to be specialized for
272     that space savings.
273   </para>
274   <para>
275     The problem is that <code>vector&lt;bool&gt;</code> doesn't
276     behave like a normal vector anymore.  There have been
277     journal articles which discuss the problems (the ones by Herb
278     Sutter in the May and July/August 1999 issues of C++ Report cover
279     it well).  Future revisions of the ISO C++ Standard will change
280     the requirement for <code>vector&lt;bool&gt;</code>
281     specialization.  In the meantime, <code>deque&lt;bool&gt;</code>
282     is recommended (although its behavior is sane, you probably will
283     not get the space savings, but the allocation scheme is different
284     than that of vector).
285   </para>
286   <para>
287     <emphasis>Extremely weird solutions.  </emphasis> If
288     you have access to the compiler and linker at runtime, you can do
289     something insane, like figuring out just how many bits you need,
290     then writing a temporary source code file.  That file contains an
291     instantiation of <code>bitset</code> for the required number of
292     bits, inside some wrapper functions with unchanging signatures.
293     Have your program then call the compiler on that file using
294     Position Independent Code, then open the newly-created object
295     file and load those wrapper functions.  You'll have an
296     instantiation of <code>bitset&lt;N&gt;</code> for the exact
297     <code>N</code> that you need at the time.  Don't forget to delete
298     the temporary files.  (Yes, this <emphasis>can</emphasis> be, and
299     <emphasis>has been</emphasis>, done.)
300   </para>
301   <!-- I wonder if this next paragraph will get me in trouble... -->
302   <para>
303     This would be the approach of either a visionary genius or a
304     raving lunatic, depending on your programming and management
305     style.  Probably the latter.
306   </para>
307   <para>
308     Which of the above techniques you use, if any, are up to you and
309     your intended application.  Some time/space profiling is
310     indicated if it really matters (don't just guess).  And, if you
311     manage to do anything along the lines of the third category, the
312     author would love to hear from you...
313   </para>
314   <para>
315     Also note that the implementation of bitset used in libstdc++ has
316     <link linkend="manual.ext.containers.sgi">some extensions</link>.
317   </para>
318
319    </section>
320    <section xml:id="associative.bitset.type_string" xreflabel="Type String"><info><title>Type String</title></info>
321
322      <para>
323      </para>
324   <para>
325     Bitmasks do not take char* nor const char* arguments in their
326     constructors.  This is something of an accident, but you can read
327     about the problem: follow the library's <quote>Links</quote> from
328     the homepage, and from the C++ information <quote>defect
329     reflector</quote> link, select the library issues list.  Issue
330     number 116 describes the problem.
331   </para>
332   <para>
333     For now you can simply make a temporary string object using the
334     constructor expression:
335   </para>
336   <programlisting>
337      std::bitset&lt;5&gt; b ( std::string(<quote>10110</quote>) );
338   </programlisting>
339
340   <para>
341     instead of
342   </para>
343
344    <programlisting>
345      std::bitset&lt;5&gt; b ( <quote>10110</quote> );    // invalid
346    </programlisting>
347    </section>
348  </section>
349
350</section>
351
352<!-- Sect1 03 : Unordered Associative -->
353<section xml:id="std.containers.unordered" xreflabel="Unordered">
354  <info><title>Unordered Associative</title></info>
355  <?dbhtml filename="unordered_associative.html"?>
356
357  <section xml:id="containers.unordered.hash" xreflabel="Hash">
358    <info><title>Hash Code</title></info>
359
360  <section xml:id="containers.unordered.cache" xreflabel="Cache">
361    <info><title>Hash Code Caching Policy</title></info>
362
363    <para>
364      The unordered containers in libstdc++ may cache the hash code for each
365      element alongside the element itself. In some cases not recalculating
366      the hash code every time it's needed can improve performance, but the
367      additional memory overhead can also reduce performance, so whether an
368      unordered associative container caches the hash code or not depends on
369      a number of factors. The caching policy for GCC 4.8 is described below.
370    </para>
371    <para>
372      The C++ standard requires that <code>erase</code> and <code>swap</code>
373      operations must not throw exceptions. Those operations might need an
374      element's hash code, but cannot use the hash function if it could
375      throw.
376      This means the hash codes will be cached unless the hash function
377      has a non-throwing exception specification such as <code>noexcept</code>
378      or <code>throw()</code>.
379    </para>
380    <para>
381      Secondly, libstdc++ also needs the hash code in the implementation of
382      <code>local_iterator</code> and <code>const_local_iterator</code> in
383      order to know when the iterator has reached the end of the bucket.
384      This means that the local iterator types will embed a copy of the hash
385      function when possible.
386      Because the local iterator types must be DefaultConstructible and
387      CopyAssignable, if the hash function type does not model those concepts
388      then it cannot be embedded and so the hash code must be cached.
389      Note that a hash function might not be safe to use when
390      default-constructed (e.g if it a function pointer) so a hash
391      function that is contained in a local iterator won't be used until
392      the iterator is valid, so the hash function has been copied from a
393      correctly-initialized object.
394    </para>
395    <para>
396      If the hash function is non-throwing, DefaultConstructible and
397      CopyAssignable then libstdc++ doesn't need to cache the hash code for
398      correctness, but might still do so for performance if computing a
399      hash code is an expensive operation, as it may be for arbitrarily
400      long strings.
401      As an extension libstdc++ provides a trait type to describe whether
402      a hash function is fast. By default hash functions are assumed to be
403      fast unless the trait is specialized for the hash function and the
404      trait's value is false, in which case the hash code will always be
405      cached.
406      The trait can be specialized for user-defined hash functions like so:
407    </para>
408    <programlisting>
409      #include &lt;unordered_set&gt;
410
411      struct hasher
412      {
413        std::size_t operator()(int val) const noexcept
414        {
415          // Some very slow computation of a hash code from an int !
416          ...
417        }
418      }
419
420      namespace std
421      {
422        template&lt;&gt;
423          struct __is_fast_hash&lt;hasher&gt; : std::false_type
424          { };
425      }
426    </programlisting>
427  </section>
428</section>
429
430</section>
431
432<!-- Sect1 04 : Interacting with C -->
433<section xml:id="std.containers.c" xreflabel="Interacting with C"><info><title>Interacting with C</title></info>
434<?dbhtml filename="containers_and_c.html"?>
435
436
437  <section xml:id="containers.c.vs_array" xreflabel="Containers vs. Arrays"><info><title>Containers vs. Arrays</title></info>
438
439   <para>
440     You're writing some code and can't decide whether to use builtin
441     arrays or some kind of container.  There are compelling reasons
442     to use one of the container classes, but you're afraid that
443     you'll eventually run into difficulties, change everything back
444     to arrays, and then have to change all the code that uses those
445     data types to keep up with the change.
446   </para>
447   <para>
448     If your code makes use of the standard algorithms, this isn't as
449     scary as it sounds.  The algorithms don't know, nor care, about
450     the kind of <quote>container</quote> on which they work, since
451     the algorithms are only given endpoints to work with.  For the
452     container classes, these are iterators (usually
453     <code>begin()</code> and <code>end()</code>, but not always).
454     For builtin arrays, these are the address of the first element
455     and the <link linkend="iterators.predefined.end">past-the-end</link> element.
456   </para>
457   <para>
458     Some very simple wrapper functions can hide all of that from the
459     rest of the code.  For example, a pair of functions called
460     <code>beginof</code> can be written, one that takes an array,
461     another that takes a vector.  The first returns a pointer to the
462     first element, and the second returns the vector's
463     <code>begin()</code> iterator.
464   </para>
465   <para>
466     The functions should be made template functions, and should also
467     be declared inline.  As pointed out in the comments in the code
468     below, this can lead to <code>beginof</code> being optimized out
469     of existence, so you pay absolutely nothing in terms of increased
470     code size or execution time.
471   </para>
472   <para>
473     The result is that if all your algorithm calls look like
474   </para>
475   <programlisting>
476   std::transform(beginof(foo), endof(foo), beginof(foo), SomeFunction);
477   </programlisting>
478   <para>
479     then the type of foo can change from an array of ints to a vector
480     of ints to a deque of ints and back again, without ever changing
481     any client code.
482   </para>
483
484<programlisting>
485// beginof
486template&lt;typename T&gt;
487  inline typename vector&lt;T&gt;::iterator
488  beginof(vector&lt;T&gt; &amp;v)
489  { return v.begin(); }
490
491template&lt;typename T, unsigned int sz&gt;
492  inline T*
493  beginof(T (&amp;array)[sz]) { return array; }
494
495// endof
496template&lt;typename T&gt;
497  inline typename vector&lt;T&gt;::iterator
498  endof(vector&lt;T&gt; &amp;v)
499  { return v.end(); }
500
501template&lt;typename T, unsigned int sz&gt;
502  inline T*
503  endof(T (&amp;array)[sz]) { return array + sz; }
504
505// lengthof
506template&lt;typename T&gt;
507  inline typename vector&lt;T&gt;::size_type
508  lengthof(vector&lt;T&gt; &amp;v)
509  { return v.size(); }
510
511template&lt;typename T, unsigned int sz&gt;
512  inline unsigned int
513  lengthof(T (&amp;)[sz]) { return sz; }
514</programlisting>
515
516   <para>
517     Astute readers will notice two things at once: first, that the
518     container class is still a <code>vector&lt;T&gt;</code> instead
519     of a more general <code>Container&lt;T&gt;</code>.  This would
520     mean that three functions for <code>deque</code> would have to be
521     added, another three for <code>list</code>, and so on.  This is
522     due to problems with getting template resolution correct; I find
523     it easier just to give the extra three lines and avoid confusion.
524   </para>
525   <para>
526     Second, the line
527   </para>
528   <programlisting>
529    inline unsigned int lengthof (T (&amp;)[sz]) { return sz; }
530   </programlisting>
531   <para>
532     looks just weird!  Hint:  unused parameters can be left nameless.
533   </para>
534  </section>
535
536</section>
537
538</chapter>
539