[mirror_edk2.git] / AppPkg / Applications / Python / Python-2.7.2 / Demo / metaclasses / index.html

<HTML>\r
\r
<HEAD>\r
<TITLE>Metaclasses in Python 1.5</TITLE>\r
</HEAD>\r
\r
<BODY BGCOLOR="FFFFFF">\r
\r
<H1>Metaclasses in Python 1.5</H1>\r
<H2>(A.k.a. The Killer Joke :-)</H2>\r
\r
<HR>\r
\r
(<i>Postscript:</i> reading this essay is probably not the best way to\r
understand the metaclass hook described here.  See a <A\r
HREF="meta-vladimir.txt">message posted by Vladimir Marangozov</A>\r
which may give a gentler introduction to the matter.  You may also\r
want to search Deja News for messages with "metaclass" in the subject\r
posted to comp.lang.python in July and August 1998.)\r
\r
<HR>\r
\r
<P>In previous Python releases (and still in 1.5), there is something\r
called the ``Don Beaudry hook'', after its inventor and champion.\r
This allows C extensions to provide alternate class behavior, thereby\r
allowing the Python class syntax to be used to define other class-like\r
entities.  Don Beaudry has used this in his infamous <A\r
HREF="http://maigret.cog.brown.edu/pyutil/">MESS</A> package; Jim\r
Fulton has used it in his <A\r
HREF="http://www.digicool.com/releases/ExtensionClass/">Extension\r
Classes</A> package.  (It has also been referred to as the ``Don\r
Beaudry <i>hack</i>,'' but that's a misnomer.  There's nothing hackish\r
about it -- in fact, it is rather elegant and deep, even though\r
there's something dark to it.)\r
\r
<P>(On first reading, you may want to skip directly to the examples in\r
the section "Writing Metaclasses in Python" below, unless you want\r
your head to explode.)\r
\r
<P>\r
\r
<HR>\r
\r
<P>Documentation of the Don Beaudry hook has purposefully been kept\r
minimal, since it is a feature of incredible power, and is easily\r
abused.  Basically, it checks whether the <b>type of the base\r
class</b> is callable, and if so, it is called to create the new\r
class.\r
\r
<P>Note the two indirection levels.  Take a simple example:\r
\r
<PRE>\r
class B:\r
    pass\r
\r
class C(B):\r
    pass\r
</PRE>\r
\r
Take a look at the second class definition, and try to fathom ``the\r
type of the base class is callable.''\r
\r
<P>(Types are not classes, by the way.  See questions 4.2, 4.19 and in\r
particular 6.22 in the <A\r
HREF="http://www.python.org/cgi-bin/faqw.py" >Python FAQ</A>\r
for more on this topic.)\r
\r
<P>\r
\r
<UL>\r
\r
<LI>The <b>base class</b> is B; this one's easy.<P>\r
\r
<LI>Since B is a class, its type is ``class''; so the <b>type of the\r
base class</b> is the type ``class''.  This is also known as\r
types.ClassType, assuming the standard module <code>types</code> has\r
been imported.<P>\r
\r
<LI>Now is the type ``class'' <b>callable</b>?  No, because types (in\r
core Python) are never callable.  Classes are callable (calling a\r
class creates a new instance) but types aren't.<P>\r
\r
</UL>\r
\r
<P>So our conclusion is that in our example, the type of the base\r
class (of C) is not callable.  So the Don Beaudry hook does not apply,\r
and the default class creation mechanism is used (which is also used\r
when there is no base class).  In fact, the Don Beaudry hook never\r
applies when using only core Python, since the type of a core object\r
is never callable.\r
\r
<P>So what do Don and Jim do in order to use Don's hook?  Write an\r
extension that defines at least two new Python object types.  The\r
first would be the type for ``class-like'' objects usable as a base\r
class, to trigger Don's hook.  This type must be made callable.\r
That's why we need a second type.  Whether an object is callable\r
depends on its type.  So whether a type object is callable depends on\r
<i>its</i> type, which is a <i>meta-type</i>.  (In core Python there\r
is only one meta-type, the type ``type'' (types.TypeType), which is\r
the type of all type objects, even itself.)  A new meta-type must\r
be defined that makes the type of the class-like objects callable.\r
(Normally, a third type would also be needed, the new ``instance''\r
type, but this is not an absolute requirement -- the new class type\r
could return an object of some existing type when invoked to create an\r
instance.)\r
\r
<P>Still confused?  Here's a simple device due to Don himself to\r
explain metaclasses.  Take a simple class definition; assume B is a\r
special class that triggers Don's hook:\r
\r
<PRE>\r
class C(B):\r
    a = 1\r
    b = 2\r
</PRE>\r
\r
This can be though of as equivalent to:\r
\r
<PRE>\r
C = type(B)('C', (B,), {'a': 1, 'b': 2})\r
</PRE>\r
\r
If that's too dense for you, here's the same thing written out using\r
temporary variables:\r
\r
<PRE>\r
creator = type(B)               # The type of the base class\r
name = 'C'                      # The name of the new class\r
bases = (B,)                    # A tuple containing the base class(es)\r
namespace = {'a': 1, 'b': 2}    # The namespace of the class statement\r
C = creator(name, bases, namespace)\r
</PRE>\r
\r
This is analogous to what happens without the Don Beaudry hook, except\r
that in that case the creator function is set to the default class\r
creator.\r
\r
<P>In either case, the creator is called with three arguments.  The\r
first one, <i>name</i>, is the name of the new class (as given at the\r
top of the class statement).  The <i>bases</i> argument is a tuple of\r
base classes (a singleton tuple if there's only one base class, like\r
the example).  Finally, <i>namespace</i> is a dictionary containing\r
the local variables collected during execution of the class statement.\r
\r
<P>Note that the contents of the namespace dictionary is simply\r
whatever names were defined in the class statement.  A little-known\r
fact is that when Python executes a class statement, it enters a new\r
local namespace, and all assignments and function definitions take\r
place in this namespace.  Thus, after executing the following class\r
statement:\r
\r
<PRE>\r
class C:\r
    a = 1\r
    def f(s): pass\r
</PRE>\r
\r
the class namespace's contents would be {'a': 1, 'f': &lt;function f\r
...&gt;}.\r
\r
<P>But enough already about writing Python metaclasses in C; read the\r
documentation of <A\r
HREF="http://maigret.cog.brown.edu/pyutil/">MESS</A> or <A\r
HREF="http://www.digicool.com/papers/ExtensionClass.html" >Extension\r
Classes</A> for more information.\r
\r
<P>\r
\r
<HR>\r
\r
<H2>Writing Metaclasses in Python</H2>\r
\r
<P>In Python 1.5, the requirement to write a C extension in order to\r
write metaclasses has been dropped (though you can still do\r
it, of course).  In addition to the check ``is the type of the base\r
class callable,'' there's a check ``does the base class have a\r
__class__ attribute.''  If so, it is assumed that the __class__\r
attribute refers to a class.\r
\r
<P>Let's repeat our simple example from above:\r
\r
<PRE>\r
class C(B):\r
    a = 1\r
    b = 2\r
</PRE>\r
\r
Assuming B has a __class__ attribute, this translates into:\r
\r
<PRE>\r
C = B.__class__('C', (B,), {'a': 1, 'b': 2})\r
</PRE>\r
\r
This is exactly the same as before except that instead of type(B),\r
B.__class__ is invoked.  If you have read <A HREF=\r
"http://www.python.org/cgi-bin/faqw.py?req=show&file=faq06.022.htp"\r
>FAQ question 6.22</A> you will understand that while there is a big\r
technical difference between type(B) and B.__class__, they play the\r
same role at different abstraction levels.  And perhaps at some point\r
in the future they will really be the same thing (at which point you\r
would be able to derive subclasses from built-in types).\r
\r
<P>At this point it may be worth mentioning that C.__class__ is the\r
same object as B.__class__, i.e., C's metaclass is the same as B's\r
metaclass.  In other words, subclassing an existing class creates a\r
new (meta)inststance of the base class's metaclass.\r
\r
<P>Going back to the example, the class B.__class__ is instantiated,\r
passing its constructor the same three arguments that are passed to\r
the default class constructor or to an extension's metaclass:\r
<i>name</i>, <i>bases</i>, and <i>namespace</i>.\r
\r
<P>It is easy to be confused by what exactly happens when using a\r
metaclass, because we lose the absolute distinction between classes\r
and instances: a class is an instance of a metaclass (a\r
``metainstance''), but technically (i.e. in the eyes of the python\r
runtime system), the metaclass is just a class, and the metainstance\r
is just an instance.  At the end of the class statement, the metaclass\r
whose metainstance is used as a base class is instantiated, yielding a\r
second metainstance (of the same metaclass).  This metainstance is\r
then used as a (normal, non-meta) class; instantiation of the class\r
means calling the metainstance, and this will return a real instance.\r
And what class is that an instance of?  Conceptually, it is of course\r
an instance of our metainstance; but in most cases the Python runtime\r
system will see it as an instance of a a helper class used by the\r
metaclass to implement its (non-meta) instances...\r
\r
<P>Hopefully an example will make things clearer.  Let's presume we\r
have a metaclass MetaClass1.  It's helper class (for non-meta\r
instances) is callled HelperClass1.  We now (manually) instantiate\r
MetaClass1 once to get an empty special base class:\r
\r
<PRE>\r
BaseClass1 = MetaClass1("BaseClass1", (), {})\r
</PRE>\r
\r
We can now use BaseClass1 as a base class in a class statement:\r
\r
<PRE>\r
class MySpecialClass(BaseClass1):\r
    i = 1\r
    def f(s): pass\r
</PRE>\r
\r
At this point, MySpecialClass is defined; it is a metainstance of\r
MetaClass1 just like BaseClass1, and in fact the expression\r
``BaseClass1.__class__ == MySpecialClass.__class__ == MetaClass1''\r
yields true.\r
\r
<P>We are now ready to create instances of MySpecialClass.  Let's\r
assume that no constructor arguments are required:\r
\r
<PRE>\r
x = MySpecialClass()\r
y = MySpecialClass()\r
print x.__class__, y.__class__\r
</PRE>\r
\r
The print statement shows that x and y are instances of HelperClass1.\r
How did this happen?  MySpecialClass is an instance of MetaClass1\r
(``meta'' is irrelevant here); when an instance is called, its\r
__call__ method is invoked, and presumably the __call__ method defined\r
by MetaClass1 returns an instance of HelperClass1.\r
\r
<P>Now let's see how we could use metaclasses -- what can we do\r
with metaclasses that we can't easily do without them?  Here's one\r
idea: a metaclass could automatically insert trace calls for all\r
method calls.  Let's first develop a simplified example, without\r
support for inheritance or other ``advanced'' Python features (we'll\r
add those later).\r
\r
<PRE>\r
import types\r
\r
class Tracing:\r
    def __init__(self, name, bases, namespace):\r
        """Create a new class."""\r
        self.__name__ = name\r
        self.__bases__ = bases\r
        self.__namespace__ = namespace\r
    def __call__(self):\r
        """Create a new instance."""\r
        return Instance(self)\r
\r
class Instance:\r
    def __init__(self, klass):\r
        self.__klass__ = klass\r
    def __getattr__(self, name):\r
        try:\r
            value = self.__klass__.__namespace__[name]\r
        except KeyError:\r
            raise AttributeError, name\r
        if type(value) is not types.FunctionType:\r
            return value\r
        return BoundMethod(value, self)\r
\r
class BoundMethod:\r
    def __init__(self, function, instance):\r
        self.function = function\r
        self.instance = instance\r
    def __call__(self, *args):\r
        print "calling", self.function, "for", self.instance, "with", args\r
        return apply(self.function, (self.instance,) + args)\r
\r
Trace = Tracing('Trace', (), {})\r
\r
class MyTracedClass(Trace):\r
    def method1(self, a):\r
        self.a = a\r
    def method2(self):\r
        return self.a\r
\r
aninstance = MyTracedClass()\r
\r
aninstance.method1(10)\r
\r
print "the answer is %d" % aninstance.method2()\r
</PRE>\r
\r
Confused already?  The intention is to read this from top down.  The\r
Tracing class is the metaclass we're defining.  Its structure is\r
really simple.\r
\r
<P>\r
\r
<UL>\r
\r
<LI>The __init__ method is invoked when a new Tracing instance is\r
created, e.g. the definition of class MyTracedClass later in the\r
example.  It simply saves the class name, base classes and namespace\r
as instance variables.<P>\r
\r
<LI>The __call__ method is invoked when a Tracing instance is called,\r
e.g. the creation of aninstance later in the example.  It returns an\r
instance of the class Instance, which is defined next.<P>\r
\r
</UL>\r
\r
<P>The class Instance is the class used for all instances of classes\r
built using the Tracing metaclass, e.g. aninstance.  It has two\r
methods:\r
\r
<P>\r
\r
<UL>\r
\r
<LI>The __init__ method is invoked from the Tracing.__call__ method\r
above to initialize a new instance.  It saves the class reference as\r
an instance variable.  It uses a funny name because the user's\r
instance variables (e.g. self.a later in the example) live in the same\r
namespace.<P>\r
\r
<LI>The __getattr__ method is invoked whenever the user code\r
references an attribute of the instance that is not an instance\r
variable (nor a class variable; but except for __init__ and\r
__getattr__ there are no class variables).  It will be called, for\r
example, when aninstance.method1 is referenced in the example, with\r
self set to aninstance and name set to the string "method1".<P>\r
\r
</UL>\r
\r
<P>The __getattr__ method looks the name up in the __namespace__\r
dictionary.  If it isn't found, it raises an AttributeError exception.\r
(In a more realistic example, it would first have to look through the\r
base classes as well.)  If it is found, there are two possibilities:\r
it's either a function or it isn't.  If it's not a function, it is\r
assumed to be a class variable, and its value is returned.  If it's a\r
function, we have to ``wrap'' it in instance of yet another helper\r
class, BoundMethod.\r
\r
<P>The BoundMethod class is needed to implement a familiar feature:\r
when a method is defined, it has an initial argument, self, which is\r
automatically bound to the relevant instance when it is called.  For\r
example, aninstance.method1(10) is equivalent to method1(aninstance,\r
10).  In the example if this call, first a temporary BoundMethod\r
instance is created with the following constructor call: temp =\r
BoundMethod(method1, aninstance); then this instance is called as\r
temp(10).  After the call, the temporary instance is discarded.\r
\r
<P>\r
\r
<UL>\r
\r
<LI>The __init__ method is invoked for the constructor call\r
BoundMethod(method1, aninstance).  It simply saves away its\r
arguments.<P>\r
\r
<LI>The __call__ method is invoked when the bound method instance is\r
called, as in temp(10).  It needs to call method1(aninstance, 10).\r
However, even though self.function is now method1 and self.instance is\r
aninstance, it can't call self.function(self.instance, args) directly,\r
because it should work regardless of the number of arguments passed.\r
(For simplicity, support for keyword arguments has been omitted.)<P>\r
\r
</UL>\r
\r
<P>In order to be able to support arbitrary argument lists, the\r
__call__ method first constructs a new argument tuple.  Conveniently,\r
because of the notation *args in __call__'s own argument list, the\r
arguments to __call__ (except for self) are placed in the tuple args.\r
To construct the desired argument list, we concatenate a singleton\r
tuple containing the instance with the args tuple: (self.instance,) +\r
args.  (Note the trailing comma used to construct the singleton\r
tuple.)  In our example, the resulting argument tuple is (aninstance,\r
10).\r
\r
<P>The intrinsic function apply() takes a function and an argument\r
tuple and calls the function for it.  In our example, we are calling\r
apply(method1, (aninstance, 10)) which is equivalent to calling\r
method(aninstance, 10).\r
\r
<P>From here on, things should come together quite easily.  The output\r
of the example code is something like this:\r
\r
<PRE>\r
calling &lt;function method1 at ae8d8&gt; for &lt;Instance instance at 95ab0&gt; with (10,)\r
calling &lt;function method2 at ae900&gt; for &lt;Instance instance at 95ab0&gt; with ()\r
the answer is 10\r
</PRE>\r
\r
<P>That was about the shortest meaningful example that I could come up\r
with.  A real tracing metaclass (for example, <A\r
HREF="#Trace">Trace.py</A> discussed below) needs to be more\r
complicated in two dimensions.\r
\r
<P>First, it needs to support more advanced Python features such as\r
class variables, inheritance, __init__ methods, and keyword arguments.\r
\r
<P>Second, it needs to provide a more flexible way to handle the\r
actual tracing information; perhaps it should be possible to write\r
your own tracing function that gets called, perhaps it should be\r
possible to enable and disable tracing on a per-class or per-instance\r
basis, and perhaps a filter so that only interesting calls are traced;\r
it should also be able to trace the return value of the call (or the\r
exception it raised if an error occurs).  Even the Trace.py example\r
doesn't support all these features yet.\r
\r
<P>\r
\r
<HR>\r
\r
<H1>Real-life Examples</H1>\r
\r
<P>Have a look at some very preliminary examples that I coded up to\r
teach myself how to write metaclasses:\r
\r
<DL>\r
\r
<DT><A HREF="Enum.py">Enum.py</A>\r
\r
<DD>This (ab)uses the class syntax as an elegant way to define\r
enumerated types.  The resulting classes are never instantiated --\r
rather, their class attributes are the enumerated values.  For\r
example:\r
\r
<PRE>\r
class Color(Enum):\r
    red = 1\r
    green = 2\r
    blue = 3\r
print Color.red\r
</PRE>\r
\r
will print the string ``Color.red'', while ``Color.red==1'' is true,\r
and ``Color.red + 1'' raise a TypeError exception.\r
\r
<P>\r
\r
<DT><A NAME=Trace></A><A HREF="Trace.py">Trace.py</A>\r
\r
<DD>The resulting classes work much like standard\r
classes, but by setting a special class or instance attribute\r
__trace_output__ to point to a file, all calls to the class's methods\r
are traced.  It was a bit of a struggle to get this right.  This\r
should probably redone using the generic metaclass below.\r
\r
<P>\r
\r
<DT><A HREF="Meta.py">Meta.py</A>\r
\r
<DD>A generic metaclass.  This is an attempt at finding out how much\r
standard class behavior can be mimicked by a metaclass.  The\r
preliminary answer appears to be that everything's fine as long as the\r
class (or its clients) don't look at the instance's __class__\r
attribute, nor at the class's __dict__ attribute.  The use of\r
__getattr__ internally makes the classic implementation of __getattr__\r
hooks tough; we provide a similar hook _getattr_ instead.\r
(__setattr__ and __delattr__ are not affected.)\r
(XXX Hm.  Could detect presence of __getattr__ and rename it.)\r
\r
<P>\r
\r
<DT><A HREF="Eiffel.py">Eiffel.py</A>\r
\r
<DD>Uses the above generic metaclass to implement Eiffel style\r
pre-conditions and post-conditions.\r
\r
<P>\r
\r
<DT><A HREF="Synch.py">Synch.py</A>\r
\r
<DD>Uses the above generic metaclass to implement synchronized\r
methods.\r
\r
<P>\r
\r
<DT><A HREF="Simple.py">Simple.py</A>\r
\r
<DD>The example module used above.\r
\r
<P>\r
\r
</DL>\r
\r
<P>A pattern seems to be emerging: almost all these uses of\r
metaclasses (except for Enum, which is probably more cute than useful)\r
mostly work by placing wrappers around method calls.  An obvious\r
problem with that is that it's not easy to combine the features of\r
different metaclasses, while this would actually be quite useful: for\r
example, I wouldn't mind getting a trace from the test run of the\r
Synch module, and it would be interesting to add preconditions to it\r
as well.  This needs more research.  Perhaps a metaclass could be\r
provided that allows stackable wrappers...\r
\r
<P>\r
\r
<HR>\r
\r
<H2>Things You Could Do With Metaclasses</H2>\r
\r
<P>There are lots of things you could do with metaclasses.  Most of\r
these can also be done with creative use of __getattr__, but\r
metaclasses make it easier to modify the attribute lookup behavior of\r
classes.  Here's a partial list.\r
\r
<P>\r
\r
<UL>\r
\r
<LI>Enforce different inheritance semantics, e.g. automatically call\r
base class methods when a derived class overrides<P>\r
\r
<LI>Implement class methods (e.g. if the first argument is not named\r
'self')<P>\r
\r
<LI>Implement that each instance is initialized with <b>copies</b> of\r
all class variables<P>\r
\r
<LI>Implement a different way to store instance variables (e.g. in a\r
list kept outside the instance but indexed by the instance's id())<P>\r
\r
<LI>Automatically wrap or trap all or certain methods\r
\r
<UL>\r
\r
<LI>for tracing\r
\r
<LI>for precondition and postcondition checking\r
\r
<LI>for synchronized methods\r
\r
<LI>for automatic value caching\r
\r
</UL>\r
<P>\r
\r
<LI>When an attribute is a parameterless function, call it on\r
reference (to mimic it being an instance variable); same on assignment<P>\r
\r
<LI>Instrumentation: see how many times various attributes are used<P>\r
\r
<LI>Different semantics for __setattr__ and __getattr__ (e.g.  disable\r
them when they are being used recursively)<P>\r
\r
<LI>Abuse class syntax for other things<P>\r
\r
<LI>Experiment with automatic type checking<P>\r
\r
<LI>Delegation (or acquisition)<P>\r
\r
<LI>Dynamic inheritance patterns<P>\r
\r
<LI>Automatic caching of methods<P>\r
\r
</UL>\r
\r
<P>\r
\r
<HR>\r
\r
<H4>Credits</H4>\r
\r
<P>Many thanks to David Ascher and Donald Beaudry for their comments\r
on earlier draft of this paper.  Also thanks to Matt Conway and Tommy\r
Burnette for putting a seed for the idea of metaclasses in my\r
mind, nearly three years ago, even though at the time my response was\r
``you can do that with __getattr__ hooks...'' :-)\r
\r
<P>\r
\r
<HR>\r
\r
</BODY>\r
\r
</HTML>\r