UTF-8 and Unicode FAQ for Unix/Linux

by Markus Kuhn

This text is a very comprehensive one-stop information resource on how you can use Unicode/UTF-8 on POSIX systems (Linux, Unix). You will find here both introductory information for every user, as well as detailed references for the experienced developer.

Unicode has started to replace ASCII, ISO 8859 and EUC at all levels. It enables users to handle not only practically any script and language used on this planet, it also supports a comprehensive set of mathematical and technical symbols to simplify scientific information exchange.

With the UTF-8 encoding, Unicode can be used in a convenient and backwards compatible way in environments that were designed entirely around ASCII, like Unix. UTF-8 is the way in which Unicode is used under Unix, Linux, and similar systems. It is now time to make sure that you are well familiar with it and that your software supports UTF-8 smoothly.

Contents

What are UCS and ISO 10646?

The international standard ISO 10646 defines the Universal Character Set (UCS). UCS is a superset of all other character set standards. It guarantees round-trip compatibility to other character sets. This means simply that no information is lost if you convert any text string to UCS and then back to its original encoding.

UCS contains the characters required to represent practically all known languages. This includes not only the Latin, Greek, Cyrillic, Hebrew, Arabic, Armenian, and Georgian scripts, but also Chinese, Japanese and Korean Han ideographs as well as scripts such as Hiragana, Katakana, Hangul, Devanagari, Bengali, Gurmukhi, Gujarati, Oriya, Tamil, Telugu, Kannada, Malayalam, Thai, Lao, Khmer, Bopomofo, Tibetan, Runic, Ethiopic, Canadian Syllabics, Cherokee, Mongolian, Ogham, Myanmar, Sinhala, Thaana, Yi, and others. For scripts not yet covered, research on how to best encode them for computer usage is still going on and they will be added eventually. This includes not only historic scripts such as Cuneiform, Hieroglyphs and various Indo-European notations, but even some selected artistic scripts such as Tolkien’s Tengwar and Cirth. UCS also covers a large number of graphical, typographical, mathematical and scientific symbols, including those provided by TeX, PostScript, APL, the International Phonetic Alphabet (IPA), MS-DOS, MS-Windows, Macintosh, OCR fonts, as well as many word processing and publishing systems. The standard continues to be maintained and updated. Ever more exotic and specialized symbols and characters will be added for many years to come.

ISO 10646 originally defined a 31-bit character set. The subsets of 216 characters where the elements differ (in a 32-bit integer representation) only in the 16 least-significant bits are called the planes of UCS.

The most commonly used characters, including all those found in major older encoding standards, have been placed into the first plane (0x0000 to 0xFFFD), which is called the Basic Multilingual Plane (BMP) or Plane 0. The characters that were later added outside the 16-bit BMP are mostly for specialist applications such as historic scripts and scientific notation. Current plans are that there will never be characters assigned outside the 21-bit code space from 0x000000 to 0x10FFFF, which covers a bit over one million potential future characters. The ISO 10646-1 standard was first published in 1993 and defines the architecture of the character set and the content of the BMP. A second part ISO 10646-2 was added in 2001 and defines characters encoded outside the BMP. In the 2003 edition, the two parts were combined into a single ISO 10646 standard. New characters are still being added on a continuous basis, but the existing characters will not be changed any more and are stable.

UCS assigns to each character not only a code number but also an official name. A hexadecimal number that represents a UCS or Unicode value is commonly preceded by “U+” as in U+0041 for the character “Latin capital letter A”. The UCS characters U+0000 to U+007F are identical to those in US-ASCII (ISO 646 IRV) and the range U+0000 to U+00FF is identical to ISO 8859-1 (Latin-1). The range U+E000 to U+F8FF and also larger ranges outside the BMP are reserved for private use. UCS also defines several methods for encoding a string of characters as a sequence of bytes, such as UTF-8 and UTF-16.

The full reference for the UCS standard is

International Standard ISO/IEC 10646, Information technology — Universal Multiple-Octet Coded Character Set (UCS) . Third edition, International Organization for Standardization, Geneva, 2003.

The standard can be ordered online from ISO as a set of PDF files on CD-ROM for 112 CHF.

What are combining characters?

Some code points in UCS have been assigned to combining characters. These are similar to the non-spacing accent keys on a typewriter. A combining character is not a full character by itself. It is an accent or other diacritical mark that is added to the previous character. This way, it is possible to place any accent on any character. The most important accented characters, like those used in the orthographies of common languages, have codes of their own in UCS to ensure backwards compatibility with older character sets. They are known as precomposed characters. Precomposed characters are available in UCS for backwards compatibility with older encodings that have no combining characters, such as ISO 8859. The combining-character mechanism allows one to add accents and other diacritical marks to any character. This is especially important for scientific notations such as mathematical formulae and the International Phonetic Alphabet, where any possible combination of a base character and one or several diacritical marks could be needed.

Combining characters follow the character which they modify. For example, the German umlaut character Ä (“Latin capital letter A with diaeresis”) can either be represented by the precomposed UCS code U+00C4, or alternatively by the combination of a normal “Latin capital letter A” followed by a “combining diaeresis”: U+0041 U+0308. Several combining characters can be applied when it is necessary to stack multiple accents or add combining marks both above and below the base character. The Thai script, for example, needs up to two combining characters on a single base character.

What are UCS implementation levels?

Not all systems can be expected to support all the advanced mechanisms of UCS, such as combining characters. Therefore, ISO 10646 specifies the following three implementation levels:

Level 1
Combining characters and Hangul Jamo characters are not supported.
[Hangul Jamo are an alternative representation of precomposed modern Hangul syllables as a sequence of consonants and vowels. They are required to fully support the Korean script including Middle Korean.]
Level 2
Like level 1, however in some scripts, a fixed list of combining characters is now allowed (e.g., for Hebrew, Arabic, Devanagari, Bengali, Gurmukhi, Gujarati, Oriya, Tamil, Telugo, Kannada, Malayalam, Thai and Lao). These scripts cannot be represented adequately in UCS without support for at least certain combining characters.
Level 3
All UCS characters are supported, such that, for example, mathematicians can place a tilde or an arrow (or both) on any character.

Has UCS been adopted as a national standard?

Yes, a number of countries have published national adoptions of ISO 10646, sometimes after adding additional annexes with cross-references to older national standards, implementation guidelines, and specifications of various national implementation subsets:

What is Unicode?

In the late 1980s, there have been two independent attempts to create a single unified character set. One was the ISO 10646 project of the International Organization for Standardization (ISO), the other was the Unicode Project organized by a consortium of (initially mostly US) manufacturers of multi-lingual software. Fortunately, the participants of both projects realized in around 1991 that two different unified character sets is not exactly what the world needs. They joined their efforts and worked together on creating a single code table. Both projects still exist and publish their respective standards independently, however the Unicode Consortium and ISO/IEC JTC1/SC2 have agreed to keep the code tables of the Unicode and ISO 10646 standards compatible and they closely coordinate any further extensions. Unicode 1.1 corresponded to ISO 10646-1:1993, Unicode 3.0 corresponded to ISO 10646-1:2000, Unicode 3.2 added ISO 10646-2:2001, and Unicode 4.0 corresponds to ISO 10646:2003. All Unicode versions since 2.0 are compatible, only new characters will be added, no existing characters will be removed or renamed in the future.

The Unicode Standard can be ordered like any normal book, for instance via amazon.com for around 75 USD:

The Unicode Consortium: The Unicode Standard, Version 4.0,
Addison-Wesley, 2003,
ISBN 0-321-18578-1.

If you work frequently with text processing and character sets, you definitely should get a copy. Unicode 4.0 is also available online.

So what is the difference between Unicode and ISO 10646?

The Unicode Standard published by the Unicode Consortium corresponds to ISO 10646 at implementation level 3. All characters are at the same positions and have the same names in both standards.

The Unicode Standard defines in addition much more semantics associated with some of the characters and is in general a better reference for implementors of high-quality typographic publishing systems. Unicode specifies algorithms for rendering presentation forms of some scripts (say Arabic), handling of bi-directional texts that mix for instance Latin and Hebrew, algorithms for sorting and string comparison, and much more.

The ISO 10646 standard on the other hand is not much more than a simple character set table, comparable to the old ISO 8859 standards. It specifies some terminology related to the standard, defines some encoding alternatives, and it contains specifications of how to use UCS in connection with other established ISO standards such as ISO 6429 and ISO 2022. There are other closely related ISO standards, for instance ISO 14651 on sorting UCS strings. A nice feature of the ISO 10646-1 standard is that it provides CJK example glyphs in five different style variants, while the Unicode standard shows the CJK ideographs only in a Chinese variant.

What is UTF-8?

UCS and Unicode are first of all just code tables that assign integer numbers to characters. There exist several alternatives for how a sequence of such characters or their respective integer values can be represented as a sequence of bytes. The two most obvious encodings store Unicode text as sequences of either 2 or 4 bytes sequences. The official terms for these encodings are UCS-2 and UCS-4, respectively. Unless otherwise specified, the most significant byte comes first in these (Bigendian convention). An ASCII or Latin-1 file can be transformed into a UCS-2 file by simply inserting a 0x00 byte in front of every ASCII byte. If we want to have a UCS-4 file, we have to insert three 0x00 bytes instead before every ASCII byte.

Using UCS-2 (or UCS-4) under Unix would lead to very severe problems. Strings with these encodings can contain as parts of many wide characters bytes like “\0” or “/” which have a special meaning in filenames and other C library function parameters. In addition, the majority of UNIX tools expects ASCII files and cannot read 16-bit words as characters without major modifications. For these reasons, UCS-2 is not a suitable external encoding of Unicode in filenames, text files, environment variables, etc.

The UTF-8 encoding defined in ISO 10646-1:2000 Annex D and also described in RFC 3629 as well as section 3.9 of the Unicode 4.0 standard does not have these problems. It is clearly the way to go for using Unicode under Unix-style operating systems.

UTF-8 has the following properties:

The following byte sequences are used to represent a character. The sequence to be used depends on the Unicode number of the character:

U-00000000 – U-0000007F: 0xxxxxxx
U-00000080 – U-000007FF: 110xxxxx 10xxxxxx
U-00000800 – U-0000FFFF: 1110xxxx 10xxxxxx 10xxxxxx
U-00010000 – U-001FFFFF: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
U-00200000 – U-03FFFFFF: 111110xx 10xxxxxx 10xxxxxx 10xxxxxx 10xxxxxx
U-04000000 – U-7FFFFFFF: 1111110x 10xxxxxx 10xxxxxx 10xxxxxx 10xxxxxx 10xxxxxx

The xxx bit positions are filled with the bits of the character code number in binary representation. The rightmost x bit is the least-significant bit. Only the shortest possible multibyte sequence which can represent the code number of the character can be used. Note that in multibyte sequences, the number of leading 1 bits in the first byte is identical to the number of bytes in the entire sequence.

Examples: The Unicode character U+00A9 = 1010 1001 (copyright sign) is encoded in UTF-8 as

    11000010 10101001 = 0xC2 0xA9

and character U+2260 = 0010 0010 0110 0000 (not equal to) is encoded as:

    11100010 10001001 10100000 = 0xE2 0x89 0xA0

The official name and spelling of this encoding is UTF-8, where UTF stands for UCS Transformation Format. Please do not write UTF-8 in any documentation text in other ways (such as utf8 or UTF_8), unless of course you refer to a variable name and not the encoding itself.

An important note for developers of UTF-8 decoding routines: For security reasons, a UTF-8 decoder must not accept UTF-8 sequences that are longer than necessary to encode a character. For example, the character U+000A (line feed) must be accepted from a UTF-8 stream only in the form 0x0A, but not in any of the following five possible overlong forms:

  0xC0 0x8A
  0xE0 0x80 0x8A
  0xF0 0x80 0x80 0x8A
  0xF8 0x80 0x80 0x80 0x8A
  0xFC 0x80 0x80 0x80 0x80 0x8A

Any overlong UTF-8 sequence could be abused to bypass UTF-8 substring tests that look only for the shortest possible encoding. All overlong UTF-8 sequences start with one of the following byte patterns:

1100000x (10xxxxxx)
11100000 100xxxxx (10xxxxxx)
11110000 1000xxxx (10xxxxxx 10xxxxxx)
11111000 10000xxx (10xxxxxx 10xxxxxx 10xxxxxx)
11111100 100000xx (10xxxxxx 10xxxxxx 10xxxxxx 10xxxxxx)

Also note that the code positions U+D800 to U+DFFF (UTF-16 surrogates) as well as U+FFFE and U+FFFF must not occur in normal UTF-8 or UCS-4 data. UTF-8 decoders should treat them like malformed or overlong sequences for safety reasons.

Markus Kuhn’s UTF-8 decoder stress test file contains a systematic collection of malformed and overlong UTF-8 sequences and will help you to verify the robustness of your decoder.

Who invented UTF-8?

The encoding known today as UTF-8 was invented by Ken Thompson. It was born during the evening hours of 1992-09-02 in a New Jersey diner, where he designed it in the presence of Rob Pike on a placemat (see Rob Pike’s UTF-8 history). It replaced an earlier attempt to design a FSS/UTF (file system safe UCS transformation format) that was circulated in an X/Open working document in August 1992 by Gary Miller (IBM), Greger Leijonhufvud and John Entenmann (SMI) as a replacement for the division-heavy UTF-1 encoding from the first edition of ISO 10646-1. By the end of the first week of September 1992, Pike and Thompson had turned AT&T Bell Lab’s Plan 9 into the world’s first operating system to use UTF-8. They reported about their experience at the USENIX Winter 1993 Technical Conference, San Diego, January 25-29, 1993, Proceedings, pp. 43-50. FSS/UTF was briefly also referred to as UTF-2 and later renamed into UTF-8, and pushed through the standards process by the X/Open Joint Internationalization Group XOJIG.

Where do I find nice UTF-8 example files?

A few interesting UTF-8 example files for tests and demonstrations are:

What different encodings are there?

Both the UCS and Unicode standards are first of all large tables that assign to every character an integer number. If you use the term “UCS”, “ISO 10646”, or “Unicode”, this just refers to a mapping between characters and integers. This does not yet specify how to store these integers as a sequence of bytes in memory.

ISO 10646-1 defines the UCS-2 and UCS-4 encodings. These are sequences of 2 bytes and 4 bytes per character, respectively. ISO 10646 was from the beginning designed as a 31-bit character set (with possible code positions ranging from U-00000000 to U-7FFFFFFF), however it took until 2001 for the first characters to be assigned beyond the Basic Multilingual Plane (BMP), that is beyond the first 216 character positions (see ISO 10646-2 and Unicode 3.1). UCS-4 can represent all UCS and Unicode characters, UCS-2 can represent only those from the BMP (U+0000 to U+FFFF).

“Unicode” originally implied that the encoding was UCS-2 and it initially didn’t make any provisions for characters outside the BMP (U+0000 to U+FFFF). When it became clear that more than 64k characters would be needed for certain special applications (historic alphabets and ideographs, mathematical and musical typesetting, etc.), Unicode was turned into a sort of 21-bit character set with possible code points in the range U-00000000 to U-0010FFFF. The 2×1024 surrogate characters (U+D800 to U+DFFF) were introduced into the BMP to allow 1024×1024 non-BMP characters to be represented as a sequence of two 16-bit surrogate characters. This way UTF-16 was born, which represents the extended “21-bit” Unicode in a way backwards compatible with UCS-2. The term UTF-32 was introduced in Unicode to describe a 4-byte encoding of the extended “21-bit” Unicode. UTF-32 is the exact same thing as UCS-4, except that by definition UTF-32 is never used to represent characters above U-0010FFFF, while UCS-4 can cover all 231 code positions up to U-7FFFFFFF. The ISO 10646 working group has agreed to modify their standard to exclude code positions beyond U-0010FFFF, in order to turn the new UCS-4 and UTF-32 into practically the same thing.

In addition to all that, UTF-8 was introduced to provide an ASCII backwards compatible multi-byte encoding. The definitions of UTF-8 in UCS and Unicode differed originally slightly, because in UCS, up to 6-byte long UTF-8 sequences were possible to represent characters up to U-7FFFFFFF, while in Unicode only up to 4-byte long UTF-8 sequences are defined to represent characters up to U-0010FFFF. (The difference was in essence the same as between UCS-4 and UTF-32.)

No endianess is implied by the encoding names UCS-2, UCS-4, UTF-16, and UTF-32, though ISO 10646-1 says that Bigendian should be preferred unless otherwise agreed. It has become customary to append the letters “BE” (Bigendian, high-byte first) and “LE” (Littleendian, low-byte first) to the encoding names in order to explicitly specify a byte order.

In order to allow the automatic detection of the byte order, it has become customary on some platforms (notably Win32) to start every Unicode file with the character U+FEFF (ZERO WIDTH NO-BREAK SPACE), also known as the Byte-Order Mark (BOM). Its byte-swapped equivalent U+FFFE is not a valid Unicode character, therefore it helps to unambiguously distinguish the Bigendian and Littleendian variants of UTF-16 and UTF-32.

A full featured character encoding converter will have to provide the following 13 encoding variants of Unicode and UCS:

UCS-2, UCS-2BE, UCS-2LE, UCS-4, UCS-4LE, UCS-4BE, UTF-8, UTF-16, UTF-16BE, UTF-16LE, UTF-32, UTF-32BE, UTF-32LE

Where no byte order is explicitly specified, use the byte order of the CPU on which the conversion takes place and in an input stream swap the byte order whenever U+FFFE is encountered. The difference between outputting UCS-4 versus UTF-32 and UTF-16 versus UCS-2 lies in handling out-of-range characters. The fallback mechanism for non-representable characters has to be activated in UTF-32 (for characters > U-0010FFFF) or UCS-2 (for characters > U+FFFF) even where UCS-4 or UTF-16 respectively would offer a representation.

Really just of historic interest are UTF-1, UTF-7, SCSU and a dozen other less widely publicised UCS encoding proposals with various properties, none of which ever enjoyed any significant use. Their use should be avoided.

A good encoding converter will also offer options for adding or removing the BOM:

It has also been suggested to use the UTF-8 encoded BOM (0xEF 0xBB 0xBF) as a signature to mark the beginning of a UTF-8 file. This practice should definitely not be used on POSIX systems for several reasons:

In addition to the encoding alternatives, Unicode also specifies various Normalization Forms, which provide reasonable subsets of Unicode, especially to remove encoding ambiguities caused by the presence of precomposed and compatibility characters:

A full-featured character encoding converter should also offer conversion between normalization forms. Care should be used with mapping to NFKD or NFKC, as semantic information might be lost (for instance U+00B2 (SUPERSCRIPT TWO) maps to 2) and extra mark-up information might have to be added to preserve it (e.g., <SUP>2</SUP> in HTML).

What programming languages support Unicode?

More recent programming languages that were developed after around 1993 already have special data types for Unicode/ISO 10646-1 characters. This is the case with Ada95, Java, TCL, Perl, Python, C# and others.

ISO C 90 specifies mechanisms to handle multi-byte encoding and wide characters. These facilities were improved with Amendment 1 to ISO C 90 in 1994 and even further improvements were made in the ISO C 99 standard. These facilities were designed originally with various East-Asian encodings in mind. They are on one side slightly more sophisticated than what would be necessary to handle UCS (handling of “shift sequences”), but also lack support for more advanced aspects of UCS (combining characters, etc.). UTF-8 is an example of what the ISO C standard calls multi-byte encoding. The type wchar_t, which in modern environments is usually a signed 32-bit integer, can be used to hold Unicode characters.

Unfortunately, wchar_t was already widely used for various Asian 16-bit encodings throughout the 1990s. Therefore, the ISO C 99 standard was bound by backwards compatibility. It could not be changed to require wchar_t to be used with UCS, like Java and Ada95 managed to do. However, the C compiler can at least signal to an application that wchar_t is guaranteed to hold UCS values in all locales. To do so, it defines the macro __STDC_ISO_10646__ to be an integer constant of the form yyyymmL. The year and month refer to the version of ISO/IEC 10646 and its amendments that have been implemented. For example, __STDC_ISO_10646__ == 200009L if the implementation covers ISO/IEC 10646-1:2000.

How should Unicode be used under Linux?

Before UTF-8 emerged, Linux users all over the world had to use various different language-specific extensions of ASCII. Most popular were ISO 8859-1 and ISO 8859-2 in Europe, ISO 8859-7 in Greece, KOI-8 / ISO 8859-5 / CP1251 in Russia, EUC and Shift-JIS in Japan, BIG5 in Taiwan, etc. This made the exchange of files difficult and application software had to worry about various small differences between these encodings. Support for these encodings was usually incomplete, untested, and unsatisfactory, because the application developers rarely used all these encodings themselves.

Because of these difficulties, major Linux distributors and application developers are now phasing out these older legacy encodings in favour of UTF-8. UTF-8 support has improved dramatically over the last few years and many people now use UTF-8 on a daily basis in

and in any other places where byte sequences used to be interpreted in ASCII.

In UTF-8 mode, terminal emulators such as xterm or the Linux console driver transform every keystroke into the corresponding UTF-8 sequence and send it to the stdin of the foreground process. Similarly, any output of a process on stdout is sent to the terminal emulator, where it is processed with a UTF-8 decoder and then displayed using a 16-bit font.

Full Unicode functionality with all bells and whistles (e.g. high-quality typesetting of the Arabic and Indic scripts) can only be expected from sophisticated multi-lingual word-processing packages. What Linux supports today on a broad base is far simpler and mainly aimed at replacing the old 8- and 16-bit character sets. Linux terminal emulators and command line tools usually only support a Level 1 implementation of ISO 10646-1 (no combining characters), and only scripts such as Latin, Greek, Cyrillic, Armenian, Georgian, CJK, and many scientific symbols are supported that need no further processing support. At this level, UCS support is very comparable to ISO 8859 support and the only significant difference is that we have now thousands of different characters available, that characters can be represented by multibyte sequences, and that ideographic Chinese/Japanese/Korean characters require two terminal character positions (double-width).

Level 2 support in the form of combining characters for selected scripts (in particular Thai) and Hangul Jamo is in parts also available (i.e., some fonts, terminal emulators and editors support it via simple overstringing), but precomposed characters should be preferred over combining character sequences where available. More formally, the preferred way of encoding text in Unicode under Linux should be Normalization Form C as defined in Unicode Technical Report #15.

One influential non-POSIX PC operating system vendor (whom we shall leave unnamed here) suggested that all Unicode files should start with the character ZERO WIDTH NOBREAK SPACE (U+FEFF), which is in this role also referred to as the “signature” or “byte-order mark (BOM)”, in order to identify the encoding and byte-order used in a file. Linux/Unix does not use any BOMs and signatures. They would break far too many existing ASCII syntax conventions (such as scripts starting with #!). On POSIX systems, the selected locale identifies already the encoding expected in all input and output files of a process. It has also been suggested to call UTF-8 files without a signature “UTF-8N” files, but this non-standard term is usually not used in the POSIX world.

Before you switch to UTF-8 under Linux, update your installation to a recent distribution with up-to-date UTF-8 support. This is particular the case if you use an installation older than SuSE 9.1 or Red Hat 8.0. Before these, UTF-8 support was not yet mature enough to be recommendable for daily use.

Red Hat Linux 8.0 (September 2002) was the first distribution to take the leap of switching to UTF-8 as the default encoding for most locales. The only exceptions were Chinese/Japanese/Korean locales, for which there were at the time still too many specialized tools available that did not yet support UTF-8. This first mass deployment of UTF-8 under Linux caused most remaining issues to be ironed out rather quickly during 2003. SuSE Linux then switched its default locales to UTF-8 as well, as of version 9.1 (May 2004). It was followed by Ubuntu Linux, the first Debian-derivative that switched to UTF-8 as the system-wide default encoding. With the migration of the three most popular Linux distributions, UTF-8 related bugs have now been fixed in practically all well-maintained Linux tools. Other distributions can be expected to follow soon.

How do I have to modify my software?

If you are a developer, there are several approaches to add UTF-8 support. We can split them into two categories, which I will call soft and hard conversion. In soft conversion, data is kept in its UTF-8 form everywhere and only very few software changes are necessary. In hard conversion, any UTF-8 data that the program reads will be converted into wide-character arrays and will be handled as such everywhere inside the application. Strings will only be converted back to UTF-8 at output time. Internally, a character remains a fixed-size memory object.

We can also distinguish hard-wired and locale-dependent approaches for supporting UTF-8, depending on how much the string processing relies on the standard library. C offers a number of string processing functions designed to handle arbitrary locale-specific multibyte encodings. An application programmer who relies entirely on these can remain unaware of the actual details of the UTF-8 encoding. Chances are then that by merely changing the locale setting, several other multi-byte encodings (such as EUC) will automatically be supported as well. The other way a programmer can go is to hardcode knowledge about UTF-8 into the application. This may lead in some situations to significant performance improvements. It may be the best approach for applications that will only be used with ASCII and UTF-8.

Even where support for every multi-byte encoding supported by libc is desired, it may well be worth to add extra code optimized for UTF-8. Thanks to UTF-8’s self-synchronizing features, it can be processed very efficiently. The locale-dependent libc string functions can be two orders of magnitude slower than equivalent hardwired UTF-8 procedures. A bad teaching example was GNU grep 2.5.1, which relied entirely on locale-dependent libc functions such as mbrlen() for its generic multi-byte encoding support. This made it about 100× slower in multibyte mode than in single-byte mode! Other applications with hardwired support for UTF-8 regular expressions (e.g., Perl 5.8) do not suffer this dramatic slowdown.

Most applications can do very fine with just soft conversion. This is what makes the introduction of UTF-8 on Unix feasible at all. To name two trivial examples, programs such as cat and echo do not have to be modified at all. They can remain completely ignorant as to whether their input and output is ISO 8859-2 or UTF-8, because they handle just byte streams without processing them. They only recognize ASCII characters and control codes such as '\n' which do not change in any way under UTF-8. Therefore the UTF-8 encoding and decoding is done for these applications completely in the terminal emulator.

A small modification will be necessary for any program that determines the number of characters in a string by counting the bytes. With UTF-8, as with other multi-byte encodings, where the length of a text string is of concern, programmers have to distinguish clearly between

  1. the number of bytes,
  2. the number of characters,
  3. the display width (e.g., the number of cursor position cells in a VT100 terminal emulator)
of a string.

C’s strlen(s) function always counts the number of bytes. This is the number relevant, for example, for memory management (determination of string buffer sizes). Where the output of strlen is used for such purposes, no change will be necessary.

The number of characters can be counted in C in a portable way using mbstowcs(NULL,s,0). This works for UTF-8 like for any other supported encoding, as long as the appropriate locale has been selected. A hard-wired technique to count the number of characters in a UTF-8 string is to count all bytes except those in the range 0x80 – 0xBF, because these are just continuation bytes and not characters of their own. However, the need to count characters arises surprisingly rarely in applications.

In applications written for ASCII or ISO 8859, a far more common use of strlen is to predict the number of columns that the cursor of the terminal will advance if a string is printed. With UTF-8, neither a byte nor a character count will predict the display width, because ideographic characters (Chinese, Japanese, Korean) will occupy two column positions, whereas control and combining characters occupy none. To determine the width of a string on the terminal screen, it is necessary to decode the UTF-8 sequence and then use the wcwidth function to test the display width of each character, or wcswidth to measure the entire string.

For instance, the ls program had to be modified, because without knowing the column widths of filenames, it cannot format the table layout in which it presents directories to the user. Similarly, all programs that assume somehow that the output is presented in a fixed-width font and format it accordingly have to learn how to count columns in UTF-8 text. Editor functions such as deleting a single character have to be slightly modified to delete all bytes that might belong to one character. Affected were for instance editors (vi, emacs, readline, etc.) as well as programs that use the ncurses library.

Any Unix-style kernel can do fine with soft conversion and needs only very minor modifications to fully support UTF-8. Most kernel functions that handle strings (e.g. file names, environment variables, etc.) are not affected at all by the encoding. Modifications were necessary in Linux the following places:

C support for Unicode and UTF-8

Starting with GNU glibc 2.2, the type wchar_t is officially intended to be used only for 32-bit ISO 10646 values, independent of the currently used locale. This is signalled to applications by the definition of the __STDC_ISO_10646__ macro as required by ISO C99. The ISO C multi-byte conversion functions (mbsrtowcs(), wcsrtombs(), etc.) are fully implemented in glibc 2.2 or higher and can be used to convert between wchar_t and any locale-dependent multibyte encoding, including UTF-8, ISO 8859-1, etc.

For example, you can write

  #include <stdio.h>
  #include <locale.h>

  int main()
  {
    if (!setlocale(LC_CTYPE, "")) {
      fprintf(stderr, "Can't set the specified locale! "
              "Check LANG, LC_CTYPE, LC_ALL.\n");
      return 1;
    }
    printf("%ls\n", L"Schöne Grüße");
    return 0;
  }

Call this program with the locale setting LANG=de_DE and the output will be in ISO 8859-1. Call it with LANG=de_DE.UTF-8 and the output will be in UTF-8. The %ls format specifier in printf calls wcsrtombs in order to convert the wide character argument string into the locale-dependent multi-byte encoding.

Many of C’s string functions are locale-independent and they just look at zero-terminated byte sequences:

  strcpy strncpy strcat strncat strcmp strncmp strdup strchr strrchr
  strcspn strspn strpbrk strstr strtok

Some of these (e.g. strcpy) can equally be used for single-byte (ISO 8859-1) and multi-byte (UTF-8) encoded character sets, as they need no notion of how many byte long a character is, while others (e.g., strchr) depend on one character being encoded in a single char value and are of less use for UTF-8 (strchr still works fine if you just search for an ASCII character in a UTF-8 string).

Other C functions are locale dependent and work in UTF-8 locales just as well:

  strcoll strxfrm

How should the UTF-8 mode be activated?

If your application is soft converted and does not use the standard locale-dependent C multibyte routines (mbsrtowcs(), wcsrtombs(), etc.) to convert everything into wchar_t for processing, then it might have to find out in some way, whether it is supposed to assume that the text data it handles is in some 8-bit encoding (like ISO 8859-1, where 1 byte = 1 character) or UTF-8. Once everyone uses only UTF-8, you can just make it the default, but until then both the classical 8-bit sets and UTF-8 may still have to be supported.

The first wave of applications with UTF-8 support used a whole lot of different command line switches to activate their respective UTF-8 modes, for instance the famous xterm -u8. That turned out to be a very bad idea. Having to remember a special command line option or other configuration mechanism for every application is very tedious, which is why command line options are not the proper way of activating a UTF-8 mode.

The proper way to activate UTF-8 is the POSIX locale mechanism. A locale is a configuration setting that contains information about culture-specific conventions of software behaviour, including the character encoding, the date/time notation, alphabetic sorting rules, the measurement system and common office paper size, etc. The names of locales usually consist of ISO 639-1 language and ISO 3166-1 country codes, sometimes with additional encoding names or other qualifiers.

You can get a list of all locales installed on your system (usually in /usr/lib/locale/) with the command locale -a. Set the environment variable LANG to the name of your preferred locale. When a C program executes the setlocale(LC_CTYPE, "") function, the library will test the environment variables LC_ALL, LC_CTYPE, and LANG in that order, and the first one of these that has a value will determine which locale data is loaded for the LC_CTYPE category (which controls the multibyte conversion functions). The locale data is split up into separate categories. For example, LC_CTYPE defines the character encoding and LC_COLLATE defines the string sorting order. The LANG environment variable is used to set the default locale for all categories, but the LC_* variables can be used to override individual categories. Do not worry too much about the country identifiers in the locales. Locales such as en_GB (English in Great Britain) and en_AU (English in Australia) differ usually only in the LC_MONETARY category (name of currency, rules for printing monetary amounts), which practically no Linux application ever uses. LC_CTYPE=en_GB and LC_CTYPE=en_AU have exactly the same effect.

You can query the name of the character encoding in your current locale with the command locale charmap. This should say UTF-8 if you successfully picked a UTF-8 locale in the LC_CTYPE category. The command locale -m provides a list with the names of all installed character encodings.

If you use exclusively C library multibyte functions to do all the conversion between the external character encoding and the wchar_t encoding that you use internally, then the C library will take care of using the right encoding according to LC_CTYPE for you and your program does not even have to know explicitly what the current multibyte encoding is.

However, if you prefer not to do everything using the libc multi-byte functions (e.g., because you think this would require too many changes in your software or is not efficient enough), then your application has to find out for itself when to activate the UTF-8 mode. To do this, on any X/Open compliant systems, where <langinfo.h> is available, you can use a line such as

  utf8_mode = (strcmp(nl_langinfo(CODESET), "UTF-8") == 0);

in order to detect whether the current locale uses the UTF-8 encoding. You have of course to add a setlocale(LC_CTYPE, "") at the beginning of your application to set the locale according to the environment variables first. The standard function call nl_langinfo(CODESET) is also what locale charmap calls to find the name of the encoding specified by the current locale for you. It is available on pretty much every modern Unix now. FreeBSD added nl_langinfo(CODESET) support with version 4.6 (2002-06). If you need an autoconf test for the availability of nl_langinfo(CODESET), here is the one Bruno Haible suggested:

======================== m4/codeset.m4 ================================
#serial AM1

dnl From Bruno Haible.

AC_DEFUN([AM_LANGINFO_CODESET],
[
  AC_CACHE_CHECK([for nl_langinfo and CODESET], am_cv_langinfo_codeset,
    [AC_TRY_LINK([#include <langinfo.h>],
      [char* cs = nl_langinfo(CODESET);],
      am_cv_langinfo_codeset=yes,
      am_cv_langinfo_codeset=no)
    ])
  if test $am_cv_langinfo_codeset = yes; then
    AC_DEFINE(HAVE_LANGINFO_CODESET, 1,
      [Define if you have <langinfo.h> and nl_langinfo(CODESET).])
  fi
])
=======================================================================

[You could also try to query the locale environment variables yourself without using setlocale(). In the sequence LC_ALL, LC_CTYPE, LANG, look for the first of these environment variables that has a value. Make the UTF-8 mode the default (still overridable by command line switches) when this value contains the substring UTF-8, as this indicates reasonably reliably that the C library has been asked to use a UTF-8 locale. An example code fragment that does this is

  char *s;
  int utf8_mode = 0;

  if (((s = getenv("LC_ALL"))   && *s) ||
      ((s = getenv("LC_CTYPE")) && *s) ||
      ((s = getenv("LANG"))     && *s)) {
    if (strstr(s, "UTF-8"))
      utf8_mode = 1;
  }

This relies of course on all UTF-8 locales having the name of the encoding in their name, which is not always the case, therefore the nl_langinfo() query is clearly the better method. If you are really concerned that calling nl_langinfo() might not be portable enough, there is also Markus Kuhn’s portable public domain nl_langinfo(CODESET) emulator for systems that do not have the real thing (and another one from Bruno Haible), and you can use the norm_charmap() function to standardize the output of the nl_langinfo(CODESET) on different platforms.]

How do I get a UTF-8 version of xterm?

The xterm version that comes with XFree86 4.0 or higher (maintained by Thomas Dickey) includes UTF-8 support. To activate it, start xterm in a UTF-8 locale and use a font with iso10646-1 encoding, for instance with

  LC_CTYPE=en_GB.UTF-8 xterm \
    -fn '-Misc-Fixed-Medium-R-SemiCondensed--13-120-75-75-C-60-ISO10646-1'

and then cat some example file, such as UTF-8-demo.txt in the newly started xterm and enjoy what you see.

If you are not using XFree86 4.0 or newer, then you can alternatively download the latest xterm development version separately and compile it yourself with “./configure --enable-wide-chars ; make” or alternatively with “xmkmf; make Makefiles; make; make install; make install.man”.

If you do not have UTF-8 locale support available, use command line option -u8 when you invoke xterm to switch input and output to UTF-8.

How much of Unicode does xterm support?

Xterm in XFree86 4.0.1 only supported Level 1 (no combining characters) of ISO 10646-1 with fixed character width and left-to-right writing direction. In other words, the terminal semantics were basically the same as for ISO 8859-1, except that it can now decode UTF-8 and can access 16-bit characters.

With XFree86 4.0.3, two important functions were added:

If the selected normal font is X × Y pixels large, then xterm will attempt to load in addition a 2X × Y pixels large font (same XLFD, except for a doubled value of the AVERAGE_WIDTH property). It will use this font to represent all Unicode characters that have been assigned the East Asian Wide (W) or East Asian FullWidth (F) property in Unicode Technical Report #11.

The following fonts coming with XFree86 4.x are suitable for display of Japanese and Korean Unicode text with terminal emulators and editors:

  6x13    -Misc-Fixed-Medium-R-SemiCondensed--13-120-75-75-C-60-ISO10646-1
  6x13B   -Misc-Fixed-Bold-R-SemiCondensed--13-120-75-75-C-60-ISO10646-1
  6x13O   -Misc-Fixed-Medium-O-SemiCondensed--13-120-75-75-C-60-ISO10646-1
  12x13ja -Misc-Fixed-Medium-R-Normal-ja-13-120-75-75-C-120-ISO10646-1

  9x18    -Misc-Fixed-Medium-R-Normal--18-120-100-100-C-90-ISO10646-1
  9x18B   -Misc-Fixed-Bold-R-Normal--18-120-100-100-C-90-ISO10646-1
  18x18ja -Misc-Fixed-Medium-R-Normal-ja-18-120-100-100-C-180-ISO10646-1
  18x18ko -Misc-Fixed-Medium-R-Normal-ko-18-120-100-100-C-180-ISO10646-1

Some simple support for nonspacing or enclosing combining characters (i.e., those with general category code Mn or Me in the Unicode database) is now also available, which is implemented by just overstriking (logical OR-ing) a base-character glyph with up to two combining-character glyphs. This produces acceptable results for accents below the base line and accents on top of small characters. It also works well, for example, for Thai and Korean Hangul Conjoining Jamo fonts that were specifically designed for use with overstriking. However, the results might not be fully satisfactory for combining accents on top of tall characters in some fonts, especially with the fonts of the “fixed” family. Therefore precomposed characters will continue to be preferable where available.

The fonts below that come with XFree86 4.x are suitable for display of Latin etc. combining characters (extra head-space). Other fonts will only look nice with combining accents on small x-high characters.

  6x12    -Misc-Fixed-Medium-R-Semicondensed--12-110-75-75-C-60-ISO10646-1
  9x18    -Misc-Fixed-Medium-R-Normal--18-120-100-100-C-90-ISO10646-1
  9x18B   -Misc-Fixed-Bold-R-Normal--18-120-100-100-C-90-ISO10646-1

The following fonts coming with XFree86 4.x are suitable for display of Thai combining characters:

  6x13    -Misc-Fixed-Medium-R-SemiCondensed--13-120-75-75-C-60-ISO10646-1
  9x15    -Misc-Fixed-Medium-R-Normal--15-140-75-75-C-90-ISO10646-1
  9x15B   -Misc-Fixed-Bold-R-Normal--15-140-75-75-C-90-ISO10646-1
  10x20   -Misc-Fixed-Medium-R-Normal--20-200-75-75-C-100-ISO10646-1
  9x18    -Misc-Fixed-Medium-R-Normal--18-120-100-100-C-90-ISO10646-1

The fonts 18x18ko, 18x18Bko, 16x16Bko, and 16x16ko are suitable for displaying Hangul Jamo (using the same simple overstriking character mechanism used for Thai).

A note for programmers of text mode applications:

With support for CJK ideographs and combining characters, the output of xterm behaves a little bit more like with a proportional font, because a Latin/Greek/Cyrillic/etc. character requires one column position, a CJK ideograph two, and a combining character zero.

The Open Group’s Single UNIX Specification specifies the two C functions wcwidth() and wcswidth() that allow an application to test how many column positions a character will occupy:

  #include <wchar.h>
  int wcwidth(wchar_t wc);
  int wcswidth(const wchar_t *pwcs, size_t n);

Markus Kuhn’s free wcwidth() implementation can be used by applications on platforms where the C library does not yet provide a suitable function.

Xterm will for the foreseeable future probably not support the following functionality, which you might expect from a more sophisticated full Unicode rendering engine:

Hebrew and Arabic users will therefore have to use application programs that reverse and left-pad Hebrew and Arabic strings before sending them to the terminal. In other words, the bidirectional processing has to be done by the application and not by xterm. The situation for Hebrew and Arabic improves over ISO 8859 at least in the form of the availability of precomposed glyphs and presentation forms. It is far from clear at the moment, whether bidirectional support should really go into xterm and how precisely this should work. Both ISO 6429 = ECMA-48 and the Unicode bidi algorithm provide alternative starting points. See also ECMA Technical Report TR/53.

If you plan to support bidirectional text output in your application, have a look at either Dov Grobgeld’s FriBidi or Mark Leisher’s Pretty Good Bidi Algorithm, two free implementations of the Unicode bidi algorithm.

Xterm currently does not support the Arabic, Syriac, or Indic text formatting algorithms, although Robert Brady has published some experimental patches towards bidi support. It is still unclear whether it is feasible or preferable to do this in a VT100 emulator at all. Applications can apply the Arabic and Hangul formatting algorithms themselves easily, because xterm allows them to output the necessary presentation forms. For Hangul, Unicode contains the presentation forms needed for modern (post-1933) Korean orthography. For Indic scripts, the X font mechanism at the moment does not even support the encoding of the necessary ligature variants, so there is little xterm could offer anyway. Applications requiring Indic or Syriac output should better use a proper Unicode X11 rendering library such as Pango instead of a VT100 emulator like xterm.

Where do I find ISO 10646-1 X11 fonts?

Quite a number of Unicode fonts have become available for X11 over the past few months, and the list is growing quickly:

Unicode X11 font names end with -ISO10646-1. This is now the officially registered value for the X Logical Font Descriptor (XLFD) fields CHARSET_REGISTRY and CHARSET_ENCODING for all Unicode and ISO 10646-1 16-bit fonts. The *-ISO10646-1 fonts contain some unspecified subset of the entire Unicode character set, and users have to make sure that whatever font they select covers the subset of characters needed by them.

The *-ISO10646-1 fonts usually also specify a DEFAULT_CHAR value that points to a special non-Unicode glyph for representing any character that is not available in the font (usually a dashed box, the size of an H, located at 0x00). This ensures that users at least see clearly that there is an unsupported character. The smaller fixed-width fonts such as 6x13 etc. for xterm will never be able to cover all of Unicode, because many scripts such as Kanji can only be represented in considerably larger pixel sizes than those widely used by European users. Typical Unicode fonts for European usage will contain only subsets of between 1000 and 3000 characters, such as the CEN MES-3 repertoire.

You might notice that in the *-ISO10646-1 fonts the shapes of the ASCII quotation marks has slightly changed to bring them in line with the standards and practice on other platforms.

What are the issues related to UTF-8 terminal emulators?

VT100 terminal emulators accept ISO 2022 (=ECMA-35) ESC sequences in order to switch between different character sets.

UTF-8 is in the sense of ISO 2022 an “other coding system” (see section 15.4 of ECMA 35). UTF-8 is outside the ISO 2022 SS2/SS3/G0/G1/G2/G3 world, so if you switch from ISO 2022 to UTF-8, all SS2/SS3/G0/G1/G2/G3 states become meaningless until you leave UTF-8 and switch back to ISO 2022. UTF-8 is a stateless encoding, i.e. a self-terminating short byte sequence determines completely which character is meant, independent of any switching state. G0 and G1 in ISO 10646-1 are those of ISO 8859-1, and G2/G3 do not exist in ISO 10646, because every character has a fixed position and no switching takes place. With UTF-8, it is not possible that your terminal remains switched to strange graphics-character mode after you accidentally dumped a binary file to it. This makes a terminal in UTF-8 mode much more robust than with ISO 2022 and it is therefore useful to have a way of locking a terminal into UTF-8 mode such that it cannot accidentally go back to the ISO 2022 world.

The ISO 2022 standard specifies a range of ESC % sequences for leaving the ISO 2022 world (designation of other coding system, DOCS), and a number of such sequences have been registered for UTF-8 in section 2.8 of the ISO 2375 International Register of Coded Character Sets:

While a terminal emulator is in UTF-8 mode, any ISO 2022 escape sequences such as for switching G2/G3 etc. are ignored. The only ISO 2022 sequence on which a terminal emulator might act in UTF-8 mode is ESC %@ for returning from UTF-8 back to the ISO 2022 scheme.

UTF-8 still allows you to use C1 control characters such as CSI, even though UTF-8 also uses bytes in the range 0x80-0x9F. It is important to understand that a terminal emulator in UTF-8 mode must apply the UTF-8 decoder to the incoming byte stream before interpreting any control characters. C1 characters are UTF-8 decoded just like any other character above U+007F.

Many text-mode applications available today expect to speak to the terminal using a legacy encoding or to use ISO 2022 sequences for switching terminal fonts. In order to use such applications within a UTF-8 terminal emulator, it is possible to use a conversion layer that will translate between ISO 2022 and UTF-8 on the fly. Examples for such utilities are Juliusz Chroboczek’s luit and pluto. If all you need is ISO 8859 support in a UTF-8 terminal, you can also use screen (version 4.0 or newer) by Michael Schröder and Jürgen Weigert. As implementation of ISO 2022 is a complex and error-prone task, better avoid implementing ISO 2022 yourself. Implement only UTF-8 and point users who need ISO 2022 at luit (or screen).

What UTF-8 enabled applications are available?

Warning: As of mid-2003, this section is becoming increasingly incomplete. UTF-8 support is now a pretty standard feature for most well-maintained packages. This list will soon have to be converted into a list of the most popular programs that still have problems with UTF-8.

Terminal emulation and communication

Editing and word processing

Programming

Mail and Internet

Printing

Misc

What patches to improve UTF-8 support are available?

Many of these already have been included in the respective main distribution.

Are there free libraries for dealing with Unicode available?

What is the status of Unicode support for various X widget libraries?

What packages with UTF-8 support are currently under development?

How does UTF-8 support work under Solaris?

Starting with Solaris 2.8, UTF-8 is at least partially supported. To use it, just set one of the UTF-8 locales, for instance by typing

 setenv LANG en_US.UTF-8
in a C shell.

Now the dtterm terminal emulator can be used to input and output UTF-8 text and the mp print filter will print UTF-8 files on PostScript printers. The en_US.UTF-8 locale is at the moment supported by Motif and CDE desktop applications and libraries, but not by OpenWindows, XView, and OPENLOOK DeskSet applications and libraries.

For more information, read Sun’s Overview of en_US.UTF-8 Locale Support web page.

Can I use UTF-8 on the Web?

Yes. There are two ways in which a HTTP server can indicate to a client that a document is encoded in UTF-8:

The currently most widely used browsers support UTF-8 well enough to generally recommend UTF-8 for use on web pages. The old Netscape 4 browser used an annoyingly large single font for displaying any UTF-8 document. Best upgrade to Mozilla, Netscape 6 or some other recent browser (Netscape 4 is generally very buggy and not maintained any more).

There is also the question of how non-ASCII characters entered into HTML forms are encoded in the subsequent HTTP GET or POST request that transfers the field contents to a CGI script on the server. Unfortunately, both standardization and implementation are still a huge mess here, as discussed in the FORM submission and i18n tutorial by Alan Flavell. We can only hope that a practice of doing all this in UTF-8 will emerge eventually. See also the discussion about Mozilla bug 18643.

How are PostScript glyph names related to UCS codes?

See Adobe’s Unicode and Glyph Names guide.

Are there any well-defined UCS subsets?

With over 40000 characters, a full and complete Unicode implementation is an enormous project. However, it is often sufficient (especially for the European market) to implement only a few hundred or thousand characters as before and still enjoy the simplicity of reaching all required characters in just one single simple encoding via Unicode. A number of different UCS subsets already have been established:

Markus Kuhn’s uniset Perl script allows convenient set arithmetic over UCS subsets for anyone who wants to define a new one or wants to check coverage of an implementation.

What issues are there to consider when converting encodings

The Unicode Consortium maintains a collection of mapping tables between Unicode and various older encoding standards. It is important to understand that the primary purpose of these tables was to demonstrate that Unicode is a superset of the mapped legacy encodings, and to document the motivation and origin behind those Unicode characters that were included into the standard primarily for round-trip compatibility reasons with older character sets. The implementation of good character encoding conversion rountines is a significantly more complex task than just blindly applying these example mapping tables! This is because some character sets distinguish characters that others unify.

The Unicode mapping tables alone are to some degree well suited to directly convert text from the older encodings to Unicode. High-end conversion tools nevertheless should provide interactive mechanisms, where characters that are unified in the legacy encoding but distinguished in Unicode can interactively or semi-automatically be disambiguated on a case-by-case basis.

Conversion in the opposite direction from Unicode to a legacy character set requires non-injective (= many-to-one) extensions of these mapping tables. Several Unicode characters have to be mapped to a single code point in many legacy encodings. The Unicode consortium currently does not maintain standard many-to-one tables for this purpose and does not define any standard behavior of coded character set conversion tools.

Here are some examples for the many-to-one mappings that have to be handled when converting from Unicode into something else:

UCS characters equivalent character in target code
U+00B5 MICRO SIGN
U+03BC GREEK SMALL LETTER MU
0xB5 ISO 8859-1
U+00C5 LATIN CAPITAL LETTER A WITH RING ABOVE
U+212B ANGSTROM SIGN
0xC5 ISO 8859-1
U+03B2 GREEK CAPITAL LETTER BETA
U+00DF LATIN SMALL LETTER SHARP S
0xE1 CP437
U+03A9 GREEK CAPITAL LETTER OMEGA
U+2126 OHM SIGN
0xEA CP437
U+03B5 GREEK SMALL LETTER EPSILON
U+2208 ELEMENT OF
0xEE CP437
U+005C REVERSE SOLIDUS
U+FF3C FULLWIDTH REVERSE SOLIDUS
0x2140 JIS X 0208

A first approximation of such many-to-one tables can be generated from available normalization information, but these then still have to be manually extended and revised. For example, it seems obvious that the character 0xE1 in the original IBM PC character set was meant to be useable as both a Greek small beta (because it is located between the code positions for alpha and gamma) and as a German sharp-s character (because that code is produced when pressing this letter on a German keyboard). Similarly 0xEE can be either the mathematical element-of sign, as well as a small epsilon. These characters are not Unicode normalization equivalents, because although they look similar in low-resolution video fonts, they are very different characters in high-quality typography. IBM’s tables for CP437 reflected one usage in some cases, Microsoft’s the other, both equally sensible. A good code converter should aim to be compatible with both, and not just blindly use the Microsoft mapping table alone when converting from Unicode.

The Unicode database does contain in field 5 the Character Decomposition Mapping that can be used to generate some of the above example mappings automatically. As a rule, the output of a Unicode-to-Something converter should not depend on whether the Unicode input has first been converted into Normalization Form C or not. For equivalence information on Chinese, Japanese, and Korean Han/Kanji/Hanja characters, use the Unihan database. In the cases of the IBM PC characters in the above examples, where the normalization tables do not offer adequate mapping, the cross-references to similar looking characters in the Unicode book are a valuable source of suggestions for equivalence mappings. In the end, which mappings are used and which not is a matter of taste and observed usage.

The Unicode consortium used to maintain mapping tables to CJK character set standards, but has declared them to be obsolete, because their presence on the Unicode web server led to the development of a number of inadequate and naive EUC converters. In particular, the (now obsolete) CJK Unicode mapping tables had to be slightly modified sometimes to preserve information in combination encodings. For example, the standard mappings provide round-trip compatibility for conversion chains ASCII to Unicode to ASCII as well as for JIS X 0208 to Unicode to JIS X 0208. However, the EUC-JP encoding covers the union of ASCII and JIS X 0208, and the UCS repertoire covered by the ASCII and JIS X 0208 mapping tables overlaps for one character, namely U+005C REVERSE SOLIDUS. EUC-JP converters therefore have to use a slightly modified JIS X 0208 mapping table, such that the JIS X 0208 code 0x2140 (0xA1 0xC0 in EUC-JP) gets mapped to U+FF3C FULLWIDTH REVERSE SOLIDUS. This way, round-trip compatibility from EUC-JP to Unicode to EUC-JP can be guaranteed without any loss of information. Unicode Standard Annex #11: East Asian Width provides further guidance on this issue. Another problem area is compatibility with older conversion tables, as explained in an essay by Tomohiro Kubota.

In addition to just using standard normalization mappings, developers of code converters can also offer transliteration support. Transliteration is the conversion of a Unicode character into a graphically and/or semantically similar character in the target code, even if the two are distinct characters in Unicode after normalization. Examples of transliteration:

UCS characters equivalent character in target code
U+0022 QUOTATION MARK
U+201C LEFT DOUBLE QUOTATION MARK
U+201D RIGHT DOUBLE QUOTATION MARK
U+201E DOUBLE LOW-9 QUOTATION MARK
U+201F DOUBLE HIGH-REVERSED-9 QUOTATION MARK
0x22 ISO 8859-1

The Unicode Consortium does not provide or maintain any standard transliteration tables at this time. CEN/TC304 has a draft report “European fallback rules” on recommended ASCII fallback characters for MES-2 in the pipeline, but this is not yet mature. Which transliterations are appropriate or not can in some cases depend on language, application field, and most of all personal preference. Available Unicode transliteration tables include, for example, those found in Bruno Haible’s libiconv, the glibc 2.2 locales, and Markus Kuhn’s transtab package.

Is X11 ready for Unicode?

The X11 R7.0 release (2005) is the latest version of the X Consortium’s sample implementation of the X11 Window System standards. The bulk of the current X11 standards and parts of the sample implementation still pre-date widespread interest in Unicode under Unix.

Among the things that have already been fixed are:

There remain a number of problems in the X11 standards and some inconveniences in the sample implementation for Unicode users that still need to be fixed in one of the next X11 releases:

Several XFree86 team members have worked on these issues. X.Org, the official successor of the X Consortium and the Opengroup as the custodian of the X11 standards and the sample implementation, has taken over the results or is still considering them.

With regard to the font related problems, the solution will probably be to dump the old server-side font mechanisms entirely and use instead XFree86’s new Xft. Another related work-in-progress is Standard Type Services (ST) framework that Sun has been working on.

What are useful Perl one-liners for working with UTF-8?

These examples assume that you have Perl 5.8.1 or newer and that you work in a UTF-8 locale (i.e., “locale charmap” outputs “UTF-8”).

For Perl 5.8.0, option -C is not needed and the examples without -C will not work in a UTF-8 locale. You really should no longer use Perl 5.8.0, as its Unicode support had lots of bugs.

Print the euro sign (U+20AC) to stdout:

  perl -C -e 'print pack("U",0x20ac)."\n"'
  perl -C -e 'print "\x{20ac}\n"'           # works only from U+0100 upwards

Locate malformed UTF-8 sequences:

  perl -ne '/^(([\x00-\x7f]|[\xc0-\xdf][\x80-\xbf]|[\xe0-\xef][\x80-\xbf]{2}|[\xf0-\xf7][\x80-\xbf]{3})*)(.*)$/;print "$ARGV:$.:".($-[3]+1).":$_" if length($3)'

Convert non-ASCII characters into SGML/HTML/XML-style decimal numeric character references (e.g. Ş becomes &#350;):

  perl -C -pe 's/([^\x00-\x7f])/sprintf("&#%d;", ord($1))/ge;'

Convert (hexa)decimal numeric character references to UTF-8:

  perl -C -pe 's/&\#(\d+);/chr($1)/ge;s/&\#x([a-fA-F\d]+);/chr(hex($1))/ge;'

How can I enter Unicode characters?

There are a range of techniques for entering Unicode characters that are not present by default on your keyboard.

Application-independent methods

Application-specific methods

Are there any good mailing lists on these issues?

You should certainly be on the linux-utf8@nl.linux.org mailing list. That’s the place to meet for everyone interested in working towards better UTF-8 support for GNU/Linux or Unix systems and applications. To subscribe, send a message to mailto:linux-utf8-request@nl.linux.org?Subject=subscribe with the subject subscribe. You can also browse the linux-utf8 archive.

There is also the unicode@unicode.org mailing list, which is the best way of finding out what the authors of the Unicode standard and a lot of other gurus have to say. To subscribe, send to unicode-request@unicode.org a message with the subject line “subscribe” and the text “subscribe YOUR@EMAIL.ADDRESS unicode”.

The relevant mailing lists for discussions about Unicode support in Xlib and the X server are the fonts and i18n at xfree86.org mailing lists.

Further references

I add new material to this document quite frequently, so please come back from time to time. Suggestions for improvement are very welcome. Please help to spread the word in the free software community about the importance of UTF-8.

Special thanks to Ulrich Drepper, Bruno Haible, Robert Brady, Juliusz Chroboczek, Shuhei Amakawa, Jungshik Shi, Robert Rogers and many others for valuable comments, and to SuSE GmbH, Nürnberg, for their support.

Markus Kuhn

created 1999-06-04 – last modified 2005-12-29 – http://www.cl.cam.ac.uk/~mgk25/unicode.html