The standard class codecvt attempts to address conversions between different character encoding schemes. In particular, the standard attempts to detail conversions between the implementation-defined wide characters (hereafter referred to as wchar_t) and the standard type char that is so beloved in classic C (which can now be referred to as narrow characters.) This document attempts to describe how the GNU libstdc++ implementation deals with the conversion between wide and narrow characters, and also presents a framework for dealing with the huge number of other encodings that iconv can convert, including Unicode and UTF8. Design issues and requirements are addressed, and examples of correct usage for both the required specializations for wide and narrow characters and the implementation-provided extended functionality are given.

Around page 425 of the C++ Standard, this charming heading comes into view:

The text around the codecvt definition gives some clues:

Hmm. So, in some unspecified way, Unicode encodings and translations between other character sets should be handled by this class.

Ah ha! Another clue...

At this point, a couple points become clear:

One: The standard clearly implies that attempts to add non-required (yet useful and widely used) conversions need to do so through the third template parameter, stateT.

Two: The required conversions, by specifying mbstate_t as the third template parameter, imply an implementation strategy that is mostly (or wholly) based on the underlying C library, and the functions mcsrtombs and wcsrtombs in particular.

Probably the most frequently asked question about code conversion is: "So dudes, what's the deal with Unicode strings?" The dude part is optional, but apparently the usefulness of Unicode strings is pretty widely appreciated. Sadly, this specific encoding (And other useful encodings like UTF8, UCS4, ISO 8859-10, etc etc etc) are not mentioned in the C++ standard.

A couple of comments:

The thought that all one needs to convert between two arbitrary codesets is two types and some kind of state argument is unfortunate. In particular, encodings may be stateless. The naming of the third parameter as stateT is unfortunate, as what is really needed is some kind of generalized type that accounts for the issues that abstract encodings will need. The minimum information that is required includes:

In addition, multi-threaded and multi-locale environments also impact the design and requirements for code conversions. In particular, they affect the required specialization codecvt<wchar_t, char, mbstate_t> when implemented using standard "C" functions.

Three problems arise, one big, one of medium importance, and one small.

First, the small: mcsrtombs and wcsrtombs may not be multithread-safe on all systems required by the GNU tools. For GNU/Linux and glibc, this is not an issue.

Of medium concern, in the grand scope of things, is that the functions used to implement this specialization work on null-terminated strings. Buffers, especially file buffers, may not be null-terminated, thus giving conversions that end prematurely or are otherwise incorrect. Yikes!

The last, and fundamental problem, is the assumption of a global locale for all the "C" functions referenced above. For something like C++ iostreams (where codecvt is explicitly used) the notion of multiple locales is fundamental. In practice, most users may not run into this limitation. However, as a quality of implementation issue, the GNU C++ library would like to offer a solution that allows multiple locales and or simultaneous usage with computationally correct results. In short, libstdc++ is trying to offer, as an option, a high-quality implementation, damn the additional complexity!

For the required specialization codecvt<wchar_t, char, mbstate_t> , conversions are made between the internal character set (always UCS4 on GNU/Linux) and whatever the currently selected locale for the LC_CTYPE category implements.

The two required specializations are implemented as follows:

codecvt<char, char, mbstate_t>

This is a degenerate (i.e., does nothing) specialization. Implementing this was a piece of cake.

codecvt<char, wchar_t, mbstate_t>

This specialization, by specifying all the template parameters, pretty much ties the hands of implementors. As such, the implementation is straightforward, involving mcsrtombs for the conversions between char to wchar_t and wcsrtombs for conversions between wchar_t and char.

Neither of these two required specializations deals with Unicode characters. As such, libstdc++ implements a partial specialization of the codecvt class with and iconv wrapper class, encoding_state as the third template parameter.

This implementation should be standards conformant. First of all, the standard explicitly points out that instantiations on the third template parameter, stateT, are the proper way to implement non-required conversions. Second of all, the standard says (in Chapter 17) that partial specializations of required classes are a-ok. Third of all, the requirements for the stateT type elsewhere in the standard (see 21.1.2 traits typedefs) only indicate that this type be copy constructible.

As such, the type encoding_state is defined as a non-templatized, POD type to be used as the third type of a codecvt instantiation. This type is just a wrapper class for iconv, and provides an easy interface to iconv functionality.

There are two constructors for encoding_state:

encoding_state() : __in_desc(0), __out_desc(0)

This default constructor sets the internal encoding to some default (currently UCS4) and the external encoding to whatever is returned by nl_langinfo(CODESET).

encoding_state(const char* __int, const char* __ext)

This constructor takes as parameters string literals that indicate the desired internal and external encoding. There are no defaults for either argument.

One of the issues with iconv is that the string literals identifying conversions are not standardized. Because of this, the thought of mandating and or enforcing some set of pre-determined valid identifiers seems iffy: thus, a more practical (and non-migraine inducing) strategy was implemented: end-users can specify any string (subject to a pre-determined length qualifier, currently 32 bytes) for encodings. It is up to the user to make sure that these strings are valid on the target system.

void _M_init()

Strangely enough, this member function attempts to open conversion descriptors for a given encoding_state object. If the conversion descriptors are not valid, the conversion descriptors returned will not be valid and the resulting calls to the codecvt conversion functions will return error.

bool _M_good()

Provides a way to see if the given encoding_state object has been properly initialized. If the string literals describing the desired internal and external encoding are not valid, initialization will fail, and this will return false. If the internal and external encodings are valid, but iconv_open could not allocate conversion descriptors, this will also return false. Otherwise, the object is ready to convert and will return true.

encoding_state(const encoding_state&)

As iconv allocates memory and sets up conversion descriptors, the copy constructor can only copy the member data pertaining to the internal and external code conversions, and not the conversion descriptors themselves.

Definitions for all the required codecvt member functions are provided for this specialization, and usage of codecvt<internal character type, external character type, encoding_state> is consistent with other codecvt usage.

The std::messages facet implements message retrieval functionality equivalent to Java's java.text.MessageFormat .using either GNU gettext or IEEE 1003.1-200 functions.

The std::messages facet is probably the most vaguely defined facet in the standard library. It's assumed that this facility was built into the standard library in order to convert string literals from one locale to the other. For instance, converting the "C" locale's const char* c = "please" to a German-localized "bitte" during program execution.

This class has three public member functions, which directly correspond to three protected virtual member functions.

The public member functions are:

catalog open(const string&, const locale&) const

string_type get(catalog, int, int, const string_type&) const

void close(catalog) const

While the virtual functions are:

catalog do_open(const string&, const locale&) const

string_type do_get(catalog, int, int, const string_type&) const

void do_close(catalog) const

A couple of notes on the standard.

First, why is messages_base::catalog specified as a typedef to int? This makes sense for implementations that use catopen and define nl_catd as int, but not for others. Fortunately, it's not heavily used and so only a minor irritant. This has been reported as a possible defect in the standard (LWG 2028).

Second, by making the member functions const, it is impossible to save state in them. Thus, storing away information used in the 'open' member function for use in 'get' is impossible. This is unfortunate.

The 'open' member function in particular seems to be oddly designed. The signature seems quite peculiar. Why specify a const string& argument, for instance, instead of just const char*? Or, why specify a const locale& argument that is to be used in the 'get' member function? How, exactly, is this locale argument useful? What was the intent? It might make sense if a locale argument was associated with a given default message string in the 'open' member function, for instance. Quite murky and unclear, on reflection.

Lastly, it seems odd that messages, which explicitly require code conversion, don't use the codecvt facet. Because the messages facet has only one template parameter, it is assumed that ctype, and not codecvt, is to be used to convert between character sets.

It is implicitly assumed that the locale for the default message string in 'get' is in the "C" locale. Thus, all source code is assumed to be written in English, so translations are always from "en_US" to other, explicitly named locales.

This is a relatively simple class, on the face of it. The standard specifies very little in concrete terms, so generic implementations that are conforming yet do very little are the norm. Adding functionality that would be useful to programmers and comparable to Java's java.text.MessageFormat takes a bit of work, and is highly dependent on the capabilities of the underlying operating system.

Three different mechanisms have been provided, selectable via configure flags:

A new, standards-conformant non-virtual member function signature was added for 'open' so that a directory could be specified with a given message catalog. This simplifies calling conventions for the gnu model.