Boost C++ Libraries

...one of the most highly regarded and expertly designed C++ library projects in the world. Herb Sutter and Andrei Alexandrescu, C++ Coding Standards

PrevUpHomeNext

Design Discussion

Unicode Support

This section focuses on some of the design questions.

Unicode Support

Unicode support was one of the features specifically requested during the formal review. Throughout this document "Unicode support" is a synonym for "wchar_t" support, assuming that "wchar_t" always uses Unicode encoding. Also, when talking about "ascii" (in lowercase) we'll not mean strict 7-bit ASCII encoding, but rather "char" strings in local 8-bit encoding.

Generally, "Unicode support" can mean many things, but for the program_options library it means that:

  • Each parser should accept either char* or wchar_t*, correctly split the input into option names and option values and return the data.

  • For each option, it should be possible to specify whether the conversion from string to value uses ascii or Unicode.

  • The library guarantees that:

    • ascii input is passed to an ascii value without change

    • Unicode input is passed to a Unicode value without change

    • ascii input passed to a Unicode value, and Unicode input passed to an ascii value will be converted using a codecvt facet (which may be specified by the user).

The important point is that it's possible to have some "ascii options" together with "Unicode options". There are two reasons for this. First, for a given type you might not have the code to extract the value from Unicode string and it's not good to require that such code be written. Second, imagine a reusable library which has some options and exposes options description in its interface. If all options are either ascii or Unicode, and the library does not use any Unicode strings, then the author is likely to use ascii options, making the library unusable inside Unicode applications. Essentially, it would be necessary to provide two versions of the library -- ascii and Unicode.

Another important point is that ascii strings are passed though without modification. In other words, it's not possible to just convert ascii to Unicode and process the Unicode further. The problem is that the default conversion mechanism -- the codecvt facet -- might not work with 8-bit input without additional setup.

The Unicode support outlined above is not complete. For example, we don't support Unicode option names. Unicode support is hard and requires a Boost-wide solution. Even comparing two arbitrary Unicode strings is non-trivial. Finally, using Unicode in option names is related to internationalization, which has it's own complexities. E.g. if option names depend on current locale, then all program parts and other parts which use the name must be internationalized too.

The primary question in implementing the Unicode support is whether to use templates and std::basic_string or to use some internal encoding and convert between internal and external encodings on the interface boundaries.

The choice, mostly, is between code size and execution speed. A templated solution would either link library code into every application that uses the library (thereby making shared library impossible), or provide explicit instantiations in the shared library (increasing its size). The solution based on internal encoding would necessarily make conversions in a number of places and will be somewhat slower. Since speed is generally not an issue for this library, the second solution looks more attractive, but we'll take a closer look at individual components.

For the parsers component, we have three choices:

  • Use a fully templated implementation: given a string of a certain type, a parser will return a parsed_options instance with strings of the same type (i.e. the parsed_options class will be templated).

  • Use internal encoding: same as above, but strings will be converted to and from the internal encoding.

  • Use and partly expose the internal encoding: same as above, but the strings in the parsed_options instance will be in the internal encoding. This might avoid a conversion if parsed_options instance is passed directly to other components, but can be also dangerous or confusing for a user.

The second solution appears to be the best -- it does not increase the code size much and is cleaner than the third. To avoid extra conversions, the Unicode version of parsed_options can also store strings in internal encoding.

For the options descriptions component, we don't have much choice. Since it's not desirable to have either all options use ascii or all of them use Unicode, but rather have some ascii and some Unicode options, the interface of the value_semantic must work with both. The only way is to pass an additional flag telling if strings use ascii or internal encoding. The instance of value_semantic can then convert into some other encoding if needed.

For the storage component, the only affected function is store. For Unicode input, the store function should convert the value to the internal encoding. It should also inform the value_semantic class about the used encoding.

Finally, what internal encoding should we use? The alternatives are: std::wstring (using UCS-4 encoding) and std::string (using UTF-8 encoding). The difference between alternatives is:

  • Speed: UTF-8 is a bit slower

  • Space: UTF-8 takes less space when input is ascii

  • Code size: UTF-8 requires additional conversion code. However, it allows one to use existing parsers without converting them to std::wstring and such conversion is likely to create a number of new instantiations.

There's no clear leader, but the last point seems important, so UTF-8 will be used.

Choosing the UTF-8 encoding allows the use of existing parsers, because 7-bit ascii characters retain their values in UTF-8, so searching for 7-bit strings is simple. However, there are two subtle issues:

  • We need to assume the character literals use ascii encoding and that inputs use Unicode encoding.

  • A Unicode character (say '=') can be followed by 'composing character' and the combination is not the same as just '=', so a simple search for '=' might find the wrong character.

Neither of these issues appear to be critical in practice, since ascii is almost universal encoding and since composing characters following '=' (and other characters with special meaning to the library) are not likely to appear.


PrevUpHomeNext