PEP: 253 Title: Subtyping Built-in Types Author: Guido van Rossum
<guido@python.org> Status: Final Type: Standards Track Content-Type:
text/x-rst Created: 14-May-2001 Python-Version: 2.2 Post-History:

Abstract

This PEP proposes additions to the type object API that will allow the
creation of subtypes of built-in types, in C and in Python.

[Editor's note: the ideas described in this PEP have been incorporated
into Python. The PEP no longer accurately describes the implementation.]

Introduction

Traditionally, types in Python have been created statically, by
declaring a global variable of type PyTypeObject and initializing it
with a static initializer. The slots in the type object describe all
aspects of a Python type that are relevant to the Python interpreter. A
few slots contain dimensional information (like the basic allocation
size of instances), others contain various flags, but most slots are
pointers to functions to implement various kinds of behaviors. A NULL
pointer means that the type does not implement the specific behavior; in
that case the system may provide a default behavior or raise an
exception when the behavior is invoked for an instance of the type. Some
collections of function pointers that are usually defined together are
obtained indirectly via a pointer to an additional structure containing
more function pointers.

While the details of initializing a PyTypeObject structure haven't been
documented as such, they are easily gleaned from the examples in the
source code, and I am assuming that the reader is sufficiently familiar
with the traditional way of creating new Python types in C.

This PEP will introduce the following features:

-   a type can be a factory function for its instances
-   types can be subtyped in C
-   types can be subtyped in Python with the class statement
-   multiple inheritance from types is supported (insofar as practical
    -- you still can't multiply inherit from list and dictionary)
-   the standard coercion functions (int, tuple, str etc.) will be
    redefined to be the corresponding type objects, which serve as their
    own factory functions
-   a class statement can contain a __metaclass__ declaration,
    specifying the metaclass to be used to create the new class
-   a class statement can contain a __slots__ declaration, specifying
    the specific names of the instance variables supported

This PEP builds on PEP 252, which adds standard introspection to types;
for example, when a particular type object initializes the tp_hash slot,
that type object has a __hash__ method when introspected. PEP 252 also
adds a dictionary to type objects which contains all methods. At the
Python level, this dictionary is read-only for built-in types; at the C
level, it is accessible directly (but it should not be modified except
as part of initialization).

For binary compatibility, a flag bit in the tp_flags slot indicates the
existence of the various new slots in the type object introduced below.
Types that don't have the Py_TPFLAGS_HAVE_CLASS bit set in their
tp_flags slot are assumed to have NULL values for all the subtyping
slots. (Warning: the current implementation prototype is not yet
consistent in its checking of this flag bit. This should be fixed before
the final release.)

In current Python, a distinction is made between types and classes. This
PEP together with PEP 254 will remove that distinction. However, for
backwards compatibility the distinction will probably remain for years
to come, and without PEP 254, the distinction is still large: types
ultimately have a built-in type as a base class, while classes
ultimately derive from a user-defined class. Therefore, in the rest of
this PEP, I will use the word type whenever I can -- including base type
or supertype, derived type or subtype, and metatype. However, sometimes
the terminology necessarily blends, for example an object's type is
given by its __class__ attribute, and subtyping in Python is spelled
with a class statement. If further distinction is necessary,
user-defined classes can be referred to as "classic" classes.

About metatypes

Inevitably the discussion comes to metatypes (or metaclasses). Metatypes
are nothing new in Python: Python has always been able to talk about the
type of a type:

    >>> a = 0
    >>> type(a)
    <type 'int'>
    >>> type(type(a))
    <type 'type'>
    >>> type(type(type(a)))
    <type 'type'>
    >>>

In this example, type(a) is a "regular" type, and type(type(a)) is a
metatype. While as distributed all types have the same metatype
(PyType_Type, which is also its own metatype), this is not a
requirement, and in fact a useful and relevant 3rd party extension
(ExtensionClasses by Jim Fulton) creates an additional metatype. The
type of classic classes, known as types.ClassType, can also be
considered a distinct metatype.

A feature closely connected to metatypes is the "Don Beaudry hook",
which says that if a metatype is callable, its instances (which are
regular types) can be subclassed (really subtyped) using a Python class
statement. I will use this rule to support subtyping of built-in types,
and in fact it greatly simplifies the logic of class creation to always
simply call the metatype. When no base class is specified, a default
metatype is called -- the default metatype is the "ClassType" object, so
the class statement will behave as before in the normal case. (This
default can be changed per module by setting the global variable
__metaclass__.)

Python uses the concept of metatypes or metaclasses in a different way
than Smalltalk. In Smalltalk-80, there is a hierarchy of metaclasses
that mirrors the hierarchy of regular classes, metaclasses map 1-1 to
classes (except for some funny business at the root of the hierarchy),
and each class statement creates both a regular class and its metaclass,
putting class methods in the metaclass and instance methods in the
regular class.

Nice though this may be in the context of Smalltalk, it's not compatible
with the traditional use of metatypes in Python, and I prefer to
continue in the Python way. This means that Python metatypes are
typically written in C, and may be shared between many regular types.
(It will be possible to subtype metatypes in Python, so it won't be
absolutely necessary to write C to use metatypes; but the power of
Python metatypes will be limited. For example, Python code will never be
allowed to allocate raw memory and initialize it at will.)

Metatypes determine various policies for types, such as what happens
when a type is called, how dynamic types are (whether a type's __dict__
can be modified after it is created), what the method resolution order
is, how instance attributes are looked up, and so on.

I'll argue that left-to-right depth-first is not the best solution when
you want to get the most use from multiple inheritance.

I'll argue that with multiple inheritance, the metatype of the subtype
must be a descendant of the metatypes of all base types.

I'll come back to metatypes later.

Making a type a factory for its instances

Traditionally, for each type there is at least one C factory function
that creates instances of the type (PyTuple_New(), PyInt_FromLong() and
so on). These factory functions take care of both allocating memory for
the object and initializing that memory. As of Python 2.0, they also
have to interface with the garbage collection subsystem, if the type
chooses to participate in garbage collection (which is optional, but
strongly recommended for so-called "container" types: types that may
contain references to other objects, and hence may participate in
reference cycles).

In this proposal, type objects can be factory functions for their
instances, making the types directly callable from Python. This mimics
the way classes are instantiated. The C APIs for creating instances of
various built-in types will remain valid and in some cases more
efficient. Not all types will become their own factory functions.

The type object has a new slot, tp_new, which can act as a factory for
instances of the type. Types are now callable, because the tp_call slot
is set in PyType_Type (the metatype); the function looks for the tp_new
slot of the type that is being called.

Explanation: the tp_call slot of a regular type object (such as
PyInt_Type or PyList_Type) defines what happens when instances of that
type are called; in particular, the tp_call slot in the function type,
PyFunction_Type, is the key to making functions callable. As another
example, PyInt_Type.tp_call is NULL, because integers are not callable.
The new paradigm makes type objects callable. Since type objects are
instances of their metatype (PyType_Type), the metatype's tp_call slot
(PyType_Type.tp_call) points to a function that is invoked when any type
object is called. Now, since each type has to do something different to
create an instance of itself, PyType_Type.tp_call immediately defers to
the tp_new slot of the type that is being called. PyType_Type itself is
also callable: its tp_new slot creates a new type. This is used by the
class statement (formalizing the Don Beaudry hook, see above). And what
makes PyType_Type callable? The tp_call slot of its metatype -- but
since it is its own metatype, that is its own tp_call slot!

If the type's tp_new slot is NULL, an exception is raised. Otherwise,
the tp_new slot is called. The signature for the tp_new slot is

    PyObject *tp_new(PyTypeObject *type,
                     PyObject *args,
                     PyObject *kwds)

where 'type' is the type whose tp_new slot is called, and 'args' and
'kwds' are the sequential and keyword arguments to the call, passed
unchanged from tp_call. (The 'type' argument is used in combination with
inheritance, see below.)

There are no constraints on the object type that is returned, although
by convention it should be an instance of the given type. It is not
necessary that a new object is returned; a reference to an existing
object is fine too. The return value should always be a new reference,
owned by the caller.

Once the tp_new slot has returned an object, further initialization is
attempted by calling the tp_init() slot of the resulting object's type,
if not NULL. This has the following signature:

    int tp_init(PyObject *self,
                PyObject *args,
                PyObject *kwds)

It corresponds more closely to the __init__() method of classic classes,
and in fact is mapped to that by the slot/special-method correspondence
rules. The difference in responsibilities between the tp_new() slot and
the tp_init() slot lies in the invariants they ensure. The tp_new() slot
should ensure only the most essential invariants, without which the C
code that implements the objects would break. The tp_init() slot should
be used for overridable user-specific initializations. Take for example
the dictionary type. The implementation has an internal pointer to a
hash table which should never be NULL. This invariant is taken care of
by the tp_new() slot for dictionaries. The dictionary tp_init() slot, on
the other hand, could be used to give the dictionary an initial set of
keys and values based on the arguments passed in.

Note that for immutable object types, the initialization cannot be done
by the tp_init() slot: this would provide the Python user with a way to
change the initialization. Therefore, immutable objects typically have
an empty tp_init() implementation and do all their initialization in
their tp_new() slot.

You may wonder why the tp_new() slot shouldn't call the tp_init() slot
itself. The reason is that in certain circumstances (like support for
persistent objects), it is important to be able to create an object of a
particular type without initializing it any further than necessary. This
may conveniently be done by calling the tp_new() slot without calling
tp_init(). It is also possible that tp_init() is not called, or called
more than once -- its operation should be robust even in these anomalous
cases.

For some objects, tp_new() may return an existing object. For example,
the factory function for integers caches the integers -1 through 99.
This is permissible only when the type argument to tp_new() is the type
that defined the tp_new() function (in the example, if
type == &PyInt_Type), and when the tp_init() slot for this type does
nothing. If the type argument differs, the tp_new() call is initiated by
a derived type's tp_new() to create the object and initialize the base
type portion of the object; in this case tp_new() should always return a
new object (or raise an exception).

Both tp_new() and tp_init() should receive exactly the same 'args' and
'kwds' arguments, and both should check that the arguments are
acceptable, because they may be called independently.

There's a third slot related to object creation: tp_alloc(). Its
responsibility is to allocate the memory for the object, initialize the
reference count (ob_refcnt) and the type pointer (ob_type), and
initialize the rest of the object to all zeros. It should also register
the object with the garbage collection subsystem if the type supports
garbage collection. This slot exists so that derived types can override
the memory allocation policy (like which heap is being used) separately
from the initialization code. The signature is:

    PyObject *tp_alloc(PyTypeObject *type, int nitems)

The type argument is the type of the new object. The nitems argument is
normally zero, except for objects with a variable allocation size
(basically strings, tuples, and longs). The allocation size is given by
the following expression:

    type->tp_basicsize  +  nitems * type->tp_itemsize

The tp_alloc slot is only used for subclassable types. The tp_new()
function of the base class must call the tp_alloc() slot of the type
passed in as its first argument. It is the tp_new() function's
responsibility to calculate the number of items. The tp_alloc() slot
will set the ob_size member of the new object if the type->tp_itemsize
member is nonzero.

(Note: in certain debugging compilation modes, the type structure used
to have members named tp_alloc and a tp_free slot already, counters for
the number of allocations and deallocations. These are renamed to
tp_allocs and tp_deallocs.)

Standard implementations for tp_alloc() and tp_new() are available.
PyType_GenericAlloc() allocates an object from the standard heap and
initializes it properly. It uses the above formula to determine the
amount of memory to allocate, and takes care of GC registration. The
only reason not to use this implementation would be to allocate objects
from a different heap (as is done by some very small frequently used
objects like ints and tuples). PyType_GenericNew() adds very little: it
just calls the type's tp_alloc() slot with zero for nitems. But for
mutable types that do all their initialization in their tp_init() slot,
this may be just the ticket.

Preparing a type for subtyping

The idea behind subtyping is very similar to that of single inheritance
in C++. A base type is described by a structure declaration (similar to
the C++ class declaration) plus a type object (similar to the C++
vtable). A derived type can extend the structure (but must leave the
names, order and type of the members of the base structure unchanged)
and can override certain slots in the type object, leaving others the
same. (Unlike C++ vtables, all Python type objects have the same memory
layout.)

The base type must do the following:

-   Add the flag value Py_TPFLAGS_BASETYPE to tp_flags.
-   Declare and use tp_new(), tp_alloc() and optional tp_init() slots.
-   Declare and use tp_dealloc() and tp_free().
-   Export its object structure declaration.
-   Export a subtyping-aware type-checking macro.

The requirements and signatures for tp_new(), tp_alloc() and tp_init()
have already been discussed above: tp_alloc() should allocate the memory
and initialize it to mostly zeros; tp_new() should call the tp_alloc()
slot and then proceed to do the minimally required initialization;
tp_init() should be used for more extensive initialization of mutable
objects.

It should come as no surprise that there are similar conventions at the
end of an object's lifetime. The slots involved are tp_dealloc()
(familiar to all who have ever implemented a Python extension type) and
tp_free(), the new kid on the block. (The names aren't quite symmetric;
tp_free() corresponds to tp_alloc(), which is fine, but tp_dealloc()
corresponds to tp_new(). Maybe the tp_dealloc slot should be renamed?)

The tp_free() slot should be used to free the memory and unregister the
object with the garbage collection subsystem, and can be overridden by a
derived class; tp_dealloc() should deinitialize the object (usually by
calling Py_XDECREF() for various sub-objects) and then call tp_free() to
deallocate the memory. The signature for tp_dealloc() is the same as it
always was:

    void tp_dealloc(PyObject *object)

The signature for tp_free() is the same:

    void tp_free(PyObject *object)

(In a previous version of this PEP, there was also a role reserved for
the tp_clear() slot. This turned out to be a bad idea.)

To be usefully subtyped in C, a type must export the structure
declaration for its instances through a header file, as it is needed to
derive a subtype. The type object for the base type must also be
exported.

If the base type has a type-checking macro (like PyDict_Check()), this
macro should be made to recognize subtypes. This can be done by using
the new PyObject_TypeCheck(object, type) macro, which calls a function
that follows the base class links.

The PyObject_TypeCheck() macro contains a slight optimization: it first
compares object->ob_type directly to the type argument, and if this is a
match, bypasses the function call. This should make it fast enough for
most situations.

Note that this change in the type-checking macro means that C functions
that require an instance of the base type may be invoked with instances
of the derived type. Before enabling subtyping of a particular type, its
code should be checked to make sure that this won't break anything. It
has proved useful in the prototype to add another type-checking macro
for the built-in Python object types, to check for exact type match too
(for example, PyDict_Check(x) is true if x is an instance of dictionary
or of a dictionary subclass, while PyDict_CheckExact(x) is true only if
x is a dictionary).

Creating a subtype of a built-in type in C

The simplest form of subtyping is subtyping in C. It is the simplest
form because we can require the C code to be aware of some of the
problems, and it's acceptable for C code that doesn't follow the rules
to dump core. For added simplicity, it is limited to single inheritance.

Let's assume we're deriving from a mutable base type whose tp_itemsize
is zero. The subtype code is not GC-aware, although it may inherit
GC-awareness from the base type (this is automatic). The base type's
allocation uses the standard heap.

The derived type begins by declaring a type structure which contains the
base type's structure. For example, here's the type structure for a
subtype of the built-in list type:

    typedef struct {
        PyListObject list;
        int state;
    } spamlistobject;

Note that the base type structure member (here PyListObject) must be the
first member of the structure; any following members are additions. Also
note that the base type is not referenced via a pointer; the actual
contents of its structure must be included! (The goal is for the memory
layout of the beginning of the subtype instance to be the same as that
of the base type instance.)

Next, the derived type must declare a type object and initialize it.
Most of the slots in the type object may be initialized to zero, which
is a signal that the base type slot must be copied into it. Some slots
that must be initialized properly:

-   The object header must be filled in as usual; the type should be
    &PyType_Type.
-   The tp_basicsize slot must be set to the size of the subtype
    instance struct (in the above example: sizeof(spamlistobject)).
-   The tp_base slot must be set to the address of the base type's type
    object.
-   If the derived slot defines any pointer members, the tp_dealloc slot
    function requires special attention, see below; otherwise, it can be
    set to zero, to inherit the base type's deallocation function.
-   The tp_flags slot must be set to the usual Py_TPFLAGS_DEFAULT value.
-   The tp_name slot must be set; it is recommended to set tp_doc as
    well (these are not inherited).

If the subtype defines no additional structure members (it only defines
new behavior, no new data), the tp_basicsize and the tp_dealloc slots
may be left set to zero.

The subtype's tp_dealloc slot deserves special attention. If the derived
type defines no additional pointer members that need to be DECREF'ed or
freed when the object is deallocated, it can be set to zero. Otherwise,
the subtype's tp_dealloc() function must call Py_XDECREF() for any
PyObject * members and the correct memory freeing function for any other
pointers it owns, and then call the base class's tp_dealloc() slot. This
call has to be made via the base type's type structure, for example,
when deriving from the standard list type:

    PyList_Type.tp_dealloc(self);

If the subtype wants to use a different allocation heap than the base
type, the subtype must override both the tp_alloc() and the tp_free()
slots. These will be called by the base class's tp_new() and
tp_dealloc() slots, respectively.

To complete the initialization of the type, PyType_InitDict() must be
called. This replaces slots initialized to zero in the subtype with the
value of the corresponding base type slots. (It also fills in tp_dict,
the type's dictionary, and does various other initializations necessary
for type objects.)

A subtype is not usable until PyType_InitDict() is called for it; this
is best done during module initialization, assuming the subtype belongs
to a module. An alternative for subtypes added to the Python core (which
don't live in a particular module) would be to initialize the subtype in
their constructor function. It is allowed to call PyType_InitDict() more
than once; the second and further calls have no effect. To avoid
unnecessary calls, a test for tp_dict==NULL can be made.

(During initialization of the Python interpreter, some types are
actually used before they are initialized. As long as the slots that are
actually needed are initialized, especially tp_dealloc, this works, but
it is fragile and not recommended as a general practice.)

To create a subtype instance, the subtype's tp_new() slot is called.
This should first call the base type's tp_new() slot and then initialize
the subtype's additional data members. To further initialize the
instance, the tp_init() slot is typically called. Note that the tp_new()
slot should not call the tp_init() slot; this is up to tp_new()'s caller
(typically a factory function). There are circumstances where it is
appropriate not to call tp_init().

If a subtype defines a tp_init() slot, the tp_init() slot should
normally first call the base type's tp_init() slot.

(XXX There should be a paragraph or two about argument passing here.)

Subtyping in Python

The next step is to allow subtyping of selected built-in types through a
class statement in Python. Limiting ourselves to single inheritance for
now, here is what happens for a simple class statement:

    class C(B):
        var1 = 1
        def method1(self): pass
        # etc.

The body of the class statement is executed in a fresh environment
(basically, a new dictionary used as local namespace), and then C is
created. The following explains how C is created.

Assume B is a type object. Since type objects are objects, and every
object has a type, B has a type. Since B is itself a type, we also call
its type its metatype. B's metatype is accessible via type(B) or
B.__class__ (the latter notation is new for types; it is introduced in
PEP 252). Let's say this metatype is M (for Metatype). The class
statement will create a new type, C. Since C will be a type object just
like B, we view the creation of C as an instantiation of the metatype,
M. The information that needs to be provided for the creation of a
subclass is:

-   its name (in this example the string "C");
-   its bases (a singleton tuple containing B);
-   the results of executing the class body, in the form of a dictionary
    (for example {"var1": 1, "method1": <functionmethod1 at ...>, ...}).

The class statement will result in the following call:

    C = M("C", (B,), dict)

where dict is the dictionary resulting from execution of the class body.
In other words, the metatype (M) is called.

Note that even though the example has only one base, we still pass in a
(singleton) sequence of bases; this makes the interface uniform with the
multiple-inheritance case.

In current Python, this is called the "Don Beaudry hook" after its
inventor; it is an exceptional case that is only invoked when a base
class is not a regular class. For a regular base class (or when no base
class is specified), current Python calls PyClass_New(), the C level
factory function for classes, directly.

Under the new system this is changed so that Python always determines a
metatype and calls it as given above. When one or more bases are given,
the type of the first base is used as the metatype; when no base is
given, a default metatype is chosen. By setting the default metatype to
PyClass_Type, the metatype of "classic" classes, the classic behavior of
the class statement is retained. This default can be changed per module
by setting the global variable __metaclass__.

There are two further refinements here. First, a useful feature is to be
able to specify a metatype directly. If the class suite defines a
variable __metaclass__, that is the metatype to call. (Note that setting
__metaclass__ at the module level only affects class statements without
a base class and without an explicit __metaclass__ declaration; but
setting __metaclass__ in a class suite overrides the default metatype
unconditionally.)

Second, with multiple bases, not all bases need to have the same
metatype. This is called a metaclass conflict[1]. Some metaclass
conflicts can be resolved by searching through the set of bases for a
metatype that derives from all other given metatypes. If such a metatype
cannot be found, an exception is raised and the class statement fails.

This conflict resolution can be implemented by the metatype
constructors: the class statement just calls the metatype of the first
base (or that specified by the __metaclass__ variable), and this
metatype's constructor looks for the most derived metatype. If that is
itself, it proceeds; otherwise, it calls that metatype's constructor.
(Ultimate flexibility: another metatype might choose to require that all
bases have the same metatype, or that there's only one base class, or
whatever.)

(In[2], a new metaclass is automatically derived that is a subclass of
all given metaclasses. But since it is questionable in Python how
conflicting method definitions of the various metaclasses should be
merged, I don't think this is feasible. Should the need arise, the user
can derive such a metaclass manually and specify it using the
__metaclass__ variable. It is also possible to have a new metaclass that
does this.)

Note that calling M requires that M itself has a type: the
meta-metatype. And the meta-metatype has a type, the meta-meta-metatype.
And so on. This is normally cut short at some level by making a metatype
be its own metatype. This is indeed what happens in Python: the ob_type
reference in PyType_Type is set to &PyType_Type. In the absence of third
party metatypes, PyType_Type is the only metatype in the Python
interpreter.

(In a previous version of this PEP, there was one additional meta-level,
and there was a meta-metatype called "turtle". This turned out to be
unnecessary.)

In any case, the work for creating C is done by M's tp_new() slot. It
allocates space for an "extended" type structure, containing: the type
object; the auxiliary structures (as_sequence etc.); the string object
containing the type name (to ensure that this object isn't deallocated
while the type object is still referencing it); and some auxiliary
storage (to be described later). It initializes this storage to zeros
except for a few crucial slots (for example, tp_name is set to point to
the type name) and then sets the tp_base slot to point to B. Then
PyType_InitDict() is called to inherit B's slots. Finally, C's tp_dict
slot is updated with the contents of the namespace dictionary (the third
argument to the call to M).

Multiple inheritance

The Python class statement supports multiple inheritance, and we will
also support multiple inheritance involving built-in types.

However, there are some restrictions. The C runtime architecture doesn't
make it feasible to have a meaningful subtype of two different built-in
types except in a few degenerate cases. Changing the C runtime to
support fully general multiple inheritance would be too much of an
upheaval of the code base.

The main problem with multiple inheritance from different built-in types
stems from the fact that the C implementation of built-in types accesses
structure members directly; the C compiler generates an offset relative
to the object pointer and that's that. For example, the list and
dictionary type structures each declare a number of different but
overlapping structure members. A C function accessing an object
expecting a list won't work when passed a dictionary, and vice versa,
and there's not much we could do about this without rewriting all code
that accesses lists and dictionaries. This would be too much work, so we
won't do this.

The problem with multiple inheritance is caused by conflicting structure
member allocations. Classes defined in Python normally don't store their
instance variables in structure members: they are stored in an instance
dictionary. This is the key to a partial solution. Suppose we have the
following two classes:

    class A(dictionary):
        def foo(self): pass

    class B(dictionary):
        def bar(self): pass

    class C(A, B): pass

(Here, 'dictionary' is the type of built-in dictionary objects, a.k.a.
type({}) or {}.__class__ or types.DictType.) If we look at the structure
layout, we find that an A instance has the layout of a dictionary
followed by the __dict__ pointer, and a B instance has the same layout;
since there are no structure member layout conflicts, this is okay.

Here's another example:

    class X(object):
        def foo(self): pass

    class Y(dictionary):
        def bar(self): pass

    class Z(X, Y): pass

(Here, 'object' is the base for all built-in types; its structure layout
only contains the ob_refcnt and ob_type members.) This example is more
complicated, because the __dict__ pointer for X instances has a
different offset than that for Y instances. Where is the __dict__
pointer for Z instances? The answer is that the offset for the __dict__
pointer is not hardcoded, it is stored in the type object.

Suppose on a particular machine an 'object' structure is 8 bytes long,
and a 'dictionary' struct is 60 bytes, and an object pointer is 4 bytes.
Then an X structure is 12 bytes (an object structure followed by a
__dict__ pointer), and a Y structure is 64 bytes (a dictionary structure
followed by a __dict__ pointer). The Z structure has the same layout as
the Y structure in this example. Each type object (X, Y and Z) has a
"__dict__ offset" which is used to find the __dict__ pointer. Thus, the
recipe for looking up an instance variable is:

1.  get the type of the instance
2.  get the __dict__ offset from the type object
3.  add the __dict__ offset to the instance pointer
4.  look in the resulting address to find a dictionary reference
5.  look up the instance variable name in that dictionary

Of course, this recipe can only be implemented in C, and I have left out
some details. But this allows us to use multiple inheritance patterns
similar to the ones we can use with classic classes.

XXX I should write up the complete algorithm here to determine base
class compatibility, but I can't be bothered right now. Look at
best_base() in typeobject.c in the implementation mentioned below.

MRO: Method resolution order (the lookup rule)

With multiple inheritance comes the question of method resolution order:
the order in which a class or type and its bases are searched looking
for a method of a given name.

In classic Python, the rule is given by the following recursive
function, also known as the left-to-right depth-first rule:

    def classic_lookup(cls, name):
        if cls.__dict__.has_key(name):
            return cls.__dict__[name]
        for base in cls.__bases__:
            try:
                return classic_lookup(base, name)
            except AttributeError:
                pass
        raise AttributeError, name

The problem with this becomes apparent when we consider a "diamond
diagram":

    class A:
      ^ ^  def save(self): ...
     /   \
    /     \
    /       \
    /         \
    class B     class C:
    ^         ^  def save(self): ...
    \       /
    \     /
     \   /
      \ /
    class D

Arrows point from a subtype to its base type(s). This particular diagram
means B and C derive from A, and D derives from B and C (and hence also,
indirectly, from A).

Assume that C overrides the method save(), which is defined in the base
A. (C.save() probably calls A.save() and then saves some of its own
state.) B and D don't override save(). When we invoke save() on a D
instance, which method is called? According to the classic lookup rule,
A.save() is called, ignoring C.save()!

This is not good. It probably breaks C (its state doesn't get saved),
defeating the whole purpose of inheriting from C in the first place.

Why was this not a problem in classic Python? Diamond diagrams are
rarely found in classic Python class hierarchies. Most class hierarchies
use single inheritance, and multiple inheritance is usually confined to
mix-in classes. In fact, the problem shown here is probably the reason
why multiple inheritance is unpopular in classic Python.

Why will this be a problem in the new system? The 'object' type at the
top of the type hierarchy defines a number of methods that can usefully
be extended by subtypes, for example __getattr__().

(Aside: in classic Python, the __getattr__() method is not really the
implementation for the get-attribute operation; it is a hook that only
gets invoked when an attribute cannot be found by normal means. This has
often been cited as a shortcoming -- some class designs have a
legitimate need for a __getattr__() method that gets called for all
attribute references. But then of course this method has to be able to
invoke the default implementation directly. The most natural way is to
make the default implementation available as
object.__getattr__(self, name).)

Thus, a classic class hierarchy like this:

    class B     class C:
        ^         ^  def __getattr__(self, name): ...
         \       /
          \     /
           \   /
            \ /
          class D

will change into a diamond diagram under the new system:

    object:
      ^ ^  __getattr__()
     /   \
    /     \
    /       \
    /         \
    class B     class C:
    ^         ^  def __getattr__(self, name): ...
    \       /
    \     /
     \   /
      \ /
    class D

and while in the original diagram C.__getattr__() is invoked, under the
new system with the classic lookup rule, object.__getattr__() would be
invoked!

Fortunately, there's a lookup rule that's better. It's a bit difficult
to explain, but it does the right thing in the diamond diagram, and it
is the same as the classic lookup rule when there are no diamonds in the
inheritance graph (when it is a tree).

The new lookup rule constructs a list of all classes in the inheritance
diagram in the order in which they will be searched. This construction
is done at class definition time to save time. To explain the new lookup
rule, let's first consider what such a list would look like for the
classic lookup rule. Note that in the presence of diamonds the classic
lookup visits some classes multiple times. For example, in the ABCD
diamond diagram above, the classic lookup rule visits the classes in
this order:

    D, B, A, C, A

Note how A occurs twice in the list. The second occurrence is redundant,
since anything that could be found there would already have been found
when searching the first occurrence.

We use this observation to explain our new lookup rule. Using the
classic lookup rule, construct the list of classes that would be
searched, including duplicates. Now for each class that occurs in the
list multiple times, remove all occurrences except for the last. The
resulting list contains each ancestor class exactly once (including the
most derived class, D in the example).

Searching for methods in this order will do the right thing for the
diamond diagram. Because of the way the list is constructed, it does not
change the search order in situations where no diamond is involved.

Isn't this backwards incompatible? Won't it break existing code? It
would, if we changed the method resolution order for all classes.
However, in Python 2.2, the new lookup rule will only be applied to
types derived from built-in types, which is a new feature. Class
statements without a base class create "classic classes", and so do
class statements whose base classes are themselves classic classes. For
classic classes the classic lookup rule will be used. (To experiment
with the new lookup rule for classic classes, you will be able to
specify a different metaclass explicitly.) We'll also provide a tool
that analyzes a class hierarchy looking for methods that would be
affected by a change in method resolution order.

XXX Another way to explain the motivation for the new MRO, due to Damian
Conway: you never use the method defined in a base class if it is
defined in a derived class that you haven't explored yet (using the old
search order).

XXX To be done

Additional topics to be discussed in this PEP:

-   backwards compatibility issues!!!
-   class methods and static methods
-   cooperative methods and super()
-   mapping between type object slots (tp_foo) and special methods
    (__foo__) (actually, this may belong in PEP 252)
-   built-in names for built-in types (object, int, str, list etc.)
-   __dict__ and __dictoffset__
-   __slots__
-   the HEAPTYPE flag bit
-   GC support
-   API docs for all the new functions
-   how to use __new__
-   writing metaclasses (using mro() etc.)
-   high level user overview

open issues

-   do we need __del__?
-   assignment to __dict__, __bases__
-   inconsistent naming (e.g.
    tp_dealloc/tp_new/tp_init/tp_alloc/tp_free)
-   add builtin alias 'dict' for 'dictionary'?
-   when subclasses of dict/list etc. are passed to system functions,
    the __getitem__ overrides (etc.) aren't always used

Implementation

A prototype implementation of this PEP (and for PEP 252) is available
from CVS, and in the series of Python 2.2 alpha and beta releases. For
some examples of the features described here, see the file
Lib/test/test_descr.py and the extension module Modules/xxsubtype.c.

References

Copyright

This document has been placed in the public domain.

[1] "Putting Metaclasses to Work", by Ira R. Forman and Scott H.
Danforth, Addison-Wesley 1999.
(http://www.aw.com/product/0,2627,0201433052,00.html)

[2] "Putting Metaclasses to Work", by Ira R. Forman and Scott H.
Danforth, Addison-Wesley 1999.
(http://www.aw.com/product/0,2627,0201433052,00.html)