<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE appendix PUBLIC "-//Boost//DTD BoostBook XML V1.0//EN"
  "http://www.boost.org/tools/boostbook/dtd/boostbook.dtd">

  <chapter id="bbv2.extender">
    <title>Extender Manual</title>

    <section id="bbv2.extender.intro">
      <title>Introduction</title>

      <para>
        This section explains how to extend Boost.Build to accomodate your
        local requirements&mdash;primarily to add support for non-standard
        tools you have. Before we start, be sure you have read and understoon
        the concept of metatarget, <xref linkend="bbv2.overview.concepts"/>,
        which is critical to understanding the remaining material.
      </para>

      <para>
        The current version of Boost.Build has three levels of targets, listed
        below.
      </para>

      <variablelist>
        
        <varlistentry>
          <term>metatarget</term>
          <listitem>
            <para>
              Object that is created from declarations in Jamfiles. May
              be called with a set of properties to produce concrete
              targets.
            </para>
          </listitem>
        </varlistentry>
        
        <varlistentry>
          <term>concrete target</term>
          <listitem>
            <para>
              Object that corresponds to a file or an action.
            </para>
          </listitem>
        </varlistentry>
        
        <varlistentry>
          <term>jam target</term>
          <listitem>
            <para>
              Low-level concrete target that is specific to Boost.Jam build
              engine. Essentially a string&mdash;most often a name of file.                
            </para>
          </listitem>
        </varlistentry>
        
      </variablelist>

      <para>
        In most cases, you will only have to deal with concrete targets and
        the process that creates concrete targets from
        metatargets. Extending metatarget level is rarely required. The jam
        targets are typically only used inside the command line patterns.
      </para>

      <warning>
        <para>All of the Boost.Jam target-related builtin functions, like
        <code>DEPENDS</code> or <code>ALWAYS</code> operate on jam
        targets. Applying them to metatargets or concrete targets has no
        effect.</para>
      </warning>
            
      <section id="bbv2.extender.overview.metatargets">
        <title>Metatargets</title>
        
        <para>Metatarget is an object that records information specified
        in Jamfile, such as metatarget kind, name, sources and properties,
        and can be called with specific properties to generate concrete
        targets. At the code level it is represented by an instance of
        class derived from <classname>abstract-target</classname>. 
        <footnote><para>This name is historic, and will be eventuall changed to 
        <code>metatarget</code></para></footnote>
        </para>
        
        <para>The <methodname>generate</methodname> method takes the build properties
        (as an instance of the <classname>property-set</classname> class) and returns
        a list containing:</para>
        <itemizedlist>
          <listitem><para>As front element&mdash;Usage-requirements from this invocation
          (an instance of <classname>property-set</classname>)</para></listitem>
          <listitem><para>As subsequent elements&mdash;created concrete targets (
          instances of the <classname>virtual-target</classname> class.)</para></listitem>
        </itemizedlist>
        
        <para>It's possible to lookup a metataget by target-id using the 
        <code>targets.resolve-reference</code> function, and the
        <code>targets.generate-from-reference</code> function can both
        lookup and generate a metatarget.</para>
        
        <para>The <classname>abstract-target</classname> class has three immediate
        derived classes:</para>
        <itemizedlist>
          
          <listitem><para><classname>project-target</classname> that
          corresponds to a project and is not intended for further
          subclassing. The <methodname>generate</methodname> method of this
          class builds all targets in the project that are not marked as
          explicit.</para></listitem>
          
          <listitem><para><classname>main-target</classname> corresponds to a target in a project
          and contains one or more target alternatives. This class also should not be 
          subclassed. The <methodname>generate</methodname> method of this class selects 
          an alternative to build, and calls the <methodname>generate</methodname> method of that
          alternative.</para></listitem>
          
          <listitem><para><classname>basic-target</classname> corresponds to a
          specific target alternative. This is base class, with a number of
          derived classes.  The <methodname>generate</methodname> method
          processes the target requirements and requested build properties to
          determine final properties for the target, builds all sources, and
          finally calls the abstract <classname>construct</classname> method with the list
          of source virtual targets, and the final properties.
          </para></listitem>
          
        </itemizedlist>
        
        <para>The instances of the <classname>project-target</classname> and
        <classname>main-target</classname> classes are created
        implicitly&mdash;when loading a new Jamfiles, or when a new target
        alternative with as-yet unknown name is created.  The instances of the
        classes derived from <classname>basic-target</classname> are typically
        created when Jamfile calls a <firstterm>metatarget rule</firstterm>,
        such as such as <code>exe</code>.
        </para>
        
        <para>It it permissible to create a custom class derived from
        <classname>basic-target</classname> and create new metatarget rule
        that creates instance of such target. However, in the majority
        of cases, a specific subclass of <classname>basic-target</classname>&mdash;
        <classname>typed-target</classname> is used. That class is associated
        with a <firstterm>type</firstterm> and relays to <firstterm>generators</firstterm>
        to construct concrete targets of that type. This process will be explained below.
        When a new type is declared, a new metatarget rule is automatically defined.
        That rule creates new instance of type-target, associated with that type.                
        </para>
        
      </section>

      <section id="bbv2.extender.overview.targets">
        <title>Concrete targets</title>

        <para>Concrete targets are represented by instance of classes derived
        from <classname>virtual-target</classname>. The most commonly used
        subclass is <classname>file-target</classname>. A file target is associated
        with an action that creates it&mdash; an instance of the <classname>action</classname>
        class. The action, in turn, hold a list of source targets. It also holds the 
        <classname>property-set</classname> instance with the build properties that
        should be used for the action.</para>

        <para>Here's an example of creating a target from another target, <code>source</code></para>
<programlisting>
local a = [ new action $(source) : common.copy : $(property-set) ] ;
local t = [ new file-target $(name) : CPP : $(project) : $(a) ] ;
</programlisting>
        <para>The first line creates an instance of the <classname>action></classname> class.
        The first parameter is the list of sources. The second parameter is the name
        a jam-level <link linkend="bbv2.overview.jam_language.actions">action</link>.
        The third parameter is the property-set applying to this action. The second line
        creates a target. We specifie a name, a type and a project. We also pass the
        action object created earlier.  If the action creates several targets, we can repeat
        the second line several times.</para>

        <para>In some cases, code that creates concrete targets may be invoked more than
        once with the same properties. Returning to different instance of <classname>file-target</classname>
        that correspond to the same file clearly will result in problems. Therefore, whenever
        returning targets you should pass them via the <code>virtual-target.register</code>
        function, that will replace targets with previously created identical ones, as
        necessary.<footnote><para>This create-then-register pattern is caused by limitations 
        of the Boost.Jam language. Python port is likely to never create duplicate targets.</para></footnote>
        Here are a couple of examples:
<programlisting>
return [ virtual-target.register $(t) ] ;
return [ sequence.transform virtual-target.register : $(targets) ] ;
</programlisting>
        </para>

      </section>
      
      <section id="bbv2.extender.overview.generators">
        <title>Generators</title>

        <para>In theory, every kind of metatarget in Boost.Build (like <code>exe</code>, 
        <code>lib</code> or <code>obj</code>) could be implemented
        by writing a new metatarget class that, independently of the other code, figures
        what files to produce and what commands to use. However, that would be rather inflexible.
        For example, adding support for a new compiler would require editing several metatargets.
        </para>

        <para>In practice, most files have specific types, and most tools
        consume and produce files of specific type. To take advantage of this
        fact, Boost.Build defines concept of target type and
        <indexterm><primary>generators</primary></indexterm>
        <firstterm>generators</firstterm>, and has special metatarget class
        <classname>typed-target</classname>.  Target type is merely an
        identifier. It is associated with a set of file extensions that
        correspond to that type. Generator is an abstraction of a tool. It advertises
        the types it produces and, if called with a set of input target, tries to construct
        output targets of the advertised types. Finally, <classname>typed-target</classname>
        is associated with specific target type, and relays the generator (or generators)
        for that type.
        </para>

        <para>A generator is an instance of a class derived from <classname>generator</classname>.
        The <classname>generator</classname> class itself is suitable for common cases.
        You can define derived classes for custom scenarios.</para>

        <!--
        <para>Given a set of generators, the fundamental operation is to
        construct a target of a given type, with given properties, from a
        set of targets. That operation is performed by rule
        <literal>generators.construct</literal> and the used algorithm is described
        below.</para>

        <section>
          <title>Selecting and ranking viable generators</title>

          <para>Each generator, in addition to target types that it can
          produce, have attribute that affects its applicability in
          particular sitiation. Those attributes are:</para>

          <orderedlist>
            <listitem>
              <simpara>
                Required properties, which are properties absolutely
                necessary for the generator to work. For example, generator
                encapsulating the gcc compiler would have &lt;toolset&gt;gcc as
                required property.
              </simpara>
            </listitem>

            <listitem>
              <simpara>
                Optional properties, which increase the generators
                suitability for a particual build.
              </simpara>
            </listitem>
          </orderedlist>

          <para>
            Generator's required and optional properties may not include
            either free or incidental properties. (Allowing this would
            greatly complicate caching targets).
          </para>

          <para>When trying to construct a target, the first step is to select
          all possible generators for the requested target type, which
          required properties are a subset of requested properties.
          Generators that were already selected up the call stack are
          excluded. In addition, if any composing generators were selected
          up the call stack, all other composing generators are ignored
          (TODO: define composing generators). The found generators
          are assigned a rank, which is the number of optional properties
          present in requested properties. Finally, generators with highest
          rank are selected for futher processing.</para>

        </section>
        <section>
          <title>Running generators</title>

          <para>When generators are selected, each is run to produce a list of
          created targets. This list might include targets that are not of
          requested types, because generators create the same targets as
          some tool, and tool's behaviour is fixed. (Note: should specify
          that in some cases we actually want extra targets). If generator
          fails, it returns an empty list. Generator is free to call
          'construct' again, to convert sources to the types it can handle.
          It also can pass modified properties to 'construct'. However, a
          generator is not allowed to modify any propagated properties,
          otherwise when actually consuming properties we might discover
          that the set of propagated properties is different from what was
          used for building sources.</para>

          <para>For all targets that are not of requested types, we try to
          convert them to requested type, using a second call to
          <literal>construct</literal>. This is done in order to support
          transformation sequences where single source file expands to
          several later. See <ulink url=
          "http://groups.yahoo.com/group/jamboost/message/1667">this
          message</ulink> for details.</para>

        </section>

        -->

        <!-- FIXME: review the below content. Maybe, some of it is
             still useful.
        <section>
          <title>Property adjustment</title>

          <para>Because target location is determined by the build system, it
          is sometimes necessary to adjust properties, in order to not
          break actions. For example, if there's an action that generates
          a header, say "a_parser.h", and a source file "a.cpp" which
          includes that file, we must make everything work as if a_parser.h
          is generated in the same directory where it would be generated
          without any subvariants.</para>

          <para>Correct property adjustment can be done only after all targets
          are created, so the approach taken is:</para>

          <orderedlist>
            <listitem>
              <para>
                When dependency graph is constructed, each action can be
                assigned a rule for property adjustment.
              </para>
            </listitem>

            <listitem>
              <para>
                When virtual target is actualized, that rule is run and
                return the final set of properties. At this stage it can use
                information of all created virtual targets.
              </para>
            </listitem>
          </orderedlist>

          <para>In case of quoted includes, no adjustment can give 100% correct
          results. If target dirs are not changed by build system, quoted
          includes are searched in "." and then in include path, while angle
          includes are searched only in include path. When target dirs are
          changed, we'd want to make quoted includes to be search in "." then in
          additional dirs and then in the include path and make angle includes
          be searched in include path, probably with additional paths added at
          some position. Unless, include path already has "." as the first
          element, this is not possible. So, either generated headers should not
          be included with quotes, or first element of include path should be
          ".", which essentially erases the difference between quoted and angle
          includes. <emphasis role="bold">Note:</emphasis> the only way to get
          "." as include path into compiler command line is via verbatim
          compiler option. In all other case, Boost.Build will convert "." into
          directory where it occurs.</para> 

        </section>

        -->

      </section>
      
    </section>

    <section id="bbv2.extender.example">
      <title>Example: 1-to-1 generator</title>

      <para>Say you're writing an application that generates C++ code. If
      you ever did this, you know that it's not nice. Embedding large
      portions of C++ code in string literals is very awkward. A much
      better solution is:</para>

      <orderedlist>
        <listitem>
          <simpara>
            Write the template of the code to be generated, leaving
            placeholders at the points that will change
          </simpara>
        </listitem>
        
        <listitem>
          <simpara>
            Access the template in your application and replace
            placeholders with appropriate text.
          </simpara>
        </listitem>
        
        <listitem>
          <simpara>Write the result.</simpara>
        </listitem>
      </orderedlist>
      
      <para>It's quite easy to achieve. You write special verbatim files that are
      just C++, except that the very first line of the file contains the name of a
      variable that should be generated. A simple tool is created that takes a
      verbatim file and creates a cpp file with a single <code>char*</code> variable
      whose name is taken from the first line of the verbatim file and whose value
      is the file's properly quoted content.</para>

      <para>Let's see what Boost.Build can do.</para>
      
      <para>First off, Boost.Build has no idea about "verbatim files". So, you must
      register a new target type. The following code does it:</para>
      
<programlisting>
import type ;
type.register VERBATIM : verbatim ;
</programlisting>

      <para>The first parameter to <functionname>type.register</functionname> gives
      the name of the declared type. By convention, it's uppercase. The second
      parameter is the suffix for files of this type. So, if Boost.Build sees
      <filename>code.verbatim</filename> in a list of sources, it knows that it's of
      type <code>VERBATIM</code>.</para>

      <para>Next, you tell Boost.Build that the verbatim files can be
      transformed into C++ files in one build step.  A
      <firstterm>generator</firstterm> is a template for a build step that
      transforms targets of one type (or set of types) into another.  Our
      generator will be called <code>verbatim.inline-file</code>; it
      transforms <code>VERBATIM</code> files into <code>CPP</code> files:

<programlisting>
import generators ;
generators.register-standard verbatim.inline-file : VERBATIM : CPP ;
</programlisting>
  </para>

       <para>Lastly, you have to inform Boost.Build about the shell
       commands used to make that transformation.  That's done with an
       <code>actions</code> declaration.

<programlisting>
actions inline-file
{
    "./inline-file.py" $(&lt;) $(&gt;)
}
</programlisting>

<!-- You need to explain all the parameters to an "actions" and
     describe the accompanying rule declaration: the user has no clue
     what $(<) and $(>) are, and doesn't know about the third
     parameter that gets passed to the rule. -->

<!-- We use verbatim.inline-file in one place and just inline-file in
             another. Is this confusing for user?
        -->
</para>

  <para>
    Now, we're ready to tie it all together. Put all the code above in file
    <filename>verbatim.jam</filename>, add <code>import verbatim ;</code> to
    <filename>Jamroot.jam</filename>, and it's possible to write the following
    in your Jamfile:
  </para>

<programlisting>
exe codegen : codegen.cpp class_template.verbatim usage.verbatim ;
</programlisting>

  <para>
    The listed verbatim files will be automatically converted into C++ source
    files, compiled and then linked to the codegen executable.
  </para>

  <para>
    In subsequent sections, we will extend this example, and review all the
    mechanisms in detail. The complete code is available in the
    <filename>example/customization</filename> directory.
  </para>
  </section>

  <section id="bbv2.extending.targets">
    <title>Target types</title>
      <para>The first thing we did in the <link
          linkend="bbv2.extender.intro">intruduction</link> was declaring a
      new target type:
<programlisting>
import type ;
type.register VERBATIM : verbatim ;
</programlisting>
        The type is the most important property of a target. Boost.Build can
        automatically generate necessary build actions only because you
        specify the desired type (using the different main target rules), and
        because Boost.Build can guess the type of sources from their
        extensions.
      </para>

      <para>The first two parameters for the <code>type.register</code> rule
        are the name of new type and the list of extensions associated with
        it. A file with an extension from the list will have the given target
        type. In the case where a target of the declared type is generated
        from other sources, the first specified extension will be used.
      </para>

    <para>Sometimes you want to change the suffix used for generated targets
      depending on build properties, such as toolset. For example, some compiler
      uses extension <literal>elf</literal> for executable files. You can use the
      <code>type.set-generated-target-suffix</code> rule:
<programlisting>
type.set-generated-target-suffix EXE : &lt;toolset&gt;elf : elf ;
</programlisting>
    </para>

    <para>A new target type can be inherited from an existing one.
<programlisting>
type.register PLUGIN : : SHARED_LIB ;
</programlisting>
      The above code defines a new type derived from
      <code>SHARED_LIB</code>. Initially, the new type inherits all the
      properties of the base type - in particular generators and suffix.
      Typically, you'll change the new type in some way. For example, using
      <code>type.set-generated-target-suffix</code> you can set the suffix for
      the new type. Or you can write special a generator for the new type. For
      example, it can generate additional metainformation for the plugin.
      In either way, the <code>PLUGIN</code> type can be used whenever
      <code>SHARED_LIB</code> can. For example, you can directly link plugins
      to an application.
    </para>

    <para>A type can be defined as "main", in which case Boost.Build will
      automatically declare a main target rule for building targets of that
      type. More details can be found <link
      linkend="bbv2.extending.rules.main-type">later</link>.
    </para>

      <section id="bbv2.extending.scanners">
        <title>Scanners</title>
        <para>
          Sometimes, a file can refer to other files via some include system. To
          make Boost.Build track dependencies between included files, you need
          to provide a scanner. The primary limitation is that only one scanner
          can be assigned to a target type.
        </para>

        <para>First, we need to declare a new class for the scanner:
<programlisting>
class verbatim-scanner : common-scanner
{
    rule pattern ( )
    {
        return "//###include[ ]*\"([^\"]*)\"" ;
    }
}
</programlisting>
          All the complex logic is in the <code>common-scanner</code>
          class, and you only need to override the method that returns
          the regular expression to be used for scanning. The
          parentheses in the regular expression indicate which part
          of the string is the name of the included file.  Only the
          first parenthesized group in the regular expression will be
          recognized; if you can't express everything you want that
          way, you can return multiple regular expressions, each of
          which contains a parenthesized group to be matched.
        </para>

        <para>After that, we need to register our scanner class:
<programlisting>
scanner.register verbatim-scanner : include ;
</programlisting>
            The value of the second parameter, in this case
            <code>include</code>, specifies the properties that contain the list
            of paths that should be searched for the included files.
         </para>

        <para>Finally, we assign the new scanner to the <code>VERBATIM</code>
        target type:
<programlisting>
type.set-scanner VERBATIM : verbatim-scanner ;
</programlisting>
          That's enough for scanning include dependencies.
        </para>

      </section>

  </section>

  <section id="bbv2.extending.tools">
    <title>Tools and generators</title>
      <para>
        This section will describe how Boost.Build can be extended to support
        new tools.
      </para>

      <para>For each additional tool, a Boost.Build object called generator
        must be created. That object has specific types of targets that it
        accepts and produces. Using that information, Boost.Build is able
        to automatically invoke the generator. For example, if you declare a
        generator that takes a target of the type <literal>D</literal> and
        produces a target of the type <literal>OBJ</literal>, when placing a
        file with extention <literal>.d</literal> in a list of sources will
        cause Boost.Build to invoke your generator, and then to link the
        resulting object file into an application. (Of course, this requires
        that you specify that the <literal>.d</literal> extension corresponds
        to the <literal>D</literal> type.)
      </para>

      <para>Each generator should be an instance of a class derived from the
        <code>generator</code> class. In the simplest case, you don't need to
        create a derived class, but simply create an instance of the
        <code>generator</code> class. Let's review the example we've seen in the
        <link linkend="bbv2.extender.intro">introduction</link>.
        <!-- Is the following supposed to be verbatim.jam?  Tell the
             user so.  You also need to describe the meanings of $(<)
             and $(>); this is the first time they're encountered. -->
<programlisting>
import generators ;
generators.register-standard verbatim.inline-file : VERBATIM : CPP ;
actions inline-file
{
    "./inline-file.py" $(&lt;) $(&gt;)
}
</programlisting>
      </para>

      <para>We declare a standard generator, specifying its id, the source type
        and the target type. When invoked, the generator will create a target
        of type <literal>CPP</literal> with a source target of
        type <literal>VERBATIM</literal> as the only source. But what command
        will be used to actually generate the file? In bjam, actions are
        specified using named "actions" blocks and the name of the action
        block should be specified when creating targets. By convention,
        generators use the same name of the action block as their own id. So,
        in above example, the "inline-file" actions block will be used to
        convert the source into the target.
      </para>

      <para>
        There are two primary kinds of generators: standard and composing,
        which are registered with the
        <code>generators.register-standard</code> and the
        <code>generators.register-composing</code> rules, respectively. For
        example:
<programlisting>
generators.register-standard verbatim.inline-file : VERBATIM : CPP ;
generators.register-composing mex.mex : CPP LIB : MEX ;
</programlisting>
        The first (standard) generator takes a <emphasis>single</emphasis>
        source of type <code>VERBATIM</code> and produces a result. The second
        (composing) generator takes any number of sources, which can have either
        the <code>CPP</code> or the <code>LIB</code> type. Composing generators
        are typically used for generating top-level target type. For example,
        the first generator invoked when building an <code>exe</code> target is
        a composing generator corresponding to the proper linker.
      </para>

      <para>You should also know about two specific functions for registering
        generators: <code>generators.register-c-compiler</code> and
        <code>generators.register-linker</code>. The first sets up header
        dependecy scanning for C files, and the seconds handles various
        complexities like searched libraries. For that reason, you should always
        use those functions when adding support for compilers and linkers.
      </para>

      <para>(Need a note about UNIX)</para>
      <!-- What kind of note?  Either write the note or don't, but remove this dross. -->
      <bridgehead>Custom generator classes</bridgehead>

      <para>The standard generators allows you to specify source and target
        types, an action, and a set of flags. If you need anything more complex,
        <!-- What sort of flags?  Command-line flags?  What does the system do with them? -->
        you need to create a new generator class with your own logic. Then,
        you have to create an instance of that class and register it. Here's
        an example how you can create your own generator class:
<programlisting>
class custom-generator : generator
{
    rule __init__ ( * : * )
    {
        generator.__init__ $(1) : $(2) : $(3) : $(4) : $(5) : $(6) : $(7) : $(8) : $(9) ;
    }
<!-- What is the point of this __init__ function?? -->
}

generators.register
  [ new custom-generator verbatim.inline-file : VERBATIM : CPP ] ;
</programlisting>
        This generator will work exactly like the
        <code>verbatim.inline-file</code> generator we've defined above, but
        it's possible to customize the behaviour by overriding methods of the
        <code>generator</code> class.
      </para>

      <para>There are two methods of interest. The <code>run</code> method is
        responsible for the overall process - it takes a number of source targets,
        converts them to the right types, and creates the result. The
        <code>generated-targets</code> method is called when all sources are
        converted to the right types to actually create the result.
      </para>

      <para>The <code>generated-targets</code> method can be overridden when you
        want to add additional properties to the generated targets or use
        additional sources. For a real-life example, suppose you have a program
        analysis tool that should be given a name of executable and the list of
        all sources. Naturally, you don't want to list all source files
        manually. Here's how the <code>generated-targets</code> method can find
        the list of sources automatically:
<programlisting>
class itrace-generator : generator {
....
    rule generated-targets ( sources + : property-set : project name ? )
    {
        local leaves ;
        local temp = [ virtual-target.traverse $(sources[1]) : : include-sources ] ;<!-- You must explain include-sources! -->
        for local t in $(temp)
        {
            if ! [ $(t).action<!-- In what namespace is this evaluated? --> ]
            {
                leaves += $(t) ;
            }
        }
        return [ generator.generated-targets $(sources) $(leafs)
          : $(property-set) : $(project) $(name) ] ;
    }
}
generators.register [ new itrace-generator nm.itrace : EXE : ITRACE ] ;
</programlisting>
        The <code>generated-targets</code> method will be called with a single
        source target of type <literal>EXE</literal>. The call to
        <code>virtual-target.traverse</code> will return all targets the
        executable depends on, and we further find files that are not
        produced from anything. <!-- What does "not produced from anything" mean? -->
        The found targets are added to the sources.
      </para>

      <para>The <code>run</code> method can be overriden to completely
        customize the way the generator works. In particular, the conversion of
        sources to the desired types can be completely customized. Here's
        another real example. Tests for the Boost Python library usually
        consist of two parts: a Python program and a C++ file. The C++ file is
        compiled to Python extension that is loaded by the Python
        program. But in the likely case that both files have the same name,
        the created Python extension must be renamed. Otherwise, the Python
        program will import itself, not the extension. Here's how it can be
        done:
<programlisting>
rule run ( project name ? : property-set : sources * )
{
    local python ;
    for local s in $(sources)
    {
        if [ $(s).type ] = PY
        {
            python = $(s) ;
        }
    }
    <!-- This is horrible code.  Use a filter function, or at _least_ consolidate the two loops! -->
    local libs ;
    for local s in $(sources)
    {
        if [ type.is-derived [ $(s).type ] LIB ]
        {
            libs += $(s) ;
        }
    }

    local new-sources ;
    for local s in $(sources)
    {
        if [ type.is-derived [ $(s).type ] CPP ]
        {
            local name = [ $(s).name ] ;    # get the target's basename
            if $(name) = [ $(python).name ]
            {
                name = $(name)_ext ;        # rename the target
            }
            new-sources += [ generators.construct $(project) $(name) :
              PYTHON_EXTENSION : $(property-set) : $(s) $(libs) ] ;
        }
    }

    result = [ construct-result $(python) $(new-sources) : $(project) $(name)
                 : $(property-set) ] ;
}
</programlisting>
        <!-- Why are we doing this with a generator??? It seems
             insane.  We could just use a nice front-end rule that
             calls some normal target-creation rules. No? -->

        First, we separate all source into python files, libraries and C++
        sources. For each C++ source we create a separate Python extension by
        calling <code>generators.construct</code> and passing the C++ source
        and the libraries. At this point, we also change the extension's name,
        if necessary.
      </para>


    </section>

    <section id="bbv2.extending.features">
      <title>Features</title>
      <para>
        Often, we need to control the options passed the invoked tools. This
        is done with features. Consider an example:
<programlisting>
# Declare a new free feature
import feature : feature ;
feature verbatim-options : : free ;

# Cause the value of the 'verbatim-options' feature to be
# available as 'OPTIONS' variable inside verbatim.inline-file
import toolset : flags ;
flags verbatim.inline-file OPTIONS &lt;verbatim-options&gt; ;<!-- You must tell the reader what the syntax of the flags rule is -->

# Use the "OPTIONS" variable
actions inline-file
{
    "./inline-file.py" $(OPTIONS) $(&lt;) $(&gt;)
}
</programlisting>
        We first define a new feature. Then, the <code>flags</code> invocation
        says that whenever verbatin.inline-file action is run, the value of
        the <code>verbatim-options</code> feature will be added to the
        <code>OPTIONS</code> variable, and can be used inside the action body.
        You'd need to consult online help (--help) to find all the features of
        the <code>toolset.flags</code> rule.
        <!-- It's been a while since I wrote these notes, so I don't
             remember what I meant.  But right here, I wrote "bad" and
             circled it.  Maybe you can figure out what I meant. ;-)
             -->
      </para>

    <para>
      Although you can define any set of features and interpret their values
      in any way, Boost.Build suggests the following coding standard for
      designing features.
    </para>

    <para>Most features should have a fixed set of values that is portable
      (tool neutral) across the class of tools they are designed to work
      with. The user does not have to adjust the values for a exact tool.  For
      example, <code>&lt;optimization&gt;speed</code> has the same meaning for
      all C++ compilers and the user does not have to worry about the exact
      options passed to the compiler's command line.
    </para>

    <para>
      Besides such portable features there are special 'raw' features that
      allow the user to pass any value to the command line parameters for a
      particular tool, if so desired. For example, the
      <code>&lt;cxxflags&gt;</code> feature allows you to pass any command line
      options to a C++ compiler. The <code>&lt;include&gt;</code> feature
      allows you to pass any string preceded by <code>-I</code> and the interpretation
      is tool-specific. <!-- It's really tool-specific?  That surprises me --> (See <xref
      linkend="bbv2.faq.external"/> for an example of very smart usage of that
      feature).  Of course one should always strive to use portable
      features, but these are still be provided as a backdoor just to make
      sure Boost.Build does not take away any control from the user.
    </para>

    <para>
      Using portable features is a good idea because:
      <itemizedlist>
        <listitem>
          <para>When a portable feature is given a fixed set of
          values, you can build your project with two different
          settings of the feature and Boost.Build will automatically
          use two different directories for generated files.
          Boost.Build does not try to separate targets built with
          different raw options.
            <!-- It's a computer program.  It doesn't "care" about options -->
          </para>
        </listitem>

        <listitem>
          <para>Unlike with “raw” features, you don't need to use
          specific command-line flags in your Jamfile, and it will be
          more likely to work with other tools.
          </para>
        </listitem>
      </itemizedlist>
    </para>

      <bridgehead>Steps for adding a feauture</bridgehead>
      <!-- This section is redundant with the previous one -->
      <para>Adding a feature requires three steps:

        <orderedlist>
          <listitem><para>Declaring a feature. For that, the "feature.feature"
              rule is used. You have to decide on the set of <link
              linkend="bbv2.reference.features.attributes">feature
              attributes</link>:

              <itemizedlist>
                <listitem><para>if you want a feature value set for one target
                to automaticaly propagate to its dependant targets then make it
                “propagated”. <!-- Examples needed. --></para></listitem>

                <listitem><para>if a feature does not have a fixed list of
                values, it must be “free.”  For example, the <code>include
                </code> feature is a free feature.</para></listitem>

                <listitem><para>if a feature is used to refer to a path relative
                to the Jamfile, it must be a “path” feature. Such features will
                also get their values automatically converted to Boost Build's
                internal path representation. For example, <code>include</code>
                is a path feature.</para></listitem>

                <listitem><para>if feature is used to refer to some target, it
                must be a “dependency” feature. <!-- for example? --></para>

                <!-- Any other feature attributes? -->
                </listitem>
              </itemizedlist>
              </para>
          </listitem>

          <listitem><para>Representing the feature value in a
          target-specific variable. Build actions are command
          templates modified by Boost.Jam variable expansions.  The
          <code>toolset.flags</code> rule sets a target-specific
          variable to the value of a feature.</para></listitem>

                <listitem><para>Using the variable. The variable set in step 2 can
              be used in a build action to form command parameters or
              files.</para></listitem>

        </orderedlist>
      </para>

      <bridgehead>Another example</bridgehead>

      <para>Here's another example.
        Let's see how we can make a feature that refers to a target. For example,
        when linking dynamic libraries on Windows, one sometimes needs to
        specify a "DEF file", telling what functions should be exported. It
        would be nice to use this file like this:
<programlisting>
        lib a : a.cpp : &lt;def-file&gt;a.def ;
</programlisting>
<!-- Why would that be nice?  It seems to me that having a.def in the sources is the obvious and much nicer thing to do:

        lib a : a.cpp a.def ;
-->
        Actually, this feature is already supported, but anyway...
        <!-- Something about saying that is very off-putting.  I'm
             sorry that I can't put my finger on it -->
      </para>

      <orderedlist>
        <listitem>
          <para>Since the feature refers to a target, it must be "dependency".
<programlisting>
feature def-file : : free dependency ;
</programlisting>
            </para></listitem>

        <listitem><para>One of the toolsets that cares about
        <!-- The toolset doesn't "care." What do your really mean? -->
        DEF files is msvc. The following line should be added to it.
        <!-- Are you saying the msvc toolset is broken (or that it
             doesn't use DEF files) as-shipped and the reader needs to
             fix it? -->

<programlisting>
flags msvc.link DEF_FILE &lt;def-file&gt; ;
</programlisting>
            <!-- And that line does... what? -->
            </para></listitem>

        <listitem><para>Since the DEF_FILE variable is not used by the
msvc.link action,
<!-- It's not?  You just told us that MSVC "cares" about DEF files. I
     presume that means that it uses them in some appropriate way? -->
we need to modify it to be:

<programlisting>
actions link bind DEF_FILE
{
    $(.LD) .... /DEF:$(DEF_FILE) ....
}
</programlisting>
            </para>


          <para> Note the <code>bind DEF_FILE</code> part. It tells
          bjam to translate the internal target name in
          <varname>DEF_FILE</varname> to a corresponding filename in
          the <code>link</code> action.  Without it the expansion of
          <code>$(DEF_FILE)</code> would be a strange symbol that is
          not likely to make sense for the linker.
          </para>

          <!-- I have a note here that says: "none of this works for
               targets in general, only source files."  I'm not sure
               what I meant by that; maybe you can figure it out. -->
          <para>
            We are almost done, but we should stop for a small workaround. Add the following
            code to msvc.jam

<programlisting>
rule link
{
    DEPENDS $(&lt;) : [ on $(&lt;) return $(DEF_FILE) ] ;
}
</programlisting>
<!-- You *must* explain the part in [...] above. It's completely opaque to the casual reader -->

            This is needed to accomodate some bug in bjam, which hopefully
            will be fixed one day.
            <!-- This is *NOT* a bug!!  Anyway, BBv2 shouild handle this automatically. Why doesn't it? -->
</para></listitem>

      </orderedlist>

      <bridgehead>Variants and composite features.</bridgehead>

      <para>Sometimes you want to create a shortcut for some set of
        features. For example, <code>release</code> is a value of
        <code>&lt;variant&gt;</code> and is a shortcut for a set of features.
      </para>

      <para>It is possible to define your own build variants. For example:
<programlisting>
variant crazy : &lt;optimization&gt;speed &lt;inlining&gt;off
                &lt;debug-symbols&gt;on &lt;profiling&gt;on ;
</programlisting>
        will define a new variant with the specified set of properties. You
        can also extend an existing variant:
<programlisting>
variant super_release : release : &lt;define&gt;USE_ASM ;
</programlisting>
        In this case, <code>super_release</code> will expand to all properties
        specified by <code>release</code>, and the additional one you've specified.
      </para>

      <para>You are not restricted to using the <code>variant</code> feature
      only.
      <!-- What do you mean by that?  How is defining a new feature related to what came before? -->
      Here's example that defines a brand new feature:
<programlisting>
feature parallelism : mpi fake none : composite link-incompatible ;
feature.compose &lt;parallelism&gt;mpi : &lt;library&gt;/mpi//mpi/&lt;parallelism&gt;none ;
feature.compose &lt;parallelism&gt;fake : &lt;library&gt;/mpi//fake/&lt;parallelism&gt;none ;
</programlisting>
<!-- The use of the <library>/mpi//mpi/<parallelism>none construct
     above is at best confusing and unexplained -->
        This will allow you to specify the value of feature
        <code>parallelism</code>, which will expand to link to the necessary
        library.
      </para>

  </section>

  <section id="bbv2.extending.rules">
    <title>Main target rules</title>
    <para>
      A main target rule (e.g “<functionname>exe</functionname>”
      Or “<functionname>lib</functionname>”) creates a top-level target. It's quite likely that you'll want to declare your own and
      there are two ways to do that.
      <!-- Why did "that" get changed to "this" above? -->
    </para>

    <para id="bbv2.extending.rules.main-type">The first way applies when
<!-- This is not a "way of defining a main target rule."  Rephrase this and the previous sentence. -->
      your target rule should just produce a target of specific type. In that case, a
      rule is already defined for you! When you define a new type, Boost.Build
      automatically defines a corresponding rule. The name of the rule is
      obtained from the name of the type, by downcasing all letters and
      replacing underscores with dashes.
      <!-- This strikes me as needless complexity, and confusing.  Why
           do we have the uppercase-underscore convention for target
           types?  If we just dropped that, the rule names could be
           the same as the type names. -->
      For example, if you create a module
      <filename>obfuscate.jam</filename> containing:

<programlisting>
import type ;
type.register OBFUSCATED_CPP  : ocpp ;

import generators ;
generators.register-standard obfuscate.file : CPP : OBFUSCATED_CPP ;
</programlisting>
      and import that module, you'll be able to use the rule "obfuscated-cpp"
      in Jamfiles, which will convert source to the OBFUSCATED_CPP type.
    </para>

    <para>
      The second way is to write a wrapper rule that calls any of the existing
      rules. For example, suppose you have only one library per directory and
      want all cpp files in the directory to be compiled into that library. You
      can achieve this effect using:
<programlisting>
lib codegen : [ glob *.cpp ] ;
</programlisting>
      If you want to make it even simpler, you could add the following
      definition to the <filename>Jamroot.jam</filename> file:
<programlisting>
rule glib ( name : extra-sources * : requirements * )
{
    lib $(name) : [ glob *.cpp ] $(extra-sources) : $(requirements) ;
}
</programlisting>
      allowing you to reduce the Jamfile to just
<programlisting>
glib codegen ;
</programlisting>
    </para>

    <para>
      Note that because you can associate a custom generator with a target type,
      the logic of building can be rather complicated. For example, the
      <code>boostbook</code> module declares a target type
      <code>BOOSTBOOK_MAIN</code> and a custom generator for that type. You can
      use that as example if your main target rule is non-trivial.
    </para>
  </section>

  <section id="bbv2.extending.toolset_modules">

    <title>Toolset modules</title>

    <para>
      If your extensions will be used only on one project, they can be placed in
      a separate <filename>.jam</filename> file and imported by your
      <filename>Jamroot.jam</filename>. If the extensions will be used on many
      projects, users will thank you for a finishing touch.
    </para>

    <para>The <code>using</code> rule provides a standard mechanism
    for loading and configuring extensions.  To make it work, your module
    <!-- "module" hasn't been defined yet.  Furthermore you haven't
         said anything about where that module file must be
         placed. -->
    should provide an <code>init</code> rule. The rule will be called
    with the same parameters that were passed to the
    <code>using</code> rule. The set of allowed parameters is
    determined by you. For example, you can allow the user to specify
    paths, tool versions, and other options.
    <!-- But it's not entirely arbitrary.  We have a standard
         parameter order which you should describe here for
         context. -->
    </para>

    <para>Here are some guidelines that help to make Boost.Build more
      consistent:
      <itemizedlist>
        <listitem><para>The <code>init</code> rule should never fail. Even if
          the user provided an incorrect path, you should emit a warning and go
          on. Configuration may be shared between different machines, and
          wrong values on one machine can be OK on another.
          <!-- So why shouldn't init fail on machines where it's wrong?? -->
          </para></listitem>

        <listitem><para>Prefer specifying the command to be executed
        to specifying the tool's installation path. First of all, this
        gives more control: it's possible to specify
<programlisting>
/usr/bin/g++-snapshot
time g++
<!-- Is this meant to be a single command?  If not, insert "or" -->
</programlisting>
            as the command. Second, while some tools have a logical
            "installation root", it's better if the user doesn't have to remember whether
            a specific tool requires a full command or a path.
            <!-- But many tools are really collections: e.g. a
                 compiler, a linker, and others.  The idea that the
                 "command to invoke" has any significance may be
                 completely bogus.  Plus if you want to allow "time
                 /usr/bin/g++" the toolset may need to somehow parse
                 the command and find the path when it needs to invoke
                 some related executable.  And in that case, will the
                 command be ignored?  This scheme doesn't scale and
                 should be fixed. -->
          </para></listitem>

        <listitem><para>Check for multiple initialization. A user can try to
            initialize the module several times. You need to check for this
            and decide what to do. Typically, unless you support several
            versions of a tool, duplicate initialization is a user error.
            <!-- Why should that be typical? -->
            If the
            tool's version can be specified during initialization, make sure the
            version is either always specified, or never specified (in which
            case the tool is initialied only once). For example, if you allow:
<programlisting>
using yfc ;
using yfc : 3.3 ;
using yfc : 3.4 ;
</programlisting>
            Then it's not clear if the first initialization corresponds to
            version 3.3 of the tool, version 3.4 of the tool, or some other
            version. This can lead to building twice with the same version.
            <!-- That would not be so terrible, and is much less harmful
                 than this restriction, IMO.  It makes site-config
                 harder to maintain than necessary. -->
            </para></listitem>

        <listitem><para>If possible, <code>init</code> must be callable
          with no parameters. In which case, it should try to autodetect all
          the necessary information, for example, by looking for a tool in
          <envar>PATH</envar> or in common installation locations. Often this
          is possible and allows the user to simply write:
<programlisting>
using yfc ;
</programlisting>
          </para></listitem>

        <listitem><para>Consider using facilities in the
          <code>tools/common</code> module. You can take a look at how
          <code>tools/gcc.jam</code> uses that module in the <code>init</code> rule.
          </para></listitem>

      </itemizedlist>
    </para>


  </section>

  </chapter>

<!--
     Local Variables:
     sgml-indent-data: t
     sgml-parent-document: ("userman.xml" "chapter")
     sgml-set-face: t
     End:
-->