Libxml Internationalization support

Location: http://xmlsoft.org/encoding.html

Mailing-list archive: http://xmlsoft.org/messages/

Version: $Revision$

Table of Content:

What does internationalization support mean ?
The internal encoding, how and why
How is it implemented ?
Default supported encodings
How to extend the existing support

What does internationalization support mean ?

XML was designed from the start to allow the support of any character set by using Unicode. Any conformant XML parser has to support the UTF-8 and UTF-16 default encodings which can both express the full unicode ranges. UTF8 is a variable length encoding whose greatest point are to resuse the same emcoding for ASCII and to save space for Western encodings, but it is a bit more complex to handle in practice. UTF-16 use 2 bytes per characters (and sometimes combines two pairs), it makes implementation easier, but looks a bit overkill for Western languages encoding. Moreover the XML specification allows document to be encoded in other encodings at the condition that they are clearly labelled as such. For example the following is a wellformed XML document encoded in ISO-8859 1 and using accentuated letter that we French likes for both markup and content:

<?xml version="1.0" encoding="ISO-8859-1"?>
<très>là</très>

Having internationalization support in libxml means the foolowing:

the document is properly parsed
informations about it's encoding are saved
it can be modified
it can be saved in its original encoding
it can also be saved in another encoding supported by libxml (for example straight UTF8 or even an ASCII form)

Another very important point is that the whole libxml API, with the exception of a few routines to read with a specific encoding or save to a specific encoding, is completely agnostic about the original encoding of the document.

It should be noted too that the HTML parser embedded in libxml now obbey the same rules too, the following document will be (as of 2.2.2) handled in an internationalized fashion by libxml too:

<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN"
                      "http://www.w3.org/TR/REC-html40/loose.dtd">
<html lang="fr">
<head>
  <META HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=ISO-8859-1">
</head>
<body>
<p>W3C crée des standards pour le Web.</body>
</html>

The internal encoding, how and why

One of the core decision was to force all documents to be converted to a default internal encoding, and that encoding to be UTF-8, here are the rationale for those choices:

keeping the native encoding in the internal form would force the libxml users (or the code associated) to be fully aware of the encoding of the original document, for examples when adding a text node to a document, the content would have to be provided in the document encoding, i.e. the client code would have to check it before hand, make sure it's conformant to the encoding, etc ... Very hard in practice, though in some specific cases this may make sense.
the second decision was which encoding. From the XML spec only UTF8 and UTF16 really makes sense as being the two only encodings for which there is amndatory support. UCS-4 (32 bits fixed size encoding) could be considered an intelligent choice too since it's a direct Unicode mapping support. I selected UTF-8 on the basis of efficiency and compatibility with surrounding software:
- UTF-8 while a bit more complex to convert from/to (i.e. slightly more costly to import and export CPU wise) is also far more compact than UTF-16 (and UCS-4) for a majority of the documents I see it used for right now (RPM RDF catalogs, advogato data, various configuration file formats, etc.) and the key point for today's computer architecture is efficient uses of caches. If one nearly double the memory requirement to store the same amount of data, this will trash caches (main memory/external caches/internal caches) and my take is that this harms the system far more than the CPU requirements needed for the conversion to UTF-8
- Most of libxml version 1 users were using it with straight ASCII most of the time, doing the conversion with an internal encoding requiring all their code to be rewritten was a serious show-stopper for using UTF-16 or UCS-4.
- UTF-8 is being used as the de-facto internal encoding standard for related code like the pango upcoming Gnome text widget, and a lot of Unix code (yep another place where Unix programmer base takes a different approach from Microsoft - they are using UTF-16)

What does this mean in practice for the libxml user:

xmlChar, the libxml data type is a byte, those bytes must be assembled as UTF-8 valid strings. The proper way to terminate an xmlChar * string is simply to append 0 byte, as usual.
One just need to make sure that when using chars outside the ASCII set, the values has been properly converted to UTF-8

How is it implemented ?

Let's describe how all this works within libxml, basically the I18N (internationalization) support get triggered only during I/O operation, i.e. when reading a document or saving one. Let's look first at the reading sequence:

when a document is processed, we usually don't know the encoding, a simple heuristic allows to detect UTF-18 and UCS-4 from whose where the ASCII range (0-0x7F) maps with ASCII
the xml declaration if available is parsed, including the encoding declaration. At that point, if the autodetected encoding is different from the one declared a call to xmlSwitchEncoding() is issued.
If there is no encoding declaration, then the input has to be in either UTF-8 or UTF-16, if it is not then at some point when processing the input, the converter/checker of UTF-8 form will raise an encoding error. You may end-up with a garbled document, or no document at all ! Example:
```
~/XML -> ./xmllint err.xml 
err.xml:1: error: Input is not proper UTF-8, indicate encoding !
<très>là</très>
   ^
err.xml:1: error: Bytes: 0xE8 0x73 0x3E 0x6C
<très>là</très>
   ^
```
xmlSwitchEncoding() does an encoding name lookup, canonalize it, and then search the default registered encoding converters for that encoding. If it's not within the default set and iconv() support has been compiled it, it will ask iconv for such an encoder. If this fails then the parser will report an error and stops processing:
```
~/XML -> ./xmllint err2.xml 
err2.xml:1: error: Unsupported encoding UnsupportedEnc
<?xml version="1.0" encoding="UnsupportedEnc"?>
                                             ^
```
From that point the encoder process progressingly the input (it is plugged as a front-end to the I/O module) for that entity. It captures and convert on-the-fly the document to be parsed to UTF-8. The parser itself just does UTF-8 checking of this input and process it transparently. The only difference is that the encoding information has been added to the parsing context (more precisely to the input corresponding to this entity).
The result (when using DOM) is an internal form completely in UTF-8 with just an encoding information on the document node.

Ok then what's happen when saving the document (assuming you colllected/built an xmlDoc DOM like structure) ? It depends on the function called, xmlSaveFile() will just try to save in the original encoding, while xmlSaveFileTo() and xmlSaveFileEnc() can optionally save to a given encoding:

if no encoding is given, libxml will look for an encoding value associated to the document and if it exists will try to save to that encoding,
otherwise everything is written in the internal form, i.e. UTF-8
so if an encoding was specified, either at the API level or on the document, libxml will again canonalize the encoding name, lookup for a converter in the registered set or through iconv. If not found the function will return an error code
the converter is placed before the I/O buffer layer, as another kind of buffer, then libxml will simply push the UTF-8 serialization to through that buffer, which will then progressively be converted and pushed onto the I/O layer.
It is possible that the converter code fails on some input, for example trying to push an UTF-8 encoded chinese character through the UTF-8 to ISO-8859-1 converter won't work. Since the encoders are progressive they will just report the error and the number of bytes converted, at that point libxml will decode the offending character, remove it from the buffer and replace it with the associated charRef encoding { and resume the convertion. This guarante that any document will be saved without losses (except for markup names where this is not legal, this is a problem in the current version, in pactice avoid using non-ascci characters for tags or attributes names @@). A special "ascii" encoding name is used to save documents to a pure ascii form can be used when portability is really crucial

Here is a few examples based on the same test document:

~/XML -> ./xmllint isolat1 
<?xml version="1.0" encoding="ISO-8859-1"?>
<très>là</très>
~/XML -> ./xmllint --encode UTF-8 isolat1 
<?xml version="1.0" encoding="UTF-8"?>
<trÃ¨s>lÃ  </trÃ¨s>
~/XML ->

The same processing is applied (and reuse most of the code) for HTML I18N processing. Looking up and modifying the content encoding is a bit more difficult since it is located in a <meta> tag under the <head>, so a couple of functions htmlGetMetaEncoding() and htmlSetMetaEncoding() have been provided. The parser also attempts to switch encoding on the fly when detecting such a tag on input. Except for that the processing is the same (and again reuses the same code).

Default supported encodings

libxml has a set of default converters for the following encodings (located in encoding.c):

UTF-8 is supported by default (null handlers)
UTF-16, both little and big endian
ISO-Latin-1 (ISO-8859-1) covering most western languages
ASCII, useful mostly for saving
HTML, a specific handler for the conversion of UTF-8 to ASCII with HTML predefined entities like © for the Copyright sign.

More over when compiled on an Unix platfor with iconv support the full set of encodings supported by iconv can be instantly be used by libxml. On a linux machine with glibc-2.1 the list of supported encodings and aliases fill 3 full pages, and include UCS-4, the full set of ISO-Latin encodings, and the various Japanese ones.

Encoding aliases

From 2.2.3, libxml has support to register encoding names aliases. The goal is to be able to parse document whose encoding is supported but where the name differs (for example from the default set of names accepted by iconv). The following functions allow to register and handle new aliases for existing encodings. Once registered libxml will automatically lookup the aliases when handling a document:

int xmlAddEncodingAlias(const char *name, const char *alias);
int xmlDelEncodingAlias(const char *alias);
const char * xmlGetEncodingAlias(const char *alias);
void xmlCleanupEncodingAliases(void);

How to extend the existing support

Well adding support for new encoding, or overriding one of the encoders (assuming it is buggy) should not be hard, just write an input and output conversion routines to/from UTF-8, and register them using xmlNewCharEncodingHandler(name, xxxToUTF8, UTF8Toxxx), and they will be called automatically if the parser(s) encounter such an encoding name (register it uppercase, this will help). The description of the encoders, their arguments and expected return values are described in the encoding.h header.

A quick note on the topic of subverting the parser to use a different internal encoding than UTF-8, in some case people will absolutely want to keep the internal encoding different, I think it's still possible (but the encoding must be compliant with ASCII on the same subrange) though I didn't tried it. The key is to override the default conversion routines (by registering null encoders/decoders for your charsets), and bypass the UTF-8 checking of the parser by setting the parser context charset (ctxt->charset) to something different than XML_CHAR_ENCODING_UTF8, but there is no guarantee taht this will work. You may also have some troubles saving back.

Basically proper I18N support is important, this requires at least libxml-2.0.0, but a lot of features and corrections are really available only starting 2.2.

Daniel Veillard

$Id$