libs/function/doc/faq.xml

<?xml version="1.0" encoding="utf-8"?>
<!--
   Copyright (c) 2002 Douglas Gregor <doug.gregor -at- gmail.com>
  
   Distributed under the Boost Software License, Version 1.0.
   (See accompanying file LICENSE_1_0.txt or copy at
   http://www.boost.org/LICENSE_1_0.txt)
  -->
<!DOCTYPE library PUBLIC "-//Boost//DTD BoostBook XML V1.0//EN"
  "http://www.boost.org/tools/boostbook/dtd/boostbook.dtd">
<section id="function.faq" last-revision="$Date: 2006-11-03 11:41:10 -0800 (Fri, 03 Nov 2006) $">
  <title>Frequently Asked Questions</title>

<qandaset>
  <qandaentry>
    <question><para>Why can't I compare
    <classname>boost::function</classname> objects with
    <code>operator==</code> or
    <code>operator!=</code>?</para></question>

    <answer>
      <para>Comparison between <classname>boost::function</classname>
      objects cannot be implemented "well", and therefore will not be
      implemented. The typical semantics requested for <code>f ==
      g</code> given <classname>boost::function</classname> objects
      <code>f</code> and <code>g</code> are:</para>
        <itemizedlist>
          <listitem><simpara>If <code>f</code> and <code>g</code>
          store function objects of the same type, use that type's
          <code>operator==</code> to compare
          them.</simpara></listitem> 

          <listitem><simpara>If <code>f</code> and <code>g</code>
          store function objects of different types, return
          <code>false</code>.</simpara></listitem>
        </itemizedlist>
      <para>The problem occurs when the type of the function objects
      stored by both <code>f</code> and <code>g</code> doesn't have an
      <code>operator==</code>: we would like the expression <code>f ==
      g</code> to fail to compile, as occurs with, e.g., the standard
      containers. However, this is not implementable for
      <classname>boost::function</classname> because it necessarily
      "erases" some type information after it has been assigned a
      function object, so it cannot try to call
      <code>operator==</code> later: it must either find a way to call
      <code>operator==</code> now, or it will never be able to call it
      later. Note, for instance, what happens if you try to put a
      <code>float</code> value into a
      <classname>boost::function</classname> object: you will get an
      error at the assignment operator or constructor, not in
      <code>operator()</code>, because the function-call expression
      must be bound in the constructor or assignment operator.</para>

      <para>The most promising approach is to find a method of
      determining if <code>operator==</code> can be called for a
      particular type, and then supporting it only when it is
      available; in other situations, an exception would be
      thrown. However, to date there is no known way to detect if an
      arbitrary operator expression <code>f == g</code> is suitably
      defined. The best solution known has the following undesirable
      qualities:</para>

      <orderedlist>
        <listitem><simpara>Fails at compile-time for objects where
        <code>operator==</code> is not accessible (e.g., because it is
        <code>private</code>).</simpara></listitem>

        <listitem><simpara>Fails at compile-time if calling
        <code>operator==</code> is ambiguous.</simpara></listitem>

        <listitem><simpara>Appears to be correct if the
        <code>operator==</code> declaration is correct, even though
        <code>operator==</code> may not compile.</simpara></listitem>
      </orderedlist>

      <para>All of these problems translate into failures in the
      <classname>boost::function</classname> constructors or
      assignment operator, <emphasis>even if the user never invokes
      operator==</emphasis>. We can't do that to users.</para>

      <para>The other option is to place the burden on users that want
      to use <code>operator==</code>, e.g., by providing an
      <code>is_equality_comparable</code> trait they may
      specialize. This is a workable solution, but is dangerous in
      practice, because forgetting to specialize the trait will result
      in unexpected exceptions being thrown from
      <classname>boost::function</classname>'s
      <code>operator==</code>. This essentially negates the usefulness
      of <code>operator==</code> in the context in which it is most
      desired: multitarget callbacks. The
      <libraryname>Signals</libraryname> library has a way around
      this.</para>
    </answer>
  </qandaentry>

  <qandaentry>
    <question><para>I see void pointers; is this [mess] type safe?</para></question>
    <answer>
<para>Yes, <computeroutput>boost::function</computeroutput> is type
safe even though it uses void pointers and pointers to functions
returning void and taking no arguments. Essentially, all type
information is encoded in the functions that manage and invoke
function pointers and function objects. Only these functions are
instantiated with the exact type that is pointed to by the void
pointer or pointer to void function. The reason that both are required
is that one may cast between void pointers and object pointers safely
or between different types of function pointers (provided you don't
invoke a function pointer with the wrong type).  </para>
    </answer>
  </qandaentry>

  <qandaentry>
    <question><para>Why are there workarounds for void returns? C++ allows them!</para></question>
    <answer><para>Void returns are permitted by the C++ standard, as in this code snippet:
<programlisting>void f();
void g() { return f(); }</programlisting>
    </para>

    <para> This is a valid usage of <computeroutput>boost::function</computeroutput> because void returns are not used. With void returns, we would attempting to compile ill-formed code similar to:
<programlisting>int f();
void g() { return f(); }</programlisting>
</para>

<para> In essence, not using void returns allows
<computeroutput>boost::function</computeroutput> to swallow a return value. This is
consistent with allowing the user to assign and invoke functions and
function objects with parameters that don't exactly match.</para>

    </answer>
  </qandaentry>

  <qandaentry>
    <question><para>Why (function) cloning?</para></question>
    <answer>
      <para>In November and December of 2000, the issue of cloning
      vs. reference counting was debated at length and it was decided
      that cloning gave more predictable semantics. I won't rehash the
      discussion here, but if it cloning is incorrect for a particular
      application a reference-counting allocator could be used.</para>
    </answer>
  </qandaentry>

  <qandaentry>
    <question><para>How much overhead does a call through <code><classname>boost::function</classname></code> incur?</para></question>
    <answer>
      <para>The cost of <code>boost::function</code> can be reasonably
      consistently measured at around 20ns +/- 10 ns on a modern >2GHz
      platform versus directly inlining the code.</para>

      <para>However, the performance of your application may benefit
      from or be disadvantaged by <code>boost::function</code>
      depending on how your C++ optimiser optimises.  Similar to a
      standard function pointer, differences of order of 10% have been
      noted to the benefit or disadvantage of using
      <code>boost::function</code> to call a function that contains a
      tight loop depending on your compilation circumstances.</para>
     
      <para>[Answer provided by Matt Hurd. See <ulink url="http://article.gmane.org/gmane.comp.lib.boost.devel/33278"/>]</para>
    </answer>
  </qandaentry>
</qandaset> 

</section>