boost/python
2
0
Fork 0
mirror of https://github.com/boostorg/python.git synced 2026-01-24 18:12:43 +00:00
Code Issues Projects Releases Wiki Activity

added internals

Browse Source 
[SVN r19390]
This commit is contained in:
Dave Abrahams
2003-07-31 18:03:17 +00:00
parent bfa868a440
commit f48aacf477
3 changed files with 354 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
doc/index.html
View File

@@ -88,6 +88,8 @@
<dt><a href="../pyste/index.html">Pyste (Boost.Python code generator)</a></dt>
<dt><a href="internals.html">Internals Documentation</a></dt>
<dt><a href="news.html">News/Change Log</a></dt>
<dt><a href="v2/progress_reports.html">LLNL Progress Reports</a></dt>

184
doc/internals.html Executable file
View File

@@ -0,0 +1,184 @@
<?xml version="1.0" encoding="utf-8" ?>
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
<html xmlns="http://www.w3.org/1999/xhtml" xml:lang="en" lang="en">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8" />
<meta name="generator" content="Docutils 0.3.0: http://docutils.sourceforge.net/" />
<title>Boost Python Internals</title>
<meta name="copyright" content="Copyright Brett Calcott and David Abrahams 2003. All rights reserved." />
<link rel="stylesheet" href="../../../rst.css" type="text/css" />
</head>
<body>
<div class="document" id="boost-python-internals">
<h1 class="title">Boost Python Internals</h1>
<h2 class="subtitle" id="a-conversation-between-brett-calcott-and-david-abrahams">A conversation between Brett Calcott and David Abrahams</h2>
<table class="docinfo" frame="void" rules="none">
<col class="docinfo-name" />
<col class="docinfo-content" />
<tbody valign="top">
<tr><th class="docinfo-name">Copyright:</th>
<td>Copyright Brett Calcott and David Abrahams 2003. All
rights reserved.</td></tr>
</tbody>
</table>
<p>In both of these cases, I'm quite capable of reading code - but the
thing I don't get from scanning the source is a sense of the
architecture, both structurally, and temporally (er, I mean in what
order things go on).</p>
<ol class="arabic">
<li><p class="first">What happens when you do the following:</p>
<pre class="literal-block">
struct boring {};
...etc...
class_&lt;boring&gt;(&quot;boring&quot;)
;
</pre>
</li>
</ol>
<p>There seems to be a fair bit going on.</p>
<blockquote>
<ul class="simple">
<li>Python needs a new ClassType to be registered.</li>
<li>We need to construct a new type that can hold our boring struct.</li>
<li>Inward and outward converters need to be registered for the type.</li>
</ul>
</blockquote>
<p>Can you gesture in the general direction where these things are done?</p>
<blockquote>
<p>I only have time for a &quot;off-the-top-of-my-head&quot; answer at the moment;
I suggest you step through the code with a debugger after reading this
to see how it works, fill in details, and make sure I didn't forget
anything.</p>
<blockquote>
<p>A new (Python) subclass of Boost.Python.Instance (see
libs/python/src/object/class.cpp) is created by invoking
Boost.Python.class, the metatype:</p>
<pre class="literal-block">
&gt;&gt;&gt; boring = Boost.Python.class(
... 'boring'
... , bases_tuple # in this case, just ()
... , {
... '__module__' : module_name
... , '__doc__' : doc_string # optional
... }
... )
</pre>
<p>A handle to this object is stuck in the m_class_object field
of the registration associated with <tt class="literal"><span class="pre">typeid(boring)</span></tt>. The
registry will keep that object alive forever, even if you
wipe out the 'boring' attribute of the extension module
(probably not a good thing).</p>
<p>Because you didn't specify <tt class="literal"><span class="pre">class&lt;boring,</span> <span class="pre">non_copyable,</span>
<span class="pre">...&gt;</span></tt>, a to-python converter for boring is registered which
copies its argument into a value_holder held by the the
Python boring object.</p>
<p>Because you didn't specify <tt class="literal"><span class="pre">class&lt;boring</span> <span class="pre">...&gt;(no_init)</span></tt>,
an <tt class="literal"><span class="pre">__init__</span></tt> function object is added to the class
dictionary which default-constructs a boring in a
value_holder (because you didn't specify some smart pointer
or derived wrapper class as a holder) held by the Python
boring object.</p>
<p><tt class="literal"><span class="pre">register_class_from_python</span></tt> is used to register a
from-python converter for <tt class="literal"><span class="pre">shared_ptr&lt;boring&gt;</span></tt>.
<tt class="literal"><span class="pre">boost::shared_ptr</span></tt>s are special among smart pointers
because their Deleter argument can be made to manage the
whole Python object, not just the C++ object it contains, no
matter how the C++ object is held.</p>
<p>If there were any <tt class="literal"><span class="pre">bases&lt;&gt;</span></tt>, we'd also be registering the
relationship between these base classes and boring in the
up/down cast graph (<tt class="literal"><span class="pre">inheritance.[hpp/cpp]</span></tt>).</p>
<p>In earlier versions of the code, we'd be registering lvalue
from-python converters for the class here, but now
from-python conversion for wrapped classes is handled as a
special case, before consulting the registry, if the source
Python object's metaclass is the Boost.Python metaclass.</p>
<p>Hmm, that from-python converter probably ought to be handled
the way class converters are, with no explicit conversions
registered.</p>
</blockquote>
</blockquote>
<ol class="arabic" start="2">
<li><p class="first">Can you give a brief overview of the data structures that are
present in the registry</p>
<blockquote>
<p>The registry is simple: it's just a map from typeid -&gt;
registration (see boost/python/converter/registrations.hpp).
<tt class="literal"><span class="pre">lvalue_chain</span></tt> and <tt class="literal"><span class="pre">rvalue_chain</span></tt> are simple endogenous
linked lists.</p>
<p>If you want to know more, just ask.</p>
<p>If you want to know about the cast graph, ask me something specific in
a separate message.</p>
</blockquote>
<p>and an overview of the process that happens as a type makes its
way from c++ to python and back again.</p>
</li>
</ol>
<blockquote>
<p>Big subject. I suggest some background reading: look for relevant
info in the LLNL progress reports and the messages they link to.
Also,</p>
<blockquote>
<a class="reference" href="http://mail.python.org/pipermail/c++-sig/2002-May/001023.html">http://mail.python.org/pipermail/c++-sig/2002-May/001023.html</a>
<a class="reference" href="http://mail.python.org/pipermail/c++-sig/2002-December/003115.html">http://mail.python.org/pipermail/c++-sig/2002-December/003115.html</a>
<a class="reference" href="http://aspn.activestate.com/ASPN/Mail/Message/1280898">http://aspn.activestate.com/ASPN/Mail/Message/1280898</a>
<a class="reference" href="http://mail.python.org/pipermail/c++-sig/2002-July/001755.html">http://mail.python.org/pipermail/c++-sig/2002-July/001755.html</a></blockquote>
<p>from c++ to python:</p>
<blockquote>
<p>It depends on the type and the call policies in use or, for
<tt class="literal"><span class="pre">call&lt;&gt;(...)</span></tt>, <tt class="literal"><span class="pre">call_method&lt;&gt;(...)</span></tt>, or <tt class="literal"><span class="pre">object(...)</span></tt>, if
<tt class="literal"><span class="pre">ref</span></tt> or <tt class="literal"><span class="pre">ptr</span></tt> is used. There are also two basic
categories to to-python conversion, &quot;return value&quot; conversion
(for Python-&gt;C++ calls) and &quot;argument&quot; conversion (for
C++-&gt;Python calls and explicit <tt class="literal"><span class="pre">object()</span></tt> conversions). The
behavior of these two categories differs subtly in various ways
whose details I forget at the moment. You can probably find
the answers in the above references, and certainly in the code.</p>
<p>The &quot;default&quot; case is by-value (copying) conversion, which uses
to_python_value as a to-python converter.</p>
<blockquote>
<p>Since there can sensibly be only one way to convert any type
to python (disregarding the idea of scoped registries for the
moment), it makes sense that to-python conversions can be
handled by specializing a template. If the type is one of
the types handled by a built-in conversion
(builtin_converters.hpp), the corresponding template
specialization of to_python_value gets used.</p>
<p>Otherwise, to_python_value uses the <tt class="literal"><span class="pre">m_to_python</span></tt>
function in the registration for the C++ type.</p>
</blockquote>
<p>Other conversions, like by-reference conversions, are only
available for wrapped classes, and are requested explicitly by
using <tt class="literal"><span class="pre">ref(...)</span></tt>, <tt class="literal"><span class="pre">ptr(...)</span></tt>, or by specifying different
CallPolicies for a call, which can cause a different to-python
converter to be used. These conversions are never registered
anywhere, though they do need to use the registration to find
the Python class corresponding to the C++ type being referred
to. They just build a new Python instance and stick the
appropriate Holder instance in it.</p>
</blockquote>
<p>from python to C++:</p>
<blockquote>
<p>Once again I think there is a distinction between &quot;return value&quot;
and &quot;argument&quot; conversions, and I forget exactly what that is.</p>
<p>What happens depends on whether an lvalue conversion is needed
(see <a class="reference" href="http://mail.python.org/pipermail/c++-sig/2002-May/001023.html">http://mail.python.org/pipermail/c++-sig/2002-May/001023.html</a>)
All lvalue conversions are also registered in a type's rvalue
conversion chain, since when an rvalue will do, an lvalue is
certainly good enough.</p>
<p>An lvalue conversion can be done in one step (just get me the
pointer to the object - it can be <tt class="literal"><span class="pre">NULL</span></tt> if no conversion is
possible) while an rvalue conversion requires two steps to
support wrapped function overloading and multiple converters for
a given C++ target type: first tell me if a conversion is
possible, then construct the converted object as a second step.</p>
</blockquote>
</blockquote>
</div>
<hr class="footer"/>
<div class="footer">
<a class="reference" href="internals.rst">View document source</a>.
Generated on: 2003-07-31 17:58 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>
</html>

168
doc/internals.rst Executable file
View File

@@ -0,0 +1,168 @@
========================
Boost Python Internals
========================
-------------------------------------------------------
A conversation between Brett Calcott and David Abrahams
-------------------------------------------------------
:copyright: Copyright Brett Calcott and David Abrahams 2003. All
rights reserved.
In both of these cases, I'm quite capable of reading code - but the
thing I don't get from scanning the source is a sense of the
architecture, both structurally, and temporally (er, I mean in what
order things go on).
1) What happens when you do the following::
struct boring {};
...etc...
class_<boring>("boring")
;
There seems to be a fair bit going on.
- Python needs a new ClassType to be registered.
- We need to construct a new type that can hold our boring struct.
- Inward and outward converters need to be registered for the type.
Can you gesture in the general direction where these things are done?
I only have time for a "off-the-top-of-my-head" answer at the moment;
I suggest you step through the code with a debugger after reading this
to see how it works, fill in details, and make sure I didn't forget
anything.
A new (Python) subclass of Boost.Python.Instance (see
libs/python/src/object/class.cpp) is created by invoking
Boost.Python.class, the metatype::
>>> boring = Boost.Python.class(
... 'boring'
... , bases_tuple # in this case, just ()
... , {
... '__module__' : module_name
... , '__doc__' : doc_string # optional
... }
... )
A handle to this object is stuck in the m_class_object field
of the registration associated with ``typeid(boring)``. The
registry will keep that object alive forever, even if you
wipe out the 'boring' attribute of the extension module
(probably not a good thing).
Because you didn't specify ``class<boring, non_copyable,
...>``, a to-python converter for boring is registered which
copies its argument into a value_holder held by the the
Python boring object.
Because you didn't specify ``class<boring ...>(no_init)``,
an ``__init__`` function object is added to the class
dictionary which default-constructs a boring in a
value_holder (because you didn't specify some smart pointer
or derived wrapper class as a holder) held by the Python
boring object.
``register_class_from_python`` is used to register a
from-python converter for ``shared_ptr<boring>``.
``boost::shared_ptr``\ s are special among smart pointers
because their Deleter argument can be made to manage the
whole Python object, not just the C++ object it contains, no
matter how the C++ object is held.
If there were any ``bases<>``, we'd also be registering the
relationship between these base classes and boring in the
up/down cast graph (``inheritance.[hpp/cpp]``).
In earlier versions of the code, we'd be registering lvalue
from-python converters for the class here, but now
from-python conversion for wrapped classes is handled as a
special case, before consulting the registry, if the source
Python object's metaclass is the Boost.Python metaclass.
Hmm, that from-python converter probably ought to be handled
the way class converters are, with no explicit conversions
registered.
2) Can you give a brief overview of the data structures that are
present in the registry
The registry is simple: it's just a map from typeid ->
registration (see boost/python/converter/registrations.hpp).
``lvalue_chain`` and ``rvalue_chain`` are simple endogenous
linked lists.
If you want to know more, just ask.
If you want to know about the cast graph, ask me something specific in
a separate message.
and an overview of the process that happens as a type makes its
way from c++ to python and back again.
Big subject. I suggest some background reading: look for relevant
info in the LLNL progress reports and the messages they link to.
Also,
http://mail.python.org/pipermail/c++-sig/2002-May/001023.html
http://mail.python.org/pipermail/c++-sig/2002-December/003115.html
http://aspn.activestate.com/ASPN/Mail/Message/1280898
http://mail.python.org/pipermail/c++-sig/2002-July/001755.html
from c++ to python:
It depends on the type and the call policies in use or, for
``call<>(...)``, ``call_method<>(...)``, or ``object(...)``, if
``ref`` or ``ptr`` is used. There are also two basic
categories to to-python conversion, "return value" conversion
(for Python->C++ calls) and "argument" conversion (for
C++->Python calls and explicit ``object()`` conversions). The
behavior of these two categories differs subtly in various ways
whose details I forget at the moment. You can probably find
the answers in the above references, and certainly in the code.
The "default" case is by-value (copying) conversion, which uses
to_python_value as a to-python converter.
Since there can sensibly be only one way to convert any type
to python (disregarding the idea of scoped registries for the
moment), it makes sense that to-python conversions can be
handled by specializing a template. If the type is one of
the types handled by a built-in conversion
(builtin_converters.hpp), the corresponding template
specialization of to_python_value gets used.
Otherwise, to_python_value uses the ``m_to_python``
function in the registration for the C++ type.
Other conversions, like by-reference conversions, are only
available for wrapped classes, and are requested explicitly by
using ``ref(...)``, ``ptr(...)``, or by specifying different
CallPolicies for a call, which can cause a different to-python
converter to be used. These conversions are never registered
anywhere, though they do need to use the registration to find
the Python class corresponding to the C++ type being referred
to. They just build a new Python instance and stick the
appropriate Holder instance in it.
from python to C++:
Once again I think there is a distinction between "return value"
and "argument" conversions, and I forget exactly what that is.
What happens depends on whether an lvalue conversion is needed
(see http://mail.python.org/pipermail/c++-sig/2002-May/001023.html)
All lvalue conversions are also registered in a type's rvalue
conversion chain, since when an rvalue will do, an lvalue is
certainly good enough.
An lvalue conversion can be done in one step (just get me the
pointer to the object - it can be ``NULL`` if no conversion is
possible) while an rvalue conversion requires two steps to
support wrapped function overloading and multiple converters for
a given C++ target type: first tell me if a conversion is
possible, then construct the converted object as a second step.