2 Using Unicode in Erlang

2.1  Unicode Implementation

Implementing support for Unicode character sets is an ongoing process. The Erlang Enhancement Proposal (EEP) 10 outlined the basics of Unicode support and also specified a default encoding in binaries that all Unicode-aware modules should handle in the future.

The functionality described in EEP10 was implemented in Erlang/OTP as of R13A, but that was by no means the end of it. In R14B01 support for Unicode file names was added, although it was in no way complete and was by default disabled on platforms where no guarantee was given for the file name encoding. With R16A came support for UTF-8 encoded source code, among with enhancements to many of the applications to support both Unicode encoded file names as well as support for UTF-8 encoded files in several circumstances. Most notable is the support for UTF-8 in files read by file:consult/1, release handler support for UTF-8 and more support for Unicode character sets in the I/O-system.

In R17, the encoding default for Erlang source files will be switched to UTF-8 and in R18 Erlang will support atoms in the full Unicode range, meaning full Unicode function and module names

This guide outlines the current Unicode support and gives a couple of recipes for working with Unicode data.

2.2  Understanding Unicode

Experience with the Unicode support in Erlang has made it painfully clear that understanding Unicode characters and encodings is not as easy as one would expect. The complexity of the field as well as the implications of the standard requires thorough understanding of concepts rarely before thought of.

Furthermore the Erlang implementation requires understanding of concepts that never were an issue for many (Erlang) programmers. To understand and use Unicode characters requires that you study the subject thoroughly, even if you're an experienced programmer.

As an example, one could contemplate the issue of converting between upper and lower case letters. Reading the standard will make you realize that, to begin with, there's not a simple one to one mapping in all scripts. Take German as an example, where there's a letter "ß" (Sharp s) in lower case, but the uppercase equivalent is "SS". Or Greek, where "Σ" has two different lowercase forms: "ς" in word-final position and "σ" elsewhere. Or Turkish where dotted and dot-less "i" both exist in lower case and upper case forms, or Cyrillic "I" which usually has no lowercase form. Or of course languages that have no concept of upper case (or lower case). So, a conversion function will need to know not only one character at a time, but possibly the whole sentence, maybe the natural language the translation should be in and also take into account differences in input and output string length and so on. There is at the time of writing no Unicode to_upper/to_lower functionality in Erlang/OTP, but there are publicly available libraries that address these issues.

Another example is the accented characters where the same glyph has two different representations. Let's look at the Swedish "ö". There's a code point for that in the Unicode standard, but you can also write it as "o" followed by U+0308 (Combining Diaeresis, with the simplified meaning that the last letter should have a "¨" above). They have exactly the same glyph. They are for most purposes the same, but they have completely different representations. For example MacOS X converts all file names to use Combining Diaeresis, while most other programs (including Erlang) try to hide that by doing the opposite when for example listing directories. However it's done, it's usually important to normalize such characters to avoid utter confusion.

The list of examples can be made as long as the Unicode standard, I suspect. The point is that one need a kind of knowledge that was never needed when programs only took one or two languages into account. The complexity of human languages and scripts, certainly has made this a challenge when constructing a universal standard. Supporting Unicode properly in your program will require effort.

2.3  What Unicode Is

Unicode is a standard defining code points (numbers) for all known, living or dead, scripts. In principle, every known symbol used in any language has a Unicode code point.

Unicode code points are defined and published by the Unicode Consortium, which is a non profit organization.

Support for Unicode is increasing throughout the world of computing, as the benefits of one common character set are overwhelming when programs are used in a global environment.

Along with the base of the standard: the code points for all the scripts, there are a couple of encoding standards available.

It is vital to understand the difference between encodings and Unicode characters. Unicode characters are code points according to the Unicode standard, while the encodings are ways to represent such code points. An encoding is just a standard for representation, UTF-8 can for example be used to represent a very limited part of the Unicode character set (e.g. ISO-Latin-1), or the full Unicode range. It's just an encoding format.

As long as all character sets were limited to 256 characters, each character could be stored in one single byte, so there was more or less only one practical encoding for the characters. Encoding each character in one byte was so common that the encoding wasn't even named. When we now, with the Unicode system, have a lot more than 256 characters, we need a common way to represent these. The common ways of representing the code points are the encodings. This means a whole new concept to the programmer, the concept of character representation, which was before a non-issue.

Different operating systems and tools support different encodings. For example Linux and MacOS X has chosen the UTF-8 encoding, which is backwards compatible with 7-bit ASCII and therefore affects programs written in plain English the least. Windows on the other hand supports a limited version of UTF-16, namely all the code planes where the characters can be stored in one single 16-bit entity, which includes most living languages.

The most widely spread encodings are:

Bytewise representation
This is not a proper Unicode representation, but the representation used for characters before the Unicode standard. It can still be used to represent character code points in the Unicode standard that have numbers below 256, which corresponds exactly to the ISO-Latin-1 character set. In Erlang, this is commonly denoted latin1 encoding, which is slightly misleading as ISO-Latin-1 is a character code range, not an encoding.
UTF-8
Each character is stored in one to four bytes depending on code point. The encoding is backwards compatible with bytewise representation of 7-bit ASCII as all 7-bit characters are stored in one single byte in UTF-8. The characters beyond code point 127 are stored in more bytes, letting the most significant bit in the first character indicate a multi-byte character. For details on the encoding, the RFC is publicly available. Note that UTF-8 is not compatible with bytewise representation for code points between 128 and 255, so a ISO-Latin-1 bytewise representation is not generally compatible with UTF-8.
UTF-16
This encoding has many similarities to UTF-8, but the basic unit is a 16-bit number. This means that all characters occupy at least two bytes, some high numbers even four bytes. Some programs, libraries and operating systems claiming to use UTF-16 only allows for characters that can be stored in one 16-bit entity, which is usually sufficient to handle living languages. As the basic unit is more than one byte, byte-order issues occur, why UTF-16 exists in both a big-endian and little-endian variant. In Erlang, the full UTF-16 range is supported when applicable, like in the unicode module and in the bit syntax.
UTF-32
The most straight forward representation. Each character is stored in one single 32-bit number. There is no need for escapes or any variable amount of entities for one character, all Unicode code points can be stored in one single 32-bit entity. As with UTF-16, there are byte-order issues, UTF-32 can be both big- and little-endian.
UCS-4
Basically the same as UTF-32, but without some Unicode semantics, defined by IEEE and has little use as a separate encoding standard. For all normal (and possibly abnormal) usages, UTF-32 and UCS-4 are interchangeable.

Certain ranges of numbers are left unused in the Unicode standard and certain ranges are even deemed invalid. The most notable invalid range is 16#D800 - 16#DFFF, as the UTF-16 encoding does not allow for encoding of these numbers. It can be speculated that the UTF-16 encoding standard was, from the beginning, expected to be able to hold all Unicode characters in one 16-bit entity, but then had to be extended, leaving a hole in the Unicode range to cope with backward compatibility.

Additionally, the code point 16#FEFF is used for byte order marks (BOM's) and use of that character is not encouraged in other contexts than that. It actually is valid though, as the character "ZWNBS" (Zero Width Non Breaking Space). BOM's are used to identify encodings and byte order for programs where such parameters are not known in advance. Byte order marks are more seldom used than one could expect, but their use might become more widely spread as they provide the means for programs to make educated guesses about the Unicode format of a certain file.

2.4  Areas of Unicode Support

To support Unicode in Erlang, problems in several areas have been addressed. Each area is described briefly in this section and more thoroughly further down in this document:

Representation
To handle Unicode characters in Erlang, we have to have a common representation both in lists and binaries. The EEP (10) and the subsequent initial implementation in R13A settled a standard representation of Unicode characters in Erlang.
Manipulation
The Unicode characters need to be processed by the Erlang program, why library functions need to be able to handle them. In some cases functionality was added to already existing interfaces (as the string module now can handle lists with arbitrary code points), in some cases new functionality or options need to be added (as in the io-module, the file handling, the unicode module and the bit syntax). Today most modules in kernel and STDLIB, as well as the VM are Unicode aware.
File I/O
I/O is by far the most problematic area for Unicode. A file is an entity where bytes are stored and the lore of programming has been to treat characters and bytes as interchangeable. With Unicode characters, you need to decide on an encoding as soon as you want to store the data in a file. In Erlang you can open a text file with an encoding option, so that you can read characters from it rather than bytes, but you can also open a file for bytewise I/O. The I/O-system of Erlang has been designed (or at least used) in a way where you expect any I/O-server to be able to cope with any string data, but that is no longer the case when you work with Unicode characters. Handling the fact that you need to know the capabilities of the device where your data ends up is something new to the Erlang programmer. Furthermore, ports in Erlang are byte oriented, so an arbitrary string of (Unicode) characters can not be sent to a port without first converting it to an encoding of choice.
Terminal I/O
Terminal I/O is slightly easier than file I/O. The output is meant for human reading and is usually Erlang syntax (e.g. in the shell). There exists syntactic representation of any Unicode character without actually displaying the glyph (instead written as \x{HHH}), so Unicode data can usually be displayed even if the terminal as such do not support the whole Unicode range.
File names
File names can be stored as Unicode strings, in different ways depending on the underlying OS and file system. This can be handled fairly easy by a program. The problems arise when the file system is not consistent in it's encodings, like for example Linux. Linux allows files to be named with any sequence of bytes, leaving to each program to interpret those bytes. On systems where these "transparent" file names are used, Erlang has to be informed about the file name encoding by a startup flag. The default is bytewise interpretation, which is actually usually wrong, but allows for interpretation of all file names. The concept of "raw file names" can be used to handle wrongly encoded file names if one enables Unicode file name translation (+fnu) on platforms where this is not the default.
Source code encoding
When it comes to the Erlang source code, there is support for the UTF-8 encoding and bytewise encoding. The default in R16B is bytewise (or latin1) encoding. You can control the encoding by a comment like:
%% -*- coding: utf-8 -*-
in the beginning of the file. This of course requires your editor to support UTF-8 as well. The same comment is also interpreted by functions like file:consult/1, the release handler etc, so that you can have all text files in your source directories in UTF-8 encoding.
The language
Having the source code in UTF-8 also allows you to write string literals containing Unicode characters with code points > 255, although atoms, module names and function names will be restricted to the ISO-Latin-1 range until the R18 release. Binary literals where you use the /utf8 type, can also be expressed using Unicode characters > 255. Having module names using characters other than 7-bit ASCII can cause trouble on operating systems with inconsistent file naming schemes, and might also hurt portability, so it's not really recommended. It is suggested in EEP 40 that the language should also allow for Unicode characters > 255 in variable names. Whether to implement that EEP or not is yet to be decided.

2.5  Standard Unicode Representation

In Erlang, strings are actually lists of integers. A string was up until R13 defined to be encoded in the ISO-latin-1 (ISO8859-1) character set, which is, code point by code point, a sub-range of the Unicode character set.

The standard list encoding for strings was therefore easily extended to cope with the whole Unicode range: A Unicode string in Erlang is simply a list containing integers, each integer being a valid Unicode code point and representing one character in the Unicode character set.

Erlang strings in ISO-latin-1 are a subset of Unicode strings.

Only if a string contains code points < 256, can it be directly converted to a binary by using i.e. erlang:iolist_to_binary/1 or can be sent directly to a port. If the string contains Unicode characters > 255, an encoding has to be decided upon and the string should be converted to a binary in the preferred encoding using unicode:characters_to_binary/{1,2,3}. Strings are not generally lists of bytes, as they were before R13. They are lists of characters. Characters are not generally bytes, they are Unicode code points.

Binaries are more troublesome. For performance reasons, programs often store textual data in binaries instead of lists, mainly because they are more compact (one byte per character instead of two words per character, as is the case with lists). Using erlang:list_to_binary/1, an ISO-Latin-1 Erlang string could be converted into a binary, effectively using bytewise encoding - one byte per character. This was very convenient for those limited Erlang strings, but cannot be done for arbitrary Unicode lists.

As the UTF-8 encoding is widely spread and provides some backward compatibility in the 7-bit ASCII range, it is selected as the standard encoding for Unicode characters in binaries for Erlang.

The standard binary encoding is used whenever a library function in Erlang should cope with Unicode data in binaries, but is of course not enforced when communicating externally. Functions and bit-syntax exist to encode and decode both UTF-8, UTF-16 and UTF-32 in binaries. Library functions dealing with binaries and Unicode in general, however, only deal with the default encoding.

Character data may be combined from several sources, sometimes available in a mix of strings and binaries. Erlang has for long had the concept of iodata or iolists, where binaries and lists can be combined to represent a sequence of bytes. In the same way, the Unicode aware modules often allow for combinations of binaries and lists where the binaries have characters encoded in UTF-8 and the lists contain such binaries or numbers representing Unicode code points:

unicode_binary() = binary() with characters encoded in UTF-8 coding standard

chardata() = charlist() | unicode_binary()

charlist() = maybe_improper_list(char() | unicode_binary() | charlist(),
                                 unicode_binary() | nil())

The module unicode in STDLIB even supports similar mixes with binaries containing other encodings than UTF-8, but that is a special case to allow for conversions to and from external data:

external_unicode_binary() = binary() with characters coded in
  a user specified Unicode encoding other than UTF-8 (UTF-16 or UTF-32)

external_chardata() = external_charlist() | external_unicode_binary()

external_charlist() = maybe_improper_list(char() |
                                            external_unicode_binary() |
                                            external_charlist(),
                                          external_unicode_binary() | nil())

2.6  Basic Language Support

As of Erlang/OTP R16 Erlang source files can be written in either UTF-8 or bytewise encoding (a.k.a. latin1 encoding). The details on how to state the encoding of an Erlang source file can be found in epp(3). Strings and comments can be written using Unicode, but functions still have to be named using characters from the ISO-latin-1 character set and atoms are restricted to the same ISO-latin-1 range. These restrictions in the language are of course independent of the encoding of the source file. Erlang/OTP R18 is expected to handle functions named in Unicode as well as Unicode atoms.

Bit-syntax

The bit-syntax contains types for coping with binary data in the three main encodings. The types are named utf8, utf16 and utf32 respectively. The utf16 and utf32 types can be in a big- or little-endian variant:

<<Ch/utf8,_/binary>> = Bin1,
<<Ch/utf16-little,_/binary>> = Bin2,
Bin3 = <<$H/utf32-little, $e/utf32-little, $l/utf32-little, $l/utf32-little,
$o/utf32-little>>,

For convenience, literal strings can be encoded with a Unicode encoding in binaries using the following (or similar) syntax:

Bin4 = <<"Hello"/utf16>>,

String and Character Literals

For source code, there is an extension to the \OOO (backslash followed by three octal numbers) and \xHH (backslash followed by x, followed by two hexadecimal characters) syntax, namely \x{H ...} (a backslash followed by an x, followed by left curly bracket, any number of hexadecimal digits and a terminating right curly bracket). This allows for entering characters of any code point literally in a string even when the encoding of the source file is bytewise (latin1).

In the shell, if using a Unicode input device, or in source code stored in UTF-8, $ can be followed directly by a Unicode character producing an integer. In the following example the code point of a Cyrillic с is output:

7> $с.
1089

Heuristic String Detection

In certain output functions and in the output of return values in the shell, Erlang tries to heuristically detect string data in lists and binaries. Typically you will see heuristic detection in a situation like this:

1> [97,98,99].
"abc"
2> <<97,98,99>>.
<<"abc">>    
3> <<195,165,195,164,195,182>>.
<<"åäö"/utf8>>

Here the shell will detect lists containing printable characters or binaries containing printable characters either in bytewise or UTF-8 encoding. The question here is: what is a printable character? One view would be that anything the Unicode standard thinks is printable, will also be printable according to the heuristic detection. The result would be that almost any list of integers will be deemed a string, resulting in all sorts of characters being printed, maybe even characters your terminal does not have in its font set (resulting in some generic output you probably will not appreciate). Another way is to keep it backwards compatible so that only the ISO-Latin-1 character set is used to detect a string. A third way would be to let the user decide exactly what Unicode ranges are to be viewed as characters. In R16B you can select either the whole Unicode range or the ISO-Latin-1 range by supplying the startup flag +pc Range, where Range is either latin1 or unicode. For backwards compatibility, the default is latin1. This only controls how heuristic string detection is done. In the future, more ranges are expected to be added, so that one can tailor the heuristics to the language and region relevant to the user.

Lets look at an example with the two different startup options:

$ erl +pc latin1
Erlang R16B (erts-5.10.1) [source] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.10.1  (abort with ^G)  
1> [1024].
[1024]
2> [1070,1085,1080,1082,1086,1076].
[1070,1085,1080,1082,1086,1076]
3> [229,228,246].
"åäö"
4> <<208,174,208,189,208,184,208,186,208,190,208,180>>.
<<208,174,208,189,208,184,208,186,208,190,208,180>>
5> <<229/utf8,228/utf8,246/utf8>>.
<<"åäö"/utf8>>
$ erl +pc unicode
Erlang R16B (erts-5.10.1) [source] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.10.1  (abort with ^G)  
1> [1024].
"Ѐ"
2> [1070,1085,1080,1082,1086,1076].
"Юникод"
3> [229,228,246].
"åäö"
4> <<208,174,208,189,208,184,208,186,208,190,208,180>>.
<<"Юникод"/utf8>>
5> <<229/utf8,228/utf8,246/utf8>>.
<<"åäö"/utf8>>

In the examples, we can see that the default Erlang shell will only interpret characters from the ISO-Latin1 range as printable and will only detect lists or binaries with those "printable" characters as containing string data. The valid UTF-8 binary containing "Юникод", will not be printed as a string. When, on the other hand, started with all Unicode characters printable (+pc unicode), the shell will output anything containing printable Unicode data (in binaries either UTF-8 or bytewise encoded) as string data.

These heuristics are also used by io(_lib):format/2 and friends when the t modifier is used in conjunction with ~p or ~P:

$ erl +pc latin1
Erlang R16B (erts-5.10.1) [source] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.10.1  (abort with ^G)  
1> io:format("~tp~n",[{<<"åäö">>, <<"åäö"/utf8>>, <<208,174,208,189,208,184,208,186,208,190,208,180>>}]).
{<<"åäö">>,<<"åäö"/utf8>>,<<208,174,208,189,208,184,208,186,208,190,208,180>>}
ok
$ erl +pc unicode
Erlang R16B (erts-5.10.1) [source] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.10.1  (abort with ^G)  
1> io:format("~tp~n",[{<<"åäö">>, <<"åäö"/utf8>>, <<208,174,208,189,208,184,208,186,208,190,208,180>>}]).
{<<"åäö">>,<<"åäö"/utf8>>,<<"Юникод"/utf8>>}
ok

Please observe that this only affects heuristic interpretation of lists and binaries on output. For example the ~ts format sequence does always output a valid lists of characters, regardless of the +pc setting, as the programmer has explicitly requested string output.

2.7  The Interactive Shell

The interactive Erlang shell, when started towards a terminal or started using the werl command on windows, can support Unicode input and output.

On Windows, proper operation requires that a suitable font is installed and selected for the Erlang application to use. If no suitable font is available on your system, try installing the DejaVu fonts (dejavu-fonts.org), which are freely available and then select that font in the Erlang shell application.

On Unix-like operating systems, the terminal should be able to handle UTF-8 on input and output (modern versions of XTerm, KDE konsole and the Gnome terminal do for example) and your locale settings have to be proper. As an example, my LANG environment variable is set as this:

$ echo $LANG
en_US.UTF-8

Actually, most systems handle the LC_CTYPE variable before LANG, so if that is set, it has to be set to UTF-8:

$ echo $LC_CTYPE
en_US.UTF-8

The LANG or LC_CTYPE setting should be consistent with what the terminal is capable of, there is no portable way for Erlang to ask the actual terminal about its UTF-8 capacity, we have to rely on the language and character type settings.

To investigate what Erlang thinks about the terminal, the io:getopts() call can be used when the shell is started:

$ LC_CTYPE=en_US.ISO-8859-1 erl
Erlang R16B (erts-5.10.1) [source] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.10.1  (abort with ^G)
1> lists:keyfind(encoding, 1, io:getopts()).
{encoding,latin1}
2> q().
ok
$ LC_CTYPE=en_US.UTF-8 erl
Erlang R16B (erts-5.10.1) [source] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.10.1  (abort with ^G)
1> lists:keyfind(encoding, 1, io:getopts()).
{encoding,unicode}
2>

When (finally?) everything is in order with the locale settings, fonts and the terminal emulator, you probably also have discovered a way to input characters in the script you desire. For testing, the simplest way is to add some keyboard mappings for other languages, usually done with some applet in your desktop environment. In my KDE environment, I start the KDE Control Center (Personal Settings), select "Regional and Accessibility" and then "Keyboard Layout". On Windows XP, I start Control Panel->Regional and Language Options, select the Language tab and click the Details... button in the square named "Text services and input Languages". Your environment probably provides similar means of changing the keyboard layout. Make sure you have a way to easily switch back and forth between keyboards if you are not used to this, entering commands using a Cyrillic character set is, as an example, not easily done in the Erlang shell.

Now you are set up for some Unicode input and output. The simplest thing to do is of course to enter a string in the shell:

$ erl
Erlang R16B (erts-5.10.1) [source] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.10.1  (abort with ^G)
1> lists:keyfind(encoding, 1, io:getopts()).
{encoding,unicode}
2> "Юникод".
"Юникод"
3> io:format("~ts~n", [v(2)]).
Юникод
ok
4> 

While strings can be input as Unicode characters, the language elements are still limited to the ISO-latin-1 character set. Only character constants and strings are allowed to be beyond that range:

$ erl
Erlang R16B (erts-5.10.1) [source] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.10.1  (abort with ^G)
1> $ξ.
958
2> Юникод.
* 1: illegal character
2> 

2.8  Unicode File Names

Most modern operating systems support Unicode file names in some way or another. There are several different ways to do this and Erlang by default treats the different approaches differently:

Mandatory Unicode file naming

Windows and, for most common uses, MacOS X enforces Unicode support for file names. All files created in the file system have names that can consistently be interpreted. In MacOS X, all file names are retrieved in UTF-8 encoding, while Windows has selected an approach where each system call handling file names has a special Unicode aware variant, giving much the same effect. There are no file names on these systems that are not Unicode file names, why the default behavior of the Erlang VM is to work in "Unicode file name translation mode", meaning that a file name can be given as a Unicode list and that will be automatically translated to the proper name encoding for the underlying operating and file system.

Doing i.e. a file:list_dir/1 on one of these systems may return Unicode lists with code points beyond 255, depending on the content of the actual file system.

As the feature is fairly new, you may still stumble upon non core applications that cannot handle being provided with file names containing characters with code points larger than 255, but the core Erlang system should have no problems with Unicode file names.

Transparent file naming

Most Unix operating systems have adopted a simpler approach, namely that Unicode file naming is not enforced, but by convention. Those systems usually use UTF-8 encoding for Unicode file names, but do not enforce it. On such a system, a file name containing characters having code points between 128 and 255 may be named either as plain ISO-latin-1 or using UTF-8 encoding. As no consistency is enforced, the Erlang VM can do no consistent translation of all file names. If the VM would automatically select encoding based on heuristics, one could get unexpected behavior on these systems. By default, Erlang starts in "latin1" file name mode on such systems, meaning bytewise encoding in file names. This allows for list representation of all file names in the system, but, for example, a file named "Östersund.txt", will appear in file:list_dir/1 as either "Östersund.txt" (if the file name was encoded in bytewise ISO-Latin-1 by the program creating the file, or more probably as [195,150,115,116,101,114,115,117,110,100], which is a list containing UTF-8 bytes - not what you would want... If you on the other hand use Unicode file name translation on such a system, non-UTF-8 file names will simply be ignored by functions like file:list_dir/1. They can be retrieved with file:list_dir_all/1, but wrongly encoded file names will appear as "raw file names".

The Unicode file naming support was introduced with OTP release R14B01. A VM operating in Unicode file name translation mode can work with files having names in any language or character set (as long as it is supported by the underlying OS and file system). The Unicode character list is used to denote file or directory names and if the file system content is listed, you will also get Unicode lists as return value. The support lies in the Kernel and STDLIB modules, why most applications (that does not explicitly require the file names to be in the ISO-latin-1 range) will benefit from the Unicode support without change.

On operating systems with mandatory Unicode file names, this means that you more easily conform to the file names of other (non Erlang) applications, and you can also process file names that, at least on Windows, were completely inaccessible (due to having names that could not be represented in ISO-latin-1). Also you will avoid creating incomprehensible file names on MacOS X as the vfs layer of the OS will accept all your file names as UTF-8 and will not rewrite them.

For most systems, turning on Unicode file name translation is no problem even if it uses transparent file naming. Very few systems have mixed file name encodings. A consistent UTF-8 named system will work perfectly in Unicode file name mode. It was still however considered experimental in R14B01 and is still not the default on such systems. Unicode file name translation is turned on with the +fnu switch to the On Linux, a VM started without explicitly stating the file name translation mode will default to latin1 as the native file name encoding. On Windows and MacOS X, the default behavior is that of Unicode file name translation, why the file:native_name_encoding/0 by default returns utf8 on those systems (the fact that Windows actually does not use UTF-8 on the file system level can safely be ignored by the Erlang programmer). The default behavior can, as stated before, be changed using the +fnu or +fnl options to the VM, see the erl program. If the VM is started in Unicode file name translation mode, file:native_name_encoding/0 will return the atom utf8. The +fnu switch can be followed by w, i or e, to control how wrongly encoded file names are to be reported. w means that a warning is sent to the error_logger whenever a wrongly encoded file name is "skipped" in directory listings, i means that those wrongly encoded file names are silently ignored and e means that the API function will return an error whenever a wrongly encoded file (or directory) name is encountered. w is the default. Note that file:read_link/1 will always return an error if the link points to an invalid file name.

In Unicode file name mode, file names given to the BIF open_port/2 with the option {spawn_executable,...} are also interpreted as Unicode. So is the parameter list given in the args option available when using spawn_executable. The UTF-8 translation of arguments can be avoided using binaries, see the discussion about raw file names below.

It is worth noting that the file encoding options given when opening a file has nothing to do with the file name encoding convention. You can very well open files containing data encoded in UTF-8 but having file names in bytewise (latin1) encoding or vice versa.

Note

Erlang drivers and NIF shared objects still can not be named with names containing code points beyond 127. This is a known limitation to be removed in a future release. Erlang modules however can, but it is definitely not a good idea and is still considered experimental.

Notes About Raw File Names

Raw file names were introduced together with Unicode file name support in erts-5.8.2 (OTP R14B01). The reason "raw file names" was introduced in the system was to be able to consistently represent file names given in different encodings on the same system. Having the VM automatically translate a file name that is not in UTF-8 to a list of Unicode characters might seem practical, but this would open up for both duplicate file names and other inconsistent behavior. Consider a directory containing a file named "björn" in ISO-latin-1, while the Erlang VM is operating in Unicode file name mode (and therefore expecting UTF-8 file naming). The ISO-latin-1 name is not valid UTF-8 and one could be tempted to think that automatic conversion in for example file:list_dir/1 is a good idea. But what would happen if we later tried to open the file and have the name as a Unicode list (magically converted from the ISO-latin-1 file name)? The VM will convert the file name given to UTF-8, as this is the encoding expected. Effectively this means trying to open the file named <<"björn"/utf8>>. This file does not exist, and even if it existed it would not be the same file as the one that was listed. We could even create two files named "björn", one named in the UTF-8 encoding and one not. If file:list_dir/1 would automatically convert the ISO-latin-1 file name to a list, we would get two identical file names as the result. To avoid this, we need to differentiate between file names being properly encoded according to the Unicode file naming convention (i.e. UTF-8) and file names being invalid under the encoding. By the common file:list_dir/1 function, the wrongly encoded file names are simply ignored in Unicode file name translation mode, but by the file:list_dir_all/1 function, the file names with invalid encoding are returned as "raw" file names, i.e. as binaries.

The Erlang file module accepts raw file names as input. open_port({spawn_executable, ...} ...) also accepts them. As mentioned earlier, the arguments given in the option list to open_port({spawn_executable, ...} ...) undergo the same conversion as the file names, meaning that the executable will be provided with arguments in UTF-8 as well. This translation is avoided consistently with how the file names are treated, by giving the argument as a binary.

To force Unicode file name translation mode on systems where this is not the default was considered experimental in OTP R14B01 due to the fact that the initial implementation did not ignore wrongly encoded file names, so that raw file names could spread unexpectedly throughout the system. Beginning with R16B, the wrongly encoded file names are only retrieved by special functions (e.g. file:list_dir_all/1), so the impact on existing code is much lower, why it is now supported. Unicode file name translation is expected to be default in future releases.

Even if you are operating without Unicode file naming translation automatically done by the VM, you can access and create files with names in UTF-8 encoding by using raw file names encoded as UTF-8. Enforcing the UTF-8 encoding regardless of the mode the Erlang VM is started in might, in some circumstances be a good idea, as the convention of using UTF-8 file names is spreading.

Notes About MacOS X

MacOS X's vfs layer enforces UTF-8 file names in a quite aggressive way. Older versions did this by simply refusing to create non UTF-8 conforming file names, while newer versions replace offending bytes with the sequence "%HH", where HH is the original character in hexadecimal notation. As Unicode translation is enabled by default on MacOS X, the only way to come up against this is to either start the VM with the +fnl flag or to use a raw file name in bytewise (latin1) encoding. If using a raw filename, with a bytewise encoding containing characters between 127 and 255, to create a file, the file can not be opened using the same name as the one used to create it. There is no remedy for this behaviour, other than keeping the file names in the right encoding.

MacOS X also reorganizes the names of files so that the representation of accents etc is using the "combining characters", i.e. the character ö is represented as the code points [111,776], where 111 is the character o and 776 is the special accent character "combining diaeresis". This way of normalizing Unicode is otherwise very seldom used and Erlang normalizes those file names in the opposite way upon retrieval, so that file names using combining accents are not passed up to the Erlang application. In Erlang the file name "björn" is retrieved as [98,106,246,114,110], not as [98,106,117,776,114,110], even though the file system might think differently. The normalization into combining accents are redone when actually accessing files, so this can usually be ignored by the Erlang programmer.

2.9  Unicode in Environment and Parameters

Environment variables and their interpretation is handled much in the same way as file names. If Unicode file names are enabled, environment variables as well as parameters to the Erlang VM are expected to be in Unicode.

If Unicode file names are enabled, the calls to os:getenv/0, os:getenv/1 and os:putenv/2 will handle Unicode strings. On Unix-like platforms, the built-in functions will translate environment variables in UTF-8 to/from Unicode strings, possibly with code points > 255. On Windows the Unicode versions of the environment system API will be used, also allowing for code points > 255.

On Unix-like operating systems, parameters are expected to be UTF-8 without translation if Unicode file names are enabled.

2.10  Unicode-aware Modules

Most of the modules in Erlang/OTP are of course Unicode-unaware in the sense that they have no notion of Unicode and really should not have. Typically they handle non-textual or byte-oriented data (like gen_tcp etc).

Modules that actually handle textual data (like io_lib, string etc) are sometimes subject to conversion or extension to be able to handle Unicode characters.

Fortunately, most textual data has been stored in lists and range checking has been sparse, why modules like string works well for Unicode lists with little need for conversion or extension.

Some modules are however changed to be explicitly Unicode-aware. These modules include:

unicode

The module unicode is obviously Unicode-aware. It contains functions for conversion between different Unicode formats as well as some utilities for identifying byte order marks. Few programs handling Unicode data will survive without this module.

io

The io module has been extended along with the actual I/O-protocol to handle Unicode data. This means that several functions require binaries to be in UTF-8 and there are modifiers to formatting control sequences to allow for outputting of Unicode strings.

file, group, user

I/O-servers throughout the system are able to handle Unicode data and has options for converting data upon actual output or input to/from the device. As shown earlier, the shell has support for Unicode terminals and the file module allows for translation to and from various Unicode formats on disk.

The actual reading and writing of files with Unicode data is however not best done with the file module as its interface is byte oriented. A file opened with a Unicode encoding (like UTF-8), is then best read or written using the io module.

re

The re module allows for matching Unicode strings as a special option. As the library is actually centered on matching in binaries, the Unicode support is UTF-8-centered.

wx

The wx graphical library has extensive support for Unicode text

The module string works perfectly for Unicode strings as well as for ISO-latin-1 strings with the exception of the language-dependent to_upper and to_lower functions, which are only correct for the ISO-latin-1 character set. Actually they can never function correctly for Unicode characters in their current form, as there are language and locale issues as well as multi-character mappings to consider when converting text between cases. Converting case in an international environment is a big subject not yet addressed in OTP.

2.11  Unicode Data in Files

The fact that Erlang as such can handle Unicode data in many forms does not automatically mean that the content of any file can be Unicode text. The external entities such as ports or I/O-servers are not generally Unicode capable.

Ports are always byte oriented, so before sending data that you are not sure is bytewise encoded to a port, make sure to encode it in a proper Unicode encoding. Sometimes this will mean that only part of the data shall be encoded as e.g. UTF-8, some parts may be binary data (like a length indicator) or something else that shall not undergo character encoding, so no automatic translation is present.

I/O-servers behave a little differently. The I/O-servers connected to terminals (or stdout) can usually cope with Unicode data regardless of the encoding option. This is convenient when one expects a modern environment but do not want to crash when writing to a archaic terminal or pipe. Files on the other hand are more picky. A file can have an encoding option which makes it generally usable by the io-module (e.g. {encoding,utf8}), but is by default opened as a byte oriented file. The file module is byte oriented, why only ISO-Latin-1 characters can be written using that module. The io module is the one to use if Unicode data is to be output to a file with other encoding than latin1 (a.k.a. bytewise encoding). It is slightly confusing that a file opened with e.g. file:open(Name,[read,{encoding,utf8}]), cannot be properly read using file:read(File,N) but you have to use the io module to retrieve the Unicode data from it. The reason is that file:read and file:write (and friends) are purely byte oriented, and should so be, as that is the way to access files other than text files - byte by byte. Just as with ports, you can of course write encoded data into a file by "manually" converting the data to the encoding of choice (using the unicode module or the bit syntax) and then output it on a bytewise encoded (latin1) file.

The rule of thumb is that the file module should be used for files opened for bytewise access ({encoding,latin1}) and the io module should be used when accessing files with any other encoding (e.g. {encoding,uf8}).

Functions reading Erlang syntax from files generally recognize the coding: comment and can therefore handle Unicode data on input. When writing Erlang Terms to a file, you should insert such comments when applicable:

$ erl +fna +pc unicode
Erlang R16B (erts-5.10.1) [source]  [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.10.1  (abort with ^G)
1> file:write_file("test.term",<<"%% coding: utf-8\n[{\"Юникод\",4711}].\n"/utf8>>).
ok
2> file:consult("test.term").   
{ok,[[{"Юникод",4711}]]}
  

2.12  Summary of Options

The Unicode support is controlled by both command line switches, some standard environment variables and the version of OTP you are using. Most options affect mainly the way Unicode data is displayed, not the actual functionality of the API's in the standard libraries. This means that Erlang programs usually do not need to concern themselves with these options, they are more for the development environment. An Erlang program can be written so that it works well regardless of the type of system or the Unicode options that are in effect.

Here follows a summary of the settings affecting Unicode:

The LANG and LC_CTYPE environment variables

The language setting in the OS mainly affects the shell. The terminal (i.e. the group leader) will operate with {encoding, unicode} only if the environment tells it that UTF-8 is allowed. This setting should correspond to the actual terminal you are using.

The environment can also affect file name interpretation, if Erlang is started with the +fna flag.

You can check the setting of this by calling io:getopts(), which will give you an option list containing {encoding,unicode} or {encoding,latin1}.

The +pc {unicode|latin1} flag to erl(1)

This flag affects what is interpreted as string data when doing heuristic string detection in the shell and in io/io_lib:format with the "~tp" and ~tP formatting instructions, as described above.

You can check this option by calling io:printable_range/0, which in R16B will return unicode or latin1. To be compatible with future (expected) extensions to the settings, one should rather use io_lib:printable_list/1 to check if a list is printable according to the setting. That function will take into account new possible settings returned from io:printable_range/0.

The +fn{l|a|u} [{w|i|e}] flag to erl(1)

This flag affects how the file names are to be interpreted. On operating systems with transparent file naming, this has to be specified to allow for file naming in Unicode characters (and for correct interpretation of file names containing characters > 255.

+fnl means bytewise interpretation of file names, which was the usual way to represent ISO-Latin-1 file names before UTF-8 file naming got widespread. This is the default on all Unix-like operating systems except MacOS X.

+fnu means that file names are encoded in UTF-8, which is nowadays the common scheme (although not enforced).

+fna means that you automatically select between +fnl and +fnu, based on the LANG and LC_CTYPE environment variables. This is optimistic heuristics indeed, nothing enforces a user to have a terminal with the same encoding as the file system, but usually, this is the case. This might be the default behavior in a future release.

The file name translation mode can be read with the file:native_name_encoding/0 function, which returns latin1 (meaning bytewise encoding) or utf8.

epp:default_encoding/0

This function returns the default encoding for Erlang source files (if no encoding comment is present) in the currently running release. For R16 this returns latin1 (meaning bytewise encoding). In R17 and forward it is expected to return utf8.

The encoding of each file can be specified using comments as described in epp(3).

io:setopts/{1,2} and the -oldshell/-noshell flags.

When Erlang is started with -oldshell or -noshell, the I/O-server for standard_io is default set to bytewise encoding, while an interactive shell defaults to what the environment variables says.

With the io:setopts/2 function you can set the encoding of a file or other I/O-server. This can also be set when opening a file. Setting the terminal (or other standard_io server) unconditionally to the option {encoding,utf8} will for example make UTF-8 encoded characters being written to the device regardless of how Erlang was started or the users environment.

Opening files with encoding option is convenient when writing or reading text files in a known encoding.

You can retrieve the encoding setting for an I/O-server using io:getopts().

2.13  Recipes

When starting with Unicode, one often stumbles over some common issues. I try to outline some methods of dealing with Unicode data in this section.

Byte Order Marks

A common method of identifying encoding in text-files is to put a byte order mark (BOM) first in the file. The BOM is the code point 16#FEFF encoded in the same way as the rest of the file. If such a file is to be read, the first few bytes (depending on encoding) is not part of the actual text. This code outlines how to open a file which is believed to have a BOM and set the files encoding and position for further sequential reading (preferably using the io module). Note that error handling is omitted from the code:

open_bom_file_for_reading(File) ->
    {ok,F} = file:open(File,[read,binary]),
    {ok,Bin} = file:read(F,4),
    {Type,Bytes} = unicode:bom_to_encoding(Bin),
    file:position(F,Bytes),
    io:setopts(F,[{encoding,Type}]),
    {ok,F}.

The unicode:bom_to_encoding/1 function identifies the encoding from a binary of at least four bytes. It returns, along with an term suitable for setting the encoding of the file, the actual length of the BOM, so that the file position can be set accordingly. Note that file:position/2 always works on byte-offsets, so that the actual byte-length of the BOM is needed.

To open a file for writing and putting the BOM first is even simpler:

open_bom_file_for_writing(File,Encoding) ->
    {ok,F} = file:open(File,[write,binary]),
    ok = file:write(File,unicode:encoding_to_bom(Encoding)),
    io:setopts(F,[{encoding,Encoding}]),
    {ok,F}.

In both cases the file is then best processed using the io module, as the functions in io can handle code points beyond the ISO-latin-1 range.

Formatted I/O

When reading and writing to Unicode-aware entities, like the User or a file opened for Unicode translation, you will probably want to format text strings using the functions in io or io_lib. For backward compatibility reasons, these functions do not accept just any list as a string, but require a special translation modifier when working with Unicode texts. The modifier is t. When applied to the s control character in a formatting string, it accepts all Unicode code points and expect binaries to be in UTF-8:

1> io:format("~ts~n",[<<"åäö"/utf8>>]).
åäö
ok
2> io:format("~s~n",[<<"åäö"/utf8>>]).
åäö
ok

Obviously the second io:format/2 gives undesired output because the UTF-8 binary is not in latin1. For backward compatibility, the non prefixed s control character expects bytewise encoded ISO-latin-1 characters in binaries and lists containing only code points < 256.

As long as the data is always lists, the t modifier can be used for any string, but when binary data is involved, care must be taken to make the right choice of formatting characters. A bytewise encoded binary will also be interpreted as a string and printed even when using ~ts, but it might be mistaken for a valid UTF-8 string and one should therefore avoid using the ~ts control if the binary contains bytewise encoded characters and not UTF-8.

The function format/2 in io_lib behaves similarly. This function is defined to return a deep list of characters and the output could easily be converted to binary data for outputting on a device of any kind by a simple erlang:list_to_binary/1. When the translation modifier is used, the list can however contain characters that cannot be stored in one byte. The call to erlang:list_to_binary/1 will in that case fail. However, if the I/O server you want to communicate with is Unicode-aware, the list returned can still be used directly:

$ erl +pc unicode
Erlang R16B (erts-5.10.1) [source] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.10.1 (abort with ^G)
1> io_lib:format("~ts~n", ["Γιούνικοντ"]).
["Γιούνικοντ","\n"]
2> io:put_chars(io_lib:format("~ts~n", ["Γιούνικοντ"])).
Γιούνικοντ
ok

The Unicode string is returned as a Unicode list, which is recognized as such since the Erlang shell uses the Unicode encoding (and is started with all Unicode characters considered printable). The Unicode list is valid input to the io:put_chars/2 function, so data can be output on any Unicode capable device. If the device is a terminal, characters will be output in the \x{H ...} format if encoding is latin1 otherwise in UTF-8 (for the non-interactive terminal - "oldshell" or "noshell") or whatever is suitable to show the character properly (for an interactive terminal - the regular shell). The bottom line is that you can always send Unicode data to the standard_io device. Files will however only accept Unicode code points beyond ISO-latin-1 if encoding is set to something else than latin1.

Heuristic Identification of UTF-8

While it is strongly encouraged that the actual encoding of characters in binary data is known prior to processing, that is not always possible. On a typical Linux system, there is a mix of UTF-8 and ISO-latin-1 text files and there are seldom any BOM's in the files to identify them.

UTF-8 is designed in such a way that ISO-latin-1 characters with numbers beyond the 7-bit ASCII range are seldom considered valid when decoded as UTF-8. Therefore one can usually use heuristics to determine if a file is in UTF-8 or if it is encoded in ISO-latin-1 (one byte per character) encoding. The unicode module can be used to determine if data can be interpreted as UTF-8:

heuristic_encoding_bin(Bin) when is_binary(Bin) ->
    case unicode:characters_to_binary(Bin,utf8,utf8) of
	Bin ->
	    utf8;
	_ ->
	    latin1
    end.
    

If one does not have a complete binary of the file content, one could instead chunk through the file and check part by part. The return-tuple {incomplete,Decoded,Rest} from unicode:characters_to_binary/{1,2,3} comes in handy. The incomplete rest from one chunk of data read from the file is prepended to the next chunk and we therefore circumvent the problem of character boundaries when reading chunks of bytes in UTF-8 encoding:

heuristic_encoding_file(FileName) ->
    {ok,F} = file:open(FileName,[read,binary]),
    loop_through_file(F,<<>>,file:read(F,1024)).

loop_through_file(_,<<>>,eof) ->
    utf8;
loop_through_file(_,_,eof) ->
    latin1;
loop_through_file(F,Acc,{ok,Bin}) when is_binary(Bin) ->
    case unicode:characters_to_binary([Acc,Bin]) of
	{error,_,_} ->
	    latin1;
	{incomplete,_,Rest} ->
	    loop_through_file(F,Rest,file:read(F,1024));
	Res when is_binary(Res) ->
	    loop_through_file(F,<<>>,file:read(F,1024))
    end.
    

Another option is to try to read the whole file in UTF-8 encoding and see if it fails. Here we need to read the file using io:get_chars/3, as we have to succeed in reading characters with a code point over 255:

heuristic_encoding_file2(FileName) ->
    {ok,F} = file:open(FileName,[read,binary,{encoding,utf8}]),
    loop_through_file2(F,io:get_chars(F,'',1024)).

loop_through_file2(_,eof) ->
    utf8;
loop_through_file2(_,{error,_Err}) ->
    latin1;
loop_through_file2(F,Bin) when is_binary(Bin) ->
    loop_through_file2(F,io:get_chars(F,'',1024)).
    

Lists of UTF-8 Bytes

For various reasons, you may find yourself having a list of UTF-8 bytes. This is not a regular string of Unicode characters as each element in the list does not contain one character. Instead you get the "raw" UTF-8 encoding that you have in binaries. This is easily converted to a proper Unicode string by first converting byte per byte into a binary and then converting the binary of UTF-8 encoded characters back to a Unicode string:

  utf8_list_to_string(StrangeList) ->
    unicode:characters_to_list(list_to_binary(StrangeList)).
    

Double UTF-8 Encoding

When working with binaries, you may get the horrible "double UTF-8 encoding", where strange characters are encoded in your binaries or files that you did not expect. What you may have got, is a UTF-8 encoded binary that is for the second time encoded as UTF-8. A common situation is where you read a file, byte by byte, but the actual content is already UTF-8. If you then convert the bytes to UTF-8, using i.e. the unicode module or by writing to a file opened with the {encoding,utf8} option. You will have each byte in the in the input file encoded as UTF-8, not each character of the original text (one character may have been encoded in several bytes). There is no real remedy for this other than being very sure of which data is actually encoded in which format, and never convert UTF-8 data (possibly read byte by byte from a file) into UTF-8 again.

The by far most common situation where this happens, is when you get lists of UTF-8 instead of proper Unicode strings, and then convert them to UTF-8 in a binary or on a file:

  wrong_thing_to_do() ->
    {ok,Bin} = file:read_file("an_utf8_encoded_file.txt"),
    MyList = binary_to_list(Bin), %% Wrong! It is an utf8 binary!
    {ok,C} = file:open("catastrophe.txt",[write,{encoding,utf8}]), 
    io:put_chars(C,MyList), %% Expects a Unicode string, but get UTF-8
                            %% bytes in a list!
    file:close(C). %% The file catastrophe.txt contains more or less unreadable
                   %% garbage!
    

Make very sure you know what a binary contains before converting it to a string. If no other option exists, try heuristics:

  if_you_can_not_know() ->
    {ok,Bin} = file:read_file("maybe_utf8_encoded_file.txt"),
    MyList = case unicode:characters_to_list(Bin) of
      L when is_list(L) ->
        L;
      _ ->
        binary_to_list(Bin) %% The file was bytewise encoded
    end,
    %% Now we know that the list is a Unicode string, not a list of UTF-8 bytes
    {ok,G} = file:open("greatness.txt",[write,{encoding,utf8}]), 
    io:put_chars(G,MyList), %% Expects a Unicode string, which is what it gets!
    file:close(G). %% The file contains valid UTF-8 encoded Unicode characters!