1. Introduction

This library provides functions for manipulating Unicode strings and for manipulating C strings according to the Unicode standard.

It consists of the following parts:

<unistr.h>

elementary string functions

<uniconv.h>

conversion from/to legacy encodings

<unistdio.h>

formatted output to strings

<uniname.h>

character names

<unictype.h>

character classification and properties

<uniwidth.h>

string width when using nonproportional fonts

<uniwbrk.h>

word breaks

<unilbrk.h>

line breaking algorithm

<uninorm.h>

normalization (composition and decomposition)

<unicase.h>

case folding

<uniregex.h>

regular expressions (not yet implemented)

libunistring is for you if your application involves non-trivial text processing, such as upper/lower case conversions, line breaking, operations on words, or more advanced analysis of text. Text provided by the user can, in general, contain characters of all kinds of scripts. The text processing functions provided by this library handle all scripts and all languages.

libunistring is for you if your application already uses the ISO C / POSIX <ctype.h>, <wctype.h> functions and the text it operates on is provided by the user and can be in any language.

libunistring is also for you if your application uses Unicode strings as internal in-memory representation.

1.1 Unicode

Unicode is a standardized repertoire of characters that contains characters from all scripts of the world, from Latin letters to Chinese ideographs and Babylonian cuneiform glyphs. It also specifies how these characters are to be rendered on a screen or on paper, and how common text processing (word selection, line breaking, uppercasing of page titles etc.) is supposed to behave on Unicode text.

Unicode also specifies three ways of storing sequences of Unicode characters in a computer whose basic unit of data is an 8-bit byte:

UTF-8

Every character is represented as 1 to 4 bytes.

UTF-16

Every character is represented as 1 to 2 units of 16 bits.

UTF-32, a.k.a. UCS-4

Every character is represented as 1 unit of 32 bits.

For encoding Unicode text in a file, UTF-8 is usually used. For encoding Unicode strings in memory for a program, either of the three encoding forms can be reasonably used.

Unicode is widely used on the web. Prior to the use of Unicode, web pages were in many different encodings (ISO-8859-1 for English, French, Spanish, ISO-8859-2 for Polish, ISO-8859-7 for Greek, KOI8-R for Russian, GB2312 or BIG5 for Chinese, ISO-2022-JP-2 or EUC-JP or Shift_JIS for Japanese, and many many others). It was next to impossible to create a document that contained Chinese and Polish text in the same document. Due to the many encodings for Japanese, even the processing of pure Japanese text was error prone.

References:

• The Unicode standard: http://www.unicode.org/
• Definition of UTF-8: http://www.rfc-editor.org/rfc/rfc3629.txt
• Definition of UTF-16: http://www.rfc-editor.org/rfc/rfc2781.txt
• Markus Kuhn's UTF-8 and Unicode FAQ: http://www.cl.cam.ac.uk/~mgk25/unicode.html

1.2 Unicode and Internationalization

Internationalization is the process of changing the source code of a program so that it can meet the expectations of users in any culture, if culture specific data (translations, images etc.) are provided.

Use of Unicode is not strictly required for internationalization, but it makes internationalization much easier, because operations that need to look at specific characters (like hyphenation, spell checking, or the automatic conversion of double-quotes to opening and closing double-quote characters) don't need to consider multiple possible encodings of the text.

Use of Unicode also enables multilingualization: the ability of having text in multiple languages present in the same document or even in the same line of text.

But use of Unicode is not everything. Internationalization usually consists of three features:

• Use of Unicode where needed for text processing. This is what this library is for.
• Use of message catalogs for messages shown to the user, This is what GNU gettext is about.
• Use of locale specific conventions for date and time formats, for numeric formatting, or for sorting of text. This can be done adequately with the POSIX APIs and the implementation of locales in the GNU C library.

1.3 Locale encodings

A locale is a set of cultural conventions. According to POSIX, for a program, at any moment, there is one locale being designated as the “current locale”. (Actually, POSIX supports also one locale per thread, but this feature is not yet universally implemented and not widely used.) The locale is partitioned into several aspects, called the “categories” of the locale. The main various aspects are:

• The character encoding and the character properties. This is the LC_CTYPE category.
• The sorting rules for text. This is the LC_COLLATE category.
• The language specific translations of messages. This is the LC_MESSAGES category.
• The formatting rules for numbers, such as the decimal separator. This is the LC_NUMERIC category.
• The formatting rules for amounts of money. This is the LC_MONETARY category.
• The formatting of date and time. This is the LC_TIME category.

In particular, the LC_CTYPE category of the current locale determines the character encoding. This is the encoding of ‘char *’ strings. We also call it the “locale encoding”. GNU libunistring has a function, locale_charset, that returns a standardized (platform independent) name for this encoding.

All locale encodings used on glibc systems are essentially ASCII compatible: Most graphic ASCII characters have the same representation, as a single byte, in that encoding as in ASCII.

Among the possible locale encodings are UTF-8 and GB18030. Both allow to represent any Unicode character as a sequence of bytes. UTF-8 is used in most of the world, whereas GB18030 is used in the People's Republic of China, because it is backward compatible with the GB2312 encoding that was used in this country earlier.

The legacy locale encodings, ISO-8859-15 (which supplanted ISO-8859-1 in most of Europe), ISO-8859-2, KOI8-R, EUC-JP, etc., are still in use in many places, though.

UTF-16 and UTF-32 are not used as locale encodings, because they are not ASCII compatible.

1.4 Choice of in-memory representation of strings

There are three ways of representing strings in memory of a running program.

• As ‘char *’ strings. Such strings are represented in locale encoding. This approach is employed when not much text processing is done by the program. When some Unicode aware processing is to be done, a string is converted to Unicode on the fly and back to locale encoding afterwards.
• As UTF-8 or UTF-16 or UTF-32 strings. This implies that conversion from locale encoding to Unicode is performed on input, and in the opposite direction on output. This approach is employed when the program does a significant amount of text processing, or when the program has multiple threads operating on the same data but in different locales.
• As ‘wchar_t *’, a.k.a. “wide strings”. This approach is misguided, see The wchar_t mess.

1.5 ‘char *’ strings

The classical C strings, with its C library support standardized by ISO C and POSIX, can be used in internationalized programs with some precautions. The problem with this API is that many of the C library functions for strings don't work correctly on strings in locale encodings, leading to bugs that only people in some cultures of the world will experience.

The first problem with the C library API is the support of multibyte locales. According to the locale encoding, in general, every character is represented by one or more bytes (up to 4 bytes in practice — but use MB_LEN_MAX instead of the number 4 in the code). When every character is represented by only 1 byte, we speak of an “unibyte locale”, otherwise of a “multibyte locale”. It is important to realize that the majority of Unix installations nowadays use UTF-8 or GB18030 as locale encoding; therefore, the majority of users are using multibyte locales.

The important fact to remember is:

As a consequence:

• The <ctype.h> API is useless in this context; it does not work in multibyte locales.
• The strlen function does not return the number of characters in a string. Nor does it return the number of screen columns occupied by a string after it is output. It merely returns the number of bytes occupied by a string.
• Truncating a string, for example, with strncpy, can have the effect of truncating it in the middle of a multibyte character. Such a string will, when output, have a garbled character at its end, often represented by a hollow box.
• strchr and strrchr do not work with multibyte strings if the locale encoding is GB18030 and the character to be searched is a digit.
• strstr does not work with multibyte strings if the locale encoding is different from UTF-8.
• strcspn, strpbrk, strspn cannot work correctly in multibyte locales: they assume the second argument is a list of single-byte characters. Even in this simple case, they do not work with multibyte strings if the locale encoding is GB18030 and one of the characters to be searched is a digit.
• strsep and strtok_r do not work with multibyte strings unless all of the delimiter characters are ASCII characters < 0x30.
• The strcasecmp, strncasecmp, and strcasestr functions do not work with multibyte strings.

The workarounds can be found in GNU gnulib http://www.gnu.org/software/gnulib/.

• gnulib has modules ‘mbchar’, ‘mbiter’, ‘mbuiter’ that represent multibyte characters and allow to iterate across a multibyte string with the same ease as through a unibyte string.
• gnulib has functions mbslen and mbswidth that can be used instead of strlen when the number of characters or the number of screen columns of a string is requested.
• gnulib has functions mbschr and mbsrrchr that are like strchr and strrchr, but work in multibyte locales.
• gnulib has a function mbsstr, like strstr, but works in multibyte locales.
• gnulib has functions mbscspn, mbspbrk, mbsspn that are like strcspn, strpbrk, strspn, but work in multibyte locales.
• gnulib has functions mbssep and mbstok_r that are like strsep and strtok_r but work in multibyte locales.
• gnulib has functions mbscasecmp, mbsncasecmp, mbspcasecmp, and mbscasestr that are like strcasecmp, strncasecmp, and strcasestr, but work in multibyte locales. Still, the function ulc_casecmp is preferable to these functions; see below.

The second problem with the C library API is that it has some assumptions built-in that are not valid in some languages:

• It assumes that there are only two forms of every character: uppercase and lowercase. This is not true for Croatian, where the character LETTER DZ WITH CARON comes in three forms: LATIN CAPITAL LETTER DZ WITH CARON (DZ), LATIN CAPITAL LETTER D WITH SMALL LETTER Z WITH CARON (Dz), LATIN SMALL LETTER DZ WITH CARON (dz).
• It assumes that uppercasing of 1 character leads to 1 character. This is not true for German, where the LATIN SMALL LETTER SHARP S, when uppercased, becomes ‘SS’.
• It assumes that there is 1:1 mapping between uppercase and lowercase forms. This is not true for the Greek sigma: GREEK CAPITAL LETTER SIGMA is the uppercase of both GREEK SMALL LETTER SIGMA and GREEK SMALL LETTER FINAL SIGMA.
• It assumes that the upper/lowercase mappings are position independent. This is not true for the Greek sigma and the Lithuanian i.

The correct way to deal with this problem is

1. to provide functions for titlecasing, as well as for upper- and lowercasing,
2. to view case transformations as functions that operates on strings, rather than on characters.

This is implemented in this library, through the functions declared in <unicase.h>, see Case mappings <unicase.h>.

1.6 The wchar_t mess

The ISO C and POSIX standard creators made an attempt to fix the first problem mentioned in the previous section. They introduced

• a type ‘wchar_t’, designed to encapsulate an entire character,
• a “wide string” type ‘wchar_t *’, and
• functions declared in <wctype.h> that were meant to supplant the ones in <ctype.h>.

Unfortunately, this API and its implementation has numerous problems:

• On AIX and Windows platforms, wchar_t is a 16-bit type. This means that it can never accommodate an entire Unicode character. Either the wchar_t * strings are limited to characters in UCS-2 (the “Basic Multilingual Plane” of Unicode), or — if wchar_t * strings are encoded in UTF-16 — a wchar_t represents only half of a character in the worst case, making the <wctype.h> functions pointless.
• On Solaris and FreeBSD, the wchar_t encoding is locale dependent and undocumented. This means, if you want to know any property of a wchar_t character, other than the properties defined by <wctype.h> — such as whether it's a dash, currency symbol, paragraph separator, or similar —, you have to convert it to char * encoding first, by use of the function wctomb.
• When you read a stream of wide characters, through the functions fgetwc and fgetws, and when the input stream/file is not in the expected encoding, you have no way to determine the invalid byte sequence and do some corrective action. If you use these functions, your program becomes “garbage in - more garbage out” or “garbage in - abort”.

As a consequence, it is better to use multibyte strings, as explained in the previous section. Such multibyte strings can bypass limitations of the wchar_t type, if you use functions defined in gnulib and libunistring for text processing. They can also faithfully transport malformed characters that were present in the input, without requiring the program to produce garbage or abort.

1.7 Unicode strings

libunistring supports Unicode strings in three representations:

• UTF-8 strings, through the type ‘uint8_t *’. The units are bytes (uint8_t).
• UTF-16 strings, through the type ‘uint16_t *’, The units are 16-bit memory words (uint16_t).
• UTF-32 strings, through the type ‘uint32_t *’. The units are 32-bit memory words (uint32_t).

As with C strings, there are two variants:

• Unicode strings with a terminating NUL character are represented as a pointer to the first unit of the string. There is a unit containing a 0 value at the end. It is considered part of the string for all memory allocation purposes, but is not considered part of the string for all other logical purposes.
• Unicode strings where embedded NUL characters are allowed. These are represented by a pointer to the first unit and the number of units (not bytes!) of the string. In this setting, there is no trailing zero-valued unit used as “end marker”.

This document was generated by Bruno Haible on August, 17 2009 using texi2html 1.78a.