Cyrus IMAP Server: Internationalization


Cyrus currently transcodes characters to a canonical UTF-8 form for searching. The base spec of IMAP4 only requires understanding multiple character sets to properly implement SEARCH. Since the base spec came out, several extensions have been proposed that require further charset support: SORT, THREAD, and the Sieve subsystem. As of this writing, Cyrus doesn’t correctly support these other commands.

Cyrus currently only believes in 16-bit characters. Technically, Unicode has up to 21-bit characters (expressible in UTF-16 and 3-byte UTF-8) and ISO 10646 allows up to 31-bit characters (though ISO’s current policy is to not allocate any characters outside of the 21-bit Unicode range). The lower 16-bit characters make up the basic multilingual plane (BMP) where the majority of languages live. This restriction is apparent in charset.c:writeutf8(), the UTF-8 decoders, and the Unicode canonicalization table used by Cyrus. Since Cyrus’s known character sets (except for UTF-8) don’t contain any characters off of the BMP this isn’t seen to be a major problem.

Throughout this text, Unicode and ISO 10646 will be used interchangeable to refer to the 16-bit character set of the BMP, regardless of encoding. “Character”, unless otherwise specified, refers to a single Unicode character ffff or under.

cyrus canonical form

Since when users search e-mail messages it’s much easier for them to eliminate false positives than realize there are hits that aren’t displayed, the Cyrus searching algorithm errs on the side of more matches. Before comparing any two strings, Cyrus puts them in a canonical form. Logically, the process works as follows:

  • the input string is translated into a sequence of Unicode characters.
  • each character is transformed into lowercase. (For some characters, a single uppercase character may transform into multiple lowercase characters.)
  • each character is fully decomposed.
  • all whitespace (Unicode general categories starting with Z) is removed.
  • combining diacritical marks, such as the accent on é, are removed. (These are Unicode characters 0300-03ff.)
  • certain characters are expanded to alternative spellings using ASCII characters, such as “æ” to “ae”.
  • the output characters are then encoded in UTF-8.

The actual transcoding does all of these steps at once with the aid of tables, carefully built at compile-time.

The central part of Cyrus’s internationalization support is it’s transcoding routines in lib/charset.[ch], and lib/chartable.[ch]. Cyrus’s transcoding routines are very elegant and very compact, thus somewhat intimidating. During compilation, Cyrus builds up a large number of tables (see mkchartable) and uses them so that it never has to consider more than a single octet at a time while outputting the Cyrus canonical form for an input string.

external interface

lib/charset.h is the public interface for Cyrus lib clients to get character canonicalization and searching support. In contains the following functions:

char *charset_convert(const char *s, int charset, char *buf, int bufsz)
Given a string s in charset charset, decode it into canonical form in buf. buf must be reallocable and currently at least size bufsz.
char *charset_decode_mimeheader(const char *s, char *buf, int bufsz)
Given a string s containing possible MIME encoded substrings (per RFC 2047), decode into canonical form in buf. buf must be reallocable and currently at least size bufsz.
charset_index charset_lookupname(const char *name)
Given name return the Cyrus charset index. 0 always represents US-ASCII. The returned charset_index may be saved in a file; it is stable and is an integer. If this version of Cyrus does not support the charset, CHARSET_UNKNOWN_CHARSET is returned.
comp_pat *charset_compilepat(const char *s)
Compiles a NUL-terminated canonicalized string s into a Boyer-Moore table for fast searching. I’ll describe these compiled patterns later.
void charset_freepat(comp_pat *pat)
Frees a pattern previously return by charset_compilepat().
int charset_searchstring(const char *substr, comp_pat *pat,     const char *s, int len)
Searches for a canonicalized string substr in the canonicalized string s. s is of length len. substr must have been previously compiled into pat. Returns non-zero for a hit, zero for no match.
int charset_searchfile(const char *substr, comp_pat *pat,                               const char *msg_base, int mapnl, int len,                               charset_index charset, int encoding)
Searches for the canonicalized string substr with compiled pattern pat in a large buffer starting at msg_base of length len. The large buffer is of charset charset with the encoding encoding. charset_searchfile() will dynamically unencode and canonicalize the search text looking for substr. (If mapnl is set, the buffer has only \n instead of \r\n, but the length assumes that each \n is dynamically converted to \r\n. This feature is deprecated.)
char *charset_decode_mimebody(const char *msg_base, int len,                                      int encoding, char **buf, int *bufsz,                                      int *outlen)
Decode the MIME body part (per RFC 2045) located in the large buffer starting at msg_base of length len. The large buffer is of encoding encoding. charset_decode_mimebody() will decode into buf. buf must be reallocable and currently at least size bufsz. The number of decoded bytes is returned in outlen.
Used by squatter and possibly other text indexing engines, but not described here.

the TRANSLATE macro: using the transcoding tables

The external interface is implemented with the help of the START and TRANSLATE macros:

void START(struct decode_state *state, const unsigned char (*table)[256][4])
START initializes state to be ready for transcoding of the charset translation table given with table. The starting active table is always the first one in the list passed in.
void TRANSLATE(struct decode_state *state, unsigned char input, unsigned char *outbuf, unsigned outindex)
TRANSLATE takes four parameters: state is the current state of the translation; it must have been initialized with START and is modified by TRANSLATE; input is one octet of input from the stream to be transcoded; outbuf is a pointer to the start of the buffer to write output characters; outindex is the index where this translation should be written. The size of outbuf must be at least outindex + charset_max_translation.

Each charset consists of a set of one or more tables; the table parameter passed into START is the first of these tables and the others are adjacent in memory. Characters are transcoded by indexing into the active table with input and examining the 4 octet translation. The 4 octet translation may consist of 0–3 character translations followed by a control code or a series of control codes. In effect, the translation for a given octet is a mini-program that consists either of UTF-8 octets or control codes. One of the control codes RET, END, JSR, or JMP must occur in the 4 octet translation.

control codes

Control codes are represented by uppercase US-ASCII characters since no uppercase characters can appear in the output translation (recall that Cyrus canonical form downcases). Any uppercase US-ASCII character ([A .. Z]) is thus interpreted specially by the TRANSLATE virtual machine. Any other octet encountered as an output translation is presumed to be part of the UTF-8 output sequence and copied to the output.

The names of control codes are actually C pre-processor defines to uppercase US-ASCII characters. As the mnenomics are easier to understand, I use them in discussing their semantics.

control code reference

TRANSLATE recognizes the following “normal” control codes:

This is the first octet of the four octet sequence, indicating that the desired translation is larger than 3 UTF-8 octets. The next two octets represent an offset to look up in the special chartables_long_translations[] table. After that translation is copied to the outbuf, the final octet is interpreted (it must be either a RET or an END).

The TRANSLATE virtual machine has a stack, fixed at size 1. A JSR copies address of the current active table to the stack and transitions to the active table given by the next two octets. (For instance, table 1 would be the next table after the table given as a parameter to START.) Translation of the current octet stops after encountering a JSR.

JSRs are useful for converting a two octet input character: the first octet in the character will make a JSR to some table; the second octet will produce a translation and RET to the current table.

Since the virtual machine has a fixed size stack, it would be highly unusual for the virtual machine to encounter two different JSRs without an intervening RET.

Similar to JSR, but does not change the stack. It is the equivalent of a goto. JMPs are useful to deal with modal input character sets (such as an escape in ISO-2022-JP, see how the tables are generated).
Indicates that we are done translating this input octet and we should return to the previous active table. It might appear as the first of the 4 translation octets, in which case this input character translates into nothing (it might be whitespace, for instance).
Indicates we are done translating this input octet. When TRANSLATE is next called, that input octet will be interpreted against the current active table; the stack does not change.

In addition, it recognizes the following “special” control codes for charsets that aren’t easily represented by a set of tables, UTF-8 and UTF-7:

UTF-7 consists of US-ASCII characters and a special escape character that indicates a transition to base-64 encoded UTF-16 characters. The virtual machine has built in code to handle the base64 decoding. In UTF-7’s base64, 8 input octets result in 3 characters, so the tables would be rather large.
This indicates that the current octet is the continuation of the base-64 section.
One and two character UTF-8 sequences are handled normally in the charset code. To keep the table size down, 3 octet sequences are handled specially. U83 indicates that the current input octet is the start of a three character sequence. It is also an implicit jump to the 2nd table in the UTF-8 sequence, ending this translation.
This input octet 2nd of 3-octet UTF-8 input, with an implicit jump to the 3rd table.
3rd octet of a 3-octet UTF-8 input. This produces the output characters and has an implicit jump to the 1st table of UTF-8.

Finally, it’s useful to mention the special character EMPTY which is guaranteed not to match any character. It is also represented by an uppercase US-ASCII character.

searching and compiled patterns


brief description of boyer-moore xxx

cyrus implementation

why two arrays? us-ascii optimization, really kinda useless now xxx

meta-data stored at the end xxx

generating the tables: mkchartable

The program mkchartable is used to generate the charset transcoding tables used by TRANSLATE. These tables are carefully constructed so no more than a single octet need be examined at a time; this octet results in either an output stream of UTF-8 characters being generated or some sort of state change.

mkchartable uses three different sorts of input files to generate these tables. These files are located in the lib/charset directory.

charset tables

Each charset file maps a single charset to the corresponding Unicode characters. For the US-ASCII and ISO-8859-x character sets this is trivial: each input byte corresponds to a single Unicode character. (Actually, some ISO-8859-x octets do not map to any Unicode character. In that case, the file either does not contain that octet or map it to “????”.)

Other character sets are trickier. For instance, GB-2312 has both single and double byte characters, but is still a simple map from input character to output character. More complicated are modal character encodings. For instance, ISO-2022-JP starts in US-ASCII mode and uses 1B as an escape character followed by another two characters to select a new mode.

The input charset labels modes with “:” followed by the mode name. The starting mode “US-ASCII” in ISO-2022-JP is preceded by “:US-ASCII”. Mode transitions are denoted by a Unicode conversion of “>newmode” or “:newmode”. To denote that the octet 42 transitions into the “US-ASCII” mode, the charset file has “42 >US-ASCII”. The mode names themselves are arbitrary labels and have no effect on the output.

The input charset labels modes with ”:” followed by the mode name. The mode name is optionally followed by a space and the “<” character. If the “<” character is present, then all translations will be followed by a RET instruction instead of an END instruction.

The transition “>newmode” results in a JSR instruction being generated. A JMP instruction is generated by a transition of “:newmode”.

The input byte can be specified as “*”. This is used to define the “default action” which is used for input bytes that are not otherwise defined for the mode. If the default action is not explicitly stated, it is a translation to EMPTY.

unicode data table

The unidata2.txt file is verbatim from the Unicode standard. More recent versions should be available from their website. Each entry in the file describers a Unicode character by the following properties, separated by semicolons:

  • code point (16-bit character value) in hex
  • character name (unused by Cyrus)
  • general category, such as whitespace or punctuation
  • the canonical combining class (unused)
  • bidirection category (unused)
  • character decomposition
  • decimal digit value (unused)
  • digit value (unused, and, no, I don’t know the difference)
  • numeric value including fractions (unused)
  • mirrored character (unused)
  • Unicode 1.0 name (unused)
  • comment (unused)
  • upper case equivalent (unused)
  • lower case equivalent

In general, Cyrus uses the lower case equivalent if there is one, and the decomposed value otherwise.

unicode fixup table

The unifix.txt file contains Cyrus-specific mappings for characters. It overrides the unidata2.txt table. Each rule in the file is explained with a comment. It’s helpful to remember that the Unicode general categories starting with Z represent whitespace, and whitespace is always removed.

generating chartable.c

how mkchartable works: collapses the encoding modes, the unicode translations, and other normalizations into the output tables described above xxx

for the future

Sieve/ACAP comparators

adjustable normalization?

The use of uppercase US-ASCII characters is one of the annoyances in trying to generalize the charset transcoding. If we continue to restrict the characters under consideration to the BMP, switching to UTF-8 control codes that start 4 or 5 byte sequences is possible.

Another possibility is to use a NUL character as an escape sequence, though this increases the size of each control code by 1 octet.

handle >2 octet input characters

make UTF-8 more regular

consider whether we really need U83, U83_2, U83_3. also consider changing { U83, 0, 0, 0 } translations to { U83, JMP, 0, 1 } sequences to at least eliminate the implicit jump.

require minimal UTF-8 characters



  • [UNICODE] Unicode / ISO 10646
  • [UTF-8] utf-8 RFC
  • [UTF-7] utf-7 RFC
  • [BM] boyer-moore
  • [ACAP] the comparators reference. see section XXX of RFC 2244.