Misc/NEWS

+++++++++++
Python News
+++++++++++

(editors: check NEWS.help for information about editing NEWS using ReST.)

What's New in Python 2.3 beta 1?
================================

*Release date: XX-XXX-2003*

Core and builtins
-----------------

- On 64-bit systems, a dictionary could contain duplicate long/int keys
  if the key value was larger than 2**32.  See SF bug #689659.

Extension modules
-----------------

- Made user requested changes to the itertools module.
  Subsumed the times() function into repeat().
  Added chain() and cycle().

Library
-------

- os.listdir() now returns Unicode strings on platforms that set
  Py_FileSystemDefaultEncoding, for file names that are not representable
  in ASCII.  (This currently only affects MacOS X; on Windows versions
  with wide file name support os.listdir() already returned Unicode
  strings.)

Tools/Demos
-----------

TBD

Build
-----

- Fix build problems when _iconv_codec failed.  (SF bug #690012.)

- Fix problem building on OSF1 because the compiler only accepted
  preprocessor directives that start in column 1.  (SF bug #691793.)

C API
-----

TBD

New platforms
-------------

TBD

Tests
-----

- test_timeout now requires -u network to be passed to regrtest to run.
  See SF bug #692988.

Windows
-------

TBD

Mac
---

- os.listdir() now may return Unicode strings on MacOS X. See the general
  news item under "Library".

- A new method MacOS.WMAvailable() returns true if it is safe to access
  the window manager, false otherwise.
  
- EasyDialogs dialogs are now movable-modal.


What's New in Python 2.3 alpha 2?
=================================

*Release date: 19-Feb-2003*

Core and builtins
-----------------

- Negative positions returned from PEP 293 error callbacks are now
  treated as being relative to the end of the input string. Positions
  that are out of bounds raise an IndexError.

- sys.path[0] (the directory from which the script is loaded) is now
  turned into an absolute pathname, unless it is the empty string.
  (SF patch #664376.)

- Finally fixed the bug in compile() and exec where a string ending
  with an indented code block but no newline would raise SyntaxError.
  This would have been a four-line change in parsetok.c...  Except
  codeop.py depends on this behavior, so a compilation flag had to be
  invented that causes the tokenizer to revert to the old behavior;
  this required extra changes to 2 .h files, 2 .c files, and 2 .py
  files.  (Fixes SF bug #501622.)

- If a new-style class defines neither __new__ nor __init__, its
  constructor would ignore all arguments.  This is changed now: the
  constructor refuses arguments in this case.  This might break code
  that worked under Python 2.2.  The simplest fix is to add a no-op
  __init__: ``def __init__(self, *args, **kw): pass``.

- Through a bytecode optimizer bug (and I bet you didn't even know
  Python *had* a bytecode optimizer :-), "unsigned" hex/oct constants
  with a leading minus sign would come out with the wrong sign.
  ("Unsigned" hex/oct constants are those with a face value in the
  range sys.maxint+1 through sys.maxint*2+1, inclusive; these have
  always been interpreted as negative numbers through sign folding.)
  E.g. 0xffffffff is -1, and -(0xffffffff) is 1, but -0xffffffff would
  come out as -4294967295.  This was the case in Python 2.2 through
  2.2.2 and 2.3a1, and in Python 2.4 it will once again have that
  value, but according to PEP 237 it really needs to be 1 now.  This
  will be backported to Python 2.2.3 a well.  (SF #660455)

- int(s, base) sometimes sign-folds hex and oct constants; it only
  does this when base is 0 and s.strip() starts with a '0'.  When the
  sign is actually folded, as in int("0xffffffff", 0) on a 32-bit
  machine, which returns -1, a FutureWarning is now issued; in Python
  2.4, this will return 4294967295L, as do int("+0xffffffff", 0) and
  int("0xffffffff", 16) right now.  (PEP 347)

- super(X, x): x may now be a proxy for an X instance, i.e.
  issubclass(x.__class__, X) but not issubclass(type(x), X).

- isinstance(x, X): if X is a new-style class, this is now equivalent
  to issubclass(type(x), X) or issubclass(x.__class__, X).  Previously
  only type(x) was tested.  (For classic classes this was already the
  case.)

- compile(), eval() and the exec statement now fully support source code
  passed as unicode strings.

- int subclasses can be initialized with longs if the value fits in an int.
  See SF bug #683467.

- long(string, base) takes time linear in len(string) when base is a power
  of 2 now.  It used to take time quadratic in len(string).

- filter returns now Unicode results for Unicode arguments.

- raw_input can now return Unicode objects.

- List objects' sort() method now accepts None as the comparison function.
  Passing None is semantically identical to calling sort() with no
  arguments.

- Fixed crash when printing a subclass of str and __str__ returned self.
  See SF bug #667147.

- Fixed an invalid RuntimeWarning and an undetected error when trying
  to convert a long integer into a float which couldn't fit.
  See SF bug #676155.

- Function objects now have a __module__ attribute that is bound to
  the name of the module in which the function was defined.  This
  applies for C functions and methods as well as functions and methods
  defined in Python.  This attribute is used by pickle.whichmodule(),
  which changes the behavior of whichmodule slightly.  In Python 2.2
  whichmodule() returns "__main__" for functions that are not defined
  at the top-level of a module (examples: methods, nested functions).
  Now whichmodule() will return the proper module name.

Extension modules
-----------------

- operator.isNumberType() now checks that the object has a nb_int or
  nb_float slot, rather than simply checking whether it has a non-NULL
  tp_as_number pointer.

- The imp module now has ways to acquire and release the "import
  lock": imp.acquire_lock() and imp.release_lock().  Note: this is a
  reentrant lock, so releasing the lock only truly releases it when
  this is the last release_lock() call.  You can check with
  imp.lock_held().  (SF bug #580952 and patch #683257.)

- Change to cPickle to match pickle.py (see below and PEP 307).

- Fix some bugs in the parser module.  SF bug #678518.

- Thanks to Scott David Daniels, a subtle bug in how the zlib
  extension implemented flush() was fixed.  Scott also rewrote the
  zlib test suite using the unittest module.  (SF bug #640230 and
  patch #678531.)

- Added an itertools module containing high speed, memory efficient
  looping constructs inspired by tools from Haskell and SML.

- The SSL module now handles sockets with a timeout set correctly (SF
  patch #675750, fixing SF bug #675552).

- A new module _iconv_codec has been added, to expose the iconv(3)
  library.

- os/posixmodule has grown the sysexits.h constants (EX_OK and friends).

- Fixed broken threadstate swap in readline that could cause fatal
  errors when a readline hook was being invoked while a background
  thread was active.  (SF bugs #660476 and #513033.)

- fcntl now exposes the strops.h I_* constants.

- Fix a crash on Solaris that occurred when calling close() on
  an mmap'ed file which was already closed.  (SF patch #665913)

- Fixed serveral serious bugs in the zipimport implementation.

- datetime changes:

  The datetime and datetimetz classes have been collapsed into a single
  datetime class, and likewise the time and timetz classes into a single
  time class.  Previously, a datetimetz object with tzinfo=None acted
  exactly like a datetime object, and similarly for timetz.  This wasn't
  enough of a difference to justify distinct classes, and life is simpler
  now.

  today() and now() now round system timestamps to the closest
  microsecond <http://www.python.org/sf/661086>.  This repairs an
  irritation most likely seen on Windows systems.

  In dt.astimezone(tz), if tz.utcoffset(dt) returns a duration,
  ValueError is raised if tz.dst(dt) returns None (2.3a1 treated it
  as 0 instead, but a tzinfo subclass wishing to participate in
  time zone conversion has to take a stand on whether it supports
  DST; if you don't care about DST, then code dst() to return 0 minutes,
  meaning that DST is never in effect).

  The tzinfo methods utcoffset() and dst() must return a timedelta object
  (or None) now.  In 2.3a1 they could also return an int or long, but that
  was an unhelpfully redundant leftover from an earlier version wherein
  they couldn't return a timedelta.  TOOWTDI.

  The example tzinfo class for local time had a bug.  It was replaced
  by a later example coded by Guido.

  datetime.astimezone(tz) no longer raises an exception when the
  input datetime has no UTC equivalent in tz.  For typical "hybrid" time
  zones (a single tzinfo subclass modeling both standard and daylight
  time), this case can arise one hour per year, at the hour daylight time
  ends.  See new docs for details.  In short, the new behavior mimics
  the local wall clock's behavior of repeating an hour in local time.

  dt.astimezone() can no longer be used to convert between naive and aware
  datetime objects.  If you merely want to attach, or remove, a tzinfo
  object, without any conversion of date and time members, use
  dt.replace(tzinfo=whatever) instead, where "whatever" is None or a
  tzinfo subclass instance.

  A new method tzinfo.fromutc(dt) can be overridden in tzinfo subclasses
  to give complete control over how a UTC time is to be converted to
  a local time.  The default astimezone() implementation calls fromutc()
  as its last step, so a tzinfo subclass can affect that too by overriding
  fromutc().  It's expected that the default fromutc() implementation will
  be suitable as-is for "almost all" time zone subclasses, but the
  creativity of political time zone fiddling appears unbounded -- fromutc()
  allows the highly motivated to emulate any scheme expressible in Python.

  datetime.now():  The optional tzinfo argument was undocumented (that's
  repaired), and its name was changed to tz ("tzinfo" is overloaded enough
  already).  With a tz argument, now(tz) used to return the local date
  and time, and attach tz to it, without any conversion of date and time
  members.  This was less than useful.  Now now(tz) returns the current
  date and time as local time in tz's time zone, akin to ::

      tz.fromutc(datetime.utcnow().replace(tzinfo=utc))

  where "utc" is an instance of a tzinfo subclass modeling UTC.  Without
  a tz argument, now() continues to return the current local date and time,
  as a naive datetime object.

  datetime.fromtimestamp():  Like datetime.now() above, this had less than
  useful behavior when the optional tinzo argument was specified.  See
  also SF bug report <http://www.python.org/sf/660872>.

  date and datetime comparison:  In order to prevent comparison from
  falling back to the default compare-object-addresses strategy, these
  raised TypeError whenever they didn't understand the other object type.
  They still do, except when the other object has a "timetuple" attribute,
  in which case they return NotImplemented now.  This gives other
  datetime objects (e.g., mxDateTime) a chance to intercept the
  comparison.

  date, time, datetime and timedelta comparison:  When the exception
  for mixed-type comparisons in the last paragraph doesn't apply, if
  the comparison is == then False is returned, and if the comparison is
  != then True is returned.  Because dict lookup and the "in" operator
  only invoke __eq__, this allows, for example, ::

      if some_datetime in some_sequence:

  and ::

      some_dict[some_timedelta] = whatever

  to work as expected, without raising TypeError just because the
  sequence is heterogeneous, or the dict has mixed-type keys.  [This
  seems like a good idea to implement for all mixed-type comparisons
  that don't want to allow falling back to address comparison.]

  The constructors building a datetime from a timestamp could raise
  ValueError if the platform C localtime()/gmtime() inserted "leap
  seconds".  Leap seconds are ignored now.  On such platforms, it's
  possible to have timestamps that differ by a second, yet where
  datetimes constructed from them are equal.

  The pickle format of date, time and datetime objects has changed
  completely.  The undocumented pickler and unpickler functions no
  longer exist.  The undocumented __setstate__() and __getstate__()
  methods no longer exist either.

Library
-------

- The logging module was updated slightly; the WARN level was renamed
  to WARNING, and the matching function/method warn() to warning().

- The pickle and cPickle modules were updated with a new pickling
  protocol (documented by pickletools.py, see below) and several
  extensions to the pickle customization API (__reduce__, __setstate__
  etc.).  The copy module now uses more of the pickle customization
  API to copy objects that don't implement __copy__ or __deepcopy__.
  See PEP 307 for details.

- The distutils "register" command now uses http://www.python.org/pypi
  as the default repository.  (See PEP 301.)

- the platform dependent path related variables sep, altsep, extsep,
  pathsep, curdir, pardir and defpath are now defined in the platform
  dependent path modules (e.g. ntpath.py) rather than os.py, so these
  variables are now available via os.path.  They continue to be 
  available from the os module.  
  (see <http://www.python.org/sf/680789>).

- array.array was added to the types repr.py knows about (see
  <http://www.python.org/sf/680789>).

- The new pickletools.py contains lots of documentation about pickle
  internals, and supplies some helpers for working with pickles, such as
  a symbolic pickle disassembler.

- Xmlrpclib.py now supports the builtin boolean type.

- py_compile has a new 'doraise' flag and a new PyCompileError
  exception.

- SimpleXMLRPCServer now supports CGI through the CGIXMLRPCRequestHandler
  class.

- The sets module now raises TypeError in __cmp__, to clarify that
  sets are not intended to be three-way-compared; the comparison
  operators are overloaded as subset/superset tests.

- Bastion.py and rexec.py are disabled.  These modules are not safe in
  Python 2.2. or 2.3.

- realpath is now exported when doing ``from poxixpath import *``.
  It is also exported for ntpath, macpath, and os2emxpath.
  See SF bug #659228.

- New module tarfile from Lars Gustäbel provides a comprehensive interface
  to tar archive files with transparent gzip and bzip2 compression.
  See SF patch #651082.

- urlparse can now parse imap:// URLs.  See SF feature request #618024.

- Tkinter.Canvas.scan_dragto() provides an optional parameter to support
  the gain value which is passed to Tk.  SF bug# 602259.

- Fix logging.handlers.SysLogHandler protocol when using UNIX domain sockets.
  See SF patch #642974.

- The dospath module was deleted.  Use the ntpath module when manipulating
  DOS paths from other platforms.

Tools/Demos
-----------

- Two new scripts (db2pickle.py and pickle2db.py) were added to the
  Tools/scripts directory to facilitate conversion from the old bsddb module
  to the new one.  While the user-visible API of the new module is
  compatible with the old one, it's likely that the version of the
  underlying database library has changed.  To convert from the old library,
  run the db2pickle.py script using the old version of Python to convert it
  to a pickle file.  After upgrading Python, run the pickle2db.py script
  using the new version of Python to reconstitute your database.  For
  example:

    % python2.2 db2pickle.py -h some.db > some.pickle
    % python2.3 pickle2db.py -h some.db.new < some.pickle

  Run the scripts without any args to get a usage message.


Build
-----

- The audio driver tests (test_ossaudiodev.py and
  test_linuxaudiodev.py) are no longer run by default.  This is
  because they don't always work, depending on your hardware and
  software.  To run these tests, you must use an invocation like ::

    ./python Lib/test/regrtest.py -u audio test_ossaudiodev

- On systems which build using the configure script, compiler flags which
  used to be lumped together using the OPT flag have been split into two
  groups, OPT and BASECFLAGS.  OPT is meant to carry just optimization- and
  debug-related flags like "-g" and "-O3".  BASECFLAGS is meant to carry
  compiler flags that are required to get a clean compile.  On some
  platforms (many Linux flavors in particular) BASECFLAGS will be empty by
  default.  On others, such as Mac OS X and SCO, it will contain required
  flags.  This change allows people building Python to override OPT without
  fear of clobbering compiler flags which are required to get a clean build.

- On Darwin/Mac OS X platforms, /sw/lib and /sw/include are added to the
  relevant search lists in setup.py.  This allows users building Python to
  take advantage of the many packages available from the fink project
  <http://fink.sf.net/>.

- A new Makefile target, scriptsinstall, installs a number of useful scripts
  from the Tools/scripts directory.

C API
-----

- PyEval_GetFrame() is now declared to return a ``PyFrameObject *``
  instead of a plain ``PyObject *``.  (SF patch #686601.)

- PyNumber_Check() now checks that the object has a nb_int or nb_float
  slot, rather than simply checking whether it has a non-NULL
  tp_as_number pointer.

- A C type that inherits from a base type that defines tp_as_buffer
  will now inherit the tp_as_buffer pointer if it doesn't define one.
  (SF #681367)

- The PyArg_Parse functions now issue a DeprecationWarning if a float
  argument is provided when an integer is specified (this affects the 'b',
  'B', 'h', 'H', 'i', and 'l' codes).  Future versions of Python will
  raise a TypeError.

Tests
-----

- Several tests weren't being run rom regrtest.py (test_timeout.py,
  test_tarfile.py, test_netrc.py, test_multifile.py,
  test_importhooks.py and test_imp.py).  Now they are.  (Note to
  developers: please read Lib/test/README when creating a new test, to
  make sure to do it right!  All tests need to use either unittest or
  pydoc.)

- Added test_posix.py, a test suite for the posix module.

- Added test_hexoct.py, a test suite for hex/oct constant folding.

Windows
-------

- The timeout code for socket connect() didn't work right; this has
  now been fixed.  test_timeout.py should pass (at least most of the
  time).

- distutils' msvccompiler class now passes the preprocessor options to
  the resource compiler.  See SF patch #669198.

- The bsddb module now ships with Sleepycat's 4.1.25.NC, the latest
  release without strong cryptography.

- sys.path[0], if it contains a directory name, is now always an
  absolute pathname. (SF patch #664376.)

- The new logging package is now installed by the Windows installer.  It
  wasn't in 2.3a1 due to oversight.

Mac
---

- There are new dialogs EasyDialogs.AskFileForOpen, AskFileForSave
  and AskFolder. The old macfs.StandardGetFile and friends are deprecated.

- Most of the standard library now uses pathnames or FSRefs in preference
  of FSSpecs, and use the underlying Carbon.File and Carbon.Folder modules
  in stead of macfs. macfs will probably be deprecated in the future.

- Type Carbon.File.FSCatalogInfo and supporting methods have been implemented.
  This also makes macfs.FSSpec.SetDates() work again.
  
- There is a new module pimp, the package install manager for Python, and
  accompanying applet PackageManager. These allow you to easily download
  and install pretested extension packages either in source or binary
  form. Only in MacPython-OSX.
  
- Applets are now built with bundlebuilder in MacPython-OSX, which should make
  them more robust and also provides a path towards BuildApplication. The
  downside of this change is that applets can no longer be run from the
  Terminal window, this will hopefully be fixed in the 2.3b1.


What's New in Python 2.3 alpha 1?
=================================

*Release date: 31-Dec-2002*

Type/class unification and new-style classes
--------------------------------------------

- One can now assign to __bases__ and __name__ of new-style classes.

- dict() now accepts keyword arguments so that dict(one=1, two=2)
  is the equivalent of {"one": 1, "two": 2}.  Accordingly,
  the existing (but undocumented) 'items' keyword argument has
  been eliminated.  This means that dict(items=someMapping) now has
  a different meaning than before.

- int() now returns a long object if the argument is outside the
  integer range, so int("4" * 1000), int(1e200) and int(1L<<1000) will
  all return long objects instead of raising an OverflowError.

- Assignment to __class__ is disallowed if either the old or the new
  class is a statically allocated type object (such as defined by an
  extension module).  This prevents anomalies like 2.__class__ = bool.

- New-style object creation and deallocation have been sped up
  significantly; they are now faster than classic instance creation
  and deallocation.

- The __slots__ variable can now mention "private" names, and the
  right thing will happen (e.g. __slots__ = ["__foo"]).

- The built-ins slice() and buffer() are now callable types.  The
  types classobj (formerly class), code, function, instance, and
  instancemethod (formerly instance-method), which have no built-in
  names but are accessible through the types module, are now also
  callable.  The type dict-proxy is renamed to dictproxy.

- Cycles going through the __class__ link of a new-style instance are
  now detected by the garbage collector.

- Classes using __slots__ are now properly garbage collected.
  [SF bug 519621]

- Tightened the __slots__ rules: a slot name must be a valid Python
  identifier.

- The constructor for the module type now requires a name argument and
  takes an optional docstring argument.  Previously, this constructor
  ignored its arguments.  As a consequence, deriving a class from a
  module (not from the module type) is now illegal; previously this
  created an unnamed module, just like invoking the module type did.
  [SF bug 563060]

- A new type object, 'basestring', is added.  This is a common base type
  for 'str' and 'unicode', and can be used instead of
  types.StringTypes, e.g. to test whether something is "a string":
  isinstance(x, basestring) is True for Unicode and 8-bit strings.  This
  is an abstract base class and cannot be instantiated directly.

- Changed new-style class instantiation so that when C's __new__
  method returns something that's not a C instance, its __init__ is
  not called.  [SF bug #537450]

- Fixed super() to work correctly with class methods.  [SF bug #535444]

- If you try to pickle an instance of a class that has __slots__ but
  doesn't define or override __getstate__, a TypeError is now raised.
  This is done by adding a bozo __getstate__ to the class that always
  raises TypeError.  (Before, this would appear to be pickled, but the
  state of the slots would be lost.)

Core and builtins
-----------------

- Import from zipfiles is now supported.  The name of a zipfile placed
  on sys.path causes the import statement to look for importable Python
  modules (with .py, pyc and .pyo extensions) and packages inside the
  zipfile.  The zipfile import follows the specification (though not
  the sample implementation) of PEP 273.  The semantics of __path__ are
  compatible with those that have been implemented in Jython since
  Jython 2.1.

- PEP 302 has been accepted.  Although it was inititally developed to
  support zipimport, it offers a new, general import hook mechanism.
  Several new variables have been added to the sys module:
  sys.meta_path, sys.path_hooks, and sys.path_importer_cache; these
  make extending the import statement much more convenient than
  overriding the __import__ built-in function.  For a description of
  these, see PEP 302.

- A frame object's f_lineno attribute can now be written to from a
  trace function to change which line will execute next.  A command to
  exploit this from pdb has been added.  [SF patch #643835]

- The _codecs support module for codecs.py was turned into a builtin
  module to assure that at least the builtin codecs are available
  to the Python parser for source code decoding according to PEP 263.

- issubclass now supports a tuple as the second argument, just like
  isinstance does. ``issubclass(X, (A, B))`` is equivalent to
  ``issubclass(X, A) or issubclass(X, B)``.

- Thanks to Armin Rigo, the last known way to provoke a system crash
  by cleverly arranging for a comparison function to mutate a list
  during a list.sort() operation has been fixed.  The effect of
  attempting to mutate a list, or even to inspect its contents or
  length, while a sort is in progress, is not defined by the language.
  The C implementation of Python 2.3 attempts to detect mutations,
  and raise ValueError if one occurs, but there's no guarantee that
  all mutations will be caught, or that any will be caught across
  releases or implementations.

- Unicode file name processing for Windows (PEP 277) is implemented.
  All platforms now have an os.path.supports_unicode_filenames attribute,
  which is set to True on Windows NT/2000/XP, and False elsewhere.

- Codec error handling callbacks (PEP 293) are implemented.
  Error handling in unicode.encode or str.decode can now be customized.

- A subtle change to the semantics of the built-in function intern():
  interned strings are no longer immortal.  You must keep a reference
  to the return value intern() around to get the benefit.

- Use of 'None' as a variable, argument or attribute name now
  issues a SyntaxWarning.  In the future, None may become a keyword.

- SET_LINENO is gone.  co_lnotab is now consulted to determine when to
  call the trace function.  C code that accessed f_lineno should call
  PyCode_Addr2Line instead (f_lineno is still there, but only kept up
  to date when there is a trace function set).

- There's a new warning category, FutureWarning.  This is used to warn
  about a number of situations where the value or sign of an integer
  result will change in Python 2.4 as a result of PEP 237 (integer
  unification).  The warnings implement stage B0 mentioned in that
  PEP.  The warnings are about the following situations:

    - Octal and hex literals without 'L' prefix in the inclusive range
      [0x80000000..0xffffffff]; these are currently negative ints, but
      in Python 2.4 they will be positive longs with the same bit
      pattern.

    - Left shifts on integer values that cause the outcome to lose
      bits or have a different sign than the left operand.  To be
      precise: x<<n where this currently doesn't yield the same value
      as long(x)<<n; in Python 2.4, the outcome will be long(x)<<n.

    - Conversions from ints to string that show negative values as
      unsigned ints in the inclusive range [0x80000000..0xffffffff];
      this affects the functions hex() and oct(), and the string
      formatting codes %u, %o, %x, and %X.  In Python 2.4, these will
      show signed values (e.g. hex(-1) currently returns "0xffffffff";
      in Python 2.4 it will return "-0x1").

- The bits manipulated under the cover by sys.setcheckinterval() have
  been changed.  Both the check interval and the ticker used to be
  per-thread values.  They are now just a pair of global variables.
  In addition, the default check interval was boosted from 10 to 100
  bytecode instructions.  This may have some effect on systems that
  relied on the old default value.  In particular, in multi-threaded
  applications which try to be highly responsive, response time will
  increase by some (perhaps imperceptible) amount.

- When multiplying very large integers, a version of the so-called
  Karatsuba algorithm is now used.  This is most effective if the
  inputs have roughly the same size.  If they both have about N digits,
  Karatsuba multiplication has O(N**1.58) runtime (the exponent is
  log_base_2(3)) instead of the previous O(N**2).  Measured results may
  be better or worse than that, depending on platform quirks.  Besides
  the O() improvement in raw instruction count, the Karatsuba algorithm
  appears to have much better cache behavior on extremely large integers
  (starting in the ballpark of a million bits).  Note that this is a
  simple implementation, and there's no intent here to compete with,
  e.g., GMP.  It gives a very nice speedup when it applies, but a package
  devoted to fast large-integer arithmetic should run circles around it.

- u'%c' will now raise a ValueError in case the argument is an
  integer outside the valid range of Unicode code point ordinals.

- The tempfile module has been overhauled for enhanced security.  The
  mktemp() function is now deprecated; new, safe replacements are
  mkstemp() (for files) and mkdtemp() (for directories), and the
  higher-level functions NamedTemporaryFile() and TemporaryFile().
  Use of some global variables in this module is also deprecated; the
  new functions have keyword arguments to provide the same
  functionality.  All Lib, Tools and Demo modules that used the unsafe
  interfaces have been updated to use the safe replacements.  Thanks
  to Zack Weinberg!

- When x is an object whose class implements __mul__ and __rmul__,
  1.0*x would correctly invoke __rmul__, but 1*x would erroneously
  invoke __mul__.  This was due to the sequence-repeat code in the int
  type.  This has been fixed now.

- Previously, "str1 in str2" required str1 to be a string of length 1.
  This restriction has been relaxed to allow str1 to be a string of
  any length.  Thus "'el' in 'hello world'" returns True now.

- File objects are now their own iterators.  For a file f, iter(f) now
  returns f (unless f is closed), and f.next() is similar to
  f.readline() when EOF is not reached; however, f.next() uses a
  readahead buffer that messes up the file position, so mixing
  f.next() and f.readline() (or other methods) doesn't work right.
  Calling f.seek() drops the readahead buffer, but other operations
  don't.  It so happens that this gives a nice additional speed boost
  to "for line in file:"; the xreadlines method and corresponding
  module are now obsolete.  Thanks to Oren Tirosh!

- Encoding declarations (PEP 263, phase 1) have been implemented.  A
  comment of the form "# -*- coding: <encodingname> -*-" in the first
  or second line of a Python source file indicates the encoding.

- list.sort() has a new implementation.  While cross-platform results
  may vary, and in data-dependent ways, this is much faster on many
  kinds of partially ordered lists than the previous implementation,
  and reported to be just as fast on randomly ordered lists on
  several major platforms.  This sort is also stable (if A==B and A
  precedes B in the list at the start, A precedes B after the sort too),
  although the language definition does not guarantee stability.  A
  potential drawback is that list.sort() may require temp space of
  len(list)*2 bytes (``*4`` on a 64-bit machine).  It's therefore possible
  for list.sort() to raise MemoryError now, even if a comparison function
  does not.  See <http://www.python.org/sf/587076> for full details.

- All standard iterators now ensure that, once StopIteration has been
  raised, all future calls to next() on the same iterator will also
  raise StopIteration.  There used to be various counterexamples to
  this behavior, which could caused confusion or subtle program
  breakage, without any benefits.  (Note that this is still an
  iterator's responsibility; the iterator framework does not enforce
  this.)

- Ctrl+C handling on Windows has been made more consistent with
  other platforms.  KeyboardInterrupt can now reliably be caught,
  and Ctrl+C at an interactive prompt no longer terminates the
  process under NT/2k/XP (it never did under Win9x).  Ctrl+C will
  interrupt time.sleep() in the main thread, and any child processes
  created via the popen family (on win2k; we can't make win9x work
  reliably) are also interrupted (as generally happens on for Linux/Unix.)
  [SF bugs 231273, 439992 and 581232]

- sys.getwindowsversion() has been added on Windows.  This
  returns a tuple with information about the version of Windows
  currently running.

- Slices and repetitions of buffer objects now consistently return
  a string.  Formerly, strings would be returned most of the time,
  but a buffer object would be returned when the repetition count
  was one or when the slice range was all inclusive.

- Unicode objects in sys.path are no longer ignored but treated
  as directory names.

- Fixed string.startswith and string.endswith builtin methods
  so they accept negative indices.  [SF bug 493951]

- Fixed a bug with a continue inside a try block and a yield in the
  finally clause.  [SF bug 567538]

- Most builtin sequences now support "extended slices", i.e. slices
  with a third "stride" parameter.  For example, "hello world"[::-1]
  gives "dlrow olleh".

- A new warning PendingDeprecationWarning was added to provide
  direction on features which are in the process of being deprecated.
  The warning will not be printed by default.  To see the pending
  deprecations, use -Walways::PendingDeprecationWarning::
  as a command line option or warnings.filterwarnings() in code.

- Deprecated features of xrange objects have been removed as
  promised.  The start, stop, and step attributes and the tolist()
  method no longer exist.  xrange repetition and slicing have been
  removed.

- New builtin function enumerate(x), from PEP 279.  Example:
  enumerate("abc") is an iterator returning (0,"a"), (1,"b"), (2,"c").
  The argument can be an arbitrary iterable object.

- The assert statement no longer tests __debug__ at runtime.  This means
  that assert statements cannot be disabled by assigning a false value
  to __debug__.

- A method zfill() was added to str and unicode, that fills a numeric
  string to the left with zeros.  For example,
  "+123".zfill(6) -> "+00123".

- Complex numbers supported divmod() and the // and % operators, but
  these make no sense.  Since this was documented, they're being
  deprecated now.

- String and unicode methods lstrip(), rstrip() and strip() now take
  an optional argument that specifies the characters to strip.  For
  example, "Foo!!!?!?!?".rstrip("?!") -> "Foo".

- There's a new dictionary constructor (a class method of the dict
  class), dict.fromkeys(iterable, value=None).  It constructs a
  dictionary with keys taken from the iterable and all values set to a
  single value.  It can be used for building sets and for removing
  duplicates from sequences.

- Added a new dict method pop(key).  This removes and returns the
  value corresponding to key.  [SF patch #539949]

- A new built-in type, bool, has been added, as well as built-in
  names for its two values, True and False.  Comparisons and sundry
  other operations that return a truth value have been changed to
  return a bool instead.  Read PEP 285 for an explanation of why this
  is backward compatible.

- Fixed two bugs reported as SF #535905: under certain conditions,
  deallocating a deeply nested structure could cause a segfault in the
  garbage collector, due to interaction with the "trashcan" code;
  access to the current frame during destruction of a local variable
  could access a pointer to freed memory.

- The optional object allocator ("pymalloc") has been enabled by
  default.  The recommended practice for memory allocation and
  deallocation has been streamlined.  A header file is included,
  Misc/pymemcompat.h, which can be bundled with 3rd party extensions
  and lets them use the same API with Python versions from 1.5.2
  onwards.

- PyErr_Display will provide file and line information for all exceptions
  that have an attribute print_file_and_line, not just SyntaxErrors.

- The UTF-8 codec will now encode and decode Unicode surrogates
  correctly and without raising exceptions for unpaired ones.

- Universal newlines (PEP 278) is implemented.  Briefly, using 'U'
  instead of 'r' when opening a text file for reading changes the line
  ending convention so that any of '\r', '\r\n', and '\n' is
  recognized (even mixed in one file); all three are converted to
  '\n', the standard Python line end character.

- file.xreadlines() now raises a ValueError if the file is closed:
  Previously, an xreadlines object was returned which would raise
  a ValueError when the xreadlines.next() method was called.

- sys.exit() inadvertently allowed more than one argument.
  An exception will now be raised if more than one argument is used.

- Changed evaluation order of dictionary literals to conform to the
  general left to right evaluation order rule. Now {f1(): f2()} will
  evaluate f1 first.

- Fixed bug #521782: when a file was in non-blocking mode, file.read()
  could silently lose data or wrongly throw an unknown error.

- The sq_repeat, sq_inplace_repeat, sq_concat and sq_inplace_concat
  slots are now always tried after trying the corresponding nb_* slots.
  This fixes a number of minor bugs (see bug #624807).

- Fix problem with dynamic loading on 64-bit AIX (see bug #639945).

Extension modules
-----------------

- Added three operators to the operator module:
    operator.pow(a,b) which is equivalent to:  a**b.
    operator.is_(a,b) which is equivalent to:  a is b.
    operator.is_not(a,b) which is equivalent to:  a is not b.

- posix.openpty now works on all systems that have /dev/ptmx.

- A module zipimport exists to support importing code from zip
  archives.

- The new datetime module supplies classes for manipulating dates and
  times.  The basic design came from the Zope "fishbowl process", and
  favors practical commercial applications over calendar esoterica.  See

      http://www.zope.org/Members/fdrake/DateTimeWiki/FrontPage

- _tkinter now returns Tcl objects, instead of strings. Objects which
  have Python equivalents are converted to Python objects, other objects
  are wrapped. This can be configured through the wantobjects method,
  or Tkinter.wantobjects.

- The PyBSDDB wrapper around the Sleepycat Berkeley DB library has
  been added as the package bsddb.  The traditional bsddb module is
  still available in source code, but not built automatically anymore,
  and is now named bsddb185.  This supports Berkeley DB versions from
  3.0 to 4.1.  For help converting your databases from the old module (which
  probably used an obsolete version of Berkeley DB) to the new module, see
  the db2pickle.py and pickle2db.py scripts described in the Tools/Demos
  section above.

- unicodedata was updated to Unicode 3.2. It supports normalization
  and names for Hangul syllables and CJK unified ideographs.

- resource.getrlimit() now returns longs instead of ints.

- readline now dynamically adjusts its input/output stream if
  sys.stdin/stdout changes.

- The _tkinter module (and hence Tkinter) has dropped support for
  Tcl/Tk 8.0 and 8.1.  Only Tcl/Tk versions 8.2, 8.3 and 8.4 are
  supported.

- cPickle.BadPickleGet is now a class.

- The time stamps in os.stat_result are floating point numbers
  after stat_float_times has been called.

- If the size passed to mmap.mmap() is larger than the length of the
  file on non-Windows platforms, a ValueError is raised. [SF bug 585792]

- The xreadlines module is slated for obsolescence.

- The strptime function in the time module is now always available (a
  Python implementation is used when the C library doesn't define it).

- The 'new' module is no longer an extension, but a Python module that
  only exists for backwards compatibility.  Its contents are no longer
  functions but callable type objects.

- The bsddb.*open functions can now take 'None' as a filename.
  This will create a temporary in-memory bsddb that won't be
  written to disk.

- posix.getloadavg, posix.lchown, posix.killpg, posix.mknod, and
  posix.getpgid have been added where available.

- The locale module now exposes the C library's gettext interface. It
  also has a new function getpreferredencoding.

- A security hole ("double free") was found in zlib-1.1.3, a popular
  third party compression library used by some Python modules.  The
  hole was quickly plugged in zlib-1.1.4, and the Windows build of
  Python now ships with zlib-1.1.4.

- pwd, grp, and resource return enhanced tuples now, with symbolic
  field names.

- array.array is now a type object. A new format character
  'u' indicates Py_UNICODE arrays. For those, .tounicode and
  .fromunicode methods are available. Arrays now support __iadd__
  and __imul__.

- dl now builds on every system that has dlfcn.h.  Failure in case
  of sizeof(int)!=sizeof(long)!=sizeof(void*) is delayed until dl.open
  is called.

- signal.sigpending, signal.sigprocmask and signal.sigsuspend have
  been added where available.

- The sys module acquired a new attribute, api_version, which evaluates
  to the value of the PYTHON_API_VERSION macro with which the
  interpreter was compiled.

- Fixed bug #470582: sre module would return a tuple (None, 'a', 'ab')
  when applying the regular expression '^((a)c)?(ab)$' on 'ab'. It now
  returns (None, None, 'ab'), as expected. Also fixed handling of
  lastindex/lastgroup match attributes in similar cases. For example,
  when running the expression r'(a)(b)?b' over 'ab', lastindex must be
  1, not 2.

- Fixed bug #581080: sre scanner was not checking the buffer limit
  before increasing the current pointer. This was creating an infinite
  loop in the search function, once the pointer exceeded the buffer
  limit.

- The os.fdopen function now enforces a file mode starting with the
  letter 'r', 'w' or 'a', otherwise a ValueError is raised. This fixes
  bug #623464.

- The linuxaudiodev module is now deprecated; it is being replaced by
  ossaudiodev.  The interface has been extended to cover a lot more of
  OSS (see www.opensound.com), including most DSP ioctls and the
  OSS mixer API.  Documentation forthcoming in 2.3a2.

Library
-------

- imaplib.py now supports SSL (Tino Lange and Piers Lauder).

- Freeze's modulefinder.py has been moved to the standard library;
  slightly improved so it will issue less false missing submodule
  reports (see sf path #643711 for details).  Documentation will follow
  with Python 2.3a2.

- os.path exposes getctime.

- unittest.py now has two additional methods called assertAlmostEqual()
  and failIfAlmostEqual().  They implement an approximate comparision
  by rounding the difference between the two arguments and comparing
  the result to zero.  Approximate comparision is essential for
  unit tests of floating point results.

- calendar.py now depends on the new datetime module rather than
  the time module.  As a result, the range of allowable dates
  has been increased.

- pdb has a new 'j(ump)' command to select the next line to be
  executed.

- The distutils created windows installers now can run a
  postinstallation script.

- doctest.testmod can now be called without argument, which means to
  test the current module.

- When cancelling a server that implemented threading with a keyboard
  interrupt, the server would shut down but not terminate (waiting on
  client threads). A new member variable, daemon_threads, was added to
  the ThreadingMixIn class in SocketServer.py to make it explicit that
  this behavior needs to be controlled.

- A new module, optparse, provides a fancy alternative to getopt for
  command line parsing.  It is a slightly modified version of Greg
  Ward's Optik package.

- UserDict.py now defines a DictMixin class which defines all dictionary
  methods for classes that already have a minimum mapping interface.
  This greatly simplifies writing classes that need to be substitutable
  for dictionaries (such as the shelve module).

- shelve.py now subclasses from UserDict.DictMixin.  Now shelve supports
  all dictionary methods.  This eases the transition to persistent
  storage for scripts originally written with dictionaries in mind.

- shelve.open and the various classes in shelve.py now accept an optional
  binary flag, which defaults to False.  If True, the values stored in the
  shelf are binary pickles.

- A new package, logging, implements the logging API defined by PEP
  282.  The code is written by Vinay Sajip.

- StreamReader, StreamReaderWriter and StreamRecoder in the codecs
  modules are iterators now.

- gzip.py now handles files exceeding 2GB.  Files over 4GB also work
  now (provided the OS supports it, and Python is configured with large
  file support), but in that case the underlying gzip file format can
  record only the least-significant 32 bits of the file size, so that
  some tools working with gzipped files may report an incorrect file
  size.

- xml.sax.saxutils.unescape has been added, to replace entity references
  with their entity value.

- Queue.Queue.{put,get} now support an optional timeout argument.

- Various features of Tk 8.4 are exposed in Tkinter.py. The multiple
  option of tkFileDialog is exposed as function askopenfile{,name}s.

- Various configure methods of Tkinter have been stream-lined, so that
  tag_configure, image_configure, window_configure now return a
  dictionary when invoked with no argument.

- Importing the readline module now no longer has the side effect of
  calling setlocale(LC_CTYPE, "").  The initial "C" locale, or
  whatever locale is explicitly set by the user, is preserved.  If you
  want repr() of 8-bit strings in your preferred encoding to preserve
  all printable characters of that encoding, you have to add the
  following code to your $PYTHONSTARTUP file or to your application's
  main():

    import locale
    locale.setlocale(locale.LC_CTYPE, "")

- shutil.move was added. shutil.copytree now reports errors as an
  exception at the end, instead of printing error messages.

- Encoding name normalization was generalized to not only
  replace hyphens with underscores, but also all other non-alphanumeric
  characters (with the exception of the dot which is used for Python
  package names during lookup). The aliases.py mapping was updated
  to the new standard.

- mimetypes has two new functions: guess_all_extensions() which
  returns a list of all known extensions for a mime type, and
  add_type() which adds one mapping between a mime type and
  an extension to the database.

- New module: sets, defines the class Set that implements a mutable
  set type using the keys of a dict to represent the set.  There's
  also a class ImmutableSet which is useful when you need sets of sets
  or when you need to use sets as dict keys, and a class BaseSet which
  is the base class of the two.

- Added random.sample(population,k) for random sampling without replacement.
  Returns a k length list of unique elements chosen from the population.

- random.randrange(-sys.maxint-1, sys.maxint) no longer raises
  OverflowError.  That is, it now accepts any combination of 'start'
  and 'stop' arguments so long as each is in the range of Python's
  bounded integers.

- Thanks to Raymond Hettinger, random.random() now uses a new core
  generator.  The Mersenne Twister algorithm is implemented in C,
  threadsafe, faster than the previous generator, has an astronomically
  large period (2**19937-1), creates random floats to full 53-bit
  precision, and may be the most widely tested random number generator
  in existence.

  The random.jumpahead(n) method has different semantics for the new
  generator.  Instead of jumping n steps ahead, it uses n and the
  existing state to create a new state.  This means that jumpahead()
  continues to support multi-threaded code needing generators of
  non-overlapping sequences.  However, it will break code which relies
  on jumpahead moving a specific number of steps forward.

  The attributes random.whseed and random.__whseed have no meaning for
  the new generator.  Code using these attributes should switch to a
  new class, random.WichmannHill which is provided for backward
  compatibility and to make an alternate generator available.

- New "algorithms" module: heapq, implements a heap queue.  Thanks to
  Kevin O'Connor for the code and François Pinard for an entertaining
  write-up explaining the theory and practical uses of heaps.

- New encoding for the Palm OS character set: palmos.

- binascii.crc32() and the zipfile module had problems on some 64-bit
  platforms.  These have been fixed.  On a platform with 8-byte C longs,
  crc32() now returns a signed-extended 4-byte result, so that its value
  as a Python int is equal to the value computed a 32-bit platform.

- xml.dom.minidom.toxml and toprettyxml now take an optional encoding
  argument.

- Some fixes in the copy module: when an object is copied through its
  __reduce__ method, there was no check for a __setstate__ method on
  the result [SF patch 565085]; deepcopy should treat instances of
  custom metaclasses the same way it treats instances of type 'type'
  [SF patch 560794].

- Sockets now support timeout mode.  After s.settimeout(T), where T is
  a float expressing seconds, subsequent operations raise an exception
  if they cannot be completed within T seconds.  To disable timeout
  mode, use s.settimeout(None).  There's also a module function,
  socket.setdefaulttimeout(T), which sets the default for all sockets
  created henceforth.

- getopt.gnu_getopt was added.  This supports GNU-style option
  processing, where options can be mixed with non-option arguments.

- Stop using strings for exceptions.  String objects used for
  exceptions are now classes deriving from Exception.  The objects
  changed were: Tkinter.TclError, bdb.BdbQuit, macpath.norm_error,
  tabnanny.NannyNag, and xdrlib.Error.

- Constants BOM_UTF8, BOM_UTF16, BOM_UTF16_LE, BOM_UTF16_BE,
  BOM_UTF32, BOM_UTF32_LE and BOM_UTF32_BE that represent the Byte
  Order Mark in UTF-8, UTF-16 and UTF-32 encodings for little and
  big endian systems were added to the codecs module. The old names
  BOM32_* and BOM64_* were off by a factor of 2.

- Added conversion functions math.degrees() and math.radians().

- math.log() now takes an optional argument:  math.log(x[, base]).

- ftplib.retrlines() now tests for callback is None rather than testing
  for False.  Was causing an error when given a callback object which
  was callable but also returned len() as zero.  The change may
  create new breakage if the caller relied on the undocumented behavior
  and called with callback set to [] or some other False value not
  identical to None.

- random.gauss() uses a piece of hidden state used by nothing else,
  and the .seed() and .whseed() methods failed to reset it.  In other
  words, setting the seed didn't completely determine the sequence of
  results produced by random.gauss().  It does now.  Programs repeatedly
  mixing calls to a seed method with calls to gauss() may see different
  results now.

- The pickle.Pickler class grew a clear_memo() method to mimic that
  provided by cPickle.Pickler.

- difflib's SequenceMatcher class now does a dynamic analysis of
  which elements are so frequent as to constitute noise.  For
  comparing files as sequences of lines, this generally works better
  than the IS_LINE_JUNK function, and function ndiff's linejunk
  argument defaults to None now as a result.  A happy benefit is
  that SequenceMatcher may run much faster now when applied
  to large files with many duplicate lines (for example, C program
  text with lots of repeated "}" and "return NULL;" lines).

- New Text.dump() method in Tkinter module.

- New distutils commands for building packagers were added to
  support pkgtool on Solaris and swinstall on HP-UX.

- distutils now has a new abstract binary packager base class
  command/bdist_packager, which simplifies writing packagers.
  This will hopefully provide the missing bits to encourage
  people to submit more packagers, e.g. for Debian, FreeBSD
  and other systems.

- The UTF-16, -LE and -BE stream readers now raise a
  NotImplementedError for all calls to .readline(). Previously, they
  used to just produce garbage or fail with an encoding error --
  UTF-16 is a 2-byte encoding and the C lib's line reading APIs don't
  work well with these.

- compileall now supports quiet operation.

- The BaseHTTPServer now implements optional HTTP/1.1 persistent
  connections.

- socket module: the SSL support was broken out of the main
  _socket module C helper and placed into a new _ssl helper
  which now gets imported by socket.py if available and working.

- encodings package: added aliases for all supported IANA character
  sets

- ftplib: to safeguard the user's privacy, anonymous login will use
  "anonymous@" as default password, rather than the real user and host
  name.

- webbrowser: tightened up the command passed to os.system() so that
  arbitrary shell code can't be executed because a bogus URL was
  passed in.

- gettext.translation has an optional fallback argument, and
  gettext.find an optional all argument. Translations will now fallback
  on a per-message basis. The module supports plural forms, by means
  of gettext.[d]ngettext and Translation.[u]ngettext.

- distutils bdist commands now offer a --skip-build option.

- warnings.warn now accepts a Warning instance as first argument.

- The xml.sax.expatreader.ExpatParser class will no longer create
  circular references by using itself as the locator that gets passed
  to the content handler implementation.  [SF bug #535474]

- The email.Parser.Parser class now properly parses strings regardless
  of their line endings, which can be any of \r, \n, or \r\n (CR, LF,
  or CRLF).  Also, the Header class's constructor default arguments
  has changed slightly so that an explicit maxlinelen value is always
  honored, and so unicode conversion error handling can be specified.

- distutils' build_ext command now links C++ extensions with the C++
  compiler available in the Makefile or CXX environment variable, if
  running under \*nix.

- New module bz2: provides a comprehensive interface for the bz2 compression
  library.  It implements a complete file interface, one-shot (de)compression
  functions, and types for sequential (de)compression.

- New pdb command 'pp' which is like 'p' except that it pretty-prints
  the value of its expression argument.

- Now bdist_rpm distutils command understands a verify_script option in
  the config file, including the contents of the referred filename in
  the "%verifyscript" section of the rpm spec file.

- Fixed bug #495695: webbrowser module would run graphic browsers in a
  unix environment even if DISPLAY was not set. Also, support for
  skipstone browser was included.

- Fixed bug #636769: rexec would run unallowed code if subclasses of
  strings were used as parameters for certain functions.

Tools/Demos
-----------

- pygettext.py now supports globbing on Windows, and accepts module
  names in addition to accepting file names.

- The SGI demos (Demo/sgi) have been removed.  Nobody thought they
  were interesting any more.  (The SGI library modules and extensions
  are still there; it is believed that at least some of these are
  still used and useful.)

- IDLE supports the new encoding declarations (PEP 263); it can also
  deal with legacy 8-bit files if they use the locale's encoding. It
  allows non-ASCII strings in the interactive shell and executes them
  in the locale's encoding.

- freeze.py now produces binaries which can import shared modules,
  unlike before when this failed due to missing symbol exports in
  the generated binary.

Build
-----

- On Unix, IDLE is now installed automatically.

- The fpectl module is not built by default; it's dangerous or useless
  except in the hands of experts.

- The public Python C API will generally be declared using PyAPI_FUNC
  and PyAPI_DATA macros, while Python extension module init functions
  will be declared with PyMODINIT_FUNC.  DL_EXPORT/DL_IMPORT macros
  are deprecated.

- A bug was fixed that could cause COUNT_ALLOCS builds to segfault, or
  get into infinite loops, when a new-style class got garbage-collected.
  Unfortunately, to avoid this, the way COUNT_ALLOCS works requires
  that new-style classes be immortal in COUNT_ALLOCS builds.  Note that
  COUNT_ALLOCS is not enabled by default, in either release or debug
  builds, and that new-style classes are immortal only in COUNT_ALLOCS
  builds.

- Compiling out the cyclic garbage collector is no longer an option.
  The old symbol WITH_CYCLE_GC is now ignored, and Python.h arranges
  that it's always defined (for the benefit of any extension modules
  that may be conditionalizing on it).  A bonus is that any extension
  type participating in cyclic gc can choose to participate in the
  Py_TRASHCAN mechanism now too; in the absence of cyclic gc, this used
  to require editing the core to teach the trashcan mechanism about the
  new type.

- According to Annex F of the current C standard,

    The Standard C macro HUGE_VAL and its float and long double analogs,
    HUGE_VALF and HUGE_VALL, expand to expressions whose values are
    positive infinities.

  Python only uses the double HUGE_VAL, and only to #define its own symbol
  Py_HUGE_VAL.  Some platforms have incorrect definitions for HUGE_VAL.
  pyport.h used to try to worm around that, but the workarounds triggered
  other bugs on other platforms, so we gave up.  If your platform defines
  HUGE_VAL incorrectly, you'll need to #define Py_HUGE_VAL to something
  that works on your platform.  The only instance of this I'm sure about
  is on an unknown subset of Cray systems, described here:

  http://www.cray.com/swpubs/manuals/SN-2194_2.0/html-SN-2194_2.0/x3138.htm

  Presumably 2.3a1 breaks such systems.  If anyone uses such a system, help!

- The configure option --without-doc-strings can be used to remove the
  doc strings from the builtin functions and modules; this reduces the
  size of the executable.

- The universal newlines option (PEP 278) is on by default.  On Unix
  it can be disabled by passing --without-universal-newlines to the
  configure script.  On other platforms, remove
  WITH_UNIVERSAL_NEWLINES from pyconfig.h.

- On Unix, a shared libpython2.3.so can be created with --enable-shared.

- All uses of the CACHE_HASH, INTERN_STRINGS, and DONT_SHARE_SHORT_STRINGS
  preprocessor symbols were eliminated.  The internal decisions they
  controlled stopped being experimental long ago.

- The tools used to build the documentation now work under Cygwin as
  well as Unix.

- The bsddb and dbm module builds have been changed to try and avoid version
  skew problems and disable linkage with Berkeley DB 1.85 unless the
  installer knows what s/he's doing.  See the section on building these
  modules in the README file for details.

C API
-----

- PyNumber_Check() now returns true for string and unicode objects.
  This is a result of these types having a partially defined
  tp_as_number slot.  (This is not a feature, but an indication that
  PyNumber_Check() is not very useful to determine numeric behavior.
  It may be deprecated.)

- The string object's layout has changed: the pointer member
  ob_sinterned has been replaced by an int member ob_sstate.  On some
  platforms (e.g. most 64-bit systems) this may change the offset of
  the ob_sval member, so as a precaution the API_VERSION has been
  incremented.  The apparently unused feature of "indirect interned
  strings", supported by the ob_sinterned member, is gone.  Interned
  strings are now usually mortal; theres a new API,
  PyString_InternImmortal() that creates immortal interned strings.
  (The ob_sstate member can only take three values; however, while
  making it a char saves a few bytes per string object on average, in
  it also slowed things down a bit because ob_sval was no longer
  aligned.)

- The Py_InitModule*() functions now accept NULL for the 'methods'
  argument.  Modules without global functions are becoming more common
  now that factories can be types rather than functions.

- New C API PyUnicode_FromOrdinal() which exposes unichr() at C
  level.

- New functions PyErr_SetExcFromWindowsErr() and
  PyErr_SetExcFromWindowsErrWithFilename(). Similar to
  PyErr_SetFromWindowsErrWithFilename() and
  PyErr_SetFromWindowsErr(), but they allow to specify
  the exception type to raise. Available on Windows.

- Py_FatalError() is now declared as taking a const char* argument.  It
  was previously declared without const.  This should not affect working
  code.

- Added new macro PySequence_ITEM(o, i) that directly calls
  sq_item without rechecking that o is a sequence and without
  adjusting for negative indices.

- PyRange_New() now raises ValueError if the fourth argument is not 1.
  This is part of the removal of deprecated features of the xrange
  object.

- PyNumber_Coerce() and PyNumber_CoerceEx() now also invoke the type's
  coercion if both arguments have the same type but this type has the
  CHECKTYPES flag set.  This is to better support proxies.

- The type of tp_free has been changed from "``void (*)(PyObject *)``" to
  "``void (*)(void *)``".

- PyObject_Del, PyObject_GC_Del are now functions instead of macros.

- A type can now inherit its metatype from its base type.  Previously,
  when PyType_Ready() was called, if ob_type was found to be NULL, it
  was always set to &PyType_Type; now it is set to base->ob_type,
  where base is tp_base, defaulting to &PyObject_Type.

- PyType_Ready() accidentally did not inherit tp_is_gc; now it does.

- The PyCore_* family of APIs have been removed.

- The "u#" parser marker will now pass through Unicode objects as-is
  without going through the buffer API.

- The enumerators of cmp_op have been renamed to use the prefix ``PyCmp_``.

- An old #define of ANY as void has been removed from pyport.h.  This
  hasn't been used since Python's pre-ANSI days, and the #define has
  been marked as obsolete since then.  SF bug 495548 says it created
  conflicts with other packages, so keeping it around wasn't harmless.

- Because Python's magic number scheme broke on January 1st, we decided
  to stop Python development.  Thanks for all the fish!

- Some of us don't like fish, so we changed Python's magic number
  scheme to a new one. See Python/import.c for details.

New platforms
-------------

- OpenVMS is now supported.

- AtheOS is now supported.

- the EMX runtime environment on OS/2 is now supported.

- GNU/Hurd is now supported.

Tests
-----

- The regrtest.py script's -u option now provides a way to say "allow
  all resources except this one."  For example, to allow everything
  except bsddb, give the option '-uall,-bsddb'.

Windows
-------

- The Windows distribution now ships with version 4.0.14 of the
  Sleepycat Berkeley database library.  This should be a huge
  improvement over the previous Berkeley DB 1.85, which had many
  bugs.
  XXX What are the licensing issues here?
  XXX If a user has a database created with a previous verion of
  XXX     Python, what must they do to convert it?
  XXX I'm still not sure how to link this thing (see PCbuild/readme.txt).
  XXX The version # is likely to change before 2.3a1.

- The Windows distribution now ships with a Secure Sockets Library (SLL)
   module (_ssl.pyd)

- The Windows distribution now ships with Tcl/Tk version 8.4.1 (it
  previously shipped with Tcl/Tk 8.3.2).

- When Python is built under a Microsoft compiler, sys.version now
  includes the compiler version number (_MSC_VER).  For example, under
  MSVC 6, sys.version constains the substring "MSC v.1200 ".  1200 is
  the value of _MSC_VER under MSVC 6.

- Sometimes the uninstall executable (UNWISE.EXE) vanishes.  One cause
  of that has been fixed in the installer (disabled Wise's "delete in-
  use files" uninstall option).

- Fixed a bug in urllib's proxy handling in Windows.  [SF bug #503031]

- The installer now installs Start menu shortcuts under (the local
  equivalent of) "All Users" when doing an Admin install.

- file.truncate([newsize]) now works on Windows for all newsize values.
  It used to fail if newsize didn't fit in 32 bits, reflecting a
  limitation of MS _chsize (which is no longer used).

- os.waitpid() is now implemented for Windows, and can be used to block
  until a specified process exits.  This is similar to, but not exactly
  the same as, os.waitpid() on POSIX systems.  If you're waiting for
  a specific process whose pid was obtained from one of the spawn()
  functions, the same Python os.waitpid() code works across platforms.
  See the docs for details.  The docs were changed to clarify that
  spawn functions return, and waitpid requires, a process handle on
  Windows (not the same thing as a Windows process id).

- New tempfile.TemporaryFile implementation for Windows:  this doesn't
  need a TemporaryFileWrapper wrapper anymore, and should be immune
  to a nasty problem:  before 2.3, if you got a temp file on Windows, it
  got wrapped in an object whose close() method first closed the
  underlying file, then deleted the file.  This usually worked fine.
  However, the spawn family of functions on Windows create (at a low C
  level) the same set of open files in the spawned process Q as were
  open in the spawning process P.  If a temp file f was among them, then
  doing f.close() in P first closed P's C-level file handle on f, but Q's
  C-level file handle on f remained open, so the attempt in P to delete f
  blew up with a "Permission denied" error (Windows doesn't allow
  deleting open files).  This was surprising, subtle, and difficult to
  work around.

- The os module now exports all the symbolic constants usable with the
  low-level os.open() on Windows:  the new constants in 2.3 are
  O_NOINHERIT, O_SHORT_LIVED, O_TEMPORARY, O_RANDOM and O_SEQUENTIAL.
  The others were also available in 2.2:  O_APPEND, O_BINARY, O_CREAT,
  O_EXCL, O_RDONLY, O_RDWR, O_TEXT, O_TRUNC and O_WRONLY.  Contrary
  to Microsoft docs, O_SHORT_LIVED does not seem to imply O_TEMPORARY
  (so specify both if you want both; note that neither is useful unless
  specified with O_CREAT too).

Mac
----

- Mac/Relnotes is gone, the release notes are now here.

- Python (the OSX-only, unix-based version, not the OS9-compatible CFM
  version) now fully supports unicode strings as arguments to various file
  system calls, eg. open(), file(), os.stat() and os.listdir().

- The current naming convention for Python on the Macintosh is that MacPython
  refers to the unix-based OSX-only version, and MacPython-OS9 refers to the
  CFM-based version that runs on both OS9 and OSX.

- All MacPython-OS9 functionality is now available in an OSX unix build,
  including the Carbon modules, the IDE, OSA support, etc. A lot of this
  will only work correctly in a framework build, though, because you cannot
  talk to the window manager unless your application is run from a .app
  bundle. There is a command line tool "pythonw" that runs your script
  with an interpreter living in such a .app bundle, this interpreter should
  be used to run any Python script using the window manager (including
  Tkinter or wxPython scripts).

- Most of Mac/Lib has moved to Lib/plat-mac, which is again used both in
  MacPython-OSX and MacPython-OS9. The only modules remaining in Mac/Lib
  are specifically for MacPython-OS9 (CFM support, preference resources, etc).

- A new utility PythonLauncher will start a Python interpreter when a .py or
  .pyw script is double-clicked in the Finder. By default .py scripts are
  run with a normal Python interpreter in a Terminal window and .pyw
  files are run with a window-aware pythonw interpreter without a Terminal
  window, but all this can be customized.

- MacPython-OS9 is now Carbon-only, so it runs on Mac OS 9 or Mac OS X and
  possibly on Mac OS 8.6 with the right CarbonLib installed, but not on earlier
  releases.

- Many tools such as BuildApplet.py and gensuitemodule.py now support a command
  line interface too.

- All the Carbon classes are now PEP253 compliant, meaning that you can
  subclass them from Python. Most of the attributes have gone, you should
  now use the accessor function call API, which is also what Apple's
  documentation uses. Some attributes such as grafport.visRgn are still
  available for convenience.

- New Carbon modules File (implementing the APIs in Files.h and Aliases.h)
  and Folder (APIs from Folders.h). The old macfs builtin module is
  gone, and replaced by a Python wrapper around the new modules.

- Pathname handling should now be fully consistent: MacPython-OSX always uses
  unix pathnames and MacPython-OS9 always uses colon-separated Mac pathnames
  (also when running on Mac OS X).

- New Carbon modules Help and AH give access to the Carbon Help Manager.
  There are hooks in the IDE to allow accessing the Python documentation
  (and Apple's Carbon and Cocoa documentation) through the Help Viewer.
  See Mac/OSX/README for converting the Python documentation to a
  Help Viewer comaptible form and installing it.

- OSA support has been redesigned and the generated Python classes now
  mirror the inheritance defined by the underlying OSA classes.

- MacPython no longer maps both \r and \n to \n on input for any text file.
  This feature has been replaced by universal newline support (PEP278).

- The default encoding for Python sourcefiles in MacPython-OS9 is no longer
  mac-roman (or whatever your local Mac encoding was) but "ascii", like on
  other platforms. If you really need sourcefiles with Mac characters in them
  you can change this in site.py.


What's New in Python 2.2 final?
===============================

*Release date: 21-Dec-2001*

Type/class unification and new-style classes
--------------------------------------------

- pickle.py, cPickle: allow pickling instances of new-style classes
  with a custom metaclass.

Core and builtins
-----------------

- weakref proxy object: when comparing, unwrap both arguments if both
  are proxies.

Extension modules
-----------------

- binascii.b2a_base64(): fix a potential buffer overrun when encoding
  very short strings.

- cPickle: the obscure "fast" mode was suspected of causing stack
  overflows on the Mac.  Hopefully fixed this by setting the recursion
  limit much smaller.  If the limit is too low (it only affects
  performance), you can change it by defining PY_CPICKLE_FAST_LIMIT
  when compiling cPickle.c (or in pyconfig.h).

Library
-------

- dumbdbm.py: fixed a dumb old bug (the file didn't get synched at
  close or delete time).

- rfc822.py: fixed a bug where the address '<>' was converted to None
  instead of an empty string (also fixes the email.Utils module).

- xmlrpclib.py: version 1.0.0; uses precision for doubles.

- test suite: the pickle and cPickle tests were not executing any code
  when run from the standard regression test.

Tools/Demos
-----------

Build
-----

C API
-----

New platforms
-------------

Tests
-----

Windows
-------

- distutils package: fixed broken Windows installers (bdist_wininst).

- tempfile.py: prevent mysterious warnings when TemporaryFileWrapper
  instances are deleted at process exit time.

- socket.py: prevent mysterious warnings when socket instances are
  deleted at process exit time.

- posixmodule.c: fix a Windows crash with stat() of a filename ending
  in backslash.

Mac
----

- The Carbon toolbox modules have been upgraded to Universal Headers
  3.4, and experimental CoreGraphics and CarbonEvents modules have
  been added.  All only for framework-enabled MacOSX.


What's New in Python 2.2c1?
===========================

*Release date: 14-Dec-2001*

Type/class unification and new-style classes
--------------------------------------------

- Guido's tutorial introduction to the new type/class features has
  been extensively updated.  See

      http://www.python.org/2.2/descrintro.html

  That remains the primary documentation in this area.

- Fixed a leak: instance variables declared with __slots__ were never
  deleted!

- The "delete attribute" method of descriptor objects is called
  __delete__, not __del__.  In previous releases, it was mistakenly
  called __del__, which created an unfortunate overloading condition
  with finalizers.  (The "get attribute" and "set attribute" methods
  are still called __get__ and __set__, respectively.)

- Some subtle issues with the super built-in were fixed:

  (a) When super itself is subclassed, its __get__ method would still
      return an instance of the base class (i.e., of super).

  (b) super(C, C()).__class__ would return C rather than super.  This
      is confusing.  To fix this, I decided to change the semantics of
      super so that it only applies to code attributes, not to data
      attributes.  After all, overriding data attributes is not
      supported anyway.

  (c) The __get__ method didn't check whether the argument was an
      instance of the type used in creation of the super instance.

- Previously, hash() of an instance of a subclass of a mutable type
  (list or dictionary) would return some value, rather than raising
  TypeError.  This has been fixed.  Also, directly calling
  dict.__hash__ and list.__hash__ now raises the same TypeError
  (previously, these were the same as object.__hash__).

- New-style objects now support deleting their __dict__.  This is for
  all intents and purposes equivalent to assigning a brand new empty
  dictionary, but saves space if the object is not used further.

Core and builtins
-----------------

- -Qnew now works as documented in PEP 238:  when -Qnew is passed on
  the command line, all occurrences of "/" use true division instead
  of classic division.  See the PEP for details.  Note that "all"
  means all instances in library and 3rd-party modules, as well as in
  your own code.  As the PEP says, -Qnew is intended for use only in
  educational environments with control over the libraries in use.
  Note that test_coercion.py in the standard Python test suite fails
  under -Qnew; this is expected, and won't be repaired until true
  division becomes the default (in the meantime, test_coercion is
  testing the current rules).

- complex() now only allows the first argument to be a string
  argument, and raises TypeError if either the second arg is a string
  or if the second arg is specified when the first is a string.

Extension modules
-----------------

- gc.get_referents was renamed to gc.get_referrers.

Library
-------

- Functions in the os.spawn() family now release the global interpreter
  lock around calling the platform spawn.  They should always have done
  this, but did not before 2.2c1.  Multithreaded programs calling
  an os.spawn function with P_WAIT will no longer block all Python threads
  until the spawned program completes.  It's possible that some programs
  relies on blocking, although more likely by accident than by design.

- webbrowser defaults to netscape.exe on OS/2 now.

- Tix.ResizeHandle exposes detach_widget, hide, and show.

- The charset alias windows_1252 has been added.

- types.StringTypes is a tuple containing the defined string types;
  usually this will be (str, unicode), but if Python was compiled
  without Unicode support it will be just (str,).

- The pulldom and minidom modules were synchronized to PyXML.

Tools/Demos
-----------

- A new script called Tools/scripts/google.py was added, which fires
  off a search on Google.

Build
-----

- Note that release builds of Python should arrange to define the
  preprocessor symbol NDEBUG on the command line (or equivalent).
  In the 2.2 pre-release series we tried to define this by magic in
  Python.h instead, but it proved to cause problems for extension
  authors.  The Unix, Windows and Mac builds now all define NDEBUG in
  release builds via cmdline (or equivalent) instead.  Ports to
  other platforms should do likewise.

- It is no longer necessary to use --with-suffix when building on a
  case-insensitive file system (such as Mac OS X HFS+). In the build
  directory an extension is used, but not in the installed python.

C API
-----

- New function PyDict_MergeFromSeq2() exposes the builtin dict
  constructor's logic for updating a dictionary from an iterable object
  producing key-value pairs.

- PyArg_ParseTupleAndKeywords() requires that the number of entries in
  the keyword list equal the number of argument specifiers.  This
  wasn't checked correctly, and PyArg_ParseTupleAndKeywords could even
  dump core in some bad cases.  This has been repaired.  As a result,
  PyArg_ParseTupleAndKeywords may raise RuntimeError in bad cases that
  previously went unchallenged.

New platforms
-------------

Tests
-----

Windows
-------

Mac
----

- In unix-Python on Mac OS X (and darwin) sys.platform is now "darwin",
  without any trailing digits.

- Changed logic for finding python home in Mac OS X framework Pythons.
  Now sys.executable points to the executable again, in stead of to
  the shared library. The latter is used only for locating the python
  home.


What's New in Python 2.2b2?
===========================

*Release date: 16-Nov-2001*

Type/class unification and new-style classes
--------------------------------------------

- Multiple inheritance mixing new-style and classic classes in the
  list of base classes is now allowed, so this works now:

      class Classic: pass
      class Mixed(Classic, object): pass

  The MRO (method resolution order) for each base class is respected
  according to its kind, but the MRO for the derived class is computed
  using new-style MRO rules if any base class is a new-style class.
  This needs to be documented.

- The new builtin dictionary() constructor, and dictionary type, have
  been renamed to dict.  This reflects a decade of common usage.

- dict() now accepts an iterable object producing 2-sequences.  For
  example, dict(d.items()) == d for any dictionary d.  The argument,
  and the elements of the argument, can be any iterable objects.

- New-style classes can now have a __del__ method, which is called
  when the instance is deleted (just like for classic classes).

- Assignment to object.__dict__ is now possible, for objects that are
  instances of new-style classes that have a __dict__ (unless the base
  class forbids it).

- Methods of built-in types now properly check for keyword arguments
  (formerly these were silently ignored).  The only built-in methods
  that take keyword arguments are __call__, __init__ and __new__.

- The socket function has been converted to a type; see below.

Core and builtins
-----------------

- Assignment to __debug__ raises SyntaxError at compile-time.  This
  was promised when 2.1c1 was released as "What's New in Python 2.1c1"
  (see below) says.

- Clarified the error messages for unsupported operands to an operator
  (like 1 + '').

Extension modules
-----------------

- mmap has a new keyword argument, "access", allowing a uniform way for
  both Windows and Unix users to create read-only, write-through and
  copy-on-write memory mappings.  This was previously possible only on
  Unix.  A new keyword argument was required to support this in a
  uniform way because the mmap() signatures had diverged across
  platforms.  Thanks to Jay T Miller for repairing this!

- By default, the gc.garbage list now contains only those instances in
  unreachable cycles that have __del__ methods; in 2.1 it contained all
  instances in unreachable cycles.  "Instances" here has been generalized
  to include instances of both new-style and old-style classes.

- The socket module defines a new method for socket objects,
  sendall().  This is like send() but may make multiple calls to
  send() until all data has been sent.  Also, the socket function has
  been converted to a subclassable type, like list and tuple (etc.)
  before it; socket and SocketType are now the same thing.

- Various bugfixes to the curses module.  There is now a test suite
  for the curses module (you have to run it manually).

- binascii.b2a_base64 no longer places an arbitrary restriction of 57
  bytes on its input.

Library
-------

- tkFileDialog exposes a Directory class and askdirectory
  convenience function.

- Symbolic group names in regular expressions must be unique.  For
  example, the regexp r'(?P<abc>)(?P<abc>)' is not allowed, because a
  single name can't mean both "group 1" and "group 2" simultaneously.
  Python 2.2 detects this error at regexp compilation time;
  previously, the error went undetected, and results were
  unpredictable.  Also in sre, the pattern.split(), pattern.sub(), and
  pattern.subn() methods have been rewritten in C.  Also, an
  experimental function/method finditer() has been added, which works
  like findall() but returns an iterator.

- Tix exposes more commands through the classes DirSelectBox,
  DirSelectDialog, ListNoteBook, Meter, CheckList, and the
  methods tix_addbitmapdir, tix_cget, tix_configure, tix_filedialog,
  tix_getbitmap, tix_getimage, tix_option_get, and tix_resetoptions.

- Traceback objects are now scanned by cyclic garbage collection, so
  cycles created by casual use of sys.exc_info() no longer cause
  permanent memory leaks (provided garbage collection is enabled).

- os.extsep -- a new variable needed by the RISCOS support.  It is the
  separator used by extensions, and is '.' on all platforms except
  RISCOS, where it is '/'.  There is no need to use this variable
  unless you have a masochistic desire to port your code to RISCOS.

- mimetypes.py has optional support for non-standard, but commonly
  found types.  guess_type() and guess_extension() now accept an
  optional 'strict' flag, defaulting to true, which controls whether
  recognize non-standard types or not.  A few non-standard types we
  know about have been added.  Also, when run as a script, there are
  new -l and -e options.

- statcache is now deprecated.

- email.Utils.formatdate() now produces the preferred RFC 2822 style
  dates with numeric timezones (it used to produce obsolete dates
  hard coded to "GMT" timezone).  An optional 'localtime' flag is
  added to produce dates in the local timezone, with daylight savings
  time properly taken into account.

- In pickle and cPickle, instead of masking errors in load() by
  transforming them into SystemError, we let the original exception
  propagate out.  Also, implement support for __safe_for_unpickling__
  in pickle, as it already was supported in cPickle.

Tools/Demos
-----------

Build
-----

- The dbm module is built using libdb1 if available.  The bsddb module
  is built with libdb3 if available.

- Misc/Makefile.pre.in has been removed by BDFL pronouncement.

C API
-----

- New function PySequence_Fast_GET_SIZE() returns the size of a non-
  NULL result from PySequence_Fast(), more quickly than calling
  PySequence_Size().

- New argument unpacking function PyArg_UnpackTuple() added.

- New functions PyObject_CallFunctionObjArgs() and
  PyObject_CallMethodObjArgs() have been added to make it more
  convenient and efficient to call functions and methods from C.

- PyArg_ParseTupleAndKeywords() no longer masks errors, so it's
  possible that this will propagate errors it didn't before.

- New function PyObject_CheckReadBuffer(), which returns true if its
  argument supports the single-segment readable buffer interface.

New platforms
-------------

- We've finally confirmed that this release builds on HP-UX 11.00,
  *with* threads, and passes the test suite.

- Thanks to a series of patches from Michael Muller, Python may build
  again under OS/2 Visual Age C++.

- Updated RISCOS port by Dietmar Schwertberger.

Tests
-----

- Added a test script for the curses module.  It isn't run automatically;
  regrtest.py must be run with '-u curses' to enable it.

Windows
-------

Mac
----

- PythonScript has been moved to unsupported and is slated to be
  removed completely in the next release.

- It should now be possible to build applets that work on both OS9 and
  OSX.

- The core is now linked with CoreServices not Carbon; as a side
  result, default 8bit encoding on OSX is now ASCII.

- Python should now build on OSX 10.1.1


What's New in Python 2.2b1?
===========================

*Release date: 19-Oct-2001*

Type/class unification and new-style classes
--------------------------------------------

- New-style classes are now always dynamic (except for built-in and
  extension types).  There is no longer a performance penalty, and I
  no longer see another reason to keep this baggage around.  One relic
  remains: the __dict__ of a new-style class is a read-only proxy; you
  must set the class's attribute to modify it.  As a consequence, the
  __defined__ attribute of new-style types no longer exists, for lack
  of need: there is once again only one __dict__ (although in the
  future a __cache__ may be resurrected with a similar function, if I
  can prove that it actually speeds things up).

- C.__doc__ now works as expected for new-style classes (in 2.2a4 it
  always returned None, even when there was a class docstring).

- doctest now finds and runs docstrings attached to new-style classes,
  class methods, static methods, and properties.

Core and builtins
-----------------

- A very subtle syntactical pitfall in list comprehensions was fixed.
  For example: [a+b for a in 'abc', for b in 'def'].  The comma in
  this example is a mistake.  Previously, this would silently let 'a'
  iterate over the singleton tuple ('abc',), yielding ['abcd', 'abce',
  'abcf'] rather than the intended ['ad', 'ae', 'af', 'bd', 'be',
  'bf', 'cd', 'ce', 'cf'].  Now, this is flagged as a syntax error.
  Note that [a for a in <singleton>] is a convoluted way to say
  [<singleton>] anyway, so it's not like any expressiveness is lost.

- getattr(obj, name, default) now only catches AttributeError, as
  documented, rather than returning the default value for all
  exceptions (which could mask bugs in a __getattr__ hook, for
  example).

- Weak reference objects are now part of the core and offer a C API.
  A bug which could allow a core dump when binary operations involved
  proxy reference has been fixed.  weakref.ReferenceError is now a
  built-in exception.

- unicode(obj) now behaves more like str(obj), accepting arbitrary
  objects, and calling a __unicode__ method if it exists.
  unicode(obj, encoding) and unicode(obj, encoding, errors) still
  require an 8-bit string or character buffer argument.

- isinstance() now allows any object as the first argument and a
  class, a type or something with a __bases__ tuple attribute for the
  second argument.  The second argument may also be a tuple of a
  class, type, or something with __bases__, in which case isinstance()
  will return true if the first argument is an instance of any of the
  things contained in the second argument tuple.  E.g.

  isinstance(x, (A, B))

  returns true if x is an instance of A or B.

Extension modules
-----------------

- thread.start_new_thread() now returns the thread ID (previously None).

- binascii has now two quopri support functions, a2b_qp and b2a_qp.

- readline now supports setting the startup_hook and the
  pre_event_hook, and adds the add_history() function.

- os and posix supports chroot(), setgroups() and unsetenv() where
  available.  The stat(), fstat(), statvfs() and fstatvfs() functions
  now return "pseudo-sequences" -- the various fields can now be
  accessed as attributes (e.g. os.stat("/").st_mtime) but for
  backwards compatibility they also behave as a fixed-length sequence.
  Some platform-specific fields (e.g. st_rdev) are only accessible as
  attributes.

- time: localtime(), gmtime() and strptime() now return a
  pseudo-sequence similar to the os.stat() return value, with
  attributes like tm_year etc.

- Decompression objects in the zlib module now accept an optional
  second parameter to decompress() that specifies the maximum amount
  of memory to use for the uncompressed data.

- optional SSL support in the socket module now exports OpenSSL
  functions RAND_add(), RAND_egd(), and RAND_status().  These calls
  are useful on platforms like Solaris where OpenSSL does not
  automatically seed its PRNG.  Also, the keyfile and certfile
  arguments to socket.ssl() are now optional.

- posixmodule (and by extension, the os module on POSIX platforms) now
  exports O_LARGEFILE, O_DIRECT, O_DIRECTORY, and O_NOFOLLOW.

Library
-------

- doctest now excludes functions and classes not defined by the module
  being tested, thanks to Tim Hochberg.

- HotShot, a new profiler implemented using a C-based callback, has
  been added.  This substantially reduces the overhead of profiling,
  but it is still quite preliminary.  Support modules and
  documentation will be added in upcoming releases (before 2.2 final).

- profile now produces correct output in situations where an exception
  raised in Python is cleared by C code (e.g. hasattr()).  This used
  to cause wrong output, including spurious claims of recursive
  functions and attribution of time spent to the wrong function.

  The code and documentation for the derived OldProfile and HotProfile
  profiling classes was removed.  The code hasn't worked for years (if
  you tried to use them, they raised exceptions).  OldProfile
  intended to reproduce the behavior of the profiler Python used more
  than 7 years ago, and isn't interesting anymore.  HotProfile intended
  to provide a faster profiler (but producing less information), and
  that's a worthy goal we intend to meet via a different approach (but
  without losing information).

- Profile.calibrate() has a new implementation that should deliver
  a much better system-specific calibration constant.  The constant can
  now be specified in an instance constructor, or as a Profile class or
  instance variable, instead of by editing profile.py's source code.
  Calibration must still be done manually (see the docs for the profile
  module).

  Note that Profile.calibrate() must be overriden by subclasses.
  Improving the accuracy required exploiting detailed knowledge of
  profiler internals; the earlier method abstracted away the details
  and measured a simplified model instead, but consequently computed
  a constant too small by a factor of 2 on some modern machines.

- quopri's encode and decode methods take an optional header parameter,
  which indicates whether output is intended for the header 'Q'
  encoding.

- The SocketServer.ThreadingMixIn class now closes the request after
  finish_request() returns.  (Not when it errors out though.)

- The nntplib module's NNTP.body() method has grown a 'file' argument
  to allow saving the message body to a file.

- The email package has added a class email.Parser.HeaderParser which
  only parses headers and does not recurse into the message's body.
  Also, the module/class MIMEAudio has been added for representing
  audio data (contributed by Anthony Baxter).

- ftplib should be able to handle files > 2GB.

- ConfigParser.getboolean() now also interprets TRUE, FALSE, YES, NO,
  ON, and OFF.

- xml.dom.minidom NodeList objects now support the length attribute
  and item() method as required by the DOM specifications.

Tools/Demos
-----------

- Demo/dns was removed.  It no longer serves any purpose; a package
  derived from it is now maintained by Anthony Baxter, see
  http://PyDNS.SourceForge.net.

- The freeze tool has been made more robust, and two new options have
  been added: -X and -E.

Build
-----

- configure will use CXX in LINKCC if CXX is used to build main() and
  the system requires to link a C++ main using the C++ compiler.

C API
-----

- The documentation for the tp_compare slot is updated to require that
  the return value must be -1, 0, 1; an arbitrary number <0 or >0 is
  not correct.  This is not yet enforced but will be enforced in
  Python 2.3; even later, we may use -2 to indicate errors and +2 for
  "NotImplemented".  Right now, -1 should be used for an error return.

- PyLong_AsLongLong() now accepts int (as well as long) arguments.
  Consequently, PyArg_ParseTuple's 'L' code also accepts int (as well
  as long) arguments.

- PyThread_start_new_thread() now returns a long int giving the thread
  ID, if one can be calculated; it returns -1 for error, 0 if no
  thread ID is calculated (this is an incompatible change, but only
  the thread module used this API).  This code has only really been
  tested on Linux and Windows; other platforms please beware (and
  report any bugs or strange behavior).

- PyUnicode_FromEncodedObject() no longer accepts Unicode objects as
  input.

New platforms
-------------

Tests
-----

Windows
-------

- Installer:  If you install IDLE, and don't disable file-extension
  registration, a new "Edit with IDLE" context (right-click) menu entry
  is created for .py and .pyw files.

- The signal module now supports SIGBREAK on Windows, thanks to Steven
  Scott.  Note that SIGBREAK is unique to Windows.  The default SIGBREAK
  action remains to call Win32 ExitProcess().  This can be changed via
  signal.signal().  For example::

      # Make Ctrl+Break raise KeyboardInterrupt, like Python's default Ctrl+C
      # (SIGINT) behavior.
      import signal
      signal.signal(signal.SIGBREAK, signal.default_int_handler)

      try:
          while 1:
              pass
      except KeyboardInterrupt:
          # We get here on Ctrl+C or Ctrl+Break now; if we had not changed
          # SIGBREAK, only on Ctrl+C (and Ctrl+Break would terminate the
          # program without the possibility for any Python-level cleanup).
          print "Clean exit"


What's New in Python 2.2a4?
===========================

*Release date: 28-Sep-2001*

Type/class unification and new-style classes
--------------------------------------------

- pydoc and inspect are now aware of new-style classes;
  e.g. help(list) at the interactive prompt now shows proper
  documentation for all operations on list objects.

- Applications using Jim Fulton's ExtensionClass module can now safely
  be used with Python 2.2.  In particular, Zope 2.4.1 now works with
  Python 2.2 (as well as with Python 2.1.1).  The Demo/metaclass
  examples also work again.  It is hoped that Gtk and Boost also work
  with 2.2a4 and beyond.  (If you can confirm this, please write
  webmaster@python.org; if there are still problems, please open a bug
  report on SourceForge.)

- property() now takes 4 keyword arguments:  fget, fset, fdel and doc.
  These map to read-only attributes 'fget', 'fset', 'fdel', and '__doc__'
  in the constructed property object.  fget, fset and fdel weren't
  discoverable from Python in 2.2a3.  __doc__ is new, and allows to
  associate a docstring with a property.

- Comparison overloading is now more completely implemented.  For
  example, a str subclass instance can properly be compared to a str
  instance, and it can properly overload comparison.  Ditto for most
  other built-in object types.

- The repr() of new-style classes has changed; instead of <type
  'M.Foo'> a new-style class is now rendered as <class 'M.Foo'>,
  *except* for built-in types, which are still rendered as <type
  'Foo'> (to avoid upsetting existing code that might parse or
  otherwise rely on repr() of certain type objects).

- The repr() of new-style objects is now always <Foo object at XXX>;
  previously, it was sometimes <Foo instance at XXX>.

- For new-style classes, what was previously called __getattr__ is now
  called __getattribute__.  This method, if defined, is called for
  *every* attribute access.  A new __getattr__ hook more similar to the
  one in classic classes is defined which is called only if regular
  attribute access raises AttributeError; to catch *all* attribute
  access, you can use __getattribute__ (for new-style classes).  If
  both are defined, __getattribute__ is called first, and if it raises
  AttributeError, __getattr__ is called.

- The __class__ attribute of new-style objects can be assigned to.
  The new class must have the same C-level object layout as the old
  class.

- The builtin file type can be subclassed now.  In the usual pattern,
  "file" is the name of the builtin type, and file() is a new builtin
  constructor, with the same signature as the builtin open() function.
  file() is now the preferred way to open a file.

- Previously, __new__ would only see sequential arguments passed to
  the type in a constructor call; __init__ would see both sequential
  and keyword arguments.  This made no sense whatsoever any more, so
  now both __new__ and __init__ see all arguments.

- Previously, hash() applied to an instance of a subclass of str or
  unicode always returned 0.  This has been repaired.

- Previously, an operation on an instance of a subclass of an
  immutable type (int, long, float, complex, tuple, str, unicode),
  where the subtype didn't override the operation (and so the
  operation was handled by the builtin type), could return that
  instance instead a value of the base type.  For example, if s was of
  a str subclass type, s[:] returned s as-is.  Now it returns a str
  with the same value as s.

- Provisional support for pickling new-style objects has been added.

Core
----

- file.writelines() now accepts any iterable object producing strings.

- PyUnicode_FromEncodedObject() now works very much like
  PyObject_Str(obj) in that it tries to use __str__/tp_str
  on the object if the object is not a string or buffer. This
  makes unicode() behave like str() when applied to non-string/buffer
  objects.

- PyFile_WriteObject now passes Unicode objects to the file's write
  method. As a result, all file-like objects which may be the target
  of a print statement must support Unicode objects, i.e. they must
  at least convert them into ASCII strings.

- Thread scheduling on Solaris should be improved; it is no longer
  necessary to insert a small sleep at the start of a thread in order
  to let other runnable threads be scheduled.

Library
-------

- StringIO.StringIO instances and cStringIO.StringIO instances support
  read character buffer compatible objects for their .write() methods.
  These objects are converted to strings and then handled as such
  by the instances.

- The "email" package has been added.  This is basically a port of the
  mimelib package <http://sf.net/projects/mimelib> with API changes
  and some implementations updated to use iterators and generators.

- difflib.ndiff() and difflib.Differ.compare() are generators now.  This
  restores the ability of Tools/scripts/ndiff.py to start producing output
  before the entire comparison is complete.

- StringIO.StringIO instances and cStringIO.StringIO instances support
  iteration just like file objects (i.e. their .readline() method is
  called for each iteration until it returns an empty string).

- The codecs module has grown four new helper APIs to access
  builtin codecs: getencoder(), getdecoder(), getreader(),
  getwriter().

- SimpleXMLRPCServer: a new module (based upon SimpleHTMLServer)
  simplifies writing XML RPC servers.

- os.path.realpath(): a new function that returns the absolute pathname
  after interpretation of symbolic links.  On non-Unix systems, this
  is an alias for os.path.abspath().

- operator.indexOf() (PySequence_Index() in the C API) now works with any
  iterable object.

- smtplib now supports various authentication and security features of
  the SMTP protocol through the new login() and starttls() methods.

- hmac: a new module implementing keyed hashing for message
  authentication.

- mimetypes now recognizes more extensions and file types.  At the
  same time, some mappings not sanctioned by IANA were removed.

- The "compiler" package has been brought up to date to the state of
  Python 2.2 bytecode generation.  It has also been promoted from a
  Tool to a standard library package.  (Tools/compiler still exists as
  a sample driver.)

Build
-----

- Large file support (LFS) is now automatic when the platform supports
  it; no more manual configuration tweaks are needed.  On Linux, at
  least, it's possible to have a system whose C library supports large
  files but whose kernel doesn't; in this case, large file support is
  still enabled but doesn't do you any good unless you upgrade your
  kernel or share your Python executable with another system whose
  kernel has large file support.

- The configure script now supplies plausible defaults in a
  cross-compilation environment.  This doesn't mean that the supplied
  values are always correct, or that cross-compilation now works
  flawlessly -- but it's a first step (and it shuts up most of
  autoconf's warnings about AC_TRY_RUN).

- The Unix build is now a bit less chatty, courtesy of the parser
  generator.  The build is completely silent (except for errors) when
  using "make -s", thanks to a -q option to setup.py.

C API
-----

- The "structmember" API now supports some new flag bits to deny read
  and/or write access to attributes in restricted execution mode.

New platforms
-------------

- Compaq's iPAQ handheld, running the "familiar" Linux distribution
  (http://familiar.handhelds.org).

Tests
-----

- The "classic" standard tests, which work by comparing stdout to
  an expected-output file under Lib/test/output/, no longer stop at
  the first mismatch.  Instead the test is run to completion, and a
  variant of ndiff-style comparison is used to report all differences.
  This is much easier to understand than the previous style of reporting.

- The unittest-based standard tests now use regrtest's test_main()
  convention, instead of running as a side-effect of merely being
  imported.  This allows these tests to be run in more natural and
  flexible ways as unittests, outside the regrtest framework.

- regrtest.py is much better integrated with unittest and doctest now,
  especially in regard to reporting errors.

Windows
-------

- Large file support now also works for files > 4GB, on filesystems
  that support it (NTFS under Windows 2000).  See "What's New in
  Python 2.2a3" for more detail.


What's New in Python 2.2a3?
===========================

*Release Date: 07-Sep-2001*

Core
----

- Conversion of long to float now raises OverflowError if the long is too
  big to represent as a C double.

- The 3-argument builtin pow() no longer allows a third non-None argument
  if either of the first two arguments is a float, or if both are of
  integer types and the second argument is negative (in which latter case
  the arguments are converted to float, so this is really the same
  restriction).

- The builtin dir() now returns more information, and sometimes much
  more, generally naming all attributes of an object, and all attributes
  reachable from the object via its class, and from its class's base
  classes, and so on from them too.  Example:  in 2.2a2, dir([]) returned
  an empty list.  In 2.2a3,

  >>> dir([])
  ['__add__', '__class__', '__contains__', '__delattr__', '__delitem__',
   '__eq__', '__ge__', '__getattr__', '__getitem__', '__getslice__',
   '__gt__', '__hash__', '__iadd__', '__imul__', '__init__', '__le__',
   '__len__', '__lt__', '__mul__', '__ne__', '__new__', '__repr__',
   '__rmul__', '__setattr__', '__setitem__', '__setslice__', '__str__',
   'append', 'count', 'extend', 'index', 'insert', 'pop', 'remove',
   'reverse', 'sort']

  dir(module) continues to return only the module's attributes, though.

- Overflowing operations on plain ints now return a long int rather
  than raising OverflowError.  This is a partial implementation of PEP
  237.  You can use -Wdefault::OverflowWarning to enable a warning for
  this situation, and -Werror::OverflowWarning to revert to the old
  OverflowError exception.

- A new command line option, -Q<arg>, is added to control run-time
  warnings for the use of classic division.  (See PEP 238.)  Possible
  values are -Qold, -Qwarn, -Qwarnall, and -Qnew.  The default is
  -Qold, meaning the / operator has its classic meaning and no
  warnings are issued.  Using -Qwarn issues a run-time warning about
  all uses of classic division for int and long arguments; -Qwarnall
  also warns about classic division for float and complex arguments
  (for use with fixdiv.py).
  [Note:  the remainder of this item (preserved below) became
  obsolete in 2.2c1 -- -Qnew has global effect in 2.2] ::

    Using -Qnew is questionable; it turns on new division by default, but
    only in the __main__ module.  You can usefully combine -Qwarn or
    -Qwarnall and -Qnew: this gives the __main__ module new division, and
    warns about classic division everywhere else.

- Many built-in types can now be subclassed.  This applies to int,
  long, float, str, unicode, and tuple.  (The types complex, list and
  dictionary can also be subclassed; this was introduced earlier.)
  Note that restrictions apply when subclassing immutable built-in
  types: you can only affect the value of the instance by overloading
  __new__.  You can add mutable attributes, and the subclass instances
  will have a __dict__ attribute, but you cannot change the "value"
  (as implemented by the base class) of an immutable subclass instance
  once it is created.

- The dictionary constructor now takes an optional argument, a
  mapping-like object, and initializes the dictionary from its
  (key, value) pairs.

- A new built-in type, super, has been added.  This facilitates making
  "cooperative super calls" in a multiple inheritance setting.  For an
  explanation, see http://www.python.org/2.2/descrintro.html#cooperation

- A new built-in type, property, has been added.  This enables the
  creation of "properties".  These are attributes implemented by
  getter and setter functions (or only one of these for read-only or
  write-only attributes), without the need to override __getattr__.
  See http://www.python.org/2.2/descrintro.html#property

- The syntax of floating-point and imaginary literals has been
  liberalized, to allow leading zeroes.  Examples of literals now
  legal that were SyntaxErrors before:

      00.0    0e3   0100j   07.5   00000000000000000008.

- An old tokenizer bug allowed floating point literals with an incomplete
  exponent, such as 1e and 3.1e-.  Such literals now raise SyntaxError.

Library
-------

- telnetlib includes symbolic names for the options, and support for
  setting an option negotiation callback. It also supports processing
  of suboptions.

- The new C standard no longer requires that math libraries set errno to
  ERANGE on overflow.  For platform libraries that exploit this new
  freedom, Python's overflow-checking was wholly broken.  A new overflow-
  checking scheme attempts to repair that, but may not be reliable on all
  platforms (C doesn't seem to provide anything both useful and portable
  in this area anymore).

- Asynchronous timeout actions are available through the new class
  threading.Timer.

- math.log and math.log10 now return sensible results for even huge
  long arguments.  For example, math.log10(10 ** 10000) ~= 10000.0.

- A new function, imp.lock_held(), returns 1 when the import lock is
  currently held.  See the docs for the imp module.

- pickle, cPickle and marshal on 32-bit platforms can now correctly read
  dumps containing ints written on platforms where Python ints are 8 bytes.
  When read on a box where Python ints are 4 bytes, such values are
  converted to Python longs.

- In restricted execution mode (using the rexec module), unmarshalling
  code objects is no longer allowed.  This plugs a security hole.

- unittest.TestResult instances no longer store references to tracebacks
  generated by test failures. This prevents unexpected dangling references
  to objects that should be garbage collected between tests.

Tools
-----

- Tools/scripts/fixdiv.py has been added which can be used to fix
  division operators as per PEP 238.

Build
-----

- If you are an adventurous person using Mac OS X you may want to look at
  Mac/OSX. There is a Makefile there that will build Python as a real Mac
  application, which can be used for experimenting with Carbon or Cocoa.
  Discussion of this on pythonmac-sig, please.

C API
-----

- New function PyObject_Dir(obj), like Python __builtin__.dir(obj).

- Note that PyLong_AsDouble can fail!  This has always been true, but no
  callers checked for it.  It's more likely to fail now, because overflow
  errors are properly detected now.  The proper way to check::

      double x = PyLong_AsDouble(some_long_object);
      if (x == -1.0 && PyErr_Occurred()) {
              /* The conversion failed. */
      }

- The GC API has been changed.  Extensions that use the old API will still
  compile but will not participate in GC.  To upgrade an extension
  module:

    - rename Py_TPFLAGS_GC to PyTPFLAGS_HAVE_GC

    - use PyObject_GC_New or PyObject_GC_NewVar to allocate objects and
      PyObject_GC_Del to deallocate them

    - rename PyObject_GC_Init to PyObject_GC_Track and PyObject_GC_Fini
      to PyObject_GC_UnTrack

    - remove PyGC_HEAD_SIZE from object size calculations

    - remove calls to PyObject_AS_GC and PyObject_FROM_GC

- Two new functions: PyString_FromFormat() and PyString_FromFormatV().
  These can be used safely to construct string objects from a
  sprintf-style format string (similar to the format string supported
  by PyErr_Format()).

New platforms
-------------

- Stephen Hansen contributed patches sufficient to get a clean compile
  under Borland C (Windows), but he reports problems running it and ran
  out of time to complete the port.  Volunteers?  Expect a MemoryError
  when importing the types module; this is probably shallow, and
  causing later failures too.

Tests
-----

Windows
-------

- Large file support is now enabled on Win32 platforms as well as on
  Win64.  This means that, for example, you can use f.tell() and f.seek()
  to manipulate files larger than 2 gigabytes (provided you have enough
  disk space, and are using a Windows filesystem that supports large
  partitions).  Windows filesystem limits:  FAT has a 2GB (gigabyte)
  filesize limit, and large file support makes no difference there.
  FAT32's limit is 4GB, and files >= 2GB are easier to use from Python now.
  NTFS has no practical limit on file size, and files of any size can be
  used from Python now.

- The w9xpopen hack is now used on Windows NT and 2000 too when COMPSPEC
  points to command.com (patch from Brian Quinlan).


What's New in Python 2.2a2?
===========================

*Release Date: 22-Aug-2001*

Build
-----

- Tim Peters developed a brand new Windows installer using Wise 8.1,
  generously donated to us by Wise Solutions.

- configure supports a new option --enable-unicode, with the values
  ucs2 and ucs4 (new in 2.2a1). With --disable-unicode, the Unicode
  type and supporting code is completely removed from the interpreter.

- A new configure option --enable-framework builds a Mac OS X framework,
  which "make frameworkinstall" will install. This provides a starting
  point for more mac-like functionality, join pythonmac-sig@python.org
  if you are interested in helping.

- The NeXT platform is no longer supported.

- The 'new' module is now statically linked.

Tools
-----

- The new Tools/scripts/cleanfuture.py can be used to automatically
  edit out obsolete future statements from Python source code.  See
  the module docstring for details.

Tests
-----

- regrtest.py now knows which tests are expected to be skipped on some
  platforms, allowing to give clearer test result output.  regrtest
  also has optional --use/-u switch to run normally disabled tests
  which require network access or consume significant disk resources.

- Several new tests in the standard test suite, with special thanks to
  Nick Mathewson.

Core
----

- The floor division operator // has been added as outlined in PEP
  238.  The / operator still provides classic division (and will until
  Python 3.0) unless "from __future__ import division" is included, in
  which case the / operator will provide true division.  The operator
  module provides truediv() and floordiv() functions.  Augmented
  assignment variants are included, as are the equivalent overloadable
  methods and C API methods.  See the PEP for a full discussion:
  <http://python.sf.net/peps/pep-0238.html>

- Future statements are now effective in simulated interactive shells
  (like IDLE).  This should "just work" by magic, but read Michael
  Hudson's "Future statements in simulated shells" PEP 264 for full
  details:  <http://python.sf.net/peps/pep-0264.html>.

- The type/class unification (PEP 252-253) was integrated into the
  trunk and is not so tentative any more (the exact specification of
  some features is still tentative).  A lot of work has done on fixing
  bugs and adding robustness and features (performance still has to
  come a long way).

- Warnings about a mismatch in the Python API during extension import
  now use the Python warning framework (which makes it possible to
  write filters for these warnings).

- A function's __dict__ (aka func_dict) will now always be a
  dictionary.  It used to be possible to delete it or set it to None,
  but now both actions raise TypeErrors.  It is still legal to set it
  to a dictionary object.  Getting func.__dict__ before any attributes
  have been assigned now returns an empty dictionary instead of None.

- A new command line option, -E, was added which disables the use of
  all environment variables, or at least those that are specifically
  significant to Python.  Usually those have a name starting with
  "PYTHON".  This was used to fix a problem where the tests fail if
  the user happens to have PYTHONHOME or PYTHONPATH pointing to an
  older distribution.

Library
-------

- New class Differ and new functions ndiff() and restore() in difflib.py.
  These package the algorithms used by the popular Tools/scripts/ndiff.py,
  for programmatic reuse.

- New function xml.sax.saxutils.quoteattr():  Quote an XML attribute
  value using the minimal quoting required for the value; more
  reliable than using xml.sax.saxutils.escape() for attribute values.

- Readline completion support for cmd.Cmd was added.

- Calling os.tempnam() or os.tmpnam() generate RuntimeWarnings.

- Added function threading.BoundedSemaphore()

- Added Ka-Ping Yee's cgitb.py module.

- The 'new' module now exposes the CO_xxx flags.

- The gc module offers the get_referents function.

New platforms
-------------

C API
-----

- Two new APIs PyOS_snprintf() and PyOS_vsnprintf() were added
  which provide a cross-platform implementations for the
  relatively new snprintf()/vsnprintf() C lib APIs. In contrast to
  the standard sprintf() and vsprintf() C lib APIs, these versions
  apply bounds checking on the used buffer which enhances protection
  against buffer overruns.

- Unicode APIs now use name mangling to assure that mixing interpreters
  and extensions using different Unicode widths is rendered next to
  impossible. Trying to import an incompatible Unicode-aware extension
  will result in an ImportError.  Unicode extensions writers must make
  sure to check the Unicode width compatibility in their extensions by
  using at least one of the mangled Unicode APIs in the extension.

- Two new flags METH_NOARGS and METH_O are available in method definition
  tables to simplify implementation of methods with no arguments and a
  single untyped argument. Calling such methods is more efficient than
  calling corresponding METH_VARARGS methods. METH_OLDARGS is now
  deprecated.

Windows
-------

- "import module" now compiles module.pyw if it exists and nothing else
  relevant is found.


What's New in Python 2.2a1?
===========================

*Release date: 18-Jul-2001*

Core
----

- TENTATIVELY, a large amount of code implementing much of what's
  described in PEP 252 (Making Types Look More Like Classes) and PEP
  253 (Subtyping Built-in Types) was added.  This will be released
  with Python 2.2a1.  Documentation will be provided separately
  through http://www.python.org/2.2/.  The purpose of releasing this
  with Python 2.2a1 is to test backwards compatibility.  It is
  possible, though not likely, that a decision is made not to release
  this code as part of 2.2 final, if any serious backwards
  incompatibilities are found during alpha testing that cannot be
  repaired.

- Generators were added; this is a new way to create an iterator (see
  below) using what looks like a simple function containing one or
  more 'yield' statements.  See PEP 255.  Since this adds a new
  keyword to the language, this feature must be enabled by including a
  future statement: "from __future__ import generators" (see PEP 236).
  Generators will become a standard feature in a future release
  (probably 2.3).  Without this future statement, 'yield' remains an
  ordinary identifier, but a warning is issued each time it is used.
  (These warnings currently don't conform to the warnings framework of
  PEP 230; we intend to fix this in 2.2a2.)

- The UTF-16 codec was modified to be more RFC compliant. It will now
  only remove BOM characters at the start of the string and then
  only if running in native mode (UTF-16-LE and -BE won't remove a
  leading BMO character).

- Strings now have a new method .decode() to complement the already
  existing .encode() method. These two methods provide direct access
  to the corresponding decoders and encoders of the registered codecs.

  To enhance the usability of the .encode() method, the special
  casing of Unicode object return values was dropped (Unicode objects
  were auto-magically converted to string using the default encoding).

  Both methods will now return whatever the codec in charge of the
  requested encoding returns as object, e.g. Unicode codecs will
  return Unicode objects when decoding is requested ("äöü".decode("latin-1")
  will return u"äöü"). This enables codec writer to create codecs
  for various simple to use conversions.

  New codecs were added to demonstrate these new features (the .encode()
  and .decode() columns indicate the type of the returned objects):

  +---------+-----------+-----------+-----------------------------+
  |Name     | .encode() | .decode() | Description                 |
  +=========+===========+===========+=============================+
  |uu       | string    | string    | UU codec (e.g. for email)   |
  +---------+-----------+-----------+-----------------------------+
  |base64   | string    | string    | base64 codec                |
  +---------+-----------+-----------+-----------------------------+
  |quopri   | string    | string    | quoted-printable codec      |
  +---------+-----------+-----------+-----------------------------+
  |zlib     | string    | string    | zlib compression            |
  +---------+-----------+-----------+-----------------------------+
  |hex      | string    | string    | 2-byte hex codec            |
  +---------+-----------+-----------+-----------------------------+
  |rot-13   | string    | Unicode   | ROT-13 Unicode charmap codec|
  +---------+-----------+-----------+-----------------------------+

- Some operating systems now support the concept of a default Unicode
  encoding for file system operations.  Notably, Windows supports 'mbcs'
  as the default.  The Macintosh will also adopt this concept in the medium
  term, although the default encoding for that platform will be other than
  'mbcs'.

  On operating system that support non-ASCII filenames, it is common for
  functions that return filenames (such as os.listdir()) to return Python
  string objects pre-encoded using the default file system encoding for
  the platform.  As this encoding is likely to be different from Python's
  default encoding, converting this name to a Unicode object before passing
  it back to the Operating System would result in a Unicode error, as Python
  would attempt to use its default encoding (generally ASCII) rather than
  the default encoding for the file system.

  In general, this change simply removes surprises when working with
  Unicode and the file system, making these operations work as you expect,
  increasing the transparency of Unicode objects in this context.
  See [????] for more details, including examples.

- Float (and complex) literals in source code were evaluated to full
  precision only when running from a .py file; the same code loaded from a
  .pyc (or .pyo) file could suffer numeric differences starting at about the
  12th significant decimal digit.  For example, on a machine with IEEE-754
  floating arithmetic,

      x = 9007199254740992.0
      print long(x)

  printed 9007199254740992 if run directly from .py, but 9007199254740000
  if from a compiled (.pyc or .pyo) file.  This was due to marshal using
  str(float) instead of repr(float) when building code objects.  marshal
  now uses repr(float) instead, which should reproduce floats to full
  machine precision (assuming the platform C float<->string I/O conversion
  functions are of good quality).

  This may cause floating-point results to change in some cases, and
  usually for the better, but may also cause numerically unstable
  algorithms to break.

- The implementation of dicts suffers fewer collisions, which has speed
  benefits.  However, the order in which dict entries appear in dict.keys(),
  dict.values() and dict.items() may differ from previous releases for a
  given dict.  Nothing is defined about this order, so no program should
  rely on it.  Nevertheless, it's easy to write test cases that rely on the
  order by accident, typically because of printing the str() or repr() of a
  dict to an "expected results" file.  See Lib/test/test_support.py's new
  sortdict(dict) function for a simple way to display a dict in sorted
  order.

- Many other small changes to dicts were made, resulting in faster
  operation along the most common code paths.

- Dictionary objects now support the "in" operator: "x in dict" means
  the same as dict.has_key(x).

- The update() method of dictionaries now accepts generic mapping
  objects.  Specifically the argument object must support the .keys()
  and __getitem__() methods.  This allows you to say, for example,
  {}.update(UserDict())

- Iterators were added; this is a generalized way of providing values
  to a for loop.  See PEP 234.  There's a new built-in function iter()
  to return an iterator.  There's a new protocol to get the next value
  from an iterator using the next() method (in Python) or the
  tp_iternext slot (in C).  There's a new protocol to get iterators
  using the __iter__() method (in Python) or the tp_iter slot (in C).
  Iterating (i.e. a for loop) over a dictionary generates its keys.
  Iterating over a file generates its lines.

- The following functions were generalized to work nicely with iterator
  arguments::

    map(), filter(), reduce(), zip()
    list(), tuple() (PySequence_Tuple() and PySequence_Fast() in C API)
    max(), min()
    join() method of strings
    extend() method of lists
    'x in y' and 'x not in y' (PySequence_Contains() in C API)
    operator.countOf() (PySequence_Count() in C API)
    right-hand side of assignment statements with multiple targets, such as ::
        x, y, z = some_iterable_object_returning_exactly_3_values

- Accessing module attributes is significantly faster (for example,
  random.random or os.path or yourPythonModule.yourAttribute).

- Comparing dictionary objects via == and != is faster, and now works even
  if the keys and values don't support comparisons other than ==.

- Comparing dictionaries in ways other than == and != is slower:  there were
  insecurities in the dict comparison implementation that could cause Python
  to crash if the element comparison routines for the dict keys and/or
  values mutated the dicts.  Making the code bulletproof slowed it down.

- Collisions in dicts are resolved via a new approach, which can help
  dramatically in bad cases.  For example, looking up every key in a dict
  d with d.keys() == [i << 16 for i in range(20000)] is approximately 500x
  faster now.  Thanks to Christian Tismer for pointing out the cause and
  the nature of an effective cure (last December! better late than never).

- repr() is much faster for large containers (dict, list, tuple).


Library
-------

- The constants ascii_letters, ascii_lowercase. and ascii_uppercase
  were added to the string module.  These a locale-independent
  constants, unlike letters, lowercase, and uppercase.  These are now
  use in appropriate locations in the standard library.

- The flags used in dlopen calls can now be configured using
  sys.setdlopenflags and queried using sys.getdlopenflags.

- Fredrik Lundh's xmlrpclib is now a standard library module.  This
  provides full client-side XML-RPC support.  In addition,
  Demo/xmlrpc/ contains two server frameworks (one SocketServer-based,
  one asyncore-based).  Thanks to Eric Raymond for the documentation.

- The xrange() object is simplified: it no longer supports slicing,
  repetition, comparisons, efficient 'in' checking, the tolist()
  method, or the start, stop and step attributes.  See PEP 260.

- A new function fnmatch.filter to filter lists of file names was added.

- calendar.py uses month and day names based on the current locale.

- strop is now *really* obsolete (this was announced before with 1.6),
  and issues DeprecationWarning when used (except for the four items
  that are still imported into string.py).

- Cookie.py now sorts key+value pairs by key in output strings.

- pprint.isrecursive(object) didn't correctly identify recursive objects.
  Now it does.

- pprint functions now much faster for large containers (tuple, list, dict).

- New 'q' and 'Q' format codes in the struct module, corresponding to C
  types "long long" and "unsigned long long" (on Windows, __int64).  In
  native mode, these can be used only when the platform C compiler supports
  these types (when HAVE_LONG_LONG is #define'd by the Python config
  process), and then they inherit the sizes and alignments of the C types.
  In standard mode, 'q' and 'Q' are supported on all platforms, and are
  8-byte integral types.

- The site module installs a new built-in function 'help' that invokes
  pydoc.help.  It must be invoked as 'help()'; when invoked as 'help',
  it displays a message reminding the user to use 'help()' or
  'help(object)'.

Tests
-----

- New test_mutants.py runs dict comparisons where the key and value
  comparison operators mutate the dicts randomly during comparison.  This
  rapidly causes Python to crash under earlier releases (not for the faint
  of heart:  it can also cause Win9x to freeze or reboot!).

- New test_pprint.py verifies that pprint.isrecursive() and
  pprint.isreadable() return sensible results.  Also verifies that simple
  cases produce correct output.

C API
-----

- Removed the unused last_is_sticky argument from the internal
  _PyTuple_Resize().  If this affects you, you were cheating.

----

**(For information about older versions, consult the HISTORY file.)**