Unicode

A minimal introduction and OCaml tips.

A minimal introduction

This introduction aims at presenting the minimum one should know to be able to work sanely with Unicode. It is not specific to OCaml.

Characters, if they exist

The purpose of Unicode is to have a universal way of representing characters of writing systems known to the world in computer systems. Defining the notion of character is a very complicated question with both philosophical and political implications. To side step these issues, we only talk about characters from a programmer's point of view and simply say that the purpose of Unicode is to assign meaning to the integers of a well-defined integer range.

This range is called the Unicode codespace, it spans from 0x0000 to 0x10FFFF and its boundaries are cast in stone. Members of this range are called Unicode code points. Note that an OCaml int value can represent them on both 32- and 64-bit platforms.

There's a lot of (non-exclusive) terminology predicates that can be applied to code points. I will only mention the most useful ones here.

First there are the reserved or unassigned code points, those are the integers to which the standard doesn't assign any meaning yet. They are reserved for future assignment and may become meaningful in newer versions of the standard. Be aware that once a code point has been assigned (aka as encoded) by the standard most of its properties may never change again, see the stability policy for details.

A very important subset of code points are the Unicode scalar values, these are the code points that belong to the ranges 0x00000xD7FF and 0xE0000x10FFFF. This is the complete Unicode codespace minus the range 0xD8000xDFFF of so called surrogate code points, a hack to be able to encode all scalar values in UTF-16 (more on that below).

Scalar values are what I call, by a total abuse of terminology, the Unicode characters; it is what a proper uchar type should represent. From a programmer's point of view they are the sole integers you will have to deal with during processing and the only code points that you are allowed to serialize and deserialize to valid Unicode byte sequences. Since OCaml 4.03 the standard library defines the Stdlib.Uchar.t type to represent them.

Unicode uses a standard notation to denote code points in running text. A code point is expressed as U+n where n is four to six uppercase hexadecimal digits with leading zeros omitted unless the code point has fewer than four digits (in printf words "U+%04X"). For example the code point bounds are expressed by U+0000 and U+10FFFF and the surrogate bounds by U+D800 and U+DFFF.

What is assigned ?

Lots of the world's scripts are encoded in the standard. The code charts give a precise idea of the coverage.

In order to be successful Unicode decided to be inclusive and to contain pre-existing international and national standards. For example the scalar values from U+0000 to U+007F correspond exactly to the code values of characters encoded by the US-ASCII standard, while those from U+0000 to U+00FF correspond exactly to the code values of ISO-8859-1 (latin1). Many other standard are injected into the codespace but their map to Unicode scalar values may not be as straightforward as the two examples given above.

One thing to be aware of is that because of the inclusive nature of the standard the same abstract character may be represented in more than one way by the standard. A simple example is the latin character "é", which can either be represented by the single scalar value U+00E9 or by the sequence of scalar values <U+0065, U+0301> that is a latin small letter "e" followed by the combining acute accent "´". This non uniqueness of representation is problematic, for example whenever you want to test sequences of scalar values for equality. Unicode solves this by defining equivalence classes between sequences of scalar values, this is called Unicode normalization and we will talk about it later.

Another issue is character spoofing. Many encoded characters resemble each other when displayed but have different scalar values and meaning. The Unicode Security FAQ has more information and pointers about these issues.

Serializing integers — UTF-X

There is more than one way of representing a large integer as a sequence of bytes. The Unicode standard defines seven encoding schemes, also known as Unicode transformation formats (UTF), that precisely define how to encode and decode scalar values — take note, scalar values, not code points — as byte sequences.

The cost of using one representation over the other depends on the character usage. For example UTF-8 is fine for latin scripts but wasteful for east-asian scripts, while the converse is true for UTF-16. I never saw any usage of UTF-32 on disk or wires, it is very wasteful. However, in memory, UTF-32 has the advantage that characters become directly indexable.

For more information see the Unicode UTF-8, UTF-16, UTF-32 and BOM FAQ.

Useful scalar values

The following scalar values are useful to know:

Equivalence and normalization

We mentioned above that concrete textual data may be represented by more than one sequence of scalar values. Latin letters with diacritics are a simple example of that. In order to be able to test two sequences of scalar values for equality we should be able to ignore these differences. The easiest way to do so is to convert them to a normal form where these differences are removed and then use binary equality to test them.

However first we need to define a notion of equality between sequences. Unicode defines two of them, which one to use depends on your processing context.

Canonical equivalence is included in compatibility equivalence: two canonically equivalent sequences are also compatibility equivalent, but the converse may not be true.

A normal form is a function mapping a sequence of scalar values to a sequence of scalar values. The Unicode standard defines four different normal forms, the one to use depends on the equivalence you want and your processing context:

Once you have two sequences in a known normal form you can compare them using binary equality. If the normal form is NFD or NFC, binary equality will entail canonical equivalence of the sequences. If the normal form is NFKC or NFKD equality will entail compatibility equivalence of the sequences. Note that normal forms are not closed under concatenation: if you concatenate two sequence of scalar values you have to renormalize the result.

For more information about normalization, see the Normalization FAQ.

Collation — sorting in alphabetical order

Normalisation forms allow to define a total order between sequences of scalar values using binary comparison. However this order is purely arbitrary. It has no meaning because the magnitude of a scalar value has, in general, no meaning. The process of ordering sequences of scalar values in a standard order like alphabetical order is called collation. Unicode defines a customizable algorithm to order two sequences of scalar values in a meaningful way, the Unicode collation algorithm. For more information and further pointers see the Unicode Collation FAQ.

OCaml tips

Characters as Uchar.t values.

Since OCaml 4.03 the standard library defines the Stdlib.Uchar.t type which represents Unicode scalar values. Support for previous OCaml versions is provided by the uchar OPAM/ocamlfind compatibility package.

Unicode text as UTF-8 encoded OCaml strings

For most OCaml programs it will be entirely sufficient to deal with Unicode by just treating the byte sequence of regular OCaml strings as valid UTF-8 encoded data.

Many libraries will already return Unicode text under this representation. Besides latin1 identifiers having been deprecated in OCaml 4.01, UTF-8 encoding your sources allows you to write UTF-8 encoded string literals directly in your programs. Be aware though that as far as OCaml's compiler is concerned these are just sequences of bytes and you can't trust these strings to be valid UTF-8 as they depend on how correctly your editor encodes them. That is you will need to validate and most likely normalize them unless you:

Checking the validity of UTF-8 strings should only be performed at the boundaries of your program: on your string literals, on data input or on the results of untrusted libraries (be careful, some libraries like Yojson will happily return invalid UTF-8 strings). This allows you to only deal with valid UTF-8 throughout your program and avoid redundant validity checks, internally or on output. The following properties of UTF-8 are useful to remember:

For checking validity or recoding the other UTF encoding schemes into UTF-8 encoded OCaml strings, the Uutf module can be used. It will also be useful if you need to fold over the scalar values of your UTF-8 encoded strings, or build new UTF-8 strings from scalar values via Stdlib.Buffer.t values. Support for the latter is however present in the OCaml Stdlib.Buffer module since OCaml 4.06.

UTF-8 and ASCII

As mentioned in Serializing integers — UTF-X, each of the 128 US-ASCII characters is represented by its own US-ASCII byte representation in UTF-8. So if you want to look for an US-ASCII character in an UTF-8 encoded string, you can just scan the bytes. But beware on the nature of your data and the algorithm you need to implement. For example to detect spaces in the string, looking for the US-ASCII space U+0020 may not be sufficient, there are a lot of other space characters like the no break space U+00A0 that are beyond the US-ASCII repertoire. Folding over the scalar values with Uutf and checking them with Uucp.White.is_white_space is a better idea. Same holds for line breaks, see for example Uutf.nln and Uutf.readlines for more information about these issues.

Equate, compare and normalize UTF-8 encoded OCaml strings

If you understood well the above section about equivalence and normalization you should realise that blindly comparing UTF-8 encoded OCaml strings using Stdlib.compare won't bring you anywhere if you don't normalize them before. The Uunf module can be used for that. Remember that concatenating normalized strings does not result in a normalized string.

Using Stdlib.compare on normalized UTF-8 encoded OCaml strings defines a total order on them that you can use with the Map or Set modules as long as you are not interested in the actual meaning of the order.

For case insensitive equality have a look at the sample code of the Case module.

Sort strings alphabetically

The only solution at the moment for collating strings is to use Camomile but be aware that it supports only Unicode 3.2 character data so don't be surprised if newer scripts don't order correctly. The official collation data also has been significantly tweaked since then.

Find user-perceived character, word, sentence and line boundaries in Unicode text.

The Uuseg module implements the Unicode text segmentation algorithms to find user-perceived character, word and sentence boundaries in Unicode text. It also provides an implementation of the Unicode Line Breaking Algorithm to find line breaks and line break opportunities.

Among other things the Uuseg_string module uses these algorithms to provide OCaml standard library formatters for best-effort formatting of UTF-8 encoded strings.

Unicode readline

A readline function as mandated by the Unicode standard is available in Uutf's sample code.

Range processing

Forget about trying to process Unicode characters using hard coded ranges of scalar values like it was possible to do with US-ASCII. The Unicode standard is not closed, it is evolving, new characters are being assigned. This makes it impossible to derive properties based simply on their integer value or position in ranges of characters. That's the reason why we have the Unicode character database and Uucp to access their properties. Using Uucp.White.is_white_space will be future proof should a new character deemed white be added to the standard (both Uucp and your program will need a recompile though).

Transcoding

Transcoding from legacy encodings to Unicode may be quite involved, use Camomile if you need to do that. There is however one translation that is very easy and direct: it is the one from ISO 8859-1 also known as latin1, the default encoding of OCaml chars. latin1 having been encoded in Unicode in the range of scalar values U+0000 to U+00FF which corresponds to latin1 code value, the translation is trivial, it is the identity:

let char_to_scalar_value c = Char.code c
let char_of_scalar_value s =
    if s > 255 then invalid_arg "" (* can't represent *) else
    Char.chr s

Pretty-printing code points in ASCII

"U+%04X" is an OCaml formatting string for printing an US-ASCII representation of an Unicode code point according to the standards' notational conventions. This is what the Fmt.Dump.uchar formatter does for Stdlib.Uchar.t values.

Writing OCaml libraries

If you write a library that deals with textual data, you should, unless technically impossible, always interact with the client of the library using Unicode. If there are other encodings involved transcode them to/from Unicode so that the client needs only to deal with Unicode, the burden of dealing with the encoding mess has to be on the library, not the client.

In this case there is no absolute need to depend on an Unicode text data structure, just use valid UTF-8 encoded data as OCaml strings and the standard library Stdlib.Uchar.t type.

Specify clearly in the documentation that all the strings returned by or given to the library must be valid UTF-8 encoded data. This validity contract is important for performance reasons: it allows both the client and the library to trust the string and forgo redundant validity checks. Remember that concatenating valid UTF-8 strings results in valid UTF-8 string.