============================================================================= Python Unicode Integration Proposal Version: 1.2 ----------------------------------------------------------------------------- Introduction: ------------- The idea of this proposal is to add native Unicode 3.0 support to Python in a way that makes use of Unicode strings as simple as possible without introducing too many pitfalls along the way. Since this goal is not easy to achieve -- strings being one of the most fundamental objects in Python --, we expect this proposal to undergo some significant refinements. Note that the current version of this proposal is still a bit unsorted due to the many different aspects of the Unicode-Python integration. The latest version of this document is always available at: http://starship.skyport.net/~lemburg/unicode-proposal.txt Older versions are available as: http://starship.skyport.net/~lemburg/unicode-proposal-X.X.txt Conventions: ------------ · In examples we use u = Unicode object and s = Python string · 'XXX' markings indicate points of discussion (PODs) General Remarks: ---------------- · Unicode encoding names should be lower case on output and case-insensitive on input (they will be converted to lower case by all APIs taking an encoding name as input). Encoding names should follow the name conventions as used by the Unicode Consortium: spaces are converted to hyphens, e.g. 'utf 16' is written as 'utf-16'. Codec modules should use the same names, but with hyphens converted to underscores, e.g. utf_8, utf_16, iso_8859_1. · The should be the widely used 'utf-8' format. This is very close to the standard 7-bit ASCII format and thus resembles the standard used programming nowadays in most aspects. Unicode Constructors: --------------------- Python should provide a built-in constructor for Unicode strings which is available through __builtins__: u = unicode(encoded_string[,encoding=][,errors="strict"]) u = u'' u = ur'' With the 'unicode-escape' encoding being defined as: · all non-escape characters represent themselves as Unicode ordinal (e.g. 'a' -> U+0061). · all existing defined Python escape sequences are interpreted as Unicode ordinals; note that \xXXXX can represent all Unicode ordinals, and \OOO (octal) can represent Unicode ordinals up to U+01FF. · a new escape sequence, \uXXXX, represents U+XXXX; it is a syntax error to have fewer than 4 digits after \u. For an explanation of possible values for errors see the Codec section below. Examples: u'abc' -> U+0061 U+0062 U+0063 u'\u1234' -> U+1234 u'abc\u1234\n' -> U+0061 U+0062 U+0063 U+1234 U+005c The 'raw-unicode-escape' encoding is defined as follows: · \uXXXX sequence represent the U+XXXX Unicode character if and only if the number of leading backslashes is odd · all other characters represent themselves as Unicode ordinal (e.g. 'b' -> U+0062) Note that you should provide some hint to the encoding you used to write your programs as pragma line in one the first few comment lines of the source file (e.g. '# source file encoding: latin-1'). If you only use 7-bit ASCII then everything is fine and no such notice is needed, but if you include Latin-1 characters not defined in ASCII, it may well be worthwhile including a hint since people in other countries will want to be able to read you source strings too. Unicode Type Object: -------------------- Unicode objects should have the type UnicodeType with type name 'unicode', made available through the standard types module. Unicode Output: --------------- Unicode objects have a method .encode([encoding=]) which returns a Python string encoding the Unicode string using the given scheme (see Codecs). print u := print u.encode() # using the str(u) := u.encode() # using the repr(u) := "u%s" % repr(u.encode('unicode-escape')) Also see Internal Argument Parsing and Buffer Interface for details on how other APIs written in C will treat Unicode objects. Unicode Ordinals: ----------------- Since Unicode 3.0 has a 32-bit ordinal character set, the implementation should provide 32-bit aware ordinal conversion APIs: ord(u[:1]) (this is the standard ord() extended to work with Unicode objects) --> Unicode ordinal number (32-bit) unichr(i) --> Unicode object for character i (provided it is 32-bit); ValueError otherwise Both APIs should go into __builtins__ just like their string counterparts ord() and chr(). Note that Unicode provides space for private encodings. Usage of these can cause different output representations on different machines. This problem is not a Python or Unicode problem, but a machine setup and maintenance one. Comparison & Hash Value: ------------------------ Unicode objects should compare equal to other objects after these other objects have been coerced to Unicode. For strings this means that they are interpreted as Unicode string using the . For the same reason, Unicode objects should return the same hash value as their UTF-8 equivalent strings. Coercion: --------- Using Python strings and Unicode objects to form new objects should always coerce to the more precise format, i.e. Unicode objects. u + s := u + unicode(s) s + u := unicode(s) + u All string methods should delegate the call to an equivalent Unicode object method call by converting all envolved strings to Unicode and then applying the arguments to the Unicode method of the same name, e.g. string.join((s,u),sep) := (s + sep) + u sep.join((s,u)) := (s + sep) + u For a discussion of %-formatting w/r to Unicode objects, see Formatting Markers. Exceptions: ----------- UnicodeError is defined in the exceptions module as subclass of ValueError. It is available at the C level via PyExc_UnicodeError. All exceptions related to Unicode encoding/decoding should be subclasses of UnicodeError. Codecs (Coder/Decoders) Lookup: ------------------------------- A Codec (see Codec Interface Definition) search registry should be implemented by a module "codecs": codecs.register(search_function) Search functions are expected to take one argument, the encoding name in all lower case letters, and return a tuple of functions (encoder, decoder, stream_reader, stream_writer) taking the following arguments: encoder and decoder: These must be functions or methods which have the same interface as the .encode/.decode methods of Codec instances (see Codec Interface). The functions/methods are expected to work in a stateless mode. stream_reader and stream_writer: These need to be factory functions with the following interface: factory(stream,errors='strict') The factory functions must return objects providing the interfaces defined by StreamWriter/StreamReader resp. (see Codec Interface). Stream codecs can maintain state. Possible values for errors are defined in the Codec section below. In case a search function cannot find a given encoding, it should return None. Aliasing support for encodings is left to the search functions to implement. The codecs module will maintain an encoding cache for performance reasons. Encodings are first looked up in the cache. If not found, the list of registered search functions is scanned. If no codecs tuple is found, a LookupError is raised. Otherwise, the codecs tuple is stored in the cache and returned to the caller. To query the Codec instance the following API should be used: codecs.lookup(encoding) This will either return the found codecs tuple or raise a LookupError. Standard Codecs: ---------------- Standard codecs should live inside an encodings/ package directory in the Standard Python Code Library. The __init__.py file of that directory should include a Codec Lookup compatible search function implementing a lazy module based codec lookup. Python should provide a few standard codecs for the most relevant encodings, e.g. 'utf-8': 8-bit variable length encoding 'utf-16': 16-bit variable length encoding (litte/big endian) 'utf-16-le': utf-16 but explicitly little endian 'utf-16-be': utf-16 but explicitly big endian 'ascii': 7-bit ASCII codepage 'iso-8859-1': ISO 8859-1 (Latin 1) codepage 'unicode-escape': See Unicode Constructors for a definition 'raw-unicode-escape': See Unicode Constructors for a definition 'native': Dump of the Internal Format used by Python Common aliases should also be provided per default, e.g. 'latin-1' for 'iso-8859-1'. Note: 'utf-16' should be implemented by using and requiring byte order marks (BOM) for file input/output. All other encodings such as the CJK ones to support Asian scripts should be implemented in seperate packages which do not get included in the core Python distribution and are not a part of this proposal. Codecs Interface Definition: ---------------------------- The following base class should be defined in the module "codecs". They provide not only templates for use by encoding module implementors, but also define the interface which is expected by the Unicode implementation. Note that the Codec Interface defined here is well suitable for a larger range of applications. The Unicode implementation expects Unicode objects on input for .encode() and .write() and character buffer compatible objects on input for .decode(). Output of .encode() and .read() should be a Python string and .decode() must return an Unicode object. First, we have the stateless encoders/decoders. These do not work in chunks as the stream codecs (see below) do, because all components are expected to be available in memory. class Codec: """ Defines the interface for stateless encoders/decoders. The .encode()/.decode() methods may implement different error handling schemes by providing the errors argument. These string values are defined: 'strict' - raise an error (or a subclass) 'ignore' - ignore the character and continue with the next 'replace' - replace with a suitable replacement character; Python will use the official U+FFFD REPLACEMENT CHARACTER for the builtin Unicode codecs. """ def encode(self,input,errors='strict'): """ Encodes the object intput and returns a tuple (output object, length consumed). errors defines the error handling to apply. It defaults to 'strict' handling. The method may not store state in the Codec instance. Use SteamCodec for codecs which have to keep state in order to make encoding/decoding efficient. """ ... def decode(self,input,errors='strict'): """ Decodes the object input and returns a tuple (output object, length consumed). input must be an object which provides the bf_getreadbuf buffer slot. Python strings, buffer objects and memory mapped files are examples of objects providing this slot. errors defines the error handling to apply. It defaults to 'strict' handling. The method may not store state in the Codec instance. Use SteamCodec for codecs which have to keep state in order to make encoding/decoding efficient. """ ... StreamWriter and StreamReader define the interface for stateful encoders/decoders which work on streams. These allow processing of the data in chunks to efficiently use memory. If you have large strings in memory, you may want to wrap them with cStringIO objects and then use these codecs on them to be able to do chunk processing as well, e.g. to provide progress information to the user. class StreamWriter(Codec): def __init__(self,stream,errors='strict'): """ Creates a StreamWriter instance. stream must be a file-like object open for writing (binary) data. The StreamWriter may implement different error handling schemes by providing the errors keyword argument. These parameters are defined: 'strict' - raise a ValueError (or a subclass) 'ignore' - ignore the character and continue with the next 'replace'- replace with a suitable replacement character """ self.stream = stream self.errors = errors def write(self,object): """ Writes the object's contents encoded to self.stream. """ data, consumed = self.encode(object,self.errors) self.stream.write(data) def reset(self): """ Flushes and resets the codec buffers used for keeping state. Calling this method should ensure that the data on the output is put into a clean state, that allows appending of new fresh data without having to rescan the whole stream to recover state. """ pass def __getattr__(self,name, getattr=getattr): """ Inherit all other methods from the underlying stream. """ return getattr(self.stream,name) class StreamReader(Codec): def __init__(self,stream,errors='strict'): """ Creates a StreamReader instance. stream must be a file-like object open for reading (binary) data. The StreamReader may implement different error handling schemes by providing the errors keyword argument. These parameters are defined: 'strict' - raise a ValueError (or a subclass) 'ignore' - ignore the character and continue with the next 'replace'- replace with a suitable replacement character; """ self.stream = stream self.errors = errors def read(self,size=-1): """ Decodes data from the stream self.stream and returns the resulting object. size indicates the approximate maximum number of bytes to read from the stream for decoding purposes. The decoder can modify this setting as appropriate. The default value -1 indicates to read and decode as much as possible. size is intended to prevent having to decode huge files in one step. The method should use a greedy read strategy meaning that it should read as much data as is allowed within the definition of the encoding and the given size, e.g. if optional encoding endings or state markers are available on the stream, these should be read too. """ # Unsliced reading: if size < 0: return self.decode(self.stream.read())[0] # Sliced reading: read = self.stream.read decode = self.decode data = read(size) i = 0 while 1: try: object, decodedbytes = decode(data) except ValueError,why: # This method is slow but should work under pretty much # all conditions; at most 10 tries are made i = i + 1 newdata = read(1) if not newdata or i > 10: raise data = data + newdata else: return object def reset(self): """ Resets the codec buffers used for keeping state. Note that no stream repositioning should take place. This method is primarely intended to be able to recover from decoding errors. """ pass def __getattr__(self,name, getattr=getattr): """ Inherit all other methods from the underlying stream. """ return getattr(self.stream,name) XXX What about .readline(), .readlines() ? These could be implemented using .read() as generic functions instead of requiring their implementation by all codecs. Also see Line Breaks. Stream codec implementors are free to combine the StreamWriter and StreamReader interfaces into one class. Even combining all these with the Codec class should be possible. Implementors are free to add additional methods to enhance the codec functionality or provide extra state information needed for them to work. The internal codec implementation will only use the above interfaces, though. It is not required by the Unicode implementation to use these base classes, only the interfaces must match; this allows writing Codecs as extensions types. As guideline, large mapping tables should be implemented using static C data in separate (shared) extension modules. That way multiple processes can share the same data. A tool to auto-convert Unicode mapping files to mapping modules should be provided to simplify support for additional mappings (see References). Whitespace: ----------- The .split() method will have to know about what is considered whitespace in Unicode. Case Conversion: ---------------- Case conversion is rather complicated with Unicode data, since there are many different conditions to respect. See http://www.unicode.org/unicode/reports/tr13/ for some guidelines on implementing case conversion. For Python, we should only implement the 1-1 conversions included in Unicode. Locale dependent and other special case conversions (see the Unicode standard file SpecialCasing.txt) should be left to user land routines and not go into the core interpreter. The methods .capitalize() and .iscapitalized() should follow the case mapping algorithm defined in the above technical report as closely as possible. Line Breaks: ------------ Line breaking should be done for all Unicode characters having the B property as well as the combinations CRLF, CR, LF (interpreted in that order) and other special line separators defined by the standard. The Unicode type should provide a .splitlines() method which returns a list of lines according to the above specification. See Unicode Methods. Unicode Character Properties: ----------------------------- A separate module "unicodedata" should provide a compact interface to all Unicode character properties defined in the standard's UnicodeData.txt file. Among other things, these properties provide ways to recognize numbers, digits, spaces, whitespace, etc. Since this module will have to provide access to all Unicode characters, it will eventually have to contain the data from UnicodeData.txt which takes up around 600kB. For this reason, the data should be stored in static C data. This enables compilation as shared module which the underlying OS can shared between processes (unlike normal Python code modules). There should be a standard Python interface for accessing this information so that other implementors can plug in their own possibly enhanced versions, e.g. ones that do decompressing of the data on-the-fly. Private Code Point Areas: ------------------------- Support for these is left to user land Codecs and not explicitly intergrated into the core. Note that due to the Internal Format being implemented, only the area between \uE000 and \uF8FF is useable for private encodings. Internal Format: ---------------- The internal format for Unicode objects should use a Python specific fixed format implemented as 'unsigned short' (or another unsigned numeric type having 16 bits). Byte order is platform dependent. This format will hold UTF-16 encodings of the corresponding Unicode ordinals. The Python Unicode implementation will address these values as if they were UCS-2 values. UCS-2 and UTF-16 are the same for all currently defined Unicode character points. UTF-16 without surrogates provides access to about 64k characters and covers all characters in the Basic Multilingual Plane (BMP) of Unicode. It is the Codec's responsibility to ensure that the data they pass to the Unicode object constructor repects this assumption. The constructor does not check the data for Unicode compliance or use of surrogates. Future implementations can extend the 32 bit restriction to the full set of all UTF-16 addressable characters (around 1M characters). The Unicode API should provide inteface routines from to the compiler's wchar_t which can be 16 or 32 bit depending on the compiler/libc/platform being used. Unicode objects should have a pointer to a cached Python string object holding the object's value using the current . This is needed for performance and internal parsing (see Internal Argument Parsing) reasons. The buffer is filled when the first conversion request to the is issued on the object. Interning is not needed (for now), since Python identifiers are defined as being ASCII only. codecs.BOM should return the byte order mark (BOM) for the format used internally. The codecs module should provide the following additional constants for convenience and reference (codecs.BOM will either be BOM_BE or BOM_LE depending on the platform): BOM_BE: '\376\377' (corresponds to Unicode U+0000FEFF in UTF-16 on big endian platforms == ZERO WIDTH NO-BREAK SPACE) BOM_LE: '\377\376' (corresponds to Unicode U+0000FFFE in UTF-16 on little endian platforms == defined as being an illegal Unicode character) BOM4_BE: '\000\000\376\377' (corresponds to Unicode U+0000FEFF in UCS-4) BOM4_LE: '\377\376\000\000' (corresponds to Unicode U+0000FFFE in UCS-4) Note that Unicode sees big endian byte order as being "correct". The swapped order is taken to be an indicator for a "wrong" format, hence the illegal character definition. The configure script should provide aid in deciding whether Python can use the native wchar_t type or not (it has to be a 16-bit unsigned type). Buffer Interface: ----------------- Implement the buffer interface using the Python string object as basis for bf_getcharbuf (corresponds to the "t#" argument parsing marker) and the internal buffer for bf_getreadbuf (corresponds to the "s#" argument parsing marker). If bf_getcharbuf is requested and the object does not yet exist, it is created first. This has the advantage of being able to write to output streams (which typically use this interface) without additional specification of the encoding to use. The internal format can also be accessed using the 'unicode-internal' codec, e.g. via u.encode('unicode-internal'). Pickle/Marshalling: ------------------- Should have native Unicode object support. The objects should be encoded using platform independent encodings. Marshal should use UTF-8 and Pickle should either choose Raw-Unicode-Escape (in text mode) or UTF-8 (in binary mode) as encoding. Using UTF-8 instead of UTF-16 has the advantage of eliminating the need to store a BOM mark. Regular Expressions: -------------------- Secret Labs AB is working on a Unicode-aware regular expression machinery. It works on plain 8-bit, UCS-2, and (optionally) UCS-4 internal character buffers. Also see http://www.unicode.org/unicode/reports/tr18/ for some remarks on how to treat Unicode REs. Formatting Markers: ------------------- Format markers are used in Python format strings. If Python strings are used as format strings, the following interpretations should be in effect: '%s': '%s' does str(u) for Unicode objects embedded in Python strings, so the output will be u.encode() In case the format string is an Unicode object, all parameters are coerced to Unicode first and then put together and formatted according to the format string. Numbers are first converted to strings and then to Unicode. '%s': Python strings are interpreted as Unicode string using the . Unicode objects are taken as is. All other string formatters should work accordingly. Example: u"%s %s" % (u"abc", "abc") == u"abc abc" Internal Argument Parsing: -------------------------- These markers are used by the PyArg_ParseTuple() APIs: 'U': Check for Unicode object and return a pointer to it 's': For Unicode objects: auto convert them to the and return a pointer to the object's buffer. 's#': Access to the Unicode object via the bf_getreadbuf buffer interface (see Buffer Interface); note that the length relates to the buffer length, not the Unicode string length (this may be different depending on the Internal Format). 't#': Access to the Unicode object via the bf_getcharbuf buffer interface (see Buffer Interface); note that the length relates to the buffer length, not necessarily to the Unicode string length (this may be different depending on the ). File/Stream Output: ------------------- Since file.write(object) and most other stream writers use the "s#" argument parsing marker for binary files and "t#" for text files, the buffer interface implementation determines the encoding to use (see Buffer Interface). For explicit handling of files using Unicode, the standard stream codecs as available through the codecs module should be used. XXX There should be a short-cut open(filename,mode,encoding) available which also assures that mode contains the 'b' character when needed. File/Stream Input: ------------------ Only the user knows what encoding the input data uses, so no special magic is applied. The user will have to explicitly convert the string data to Unicode objects as needed or use the file wrappers defined in the codecs module (see File/Stream Output). Unicode Methods & Attributes: ----------------------------- All Python string methods, plus: .encode([encoding=][,errors="strict"]) --> see Unicode Output .splitlines([include_breaks=0]) --> breaks the Unicode string into a list of (Unicode) lines; returns the lines with line breaks included, if include_breaks is true. See Line Breaks for a specification of how line breaking is done. Code Base: ---------- We should use Fredrik Lundh's Unicode object implementation as basis. It already implements most of the string methods needed and provides a well written code base which we can build upon. The object sharing implemented in Fredrik's implementation should be dropped. Test Cases: ----------- Test cases should follow those in Lib/test/test_string.py and include additional checks for the Codec Registry and the Standard Codecs. References: ----------- Unicode Consortium: http://www.unicode.org/ Unicode FAQ: http://www.unicode.org/unicode/faq/ Unicode 3.0: http://www.unicode.org/unicode/standard/versions/Unicode3.0.html Unicode-TechReports: http://www.unicode.org/unicode/reports/techreports.html Unicode-Mappings: ftp://ftp.unicode.org/Public/MAPPINGS/ Introduction to Unicode (a little outdated by still nice to read): http://www.nada.kth.se/i18n/ucs/unicode-iso10646-oview.html Encodings: Overview: http://czyborra.com/utf/ UTC-2: http://www.uazone.com/multiling/unicode/ucs2.html UTF-7: Defined in RFC2152, e.g. http://www.uazone.com/multiling/ml-docs/rfc2152.txt UTF-8: Defined in RFC2279, e.g. http://info.internet.isi.edu/in-notes/rfc/files/rfc2279.txt UTF-16: http://www.uazone.com/multiling/unicode/wg2n1035.html History of this Proposal: ------------------------- 1.2: 1.1: Added note about comparisons and hash values. Added note about case mapping algorithms. Changed stream codecs .read() and .write() method to match the standard file-like object methods (bytes consumed information is no longer returned by the methods) 1.0: changed encode Codec method to be symmetric to the decode method (they both return (object, data consumed) now and thus become interchangeable); removed __init__ method of Codec class (the methods are stateless) and moved the errors argument down to the methods; made the Codec design more generic w/r to type of input and output objects; changed StreamWriter.flush to StreamWriter.reset in order to avoid overriding the stream's .flush() method; renamed .breaklines() to .splitlines(); renamed the module unicodec to codecs; modified the File I/O section to refer to the stream codecs. 0.9: changed errors keyword argument definition; added 'replace' error handling; changed the codec APIs to accept buffer like objects on input; some minor typo fixes; added Whitespace section and included references for Unicode characters that have the whitespace and the line break characteristic; added note that search functions can expect lower-case encoding names; dropped slicing and offsets in the codec APIs 0.8: added encodings package and raw unicode escape encoding; untabified the proposal; added notes on Unicode format strings; added .breaklines() method 0.7: added a whole new set of codec APIs; added a different encoder lookup scheme; fixed some names 0.6: changed "s#" to "t#"; changed to holding a real Python string object; changed Buffer Interface to delegate requests to 's buffer interface; removed the explicit reference to the unicodec.codecs dictionary (the module can implement this in way fit for the purpose); removed the settable default encoding; move UnicodeError from unicodec to exceptions; "s#" not returns the internal data; passed the UCS-2/UTF-16 checking from the Unicode constructor to the Codecs 0.5: moved sys.bom to unicodec.BOM; added sections on case mapping, private use encodings and Unicode character properties 0.4: added Codec interface, notes on %-formatting, changed some encoding details, added comments on stream wrappers, fixed some discussion points (most important: Internal Format), clarified the 'unicode-escape' encoding, added encoding references 0.3: added references, comments on codec modules, the internal format, bf_getcharbuffer and the RE engine; added 'unicode-escape' encoding proposed by Tim Peters and fixed repr(u) accordingly 0.2: integrated Guido's suggestions, added stream codecs and file wrapping 0.1: first version ----------------------------------------------------------------------------- Written by Marc-Andre Lemburg, 1999-2000, mal@lemburg.com -----------------------------------------------------------------------------