Recode

Recode copied to clipboard

#+TITLE: README for Recode #+OPTIONS: H:2 toc:2

• Introduction ** What is Recode? Here is version 3.6 for the Recode program and library. Hereafter, Recode means the whole package, recode means the executable program. Glance through this =README= file before starting configuration. Make sure you read files =ABOUT-NLS= and =INSTALL= if you are not familiar with them already.

The Recode library converts files between character sets and usages. It recognises or produces over 200 different character sets (or about 300 if combined with an iconv library) and transliterates files between almost any pair. When exact transliteration are not possible, it gets rid of offending characters or falls back on approximations. The recode program is a handy front-end to the library.

The Recode program and library have been written by François Pinard, yet it significantly reuses tabular works from Keld Simonsen. It is an evolving package, and specifications might change in future releases.

On various Unix systems, Recode is usually compiled from sources, see the [[Installation]] section below. On Linux, it often comes bundled. Recode had been ported to other popular systems. See both [[http:/contrib.html][contrib/README]] and the [[Non-Unix ports]] section below, to find some more information about these.

** Reports and collaboration Send bug reports to [[mailto:[email protected]][this address]], or if you are comfortable with GitHub facilities, through [[https://github.com/pinard/Recode/issues][GitHub Issues]]. A bug report is an adequate description of the problem: your input, what you expected, what you got, and why this is wrong. Diffs are welcome, but they only describe a solution, from which the problem might be uneasy to infer. If needed, submit actual data files with your report. Small data files are preferred. Big files may sometimes be necessary, but do not send them on the mailing list; rather take special arrangement with the maintainer.

Your feedback will help us to make a better and more portable package. Consider documentation errors as bugs, and report them as such. If you develop anything pertaining to Recode or have suggestions, let us know and share your findings by writing at [[mailto:[email protected]][Recode forum]]. You may also choose to directly write to [[mailto:[email protected]][the author]], yet be warned that such correspondence is often visible for a while through the Recode Web site.

If you feel like receiving releases and pretest announcements for the Recode package, send a message to [[mailto:[email protected]][this Majordomo]] having, in its body, a line saying:

#+BEGIN_EXAMPLE subscribe recode-announce #+END_EXAMPLE

If you rather want to participate actively in discussions, pretesting and development for Recode, do just as above, but this time, use:

#+BEGIN_EXAMPLE subscribe recode-forum #+END_EXAMPLE

The [[https://github.com/pinard/Recode][Git repository]] for Recode gives access, through the magic of Git and GitHub, to all distribution releases, would they be actual or past, pretest or official, as well as individual files.

Please /do not/ widely redistribute releases having a letter after the version numbers, as these are meant for pretesting only, and might not be stable enough for other usages.

• Release notes ** Notes for version 3.7-beta2 Here are a few notes related to the beta2 pre-test release for the incoming Recode 3.7. I publish it to ease later exchanges of patches with testers.

• Long ago, I renamed /GNU recode/ to /Free recode/: the permission for using the /GNU/ prefix mandated a level of obedience to the FSF that once went overboard, in my opinion. After that change, I realized that some people read /Free/ as a four letter word! To be peaceful, this version changes the name again, from /Free recode/ to merely /Recode/. recode (no capital) still names the executable program specifically, or the distribution archive itself.

• Recode does not itself include libiconv anymore. However, it uses an external iconv library if one is available at installation time, like libiconv or the one provided within GNU libc. The =-x:= option to the program, or a new flag to the library recode_new_outer function, inhibits the initialisation and usage of iconv.

• The bug about loosing a few characters, here and there, when recoding big files in iconv context, seems to have been corrected. A patch for this problem has been floating around for years, but it was not solving all cases.

• Recode installation now uses Python. In particular, it creates file =build/src/iconvdecl.h= from local =iconv -l= output. Recode testing through =make check= also needs what people usually find as the python-devel package, which provides C header files for Python and distutils. The =Makemore= file has been merged within regular Makefiles and is not distributed separately anymore.

• It is likely that new bugs have been introduced through the above changes. In particular, not everything is cosy on the side of release engineering. A few files are either spuriously remade, or remade late. I'm a bit surprised by the difficulty to get this right.

• =make check= accepts a =LIMIT== option, for limiting tests to one or a few cases. See =tests/Makefile= for more information.

• PO files have been updated from the Translation Project.

** Notes for version 3.7-beta1 The beta 1 pre-test release for the incoming Recode 3.7 has been made available for those needing it right away. While it solves some serious bugs and portability problems, others are meant to be addressed only in later pre-tests. In particular, none of charset or surface issues, user requests, and various suggestions appear in this pre-test, and will not either in later pretests, until all real show-stoppers are solved first. So this is in no way a candidate for a Recode 3.7 release.

The test suite is worth more comments:

• The suite is very partial, and may not be thought as a validation suite. Before it could be used to ascertain confidence, it would need much more tests than it has already.

• Testing is notably more speedy than it used to be. For example, the previous bigauto test, which was not run by default because it ran for too long, is now executed within the standard test suite, once in non-strict mode, and a second time in strict mode.

• It does not use Autotest anymore, but rather a home grown test driver much inspired from the Codespeak project. The link between the test and the Recode library is established through a Pyrex interface, so you need to have python and python-devel installed first.

• Beware that the Pyrex interface to the Recode library is only meant for testing, for now at least. While you may play with it, it would not be wise relying on it, as the specifications might change at any time.

** Non-Unix ports Please [[mailto:[email protected]][inform us]] if you are aware of various ports to non-Unix systems not listed here, or for corrections. Please provide the goal system, a complete and stable URL, the maintainer name and address, the Recode version used as a base, and your comments.

• MSDOS (DJGPP) :: [[mailto:[email protected]][Juan Manuel Guerrero]] maintains this port, dated 2001-03 and based on Recode 3.5. The following archives hold binaries, docs and sources respectively. See [[ftp://ftp.simtel.net/pub/simtelnet/gnu/djgpp/v2gnu/rcode35b.zip][rcode35b]], [[ftp://ftp.simtel.net/pub/simtelnet/gnu/djgpp/v2gnu/rcode35d.zip][rcode35d]] and [[ftp://ftp.simtel.net/pub/simtelnet/gnu/djgpp/v2gnu/rcode35s.zip][rcode35s]]. Also see [[http:/DJGPP.html][contrib/DJGPP/README]] in the Recode distribution for more information about compiling this port.
• MSDOS (Gnuish) :: [[mailto:[email protected]][Darrel Hankerson]] maintains this port, dated 1994-11 and based on Recode 3.4. You get many GNU tools, not only Recode. The GNUish project is described in =gnuish_t.htm=. See [[http://www.simtel.net/simtel.net/][simtel]] and [[http://www.leo.org/pub/comp/platforms/pc/gnuish][gnuish]] (Germany), or for the FTP versions: [[ftp://ftp.simtel.net/simtelnet/gnu][simtel]] and [[ftp://ftp.leo.org/pub/comp/platforms/pc/gnuish][gnuish]].
• OS/2 (using emx/gcc) :: Maintainer unknown (maybe [[mailto:[email protected]][Kai Uwe Rommel]]), dated 1994-11 and based on Recode 3.4. See [[http://hobbes.nmsu.edu/pub/os2/util/convert/gnurcode.zip][gnurcode]].

• Installation ** In a hurry? You may then try:

#+BEGIN_SRC sh git clone https://github.com/pinard/Recode.git cd Recode sh after-git.sh ./configure make make install #+END_SRC

More fine-grained instructions follow.

** Prerequisites Simple installation of Recode requires the usual tools and facilities as those needed for most GNU packages. If not already bundled with your system, you also need to pre-install [[http://www.python.org][Python]], version 2.2 or better.

It is also convenient to have some iconv library already present on your system, this much extends Recode capabilities, especially in the area of Asiatic character sets. GNU libc, as found on Linux systems and a few others, already has such an iconv library. Otherwise, you might consider pre-installing the [[http://www.gnu.org/software/libiconv/][portable libiconv]], written by Bruno Haible.

** Getting a release Source files and various distributions (either latest, prestest, or archive) are available through [[https://github.com/pinard/Recode/][GitHub]].

File timestamps after checkout may trigger Make difficulties. As a way to avoid these, from the top level of the distribution, execute =sh after-patch.sh= before configuring. If you miss either sh or GNU touch, try =python after-patch.py= instead.

** Configure options Once you have an unpacked distribution, see files:

|-------------+------------------------------------------------| | File name | Description | |-------------+------------------------------------------------| | =ABOUT-NLS= | how to customise this program to your language | | =COPYING= | copying conditions for the program | | =COPYING.LIB= | copying conditions for the library | | =INSTALL= | compilation and installation instructions | | =NEWS= | major changes in the current release | | =THANKS= | partial list of contributors | |-------------+------------------------------------------------|

Besides those configure options documented in files =INSTALL= and =ABOUT-NLS=, a few extra options may be accepted after =./configure=:

• Options =--disable-shared= or =--disable-static=

  to inhibit the building of shared libraries or static libraries; the default is to always build static libraries, and to attempt building shared libraries if there is some known recipe for this.

• Option =--with-gnu-ld=

  to force the assumption that the C compiler uses GNU ld.

• Option =--with-dmalloc=

  to trigger a debugging feature for looking at memory management problems, it pre-requires Gray Watson's [[ftp://ftp.letters.com/src/dmalloc/dmalloc.tar.gz][dmalloc package]].

** Installation hints Here are a few hints which might help installing Recode on some systems. Many may be applied by temporary presetting environment variables while calling =./configure=. File =INSTALL= explains this.

• Compilation time

  Some C compilers, like Apollo's, have a hard time compiling =merged.c=. If this is your case, avoid compiler optimisation. From within the Bourne shell, you may use:

  #+BEGIN_EXAMPLE CFLAGS= ./configure #+END_EXAMPLE

  But if you want to give a real hard time to your C optimiser on =merged.c=, to get code that runs only a bit faster, merely try:

  #+BEGIN_EXAMPLE CPPFLAGS=-DINLINE_HARDER ./configure #+END_EXAMPLE

• Smallish systems

  For 80286 based systems (do some still exist?!), it has been reported that some compilers generate wrong code while optimising for /small/ models. So, from within the Bourne shell, do:

  #+BEGIN_EXAMPLE CFLAGS=-Ml LDFLAGS=-Ml ./configure #+END_EXAMPLE

  to force large memory model. For 80286 Xenix compiler, the last time it was tried a while ago, one ought to use:

  #+BEGIN_EXAMPLE CFLAGS='-Ml -F2000' LDFLAGS=-Ml ./configure #+END_EXAMPLE

  Other systems have poor pipe / popen support or thrash heavily when processes fork. In this case, just before doing =make=, edit =config.h= and ensure HAVE_PIPE is /not/ defined.

** Maintenance tools Beyond the usual Unix programs needed for configuring and installing any GNU package, you need Cython, Flex and Python to achieve simple modifications to Recode.

For more encompassing modifications, you might also need recent versions of Autoconf, automake, Flex, Gettext, Help2man, libtool, m4, GNU Make, Perl, tar and wget. Just make sure you install m4 before Autoconf, and Perl before automake or Help2man. You may also choose to establish a link in your build =doc/= directory, as explained within =doc/Makemore=.

• The future ** Motivation Recode is due for a major ovehaul. My plan is to end the 3.x series of this package, rather aiming 4.0 as a major internal rewrite.

For one thing, I want to explore some new avenues. It does not seem natural anymore, to me at least, using C code for exploring or prototyping new ideas requiring complex internal structures: encompassing changes are stretchy, work overhead is just too high. I want to add a run-time dependency between Recode and Python, with the admitted goal of shifting the internals of Recode from C to Python.

Another thing is that Recode should reuse more of the work of many competent people in the recoding area. I was brought into the business of character set conversion issues by a random set of coincidences and needs, but have never been a character set specialist myself. I rely on users to help me sketch what needs to be done. There are other tools and other maintainers who address these matters more competently than me. Recode might well rely on their work and better concentrate on user functionalities and on an overall picture.

For experimenting what Recode might become and experimenting new concepts more easily, I created a subsidiary and standalone Python project named [[https://github.com/pinard/Recodec][Recodec]], which reproduces a good part of Recode functionality. My goal is now to merge Recodec back into Recode soon, rather than slowly stretching the distance between Recode and Recodec. Recode is going to be a mix of Python, C and either Pyrex or Cython.

** Overall plan The release 3.6 for Recode was likely the last in the 3.x series. As there is still a long way before 4.0 gets ready, and /especially/ because some of my good collaborators insisted that I do so, there will likely be other Recode 3.X releases on the way to 4.0, at least to provide a selection of user-contributed patches. Also, the next releases of Recode will progressively implement the base mechanics for the transition, through a list of development steps similar to the following. By principle, the implementation should be working and usable at each devewlopment step. Moreover, for better maintainability, refactoring shall occur all along the way.

I'll likely select Cython over Pyrex, the main arguments being Unicode, Python 3 and pragmatic support, and a wide and active user base. Pyrex, the inspiration behind Cython, is amazingly well thought; I stay really admirative and grateful for Greg Ewing's work.

• The main program is written in Python, and through a Cython interface, calls the existing C API for doing the real work.
• The C API gets merely able to use Cython written steps internally, besides the actual C steps, but with no Cython steps yet.
• New Cython steps wrap many standard Python codecs, with some trickery to force Python codecs over actual, older Recode steps.
• Recode library initialization is moved from C to Python, and gets called through Cython from the C API.
• Initialization is extended to cover the Recodec Python API, which uses different tables and descriptive data.
• More steps from Recodec get moved into Recode, either coexisting with or taking over the previous wrapping of Python codecs.
• The remaining code from the Recodec engine gets moved into Recode, replacing C code having the same fonctionality.
• Special care is given to GNU libc or libiconv support, maybe going from the C side to the Python side.
• Proper documentation and decisions follow extensive comparison and diagnostic of multiple implementations of same charsets or surfaces.
• Profiling allows to fine tune when and how Cython gets used over Python; standard Python codecs might even be cythonized in Recode.
• Program and library initialization get revised to spare disk accesses and building descriptive structures, whenever possible.
• The main program directly links to the Python API rather than through the C API, while the C API becomes a separate facility.

I once thought about resorting to kludges, within a Python API interface, so the Python interpreter would not be required at all at run-time. Today, I doubt this is doable in practice, or that the implied restrictions on Cython code would be bearable. By the time, I may come to think that this is not worth the effort, anyway.

** Speed and memory So far, Recode has always been oriented towards some generality in specifications, combined with good execution speed. Generality is granted through providing recoding steps either as tables or fuller algorithms expressed as C code. Speed surely results from careful C coding of individual steps, and using Flex for more difficult recognition problems. Speed also comes from the monolithic design of a single Python-free, big executable executable holding all tables at once, relying on system paging instead of run-time opening of external data files.

Rewriting a character shuffling engine in Python is going to have consequences on both speed and memory. Python is inherently much slower than C for such problems, and program startup requires many disk accesses to load all required modules. The size of the Python interpreter is not negligible, yet Recode is not small as it stands.

Depending on how to declare things and the way to code on the Cython side, by relying less on the Python library, one may have some control over the compromise between speed and ease. With enough discipline, resisting the temptation to use many Python facilities, one can displace the equilibrium. I once dreamed of many stub or minimal routines for representing the Python library to the point of avoiding it, yet I now think it would imply too stringent limitations.

After much hesitation, I merely /decided/ that the slowdown is bearable! It was fairly tedious to make encompassing structural changes in the C version of Recode. Such changes are going to be significantly easier in Python. This might translate into shorter development cycles.

** Planned differences Whenever the Python library offers a charset or a surface which Recode also has, the Python library codec is used. In some cases, this introduces differences, those will have to be resolved one by one, either by accepting that the Python library does better, getting the Python team to improve some codecs, or overriding these from Recodec.

Other differences may occur, especially in the Asian charset area, from the fact libiconv, GNU libc recoding facilities, and various contributors to the Python codecs project, do not fully agree on how things should be done. Recodec is likely to offer configuration mechanisms to choose among various possibilities, but will not likely attempt to rule out who is right and who is wrong! ☺

Issues about reversibility and canonicity, which were much present in Recode 3.X, are fading out. While some of these were moderately easy to implement, other cases stayed pending as fairly difficult to solve without a significant loss of efficiency. I think these issues are better abandoned than forever kept as half-hearted and not wholly dependable. Any user concerned about such things might try the reverse coding to find out if the original file is recoverable, some new option might automate a (costly) reversibility test.

One drawback of the whole move is that the Global Interpreter Lock in Python gets in the way of parallel execution of the code. This would have been more of a concern if GNU libc recoding facilities were relying on the Recode library, but as things stand by now, I'm guessing that users will not be much impacted in practice.

• Other pointers ** Documentation

• IETF references

  • [[ftp://nic.ddn.mil/rfc/rfc1345.txt][Character Mnemonics & Character Sets]], by [[mailto:[email protected]][Keld Simonsen]], 1992-06.
  • [[ftp://nic.ddn.mil/rfc/rfc1642.txt][UTF-7 - A Mail-Safe Transformation Format of Unicode]], by [[mailto:[email protected]][David Goldsmith]] and [[mailto:[email protected]][Mark Davis]], 1994-07.
  • [[ftp://nic.ddn.mil/rfc/rfc2044.txt][UTF-8, a transformation format of Unicode and ISO 10646]], by [[mailto:[email protected]][François Yergeau]], 1997-10.

• Various references

  • [[ftp://ftp.unicode.org:/Public/MAPPINGS/][Unicode charset mappings]]. The Unicode consortium makes available plenty of charset mappings for converting /legacy/ charsets to Unicode.
  • [[ftp://ftp.iro.umontreal.ca/pub/contrib/pinard/accents/oqil-tome1.ps.gz][Normalisation et internationalisation: Inventaire et prospectives des normes clefs pour le traitement informatique du français.]] (392p.) or [[http://www.ceveil.qc.ca/Normes][this other copy]]. This is a report, written in French, discussing charset issues and many other topics as well. [[mailto:[email protected]][Laurent Bourbeau]] and [[mailto:[email protected]][François Pinard]], 1995-10.

• Recode specific

  • ETL presentation

    In 1999, the organisers of the [[http://www.m17n.org/conference/m17n99_all_but_registration/welcome.en.html][m17n99 conference]] in Tsukuba, Japan, were kind enough to invite me. This has been for me a fabulous trip and experience, and I met many extraordinary people in there. At the conference, I presented the Translation Project, and Recode. The Recode [[http:/m17n99.html][presentation slides]] are available.

** Programs

• libiconv :: This comprehensive [[http://www.gnu.org/software/libiconv/][charset converter library]], by [[mailto:[email protected]][Bruno Haible]], revolves around Unicode, and support Asian encodings among many others. Even Recode uses it!
• tcs :: Here is the [[ftp://research.att.com/dist/tcs.shar.Z][main recoding tool]] from the Plan9 project.
• yuedit :: This [[ftp://sunsite.unc.edu/pub/Linux/apps/editors/X/yudit-1.2.tar.gz][GUI editor]], by [[mailto:[email protected]][Gaspar Sinai]], 1999-01, handles many encodings, among which UTF-8. It also installs uniconv, a recoding program, and uniprint, a printing tool.
• ucs-fonts :: These [[http://www.cl.cam.ac.uk/~mgk25/download/ucs-fonts.tar.gz][6x13 fonts]], by [[mailto:[email protected]][Markus Kuhn]], 1998-11, covering Unicode characters besides the Asian sets, merely replace the Linux fixed 6x13 font. Works nicely with yudit.
• MtRecode :: This [[http://www.lpl.univ-aix.fr/projects/multext/MtRecode/][charset converter]] is oriented towards SGML text manipulation. It may be freely downloaded for non-commercial, non-military use. Pointer given by [[mailto:[email protected]][Jean Véronis]], 1996-06.
• sp :: This quite nice SGML [[ftp://ftp.jclark.com/pub/sp/sp-1.3.tar.gz][structure analyser]], by [[mailto:[email protected]][James Clark]], contains internal C++ modules for handling many charsets.
• b2c :: This [[http://research.de.uu.net:8080/~gnu/b2c/b2c-2.1.tar.gz][program]], by [[mailto:[email protected]][Jörg Heitkötter]], 1997-11, is able to generate interpreted character dumps, but properly embedded within complete C header files.
• PyRecode :: This [[http://www.suxers.de/PyRecode.tgz][wrapper]], by [[mailto:[email protected]][Andreas Jung]], provides Recode functionality to Python programs. Also see [[http://www.vex.net/parnassus/apyllo.py?find%3Drecode][this link]] and [[http://www.suxers.de/python/pyrecode.htm][this other link]].