python-peps/peps/pep-0346.rst

PEP: 346
Title: User Defined ("``with``") Statements
Version: $Revision$
Last-Modified: $Date$
Author: Alyssa Coghlan <ncoghlan@gmail.com>
Status: Withdrawn
Type: Standards Track
Content-Type: text/x-rst
Created: 06-May-2005
Python-Version: 2.5
Post-History:


Abstract
========

This PEP is a combination of :pep:`310`'s "Reliable Acquisition/Release
Pairs" with the "Anonymous Block Statements" of Guido's :pep:`340`.  This
PEP aims to take the good parts of :pep:`340`, blend them with parts of
:pep:`310` and rearrange the lot into an elegant whole.  It borrows from
various other PEPs in order to paint a complete picture, and is
intended to stand on its own.


Author's Note
=============

During the discussion of :pep:`340`, I maintained drafts of this PEP as
PEP 3XX on my own website (since I didn't have CVS access to update a
submitted PEP fast enough to track the activity on python-dev).

Since the first draft of this PEP, Guido wrote :pep:`343` as a simplified
version of :pep:`340`.  :pep:`343` (at the time of writing) uses the exact
same semantics for the new statements as this PEP, but uses a slightly
different mechanism to allow generators to be used to write statement
templates.  However, Guido has indicated that he intends to accept a
new PEP being written by Raymond Hettinger that will integrate :pep:`288`
and :pep:`325`, and will permit a generator decorator like the one
described in this PEP to be used to write statement templates for PEP
343. The other difference was the choice of keyword ('with' versus
'do') and Guido has stated he will organise a vote on that in the
context of :pep:`343`.

Accordingly, the version of this PEP submitted for archiving on
python.org is to be WITHDRAWN immediately after submission.  :pep:`343`
and the combined generator enhancement PEP will cover the important
ideas.


Introduction
============

This PEP proposes that Python's ability to reliably manage resources
be enhanced by the introduction of a new ``with`` statement that
allows factoring out of arbitrary ``try``/``finally`` and some
``try``/``except``/``else`` boilerplate.  The new construct is called
a 'user defined statement', and the associated class definitions are
called 'statement templates'.

The above is the main point of the PEP.  However, if that was all it
said, then :pep:`310` would be sufficient and this PEP would be
essentially redundant. Instead, this PEP recommends additional
enhancements that make it natural to write these statement templates
using appropriately decorated generators.  A side effect of those
enhancements is that it becomes important to appropriately deal
with the management of resources inside generators.

This is quite similar to :pep:`343`, but the exceptions that occur are
re-raised inside the generators frame, and the issue of generator
finalisation needs to be addressed as a result.  The template
generator decorator suggested by this PEP also creates reusable
templates, rather than the single use templates of :pep:`340`.

In comparison to :pep:`340`, this PEP eliminates the ability to suppress
exceptions, and makes the user defined statement a non-looping
construct.  The other main difference is the use of a decorator to
turn generators into statement templates, and the incorporation of
ideas for addressing iterator finalisation.

If all that seems like an ambitious operation. . . well, Guido was the
one to set the bar that high when he wrote :pep:`340` :)


Relationship with other PEPs
============================

This PEP competes directly with :pep:`310`, :pep:`340` and :pep:`343`,
as those PEPs all describe alternative mechanisms for handling
deterministic resource management.

It does not compete with :pep:`342` which splits off :pep:`340`'s
enhancements related to passing data into iterators.  The associated
changes to the ``for`` loop semantics would be combined with the
iterator finalisation changes suggested in this PEP.  User defined
statements would not be affected.

Neither does this PEP compete with the generator enhancements
described in :pep:`288`.  While this PEP proposes the ability to
inject exceptions into generator frames, it is an internal
implementation detail, and does not require making that ability
publicly available to Python code.  :pep:`288` is, in part, about
making that implementation detail easily accessible.

This PEP would, however, make the generator resource release support
described in :pep:`325` redundant - iterators which require
finalisation should provide an appropriate implementation of the
statement template protocol.


User defined statements
=======================

To steal the motivating example from :pep:`310`, correct handling of a
synchronisation lock currently looks like this::

    the_lock.acquire()
    try:
        # Code here executes with the lock held
    finally:
        the_lock.release()

Like :pep:`310`, this PEP proposes that such code be able to be written
as::

    with the_lock:
        # Code here executes with the lock held

These user defined statements are primarily designed to allow easy
factoring of ``try`` blocks that are not easily converted to
functions.  This is most commonly the case when the exception handling
pattern is consistent, but the body of the ``try`` block changes.
With a user-defined statement, it is straightforward to factor out the
exception handling into a statement template, with the body of the
``try`` clause provided inline in the user code.

The term 'user defined statement' reflects the fact that the meaning
of a ``with`` statement is governed primarily by the statement
template used, and programmers are free to create their own statement
templates, just as they are free to create their own iterators for use
in ``for`` loops.


Usage syntax for user defined statements
----------------------------------------

The proposed syntax is simple::

    with EXPR1 [as VAR1]:
        BLOCK1


Semantics for user defined statements
-------------------------------------

::

    the_stmt = EXPR1
    stmt_enter = getattr(the_stmt, "__enter__", None)
    stmt_exit = getattr(the_stmt, "__exit__", None)
    if stmt_enter is None or stmt_exit is None:
        raise TypeError("Statement template required")

    VAR1 = stmt_enter() # Omit 'VAR1 =' if no 'as' clause
    exc = (None, None, None)
    try:
        try:
            BLOCK1
        except:
            exc = sys.exc_info()
            raise
    finally:
        stmt_exit(*exc)


Other than ``VAR1``, none of the local variables shown above will be
visible to user code.  Like the iteration variable in a ``for`` loop,
``VAR1`` is visible in both ``BLOCK1`` and code following the user
defined statement.

Note that the statement template can only react to exceptions, it
cannot suppress them.  See `Rejected Options`_ for an explanation as
to why.


Statement template protocol: ``__enter__``
------------------------------------------

The ``__enter__()`` method takes no arguments, and if it raises an
exception, ``BLOCK1`` is never executed.  If this happens, the
``__exit__()`` method is not called.  The value returned by this
method is assigned to VAR1 if the ``as`` clause is used.  Object's
with no other value to return should generally return ``self`` rather
than ``None`` to permit in-place creation in the ``with`` statement.

Statement templates should use this method to set up the conditions
that are to exist during execution of the statement (e.g. acquisition
of a synchronisation lock).

Statement templates which are not always usable (e.g. closed file
objects) should raise a ``RuntimeError`` if an attempt is made to call
``__enter__()`` when the template is not in a valid state.


Statement template protocol: ``__exit__``
-----------------------------------------

The ``__exit__()`` method accepts three arguments which correspond to
the three "arguments" to the ``raise`` statement: type, value, and
traceback.  All arguments are always supplied, and will be set to
``None`` if no exception occurred.  This method will be called exactly
once by the ``with`` statement machinery if the ``__enter__()`` method
completes successfully.

Statement templates perform their exception handling in this method.
If the first argument is ``None``, it indicates non-exceptional
completion of ``BLOCK1`` - execution either reached the end of block,
or early completion was forced using a ``return``, ``break`` or
``continue`` statement.  Otherwise, the three arguments reflect the
exception that terminated ``BLOCK1``.

Any exceptions raised by the ``__exit__()`` method are propagated to
the scope containing the ``with`` statement.  If the user code in
``BLOCK1`` also raised an exception, that exception would be lost, and
replaced by the one raised by the ``__exit__()`` method.


Factoring out arbitrary exception handling
------------------------------------------

Consider the following exception handling arrangement::

    SETUP_BLOCK
    try:
        try:
            TRY_BLOCK
        except exc_type1, exc:
            EXCEPT_BLOCK1
        except exc_type2, exc:
            EXCEPT_BLOCK2
        except:
            EXCEPT_BLOCK3
        else:
            ELSE_BLOCK
    finally:
        FINALLY_BLOCK

It can be roughly translated to a statement template as follows::

    class my_template(object):

        def __init__(self, *args):
            # Any required arguments (e.g. a file name)
            # get stored in member variables
            # The various BLOCK's will need updating to reflect
            # that.

        def __enter__(self):
            SETUP_BLOCK

        def __exit__(self, exc_type, value, traceback):
            try:
                try:
                    if exc_type is not None:
                        raise exc_type, value, traceback
                except exc_type1, exc:
                    EXCEPT_BLOCK1
                except exc_type2, exc:
                    EXCEPT_BLOCK2
                except:
                    EXCEPT_BLOCK3
                else:
                    ELSE_BLOCK
            finally:
                FINALLY_BLOCK

Which can then be used as::

    with my_template(*args):
        TRY_BLOCK

However, there are two important semantic differences between this
code and the original ``try`` statement.

Firstly, in the original ``try`` statement, if a ``break``, ``return``
or ``continue`` statement is encountered in ``TRY_BLOCK``, only
``FINALLY_BLOCK`` will be executed as the statement completes.  With
the statement template, ``ELSE_BLOCK`` will also execute, as these
statements are treated like any other non-exceptional block
termination.  For use cases where it matters, this is likely to be a
good thing (see ``transaction`` in the Examples_), as this hole where
neither the ``except`` nor the ``else`` clause gets executed is easy
to forget when writing exception handlers.

Secondly, the statement template will not suppress any exceptions.
If, for example, the original code suppressed the ``exc_type1`` and
``exc_type2`` exceptions, then this would still need to be done inline
in the user code::

    try:
        with my_template(*args):
            TRY_BLOCK
    except (exc_type1, exc_type2):
        pass

However, even in these cases where the suppression of exceptions needs
to be made explicit, the amount of boilerplate repeated at the calling
site is significantly reduced (See `Rejected Options`_ for further
discussion of this behaviour).

In general, not all of the clauses will be needed.  For resource
handling (like files or synchronisation locks), it is possible to
simply execute the code that would have been part of ``FINALLY_BLOCK``
in the ``__exit__()`` method.  This can be seen in the following
implementation that makes synchronisation locks into statement
templates as mentioned at the beginning of this section::

    # New methods of synchronisation lock objects

    def __enter__(self):
        self.acquire()
        return self

    def __exit__(self, *exc_info):
        self.release()


Generators
==========

With their ability to suspend execution, and return control to the
calling frame, generators are natural candidates for writing statement
templates.  Adding user defined statements to the language does *not*
require the generator changes described in this section, thus making
this PEP an obvious candidate for a phased implementation (``with``
statements in phase 1, generator integration in phase 2).  The
suggested generator updates allow arbitrary exception handling to
be factored out like this::

    @statement_template
    def my_template(*arguments):
        SETUP_BLOCK
        try:
            try:
                yield
            except exc_type1, exc:
                EXCEPT_BLOCK1
            except exc_type2, exc:
                EXCEPT_BLOCK2
            except:
                EXCEPT_BLOCK3
            else:
                ELSE_BLOCK
        finally:
            FINALLY_BLOCK

Notice that, unlike the class based version, none of the blocks need
to be modified, as shared values are local variables of the
generator's internal frame, including the arguments passed in by the
invoking code.  The semantic differences noted earlier (all
non-exceptional block termination triggers the ``else`` clause, and
the template is unable to suppress exceptions) still apply.


Default value for ``yield``
---------------------------

When creating a statement template with a generator, the ``yield``
statement will often be used solely to return control to the body of
the user defined statement, rather than to return a useful value.

Accordingly, if this PEP is accepted, ``yield``, like ``return``, will
supply a default value of ``None`` (i.e. ``yield`` and ``yield None``
will become equivalent statements).

This same change is being suggested in :pep:`342`.  Obviously, it would
only need to be implemented once if both PEPs were accepted :)


Template generator decorator: ``statement_template``
----------------------------------------------------

As with :pep:`343`, a new decorator is suggested that wraps a generator
in an object with the appropriate statement template semantics.
Unlike :pep:`343`, the templates suggested here are reusable, as the
generator is instantiated anew in each call to ``__enter__()``.
Additionally, any exceptions that occur in ``BLOCK1`` are re-raised in
the generator's internal frame::

    class template_generator_wrapper(object):

        def __init__(self, func, func_args, func_kwds):
             self.func = func
             self.args = func_args
             self.kwds = func_kwds
             self.gen = None

        def __enter__(self):
            if self.gen is not None:
                raise RuntimeError("Enter called without exit!")
            self.gen = self.func(*self.args, **self.kwds)
            try:
                return self.gen.next()
            except StopIteration:
                raise RuntimeError("Generator didn't yield")

        def __exit__(self, *exc_info):
            if self.gen is None:
                raise RuntimeError("Exit called without enter!")
            try:
                try:
                    if exc_info[0] is not None:
                        self.gen._inject_exception(*exc_info)
                    else:
                        self.gen.next()
                except StopIteration:
                    pass
                else:
                    raise RuntimeError("Generator didn't stop")
            finally:
                self.gen = None

    def statement_template(func):
        def factory(*args, **kwds):
            return template_generator_wrapper(func, args, kwds)
        return factory


Template generator wrapper: ``__enter__()`` method
--------------------------------------------------

The template generator wrapper has an ``__enter__()`` method that
creates a new instance of the contained generator, and then invokes
``next()`` once.  It will raise a ``RuntimeError`` if the last
generator instance has not been cleaned up, or if the generator
terminates instead of yielding a value.


Template generator wrapper: ``__exit__()`` method
-------------------------------------------------

The template generator wrapper has an ``__exit__()`` method that
simply invokes ``next()`` on the generator if no exception is passed
in.  If an exception is passed in, it is re-raised in the contained
generator at the point of the last ``yield`` statement.

In either case, the generator wrapper will raise a RuntimeError if the
internal frame does not terminate as a result of the operation.  The
``__exit__()`` method will always clean up the reference to the used
generator instance, permitting ``__enter__()`` to be called again.

A ``StopIteration`` raised by the body of the user defined statement
may be inadvertently suppressed inside the ``__exit__()`` method, but
this is unimportant, as the originally raised exception still
propagates correctly.


Injecting exceptions into generators
------------------------------------

To implement the ``__exit__()`` method of the template generator
wrapper, it is necessary to inject exceptions into the internal frame
of the generator.  This is new implementation level behaviour that has
no current Python equivalent.

The injection mechanism (referred to as ``_inject_exception`` in this
PEP) raises an exception in the generator's frame with the specified
type, value and traceback information.  This means that the exception
looks like the original if it is allowed to propagate.

For the purposes of this PEP, there is no need to make this capability
available outside the Python implementation code.


Generator finalisation
----------------------

To support resource management in template generators, this PEP will
eliminate the restriction on ``yield`` statements inside the ``try``
block of a ``try``/``finally`` statement.  Accordingly, generators
which require the use of a file or some such object can ensure the
object is managed correctly through the use of ``try``/``finally`` or
``with`` statements.

This restriction will likely need to be lifted globally - it would be
difficult to restrict it so that it was only permitted inside
generators used to define statement templates.  Accordingly, this PEP
includes suggestions designed to ensure generators which are not used
as statement templates are still finalised appropriately.


Generator finalisation: ``TerminateIteration`` exception
--------------------------------------------------------

A new exception is proposed::

    class TerminateIteration(Exception): pass

The new exception is injected into a generator in order to request
finalisation.  It should not be suppressed by well-behaved code.


Generator finalisation: ``__del__()`` method
--------------------------------------------

To ensure a generator is finalised eventually (within the limits of
Python's garbage collection), generators will acquire a ``__del__()``
method with the following semantics::

    def __del__(self):
        try:
            self._inject_exception(TerminateIteration, None, None)
        except TerminateIteration:
            pass


Deterministic generator finalisation
------------------------------------

There is a simple way to provide deterministic finalisation of
generators - give them appropriate ``__enter__()`` and ``__exit__()``
methods::

    def __enter__(self):
        return self

    def __exit__(self, *exc_info):
        try:
            self._inject_exception(TerminateIteration, None, None)
        except TerminateIteration:
            pass

Then any generator can be finalised promptly by wrapping the relevant
``for`` loop inside a ``with`` statement::

    with all_lines(filenames) as lines:
        for line in lines:
            print lines

(See the Examples_ for the definition of ``all_lines``, and the reason
it requires prompt finalisation)

Compare the above example to the usage of file objects::

    with open(filename) as f:
        for line in f:
            print f


Generators as user defined statement templates
----------------------------------------------

When used to implement a user defined statement, a generator should
yield only once on a given control path.  The result of that yield
will then be provided as the result of the generator's ``__enter__()``
method.  Having a single ``yield`` on each control path ensures that
the internal frame will terminate when the generator's ``__exit__()``
method is called.  Multiple ``yield`` statements on a single control
path will result in a ``RuntimeError`` being raised by the
``__exit__()`` method when the internal frame fails to terminate
correctly.  Such an error indicates a bug in the statement template.

To respond to exceptions, or to clean up resources, it is sufficient
to wrap the ``yield`` statement in an appropriately constructed
``try`` statement.  If execution resumes after the ``yield`` without
an exception, the generator knows that the body of the ``do``
statement completed without incident.


Examples
========

1. A template for ensuring that a lock, acquired at the start of a
   block, is released when the block is left::

       # New methods on synchronisation locks
           def __enter__(self):
               self.acquire()
               return self

           def __exit__(self, *exc_info):
               lock.release()

   Used as follows::

       with myLock:
           # Code here executes with myLock held.  The lock is
           # guaranteed to be released when the block is left (even
           # if via return or by an uncaught exception).

2. A template for opening a file that ensures the file is closed when
   the block is left::

       # New methods on file objects
           def __enter__(self):
               if self.closed:
                   raise RuntimeError, "Cannot reopen closed file handle"
               return self

           def __exit__(self, *args):
               self.close()

   Used as follows::

       with open("/etc/passwd") as f:
           for line in f:
               print line.rstrip()

3. A template for committing or rolling back a database transaction::

       def transaction(db):
           try:
               yield
           except:
               db.rollback()
           else:
               db.commit()

   Used as follows::

       with transaction(the_db):
           make_table(the_db)
           add_data(the_db)
           # Getting to here automatically triggers a commit
           # Any exception automatically triggers a rollback

4. It is possible to nest blocks and combine templates::

       @statement_template
       def lock_opening(lock, filename, mode="r"):
           with lock:
               with open(filename, mode) as f:
                   yield f

   Used as follows::

       with lock_opening(myLock, "/etc/passwd") as f:
           for line in f:
               print line.rstrip()

5. Redirect stdout temporarily::

       @statement_template
       def redirected_stdout(new_stdout):
           save_stdout = sys.stdout
           try:
               sys.stdout = new_stdout
               yield
           finally:
               sys.stdout = save_stdout

   Used as follows::

       with open(filename, "w") as f:
           with redirected_stdout(f):
               print "Hello world"

6. A variant on ``open()`` that also returns an error condition::

       @statement_template
       def open_w_error(filename, mode="r"):
           try:
               f = open(filename, mode)
           except IOError, err:
               yield None, err
           else:
               try:
                   yield f, None
               finally:
                   f.close()

   Used as follows::

       do open_w_error("/etc/passwd", "a") as f, err:
           if err:
               print "IOError:", err
           else:
               f.write("guido::0:0::/:/bin/sh\n")

7. Find the first file with a specific header::

       for name in filenames:
           with open(name) as f:
               if f.read(2) == 0xFEB0:
                   break

8. Find the first item you can handle, holding a lock for the entire
   loop, or just for each iteration::

       with lock:
           for item in items:
               if handle(item):
                   break

       for item in items:
           with lock:
               if handle(item):
                   break

9. Hold a lock while inside a generator, but release it when
   returning control to the outer scope::

       @statement_template
       def released(lock):
           lock.release()
           try:
               yield
           finally:
               lock.acquire()

   Used as follows::

       with lock:
           for item in items:
               with released(lock):
                   yield item

10. Read the lines from a collection of files (e.g. processing
    multiple configuration sources)::

        def all_lines(filenames):
            for name in filenames:
                with open(name) as f:
                    for line in f:
                        yield line

    Used as follows::

        with all_lines(filenames) as lines:
            for line in lines:
                update_config(line)

11. Not all uses need to involve resource management::

        @statement_template
        def tag(*args, **kwds):
            name = cgi.escape(args[0])
            if kwds:
                kwd_pairs = ["%s=%s" % cgi.escape(key), cgi.escape(value)
                             for key, value in kwds]
                print '<%s %s>' % name, " ".join(kwd_pairs)
            else:
                print '<%s>' % name
            yield
            print '</%s>' % name

    Used as follows::

        with tag('html'):
            with tag('head'):
               with tag('title'):
                  print 'A web page'
            with tag('body'):
               for par in pars:
                  with tag('p'):
                     print par
               with tag('a', href="http://www.python.org"):
                   print "Not a dead parrot!"

12. From :pep:`343`, another useful example would be an operation that
    blocks signals.  The use could be like this::

        from signal import blocked_signals

        with blocked_signals():
            # code executed without worrying about signals

    An optional argument might be a list of signals to be blocked; by
    default all signals are blocked.  The implementation is left as an
    exercise to the reader.

13. Another use for this feature is for Decimal contexts::

        # New methods on decimal Context objects

        def __enter__(self):
            if self._old_context is not None:
                raise RuntimeError("Already suspending other Context")
            self._old_context = getcontext()
            setcontext(self)

        def __exit__(self, *args):
            setcontext(self._old_context)
            self._old_context = None

    Used as follows::

        with decimal.Context(precision=28):
           # Code here executes with the given context
           # The context always reverts after this statement


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

None, as this PEP has been withdrawn.


Rejected Options
================

Having the basic construct be a looping construct
-------------------------------------------------

The major issue with this idea, as illustrated by :pep:`340`'s
``block`` statements, is that it causes problems with factoring
``try`` statements that are inside loops, and contain ``break`` and
``continue`` statements (as these statements would then apply to the
``block`` construct, instead of the original loop).  As a key goal is
to be able to factor out arbitrary exception handling (other than
suppression) into statement templates, this is a definite problem.

There is also an understandability problem, as can be seen in the
Examples_.  In the example showing acquisition of a lock either for an
entire loop, or for each iteration of the loop, if the user defined
statement was itself a loop, moving it from outside the ``for`` loop
to inside the ``for`` loop would have major semantic implications,
beyond those one would expect.

Finally, with a looping construct, there are significant problems with
TOOWTDI, as it is frequently unclear whether a particular situation
should be handled with a conventional ``for`` loop or the new looping
construct.  With the current PEP, there is no such problem - ``for``
loops continue to be used for iteration, and the new ``do`` statements
are used to factor out exception handling.

Another issue, specifically with :pep:`340`'s anonymous block statements,
is that they make it quite difficult to write statement templates
directly (i.e. not using a generator).  This problem is addressed by
the current proposal, as can be seen by the relative simplicity of the
various class based implementations of statement templates in the
Examples_.


Allowing statement templates to suppress exceptions
---------------------------------------------------

Earlier versions of this PEP gave statement templates the ability to
suppress exceptions.  The BDFL expressed concern over the associated
complexity, and I agreed after reading an article by Raymond Chen
about the evils of hiding flow control inside macros in C code [1]_.

Removing the suppression ability eliminated a whole lot of complexity
from both the explanation and implementation of user defined
statements, further supporting it as the correct choice.  Older
versions of the PEP had to jump through some horrible hoops to avoid
inadvertently suppressing exceptions in ``__exit__()`` methods - that
issue does not exist with the current suggested semantics.

There was one example (``auto_retry``) that actually used the ability
to suppress exceptions.  This use case, while not quite as elegant,
has significantly more obvious control flow when written out in full
in the user code::

  def attempts(num_tries):
      return reversed(xrange(num_tries))

  for retry in attempts(3):
      try:
          make_attempt()
      except IOError:
          if not retry:
              raise

For what it's worth, the perverse could still write this as::

  for attempt in auto_retry(3, IOError):
      try:
          with attempt:
              make_attempt()
      except FailedAttempt:
          pass

To protect the innocent, the code to actually support that is not
included here.


Differentiating between non-exceptional exits
---------------------------------------------

Earlier versions of this PEP allowed statement templates to
distinguish between exiting the block normally, and exiting via a
``return``, ``break`` or ``continue`` statement.  The BDFL flirted
with a similar idea in :pep:`343` and its associated discussion.  This
added significant complexity to the description of the semantics, and
it required each and every statement template to decide whether or not
those statements should be treated like exceptions, or like a normal
mechanism for exiting the block.

This template-by-template decision process raised great potential for
confusion - consider if one database connector provided a transaction
template that treated early exits like an exception, whereas a second
connector treated them as normal block termination.

Accordingly, this PEP now uses the simplest solution - early exits
appear identical to normal block termination as far as the statement
template is concerned.


Not injecting raised exceptions into generators
-----------------------------------------------

:pep:`343` suggests simply invoking next() unconditionally on generators
used to define statement templates.  This means the template
generators end up looking rather unintuitive, and the retention of the
ban against yielding inside ``try``/``finally`` means that Python's
exception handling capabilities cannot be used to deal with management
of multiple resources.

The alternative which this PEP advocates (injecting raised exceptions
into the generator frame), means that multiple resources can be
managed elegantly as shown by ``lock_opening`` in the Examples_


Making all generators statement templates
-----------------------------------------

Separating the template object from the generator itself makes it
possible to have reusable generator templates.  That is, the following
code will work correctly if this PEP is accepted::

    open_it = lock_opening(parrot_lock, "dead_parrot.txt")

    with open_it as f:
        # use the file for a while

    with open_it as f:
        # use the file again

The second benefit is that iterator generators and template generators
are very different things - the decorator keeps that distinction
clear, and prevents one being used where the other is required.

Finally, requiring the decorator allows the native methods of
generator objects to be used to implement generator finalisation.


Using ``do`` as the keyword
---------------------------

``do`` was an alternative keyword proposed during the :pep:`340`
discussion.  It reads well with appropriately named functions, but it
reads poorly when used with methods, or with objects that provide
native statement template support.

When ``do`` was first suggested, the BDFL had rejected :pep:`310`'s
``with`` keyword, based on a desire to use it for a Pascal/Delphi
style ``with`` statement.  Since then, the BDFL has retracted this
objection, as he no longer intends to provide such a statement.  This
change of heart was apparently based on the C# developers reasons for
not providing the feature [2]_.


Not having a keyword
--------------------

This is an interesting option, and can be made to read quite well.
However, it's awkward to look up in the documentation for new users,
and strikes some as being too magical.  Accordingly, this PEP goes
with a keyword based suggestion.


Enhancing ``try`` statements
----------------------------

This suggestion involves give bare ``try`` statements a signature
similar to that proposed for ``with`` statements.

I think that trying to write a ``with`` statement as an enhanced
``try`` statement makes as much sense as trying to write a ``for``
loop as an enhanced ``while`` loop.  That is, while the semantics of
the former can be explained as a particular way of using the latter,
the former is not an *instance* of the latter.  The additional
semantics added around the more fundamental statement result in a new
construct, and the two different statements shouldn't be confused.

This can be seen by the fact that the 'enhanced' ``try`` statement
still needs to be explained in terms of a 'non-enhanced' ``try``
statement.  If it's something different, it makes more sense to give
it a different name.


Having the template protocol directly reflect ``try`` statements
----------------------------------------------------------------

One suggestion was to have separate methods in the protocol to cover
different parts of the structure of a generalised ``try`` statement.
Using the terms ``try``, ``except``, ``else`` and ``finally``, we
would have something like::

    class my_template(object):

        def __init__(self, *args):
            # Any required arguments (e.g. a file name)
            # get stored in member variables
            # The various BLOCK's will need to updated to reflect
            # that.

        def __try__(self):
            SETUP_BLOCK

        def __except__(self, exc, value, traceback):
            if isinstance(exc, exc_type1):
                EXCEPT_BLOCK1
            if isinstance(exc, exc_type2):
                EXCEPT_BLOCK2
            else:
                EXCEPT_BLOCK3

        def __else__(self):
            ELSE_BLOCK

        def __finally__(self):
            FINALLY_BLOCK

Aside from preferring the addition of two method slots rather than
four, I consider it significantly easier to be able to simply
reproduce a slightly modified version of the original ``try``
statement code in the ``__exit__()`` method (as shown in `Factoring
out arbitrary exception handling`_), rather than have to split the
functionality amongst several different methods (or figure out
which method to use if not all clauses are used by the template).

To make this discussion less theoretical, here is the ``transaction``
example implemented using both the two method and the four method
protocols instead of a generator.  Both implementations guarantee a
commit if a ``break``, ``return`` or ``continue`` statement is
encountered (as does the generator-based implementation in the
Examples_ section)::

    class transaction_2method(object):

        def __init__(self, db):
            self.db = db

        def __enter__(self):
            pass

        def __exit__(self, exc_type, *exc_details):
            if exc_type is None:
                self.db.commit()
            else:
                self.db.rollback()

    class transaction_4method(object):

        def __init__(self, db):
            self.db = db
            self.commit = False

        def __try__(self):
            self.commit = True

        def __except__(self, exc_type, exc_value, traceback):
            self.db.rollback()
            self.commit = False

        def __else__(self):
            pass

        def __finally__(self):
            if self.commit:
                self.db.commit()
                self.commit = False

There are two more minor points, relating to the specific method names
in the suggestion.  The name of the ``__try__()`` method is
misleading, as ``SETUP_BLOCK`` executes *before* the ``try`` statement
is entered, and the name of the ``__else__()`` method is unclear in
isolation, as numerous other Python statements include an ``else``
clause.


Iterator finalisation (WITHDRAWN)
=================================

The ability to use user defined statements inside generators is likely
to increase the need for deterministic finalisation of iterators, as
resource management is pushed inside the generators, rather than being
handled externally as is currently the case.

The PEP currently suggests handling this by making all generators
statement templates, and using ``with`` statements to handle
finalisation.  However, earlier versions of this PEP suggested the
following, more complex, solution, that allowed the *author* of a
generator to flag the need for finalisation, and have ``for`` loops
deal with it automatically.  It is included here as a long, detailed
rejected option.


Iterator protocol addition: ``__finish__``
------------------------------------------

An optional new method for iterators is proposed, called
``__finish__()``.  It takes no arguments, and should not return
anything.

The ``__finish__`` method is expected to clean up all resources the
iterator has open.  Iterators with a ``__finish__()`` method are
called 'finishable iterators' for the remainder of the PEP.


Best effort finalisation
------------------------

A finishable iterator should ensure that it provides a ``__del__``
method that also performs finalisation (e.g. by invoking the
``__finish__()`` method).  This allows Python to still make a best
effort at finalisation in the event that deterministic finalisation is
not applied to the iterator.


Deterministic finalisation
--------------------------

If the iterator used in a ``for`` loop has a ``__finish__()`` method,
the enhanced ``for`` loop semantics will guarantee that that method
will be executed, regardless of the means of exiting the loop.  This
is important for iterator generators that utilise `user defined
statements`_ or the now permitted ``try``/``finally`` statements, or
for new iterators that rely on timely finalisation to release
allocated resources (e.g. releasing a thread or database connection
back into a pool).


``for`` loop syntax
-------------------

No changes are suggested to ``for`` loop syntax.  This is just to
define the statement parts needed for the description of the
semantics::

    for VAR1 in EXPR1:
        BLOCK1
    else:
        BLOCK2


Updated ``for`` loop semantics
------------------------------

When the target iterator does not have a ``__finish__()`` method, a
``for`` loop will execute as follows (i.e. no change from the status
quo)::

    itr = iter(EXPR1)
    exhausted = False
    while True:
        try:
            VAR1 = itr.next()
        except StopIteration:
            exhausted = True
            break
        BLOCK1
    if exhausted:
        BLOCK2

When the target iterator has a ``__finish__()`` method, a ``for`` loop
will execute as follows::

    itr = iter(EXPR1)
    exhausted = False
    try:
        while True:
            try:
                VAR1 = itr.next()
            except StopIteration:
                exhausted = True
                break
            BLOCK1
        if exhausted:
            BLOCK2
    finally:
        itr.__finish__()

The implementation will need to take some care to avoid incurring the
``try``/``finally`` overhead when the iterator does not have a
``__finish__()`` method.


Generator iterator finalisation: ``__finish__()`` method
--------------------------------------------------------

When enabled with the appropriate decorator, generators will have a
``__finish__()`` method that raises ``TerminateIteration`` in the
internal frame::

    def __finish__(self):
        try:
            self._inject_exception(TerminateIteration)
        except TerminateIteration:
            pass

A decorator (e.g. ``needs_finish()``) is required to enable this
feature, so that existing generators (which are not expecting
finalisation) continue to work as expected.


Partial iteration of finishable iterators
-----------------------------------------

Partial iteration of a finishable iterator is possible, although it
requires some care to ensure the iterator is still finalised promptly
(it was made finishable for a reason!).  First, we need a class to
enable partial iteration of a finishable iterator by hiding the
iterator's ``__finish__()`` method from the ``for`` loop::

    class partial_iter(object):

        def __init__(self, iterable):
            self.iter = iter(iterable)

        def __iter__(self):
            return self

        def next(self):
            return self.itr.next()

Secondly, an appropriate statement template is needed to ensure the
iterator is finished eventually::

    @statement_template
    def finishing(iterable):
          itr = iter(iterable)
          itr_finish = getattr(itr, "__finish__", None)
          if itr_finish is None:
              yield itr
          else:
              try:
                  yield partial_iter(itr)
              finally:
                  itr_finish()

This can then be used as follows::

    do finishing(finishable_itr) as itr:
        for header_item in itr:
            if end_of_header(header_item):
                break
            # process header item
        for body_item in itr:
            # process body item

Note that none of the above is needed for an iterator that is not
finishable - without a ``__finish__()`` method, it will not be
promptly finalised by the ``for`` loop, and hence inherently allows
partial iteration.  Allowing partial iteration of non-finishable
iterators as the default behaviour is a key element in keeping this
addition to the iterator protocol backwards compatible.


Acknowledgements
================

The acknowledgements section for :pep:`340` applies, since this text grew
out of the discussion of that PEP, but additional thanks go to Michael
Hudson, Paul Moore and Guido van Rossum for writing :pep:`310` and PEP
340 in the first place, and to (in no meaningful order) Fredrik Lundh,
Phillip J. Eby, Steven Bethard, Josiah Carlson, Greg Ewing, Tim
Delaney and Arnold deVos for prompting particular ideas that made
their way into this text.


References
==========

.. [1] A rant against flow control macros
   (http://blogs.msdn.com/oldnewthing/archive/2005/01/06/347666.aspx)

.. [2] Why doesn't C# have a 'with' statement?
   (http://msdn.microsoft.com/vcsharp/programming/language/ask/withstatement/)


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
   End: