python-peps/pep-0359.txt

PEP: 359
Title: The "make" Statement
Version: $Revision$
Last-Modified: $Date$
Author: Steven Bethard <steven.bethard@gmail.com>
Status: Withdrawn
Type: Standards Track
Content-Type: text/x-rst
Created: 05-Apr-2006
Python-Version: 2.6
Post-History: 05-Apr-2006, 06-Apr-2006, 13-Apr-2006


Abstract
========

This PEP proposes a generalization of the class-declaration syntax,
the ``make`` statement.  The proposed syntax and semantics parallel
the syntax for class definition, and so::

   make <callable> <name> <tuple>:
       <block>

is translated into the assignment::

   <name> = <callable>("<name>", <tuple>, <namespace>)

where ``<namespace>`` is the dict created by executing ``<block>``.
This is mostly syntactic sugar for::

   class <name> <tuple>:
       __metaclass__ = <callable>
       <block>

and is intended to help more clearly express the intent of the
statement when something other than a class is being created.  Of
course, other syntax for such a statement is possible, but it is hoped
that by keeping a strong parallel to the class statement, an
understanding of how classes and metaclasses work will translate into
an understanding of how the make-statement works as well.

The PEP is based on a suggestion [1]_ from Michele Simionato on the
python-dev list.


Withdrawal Notice
=================

This PEP was withdrawn at Guido's request [2]_. Guido didn't like it,
and in particular didn't like how the property use-case puts the
instance methods of a property at a different level than other
instance methods and requires fixed names for the property functions.


Motivation
==========

Class statements provide two nice facilities to Python:

(1) They execute a block of statements and provide the resulting
    bindings as a dict to the metaclass.

(2) They encourage DRY (don't repeat yourself) by allowing the class
    being created to know the name it is being assigned.

Thus in a simple class statement like::

     class C(object):
         x = 1
         def foo(self):
             return 'bar'

the metaclass (``type``) gets called with something like::

    C = type('C', (object,), {'x':1, 'foo':<function foo at ...>})

The class statement is just syntactic sugar for the above assignment
statement, but clearly a very useful sort of syntactic sugar.  It
avoids not only the repetition of ``C``, but also simplifies the
creation of the dict by allowing it to be expressed as a series of
statements.

Historically, type instances (a.k.a. class objects) have been the
only objects blessed with this sort of syntactic support.  The make
statement aims to extend this support to other sorts of objects where
such syntax would also be useful.


Example: simple namespaces
--------------------------

Let's say I have some attributes in a module that I access like::

    mod.thematic_roletype
    mod.opinion_roletype

    mod.text_format
    mod.html_format

and since "Namespaces are one honking great idea", I'd like to be able
to access these attributes instead as::

    mod.roletypes.thematic
    mod.roletypes.opinion

    mod.format.text
    mod.format.html

I currently have two main options:

(1) Turn the module into a package, turn ``roletypes`` and ``format``
    into submodules, and move the attributes to the submodules.

(2) Create ``roletypes`` and ``format`` classes, and move the
    attributes to the classes.

The former is a fair chunk of refactoring work, and produces two tiny
modules without much content.  The latter keeps the attributes local
to the module, but creates classes when there is no intention of ever
creating instances of those classes.

In situations like this, it would be nice to simply be able to declare
a "namespace" to hold the few attributes.  With the new make
statement, I could introduce my new namespaces with something like::

    make namespace roletypes:
        thematic = ...
        opinion = ...

    make namespace format:
        text = ...
        html = ...

and keep my attributes local to the module without making classes that
are never intended to be instantiated.  One definition of namespace
that would make this work is::

    class namespace(object):
        def __init__(self, name, args, kwargs):
            self.__dict__.update(kwargs)

Given this definition, at the end of the make-statements above,
``roletypes`` and ``format`` would be namespace instances.


Example: GUI objects
--------------------

In GUI toolkits, objects like frames and panels are often associated
with attributes and functions.  With the make-statement, code that
looks something like::

    root = Tkinter.Tk()
    frame = Tkinter.Frame(root)
    frame.pack()
    def say_hi():
        print "hi there, everyone!"
    hi_there = Tkinter.Button(frame, text="Hello", command=say_hi)
    hi_there.pack(side=Tkinter.LEFT)
    root.mainloop()

could be rewritten to group the Button's function with its
declaration::

    root = Tkinter.Tk()
    frame = Tkinter.Frame(root)
    frame.pack()
    make Tkinter.Button hi_there(frame):
        text = "Hello"
        def command():
            print "hi there, everyone!"
    hi_there.pack(side=Tkinter.LEFT)
    root.mainloop()


Example: custom descriptors
---------------------------

Since descriptors are used to customize access to an attribute, it's
often useful to know the name of that attribute.  Current Python
doesn't give an easy way to find this name and so a lot of custom
descriptors, like Ian Bicking's setonce descriptor [3]_, have to hack
around this somehow.  With the make-statement, you could create a
``setonce`` attribute like::

    class A(object):
        ...
        make setonce x:
            "A's x attribute"
        ...

where the ``setonce`` descriptor would be defined like::

    class setonce(object):

        def __init__(self, name, args, kwargs):
            self._name = '_setonce_attr_%s' % name
            self.__doc__ = kwargs.pop('__doc__', None)

        def __get__(self, obj, type=None):
            if obj is None:
                return self
            return getattr(obj, self._name)

        def __set__(self, obj, value):
            try:
                getattr(obj, self._name)
            except AttributeError:
                setattr(obj, self._name, value)
            else:
                raise AttributeError("Attribute already set")

        def set(self, obj, value):
            setattr(obj, self._name, value)

        def __delete__(self, obj):
            delattr(obj, self._name)

Note that unlike the original implementation, the private attribute
name is stable since it uses the name of the descriptor, and therefore
instances of class A are pickleable.


Example: property namespaces
----------------------------

Python's property type takes three function arguments and a docstring
argument which, though relevant only to the property, must be declared
before it and then passed as arguments to the property call, e.g.::

    class C(object):
        ...
        def get_x(self):
            ...
        def set_x(self):
            ...
        x = property(get_x, set_x, "the x of the frobulation")

This issue has been brought up before, and Guido [4]_ and others [5]_
have briefly mused over alternate property syntaxes to make declaring
properties easier.  With the make-statement, the following syntax
could be supported::

    class C(object):
        ...
        make block_property x:
            '''The x of the frobulation'''
            def fget(self):
                ...
            def fset(self):
                ...

with the following definition of ``block_property``::

    def block_property(name, args, block_dict):
        fget = block_dict.pop('fget', None)
        fset = block_dict.pop('fset', None)
        fdel = block_dict.pop('fdel', None)
        doc = block_dict.pop('__doc__', None)
        assert not block_dict
        return property(fget, fset, fdel, doc)


Example: interfaces
-------------------

Guido [6]_ and others have occasionally suggested introducing
interfaces into python.  Most suggestions have offered syntax along
the lines of::

    interface IFoo:
        """Foo blah blah"""

        def fumble(name, count):
            """docstring"""

but since there is currently no way in Python to declare an interface
in this manner, most implementations of Python interfaces use class
objects instead, e.g. Zope's::

    class IFoo(Interface):
        """Foo blah blah"""

        def fumble(name, count):
            """docstring"""

With the new make-statement, these interfaces could instead be
declared as::

    make Interface IFoo:
        """Foo blah blah"""

        def fumble(name, count):
            """docstring"""

which makes the intent (that this is an interface, not a class) much
clearer.


Specification
=============

Python will translate a make-statement::

    make <callable> <name> <tuple>:
        <block>

into the assignment::

    <name> = <callable>("<name>", <tuple>, <namespace>)

where ``<namespace>`` is the dict created by executing ``<block>``.
The ``<tuple>`` expression is optional; if not present, an empty tuple
will be assumed.

A patch is available implementing these semantics [7]_.

The make-statement introduces a new keyword, ``make``.  Thus in Python
2.6, the make-statement will have to be enabled using ``from
__future__ import make_statement``.


Open Issues
===========

Keyword
-------

Does the ``make`` keyword break too much code?  Originally, the make
statement used the keyword ``create`` (a suggestion due to Nick
Coghlan).  However, investigations into the standard library [8]_ and
Zope+Plone code [9]_ revealed that ``create`` would break a lot more
code, so ``make`` was adopted as the keyword instead.  However, there
are still a few instances where ``make`` would break code.  Is there a
better keyword for the statement?

Some possible keywords and their counts in the standard library (plus
some installed packages):

* make - 2 (both in tests)
* create - 19 (including existing function in imaplib)
* build - 83 (including existing class in distutils.command.build)
* construct - 0
* produce - 0


The make-statement as an alternate constructor
----------------------------------------------

Currently, there are not many functions which have the signature
``(name, args, kwargs)``.  That means that something like::

    make dict params:
        x = 1
        y = 2

is currently impossible because the dict constructor has a different
signature.  Does this sort of thing need to be supported?  One
suggestion, by Carl Banks, would be to add a ``__make__`` magic method
that if found would be called instead of ``__call__``.  For types,
the ``__make__`` method would be identical to ``__call__`` and thus
unnecessary, but dicts could support the make-statement by defining a
``__make__`` method on the dict type that looks something like::

    def __make__(cls, name, args, kwargs):
        return cls(**kwargs)

Of course, rather than adding another magic method, the dict type
could just grow a classmethod something like ``dict.fromblock`` that
could be used like::

    make dict.fromblock params:
        x = 1
        y = 2

So the question is, will many types want to use the make-statement as
an alternate constructor?  And if so, does that alternate constructor
need to have the same name as the original constructor?


Customizing the dict in which the block is executed
---------------------------------------------------

Should users of the make-statement be able to determine in which dict
object the code is executed?  This would allow the make-statement to
be used in situations where a normal dict object would not suffice,
e.g. if order and repeated names must be allowed.  Allowing this sort
of customization could allow XML to be written without repeating
element names, and with nesting of make-statements corresponding to
nesting of XML elements::

    make Element html:
        make Element body:
            text('before first h1')
            make Element h1:
                attrib(style='first')
                text('first h1')
                tail('after first h1')
            make Element h1:
                attrib(style='second')
                text('second h1')
                tail('after second h1')

If the make-statement tried to get the dict in which to execute its
block by calling the callable's ``__make_dict__`` method, the
following code would allow the make-statement to be used as above::

    class Element(object):

        class __make_dict__(dict):

            def __init__(self, *args, **kwargs):
                self._super = super(Element.__make_dict__, self)
                self._super.__init__(*args, **kwargs)
                self.elements = []
                self.text = None
                self.tail = None
                self.attrib = {}

            def __getitem__(self, name):
                try:
                    return self._super.__getitem__(name)
                except KeyError:
                    if name in ['attrib', 'text', 'tail']:
                        return getattr(self, 'set_%s' % name)
                    else:
                        return globals()[name]

            def __setitem__(self, name, value):
                self._super.__setitem__(name, value)
                self.elements.append(value)

            def set_attrib(self, **kwargs):
                self.attrib = kwargs

            def set_text(self, text):
                self.text = text

            def set_tail(self, text):
                self.tail = text

        def __new__(cls, name, args, edict):
            get_element = etree.ElementTree.Element
            result = get_element(name, attrib=edict.attrib)
            result.text = edict.text
            result.tail = edict.tail
            for element in edict.elements:
                result.append(element)
            return result

Note, however, that the code to support this is somewhat fragile --
it has to magically populate the namespace with ``attrib``, ``text``
and ``tail``, and it assumes that every name binding inside the make
statement body is creating an Element.  As it stands, this code would
break with the introduction of a simple for-loop to any one of the
make-statement bodies, because the for-loop would bind a name to a
non-Element object.  This could be worked around by adding some sort
of isinstance check or attribute examination, but this still results
in a somewhat fragile solution.

It has also been pointed out that the with-statement can provide
equivalent nesting with a much more explicit syntax::

    with Element('html') as html:
        with Element('body') as body:
            body.text = 'before first h1'
            with Element('h1', style='first') as h1:
                h1.text = 'first h1'
                h1.tail = 'after first h1'
            with Element('h1', style='second') as h1:
                h1.text = 'second h1'
                h1.tail = 'after second h1'

And if the repetition of the element names here is too much of a DRY
violoation, it is also possible to eliminate all as-clauses except for
the first by adding a few methods to Element. [10]_

So are there real use-cases for executing the block in a dict of a
different type?  And if so, should the make-statement be extended to
support them?


Optional Extensions
===================

Remove the make keyword
-------------------------

It might be possible to remove the make keyword so that such
statements would begin with the callable being called, e.g.::

    namespace ns:
        badger = 42
        def spam():
            ...

    interface C(...):
        ...

However, almost all other Python statements begin with a keyword, and
removing the keyword would make it harder to look up this construct in
the documentation.  Additionally, this would add some complexity in
the grammar and so far I (Steven Bethard) have not been able to
implement the feature without the keyword.


Removing __metaclass__ in Python 3000
-------------------------------------

As a side-effect of its generality, the make-statement mostly
eliminates the need for the ``__metaclass__`` attribute in class
objects.  Thus in Python 3000, instead of::

   class <name> <bases-tuple>:
       __metaclass__ = <metaclass>
       <block>

metaclasses could be supported by using the metaclass as the callable
in a make-statement::

   make <metaclass> <name> <bases-tuple>:
       <block>

Removing the ``__metaclass__`` hook would simplify the BUILD_CLASS
opcode a bit.


Removing class statements in Python 3000
----------------------------------------

In the most extreme application of make-statements, the class
statement itself could be deprecated in favor of ``make type``
statements.


References
==========

.. [1] Michele Simionato's original suggestion
   (http://mail.python.org/pipermail/python-dev/2005-October/057435.html)

.. [2] Guido requests withdrawal
   (http://mail.python.org/pipermail/python-3000/2006-April/000936.html)

.. [3] Ian Bicking's setonce descriptor
   (http://blog.ianbicking.org/easy-readonly-attributes.html)

.. [4] Guido ponders property syntax
   (http://mail.python.org/pipermail/python-dev/2005-October/057404.html)

.. [5] Namespace-based property recipe
   (http://aspn.activestate.com/ASPN/Cookbook/Python/Recipe/442418)

.. [6] Python interfaces
   (http://www.artima.com/weblogs/viewpost.jsp?thread=86641)

.. [7] Make Statement patch
   (http://ucsu.colorado.edu/~bethard/py/make_statement.patch)

.. [8] Instances of create in the stdlib
   (http://mail.python.org/pipermail/python-list/2006-April/335159.html)

.. [9] Instances of create in Zope+Plone
   (http://mail.python.org/pipermail/python-list/2006-April/335284.html)

.. [10] Eliminate as-clauses in with-statement XML
   (http://mail.python.org/pipermail/python-list/2006-April/336774.html)


Copyright
=========

This document has been placed in the public domain.


..
   Local Variables:
   mode: indented-text
   indent-tabs-mode: nil
   sentence-end-double-space: t
   fill-column: 70
   coding: utf-8
   End: