What's new in this release? =========================== Below is a list of all relevant changes since the release 1.4, up till the release of 1.5a3. At the end is a list of changes made since 1.5a3 up to the release of 1.5a4. A note on attributions: while I have sprinkled some names throughout here, I'm grateful to many more people who remain anonymous. You may find your name in the ACKS file. If you believe you deserve more credit, let me know and I'll add you to the list! Security -------- - If you are using the setuid script C wrapper (Misc/setuid-prog.c), please use the new version. The old version has a huge security leak. Miscellaneous ------------- - Because of various (small) incompatible changes in the Python bytecode interpreter, the magic number for .pyc files has changed again. - The default module search path is now much saner. Both on Unix and Windows, it is essentially derived from the path to the executable (which can be overridden by setting the environment variable $PYTHONHOME). The value of $PYTHONPATH on Windows is now inserted in front of the default path, like in Unix (instead of overriding the default path). On Windows, the directory containing the executable is added to the end of the path. - A new version of python-mode.el for Emacs has been included. Also, a new file ccpy-style.el has been added to configure Emacs cc-mode for the preferred style in Python C sources. - On Unix, when using sys.argv[0] to insert the script directory in front of sys.path, expand a symbolic link. You can now install a program in a private directory and have a symbolic link to it in a public bin directory, and it will put the private directory in the module search path. Note that the symlink is expanded in sys.path[0] but not in sys.argv[0], so you can still tell the name by which you were invoked. - It is now recommended to use ``#!/usr/bin/env python'' instead of ``#!/usr/local/bin/python'' at the start of executable scripts, except for CGI scripts. It has been determined that the use of /usr/bin/env is more portable than that of /usr/local/bin/python -- scripts almost never have to be edited when the Python interpreter lives in a non-standard place. Note that this doesn't work for CGI scripts since the python executable often doesn't live in the HTTP server's default search path. - The silly -s command line option and the corresponding PYTHONSUPPRESS environment variable (and the Py_SuppressPrint global flag in the Python/C API) are gone. - Most problems on 64-bit platforms should now be fixed. Andrew Kuchling helped. Some uncommon extension modules are still not clean (image and audio ops?). - Fixed a bug where multiple anonymous tuple arguments would be mixed up when using the debugger or profiler (reported by Just van Rossum). The simplest example is ``def f((a,b),(c,d)): print a,b,c,d''; this would print the wrong value when run under the debugger or profiler. - The hacks that the dictionary implementation used to speed up repeated lookups of the same C string were removed; these were a source of subtle problems and don't seem to serve much of a purpose any longer. - All traces of support for the long dead access statement have been removed from the sources. - Plugged the two-byte memory leak in the tokenizer when reading an interactive EOF. - There's a -O option to the interpreter that removes SET_LINENO instructions and assert statements (see below); it uses and produces .pyo files instead of .pyc files. The speedup is only a few percent in most cases. The line numbers are still available in the .pyo file, as a separate table (which is also available in .pyc files). However, the removal of the SET_LINENO instructions means that the debugger (pdb) can't set breakpoints on lines in -O mode. The traceback module contains a function to extract a line number from the code object referenced in a traceback object. In the future it should be possible to write external bytecode optimizers that create better optimized .pyo files, and there should be more control over optimization; consider the -O option a "teaser". Without -O, the assert statement actually generates code that first checks __debug__; if this variable is false, the assertion is not checked. __debug__ is a built-in variable whose value is initialized to track the -O flag (it's true iff -O is not specified). With -O, no code is generated for assert statements, nor for code of the form ``if __debug__: ''. Sorry, no further constant folding happens. Performance ----------- - It's much faster (almost twice for pystone.py -- see Tools/scripts). See the entry on string interning below. - Some speedup by using separate free lists for method objects (both the C and the Python variety) and for floating point numbers. - Big speedup by allocating frame objects with a single malloc() call. The Python/C API for frames is changed (you shouldn't be using this anyway). - Significant speedup by inlining some common opcodes for common operand types (e.g. i+i, i-i, and list[i]). Fredrik Lundh. - Small speedup by reordering the method tables of some common objects (e.g. list.append is now first). - Big optimization to the read() method of file objects. A read() without arguments now attempts to use fstat to allocate a buffer of the right size; for pipes and sockets, it will fall back to doubling the buffer size. While that the improvement is real on all systems, it is most dramatic on Windows. Documentation ------------- - Many new pieces of library documentation were contributed, mostly by Andrew Kuchling. Even cmath is now documented! There's also a chapter of the library manual, "libundoc.tex", which provides a listing of all undocumented modules, plus their status (e.g. internal, obsolete, or in need of documentation). Also contributions by Sue Williams, Skip Montanaro, and some module authors who succumbed to pressure to document their own contributed modules :-). Note that printing the documentation now kills fewer trees -- the margins have been reduced. - I have started documenting the Python/C API. Unfortunately this project hasn't been completed yet. It will be complete before the final release of Python 1.5, though. At the moment, it's better to read the LaTeX source than to attempt to run it through LaTeX and print the resulting dvi file. - The posix module (and hence os.py) now has doc strings! Thanks to Neil Schemenauer. I received a few other contributions of doc strings. In most other places, doc strings are still wishful thinking... Language changes ---------------- - Private variables with leading double underscore are now a permanent feature of the language. (These were experimental in release 1.4. I have favorable experience using them; I can't label them "experimental" forever.) - There's new string literal syntax for "raw strings". Prefixing a string literal with the letter r (or R) disables all escape processing in the string; for example, r'\n' is a two-character string consisting of a backslash followed by the letter n. This combines with all forms of string quotes; it is actually useful for triple quoted doc strings which might contain references to \n or \t. An embedded quote prefixed with a backslash does not terminate the string, but the backslash is still included in the string; for example, r'\'' is a two-character string consisting of a backslash and a quote. (Raw strings are also affectionately known as Robin strings, after their inventor, Robin Friedrich.) - There's a simple assert statement, and a new exception AssertionError. For example, ``assert foo > 0'' is equivalent to ``if not foo > 0: raise AssertionError''. Sorry, the text of the asserted condition is not available; it would be too complicated to generate code for this (since the code is generated from a parse tree). However, the text is displayed as part of the traceback! - The raise statement has a new feature: when using "raise SomeClass, somevalue" where somevalue is not an instance of SomeClass, it instantiates SomeClass(somevalue). In 1.5a4, if somevalue is an instance of a *derived* class of SomeClass, the exception class raised is set to somevalue.__class__, and SomeClass is ignored after that. - Duplicate keyword arguments are now detected at compile time; f(a=1,a=2) is now a syntax error. Changes to builtin features --------------------------- - There's a new exception FloatingPointError (used only by Lee Busby's patches to catch floating point exceptions, at the moment). - The obsolete exception ConflictError (presumably used by the long obsolete access statement) has been deleted. - There's a new function sys.exc_info() which returns the tuple (sys.exc_type, sys.exc_value, sys.exc_traceback) in a thread-safe way. - There's a new variable sys.executable, pointing to the executable file for the Python interpreter. - The sort() methods for lists no longer uses the C library qsort(); I wrote my own quicksort implementation, with lots of help (in the form of a kind of competition) from Tim Peters. This solves a bug in dictionary comparisons on some Solaris versions when Python is built with threads, and makes sorting lists even faster. - The semantics of comparing two dictionaries have changed, to make comparison of unequal dictionaries faster. A shorter dictionary is always considered smaller than a larger dictionary. For dictionaries of the same size, the smallest differing element determines the outcome (which yields the same results as before in this case, without explicit sorting). Thanks to Aaron Watters for suggesting something like this. - The semantics of try-except have changed subtly so that calling a function in an exception handler that itself raises and catches an exception no longer overwrites the sys.exc_* variables. This also alleviates the problem that objects referenced in a stack frame that caught an exception are kept alive until another exception is caught -- the sys.exc_* variables are restored to their previous value when returning from a function that caught an exception. - There's a new "buffer" interface. Certain objects (e.g. strings and arrays) now support the "buffer" protocol. Buffer objects are acceptable whenever formerly a string was required for a write operation; mutable buffer objects can be the target of a read operation using the call f.readinto(buffer). A cool feature is that regular expression matching now also work on array objects. Contribution by Jack Jansen. (Needs documentation.) - String interning: dictionary lookups are faster when the lookup string object is the same object as the key in the dictionary, not just a string with the same value. This is done by having a pool of "interned" strings. Most names generated by the interpreter are now automatically interned, and there's a new built-in function intern(s) that returns the interned version of a string. Interned strings are not a different object type, and interning is totally optional, but by interning most keys a speedup of about 15% was obtained for the pystone benchmark. - Dictionary objects have several new methods; clear() and copy() have the obvious semantics, while update(d) merges the contents of another dictionary d into this one, overriding existing keys. The dictionary implementation file is now called dictobject.c rather than the confusing mappingobject.c. - The intrinsic function dir() is much smarter; it looks in __dict__, __members__ and __methods__. - The intrinsic functions int(), long() and float() can now take a string argument and then do the same thing as string.atoi(), string.atol(), and string.atof(). No second 'base' argument is allowed, and complex() does not take a string (nobody cared enough). - When a module is deleted, its globals are now deleted in two phases. In the first phase, all variables whose name begins with exactly one underscore are replaced by None; in the second phase, all variables are deleted. This makes it possible to have global objects whose destructors depend on other globals. The deletion order within each phase is still random. - It is no longer an error for a function to be called without a global variable __builtins__ -- an empty directory will be provided by default. - Guido's corollary to the "Don Beaudry hook": it is now possible to do metaprogramming by using an instance as a base class. Not for the faint of heart; and undocumented as yet, but basically if a base class is an instance, its class will be instantiated to create the new class. Jim Fulton will love it -- it also works with instances of his "extension classes", since it is triggered by the presence of a __class__ attribute on the purported base class. See Demo/metaclasses/index.html for an explanation and see that directory for examples. - Another change is that the Don Beaudry hook is now invoked when *any* base class is special. (Up to 1.5a3, the *last* special base class is used; in 1.5a4, the more rational choice of the *first* special base class is used.) - New optional parameter to the readlines() method of file objects. This indicates the number of bytes to read (the actual number of bytes read will be somewhat larger due to buffering reading until the end of the line). Some optimizations have also been made to speed it up (but not as much as read()). - Complex numbers no longer have the ".conj" pseudo attribute; use z.conjugate() instead, or complex(z.real, -z.imag). Complex numbers now *do* support the __members__ and __methods__ special attributes. - The complex() function now looks for a __complex__() method on class instances before giving up. - Long integers now support arbitrary shift counts, so you can now write 1L<<1000000, memory permitting. (Python 1.4 reports "outrageous shift count for this.) - The hex() and oct() functions have been changed so that for regular integers, they never emit a minus sign. For example, on a 32-bit machine, oct(-1) now returns '037777777777' and hex(-1) returns '0xffffffff'. While this may seem inconsistent, it is much more useful. (For long integers, a minus sign is used as before, to fit the result in memory :-) - The hash() function computes better hashes for several data types, including strings, floating point numbers, and complex numbers. New extension modules --------------------- - New extension modules cStringIO.c and cPickle.c, written by Jim Fulton and other folks at Digital Creations. These are much more efficient than their Python counterparts StringIO.py and pickle.py, but don't support subclassing. cPickle.c clocks up to 1000 times faster than pickle.py; cStringIO.c's improvement is less dramatic but still significant. - New extension module zlibmodule.c, interfacing to the free zlib library (gzip compatible compression). There's also a module gzip.py which provides a higher level interface. Written by Andrew Kuchling and Jeremy Hylton. - New module readline; see the "miscellaneous" section above. - New Unix extension module resource.c, by Jeremy Hylton, provides access to getrlimit(), getrusage(), setrusage(), getpagesize(), and related symbolic constants. - New extension puremodule.c, by Barry Warsaw, which interfaces to the Purify(TM) C API. See also the file Misc/PURIFY.README. It is also possible to enable Purify by simply setting the PURIFY Makefile variable in the Modules/Setup file. Changes in extension modules ---------------------------- - The struct extension module has several new features to control byte order and word size. It supports reading and writing IEEE floats even on platforms where this is not the native format. It uses uppercase format codes for unsigned integers of various sizes (always using Python long ints for 'I' and 'L'), 's' with a size prefix for strings, and 'p' for "Pascal strings" (with a leading length byte, included in the size; blame Hannu Krosing; new in 1.5a4). A prefix '>' forces big-endian data and '<' forces little-endian data; these also select standard data sizes and disable automatic alignment (use pad bytes as needed). - The array module supports uppercase format codes for unsigned data formats (like the struct module). - The fcntl extension module now exports the needed symbolic constants. (Formerly these were in FCNTL.py which was not available or correct for all platforms.) - The extension modules dbm, gdbm and bsddb now check that the database is still open before making any new calls. - The dbhash module is no more. Use bsddb instead. (There's a third party interface for the BSD 2.x code somewhere on the web; support for bsddb will be deprecated.) - The gdbm module now supports a sync() method. - The socket module now has some new functions: getprotobyname(), and the set {ntoh,hton}{s,l}(). - Various modules now export their type object: socket.SocketType, array.ArrayType. - The socket module's accept() method now returns unknown addresses as a tuple rather than raising an exception. (This can happen in promiscuous mode.) Theres' also a new function getprotobyname(). - The pthread support for the thread module now works on most platforms. - STDWIN is now officially obsolete. Support for it will eventually be removed from the distribution. - The binascii extension module is now hopefully fully debugged. (XXX Oops -- Fredrik Lundh promised me a uuencode fix that I never received.) - audioop.c: added a ratecv() function; better handling of overflow in add(). - posixmodule.c: now exports the O_* flags (O_APPEND etc.). On Windows, also O_TEXT and O_BINARY. The 'error' variable (the exception is raises) is renamed -- its string value is now "os.error", so newbies don't believe they have to import posix (or nt) to catch it when they see os.error reported as posix.error. The execve() function now accepts any mapping object for the environment. - A new version of the al (audio library) module for SGI was contributed by Sjoerd Mullender. - The regex module has a new function get_syntax() which retrieves the syntax setting set by set_syntax(). The code was also sanitized, removing worries about unclean error handling. See also below for its successor, re.py. - The "new" module (which creates new objects of various types) once again has a fully functioning new.function() method. Dangerous as ever! Also, new.code() has several new arguments. - A problem has been fixed in the rotor module: on systems with signed characters, rotor-encoded data was not portable when the key contained 8-bit characters. Also, setkey() now requires its argument rather than having broken code to default it. - The sys.builtin_module_names variable is now a tuple. Another new variables in sys is sys.executable (the full path to the Python binary, if known). - The specs for time.strftime() have undergone some revisions. It appears that not all format characters are supported in the same way on all platforms. Rather than reimplement it, we note these differences in the documentation, and emphasize the shared set of features. There's also a thorough test set (that occasionally finds problems in the C library implementation, e.g. on some Linuxes), thanks to Skip Montanaro. - The nis module seems broken when used with NIS+; unfortunately nobody knows how to fix it. It should still work with old NIS. New library modules ------------------- - New (still experimental) Perl-style regular expression module, re.py, which uses a new interface for matching as well as a new syntax; the new interface avoids the thread-unsafety of the regex interface. This comes with a helper extension reopmodule.c and vastly rewritten regexpr.c. Most work on this was done by Jeffrey Ollie, Tim Peters, and Andrew Kuchling. See the documentation libre.tex. In 1.5, the old regex module is still fully supported; in the future, it will become obsolete. - New module gzip.py; see zlib above. - New module keyword.py exports knowledge about Python's built-in keywords. (New version by Ka-Ping Yee.) - New module pprint.py (with documentation) which supports pretty-printing of lists, tuples, & dictionaries recursively. By Fred Drake. - New module code.py. The function code.compile_command() can determine whether an interactively entered command is complete or not, distinguishing incomplete from invalid input. (XXX Unfortunately, this seems broken at this moment, and I don't have the time to fix it. It's probably better to add an explicit interface to the parser for this.) - There is now a library module xdrlib.py which can read and write the XDR data format as used by Sun RPC, for example. It uses the struct module. Changes in library modules -------------------------- - Module codehack.py is now completely obsolete. - The pickle.py module has been updated to make it compatible with the new binary format that cPickle.c produces. By default it produces the old all-ASCII format compatible with the old pickle.py, still much faster than pickle.py; it will read both formats automatically. A few other updates have been made. - A new helper module, copy_reg.py, is provided to register extensions to the pickling code. - Revamped module tokenize.py is much more accurate and has an interface that makes it a breeze to write code to colorize Python source code. Contributed by Ka-Ping Yee. - In ihooks.py, ModuleLoader.load_module() now closes the file under all circumstances. - The tempfile.py module has a new class, TemporaryFile, which creates an open temporary file that will be deleted automatically when closed. This works on Windows and MacOS as well as on Unix. (Jim Fulton.) - Changes to the cgi.py module: Most imports are now done at the top of the module, which provides a speedup when using ni (Jim Fulton). The problem with file upload to a Windows platform is solved by using the new tempfile.TemporaryFile class; temporary files are now always opened in binary mode (Jim Fulton). The cgi.escape() function now takes an optional flag argument that quotes '"' to '"'. It is now possible to invoke cgi.py from a command line script, to test cgi scripts more easily outside an http server. There's an optional limit to the size of uploads to POST (Skip Montanaro). Added a 'strict_parsing' option to all parsing functions (Jim Fulton). The function parse_qs() now uses urllib.unquote() on the name as well as the value of fields (Clarence Gardner). The FieldStorage class now has a __len__() method. - httplib.py: the socket object is no longer closed; all HTTP/1.* responses are now accepted; and it is now thread-safe (by not using the regex module). - BaseHTTPModule.py: treat all HTTP/1.* versions the same. - The popen2.py module is now rewritten using a class, which makes access to the standard error stream and the process id of the subprocess possible. - Added timezone support to the rfc822.py module, in the form of a getdate_tz() method and a parsedate_tz() function; also a mktime_tz(). Also added recognition of some non-standard date formats, by Lars Wirzenius, and RFC 850 dates (Chris Lawrence). - mhlib.py: various enhancements, including almost compatible parsing of message sequence specifiers without invoking a subprocess. Also added a createmessage() method by Lars Wirzenius. - The StringIO.StringIO class now supports readline(nbytes). (Lars Wirzenius.) (Of course, you should be using cStringIO for performance.) - UserDict.py supports the new dictionary methods as well. - Improvements for whrandom.py by Tim Peters: use 32-bit arithmetic to speed it up, and replace 0 seed values by 1 to avoid degeneration. A bug was fixed in the test for invalid arguments. - Module ftplib.py: added support for parsing a .netrc file (Fred Drake). Also added an ntransfercmd() method to the FTP class, which allows access to the expected size of a transfer when available, and a parse150() function to the module which parses the corresponding 150 response. - urllib.py: the ftp cache is now limited to 10 entries. Added quote_plus() and unquote_plus() functions which are like quote() and unquote() but also replace spaces with '+' or vice versa, for encoding/decoding CGI form arguments. Catch all errors from the ftp module. HTTP requests now add the Host: header line. The proxy variable names are now mapped to lower case, for Windows. The spliturl() function no longer erroneously throws away all data past the first newline. The basejoin() function now intereprets "../" correctly. I *believe* that the problems with "exception raised in __del__" under certain circumstances have been fixed (mostly by changes elsewher in the interpreter). - In urlparse.py, there is a cache for results in urlparse.urlparse(); its size limit is set to 20. Also, new URL schemes shttp, https, and snews are "supported". - shelve.py: use cPickle and cStringIO when available. Also added a sync() method, which calls the database's sync() method if there is one. - The mimetools.py module now uses the available Python modules for decoding quoted-printable, uuencode and base64 formats, rather than creating a subprocess. - The python debugger (pdb.py, and its base class bdb.py) now support conditional breakpoints. See the docs. - The modules base64.py, uu.py and quopri.py can now be used as simple command line utilities. - Various small fixes to the nntplib.py module that I can't bother to document in detail. - Sjoerd Mullender's mimify.py module now supports base64 encoding and includes functions to handle the funny encoding you sometimes see in mail headers. It is now documented. - mailbox.py: Added BabylMailbox. Improved the way the mailbox is gotten from the environment. - Many more modules now correctly open files in binary mode when this is necessary on non-Unix platforms. - The copying functions in the undocumented module shutil.py are smarter. - The Writer classes in the formatter.py module now have a flush() method. - The sgmllib.py module accepts hyphens and periods in the middle of attribute names. While this is against the SGML standard, there is some HTML out there that uses this... - The interface for the Python bytecode disassembler module, dis.py, has been enhanced quite a bit. There's now one main function, dis.dis(), which takes almost any kind of object (function, module, class, instance, method, code object) and disassembles it; without arguments it disassembles the last frame of the last traceback. The other functions have changed slightly, too. - The imghdr.py module recognizes new image types: BMP, PNG. - The string.py module has a new function replace(str, old, new, [maxsplit]) which does substring replacements. It is actually implemented in C in the strop module. The functions [r]find() an [r]index() have an optional 4th argument indicating the end of the substring to search, alsoo implemented by their strop counterparts. (Remember, never import strop -- import string uses strop when available with zero overhead.) - The string.join() function now accepts any sequence argument, not just lists and tuples. - The string.maketrans() requires its first two arguments to be present. The old version didn't require them, but there's not much point without them, and the documentation suggests that they are required, so we fixed the code to match the documentation. - The regsub.py module has a function clear_cache(), which clears its internal cache of compiled regular expressions. Also, the cache now takes the current syntax setting into account. (However, this module is now obsolete -- use the sub() or subn() functions or methods in the re module.) - The undocumented module Complex.py has been removed, now that Python has built-in complex numbers. A similar module remains as Demo/classes/Complex.py, as an example. Changes to the build process ---------------------------- - The way GNU readline is configured is totally different. The --with-readline configure option is gone. It is now an extension module, which may be loaded dynamically. You must enable it (and specify the correct linraries to link with) in the Modules/Setup file. Importing the module installs some hooks which enable command line editing. When the interpreter shell is invoked interactively, it attempts to import the readline module; when this fails, the default input mechanism is used. The hook variables are PyOS_InputHook and PyOS_ReadlineFunctionPointer. (Code contributed by Lee Busby, with ideas from William Magro.) - New build procedure: a single library, libpython1.5.a, is now built, which contains absolutely everything except for a one-line main() program (which calls Py_Main(argc, argv) to start the interpreter shell). This makes life much simpler for applications that need to embed Python. The serial number of the build is now included in the version string (sys.version). - As far as I can tell, neither gcc -Wall nor the Microsoft compiler emits a single warning any more when compiling Python. - A number of new Makefile variables have been added for special situations, e.g. LDLAST is appended to the link command. These are used by editing the Makefile or passing them on the make command line. - A set of patches from Lee Busby has been integrated that make it possible to catch floating point exceptions. Use the configure option --with-fpectl to enable the patches; the extension modules fpectl and fpetest provide control to enable/disable and test the feature, respectively. - The support for shared libraries under AIX is now simpler and more robust. Thanks to Vladimir Marangozov for revamping his own patches! - The Modules/makesetup script now reads a file Setup.local as well as a file Setup. Most changes to the Setup script can be done by editing Setup.local instead, which makes it easier to carry a particular setup over from one release to the next. - The Modules/makesetup script now copies any "include" lines it encounters verbatim into the output Makefile. It also recognizes .cxx and .cpp as C++ source files. - The configure script is smarter about C compiler options; e.g. with gcc it uses -O2 and -g when possible, and on some other platforms it uses -Olimit 1500 to avoid a warning from the optimizer about the main loop in ceval.c (which has more than 1000 basic blocks). - The configure script now detects whether malloc(0) returns a NULL pointer or a valid block (of length zero). This avoids the nonsense of always adding one byte to all malloc() arguments on most platforms. - The configure script has a new option, --with-dec-threads, to enable DEC threads on DEC Alpha platforms. Also, --with-threads is now an alias for --with-thread (this was the Most Common Typo in configure arguments). - Many changes in Doc/Makefile; amongst others, latex2html is now used to generate HTML from all latex documents. Change to the Python/C API -------------------------- - Because some interfaces have changed, the PYTHON_API macro has been bumped. Most extensions built for the old API version will still run, but I can't guarantee this. Python prints a warning message on version mismatches; it dumps core when the version mismatch causes a serious problem :-) - I've completed the Grand Renaming, with the help of Roger Masse and Barry Warsaw. This makes reading or debugging the code much easier. Many other unrelated code reorganizations have also been carried out. The allobjects.h header file is gone; instead, you would have to include Python.h followed by rename2.h. But you're better off running Tools/scripts/fixcid.py -s Misc/RENAME on your source, so you can omit the rename2.h; it will disappear in the next release. - Various and sundry small bugs in the "abstract" interfaces have been fixed. Thanks to all the (involuntary) testers of the Python 1.4 version! Some new functions have been added, e.g. PySequence_List(o), equivalent to list(o) in Python. - New API functions PyLong_FromUnsignedLong() and PyLong_AsUnsignedLong(). - The API functions in the file cgensupport.c are no longer supported. This file has been moved to Modules and is only ever compiled when the SGI specific 'gl' module is built. - PyObject_Compare() can now raise an exception. Check with PyErr_Occurred(). The comparison function in an object type may also raise an exception. - The slice interface uses an upper bound of INT_MAX when no explicit upper bound is given (e.x. for a[1:]). It used to ask the object for its length and do the calculations. - Support for multiple independent interpreters. See Doc/api.tex, functions Py_NewInterpreter() and Py_EndInterpreter(). Since the documentation is incomplete, also see the new Demo/pysvr example (which shows how to use these in a threaded application) and the source code. - There is now a Py_Finalize() function which "de-initializes" Python. It is possible to completely restart the interpreter repeatedly by calling Py_Finalize() followed by Py_Initialize(). A change of functionality in Py_Initialize() means that it is now a fatal error to call it while the interpreter is already initialized. The old, half-hearted Py_Cleanup() routine is gone. Use of Py_Exit() is deprecated (it is nothing more than Py_Finalize() followed by exit()). - There are no known memory leaks left. While Py_Finalize() doesn't free *all* allocated memory (some of it is hard to track down), repeated calls to Py_Finalize() and Py_Initialize() do not create unaccessible heap blocks. - There is now explicit per-thread state. (Inspired by, but not the same as, Greg Stein's free threading patches.) - There is now better support for threading C applications. There are now explicit APIs to manipulate the interpreter lock. Read the source or the Demo/pysvr example; the new functions are PyEval_{Acquire,Release}{Lock,Thread}(). - The test macro DEBUG has changed to Py_DEBUG, to avoid interference with other libraries' DEBUG macros. Likewise for any other test macros that didn't yet start with Py_. - New wrappers around malloc() and friends: Py_Malloc() etc. call malloc() and call PyErr_NoMemory() when it fails; PyMem_Malloc() call just malloc(). Use of these wrappers could be essential if multiple memory allocators exist (e.g. when using certain DLL setups under Windows). (Idea by Jim Fulton.) - New C API PyImport_Import() which uses whatever __import__() hook that is installed for the current execution environment. By Jim Fulton. - It is now possible for an extension module's init function to fail non-fatally, by calling one of the PyErr_* functions and returning. - The PyInt_AS_LONG() and PyFloat_AS_DOUBLE() macros now cast their argument to the proper type, like the similar PyString macros already did. (Suggestion by Marc-Andre Lemburg.) Similar for PyList_GET_SIZE and PyList_GET_ITEM. - Some of the Py_Get* function, like Py_GetVersion() (but not yet Py_GetPath()) are now declared as returning a const char *. (More should follow.) - Changed the run-time library to check for exceptions after object comparisons. PyObject_Compare() can now return an exception; use PyErr_Occurred() to check (there is *no* special return value). - PyFile_WriteString() and Py_Flushline() now return error indicators instead of clearing exceptions. This fixes an obscure bug where using these would clear a pending exception, discovered by Just van Rossum. - There's a new function, PyArg_ParseTupleAndKeywords(), which parses an argument list including keyword arguments. Contributed by Geoff Philbrick. - PyArg_GetInt() is gone. - It's no longer necessary to include graminit.h when calling one of the extended parser API functions. The three public grammar start symbols are now in Python.h as Py_single_input, Py_file_input, and Py_eval_input. - The CObject interface has a new function, PyCObject_Import(module, name). It calls PyCObject_AsVoidPtr() on the object referenced by "module.name". Tkinter ------- - On popular demand, _tkinter once again installs a hook for readline that processes certain Tk events while waiting for the user to type (using PyOS_InputHook). - A patch by Craig McPheeters plugs the most obnoxious memory leaks, caused by command definitions referencing widget objects beyond their lifetime. - New standard dialog modules: tkColorChooser.py, tkCommonDialog.py, tkMessageBox.py, tkFileDialog.py, tkSimpleDialog.py These interface with the new Tk dialog scripts, and provide more "native platform" style file selection dialog boxes on some platforms. Contributed by Fredrik Lundh. - Tkinter.py: when the first Tk object is destroyed, it sets the hiddel global _default_root to None, so that when another Tk object is created it becomes the new default root. Other miscellaneous changes and fixes. - The Image class now has a configure method. - Added a bunch of new winfo options to Tkinter.py; we should now be up to date with Tk 4.2. The new winfo options supported are: mananger, pointerx, pointerxy, pointery, server, viewable, visualid, visualsavailable. - The broken bind() method on Canvas objects defined in the Canvas.py module has been fixed. The CanvasItem and Group classes now also have an unbind() method. - The problem with Tkinter.py falling back to trying to import "tkinter" when "_tkinter" is not found has been fixed -- it no longer tries "tkinter", ever. This makes diagnosing the problem "_tkinter not configured" much easier and will hopefully reduce the newsgroup traffic on this topic. - The ScrolledText module once again supports the 'cnf' parameter, to be compatible with the examples in Mark Lutz' book (I know, I know, too late...) - The _tkinter.c extension module has been revamped. It now support Tk versions 4.1 through 8.0; support for 4.0 has been dropped. It works well under Windows and Mac (with the latest Tk ports to those platforms). It also supports threading -- it is safe for one (Python-created) thread to be blocked in _tkinter.mainloop() while other threads modify widgets. To make the changes visible, those threads must use update_idletasks()method. (The patch for threading in 1.5a3 was broken; in 1.5a4, it is back in a different version, which requires access to the Tcl sources to get it to work -- hence it is disabled by default.) - A bug in _tkinter.c has been fixed, where Split() with a string containing an unmatched '"' could cause an exception or core dump. - Unfortunately, on Windows and Mac, Tk 8.0 no longer supports CreateFileHandler, so _tkinter.createfilehandler is not available on those platforms when using Tk 8.0 or later. I will have to rethink how to interface with Tcl's lower-level event mechanism, or with its channels (which are like Python's file-like objects). Jack Jansen has provided a fix for the Mac, so createfilehandler *is* actually supported there; maybe I can adapt his fix for Windows. Tools and Demos --------------- - A new regression test suite is provided, which tests most of the standard and built-in modules. The regression test is run by invoking the script Lib/test/regrtest.py. Barry Warsaw wrote the test harnass; he and Roger Masse contributed most of the new tests. - New tool: faqwiz -- the CGI script that is used to maintain the Python FAQ (http://grail.cnri.reston.va.us/cgi-bin/faqw.py). In Tools/faqwiz. - New tool: webchecker -- a simple extensible web robot that, when aimed at a web server, checks that server for dead links. Available are a command line utility as well as a Tkinter based GUI version. In Tools/webchecker. A simplified version of this program is dissected in my article in O'Reilly's WWW Journal, the issue on Scripting Languages (Vol 2, No 2); Scripting the Web with Python (pp 97-120). Includes a parser for robots.txt files by Skip Montanaro. - New small tools: cvsfiles.py (prints a list of all files under CVS n a particular directory tree), treesync.py (a rather Guido-specific script to synchronize two source trees, one on Windows NT, the other one on Unix under CVS but accessible from the NT box), and logmerge.py (sort a collection of RCS or CVS logs by date). In Tools/scripts. - The freeze script now also works under Windows (NT). Another feature allows the -p option to be pointed at the Python source tree instead of the installation prefix. This was loosely based on part of xfreeze by Sam Rushing and Bill Tutt. - New examples (Demo/extend) that show how to use the generic extension makefile (Misc/Makefile.pre.in). - Tools/scripts/h2py.py now supports C++ comments. - Tools/scripts/pystone.py script is upgraded to version 1.1; there was a bug in version 1.0 (distributed with Python 1.4) that leaked memory. Also, in 1.1, the LOOPS variable is incremented to 10000. - Demo/classes/Rat.py completely rewritten by Sjoerd Mullender. Windows (NT and 95) ------------------- - New project files for Developer Studio (Visual C++) 5.0 for Windows NT (the old VC++ 4.2 Makefile is also still supported, but will eventually be withdrawn due to its bulkiness). - See the note on the new module search path in the "Miscellaneous" section above. - Support for Win32s (the 32-bit Windows API under Windows 3.1) is basically withdrawn. If it still works for you, you're lucky. - There's a new extension module, msvcrt.c, which provides various low-level operations defined in the Microsoft Visual C++ Runtime Library. These include locking(), setmode(), get_osfhandle(), set_osfhandle(), and console I/O functions like kbhit(), getch() and putch(). - The -u option not only sets the standard I/O streams to unbuffered status, but also sets them in binary mode. (This can also be done using msvcrt.setmode(), by the way.) - The, sys.prefix and sys.exec_prefix variables point to the directory where Python is installed, or to the top of the source tree, if it was run from there. - The various os.path modules (posixpath, ntpath, macpath) now support passing more than two arguments to the join() function, so os.path.join(a, b, c) is the same as os.path.join(a, os.path.join(b, c)). - The ntpath module (normally used as os.path) supports ~ to $HOME expansion in expanduser(). - The freeze tool now works on Windows. - See also the Tkinter category for a sad note on _tkinter.createfilehandler(). - The truncate() method for file objects now works on Windows. - Py_Initialize() is no longer called when the DLL is loaded. You must call it yourself. (And you can't call it twice -- it's a fatal error to call it when Python is already initialized.) - The time module's clock() function now has good precision through the use of the Win32 API QueryPerformanceCounter(). - Mark Hammond will release Python 1.5 versions of PythonWin and his other Windows specific code: the win32api extensions, COM/ActiveX support, and the MFC interface. Mac --- - As always, the Macintosh port will be done by Jack Jansen. He will make a separate announcement for the Mac specific source code and the binary distribution(s) when these are ready. ====================================================================== Fixed after 1.5a3 was released ------------------------------ The following changes have been made to the source base after the release of 1.5a3. They still need to be sorted. They also need to be merged into their respective categories for the final release, but it's useful to have them separately during the alpha test cycle. - faqwiz.py: version 0.8; Recognize https:// as URL; ... feature; better install instructions; removed faqmain.py (which was an older version). - nntplib.py: Fixed some bugs reported by Lars Wirzenius (to Debian) about the treatment of lines starting with '.'. Added a minimal test function. - struct module: ignore most whitespace in format strings. - urllib.py: close the socket and temp file in URLopener.retrieve() so that multiple retrievals using the same connection work. - All standard exceptions are now classes by default; use -X to make them strings (for backward compatibility only). - There's a new standard exception hierarchy, defined in the standard library module exceptions.py (which you never need to import explicitly). See http://grail.cnri.reston.va.us/python/essays/stdexceptions.html for more info. - Three new C API functions: - int PyErr_GivenExceptionMatches(obj1, obj2) Returns 1 if obj1 and obj2 are the same object, or if obj1 is an instance of type obj2, or of a class derived from obj2 - int PyErr_ExceptionMatches(obj) Higher level wrapper around PyErr_GivenExceptionMatches() which uses PyErr_Occurred() as obj1. This will be the more commonly called function. - void PyErr_NormalizeException(typeptr, valptr, tbptr) Normalizes exceptions, and places the normalized values in the arguments. If type is not a class, this does nothing. If type is a class, then it makes sure that value is an instance of the class by: 1. if instance is of the type, or a class derived from type, it does nothing. 2. otherwise it instantiates the class, using the value as an argument. If value is None, it uses an empty arg tuple, and if the value is a tuple, it uses just that. - Another new C API function: PyErr_NewException() creates a new exception class derived from Exception; when -X is given, it creates a new string exception. - core interpreter: remove the distinction between tuple and list unpacking; allow an arbitrary sequence on the right hand side of any unpack instruction. (UNPACK_LIST and UNPACK_TUPLE now do the same thing, which should really be called UNPACK_SEQUENCE.) - classes: Allow assignments to an instance's __dict__ or __class__, so you can change ivars (including shared ivars -- shock horror) and change classes dynamically. Also make the check on read-only attributes of classes less draconic -- only the specials names __dict__, __bases__, __name__ and __{get,set,del}attr__ can't be assigned. - Two new built-in functions: issubclass() and isinstance(). Both take classes as their second arguments. The former takes a class as the first argument and returns true iff first is second, or is a subclass of second. The latter takes any object as the first argument and returns true iff first is an instance of the second, or any subclass of second. - configure: Added configuration tests for presence of alarm(), pause(), and getpwent(). - Doc/Makefile: changed latex2html targets. - classes: Reverse the search order for the Don Beaudry hook so that the first class with an applicable hook wins. Makes more sense. - Changed the checks made in Py_Initialize() and Py_Finalize(). It is now legal to call these more than once. The first call to Py_Initialize() initializes, the first call to Py_Finalize() finalizes. There's also a new API, Py_IsInitalized() which checks whether we are already initialized (in case you want to leave things as they were). - Completely disable the declarations for malloc(), realloc() and free(). Any 90's C compiler has these in header files, and the tests to decide whether to suppress the declarations kept failing on some platforms. - *Before* (instead of after) signalmodule.o is added, remove both intrcheck.o and sigcheck.o. This should get rid of warnings in ar or ld on various systems. - Added reop to PC/config.c - configure: Decided to use -Aa -D_HPUX_SOURCE on HP-UX platforms. Removed outdated HP-UX comments from README. Added Cray T3E comments. - Various renames of statically defined functions that had name conflicts on some systems, e.g. strndup (GNU libc), join (Cray), roundup (sys/types.h). - urllib.py: Interpret three slashes in file: URL as local file (for Netscape on Windows/Mac). - copy.py: Make sure the objects returned by __getinitargs__() are kept alive (in the memo) to avoid a certain kind of nasty crash. (Not easily reproducable because it requires a later call to __getinitargs__() to return a tuple that happens to be allocated at the same address.) - Added definition of AR to toplevel Makefile. Renamed @buildno temp file to buildno1. - Moved Include/assert.h to Parser/assert.h, which seems to be the only place where it's needed. - Alas, the thread support for _tkinter didn't work. Withdrew it. - Tweaked the dictionary lookup code again for some more speed (Vladimir Marangozov). - NT build: Changed the way python15.lib is included in the other projects. Per Mark Hammond's suggestion, add it to the extra libs in Settings instead of to the project's source files. - regrtest.py: Change default verbosity so that there are only three levels left: -q, default and -v. In default mode, the name of each test is now printed. -v is the same as the old -vv. -q is more quiet than the old default mode. - Removed the old FAQ from the distribution. You now have to get it from the web! - Removed the PC/make_nt.in file from the distribution; it is no longer needed. - Changed the build sequence so that shared modules are built last. This fixes things for AIX and doesn't hurt elsewhere. - Improved test for GNU MP v1 in mpzmodule.c - fileobject.c: ftell() on Linux discards all buffered data; changed read() code to use lseek() instead to get the same effect - configure.in, configure, importdl.c: NeXT sharedlib fixes - tupleobject.c: PyTuple_SetItem asserts refcnt==1 - resource.c: Different strategy regarding whether to declare getrusage() and getpagesize() -- #ifdef doesn't work, Linux has conflicting decls in its headers. Choice: only declare the return type, not the argument prototype, and not on Linux. - importdl.c, configure*: set sharedlib extensions properly for NeXT - configure*, Makefile.in, Modules/Makefile.pre.in: AIX shared libraries fixed; moved addition of PURIFY to LINKCC to configure - reopmodule.c, regexmodule.c, regexpr.c, zlibmodule.c: needed casts added to shup up various compilers. - _tkinter.c: removed buggy mac #ifndef - Doc: various Mac documentation changes, added docs for 'ic' module - PC/make_nt.in: deleted - test_time.py, test_strftime.py: tweaks to catch %Z (which may return "") - test_rotor.py: print b -> print `b` - Tkinter.py: (tagOrId) -> (tagOrId,) - Tkinter.py: the Tk class now also has a configure() method and friends (they have been moved to the Misc class to accomplish this). - dict.get(key[, default]) returns dict[key] if it exists, or default if it doesn't. The default defaults to None. This is quicker for some applications than using either has_key() or try:...except KeyError:.... - Tools/webchecker/: some small changes to webchecker.py; added websucker.py (a simple web site mirroring script). - Dictionary objects now have a get() method (also in UserDict.py). dict.get(key, default) returns dict[key] if it exists and default otherwise; default defaults to None. - Tools/scripts/logmerge.py: print the author, too. - Changes to import: support for "import a.b.c" is now built in. See http://grail.cnri.reston.va.us/python/essays/packages.html for more info. Most important deviations from "ni.py": __init__.py is executed in the package's namespace instead of as a submodule; and there's no support for "__" or "__domain__". Note that "ni.py" is not changed to match this -- it is simply declared obsolete (while at the same time, it is documented...:-( ). Unfortunately, "ihooks.py" has not been upgraded (but see "knee.py" for an example implementation of hierarchical module import written in Python). - More changes to import: the site.py module is now imported by default when Python is initialized; use -S to disable it. The site.py module extends the path with several more directories: site-packages inside the lib/python1.5/ directory, site-python in the lib/ directory, and pathnames mentioned in *.pth files found in either of those directories. See http://grail.cnri.reston.va.us/python/essays/packages.html for more info. - Changes to standard library subdirectory names: those subdirectories that are not packages have been renamed with a hypen in their name, e.g. lib-tk, lib-stdwin, plat-win, plat-linux2, plat-sunos5, dos-8x3. The test suite is now a package -- to run a test, you must now use "import test.test_foo". - A completely new re.py module is provided (thanks to Andrew Kuchling) which uses Philip Hazel's "pcre" re compiler and engine. For a while, the "old" re.py (which was new in 1.5a3!) will be kept around as re1.py. The "old" regex module and underlying parser and engine are still present -- while regex is now officially obsolete, it will probably take several major release cycles before it can be removed. - The posix module now has a strerror() function which translates an error code to a string. - The emacs.py module (which was long obsolete) has been removed. - The universal makefile Misc/Makefile.pre.in now features an "install" target. By default, installed shared libraries go into $exec_prefix/lib/python$VERSION/site-packages/. - The install-sh script is installed with the other configuration specific files (in the config/ subdirectory). - It turns out whatsound.py and sndhdr.py were identical modules. Since there's also an imghdr.py file, I propose to make sndhdr.py the official one. For compatibility, whatsound.py imports * from sndhdr.py. - Class objects have a new attribute, __module__, giving the name of the module in which they were declared. This is useful for pickle and for printing the full name of a class exception. - Many extension modules no longer issue a fatal error when their initialization fails; the importing code now checks whether an error occurred during module initialization, and correctly propagates the exception to the import statement. - Most extension modules now raise class-based exceptions (except when -X is used). - Subtle changes to PyEval_{Save,Restore}Thread(): always swap the thread state -- just don't manipulate the lock if it isn't there. - Fixed a bug in Python/getopt.c that made it do the wrong thing when an option was a single '-'. Thanks to Andrew Kuchling. - New module mimetypes.py will guess a MIME type from a filename's extension. - Windows: the DLL version is now settable via a resource rather than being hardcoded. This can be used for "branding" a binary Python distribution. - urllib.py is now threadsafe -- it now uses re instead of regex, and sys.exc_info() instead of sys.exc_{type,value}. - Many other library modules that used to use sys.exc_{type,value,traceback} are now more thread-safe by virtue of using sys.exc_info(). - The functions in popen2 have an optional buffer size parameter. Also, the command argument can now be either a string (passed to the shell) or a list of arguments (passed directly to execv). - The thread support for _tkinter.c now works. The bad news is that it requires a modified version of a file in the standard Tcl distribution, which you must compile with a -I option pointing to the standard Tcl source tree. For this reason, the thread support is disabled by default. - The errno extension module adds two tables: errorcode maps errno numbers to errno names (e.g. EINTR), and errorcode maps them to message strings. (The latter is redundant because the new call posix.strerror() now does the same, but alla...) (Marc-Andre Lemburg) - The readline extension module now provides some interfaces to internal readline routines that make it possible to write a completer in Python. An example completer, rlcompleter.py, is provided. When completing a simple identifier, it completes keywords, built-ins and globals in __main__; when completing NAME.NAME..., it evaluates (!) the expression up to the last dot and completes its attributes. It's very cool to do "import string" type "string.", hit the completion key (twice), and see the list of names defined by the string module! Tip: to use the tab key as the completion key, call readline.parse_and_bind("tab: complete") - The traceback.py module has a new function tb_lineno() by Marc-Andre Lemburg which extracts the line number from the linenumber table in the code object. Apparently the traceback object doesn't contains the right linenumber when -O is used. Rather than guessing whether -O is on or off, the module itself uses tb_lineno() unconditionally. - Fixed Demo/tkinter/matt/canvas-moving-or-creating.py: change bind() to tag_bind() so it works again. - The pystone script is now a standard library module. Example use: "import test.pystone; test.pystone.main()". - The import of the readline module in interactive mode is now also attempted when -i is specified. (Yes, I know, giving in to Marc-Andre Lemburg, who asked for this. :-) - rfc822.py: Entirely rewritten parseaddr() function by Sjoerd Mullender, to be closer to the standard. This fixes the getaddr() method. Unfortunately, getaddrlist() is as broken as ever, since it splits on commas without regard for RFC 822 quoting conventions. - pprint.py: correctly emit trailing "," in singleton tuples. - _tkinter.c: export names for its type objects, TkappType and TkttType. - pickle.py: use __module__ when defined; fix a particularly hard to reproduce bug that confuses the memo when temporary objects are returned by custom pickling interfaces; and a semantic change: when unpickling the instance variables of an instance, use inst.__dict__.update(value) instead of a for loop with setattr() over the value.keys(). This is more consistent (the pickling doesn't use getattr() either but pickles inst.__dict__) and avoids problems with instances that have a __setattr__ hook. But it *is* a semantic change (because the setattr hook is no longer used). So beware! - config.h is now installed (at last) in $exec_prefix/include/python1.5/. For most sites, this means that it is actually in $prefix/include/python1.5/, with all the other Python include files, since $prefix and $exec_prefix are the same by default. - The imp module now supports parts of the functionality to implement import of hierarchical module names. It now supports find_module() and load_module() for all types of modules. Docstrings have been added for those functions in the built-in impo module that are still relevant (some old interfaces are obsolete). For a sample implementation of hierarchical module import in Python, see the new library module knee.py. - The % operator on string objects now allows arbitrary nested parens in a %(...)X style format. (Brad Howes) - Reverse the order in which Setup and Setup.local are passed to the makesetup script. This allows variable definitions in Setup.local to override definitions in Setup. (But you'll still have to edit Setup if you want to disable modules that are enabled by default, or if such modules need non-standard options.) - Added PyImport_ImportModuleEx(name, globals, locals, fromlist); this is like PyImport_ImporModule(name) but receives the globals and locals dict and the fromlist arguments as well. (The name is a char*; the others are PyObject*s). - The 'p' format in the struct extension module alloded to above is new in 1.5a4. - The xlib extension module has been debugged and improved. - The types.py module now uses try-except in a few places to make it more likely that it can be imported in restricted mode. Some type names are undefined in that case, e.g. CodeType (inaccessible), FileType (not always accessible), and TracebackType and FrameType (inaccessible). - In urllib.py: added separate administration of temporary files created y URLopener.retrieve() so cleanup() can properly remove them. The old code removed everything in tempcache which was a bad idea if the user had passed a non-temp file into it. Also, in basejoin(), interpret relative paths starting in "../". This is necessary if the server uses symbolic links. - The Windows build procedure and project files are now based on Microsoft Visual C++ 5.x. The build now takes place in the PCbuild directory. It is much more robust, and properly builds separate Debug and Release versions. (The installer will be added shortly.) - Added casts and changed some return types in regexpr.c to avoid compiler warnings or errors on some platforms. - The AIX build tools for shared libraries now supports VPATH. (Donn Cave) - By default, disable the "portable" multimedia modules audioop, imageop, and rgbimg, since they don't work on 64-bit platforms. - Fixed a nasty bug in cStringIO.c when code was actually using the close() method (the destructors would try to free certain fields a second time). - For those who think they need it, there's a "user.py" module. This is *not* imported by default, but can be imported to run user-specific setup commands, ~/.pythonrc.py. - Various speedups suggested by Fredrik Lundh, Marc-Andre Lemburg, Vladimir Marangozov, and others. - Added os.altsep; this is '/' on DOS/Windows, and None on systems with a sane filename syntax. - os.py: Write out the dynamic OS choice, to avoid exec statements. Adding support for a new OS is now a bit more work, but I bet that 'dos' or 'nt' will cover most situations... - The obsolete exception AccessError is now really gone. - Tools/faqwiz/: New installation instructions show how to maintain multiple FAQs. Removed bootstrap script from end of faqwiz.py module. Added instructions to bootstrap script, too. Version bumped to 0.8. Added ... feature suggested by Skip Montanaro. Added leading text for Roulette, default to 'Hit Reload ...'. Fix typo in default SRCDIR. - Documentation for the relatively new modules "keyword" and "symbol" has been added (to the end of the section on the parser extension module).