boost_mirror/iterator
1
0
Fork 0
mirror of https://github.com/boostorg/iterator.git synced 2025-05-11 13:33:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Edits for clarity and correctness.

Browse Source 
[SVN r20831]
This commit is contained in:
Dave Abrahams 2003-11-17 17:15:26 +00:00
parent 104faa5ab3
commit f25ff7a1a4
2 changed files with 241 additions and 334 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

451
doc/issues.html
View File

@ -4,7 +4,7 @@
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8" />
<meta name="generator" content="Docutils 0.3.1: http://docutils.sourceforge.net/" />
<title>Issues With N1550 and N1530</title>
<title>Problem with is_writable and is_swappable in N1550</title>
<style type="text/css"><!--
/*
@ -199,21 +199,21 @@ ul.auto-toc {
--></style>
</head>
<body>
<div class="document" id="issues-with-n1550-and-n1530">
<h1 class="title">Issues With <a class="reference" href="http://www.boost-consulting.com/writing/n1550.html">N1550</a> and <a class="reference" href="http://anubis.dkuug.dk/jtc1/sc22/wg21/docs/papers/2003/n1530.html">N1530</a></h1>
<div class="document" id="problem-with-is-writable-and-is-swappable-in-n1550">
<h1 class="title">Problem with <tt class="literal"><span class="pre">is_writable</span></tt> and <tt class="literal"><span class="pre">is_swappable</span></tt> in <a class="reference" href="http://www.boost-consulting.com/writing/n1550.html">N1550</a></h1>
<table class="field-list" frame="void" rules="none">
<col class="field-name" />
<col class="field-body" />
<tbody valign="top">
<tr class="field"><th class="field-name">Author:</th><td class="field-body">David Abrahams</td>
<tr class="field"><th class="field-name">Author:</th><td class="field-body">David Abrahams and Jeremy Siek</td>
</tr>
<tr class="field"><th class="field-name">Contact:</th><td class="field-body"><a class="reference" href="mailto:dave&#64;boost-consulting.com">dave&#64;boost-consulting.com</a></td>
<tr class="field"><th class="field-name">Contact:</th><td class="field-body"><a class="reference" href="mailto:dave&#64;boost-consulting.com">dave&#64;boost-consulting.com</a>, <a class="reference" href="mailto:jsiek&#64;osl.iu.edu">jsiek&#64;osl.iu.edu</a></td>
</tr>
<tr class="field"><th class="field-name">Organization:</th><td class="field-body"><a class="reference" href="http://www.boost-consulting.com">Boost Consulting</a></td>
<tr class="field"><th class="field-name">Organization:</th><td class="field-body"><a class="reference" href="http://www.boost-consulting.com">Boost Consulting</a>, Indiana University Bloomington</td>
</tr>
<tr class="field"><th class="field-name">date:</th><td class="field-body">$Date$</td>
</tr>
<tr class="field"><th class="field-name">Copyright:</th><td class="field-body">Copyright David Abrahams 2003. Use, modification and
<tr class="field"><th class="field-name">Copyright:</th><td class="field-body">Copyright David Abrahams, Jeremy Siek 2003. Use, modification and
distribution is subject to the Boost Software License,
Version 1.0. (See accompanying file LICENSE_1_0.txt or copy
at <a class="reference" href="http://www.boost.org/LICENSE_1_0.txt">http://www.boost.org/LICENSE_1_0.txt</a>)</td>
@ -224,301 +224,194 @@ at <a class="reference" href="http://www.boost.org/LICENSE_1_0.txt">http://www.b
<p class="topic-title"><a name="table-of-contents">Table of Contents</a></p>
<ul class="simple">
<li><a class="reference" href="#introduction" id="id3" name="id3">Introduction</a></li>
<li><a class="reference" href="#the-issues" id="id4" name="id4">The Issues</a><ul>
<li><a class="reference" href="#non-uniformity-of-the-lvalue-iterator-bit" id="id5" name="id5">Non-Uniformity of the &quot;<tt class="literal"><span class="pre">lvalue_iterator</span></tt> Bit&quot;</a></li>
<li><a class="reference" href="#redundancy-of-some-explicit-access-category-flags" id="id6" name="id6">Redundancy of Some Explicit Access Category Flags</a></li>
<li><a class="reference" href="#new-access-traits-templates-wrong-for-some-iterators" id="id7" name="id7">New Access Traits Templates Wrong For Some Iterators</a><ul>
<li><a class="reference" href="#is-writable-iterator" id="id8" name="id8"><tt class="literal"><span class="pre">is_writable_iterator</span></tt></a></li>
<li><a class="reference" href="#is-swappable-iterator" id="id9" name="id9"><tt class="literal"><span class="pre">is_swappable_iterator</span></tt></a></li>
</ul>
</li>
<li><a class="reference" href="#no-use-cases-for-some-access-traits" id="id10" name="id10">No Use Cases for Some Access Traits</a><ul>
<li><a class="reference" href="#id1" id="id11" name="id11"><tt class="literal"><span class="pre">is_swappable_iterator</span></tt></a></li>
<li><a class="reference" href="#id2" id="id12" name="id12"><tt class="literal"><span class="pre">is_writable_iterator</span></tt></a></li>
</ul>
</li>
<li><a class="reference" href="#naming-issues" id="id13" name="id13">Naming Issues</a><ul>
<li><a class="reference" href="#traversal-concepts-and-tags" id="id14" name="id14">Traversal Concepts and Tags</a></li>
<li><a class="reference" href="#access-traits" id="id15" name="id15">Access Traits</a></li>
</ul>
</li>
</ul>
</li>
<li><a class="reference" href="#proposed-solution-in-progress" id="id16" name="id16">Proposed Solution (in progress)</a><ul>
<li><a class="reference" href="#overview" id="id17" name="id17">Overview</a></li>
<li><a class="reference" href="#future-enhancements" id="id18" name="id18">Future Enhancements</a></li>
<li><a class="reference" href="#impact-on-n1530-iterator-facade-and-adaptor" id="id19" name="id19">Impact on N1530 (Iterator Facade and Adaptor)</a></li>
</ul>
</li>
<li><a class="reference" href="#proposed-resolution" id="id4" name="id4">Proposed Resolution</a></li>
<li><a class="reference" href="#rationale" id="id5" name="id5">Rationale</a></li>
</ul>
</div>
<div class="section" id="introduction">
<h1><a class="toc-backref" href="#id3" name="introduction">Introduction</a></h1>
<p>Several issues with <a class="reference" href="http://www.boost-consulting.com/writing/n1550.html">N1550</a> (New Iterator Concepts) were raised in
the run-up before the fall 2003 C++ Committee meeting, in a thread
beginning with John Maddock's posting <tt class="literal"><span class="pre">c++std-lib-12187</span></tt>. In
looking at those issues, several other problems came up. This
document addresses those issues and discusses some potential
solutions and their impact on <a class="reference" href="http://anubis.dkuug.dk/jtc1/sc22/wg21/docs/papers/2003/n1530.html">N1530</a> (Iterator Facade and Adaptor).</p>
</div>
<div class="section" id="the-issues">
<h1><a class="toc-backref" href="#id4" name="the-issues">The Issues</a></h1>
<div class="section" id="non-uniformity-of-the-lvalue-iterator-bit">
<h2><a class="toc-backref" href="#id5" name="non-uniformity-of-the-lvalue-iterator-bit">Non-Uniformity of the &quot;<tt class="literal"><span class="pre">lvalue_iterator</span></tt> Bit&quot;</a></h2>
<p>The proposed <tt class="literal"><span class="pre">iterator_tag</span></tt> class template accepts an &quot;access
bits&quot; parameter which includes a bit to indicate the iterator's
<em>lvalueness</em> (whether its dereference operator returns a reference
to its <tt class="literal"><span class="pre">value_type</span></tt>. The relevant part of <a class="reference" href="http://www.boost-consulting.com/writing/n1550.html">N1550</a> says:</p>
<blockquote>
The purpose of the <tt class="literal"><span class="pre">lvalue_iterator</span></tt> part of the
<tt class="literal"><span class="pre">iterator_access</span></tt> enum is to communicate to <tt class="literal"><span class="pre">iterator_tag</span></tt>
whether the reference type is an lvalue so that the appropriate
old category can be chosen for the base class. The
<tt class="literal"><span class="pre">lvalue_iterator</span></tt> bit is not recorded in the
<tt class="literal"><span class="pre">iterator_tag::access</span></tt> data member.</blockquote>
<p>The <tt class="literal"><span class="pre">lvalue_iterator</span></tt> bit is not recorded because <a class="reference" href="http://www.boost-consulting.com/writing/n1550.html">N1550</a> aims to
improve orthogonality of the iterator concepts, and a new-style
iterator's lvalueness is detectable by examining its <tt class="literal"><span class="pre">reference</span></tt>
type. This inside/outside difference is awkward and confusing.</p>
</div>
<div class="section" id="redundancy-of-some-explicit-access-category-flags">
<h2><a class="toc-backref" href="#id6" name="redundancy-of-some-explicit-access-category-flags">Redundancy of Some Explicit Access Category Flags</a></h2>
<p>Shortly after <a class="reference" href="http://www.boost-consulting.com/writing/n1550.html">N1550</a> was accepted, we discovered that an iterator's
lvalueness can be determined knowing only its <tt class="literal"><span class="pre">value_type</span></tt>. This
predicate can be calculated even for old-style iterators (on whose
<tt class="literal"><span class="pre">reference</span></tt> type the standard places few requirements). A trait
in the Boost iterator library does it by relying on the compiler's
unwillingness to bind an rvalue to a <tt class="literal"><span class="pre">T&amp;</span></tt> function template
parameter. Similarly, it is possible to detect an iterator's
readability knowing only its <tt class="literal"><span class="pre">value_type</span></tt>. Thus, any interface
which asks the <em>user</em> to explicitly describe an iterator's
lvalue-ness or readability seems to introduce needless complexity.</p>
</div>
<div class="section" id="new-access-traits-templates-wrong-for-some-iterators">
<h2><a class="toc-backref" href="#id7" name="new-access-traits-templates-wrong-for-some-iterators">New Access Traits Templates Wrong For Some Iterators</a></h2>
<div class="section" id="is-writable-iterator">
<h3><a class="toc-backref" href="#id8" name="is-writable-iterator"><tt class="literal"><span class="pre">is_writable_iterator</span></tt></a></h3>
<p>The part of the <tt class="literal"><span class="pre">is_writable_iterator</span></tt> trait definition which
applies to old-style iterators is:</p>
<p>The <tt class="literal"><span class="pre">is_writable</span></tt> and <tt class="literal"><span class="pre">is_swappable</span></tt> traits classes in <a class="reference" href="http://www.boost-consulting.com/writing/n1550.html">N1550</a>
provide a mechanism for determining at compile time if an iterator
type is a model of the new Writable Iterator and Swappable Iterator
concepts, analogous to <tt class="literal"><span class="pre">iterator_traits&lt;X&gt;::iterator_category</span></tt>
for the old iterator concepts. For backward compatibility,
<tt class="literal"><span class="pre">is_writable</span></tt> and <tt class="literal"><span class="pre">is_swappable</span></tt> not only work with new
iterators, but they also are intended to work for old
iterators (iterators that meet the requirements for one of the
iterator concepts in the current standard). In the case of old
iterators, the writability and swapability is deduced based on the
<tt class="literal"><span class="pre">iterator_category</span></tt> and also the <tt class="literal"><span class="pre">reference</span></tt> type. The
specification for this deduction gives false positives for forward
iterators that have non-assignable value types.</p>
<p>To review, the part of the <tt class="literal"><span class="pre">is_writable</span></tt> trait definition which
applies to old iterators is:</p>
<pre class="literal-block">
if (cat is convertible to output_iterator_tag)
return true;
else if (
cat is convertible to forward_iterator_tag
return true;
else if (cat is convertible to forward_iterator_tag
and iterator_traits&lt;Iterator&gt;::reference is a
mutable reference)
return true;
else
return false;
</pre>
<p>The current forward iterator requirements place no constraints on
the iterator's <tt class="literal"><span class="pre">reference</span></tt> type, so the logic above will give
false negatives for some otherwise-writable forward iterators whose
<tt class="literal"><span class="pre">reference</span></tt> type is not a mutable reference. Also, it will
report false positives for any forward, bidirectional, or random
access iterator whose <tt class="literal"><span class="pre">reference</span></tt> is a mutable reference but
whose <tt class="literal"><span class="pre">value_type</span></tt> is not assignable (e.g. has a private
assignment operator).</p>
</div>
<div class="section" id="is-swappable-iterator">
<h3><a class="toc-backref" href="#id9" name="is-swappable-iterator"><tt class="literal"><span class="pre">is_swappable_iterator</span></tt></a></h3>
<p>Similarly, the part of <tt class="literal"><span class="pre">is_swappable_iterator</span></tt> which applies to
old-style iterators is:</p>
<pre class="literal-block">
else if (cat is convertible to forward_iterator_tag) {
if (iterator_traits&lt;Iterator&gt;::reference is a const reference)
return false;
else
return true;
} else
return true;
else
return false;
</pre>
<p>In this case false positives are possible for non-writable forward
iterators whose <tt class="literal"><span class="pre">reference</span></tt> type is not a reference, or as above,
any forward, bidirectional, or random access iterator whose
<tt class="literal"><span class="pre">reference</span></tt> is not a constant reference but whose <tt class="literal"><span class="pre">value_type</span></tt>
is not assignable (e.g., because it has a private assignment
operator).</p>
<p>False negatives can be &quot;reasoned away&quot;: since it is part of a
writable iterator's concept definition that
<tt class="literal"><span class="pre">is_writable&lt;I&gt;::value</span></tt> is <tt class="literal"><span class="pre">true</span></tt>, any iterator for which
it is <tt class="literal"><span class="pre">false</span></tt> is by definition not writable. This seems like a
perverse use of logic, though.</p>
<p>It might be reasonable to conclude that it is a defect that the
standard allows forward iterators with a <tt class="literal"><span class="pre">reference</span></tt> type other
than <tt class="literal"><span class="pre">value_type</span></tt> <em>cv</em><tt class="literal"><span class="pre">&amp;</span></tt>, but that still leaves the problem
of old-style iterators whose <tt class="literal"><span class="pre">value_type</span></tt> is not assignable. It
is not possible to correctly compute writability and swappability
for those old-style iterators without intervention
(specializations of <tt class="literal"><span class="pre">is_writable_iterator</span></tt> and
<tt class="literal"><span class="pre">is_swappable_iterator</span></tt>) from a user.</p>
<p>Suppose the <tt class="literal"><span class="pre">value_type</span></tt> of the iterator <tt class="literal"><span class="pre">It</span></tt> has a private
assignment operator:</p>
<pre class="literal-block">
class B {
public:
...
private:
B&amp; operator=(const B&amp;);
};
</pre>
<p>and suppose the <tt class="literal"><span class="pre">reference</span></tt> type of the iterator is <tt class="literal"><span class="pre">B&amp;</span></tt>. In
that case, <tt class="literal"><span class="pre">is_writable&lt;It&gt;::value</span></tt> will be true when in fact
attempting to write into <tt class="literal"><span class="pre">B</span></tt> will cause an error.</p>
<p>The same problem applies to <tt class="literal"><span class="pre">is_swappable</span></tt>.</p>
</div>
</div>
<div class="section" id="no-use-cases-for-some-access-traits">
<h2><a class="toc-backref" href="#id10" name="no-use-cases-for-some-access-traits">No Use Cases for Some Access Traits</a></h2>
<div class="section" id="id1">
<h3><a class="toc-backref" href="#id11" name="id1"><tt class="literal"><span class="pre">is_swappable_iterator</span></tt></a></h3>
<p><tt class="literal"><span class="pre">is_swappable_iterator&lt;I&gt;</span></tt> is supposed to yield true if
<tt class="literal"><span class="pre">iter_swap(x,y)</span></tt> is valid for instances <tt class="literal"><span class="pre">x</span></tt> and <tt class="literal"><span class="pre">y</span></tt> of type
<tt class="literal"><span class="pre">I</span></tt>. The only argument we have heard for
<tt class="literal"><span class="pre">is_swappable_iterator</span></tt> goes something like this:</p>
<blockquote>
<em>&quot;If</em> <tt class="literal"><span class="pre">is_swappable_iterator</span></tt> <em>yields</em> <tt class="literal"><span class="pre">false</span></tt><em>, you
could fall back to using copy construction and assignment on
the</em> <tt class="literal"><span class="pre">value_type</span></tt> <em>instead.&quot;</em></blockquote>
<p>This line of reasoning, however, falls down when closely examined.
To achieve the same effect using copy construction and assignment
on the iterator's <tt class="literal"><span class="pre">value_type</span></tt>, the iterator must be readable and
writable, and its <tt class="literal"><span class="pre">value_type</span></tt> must be copy-constructible. But
then, <tt class="literal"><span class="pre">iter_swap</span></tt> must work in that case, because its default
implementation just calls <tt class="literal"><span class="pre">swap</span></tt> on the dereferenced iterators.
The only purpose for the swappable iterator concept is to represent
iterators which do not fulfill the properties listed above, but
which are nonetheless swappable because the user has provided an
overload or specialization of <tt class="literal"><span class="pre">iter_swap</span></tt>. In other words,
generic code which wants to swap the referents of two iterators
should <em>always</em> call <tt class="literal"><span class="pre">iter_swap</span></tt> instead of doing the
assignments.</p>
</div>
<div class="section" id="id2">
<h3><a class="toc-backref" href="#id12" name="id2"><tt class="literal"><span class="pre">is_writable_iterator</span></tt></a></h3>
<p>Try to imagine a case where <tt class="literal"><span class="pre">is_writable_iterator</span></tt> can be used to
choose behavior. Since the only requirement on a writable iterator
is that we can assign into its referent, the only use for
<tt class="literal"><span class="pre">is_writable_iterator</span></tt> in selecting behavior is to modify a
sequence when the sequence is mutable, and to not modify it
<div class="section" id="proposed-resolution">
<h1><a class="toc-backref" href="#id4" name="proposed-resolution">Proposed Resolution</a></h1>
<ol class="arabic">
<li><p class="first">Remove the <tt class="literal"><span class="pre">is_writable</span></tt> and <tt class="literal"><span class="pre">is_swappable</span></tt> traits, and remove the
requirements in the Writable Iterator and Swappable Iterator concepts
that require their models to support these traits.</p>
</li>
<li><p class="first">Change the <tt class="literal"><span class="pre">is_readable</span></tt> specification to be:
<tt class="literal"><span class="pre">is_readable&lt;X&gt;::type</span></tt> is <tt class="literal"><span class="pre">true_type</span></tt> if the
result type of <tt class="literal"><span class="pre">X::operator*</span></tt> is convertible to
<tt class="literal"><span class="pre">iterator_traits&lt;X&gt;::value_type</span></tt> and is <tt class="literal"><span class="pre">false_type</span></tt>
otherwise.</p>
<p>There is no precedent for generic functions which modify their
arguments only if the arguments are non-const reference, and with
good reason: the simple fact that data is mutable does not mean
that a user <em>intends</em> it to be mutated. We provide <tt class="literal"><span class="pre">const</span></tt> and
non-<tt class="literal"><span class="pre">const</span></tt> overloads for functions like <tt class="literal"><span class="pre">operator[]</span></tt>, but
these do not modify data; they merely return a reference to data
which preserves the object's mutability properties. We can do the
same with iterators using their <tt class="literal"><span class="pre">reference</span></tt> types; the
accessibility of an assignment operator on the <tt class="literal"><span class="pre">value_type</span></tt>,
which determines writability, does not change that.</p>
<p>The one plausible argument we can imagine for
<tt class="literal"><span class="pre">is_writable_iterator</span></tt> and <tt class="literal"><span class="pre">is_swappable_iterator</span></tt> is that they
can be used to remove algorithms from an overload set using a
SFINAE technique like <a class="reference" href="http://tinyurl.com/tsr7">enable_if</a>, thus minimizing unintentional
matches due to Koenig Lookup. If it means requiring explicit
indications of writability and swappability from new-style iterator
implementors, however, it seems to be too small a gain to be worth
the cost. That's especially true since we can't get many existing
old-style iterators to meet the same requirements.</p>
</div>
</div>
<div class="section" id="naming-issues">
<h2><a class="toc-backref" href="#id13" name="naming-issues">Naming Issues</a></h2>
<div class="section" id="traversal-concepts-and-tags">
<h3><a class="toc-backref" href="#id14" name="traversal-concepts-and-tags">Traversal Concepts and Tags</a></h3>
<p>Howard Hinnant pointed out some inconsistencies with the naming of
these tag types:</p>
<p>Remove the requirement for support of the <tt class="literal"><span class="pre">is_readable</span></tt> trait from
the Readable Iterator concept.</p>
</li>
</ol>
<!-- We should give some attention to the UnaryTypeTrait concept,
which requires the trait to be derived from either true_type or
false_type (as of the last LWG meeting). -->
<ol class="arabic" start="3">
<li><p class="first">Change <tt class="literal"><span class="pre">iterator_tag</span></tt> to:</p>
<pre class="literal-block">
incrementable_iterator_tag // ++r, r++
single_pass_iterator_tag // adds a == b, a != b
forward_traversal_iterator_tag // adds multi-pass capability
bidirectional_traversal_iterator_tag // adds --r, r--
random_access_traversal_iterator_tag // adds r+n,n+r,r-n,r[n],etc.
template &lt;class Value, class Reference, class Traversal&gt;
struct iterator_tag;
</pre>
<p>Howard thought that it might be better if all tag names contained
the word &quot;traversal&quot;.</p>
<p>It's not clear that would result in the best possible names,
though. For example, incrementable iterators can only make a
single pass over their input. What really distinguishes single
pass iterators from incrementable iterators is not that they can
make a single pass, but that they are equality comparable. Forward
traversal iterators really distinguish themselves by introducing
multi-pass capability. Without entering a &quot;Parkinson's Bicycle
Shed&quot; type of discussion, it might be worth giving the names of
these tags (and the associated concepts) some extra attention.</p>
</div>
<div class="section" id="access-traits">
<h3><a class="toc-backref" href="#id15" name="access-traits">Access Traits</a></h3>
<p>The names <tt class="literal"><span class="pre">is_readable</span></tt>, <tt class="literal"><span class="pre">is_writable</span></tt>, and <tt class="literal"><span class="pre">is_swappable</span></tt>
are probably too general for their semantics. In particular, a
swappable iterator is only swappable in the same sense that a
mutable iterator is mutable: the trait refers to the iterator's
referent. It would probably be better to add the <tt class="literal"><span class="pre">_iterator</span></tt>
suffix to each of these names.</p>
</div>
</div>
</div>
<div class="section" id="proposed-solution-in-progress">
<h1><a class="toc-backref" href="#id16" name="proposed-solution-in-progress">Proposed Solution (in progress)</a></h1>
<p>We believe that <tt class="literal"><span class="pre">is_readable_iterator</span></tt> is a fine name for the
proposed <tt class="literal"><span class="pre">is_readable</span></tt> trait and will use that from here on. In
order to avoid confusion, however, and because we aren't terribly
convinced of any answer yet, we are going to phrase this solution
in terms of the existing traversal concept and tag names. We'll
propose a few possible traversal naming schemes at the end of this
section.</p>
<div class="section" id="overview">
<h2><a class="toc-backref" href="#id17" name="overview">Overview</a></h2>
<p>Following the dictum that what we can't do well probably shouldn't
be done at all, we'd like to solve many of the problems above by
eliminating details and simplifying the library as proposed. In
particular, we'd eliminate <tt class="literal"><span class="pre">is_writable</span></tt> and <tt class="literal"><span class="pre">is_swappable</span></tt>,
and remove the requirements which say that writable, and swappable
iterators must support these traits. <tt class="literal"><span class="pre">is_readable_iterator</span></tt> has
proven to be useful and will be retained, but since it can be
implemented with no special hints from the iterator, it will not be
mentioned in the readable iterator requirements. Since we don't
want to require the user to explicitly specify access category
information, we'll change <tt class="literal"><span class="pre">iterator_tag</span></tt> so that it computes the
old-style category in terms of the iterator's traversal category,
<tt class="literal"><span class="pre">reference</span></tt>, and <tt class="literal"><span class="pre">value_type</span></tt>.</p>
</div>
<div class="section" id="future-enhancements">
<h2><a class="toc-backref" href="#id18" name="future-enhancements">Future Enhancements</a></h2>
<p>For C++0x, we could consider a change to <tt class="literal"><span class="pre">iterator_traits</span></tt> which
allows the user to avoid the use of iterator_tag (or similar
devices) altogether and write a new-style iterator by specifying
only a traversal tag. This change is not being proposed as it does
not constitute a &quot;pure bolt-on&quot;:</p>
<p>The argument for <tt class="literal"><span class="pre">Value</span></tt> must be the <tt class="literal"><span class="pre">value_type</span></tt> of the
iterator, possibly const-qualified, <tt class="literal"><span class="pre">Reference</span></tt> must be the
return type of <tt class="literal"><span class="pre">operator*</span></tt> <a class="footnote-reference" href="#id2" id="id1" name="id1"><sup>*</sup></a>, and <tt class="literal"><span class="pre">Traversal</span></tt> the
traversal tag for the iterator.</p>
</li>
</ol>
<!-- I think the language above is still too informal. There is no
"the iterator", when considering iterator_tag in isolation.
Perhaps that language belongs in a non-normative note
``iterator_tag<Value,Reference,Traversal>`` is required to be
convertible to both ``Traversal`` tag and also to the
``iterator_category`` type specified by the following
pseudo-code::
old-category(Value, Reference, Traversal) =
if (Reference is a reference
and Traversal is convertible to forward_traversal_tag)
{
if (Traversal is convertible to random_access_traversal_tag)
return random_access_iterator_tag;
else if (Traversal is convertible to bidirectional_traversal_tag)
return bidirectional_iterator_tag;
else
return forward_iterator_tag;
}
else if (Traversal is convertible to single_pass_traversal_tag
and Reference is convertible to Value)
{
if (Value is const)
return input_iterator_tag;
else
return input_output_iterator_tag;
} else
return output_iterator_tag; -->
<!-- I reformatted the code for legibility; sorry. -->
<table class="footnote" frame="void" id="id2" rules="none">
<colgroup><col class="label" /><col /></colgroup>
<tbody valign="top">
<tr><td class="label"><a class="fn-backref" href="#id1" name="id2">[*]</a></td><td>Instead of saying &quot;return type of operator*&quot;, we could have
said <tt class="literal"><span class="pre">iterator_traits&lt;X&gt;::reference</span></tt>. However, the standard
specifies nothing about <tt class="literal"><span class="pre">iterator_traits&lt;X&gt;::reference</span></tt> in
many cases, which we believe is a defect. Furthermore, in some
cases it explicitly differs from the return type of
<tt class="literal"><span class="pre">operator*</span></tt>, for example see <tt class="literal"><span class="pre">istreambuf_iterator</span></tt>.</td></tr>
</tbody>
</table>
<ol class="arabic" start="4">
<li><p class="first">Change the specification of <tt class="literal"><span class="pre">traversal_category</span></tt> to:</p>
<pre class="literal-block">
iterator_traits&lt;I&gt;::iterator_category
= if (I::iterator_category is a type) // use mpl::has_xxx (SFINAE)
return I::iterator_category
// Only old-style output iterators may have a void value_type
// or difference_type
if (iterator_value_type&lt;I&gt;::type is void
|| iterator_difference_type&lt;I&gt;::type is void
)
return std::output_iterator_tag
t = iterator_traversal&lt;I&gt;::type
if (I is an lvalue iterator)
{
if (t is convertible to random_access_traversal_tag)
return std::random_access_iterator_tag
if (t is convertible to bidirectional_traversal_tag)
return std::bidirectional_iterator_tag
else if (t is convertible to forward_traversal_tag)
return std::forward_iterator_tag
}
if (t is convertible to single_pass_traversal_tag
&amp;&amp; I is a readable iterator
)
return input_output_iterator_tag // (**)
traversal-category(Iterator) =
let cat = iterator_traits&lt;Iterator&gt;::iterator_category
if (cat is convertible to incrementable_iterator_tag)
return cat; // Iterator is a new iterator
else if (cat is convertible to random_access_iterator_tag)
return random_access_traversal_tag;
else if (cat is convertible to bidirectional_iterator_tag)
return bidirectional_traversal_tag;
else if (cat is convertible to forward_iterator_tag)
return forward_traversal_tag;
else if (cat is convertible to input_iterator_tag)
return single_pass_iterator_tag;
else if (cat is convertible to output_iterator_tag)
return incrementable_iterator_tag;
else
return std::output_iterator_tag
return null_category_tag;
</pre>
</li>
</ol>
</div>
<div class="section" id="impact-on-n1530-iterator-facade-and-adaptor">
<h2><a class="toc-backref" href="#id19" name="impact-on-n1530-iterator-facade-and-adaptor">Impact on <a class="reference" href="http://anubis.dkuug.dk/jtc1/sc22/wg21/docs/papers/2003/n1530.html">N1530</a> (Iterator Facade and Adaptor)</a></h2>
<p>XXX</p>
</div>
<div class="section" id="rationale">
<h1><a class="toc-backref" href="#id5" name="rationale">Rationale</a></h1>
<ol class="arabic simple">
<li>There are two reasons for removing <tt class="literal"><span class="pre">is_writable</span></tt>
and <tt class="literal"><span class="pre">is_swappable</span></tt>. The first is that we do not know of
a way to fix the specification so that it gives the correct
answer for all iterators. Second, there was only a weak
motivation for having <tt class="literal"><span class="pre">is_writable</span></tt> and <tt class="literal"><span class="pre">is_swappable</span></tt>
there in the first place. The main motivation was simply
uniformity: we have tags for the old iterator categories
so we should have tags for the new iterator categories.
While having tags and the capability to dispatch based
on the traversal categories is often used, we see
less of a need for dispatching based on writability
and swappability, since typically algorithms
that need these capabilities have no alternative if
they are not provided.</li>
<li>We discovered that the <tt class="literal"><span class="pre">is_readable</span></tt> trait can be implemented
using only the iterator type itself and its <tt class="literal"><span class="pre">value_type</span></tt>.
Therefore we remove the requirement for <tt class="literal"><span class="pre">is_readable</span></tt> from the
Readable Iterator concept, and change the definition of
<tt class="literal"><span class="pre">is_readable</span></tt> so that it works for any iterator type.</li>
<li>With <tt class="literal"><span class="pre">is_writable</span></tt> and <tt class="literal"><span class="pre">is_swappable</span></tt> gone, and
<tt class="literal"><span class="pre">is_readable</span></tt> no longer in need of special hints,
there is no reason for the <tt class="literal"><span class="pre">iterator_tag</span></tt> class to provide
information about the access capabilities of an iterator.
This new version provides only information about the traversal
capabilities and the old iterator category tag. Instead of accessing
the traversal category as a nested typedef <tt class="literal"><span class="pre">::traversal</span></tt>,
the <tt class="literal"><span class="pre">iterator_tag</span></tt> itself will be convertible to the traversal
tag. The <tt class="literal"><span class="pre">access_bits</span></tt> parameter is no longer needed for
specifying the access member (which is now gone). However,
some access information is still needed so that we can
deduce the appropriate old iterator category. The
<tt class="literal"><span class="pre">Value</span></tt> and <tt class="literal"><span class="pre">Reference</span></tt> parameters fill this need.
Note that this solution cleans up the issue that John
Maddock raised on the reflector (<tt class="literal"><span class="pre">c++std-lib-12187</span></tt>) about the non-uniformity
of the lvalue bit.</li>
<li>The changes to the specification of <tt class="literal"><span class="pre">traversal_category</span></tt> are a
direct result of the changes to <tt class="literal"><span class="pre">iterator_tag</span></tt>.</li>
</ol>
</div>
</div>
<hr class="footer" />
<div class="footer">
<a class="reference" href="issues.rst">View document source</a>.
Generated on: 2003-11-07 19:02 UTC.
Generated on: 2003-11-17 17:15 UTC.
Generated by <a class="reference" href="http://docutils.sourceforge.net/">Docutils</a> from <a class="reference" href="http://docutils.sourceforge.net/rst.html">reStructuredText</a> source.
</div>
</body>

124
doc/issues.rst
View File

@ -23,17 +23,18 @@
==============
The ``is_writable`` and ``is_swappable`` traits classes in N1550_
provide a mechanism for determining at compile time if a type is a
model of the new Writable Iterator and Swappable Iterator concepts,
analogous to ``iterator_traits<X>::iterator_category`` for the old
iterator concepts. For backward compatibility, ``is_writable`` and
``is_swappable`` not only work with new iterators, but they also are
intended to work for old iterators (iterators that meet the
requirements for one of the iterator concepts in the current
standard). In the case of old iterators, the writability and
swapability is deduced based on the ``iterator_category`` and also the
``reference`` type. The specification for this deduction gives false
positives for forward iterators that have non-assignable value types.
provide a mechanism for determining at compile time if an iterator
type is a model of the new Writable Iterator and Swappable Iterator
concepts, analogous to ``iterator_traits<X>::iterator_category``
for the old iterator concepts. For backward compatibility,
``is_writable`` and ``is_swappable`` not only work with new
iterators, but they also are intended to work for old
iterators (iterators that meet the requirements for one of the
iterator concepts in the current standard). In the case of old
iterators, the writability and swapability is deduced based on the
``iterator_category`` and also the ``reference`` type. The
specification for this deduction gives false positives for forward
iterators that have non-assignable value types.
To review, the part of the ``is_writable`` trait definition which
applies to old iterators is::
@ -47,8 +48,8 @@ applies to old iterators is::
else
return false;
Suppose the ``value_type`` of the iterator has a private assignment
operator::
Suppose the ``value_type`` of the iterator ``It`` has a private
assignment operator::
class B {
public:
@ -57,9 +58,9 @@ operator::
B& operator=(const B&);
};
and suppose the ``reference`` type of the iterator is ``B&``. Then
``is_writable`` will deduce true when in fact attempting to write into
``B`` will cause an error.
and suppose the ``reference`` type of the iterator is ``B&``. In
that case, ``is_writable<It>::value`` will be true when in fact
attempting to write into ``B`` will cause an error.
The same problem applies to ``is_swappable``.
@ -81,64 +82,78 @@ The same problem applies to ``is_swappable``.
Remove the requirement for support of the ``is_readable`` trait from
the Readable Iterator concept.
.. We should give some attention to the UnaryTypeTrait concept,
which requires the trait to be derived from either true_type or
false_type (as of the last LWG meeting).
3. Change ``iterator_tag`` to::
template <class Value, class Reference, class Traversal>
struct iterator_tag;
The argument for ``Value`` must be the value type of the
iterator, ``Reference`` must be the return type of
``operator*`` [*]_, and ``Traversal`` the traversal tag for the
iterator.
The argument for ``Value`` must be the ``value_type`` of the
iterator, possibly const-qualified, ``Reference`` must be the
return type of ``operator*`` [*]_, and ``Traversal`` the
traversal tag for the iterator.
``iterator_tag`` is required to be convertible to both the
``Traversal`` tag and also to the appropriate old iterator category
tag, as specified by the following pseudo-code::
.. I think the language above is still too informal. There is no
"the iterator", when considering iterator_tag in isolation.
Perhaps that language belongs in a non-normative note
inherit-category(Value, Reference, Traversal) =
``iterator_tag<Value,Reference,Traversal>`` is required to be
convertible to both ``Traversal`` tag and also to the
``iterator_category`` type specified by the following
pseudo-code::
old-category(Value, Reference, Traversal) =
if (Reference is a reference
and Traversal is convertible to forward_traversal_tag) {
and Traversal is convertible to forward_traversal_tag)
{
if (Traversal is convertible to random_access_traversal_tag)
return random_access_iterator_tag;
else if (Traversal is convertible to bidirectional_traversal_tag)
return bidirectional_iterator_tag;
else
return forward_iterator_tag;
} else if (Traversal is convertible to single_pass_traversal_tag
and Reference is convertible to Value) {
}
else if (Traversal is convertible to single_pass_traversal_tag
and Reference is convertible to Value)
{
if (Value is const)
return input_iterator_tag;
else
return input_output_iterator_tag;
} else
return output_iterator_tag;
.. I reformatted the code for legibility; sorry.
.. [*] Instead of saying "return type of operator*", we could have
said ``iterator_traits<X>::reference``. However, the standard
specifies nothing about ``iterator_traits<X>::reference``,
which we believe is a defect. Once the defect is fixed,
the above could be rephrased.
specifies nothing about ``iterator_traits<X>::reference`` in
many cases, which we believe is a defect. Furthermore, in some
cases it explicitly differs from the return type of
``operator*``, for example see ``istreambuf_iterator``.
4. Change the specification of ``traversal_category`` to::
traversal-category(Iterator) =
let cat = iterator_traits<Iterator>::iterator_category
if (cat convertible to incrementable_iterator_tag)
return cat; // Iterator is a new iterator
else if (cat is convertible to random_access_iterator_tag)
return random_access_traversal_tag;
else if (cat is convertible to bidirectional_iterator_tag)
return bidirectional_traversal_tag;
else if (cat is convertible to forward_iterator_tag)
return forward_traversal_tag;
else if (cat is convertible to input_iterator_tag)
return single_pass_iterator_tag;
else if (cat is convertible to output_iterator_tag)
return incrementable_iterator_tag;
else
return null_category_tag;
let cat = iterator_traits<Iterator>::iterator_category
if (cat is convertible to incrementable_iterator_tag)
return cat; // Iterator is a new iterator
else if (cat is convertible to random_access_iterator_tag)
return random_access_traversal_tag;
else if (cat is convertible to bidirectional_iterator_tag)
return bidirectional_traversal_tag;
else if (cat is convertible to forward_iterator_tag)
return forward_traversal_tag;
else if (cat is convertible to input_iterator_tag)
return single_pass_iterator_tag;
else if (cat is convertible to output_iterator_tag)
return incrementable_iterator_tag;
else
return null_category_tag;
==========
@ -150,7 +165,7 @@ The same problem applies to ``is_swappable``.
a way to fix the specification so that it gives the correct
answer for all iterators. Second, there was only a weak
motivation for having ``is_writable`` and ``is_swappable``
there in the first place. The main motivation was simply
there in the first place. The main motivation was simply
uniformity: we have tags for the old iterator categories
so we should have tags for the new iterator categories.
While having tags and the capability to dispatch based
@ -160,12 +175,11 @@ The same problem applies to ``is_swappable``.
that need these capabilities have no alternative if
they are not provided.
2. We discovered that the ``is_readable`` trait can be
implemented without special hints from the iterator.
Therefore we remove the requirement for ``is_readable``
from the Readable Iterator concept, and change
the definition of ``is_readable`` so that it works
for any iterator type.
2. We discovered that the ``is_readable`` trait can be implemented
using only the iterator type itself and its ``value_type``.
Therefore we remove the requirement for ``is_readable`` from the
Readable Iterator concept, and change the definition of
``is_readable`` so that it works for any iterator type.
3. With ``is_writable`` and ``is_swappable`` gone, and
``is_readable`` no longer in need of special hints,
@ -181,7 +195,7 @@ The same problem applies to ``is_swappable``.
deduce the appropriate old iterator category. The
``Value`` and ``Reference`` parameters fill this need.
Note that this solution cleans up the issue that John
Maddock raised on the reflector about the non-uniformity
Maddock raised on the reflector (``c++std-lib-12187``) about the non-uniformity
of the lvalue bit.
4. The changes to the specification of ``traversal_category`` are a