Update.

[glibc.git] / manual / charset.texi

@node Character Set Handling, Locales, String and Array Utilities, Top
@c %MENU% Support for extended character sets
@chapter Character Set Handling

@ifnottex
@macro cal{text}
\text\
@end macro
@end ifnottex

Character sets used in the early days of computing had only six, seven,
or eight bits for each character: there was never a case where more than
eight bits (one byte) were used to represent a single character.  The
limitations of this approach became more apparent as more people
grappled with non-Roman character sets, where not all the characters
that make up a language's character set can be represented by @math{2^8}
choices.  This chapter shows the functionality that was added to the C
library to support multiple character sets.

@menu
* Extended Char Intro::              Introduction to Extended Characters.
* Charset Function Overview::        Overview about Character Handling
                                      Functions.
* Restartable multibyte conversion:: Restartable multibyte conversion
                                      Functions.
* Non-reentrant Conversion::         Non-reentrant Conversion Function.
* Generic Charset Conversion::       Generic Charset Conversion.
@end menu


@node Extended Char Intro
@section Introduction to Extended Characters

A variety of solutions is available to overcome the differences between
character sets with a 1:1 relation between bytes and characters and
character sets with ratios of 2:1 or 4:1.  The remainder of this
section gives a few examples to help understand the design decisions
made while developing the functionality of the @w{C library}.

@cindex internal representation
A distinction we have to make right away is between internal and
external representation.  @dfn{Internal representation} means the
representation used by a program while keeping the text in memory.
External representations are used when text is stored or transmitted
through some communication channel.  Examples of external
representations include files waiting in a directory to be
read and parsed.

Traditionally there has been no difference between the two representations.
It was equally comfortable and useful to use the same single-byte
representation internally and externally.  This comfort level decreases
with more and larger character sets.

One of the problems to overcome with the internal representation is
handling text that is externally encoded using different character
sets.  Assume a program that reads two texts and compares them using
some metric.  The comparison can be usefully done only if the texts are
internally kept in a common format.

@cindex wide character
For such a common format (@math{=} character set) eight bits are certainly
no longer enough.  So the smallest entity will have to grow: @dfn{wide
characters} will now be used.  Instead of one byte per character, two or
four will be used instead.  (Three are not good to address in memory and
more than four bytes seem not to be necessary).

@cindex Unicode
@cindex ISO 10646
As shown in some other part of this manual,
@c !!! Ahem, wide char string functions are not yet covered -- drepper
a completely new family has been created of functions that can handle wide
character texts in memory.  The most commonly used character sets for such
internal wide character representations are Unicode and @w{ISO 10646}
(also known as UCS for Universal Character Set).  Unicode was originally
planned as a 16-bit character set; whereas, @w{ISO 10646} was designed to
be a 31-bit large code space.  The two standards are practically identical.
They have the same character repertoire and code table, but Unicode specifies
added semantics.  At the moment, only characters in the first @code{0x10000}
code positions (the so-called Basic Multilingual Plane, BMP) have been
assigned, but the assignment of more specialized characters outside this
16-bit space is already in progress.  A number of encodings have been
defined for Unicode and @w{ISO 10646} characters:
@cindex UCS-2
@cindex UCS-4
@cindex UTF-8
@cindex UTF-16
UCS-2 is a 16-bit word that can only represent characters
from the BMP, UCS-4 is a 32-bit word than can represent any Unicode
and @w{ISO 10646} character, UTF-8 is an ASCII compatible encoding where
ASCII characters are represented by ASCII bytes and non-ASCII characters
by sequences of 2-6 non-ASCII bytes, and finally UTF-16 is an extension
of UCS-2 in which pairs of certain UCS-2 words can be used to encode
non-BMP characters up to @code{0x10ffff}.

To represent wide characters the @code{char} type is not suitable.  For
this reason the @w{ISO C} standard introduces a new type that is
designed to keep one character of a wide character string.  To maintain
the similarity there is also a type corresponding to @code{int} for
those functions that take a single wide character.

@comment stddef.h
@comment ISO
@deftp {Data type} wchar_t
This data type is used as the base type for wide character strings.
In other words, arrays of objects of this type are the equivalent of 
@code{char[]} for multibyte character strings.  The type is defined in 
@file{stddef.h}.

The @w{ISO C90} standard, where @code{wchar_t} was introduced, does not
say anything specific about the representation.  It only requires that
this type is capable of storing all elements of the basic character set.
Therefore it would be legitimate to define @code{wchar_t} as @code{char},
which might make sense for embedded systems.

But for GNU systems @code{wchar_t} is always 32 bits wide and, therefore,
capable of representing all UCS-4 values and, therefore, covering all of
@w{ISO 10646}.  Some Unix systems define @code{wchar_t} as a 16-bit type
and thereby follow Unicode very strictly.  This definition is perfectly
fine with the standard, but it also means that to represent all
characters from Unicode and @w{ISO 10646} one has to use UTF-16 surrogate
characters, which is in fact a multi-wide-character encoding.  But
resorting to multi-wide-character encoding contradicts the purpose of the
@code{wchar_t} type.
@end deftp

@comment wchar.h
@comment ISO
@deftp {Data type} wint_t
@code{wint_t} is a data type used for parameters and variables that
contain a single wide character.  As the name suggests this type is the
equivalent of @code{int} when using the normal @code{char} strings.  The
types @code{wchar_t} and @code{wint_t} often have the same
representation if their size is 32 bits wide but if @code{wchar_t} is
defined as @code{char} the type @code{wint_t} must be defined as
@code{int} due to the parameter promotion.

@pindex wchar.h
This type is defined in @file{wchar.h} and was introduced in
@w{Amendment 1} to @w{ISO C90}.
@end deftp

As there are for the @code{char} data type macros are available for
specifying the minimum and maximum value representable in an object of
type @code{wchar_t}.

@comment wchar.h
@comment ISO
@deftypevr Macro wint_t WCHAR_MIN
The macro @code{WCHAR_MIN} evaluates to the minimum value representable
by an object of type @code{wint_t}.

This macro was introduced in @w{Amendment 1} to @w{ISO C90}.
@end deftypevr

@comment wchar.h
@comment ISO
@deftypevr Macro wint_t WCHAR_MAX
The macro @code{WCHAR_MAX} evaluates to the maximum value representable
by an object of type @code{wint_t}.

This macro was introduced in @w{Amendment 1} to @w{ISO C90}.
@end deftypevr

Another special wide character value is the equivalent to @code{EOF}.

@comment wchar.h
@comment ISO
@deftypevr Macro wint_t WEOF
The macro @code{WEOF} evaluates to a constant expression of type
@code{wint_t} whose value is different from any member of the extended
character set.

@code{WEOF} need not be the same value as @code{EOF} and unlike
@code{EOF} it also need @emph{not} be negative.  In other words, sloppy 
code like

@smallexample
@{
  int c;
  ...
  while ((c = getc (fp)) < 0)
    ...
@}
@end smallexample

@noindent
has to be rewritten to use @code{WEOF} explicitly when wide characters
are used:

@smallexample
@{
  wint_t c;
  ...
  while ((c = wgetc (fp)) != WEOF)
    ...
@}
@end smallexample

@pindex wchar.h
This macro was introduced in @w{Amendment 1} to @w{ISO C90} and is
defined in @file{wchar.h}.
@end deftypevr


These internal representations present problems when it comes to storing
and transmittal.  Because each single wide character consists of more
than one byte, they are effected by byte-ordering.  Thus, machines with
different endianesses would see different values when accessing the same
data.  This byte ordering concern also applies for communication protocols
that are all byte-based and, thereforet require that the sender has to
decide about splitting the wide character in bytes.  A last (but not least
important) point is that wide characters often require more storage space
than a customized byte-oriented character set.

@cindex multibyte character
@cindex EBCDIC
   For all the above reasons, an external encoding that is different
from the internal encoding is often used if the latter is UCS-2 or UCS-4.
The external encoding is byte-based and can be chosen appropriately for
the environment and for the texts to be handled.  A variety of different
character sets can be used for this external encoding (information that
will not be exhaustively presented here--instead, a description of the
major groups will suffice).  All of the ASCII-based character sets
[_bkoz_: do you mean Roman character sets? If not, what do you mean
here?] fulfill one requirement: they are "filesystem safe."  This means
that the character @code{'/'} is used in the encoding @emph{only} to
represent itself.  Things are a bit different for character sets like
EBCDIC (Extended Binary Coded Decimal Interchange Code, a character set
family used by IBM), but if the operation system does not understand
EBCDIC directly the parameters-to-system calls have to be converted first
anyhow.

@itemize @bullet
@item 
The simplest character sets are single-byte character sets.  There can 
be only up to 256 characters (for @w{8 bit} character sets), which is 
not sufficient to cover all languages but might be sufficient to handle 
a specific text.  Handling of a @w{8 bit} character sets is simple.  This 
is not true for other kinds presented later, and therefore, the 
application one uses might require the use of @w{8 bit} character sets.

@cindex ISO 2022
@item
The @w{ISO 2022} standard defines a mechanism for extended character
sets where one character @emph{can} be represented by more than one
byte.  This is achieved by associating a state with the text.
Characters that can be used to change the state can be embedded in the
text.  Each byte in the text might have a different interpretation in each
state.  The state might even influence whether a given byte stands for a
character on its own or whether it has to be combined with some more
bytes.

@cindex EUC
@cindex Shift_JIS
@cindex SJIS
In most uses of @w{ISO 2022} the defined character sets do not allow
state changes that cover more than the next character.  This has the
big advantage that whenever one can identify the beginning of the byte
sequence of a character one can interpret a text correctly.  Examples of
character sets using this policy are the various EUC character sets
(used by Sun's operations systems, EUC-JP, EUC-KR, EUC-TW, and EUC-CN)
or Shift_JIS (SJIS, a Japanese encoding).

But there are also character sets using a state that is valid for more
than one character and has to be changed by another byte sequence.
Examples for this are ISO-2022-JP, ISO-2022-KR, and ISO-2022-CN.

@item
@cindex ISO 6937
Early attempts to fix 8 bit character sets for other languages using the
Roman alphabet lead to character sets like @w{ISO 6937}.  Here bytes
representing characters like the acute accent do not produce output
themselves: one has to combine them with other characters to get the
desired result.  For example, the byte sequence @code{0xc2 0x61}
(non-spacing acute accent, followed by lower-case `a') to get the ``small
a with  acute'' character.  To get the acute accent character on its own,
one has to write @code{0xc2 0x20} (the non-spacing acute followed by a
space).

Character sets like @w[ISO 6937] are used in some embedded systems such
as teletex.

@item
@cindex UTF-8
Instead of converting the Unicode or @w{ISO 10646} text used internally,
it is often also sufficient to simply use an encoding different than
UCS-2/UCS-4.  The Unicode and @w{ISO 10646} standards even specify such an
encoding: UTF-8.  This encoding is able to represent all of @w{ISO
10646} 31 bits in a byte string of length one to six.

@cindex UTF-7
There were a few other attempts to encode @w{ISO 10646} such as UTF-7,
but UTF-8 is today the only encoding that should be used.  In fact, with
any luck UTF-8 will soon be the only external encoding that has to be
supported.  It proves to be universally usable and its only disadvantage
is that it favors Roman languages by making the byte string
representation of other scripts (Cyrillic, Greek, Asian scripts) longer
than necessary if using a specific character set for these scripts.
Methods like the Unicode compression scheme can alleviate these
problems.
@end itemize

The question remaining is: how to select the character set or encoding
to use.  The answer: you cannot decide about it yourself, it is decided
by the developers of the system or the majority of the users.  Since the
goal is interoperability one has to use whatever the other people one
works with use.  If there are no constraints, the selection is based on
the requirements the expected circle of users will have.  In other words,
if a project is expected to be used in only, say, Russia it is fine to use
KOI8-R or a similar character set.  But if at the same time people from,
say, Greece are participating one should use a character set that allows
all people to collaborate.

The most widely useful solution seems to be: go with the most general
character set, namely @w{ISO 10646}.  Use UTF-8 as the external encoding
and problems about users not being able to use their own language
adequately are a thing of the past.

One final comment about the choice of the wide character representation
is necessary at this point.  We have said above that the natural choice
is using Unicode or @w{ISO 10646}.  This is not required, but at least
encouraged, by the @w{ISO C} standard.  The standard defines at least a
macro @code{__STDC_ISO_10646__} that is only defined on systems where
the @code{wchar_t} type encodes @w{ISO 10646} characters.  If this
symbol is not defined one should avoid making assumptions about the wide
character representation.  If the programmer uses only the functions
provided by the C library to handle wide character strings there should
be no compatibility problems with other systems.

@node Charset Function Overview
@section Overview about Character Handling Functions

A Unix @w{C library} contains three different sets of functions in two 
families to handle character set conversion.  One of the function families 
(the most commonly used) is specified in the @w{ISO C90} standard and, 
therefore, is portable even beyond the Unix world.  Unfortunately this 
family is the least useful one.  These functions should be avoided 
whenever possible, especially when developing libraries (as opposed to 
applications). 

The second family of functions got introduced in the early Unix standards
(XPG2) and is still part of the latest and greatest Unix standard:
@w{Unix 98}.  It is also the most powerful and useful set of functions.
But we will start with the functions defined in @w{Amendment 1} to
@w{ISO C90}.

@node Restartable multibyte conversion
@section Restartable Multibyte Conversion Functions

The @w{ISO C} standard defines functions to convert strings from a
multibyte representation to wide character strings.  There are a number
of peculiarities:

@itemize @bullet
@item
The character set assumed for the multibyte encoding is not specified
as an argument to the functions.  Instead the character set specified by
the @code{LC_CTYPE} category of the current locale is used; see
@ref{Locale Categories}.

@item
The functions handling more than one character at a time require NUL
terminated strings as the argument (i.e., converting blocks of text
does not work unless one can add a NUL byte at an appropriate place). 
The GNU C library contains some extensions to the standard that allow
specifying a size, but basically they also expect terminated strings.
@end itemize

Despite these limitations the @w{ISO C} functions can be used in many
contexts.  In graphical user interfaces, for instance, it is not
uncommon to have functions that require text to be displayed in a wide
character string if the text is not simple ASCII.  The text itself might 
come from a file with translations and the user should decide about the
current locale, which determines the translation and therefore also the
external encoding used.  In such a situation (and many others) the
functions described here are perfect.  If more freedom while performing
the conversion is necessary take a look at the @code{iconv} functions
(@pxref{Generic Charset Conversion}).

@menu
* Selecting the Conversion::     Selecting the conversion and its properties.
* Keeping the state::            Representing the state of the conversion.
* Converting a Character::       Converting Single Characters.
* Converting Strings::           Converting Multibyte and Wide Character
                                  Strings.
* Multibyte Conversion Example:: A Complete Multibyte Conversion Example.
@end menu

@node Selecting the Conversion
@subsection Selecting the conversion and its properties

We already said above that the currently selected locale for the
@code{LC_CTYPE} category decides about the conversion that is performed
by the functions we are about to describe.  Each locale uses its own
character set (given as an argument to @code{localedef}) and this is the
one assumed as the external multibyte encoding.  The wide character
character set always is UCS-4, at least on GNU systems.

A characteristic of each multibyte character set is the maximum number
of bytes that can be necessary to represent one character.  This
information is quite important when writing code that uses the
conversion functions (as shown in the examples below).
The @w{ISO C} standard defines two macros that provide this information.


@comment limits.h
@comment ISO
@deftypevr Macro int MB_LEN_MAX
@code{MB_LEN_MAX} specifies the maximum number of bytes in the multibyte
sequence for a single character in any of the supported locales.  It is
a compile-time constant and is defined in @file{limits.h}.
@pindex limits.h
@end deftypevr

@comment stdlib.h
@comment ISO
@deftypevr Macro int MB_CUR_MAX
@code{MB_CUR_MAX} expands into a positive integer expression that is the
maximum number of bytes in a multibyte character in the current locale.
The value is never greater than @code{MB_LEN_MAX}.  Unlike
@code{MB_LEN_MAX} this macro need not be a compile-time constant, and in 
the GNU C library it is not.

@pindex stdlib.h
@code{MB_CUR_MAX} is defined in @file{stdlib.h}.
@end deftypevr

Two different macros are necessary since strictly @w{ISO C90} compilers
do not allow variable length array definitions, but still it is desirable
to avoid dynamic allocation.  This incomplete piece of code shows the
problem:

@smallexample
@{
  char buf[MB_LEN_MAX];
  ssize_t len = 0;

  while (! feof (fp))
    @{
      fread (&buf[len], 1, MB_CUR_MAX - len, fp);
      /* @r{... process} buf */
      len -= used;
    @}
@}
@end smallexample

The code in the inner loop is expected to have always enough bytes in
the array @var{buf} to convert one multibyte character.  The array
@var{buf} has to be sized statically since many compilers do not allow a
variable size.  The @code{fread} call makes sure that @code{MB_CUR_MAX} 
bytes are always available in @var{buf}.  Note that it isn't
a problem if @code{MB_CUR_MAX} is not a compile-time constant.


@node Keeping the state
@subsection Representing the state of the conversion

@cindex stateful
In the introduction of this chapter it was said that certain character
sets use a @dfn{stateful} encoding.  That is, the encoded values depend 
in some way on the previous bytes in the text.

Since the conversion functions allow converting a text in more than one
step we must have a way to pass this information from one call of the
functions to another.

@comment wchar.h
@comment ISO
@deftp {Data type} mbstate_t
@cindex shift state
A variable of type @code{mbstate_t} can contain all the information
about the @dfn{shift state} needed from one call to a conversion
function to another.

@pindex wchar.h
@code{mbstate_t} is defined in @file{wchar.h}.  It was introduced in
@w{Amendment 1} to @w{ISO C90}.
@end deftp

To use objects of type @code{mbstate_t} the programmer has to define such 
objects (normally as local variables on the stack) and pass a pointer to 
the object to the conversion functions.  This way the conversion function
can update the object if the current multibyte character set is stateful.

There is no specific function or initializer to put the state object in
any specific state.  The rules are that the object should always
represent the initial state before the first use, and this is achieved by
clearing the whole variable with code such as follows:

@smallexample
@{
  mbstate_t state;
  memset (&state, '\0', sizeof (state));
  /* @r{from now on @var{state} can be used.}  */
  ...
@}
@end smallexample

When using the conversion functions to generate output it is often
necessary to test whether the current state corresponds to the initial
state.  This is necessary, for example, to decide whether to emit
escape sequences to set the state to the initial state at certain
sequence points.  Communication protocols often require this.

@comment wchar.h
@comment ISO
@deftypefun int mbsinit (const mbstate_t *@var{ps})
The @code {mbsinit} function determines whether the state object pointed 
to by @var{ps} is in the initial state.  If @var{ps} is a null pointer or 
the object is in the initial state the return value is nonzero.  Otherwise 
it is zero.

@pindex wchar.h
@code {mbsinit} was introduced in @w{Amendment 1} to @w{ISO C90} and is 
declared in @file{wchar.h}.
@end deftypefun

Code using @code {mbsinit} often looks similar to this:

@c Fix the example to explicitly say how to generate the escape sequence
@c to restore the initial state.
@smallexample
@{
  mbstate_t state;
  memset (&state, '\0', sizeof (state));
  /* @r{Use @var{state}.}  */
  ...
  if (! mbsinit (&state))
    @{
      /* @r{Emit code to return to initial state.}  */
      const wchar_t empty[] = L"";
      const wchar_t *srcp = empty;
      wcsrtombs (outbuf, &srcp, outbuflen, &state);
    @}
  ...
@}
@end smallexample

The code to emit the escape sequence to get back to the initial state is
interesting.  The @code{wcsrtombs} function can be used to determine the
necessary output code (@pxref{Converting Strings}).  Please note that on
GNU systems it is not necessary to perform this extra action for the
conversion from multibyte text to wide character text since the wide
character encoding is not stateful.  But there is nothing mentioned in
any standard that prohibits making @code{wchar_t} using a stateful
encoding.

@node Converting a Character
@subsection Converting Single Characters

The most fundamental of the conversion functions are those dealing with
single characters.  Please note that this does not always mean single
bytes.  But since there is very often a subset of the multibyte
character set that consists of single byte sequences, there are
functions to help with converting bytes.  Frequently, ASCII is a subpart 
of the multibyte character set.  In such a scenario, each ASCII character 
stands for itself, and all other characters have at least a first byte 
that is beyond the range @math{0} to @math{127}.

@comment wchar.h
@comment ISO
@deftypefun wint_t btowc (int @var{c})
The @code{btowc} function (``byte to wide character'') converts a valid
single byte character @var{c} in the initial shift state into the wide
character equivalent using the conversion rules from the currently
selected locale of the @code{LC_CTYPE} category.

If @code{(unsigned char) @var{c}} is no valid single byte multibyte
character or if @var{c} is @code{EOF}, the function returns @code{WEOF}.

Please note the restriction of @var{c} being tested for validity only in
the initial shift state.  No @code{mbstate_t} object is used from
which the state information is taken, and the function also does not use
any static state.

@pindex wchar.h
The @code{btowc} function was introduced in @w{Amendment 1} to @w{ISO C90} 
and is declared in @file{wchar.h}.
@end deftypefun

Despite the limitation that the single byte value always is interpreted
in the initial state this function is actually useful most of the time.
Most characters are either entirely single-byte character sets or they
are extension to ASCII.  But then it is possible to write code like this
(not that this specific example is very useful):

@smallexample
wchar_t *
itow (unsigned long int val)
@{
  static wchar_t buf[30];
  wchar_t *wcp = &buf[29];
  *wcp = L'\0';
  while (val != 0)
    @{
      *--wcp = btowc ('0' + val % 10);
      val /= 10;
    @}
  if (wcp == &buf[29])
    *--wcp = L'0';
  return wcp;
@}
@end smallexample

Why is it necessary to use such a complicated implementation and not
simply cast @code{'0' + val % 10} to a wide character?  The answer is
that there is no guarantee that one can perform this kind of arithmetic
on the character of the character set used for @code{wchar_t}
representation.  In other situations the bytes are not constant at
compile time and so the compiler cannot do the work.  In situations like
this it is necessary @code{btowc}.

@noindent
There also is a function for the conversion in the other direction.

@comment wchar.h
@comment ISO
@deftypefun int wctob (wint_t @var{c})
The @code{wctob} function (``wide character to byte'') takes as the
parameter a valid wide character.  If the multibyte representation for
this character in the initial state is exactly one byte long, the return
value of this function is this character.  Otherwise the return value is
@code{EOF}.

@pindex wchar.h
@code{wctob} was introduced in @w{Amendment 1} to @w{ISO C90} and
is declared in @file{wchar.h}.
@end deftypefun

There are more general functions to convert single character from
multibyte representation to wide characters and vice versa.  These
functions pose no limit on the length of the multibyte representation
and they also do not require it to be in the initial state.

@comment wchar.h
@comment ISO
@deftypefun size_t mbrtowc (wchar_t *restrict @var{pwc}, const char *restrict @var{s}, size_t @var{n}, mbstate_t *restrict @var{ps})
@cindex stateful
The @code{mbrtowc} function (``multibyte restartable to wide
character'') converts the next multibyte character in the string pointed
to by @var{s} into a wide character and stores it in the wide character
string pointed to by @var{pwc}.  The conversion is performed according
to the locale currently selected for the @code{LC_CTYPE} category.  If
the conversion for the character set used in the locale requires a state,
the multibyte string is interpreted in the state represented by the
object pointed to by @var{ps}.  If @var{ps} is a null pointer, a static,
internal state variable used only by the @code{mbrtowc} function is
used.

If the next multibyte character corresponds to the NUL wide character,
the return value of the function is @math{0} and the state object is
afterwards in the initial state.  If the next @var{n} or fewer bytes
form a correct multibyte character, the return value is the number of
bytes starting from @var{s} that form the multibyte character.  The
conversion state is updated according to the bytes consumed in the
conversion.  In both cases the wide character (either the @code{L'\0'}
or the one found in the conversion) is stored in the string pointed to
by @var{pwc} if @var{pwc} is not null.

If the first @var{n} bytes of the multibyte string possibly form a valid
multibyte character but there are more than @var{n} bytes needed to
complete it, the return value of the function is @code{(size_t) -2} and
no value is stored.  Please note that this can happen even if @var{n}
has a value greater than or equal to @code{MB_CUR_MAX} since the input 
might contain redundant shift sequences.

If the first @code{n} bytes of the multibyte string cannot possibly form
a valid multibyte character, no value is stored, the global variable
@code{errno} is set to the value @code{EILSEQ}, and the function returns
@code{(size_t) -1}.  The conversion state is afterwards undefined.

@pindex wchar.h
@code{mbrtowc} was introduced in @w{Amendment 1} to @w{ISO C90} and
is declared in @file{wchar.h}.
@end deftypefun

Use of @code{mbrtowc} is straightforward.  A function that copies a
multibyte string into a wide character string while at the same time
converting all lowercase characters into uppercase could look like this
(this is not the final version, just an example; it has no error
checking, and sometimes leaks memory):

@smallexample
wchar_t *
mbstouwcs (const char *s)
@{
  size_t len = strlen (s);
  wchar_t *result = malloc ((len + 1) * sizeof (wchar_t));
  wchar_t *wcp = result;
  wchar_t tmp[1];
  mbstate_t state;
  size_t nbytes;

  memset (&state, '\0', sizeof (state));
  while ((nbytes = mbrtowc (tmp, s, len, &state)) > 0)
    @{
      if (nbytes >= (size_t) -2)
        /* Invalid input string.  */
        return NULL;
      *result++ = towupper (tmp[0]);
      len -= nbytes;
      s += nbytes;
    @}
  return result;
@}
@end smallexample

The use of @code{mbrtowc} should be clear.  A single wide character is
stored in @code{@var{tmp}[0]}, and the number of consumed bytes is stored
in the variable @var{nbytes}.  If the conversion is successful, the 
uppercase variant of the wide character is stored in the @var{result} 
array and the pointer to the input string and the number of available 
bytes is adjusted.

The only non-obvious thing about @code{mbrtowc} might be the way memory 
is allocated for the result.  The above code uses the fact that there 
can never be more wide characters in the converted results than there are
bytes in the multibyte input string.  This method yields a pessimistic 
guess about the size of the result, and if many wide character strings 
have to be constructed this way or if the strings are long, the extra 
memory required to be allocated because the input string contains 
multibyte characters might be significant.  The allocated memory block can 
be resized to the correct size before returning it, but a better solution 
might be to allocate just the right amount of space for the result right 
away.  Unfortunately there is no function to compute the length of the wide 
character string directly from the multibyte string.  There is, however, a 
function that does part of the work.

@comment wchar.h
@comment ISO
@deftypefun size_t mbrlen (const char *restrict @var{s}, size_t @var{n}, mbstate_t *@var{ps})
The @code{mbrlen} function (``multibyte restartable length'') computes
the number of at most @var{n} bytes starting at @var{s}, which form the
next valid and complete multibyte character.

If the next multibyte character corresponds to the NUL wide character,
the return value is @math{0}.  If the next @var{n} bytes form a valid
multibyte character, the number of bytes belonging to this multibyte
character byte sequence is returned.

If the the first @var{n} bytes possibly form a valid multibyte
character but the character is incomplete, the return value is 
@code{(size_t) -2}.  Otherwise the multibyte character sequence is invalid 
and the return value is @code{(size_t) -1}.

The multibyte sequence is interpreted in the state represented by the
object pointed to by @var{ps}.  If @var{ps} is a null pointer, a state
object local to @code{mbrlen} is used.

@pindex wchar.h
@code{mbrlen} was introduced in @w{Amendment 1} to @w{ISO C90} and
is declared in @file{wchar.h}.
@end deftypefun

The attentive reader now will note that @code{mbrlen} can be implemented 
as

@smallexample
mbrtowc (NULL, s, n, ps != NULL ? ps : &internal)
@end smallexample

This is true and in fact is mentioned in the official specification.
How can this function be used to determine the length of the wide
character string created from a multibyte character string?  It is not
directly usable, but we can define a function @code{mbslen} using it:

@smallexample
size_t
mbslen (const char *s)
@{
  mbstate_t state;
  size_t result = 0;
  size_t nbytes;
  memset (&state, '\0', sizeof (state));
  while ((nbytes = mbrlen (s, MB_LEN_MAX, &state)) > 0)
    @{
      if (nbytes >= (size_t) -2)
        /* @r{Something is wrong.}  */
        return (size_t) -1;
      s += nbytes;
      ++result;
    @}
  return result;
@}
@end smallexample

This function simply calls @code{mbrlen} for each multibyte character
in the string and counts the number of function calls.  Please note that
we here use @code{MB_LEN_MAX} as the size argument in the @code{mbrlen}
call.  This is acceptable since a) this value is larger then the length of 
the longest multibyte character sequence and b) we know that the string 
@var{s} ends with a NUL byte, which cannot be part of any other multibyte 
character sequence but the one representing the NUL wide character.  
Therefore, the @code{mbrlen} function will never read invalid memory.

Now that this function is available (just to make this clear, this
function is @emph{not} part of the GNU C library) we can compute the
number of wide character required to store the converted multibyte
character string @var{s} using

@smallexample
wcs_bytes = (mbslen (s) + 1) * sizeof (wchar_t);
@end smallexample

Please note that the @code{mbslen} function is quite inefficient.  The
implementation of @code{mbstouwcs} with @code{mbslen} would have to 
perform the conversion of the multibyte character input string twice, and 
this conversion might be quite expensive.  So it is necessary to think 
about the consequences of using the easier but imprecise method before 
doing the work twice.

@comment wchar.h
@comment ISO
@deftypefun size_t wcrtomb (char *restrict @var{s}, wchar_t @var{wc}, mbstate_t *restrict @var{ps})
The @code{wcrtomb} function (``wide character restartable to
multibyte'') converts a single wide character into a multibyte string
corresponding to that wide character.

If @var{s} is a null pointer, the function resets the state stored in
the objects pointed to by @var{ps} (or the internal @code{mbstate_t}
object) to the initial state.  This can also be achieved by a call like
this:

@smallexample
wcrtombs (temp_buf, L'\0', ps)
@end smallexample

@noindent
since, if @var{s} is a null pointer, @code{wcrtomb} performs as if it
writes into an internal buffer, which is guaranteed to be large enough.

If @var{wc} is the NUL wide character, @code{wcrtomb} emits, if
necessary, a shift sequence to get the state @var{ps} into the initial
state followed by a single NUL byte, which is stored in the string 
@var{s}.

Otherwise a byte sequence (possibly including shift sequences) is written 
into the string @var{s}.  This only happens if @var{wc} is a valid wide 
character (i.e., it has a multibyte representation in the character set 
selected by locale of the @code{LC_CTYPE} category).  If @var{wc} is no 
valid wide character, nothing is stored in the strings @var{s}, 
@code{errno} is set to @code{EILSEQ}, the conversion state in @var{ps} 
is undefined and the return value is @code{(size_t) -1}.

If no error occurred the function returns the number of bytes stored in
the string @var{s}.  This includes all bytes representing shift
sequences.

One word about the interface of the function: there is no parameter
specifying the length of the array @var{s}.  Instead the function
assumes that there are at least @code{MB_CUR_MAX} bytes available since
this is the maximum length of any byte sequence representing a single
character.  So the caller has to make sure that there is enough space
available, otherwise buffer overruns can occur.

@pindex wchar.h
@code{wcrtomb} was introduced in @w{Amendment 1} to @w{ISO C90} and is
declared in @file{wchar.h}.
@end deftypefun

Using @code{wcrtomb} is as easy as using @code{mbrtowc}.  The following
example appends a wide character string to a multibyte character string.
Again, the code is not really useful (or correct), it is simply here to
demonstrate the use and some problems.

@smallexample
char *
mbscatwcs (char *s, size_t len, const wchar_t *ws)
@{
  mbstate_t state;
  /* @r{Find the end of the existing string.}  */
  char *wp = strchr (s, '\0');
  len -= wp - s;
  memset (&state, '\0', sizeof (state));
  do
    @{
      size_t nbytes;
      if (len < MB_CUR_LEN)
        @{
          /* @r{We cannot guarantee that the next}
             @r{character fits into the buffer, so}
             @r{return an error.}  */
          errno = E2BIG;
          return NULL;
        @}
      nbytes = wcrtomb (wp, *ws, &state);
      if (nbytes == (size_t) -1)
        /* @r{Error in the conversion.}  */
        return NULL;
      len -= nbytes;
      wp += nbytes;
    @}
  while (*ws++ != L'\0');
  return s;
@}
@end smallexample

First the function has to find the end of the string currently in the
array @var{s}.  The @code{strchr} call does this very efficiently since a
requirement for multibyte character representations is that the NUL byte
is never used except to represent itself (and in this context, the end
of the string).

After initializing the state object the loop is entered where the first
task is to make sure there is enough room in the array @var{s}.  We
abort if there are not at least @code{MB_CUR_LEN} bytes available.  This
is not always optimal but we have no other choice.  We might have less
than @code{MB_CUR_LEN} bytes available but the next multibyte character
might also be only one byte long.  At the time the @code{wcrtomb} call
returns it is too late to decide whether the buffer was large enough.  If 
this solution is unsuitable, there is a very slow but more accurate 
solution.

@smallexample
  ...
  if (len < MB_CUR_LEN)
    @{
      mbstate_t temp_state;
      memcpy (&temp_state, &state, sizeof (state));
      if (wcrtomb (NULL, *ws, &temp_state) > len)
        @{
          /* @r{We cannot guarantee that the next}
             @r{character fits into the buffer, so}
             @r{return an error.}  */
          errno = E2BIG;
          return NULL;
        @}
    @}
  ...
@end smallexample

Here we perform the conversion that might overflow the buffer so that 
we are afterwards in the position to make an exact decision about the 
buffer size.  Please note the @code{NULL} argument for the destination 
buffer in the new @code{wcrtomb} call; since we are not interested in the 
converted text at this point, this is a nice way to express this.  The 
most unusual thing about this piece of code certainly is the duplication 
of the conversion state object, but if a change of the state is necessary 
to emit the next multibyte character, we want to have the same shift state 
change performed in the real conversion.  Therefore, we have to preserve 
the initial shift state information.

There are certainly many more and even better solutions to this problem.
This example is only provided for educational purposes.

@node Converting Strings
@subsection Converting Multibyte and Wide Character Strings

The functions described in the previous section only convert a single
character at a time.  Most operations to be performed in real-world
programs include strings and therefore the @w{ISO C} standard also
defines conversions on entire strings.  However, the defined set of
functions is quite limited; therefore, the GNU C library contains a few
extensions that can help in some important situations.

@comment wchar.h
@comment ISO
@deftypefun size_t mbsrtowcs (wchar_t *restrict @var{dst}, const char **restrict @var{src}, size_t @var{len}, mbstate_t *restrict @var{ps})
The @code{mbsrtowcs} function (``multibyte string restartable to wide
character string'') converts an NUL-terminated multibyte character
string at @code{*@var{src}} into an equivalent wide character string,
including the NUL wide character at the end.  The conversion is started
using the state information from the object pointed to by @var{ps} or
from an internal object of @code{mbsrtowcs} if @var{ps} is a null
pointer.  Before returning, the state object is updated to match the state 
after the last converted character.  The state is the initial state if the
terminating NUL byte is reached and converted.

If @var{dst} is not a null pointer, the result is stored in the array
pointed to by @var{dst}; otherwise, the conversion result is not
available since it is stored in an internal buffer.

If @var{len} wide characters are stored in the array @var{dst} before
reaching the end of the input string, the conversion stops and @var{len}
is returned.  If @var{dst} is a null pointer, @var{len} is never checked.

Another reason for a premature return from the function call is if the
input string contains an invalid multibyte sequence.  In this case the
global variable @code{errno} is set to @code{EILSEQ} and the function
returns @code{(size_t) -1}.

@c XXX The ISO C9x draft seems to have a problem here.  It says that PS
@c is not updated if DST is NULL.  This is not said straightforward and
@c none of the other functions is described like this.  It would make sense
@c to define the function this way but I don't think it is meant like this.

In all other cases the function returns the number of wide characters
converted during this call.  If @var{dst} is not null, @code{mbsrtowcs}
stores in the pointer pointed to by @var{src} either a null pointer (if 
the NUL byte in the input string was reached) or the address of the byte
following the last converted multibyte character.

@pindex wchar.h
@code{mbsrtowcs} was introduced in @w{Amendment 1} to @w{ISO C90} and is
declared in @file{wchar.h}.
@end deftypefun

The definition of the @code{mbsrtowcs} function has one important 
limitation.  The requirement that @var{dst} has to be a NUL-terminated 
string provides problems if one wants to convert buffers with text.  A
buffer is normally no collection of NUL-terminated strings but instead a
continuous collection of lines, separated by newline characters.  Now
assume that a function to convert one line from a buffer is needed.  Since
the line is not NUL-terminated, the source pointer cannot directly point
into the unmodified text buffer.  This means, either one inserts the NUL
byte at the appropriate place for the time of the @code{mbsrtowcs}
function call (which is not doable for a read-only buffer or in a
multi-threaded application) or one copies the line in an extra buffer
where it can be terminated by a NUL byte.  Note that it is not in general 
possible to limit the number of characters to convert by setting the 
parameter @var{len} to any specific value.  Since it is not known how 
many bytes each multibyte character sequence is in length, one can only 
guess.

@cindex stateful
There is still a problem with the method of NUL-terminating a line right
after the newline character, which could lead to very strange results.
As said in the description of the @code{mbsrtowcs} function above the
conversion state is guaranteed to be in the initial shift state after
processing the NUL byte at the end of the input string.  But this NUL
byte is not really part of the text (i.e., the conversion state after
the newline in the original text could be something different than the
initial shift state and therefore the first character of the next line
is encoded using this state).  But the state in question is never
accessible to the user since the conversion stops after the NUL byte
(which resets the state).  Most stateful character sets in use today
require that the shift state after a newline be the initial state--but
this is not a strict guarantee.  Therefore, simply NUL-terminating a
piece of a running text is not always an adequate solution and, 
therefore, should never be used in generally used code.

The generic conversion interface (@pxref{Generic Charset Conversion})
does not have this limitation (it simply works on buffers, not
strings), and the GNU C library contains a set of functions that take
additional parameters specifying the maximal number of bytes that are
consumed from the input string.  This way the problem of
@code{mbsrtowcs}'s example above could be solved by determining the line
length and passing this length to the function.

@comment wchar.h
@comment ISO
@deftypefun size_t wcsrtombs (char *restrict @var{dst}, const wchar_t **restrict @var{src}, size_t @var{len}, mbstate_t *restrict @var{ps})
The @code{wcsrtombs} function (``wide character string restartable to
multibyte string'') converts the NUL-terminated wide character string at
@code{*@var{src}} into an equivalent multibyte character string and 
stores the result in the array pointed to by @var{dst}.  The NUL wide
character is also converted.  The conversion starts in the state
described in the object pointed to by @var{ps} or by a state object
locally to @code{wcsrtombs} in case @var{ps} is a null pointer.  If
@var{dst} is a null pointer, the conversion is performed as usual but the
result is not available.  If all characters of the input string were
successfully converted and if @var{dst} is not a null pointer, the 
pointer pointed to by @var{src} gets assigned a null pointer.

If one of the wide characters in the input string has no valid multibyte
character equivalent, the conversion stops early, sets the global
variable @code{errno} to @code{EILSEQ}, and returns @code{(size_t) -1}.

Another reason for a premature stop is if @var{dst} is not a null
pointer and the next converted character would require more than
@var{len} bytes in total to the array @var{dst}.  In this case (and if
@var{dest} is not a null pointer) the pointer pointed to by @var{src} is
assigned a value pointing to the wide character right after the last one
successfully converted.

Except in the case of an encoding error the return value of the 
@code{wcsrtombs} function is the number of bytes in all the multibyte 
character sequences stored in @var{dst}.  Before returning the state in 
the object pointed to by @var{ps} (or the internal object in case 
@var{ps} is a null pointer) is updated to reflect the state after the 
last conversion.  The state is the initial shift state in case the 
terminating NUL wide character was converted.

@pindex wchar.h
The @code{wcsrtombs} function was introduced in @w{Amendment 1} to 
@w{ISO C90} and is declared in @file{wchar.h}.
@end deftypefun

The restriction mentioned above for the @code{mbsrtowcs} function applies
here also.  There is no possibility of directly controlling the number of
input characters.  One has to place the NUL wide character at the correct 
place or control the consumed input indirectly via the available output 
array size (the @var{len} parameter).

@comment wchar.h
@comment GNU
@deftypefun size_t mbsnrtowcs (wchar_t *restrict @var{dst}, const char **restrict @var{src}, size_t @var{nmc}, size_t @var{len}, mbstate_t *restrict @var{ps})
The @code{mbsnrtowcs} function is very similar to the @code{mbsrtowcs}
function.  All the parameters are the same except for @var{nmc}, which is
new.  The return value is the same as for @code{mbsrtowcs}.

This new parameter specifies how many bytes at most can be used from the
multibyte character string.  In other words, the multibyte character 
string @code{*@var{src}} need not be NUL-terminated.  But if a NUL byte 
is found within the @var{nmc} first bytes of the string, the conversion 
stops here.

This function is a GNU extension.  It is meant to work around the
problems mentioned above.  Now it is possible to convert a buffer with
multibyte character text piece for piece without having to care about
inserting NUL bytes and the effect of NUL bytes on the conversion state.
@end deftypefun

A function to convert a multibyte string into a wide character string
and display it could be written like this (this is not a really useful
example):

@smallexample
void
showmbs (const char *src, FILE *fp)
@{
  mbstate_t state;
  int cnt = 0;
  memset (&state, '\0', sizeof (state));
  while (1)
    @{
      wchar_t linebuf[100];
      const char *endp = strchr (src, '\n');
      size_t n;

      /* @r{Exit if there is no more line.}  */
      if (endp == NULL)
        break;

      n = mbsnrtowcs (linebuf, &src, endp - src, 99, &state);
      linebuf[n] = L'\0';
      fprintf (fp, "line %d: \"%S\"\n", linebuf);
    @}
@}
@end smallexample

There is no problem with the state after a call to @code{mbsnrtowcs}.
Since we don't insert characters in the strings that were not in there
right from the beginning and we use @var{state} only for the conversion
of the given buffer, there is no problem with altering the state.

@comment wchar.h
@comment GNU
@deftypefun size_t wcsnrtombs (char *restrict @var{dst}, const wchar_t **restrict @var{src}, size_t @var{nwc}, size_t @var{len}, mbstate_t *restrict @var{ps})
The @code{wcsnrtombs} function implements the conversion from wide
character strings to multibyte character strings.  It is similar to
@code{wcsrtombs} but, just like @code{mbsnrtowcs}, it takes an extra
parameter, which specifies the length of the input string.

No more than @var{nwc} wide characters from the input string
@code{*@var{src}} are converted.  If the input string contains a NUL
wide character in the first @var{nwc} characters, the conversion stops at
this place.

The @code{wcsnrtombs} function is a GNU extension and just like 
@code{mbsnrtowcs} helps in situations where no NUL-terminated input 
strings are available.
@end deftypefun


@node Multibyte Conversion Example
@subsection A Complete Multibyte Conversion Example

The example programs given in the last sections are only brief and do
not contain all the error checking, etc.  Presented here is a complete
and documented example.  It features the @code{mbrtowc} function but it
should be easy to derive versions using the other functions.

@smallexample
int
file_mbsrtowcs (int input, int output)
@{
  /* @r{Note the use of @code{MB_LEN_MAX}.}
     @r{@code{MB_CUR_MAX} cannot portably be used here.}  */
  char buffer[BUFSIZ + MB_LEN_MAX];
  mbstate_t state;
  int filled = 0;
  int eof = 0;

  /* @r{Initialize the state.}  */
  memset (&state, '\0', sizeof (state));

  while (!eof)
    @{
      ssize_t nread;
      ssize_t nwrite;
      char *inp = buffer;
      wchar_t outbuf[BUFSIZ];
      wchar_t *outp = outbuf;

      /* @r{Fill up the buffer from the input file.}  */
      nread = read (input, buffer + filled, BUFSIZ);
      if (nread < 0)
        @{
          perror ("read");
          return 0;
        @}
      /* @r{If we reach end of file, make a note to read no more.} */
      if (nread == 0)
        eof = 1;

      /* @r{@code{filled} is now the number of bytes in @code{buffer}.} */
      filled += nread;

      /* @r{Convert those bytes to wide characters--as many as we can.} */
      while (1)
        @{
          size_t thislen = mbrtowc (outp, inp, filled, &state);
          /* @r{Stop converting at invalid character;}
             @r{this can mean we have read just the first part}
             @r{of a valid character.}  */
          if (thislen == (size_t) -1)
            break;
          /* @r{We want to handle embedded NUL bytes}
             @r{but the return value is 0.  Correct this.}  */
          if (thislen == 0)
            thislen = 1;
          /* @r{Advance past this character.} */
          inp += thislen;
          filled -= thislen;
          ++outp;
        @}

      /* @r{Write the wide characters we just made.}  */
      nwrite = write (output, outbuf,
                      (outp - outbuf) * sizeof (wchar_t));
      if (nwrite < 0)
        @{
          perror ("write");
          return 0;
        @}

      /* @r{See if we have a @emph{real} invalid character.} */
      if ((eof && filled > 0) || filled >= MB_CUR_MAX)
        @{
          error (0, 0, "invalid multibyte character");
          return 0;
        @}

      /* @r{If any characters must be carried forward,}
         @r{put them at the beginning of @code{buffer}.} */
      if (filled > 0)
        memmove (inp, buffer, filled);
    @}

  return 1;
@}
@end smallexample


@node Non-reentrant Conversion
@section Non-reentrant Conversion Function

The functions described in the previous chapter are defined in
@w{Amendment 1} to @w{ISO C90}, but the original @w{ISO C90} standard 
also contained functions for character set conversion.  The reason that 
these original functions are not described first is that they are almost 
entirely useless.

The problem is that all the conversion functions described in the 
original @w{ISO C90} use a local state.  Using a local state implies that 
multiple conversions at the same time (not only when using threads) 
cannot be done, and that you cannot first convert single characters and 
then strings since you cannot tell the conversion functions which state 
to use.

These original functions are therefore usable only in a very limited set 
of situations.  One must complete converting the entire string before
starting a new one, and each string/text must be converted with the same
function (there is no problem with the library itself; it is guaranteed
that no library function changes the state of any of these functions).
@strong{For the above reasons it is highly requested that the functions
described in the previous section be used in place of non-reentrant 
conversion functions.}

@menu
* Non-reentrant Character Conversion::  Non-reentrant Conversion of Single
                                         Characters.
* Non-reentrant String Conversion::     Non-reentrant Conversion of Strings.
* Shift State::                         States in Non-reentrant Functions.
@end menu

@node Non-reentrant Character Conversion
@subsection Non-reentrant Conversion of Single Characters

@comment stdlib.h
@comment ISO
@deftypefun int mbtowc (wchar_t *restrict @var{result}, const char *restrict @var{string}, size_t @var{size})
The @code{mbtowc} (``multibyte to wide character'') function when called
with non-null @var{string} converts the first multibyte character
beginning at @var{string} to its corresponding wide character code.  It
stores the result in @code{*@var{result}}.

@code{mbtowc} never examines more than @var{size} bytes.  (The idea is
to supply for @var{size} the number of bytes of data you have in hand.)

@code{mbtowc} with non-null @var{string} distinguishes three
possibilities: the first @var{size} bytes at @var{string} start with
valid multibyte characters, they start with an invalid byte sequence or
just part of a character, or @var{string} points to an empty string (a
null character).

For a valid multibyte character, @code{mbtowc} converts it to a wide
character and stores that in @code{*@var{result}}, and returns the
number of bytes in that character (always at least @math{1} and never
more than @var{size}).

For an invalid byte sequence, @code{mbtowc} returns @math{-1}.  For an
empty string, it returns @math{0}, also storing @code{'\0'} in
@code{*@var{result}}.

If the multibyte character code uses shift characters, then
@code{mbtowc} maintains and updates a shift state as it scans.  If you
call @code{mbtowc} with a null pointer for @var{string}, that
initializes the shift state to its standard initial value.  It also
returns nonzero if the multibyte character code in use actually has a
shift state.  @xref{Shift State}.
@end deftypefun

@comment stdlib.h
@comment ISO
@deftypefun int wctomb (char *@var{string}, wchar_t @var{wchar})
The @code{wctomb} (``wide character to multibyte'') function converts
the wide character code @var{wchar} to its corresponding multibyte
character sequence, and stores the result in bytes starting at
@var{string}.  At most @code{MB_CUR_MAX} characters are stored.

@code{wctomb} with non-null @var{string} distinguishes three
possibilities for @var{wchar}: a valid wide character code (one that can
be translated to a multibyte character), an invalid code, and 
@code{L'\0'}.

Given a valid code, @code{wctomb} converts it to a multibyte character,
storing the bytes starting at @var{string}.  Then it returns the number
of bytes in that character (always at least @math{1} and never more
than @code{MB_CUR_MAX}).

If @var{wchar} is an invalid wide character code, @code{wctomb} returns
@math{-1}.  If @var{wchar} is @code{L'\0'}, it returns @code{0}, also
storing @code{'\0'} in @code{*@var{string}}.

If the multibyte character code uses shift characters, then
@code{wctomb} maintains and updates a shift state as it scans.  If you
call @code{wctomb} with a null pointer for @var{string}, that
initializes the shift state to its standard initial value.  It also
returns nonzero if the multibyte character code in use actually has a
shift state.  @xref{Shift State}.

Calling this function with a @var{wchar} argument of zero when
@var{string} is not null has the side-effect of reinitializing the
stored shift state @emph{as well as} storing the multibyte character
@code{'\0'} and returning @math{0}.
@end deftypefun

Similar to @code{mbrlen} there is also a non-reentrant function that
computes the length of a multibyte character.  It can be defined in
terms of @code{mbtowc}.

@comment stdlib.h
@comment ISO
@deftypefun int mblen (const char *@var{string}, size_t @var{size})
The @code{mblen} function with a non-null @var{string} argument returns
the number of bytes that make up the multibyte character beginning at
@var{string}, never examining more than @var{size} bytes.  (The idea is
to supply for @var{size} the number of bytes of data you have in hand.)

The return value of @code{mblen} distinguishes three possibilities: the
first @var{size} bytes at @var{string} start with valid multibyte
characters, they start with an invalid byte sequence or just part of a
character, or @var{string} points to an empty string (a null character).

For a valid multibyte character, @code{mblen} returns the number of
bytes in that character (always at least @code{1} and never more than
@var{size}).  For an invalid byte sequence, @code{mblen} returns 
@math{-1}.  For an empty string, it returns @math{0}.

If the multibyte character code uses shift characters, then @code{mblen}
maintains and updates a shift state as it scans.  If you call
@code{mblen} with a null pointer for @var{string}, that initializes the
shift state to its standard initial value.  It also returns a nonzero
value if the multibyte character code in use actually has a shift state.
@xref{Shift State}.

@pindex stdlib.h
The function @code{mblen} is declared in @file{stdlib.h}.
@end deftypefun


@node Non-reentrant String Conversion
@subsection Non-reentrant Conversion of Strings

For convenience the @w{ISO C90} standard also defines functions to 
convert entire strings instead of single characters.  These functions
suffer from the same problems as their reentrant counterparts from
@w{Amendment 1} to @w{ISO C90}; see @ref{Converting Strings}.

@comment stdlib.h
@comment ISO
@deftypefun size_t mbstowcs (wchar_t *@var{wstring}, const char *@var{string}, size_t @var{size})
The @code{mbstowcs} (``multibyte string to wide character string'')
function converts the null-terminated string of multibyte characters
@var{string} to an array of wide character codes, storing not more than
@var{size} wide characters into the array beginning at @var{wstring}.
The terminating null character counts towards the size, so if @var{size}
is less than the actual number of wide characters resulting from
@var{string}, no terminating null character is stored.

The conversion of characters from @var{string} begins in the initial
shift state.

If an invalid multibyte character sequence is found, the @code{mbstowcs} 
function returns a value of @math{-1}.  Otherwise, it returns the number 
of wide characters stored in the array @var{wstring}.  This number does 
not include the terminating null character, which is present if the 
number is less than @var{size}.

Here is an example showing how to convert a string of multibyte
characters, allocating enough space for the result.

@smallexample
wchar_t *
mbstowcs_alloc (const char *string)
@{
  size_t size = strlen (string) + 1;
  wchar_t *buf = xmalloc (size * sizeof (wchar_t));

  size = mbstowcs (buf, string, size);
  if (size == (size_t) -1)
    return NULL;
  buf = xrealloc (buf, (size + 1) * sizeof (wchar_t));
  return buf;
@}
@end smallexample

@end deftypefun

@comment stdlib.h
@comment ISO
@deftypefun size_t wcstombs (char *@var{string}, const wchar_t *@var{wstring}, size_t @var{size})
The @code{wcstombs} (``wide character string to multibyte string'')
function converts the null-terminated wide character array @var{wstring}
into a string containing multibyte characters, storing not more than
@var{size} bytes starting at @var{string}, followed by a terminating
null character if there is room.  The conversion of characters begins in
the initial shift state.

The terminating null character counts towards the size, so if @var{size}
is less than or equal to the number of bytes needed in @var{wstring}, no
terminating null character is stored.

If a code that does not correspond to a valid multibyte character is
found, the @code{wcstombs} function returns a value of @math{-1}. 
Otherwise, the return value is the number of bytes stored in the array 
@var{string}.  This number does not include the terminating null character, 
which is present if the number is less than @var{size}.
@end deftypefun

@node Shift State
@subsection States in Non-reentrant Functions

In some multibyte character codes, the @emph{meaning} of any particular
byte sequence is not fixed; it depends on what other sequences have come
earlier in the same string.  Typically there are just a few sequences that 
can change the meaning of other sequences; these few are called 
@dfn{shift sequences} and we say that they set the @dfn{shift state} for
other sequences that follow.

To illustrate shift state and shift sequences, suppose we decide that
the sequence @code{0200} (just one byte) enters Japanese mode, in which
pairs of bytes in the range from @code{0240} to @code{0377} are single
characters, while @code{0201} enters Latin-1 mode, in which single bytes
in the range from @code{0240} to @code{0377} are characters, and
interpreted according to the ISO Latin-1 character set.  This is a
multibyte code that has two alternative shift states (``Japanese mode''
and ``Latin-1 mode''), and two shift sequences that specify particular
shift states.

When the multibyte character code in use has shift states, then
@code{mblen}, @code{mbtowc}, and @code{wctomb} must maintain and update
the current shift state as they scan the string.  To make this work
properly, you must follow these rules:

@itemize @bullet
@item
Before starting to scan a string, call the function with a null pointer
for the multibyte character address---for example, @code{mblen (NULL,
0)}.  This initializes the shift state to its standard initial value.

@item
Scan the string one character at a time, in order.  Do not ``back up''
and rescan characters already scanned, and do not intersperse the
processing of different strings.
@end itemize

Here is an example of using @code{mblen} following these rules:

@smallexample
void
scan_string (char *s)
@{
  int length = strlen (s);

  /* @r{Initialize shift state.}  */
  mblen (NULL, 0);

  while (1)
    @{
      int thischar = mblen (s, length);
      /* @r{Deal with end of string and invalid characters.}  */
      if (thischar == 0)
        break;
      if (thischar == -1)
        @{
          error ("invalid multibyte character");
          break;
        @}
      /* @r{Advance past this character.}  */
      s += thischar;
      length -= thischar;
    @}
@}
@end smallexample

The functions @code{mblen}, @code{mbtowc} and @code{wctomb} are not
reentrant when using a multibyte code that uses a shift state.  However,
no other library functions call these functions, so you don't have to
worry that the shift state will be changed mysteriously.


@node Generic Charset Conversion
@section Generic Charset Conversion

The conversion functions mentioned so far in this chapter all had in
common that they operate on character sets that are not directly
specified by the functions.  The multibyte encoding used is specified by
the currently selected locale for the @code{LC_CTYPE} category.  The
wide character set is fixed by the implementation (in the case of GNU C
library it is always UCS-4 encoded @w{ISO 10646}.

This has of course several problems when it comes to general character
conversion:

@itemize @bullet
@item
For every conversion where neither the source nor the destination 
character set is the character set of the locale for the @code{LC_CTYPE} 
category, one has to change the @code{LC_CTYPE} locale using 
@code{setlocale}.

Changing the @code{LC_TYPE} locale introduces major problems for the rest 
of the programs since several more functions (e.g., the character 
classification functions, @pxref{Classification of Characters}) use the 
@code{LC_CTYPE} category.

@item
Parallel conversions to and from different character sets are not
possible since the @code{LC_CTYPE} selection is global and shared by all
threads.

@item
If neither the source nor the destination character set is the character
set used for @code{wchar_t} representation, there is at least a two-step
process necessary to convert a text using the functions above.  One would 
have to select the source character set as the multibyte encoding, 
convert the text into a @code{wchar_t} text, select the destination
character set as the multibyte encoding, and convert the wide character
text to the multibyte (@math{=} destination) character set.

Even if this is possible (which is not guaranteed) it is a very tiring
work.  Plus it suffers from the other two raised points even more due to
the steady changing of the locale.
@end itemize

The XPG2 standard defines a completely new set of functions, which has
none of these limitations.  They are not at all coupled to the selected
locales, and they have no constraints on the character sets selected for
source and destination.  Only the set of available conversions limits 
them.  The standard does not specify that any conversion at all must be 
available.  Such availability is a measure of the quality of the 
implementation.

In the following text first the interface to @code{iconv} and then the
conversion function, will be described.  Comparisons with other
implementations will show what obstacles stand in the way of portable
applications.  Finally, the implementation is described in so far as might 
interest the advanced user who wants to extend conversion capabilities.

@menu
* Generic Conversion Interface::    Generic Character Set Conversion Interface.
* iconv Examples::                  A complete @code{iconv} example.
* Other iconv Implementations::     Some Details about other @code{iconv}
                                     Implementations.
* glibc iconv Implementation::      The @code{iconv} Implementation in the GNU C
                                     library.
@end menu

@node Generic Conversion Interface
@subsection Generic Character Set Conversion Interface

This set of functions follows the traditional cycle of using a resource:
open--use--close.  The interface consists of three functions, each of
which implements one step.

Before the interfaces are described it is necessary to introduce a
data type.  Just like other open--use--close interfaces the functions
introduced here work using handles and the @file{iconv.h} header
defines a special type for the handles used.

@comment iconv.h
@comment XPG2
@deftp {Data Type} iconv_t
This data type is an abstract type defined in @file{iconv.h}.  The user
must not assume anything about the definition of this type; it must be
completely opaque.

Objects of this type can get assigned handles for the conversions using
the @code{iconv} functions.  The objects themselves need not be freed, but
the conversions for which the handles stand for have to.
@end deftp

@noindent
The first step is the function to create a handle.

@comment iconv.h
@comment XPG2
@deftypefun iconv_t iconv_open (const char *@var{tocode}, const char *@var{fromcode})
The @code{iconv_open} function has to be used before starting a
conversion.  The two parameters this function takes determine the
source and destination character set for the conversion, and if the
implementation has the possibility to perform such a conversion, the
function returns a handle.

If the wanted conversion is not available, the @code{iconv_open} function 
returns @code{(iconv_t) -1}. In this case the global variable 
@code{errno} can have the following values:

@table @code
@item EMFILE
The process already has @code{OPEN_MAX} file descriptors open.
@item ENFILE
The system limit of open file is reached.
@item ENOMEM
Not enough memory to carry out the operation.
@item EINVAL
The conversion from @var{fromcode} to @var{tocode} is not supported.
@end table

It is not possible to use the same descriptor in different threads to
perform independent conversions.  The data structures associated
with the descriptor include information about the conversion state.
This must not be messed up by using it in different conversions.

An @code{iconv} descriptor is like a file descriptor as for every use a
new descriptor must be created.  The descriptor does not stand for all
of the conversions from @var{fromset} to @var{toset}.

The GNU C library implementation of @code{iconv_open} has one
significant extension to other implementations.  To ease the extension
of the set of available conversions, the implementation allows storing
the necessary files with data and code in an arbitrary number of 
directories.  How this extension must be written will be explained below
(@pxref{glibc iconv Implementation}).  Here it is only important to say
that all directories mentioned in the @code{GCONV_PATH} environment
variable are considered only if they contain a file @file{gconv-modules}.
These directories need not necessarily be created by the system
administrator.  In fact, this extension is introduced to help users
writing and using their own, new conversions.  Of course, this does not 
work for security reasons in SUID binaries; in this case only the system
directory is considered and this normally is 
@file{@var{prefix}/lib/gconv}.  The @code{GCONV_PATH} environment 
variable is examined exactly once at the first call of the 
@code{iconv_open} function.  Later modifications of the variable have no 
effect.

@pindex iconv.h
The @code{iconv_open} function was introduced early in the X/Open 
Portability Guide, @w{version 2}.  It is supported by all commercial 
Unices as it is required for the Unix branding.  However, the quality and 
completeness of the implementation varies widely.  The @code{iconv_open} 
function is declared in @file{iconv.h}.
@end deftypefun

The @code{iconv} implementation can associate large data structure with
the handle returned by @code{iconv_open}.  Therefore, it is crucial to 
free all the resources once all conversions are carried out and the 
conversion is not needed anymore.

@comment iconv.h
@comment XPG2
@deftypefun int iconv_close (iconv_t @var{cd})
The @code{iconv_close} function frees all resources associated with the
handle @var{cd}, which must have been returned by a successful call to
the @code{iconv_open} function.

If the function call was successful the return value is @math{0}.
Otherwise it is @math{-1} and @code{errno} is set appropriately.
Defined error are:

@table @code
@item EBADF
The conversion descriptor is invalid.
@end table

@pindex iconv.h
The @code{iconv_close} function was introduced together with the rest 
of the @code{iconv} functions in XPG2 and is declared in @file{iconv.h}.
@end deftypefun

The standard defines only one actual conversion function.  This has,
therefore, the most general interface: it allows conversion from one
buffer to another.  Conversion from a file to a buffer, vice versa, or
even file to file can be implemented on top of it.

@comment iconv.h
@comment XPG2
@deftypefun size_t iconv (iconv_t @var{cd}, char **@var{inbuf}, size_t *@var{inbytesleft}, char **@var{outbuf}, size_t *@var{outbytesleft})
@cindex stateful
The @code{iconv} function converts the text in the input buffer
according to the rules associated with the descriptor @var{cd} and
stores the result in the output buffer.  It is possible to call the
function for the same text several times in a row since for stateful
character sets the necessary state information is kept in the data
structures associated with the descriptor.

The input buffer is specified by @code{*@var{inbuf}} and it contains
@code{*@var{inbytesleft}} bytes.  The extra indirection is necessary for
communicating the used input back to the caller (see below).  It is
important to note that the buffer pointer is of type @code{char} and the
length is measured in bytes even if the input text is encoded in wide
characters.

The output buffer is specified in a similar way.  @code{*@var{outbuf}}
points to the beginning of the buffer with at least
@code{*@var{outbytesleft}} bytes room for the result.  The buffer
pointer again is of type @code{char} and the length is measured in
bytes.  If @var{outbuf} or @code{*@var{outbuf}} is a null pointer, the
conversion is performed but no output is available.

If @var{inbuf} is a null pointer, the @code{iconv} function performs the
necessary action to put the state of the conversion into the initial
state.  This is obviously a no-op for non-stateful encodings, but if the
encoding has a state, such a function call might put some byte sequences
in the output buffer, which perform the necessary state changes.  The
next call with @var{inbuf} not being a null pointer then simply goes on
from the initial state.  It is important that the programmer never makes
any assumption as to whether the conversion has to deal with states.  
Even if the input and output character sets are not stateful, the 
implementation might still have to keep states.  This is due to the
implementation chosen for the GNU C library as it is described below.
Therefore an @code{iconv} call to reset the state should always be
performed if some protocol requires this for the output text.

The conversion stops for one of three reasons. The first is that all
characters from the input buffer are converted.  This actually can mean
two things: either all bytes from the input buffer are consumed or
there are some bytes at the end of the buffer that possibly can form a
complete character but the input is incomplete.  The second reason for a
stop is that the output buffer is full.  And the third reason is that
the input contains invalid characters.

In all of these cases the buffer pointers after the last successful
conversion, for input and output buffer, are stored in @var{inbuf} and
@var{outbuf}, and the available room in each buffer is stored in
@var{inbytesleft} and @var{outbytesleft}.

Since the character sets selected in the @code{iconv_open} call can be
almost arbitrary, there can be situations where the input buffer contains
valid characters, which have no identical representation in the output
character set.  The behavior in this situation is undefined.  The
@emph{current} behavior of the GNU C library in this situation is to
return with an error immediately.  This certainly is not the most
desirable solution; therefore, future versions will provide better ones,
but they are not yet finished.

If all input from the input buffer is successfully converted and stored
in the output buffer, the function returns the number of non-reversible
conversions performed.  In all other cases the return value is
@code{(size_t) -1} and @code{errno} is set appropriately.  In such cases
the value pointed to by @var{inbytesleft} is nonzero.

@table @code
@item EILSEQ
The conversion stopped because of an invalid byte sequence in the input.
After the call, @code{*@var{inbuf}} points at the first byte of the
invalid byte sequence.

@item E2BIG
The conversion stopped because it ran out of space in the output buffer.

@item EINVAL
The conversion stopped because of an incomplete byte sequence at the end
of the input buffer.

@item EBADF
The @var{cd} argument is invalid.
@end table

@pindex iconv.h
The @code{iconv} function was introduced in the XPG2 standard and is 
declared in the @file{iconv.h} header.
@end deftypefun

The definition of the @code{iconv} function is quite good overall.  It
provides quite flexible functionality.  The only problems lie in the
boundary cases, which are incomplete byte sequences at the end of the
input buffer and invalid input.  A third problem, which is not really
a design problem, is the way conversions are selected.  The standard
does not say anything about the legitimate names, a minimal set of
available conversions.  We will see how this negatively impacts other
implementations, as demonstrated below.

@node iconv Examples
@subsection A complete @code{iconv} example

The example below features a solution for a common problem.  Given that
one knows the internal encoding used by the system for @code{wchar_t}
strings, one often is in the position to read text from a file and store
it in wide character buffers.  One can do this using @code{mbsrtowcs},
but then we run into the problems discussed above.

@smallexample
int
file2wcs (int fd, const char *charset, wchar_t *outbuf, size_t avail)
@{
  char inbuf[BUFSIZ];
  size_t insize = 0;
  char *wrptr = (char *) outbuf;
  int result = 0;
  iconv_t cd;

  cd = iconv_open ("WCHAR_T", charset);
  if (cd == (iconv_t) -1)
    @{
      /* @r{Something went wrong.}  */
      if (errno == EINVAL)
        error (0, 0, "conversion from '%s' to wchar_t not available",
               charset);
      else
        perror ("iconv_open");

      /* @r{Terminate the output string.}  */
      *outbuf = L'\0';

      return -1;
    @}

  while (avail > 0)
    @{
      size_t nread;
      size_t nconv;
      char *inptr = inbuf;

      /* @r{Read more input.}  */
      nread = read (fd, inbuf + insize, sizeof (inbuf) - insize);
      if (nread == 0)
        @{
          /* @r{When we come here the file is completely read.}
             @r{This still could mean there are some unused}
             @r{characters in the @code{inbuf}.  Put them back.}  */
          if (lseek (fd, -insize, SEEK_CUR) == -1)
            result = -1;

          /* @r{Now write out the byte sequence to get into the}
             @r{initial state if this is necessary.}  */
          iconv (cd, NULL, NULL, &wrptr, &avail);

          break;
        @}
      insize += nread;

      /* @r{Do the conversion.}  */
      nconv = iconv (cd, &inptr, &insize, &wrptr, &avail);
      if (nconv == (size_t) -1)
        @{
          /* @r{Not everything went right.  It might only be}
             @r{an unfinished byte sequence at the end of the}
             @r{buffer.  Or it is a real problem.}  */
          if (errno == EINVAL)
            /* @r{This is harmless.  Simply move the unused}
               @r{bytes to the beginning of the buffer so that}
               @r{they can be used in the next round.}  */
            memmove (inbuf, inptr, insize);
          else
            @{
              /* @r{It is a real problem.  Maybe we ran out of}
                 @r{space in the output buffer or we have invalid}
                 @r{input.  In any case back the file pointer to}
                 @r{the position of the last processed byte.}  */
              lseek (fd, -insize, SEEK_CUR);
              result = -1;
              break;
            @}
        @}
    @}

  /* @r{Terminate the output string.}  */
  if (avail >= sizeof (wchar_t))
    *((wchar_t *) wrptr) = L'\0';

  if (iconv_close (cd) != 0)
    perror ("iconv_close");

  return (wchar_t *) wrptr - outbuf;
@}
@end smallexample

@cindex stateful
This example shows the most important aspects of using the @code{iconv}
functions.  It shows how successive calls to @code{iconv} can be used to
convert large amounts of text.  The user does not have to care about
stateful encodings as the functions take care of everything.

An interesting point is the case where @code{iconv} returns an error and
@code{errno} is set to @code{EINVAL}.  This is not really an error in the 
transformation.  It can happen whenever the input character set contains 
byte sequences of more than one byte for some character and texts are not 
processed in one piece.  In this case there is a chance that a multibyte 
sequence is cut.  The caller can then simply read the remainder of the 
takes and feed the offending bytes together with new character from the 
input to @code{iconv} and continue the work.  The internal state kept in 
the descriptor is @emph{not} unspecified after such an event as is the 
case with the conversion functions from the @w{ISO C} standard.

The example also shows the problem of using wide character strings with
@code{iconv}.  As explained in the description of the @code{iconv}
function above, the function always takes a pointer to a @code{char}
array and the available space is measured in bytes.  In the example, the
output buffer is a wide character buffer; therefore, we use a local
variable @var{wrptr} of type @code{char *}, which is used in the
@code{iconv} calls.

This looks rather innocent but can lead to problems on platforms that
have tight restriction on alignment.  Therefore the caller of @code{iconv} 
has to make sure that the pointers passed are suitable for access of 
characters from the appropriate character set.  Since, in the
above case, the input parameter to the function is a @code{wchar_t}
pointer, this is the case (unless the user violates alignment when
computing the parameter).  But in other situations, especially when
writing generic functions where one does not know what type of character
set one uses and, therefore, treats text as a sequence of bytes, it might
become tricky.

@node Other iconv Implementations
@subsection Some Details about other @code{iconv} Implementations

This is not really the place to discuss the @code{iconv} implementation
of other systems but it is necessary to know a bit about them to write
portable programs.  The above mentioned problems with the specification
of the @code{iconv} functions can lead to portability issues.

The first thing to notice is that, due to the large number of character
sets in use, it is certainly not practical to encode the conversions
directly in the C library.  Therefore, the conversion information must
come from files outside the C library.  This is usually done in one or
both of the following ways:

@itemize @bullet
@item
The C library contains a set of generic conversion functions that can
read the needed conversion tables and other information from data files.
These files get loaded when necessary.

This solution is problematic as it requires a great deal of effort to
apply to all character sets (potentially an infinite set).  The 
differences in the structure of the different character sets is so large
that many different variants of the table-processing functions must be
developed.  In addition, the generic nature of these functions make them 
slower than specifically implemented functions.

@item
The C library only contains a framework that can dynamically load
object files and execute the conversion functions contained therein.

This solution provides much more flexibility.  The C library itself
contains only very little code and therefore reduces the general memory
footprint.  Also, with a documented interface between the C library and
the loadable modules it is possible for third parties to extend the set
of available conversion modules.  A drawback of this solution is that
dynamic loading must be available.
@end itemize

Some implementations in commercial Unices implement a mixture of these 
possibilities; the majority implement only the second solution.  Using 
loadable modules moves the code out of the library itself and keeps 
the door open for extensions and improvements, but this design is also
limiting on some platforms since not many platforms support dynamic
loading in statically linked programs.  On platforms without this
capability it is therefore not possible to use this interface in
statically linked programs.  The GNU C library has, on ELF platforms, no
problems with dynamic loading in these situations; therefore, this
point is moot.  The danger is that one gets acquainted with this 
situation and forgets about the restrictions on other systems.

A second thing to know about other @code{iconv} implementations is that
the number of available conversions is often very limited.  Some
implementations provide, in the standard release (not special
international or developer releases), at most 100 to 200 conversion
possibilities.  This does not mean 200 different character sets are
supported; for example, conversions from one character set to a set of 10 
others might count as 10 conversions.  Together with the other direction
this makes 20 conversion possibilities used up by one character set.  One 
can imagine the thin coverage these platform provide.  Some Unix vendors 
even provide only a handful of conversions, which renders them useless for 
almost all uses.

This directly leads to a third and probably the most problematic point.
The way the @code{iconv} conversion functions are implemented on all
known Unix systems and the availability of the conversion functions from
character set @math{@cal{A}} to @math{@cal{B}} and the conversion from
@math{@cal{B}} to @math{@cal{C}} does @emph{not} imply that the
conversion from @math{@cal{A}} to @math{@cal{C}} is available.

This might not seem unreasonable and problematic at first, but it is a
quite big problem as one will notice shortly after hitting it.  To show
the problem we assume to write a program that has to convert from
@math{@cal{A}} to @math{@cal{C}}.  A call like

@smallexample
cd = iconv_open ("@math{@cal{C}}", "@math{@cal{A}}");
@end smallexample

@noindent
fails according to the assumption above.  But what does the program
do now?  The conversion is necessary; therefore, simply giving up is not
an option.

This is a nuisance.  The @code{iconv} function should take care of this.
But how should the program proceed from here on?  If it tries to convert 
to character set @math{@cal{B}}, first the two @code{iconv_open}
calls

@smallexample
cd1 = iconv_open ("@math{@cal{B}}", "@math{@cal{A}}");
@end smallexample

@noindent
and

@smallexample
cd2 = iconv_open ("@math{@cal{C}}", "@math{@cal{B}}");
@end smallexample

@noindent
will succeed, but how to find @math{@cal{B}}?

Unfortunately, the answer is: there is no general solution.  On some
systems guessing might help.  On those systems most character sets can
convert to and from UTF-8 encoded @w{ISO 10646} or Unicode text. Beside 
this only some very system-specific methods can help.  Since the 
conversion functions come from loadable modules and these modules must
be stored somewhere in the filesystem, one @emph{could} try to find them
and determine from the available file which conversions are available
and whether there is an indirect route from @math{@cal{A}} to
@math{@cal{C}}.

This example shows one of the design errors of @code{iconv} mentioned 
above.  It should at least be possible to determine the list of available
conversion programmatically so that if @code{iconv_open} says there is no 
such conversion, one could make sure this also is true for indirect
routes.

@node glibc iconv Implementation
@subsection The @code{iconv} Implementation in the GNU C library

After reading about the problems of @code{iconv} implementations in the
last section it is certainly good to note that the implementation in
the GNU C library has none of the problems mentioned above.  What
follows is a step-by-step analysis of the points raised above.  The
evaluation is based on the current state of the development (as of
January 1999).  The development of the @code{iconv} functions is not
complete, but basic functionality has solidified.

The GNU C library's @code{iconv} implementation uses shared loadable
modules to implement the conversions.  A very small number of
conversions are built into the library itself but these are only rather
trivial conversions.

All the benefits of loadable modules are available in the GNU C library
implementation.  This is especially appealing since the interface is
well documented (see below), and it, therefore, is easy to write new
conversion modules.  The drawback of using loadable objects is not a
problem in the GNU C library, at least on ELF systems.  Since the
library is able to load shared objects even in statically linked
binaries, static linking need not be forbidden in case one wants to use 
@code{iconv}.

The second mentioned problem is the number of supported conversions.
Currently, the GNU C library supports more than 150 character sets.  The
way the implementation is designed the number of supported conversions
is greater than 22350 (@math{150} times @math{149}).  If any conversion
from or to a character set is missing, it can be added easily.

Particularly impressive as it may be, this high number is due to the
fact that the GNU C library implementation of @code{iconv} does not have
the third problem mentioned above (i.e., whenever there is a conversion
from a character set @math{@cal{A}} to @math{@cal{B}} and from
@math{@cal{B}} to @math{@cal{C}} it is always possible to convert from
@math{@cal{A}} to @math{@cal{C}} directly).  If the @code{iconv_open}
returns an error and sets @code{errno} to @code{EINVAL}, there is no 
known way, directly or indirectly, to perform the wanted conversion.

@cindex triangulation
Triangulation is achieved by providing for each character set a 
conversion from and to UCS-4 encoded @w{ISO 10646}.  Using @w{ISO 10646} 
as an intermediate representation it is possible to @dfn{triangulate}
(i.e., convert with an intermediate representation).

There is no inherent requirement to provide a conversion to @w{ISO
10646} for a new character set, and it is also possible to provide other
conversions where neither source nor destination character set is @w{ISO
10646}.  The existing set of conversions is simply meant to cover all 
conversions that might be of interest.

@cindex ISO-2022-JP
@cindex EUC-JP
All currently available conversions use the triangulation method above,
making conversion run unnecessarily slow.  If, for example, somebody 
often needs the conversion from ISO-2022-JP to EUC-JP, a quicker solution
would involve direct conversion between the two character sets, skipping
the input to @w{ISO 10646} first.  The two character sets of interest
are much more similar to each other than to @w{ISO 10646}.

In such a situation one easily can write a new conversion and provide it
as a better alternative.  The GNU C library @code{iconv} implementation
would automatically use the module implementing the conversion if it is
specified to be more efficient.

@subsubsection Format of @file{gconv-modules} files

All information about the available conversions comes from a file named
@file{gconv-modules}, which can be found in any of the directories along
the @code{GCONV_PATH}.  The @file{gconv-modules} files are line-oriented
text files, where each of the lines has one of the following formats:

@itemize @bullet
@item
If the first non-whitespace character is a @kbd{#} the line contains only 
comments and is ignored.

@item
Lines starting with @code{alias} define an alias name for a character 
set.  Two more words are expected on the line.  The first word 
defines the alias name, and the second defines the original name of the
character set.  The effect is that it is possible to use the alias name
in the @var{fromset} or @var{toset} parameters of @code{iconv_open} and
achieve the same result as when using the real character set name.

This is quite important as a character set has often many different
names.  There is normally an official name but this need not correspond to 
the most popular name.  Beside this many character sets have special 
names that are somehow constructed.  For example, all character sets 
specified by the ISO have an alias of the form @code{ISO-IR-@var{nnn}} 
where @var{nnn} is the registration number.  This allows programs that 
know about the registration number to construct character set names and 
use them in @code{iconv_open} calls.  More on the available names and 
aliases follows below.

@item
Lines starting with @code{module} introduce an available conversion
module.  These lines must contain three or four more words.

The first word specifies the source character set, the second word the
destination character set of conversion implemented in this module, and 
the third word is the name of the loadable module.  The filename is
constructed by appending the usual shared object suffix (normally
@file{.so}) and this file is then supposed to be found in the same
directory the @file{gconv-modules} file is in.  The last word on the line, 
which is optional, is a numeric value representing the cost of the
conversion.  If this word is missing, a cost of @math{1} is assumed.  The
numeric value itself does not matter that much; what counts are the
relative values of the sums of costs for all possible conversion paths.
Below is a more precise description of the use of the cost value.
@end itemize

Returning to the example above where one has written a module to directly
convert from ISO-2022-JP to EUC-JP and back.  All that has to be done is
to put the new module, let its name be ISO2022JP-EUCJP.so, in a directory
and add a file @file{gconv-modules} with the following content in the
same directory:

@smallexample
module  ISO-2022-JP//   EUC-JP//        ISO2022JP-EUCJP    1
module  EUC-JP//        ISO-2022-JP//   ISO2022JP-EUCJP    1
@end smallexample

To see why this is sufficient, it is necessary to understand how the
conversion used by @code{iconv} (and described in the descriptor) is
selected.  The approach to this problem is quite simple.

At the first call of the @code{iconv_open} function the program reads
all available @file{gconv-modules} files and builds up two tables: one
containing all the known aliases and another that contains the
information about the conversions and which shared object implements
them.

@subsubsection Finding the conversion path in @code{iconv}

The set of available conversions form a directed graph with weighted
edges.  The weights on the edges are the costs specified in the
@file{gconv-modules} files.  The @code{iconv_open} function uses an
algorithm suitable for search for the best path in such a graph and so
constructs a list of conversions that must be performed in succession
to get the transformation from the source to the destination character
set.

Explaining why the above @file{gconv-modules} files allows the
@code{iconv} implementation to resolve the specific ISO-2022-JP to
EUC-JP conversion module instead of the conversion coming with the
library itself is straightforward.  Since the latter conversion takes two
steps (from ISO-2022-JP to @w{ISO 10646} and then from @w{ISO 10646} to
EUC-JP), the cost is @math{1+1 = 2}.  The above @file{gconv-modules}
file, however, specifies that the new conversion modules can perform this
conversion with only the cost of @math{1}.

A mysterious item about the @file{gconv-modules} file above (and also
the file coming with the GNU C library) are the names of the character
sets specified in the @code{module} lines.  Why do almost all the names
end in @code{//}?  And this is not all: the names can actually be
regular expressions.  At this point in time this mystery should not be
revealed, unless you have the relevant spell-casting materials: ashes
from an original @w{DOS 6.2} boot disk burnt in effigy, a crucifix
blessed by St.@: Emacs, assorted herbal roots from Central America, sand
from Cebu, etc.  Sorry!  @strong{The part of the implementation where
this is used is not yet finished.  For now please simply follow the
existing examples.  It'll become clearer once it is. --drepper}

A last remark about the @file{gconv-modules} is about the names not
ending with @code{//}.  A character set named @code{INTERNAL} is often 
mentioned.  From the discussion above and the chosen name it should have 
become clear that this is the name for the representation used in the 
intermediate step of the triangulation.  We have said that this is UCS-4 
but actually that is not quite right.  The UCS-4 specification also 
includes the specification of the byte ordering used.  Since a UCS-4 value 
consists of four bytes, a stored value is effected by byte ordering.  The 
internal representation is @emph{not} the same as UCS-4 in case the byte 
ordering of the processor (or at least the running process) is not the 
same as the one required for UCS-4.  This is done for performance reasons 
as one does not want to perform unnecessary byte-swapping operations if 
one is not interested in actually seeing the result in UCS-4.  To avoid 
trouble with endianess, the internal representation consistently is named 
@code{INTERNAL} even on big-endian systems where the representations are 
identical.

@subsubsection @code{iconv} module data structures

So far this section has described how modules are located and considered 
to be used.  What remains to be described is the interface of the modules
so that one can write new ones. This section describes the interface as
it is in use in January 1999.  The interface will change a bit in the 
future but, with luck, only in an upwardly compatible way.

The definitions necessary to write new modules are publicly available
in the non-standard header @file{gconv.h}.  The following text,
therefore, describes the definitions from this header file.  First, 
however, it is necessary to get an overview.

From the perspective of the user of @code{iconv} the interface is quite
simple: the @code{iconv_open} function returns a handle that can be used 
in calls to @code{iconv}, and finally the handle is freed with a call to 
@code{iconv_close}.  The problem is that the handle has to be able to
represent the possibly long sequences of conversion steps and also the
state of each conversion since the handle is all that is passed to the
@code{iconv} function.  Therefore, the data structures are really the
elements necessary to understanding the implementation.

We need two different kinds of data structures.  The first describes the
conversion and the second describes the state etc.  There are really two
type definitions like this in @file{gconv.h}.
@pindex gconv.h

@comment gconv.h
@comment GNU
@deftp {Data type} {struct __gconv_step}
This data structure describes one conversion a module can perform.  For
each function in a loaded module with conversion functions there is
exactly one object of this type.  This object is shared by all users of
the conversion (i.e., this object does not contain any information
corresponding to an actual conversion; it only describes the conversion
itself).

@table @code
@item struct __gconv_loaded_object *__shlib_handle
@itemx const char *__modname
@itemx int __counter
All these elements of the structure are used internally in the C library
to coordinate loading and unloading the shared.  One must not expect any
of the other elements to be available or initialized.

@item const char *__from_name
@itemx const char *__to_name
@code{__from_name} and @code{__to_name} contain the names of the source and
destination character sets.  They can be used to identify the actual
conversion to be carried out since one module might implement conversions 
for more than one character set and/or direction.

@item gconv_fct __fct
@itemx gconv_init_fct __init_fct
@itemx gconv_end_fct __end_fct
These elements contain pointers to the functions in the loadable module.
The interface will be explained below.

@item int __min_needed_from
@itemx int __max_needed_from
@itemx int __min_needed_to
@itemx int __max_needed_to;
These values have to be supplied in the init function of the module.  The
@code{__min_needed_from} value specifies how many bytes a character of
the source character set at least needs.  The @code{__max_needed_from}
specifies the maximum value that also includes possible shift sequences.

The @code{__min_needed_to} and @code{__max_needed_to} values serve the
same purpose as @code{__min_needed_from} and @code{__max_needed_from} but 
this time for the destination character set.

It is crucial that these values be accurate since otherwise the
conversion functions will have problems or not work at all.

@item int __stateful
This element must also be initialized by the init function. 
@code{int __stateful} is nonzero if the source character set is stateful. 
Otherwise it is zero.

@item void *__data
This element can be used freely by the conversion functions in the
module.  @code{void *__data} can be used to communicate extra information 
from one call to another.  @code{void *__data} need not be initialized if 
not needed at all.  If @code{void *__data} element is assigned a pointer 
to dynamically allocated memory (presumably in the init function) it has 
to be made sure that the end function deallocates the memory.  Otherwise 
the application will leak memory.

It is important to be aware that this data structure is shared by all
users of this specification conversion and therefore the @code{__data}
element must not contain data specific to one specific use of the
conversion function.
@end table
@end deftp

@comment gconv.h
@comment GNU
@deftp {Data type} {struct __gconv_step_data}
This is the data structure that contains the information specific to
each use of the conversion functions.


@table @code
@item char *__outbuf
@itemx char *__outbufend
These elements specify the output buffer for the conversion step.  The
@code{__outbuf} element points to the beginning of the buffer, and
@code{__outbufend} points to the byte following the last byte in the
buffer.  The conversion function must not assume anything about the size
of the buffer but it can be safely assumed the there is room for at
least one complete character in the output buffer.

Once the conversion is finished, if the conversion is the last step, the
@code{__outbuf} element must be modified to point after the last byte
written into the buffer to signal how much output is available.  If this
conversion step is not the last one, the element must not be modified.
The @code{__outbufend} element must not be modified.

@item int __is_last
This element is nonzero if this conversion step is the last one.  This
information is necessary for the recursion.  See the description of the
conversion function internals below.  This element must never be
modified.

@item int __invocation_counter
The conversion function can use this element to see how many calls of 
the conversion function already happened.  Some character sets require a 
certain prolog when generating output, and by comparing this value with
zero, one can find out whether it is the first call and whether, 
therefore, the prolog should be emitted.  This element must never be 
modified.

@item int __internal_use
This element is another one rarely used but needed in certain
situations.  It is assigned a nonzero value in case the conversion
functions are used to implement @code{mbsrtowcs} et.al.@: (i.e., the
function is not used directly through the @code{iconv} interface).

This sometimes makes a difference as it is expected that the
@code{iconv} functions are used to translate entire texts while the
@code{mbsrtowcs} functions are normally used only to convert single
strings and might be used multiple times to convert entire texts.

But in this situation we would have problem complying with some rules of
the character set specification.  Some character sets require a prolog,
which must appear exactly once for an entire text.  If a number of
@code{mbsrtowcs} calls are used to convert the text, only the first call
must add the prolog.  However, because there is no communication between the
different calls of @code{mbsrtowcs}, the conversion functions have no
possibility to find this out.  The situation is different for sequences
of @code{iconv} calls since the handle allows access to the needed
information.

The @code{int __internal_use} element is mostly used together with 
@code{__invocation_counter} as follows:

@smallexample
if (!data->__internal_use
     && data->__invocation_counter == 0)
  /* @r{Emit prolog.}  */
  ...
@end smallexample

This element must never be modified.

@item mbstate_t *__statep
The @code{__statep} element points to an object of type @code{mbstate_t}
(@pxref{Keeping the state}).  The conversion of a stateful character
set must use the object pointed to by @code{__statep} to store 
information about the conversion state.  The @code{__statep} element 
itself must never be modified.

@item mbstate_t __state
This element must @emph{never} be used directly.  It is only part of
this structure to have the needed space allocated.
@end table
@end deftp

@subsubsection @code{iconv} module interfaces

With the knowledge about the data structures we now can describe the
conversion function itself.  To understand the interface a bit of
knowledge is necessary about the functionality in the C library that 
loads the objects with the conversions.

It is often the case that one conversion is used more than once (i.e.,
there are several @code{iconv_open} calls for the same set of character
sets during one program run).  The @code{mbsrtowcs} et.al.@: functions in
the GNU C library also use the @code{iconv} functionality, which 
increases the number of uses of the same functions even more.

Because of this multiple use of conversions, the modules do not get 
loaded exclusively for one conversion.  Instead a module once loaded can 
be used by an arbitrary number of @code{iconv} or @code{mbsrtowcs} calls 
at the same time.  The splitting of the information between conversion-
function-specific information and conversion data makes this possible. 
The last section showed the two data structures used to do this.

This is of course also reflected in the interface and semantics of the
functions that the modules must provide.  There are three functions that
must have the following names:

@table @code
@item gconv_init
The @code{gconv_init} function initializes the conversion function
specific data structure.  This very same object is shared by all
conversions that use this conversion and, therefore, no state information
about the conversion itself must be stored in here.  If a module 
implements more than one conversion, the @code{gconv_init} function will 
be called multiple times.

@item gconv_end
The @code{gconv_end} function is responsible for freeing all resources
allocated by the @code{gconv_init} function.  If there is nothing to do,
this function can be missing.  Special care must be taken if the module
implements more than one conversion and the @code{gconv_init} function
does not allocate the same resources for all conversions.

@item gconv
This is the actual conversion function.  It is called to convert one
block of text.  It gets passed the conversion step information
initialized by @code{gconv_init} and the conversion data, specific to
this use of the conversion functions.
@end table

There are three data types defined for the three module interface
functions and these define the interface.

@comment gconv.h
@comment GNU
@deftypevr {Data type} int {(*__gconv_init_fct)} (struct __gconv_step *)
This specifies the interface of the initialization function of the
module.  It is called exactly once for each conversion the module
implements.

As explained in the description of the @code{struct __gconv_step} data
structure above the initialization function has to initialize parts of
it.

@table @code
@item __min_needed_from
@itemx __max_needed_from
@itemx __min_needed_to
@itemx __max_needed_to
These elements must be initialized to the exact numbers of the minimum
and maximum number of bytes used by one character in the source and
destination character sets, respectively.  If the characters all have the
same size, the minimum and maximum values are the same.

@item __stateful
This element must be initialized to an nonzero value if the source
character set is stateful.  Otherwise it must be zero.
@end table

If the initialization function needs to communicate some information
to the conversion function, this communication can happen using the 
@code{__data} element of the @code{__gconv_step} structure.  But since 
this data is shared by all the conversions, it must not be modified by 
the conversion function.  The example below shows how this can be used.

@smallexample
#define MIN_NEEDED_FROM         1
#define MAX_NEEDED_FROM         4
#define MIN_NEEDED_TO           4
#define MAX_NEEDED_TO           4

int
gconv_init (struct __gconv_step *step)
@{
  /* @r{Determine which direction.}  */
  struct iso2022jp_data *new_data;
  enum direction dir = illegal_dir;
  enum variant var = illegal_var;
  int result;

  if (__strcasecmp (step->__from_name, "ISO-2022-JP//") == 0)
    @{
      dir = from_iso2022jp;
      var = iso2022jp;
    @}
  else if (__strcasecmp (step->__to_name, "ISO-2022-JP//") == 0)
    @{
      dir = to_iso2022jp;
      var = iso2022jp;
    @}
  else if (__strcasecmp (step->__from_name, "ISO-2022-JP-2//") == 0)
    @{
      dir = from_iso2022jp;
      var = iso2022jp2;
    @}
  else if (__strcasecmp (step->__to_name, "ISO-2022-JP-2//") == 0)
    @{
      dir = to_iso2022jp;
      var = iso2022jp2;
    @}

  result = __GCONV_NOCONV;
  if (dir != illegal_dir)
    @{
      new_data = (struct iso2022jp_data *)
        malloc (sizeof (struct iso2022jp_data));

      result = __GCONV_NOMEM;
      if (new_data != NULL)
        @{
          new_data->dir = dir;
          new_data->var = var;
          step->__data = new_data;

          if (dir == from_iso2022jp)
            @{
              step->__min_needed_from = MIN_NEEDED_FROM;
              step->__max_needed_from = MAX_NEEDED_FROM;
              step->__min_needed_to = MIN_NEEDED_TO;
              step->__max_needed_to = MAX_NEEDED_TO;
            @}
          else
            @{
              step->__min_needed_from = MIN_NEEDED_TO;
              step->__max_needed_from = MAX_NEEDED_TO;
              step->__min_needed_to = MIN_NEEDED_FROM;
              step->__max_needed_to = MAX_NEEDED_FROM + 2;
            @}

          /* @r{Yes, this is a stateful encoding.}  */
          step->__stateful = 1;

          result = __GCONV_OK;
        @}
    @}

  return result;
@}
@end smallexample

The function first checks which conversion is wanted.  The module from
which this function is taken implements four different conversions; 
which one is selected can be determined by comparing the names.  The
comparison should always be done without paying attention to the case.

Next, a data structure, which contains the necessary information about 
which conversion is selected, is allocated.  The data structure
@code{struct iso2022jp_data} is locally defined since, outside the 
module, this data is not used at all.  Please note that if all four 
conversions this modules supports are requested there are four data 
blocks.

One interesting thing is the initialization of the @code{__min_} and
@code{__max_} elements of the step data object.  A single ISO-2022-JP
character can consist of one to four bytes.  Therefore the
@code{MIN_NEEDED_FROM} and @code{MAX_NEEDED_FROM} macros are defined
this way.  The output is always the @code{INTERNAL} character set (aka
UCS-4) and therefore each character consists of exactly four bytes.  For
the conversion from @code{INTERNAL} to ISO-2022-JP we have to take into
account that escape sequences might be necessary to switch the character
sets.  Therefore the @code{__max_needed_to} element for this direction
gets assigned @code{MAX_NEEDED_FROM + 2}.  This takes into account the
two bytes needed for the escape sequences to single the switching.  The
asymmetry in the maximum values for the two directions can be explained
easily: when reading ISO-2022-JP text, escape sequences can be handled
alone (i.e., it is not necessary to process a real character since the
effect of the escape sequence can be recorded in the state information).
The situation is different for the other direction.  Since it is in
general not known which character comes next, one cannot emit escape
sequences to change the state in advance.  This means the escape
sequences that have to be emitted together with the next character.
Therefore one needs more room than only for the character itself.

The possible return values of the initialization function are:

@table @code
@item __GCONV_OK
The initialization succeeded
@item __GCONV_NOCONV
The requested conversion is not supported in the module.  This can
happen if the @file{gconv-modules} file has errors.
@item __GCONV_NOMEM
Memory required to store additional information could not be allocated.
@end table
@end deftypevr

The function called before the module is unloaded is significantly
easier.  It often has nothing at all to do; in which case it can be left
out completely.

@comment gconv.h
@comment GNU
@deftypevr {Data type} void {(*__gconv_end_fct)} (struct gconv_step *)
The task of this function is to free all resources allocated in the
initialization function.  Therefore only the @code{__data} element of
the object pointed to by the argument is of interest.  Continuing the
example from the initialization function, the finalization function
looks like this:

@smallexample
void
gconv_end (struct __gconv_step *data)
@{
  free (data->__data);
@}
@end smallexample
@end deftypevr

The most important function is the conversion function itself, which can
get quite complicated for complex character sets.  But since this is not
of interest here, we will only describe a possible skeleton for the
conversion function.

@comment gconv.h
@comment GNU
@deftypevr {Data type} int {(*__gconv_fct)} (struct __gconv_step *, struct __gconv_step_data *, const char **, const char *, size_t *, int)
The conversion function can be called for two basic reason: to convert
text or to reset the state.  From the description of the @code{iconv}
function it can be seen why the flushing mode is necessary.  What mode
is selected is determined by the sixth argument, an integer.  This 
argument being nonzero means that flushing is selected.

Common to both modes is where the output buffer can be found.  The
information about this buffer is stored in the conversion step data.  A
pointer to this information is passed as the second argument to this 
function.  The description of the @code{struct __gconv_step_data} 
structure has more information on the conversion step data.

@cindex stateful
What has to be done for flushing depends on the source character set.
If the source character set is not stateful, nothing has to be done. 
Otherwise the function has to emit a byte sequence to bring the state 
object into the initial state.  Once this all happened the other 
conversion modules in the chain of conversions have to get the same 
chance.  Whether another step follows can be determined from the 
@code{__is_last} element of the step data structure to which the first 
parameter points.

The more interesting mode is when actual text has to be converted.  The 
first step in this case is to convert as much text as possible from the 
input buffer and store the result in the output buffer.  The start of the 
input buffer is determined by the third argument, which is a pointer to a 
pointer variable referencing the beginning of the buffer.  The fourth 
argument is a pointer to the byte right after the last byte in the buffer.

The conversion has to be performed according to the current state if the
character set is stateful.  The state is stored in an object pointed to
by the @code{__statep} element of the step data (second argument).  Once
either the input buffer is empty or the output buffer is full the
conversion stops.  At this point, the pointer variable referenced by the
third parameter must point to the byte following the last processed
byte (i.e., if all of the input is consumed, this pointer and the fourth
parameter have the same value).

What now happens depends on whether this step is the last one.  If it is 
the last step, the only thing that has to be done is to update the 
@code{__outbuf} element of the step data structure to point after the
last written byte.  This update gives the caller the information on how 
much text is available in the output buffer.  In addition, the variable
pointed to by the fifth parameter, which is of type @code{size_t}, must
be incremented by the number of characters (@emph{not bytes}) that were
converted in a non-reversible way.  Then, the function can return.

In case the step is not the last one, the later conversion functions have
to get a chance to do their work.  Therefore, the appropriate conversion
function has to be called.  The information about the functions is
stored in the conversion data structures, passed as the first parameter.
This information and the step data are stored in arrays, so the next
element in both cases can be found by simple pointer arithmetic:

@smallexample
int
gconv (struct __gconv_step *step, struct __gconv_step_data *data,
       const char **inbuf, const char *inbufend, size_t *written,
       int do_flush)
@{
  struct __gconv_step *next_step = step + 1;
  struct __gconv_step_data *next_data = data + 1;
  ...
@end smallexample

The @code{next_step} pointer references the next step information and
@code{next_data} the next data record.  The call of the next function
therefore will look similar to this:

@smallexample
  next_step->__fct (next_step, next_data, &outerr, outbuf,
                    written, 0)
@end smallexample

But this is not yet all.  Once the function call returns the conversion
function might have some more to do.  If the return value of the function 
is @code{__GCONV_EMPTY_INPUT}, more room is available in the output 
buffer.  Unless the input buffer is empty the conversion, functions start 
all over again and process the rest of the input buffer.  If the return 
value is not @code{__GCONV_EMPTY_INPUT}, something went wrong and we have 
to recover from this.

A requirement for the conversion function is that the input buffer
pointer (the third argument) always point to the last character that
was put in converted form into the output buffer.  This is trivially
true after the conversion performed in the current step, but if the
conversion functions deeper downstream stop prematurely, not all
characters from the output buffer are consumed and, therefore, the input
buffer pointers must be backed off to the right position.

Correcting the input buffers is easy to do if the input and output 
character sets have a fixed width for all characters.  In this situation 
we can compute how many characters are left in the output buffer and, 
therefore, can correct the input buffer pointer appropriately with a 
similar computation.  Things are getting tricky if either character set 
has characters represented with variable length byte sequences, and it 
gets even more complicated if the conversion has to take care of the 
state.  In these cases the conversion has to be performed once again, from 
the known state before the initial conversion (i.e., if necessary the 
state of the conversion has to be reset and the conversion loop has to be 
executed again).  The difference now is that it is known how much input 
must be created, and the conversion can stop before converting the first 
unused character.  Once this is done the input buffer pointers must be 
updated again and the function can return.

One final thing should be mentioned.  If it is necessary for the
conversion to know whether it is the first invocation (in case a prolog
has to be emitted), the conversion function should increment the 
@code{__invocation_counter} element of the step data structure just 
before returning to the caller.  See the description of the @code{struct
__gconv_step_data} structure above for more information on how this can
be used.

The return value must be one of the following values:

@table @code
@item __GCONV_EMPTY_INPUT
All input was consumed and there is room left in the output buffer.
@item __GCONV_FULL_OUTPUT
No more room in the output buffer.  In case this is not the last step
this value is propagated down from the call of the next conversion
function in the chain. 
@item __GCONV_INCOMPLETE_INPUT
The input buffer is not entirely empty since it contains an incomplete
character sequence.
@end table

The following example provides a framework for a conversion function.
In case a new conversion has to be written the holes in this
implementation have to be filled and that is it.

@smallexample
int
gconv (struct __gconv_step *step, struct __gconv_step_data *data,
       const char **inbuf, const char *inbufend, size_t *written,
       int do_flush)
@{
  struct __gconv_step *next_step = step + 1;
  struct __gconv_step_data *next_data = data + 1;
  gconv_fct fct = next_step->__fct;
  int status;

  /* @r{If the function is called with no input this means we have}
     @r{to reset to the initial state.  The possibly partly}
     @r{converted input is dropped.}  */
  if (do_flush)
    @{
      status = __GCONV_OK;

      /* @r{Possible emit a byte sequence which put the state object}
         @r{into the initial state.}  */

      /* @r{Call the steps down the chain if there are any but only}
         @r{if we successfully emitted the escape sequence.}  */
      if (status == __GCONV_OK && ! data->__is_last)
        status = fct (next_step, next_data, NULL, NULL,
                      written, 1);
    @}
  else
    @{
      /* @r{We preserve the initial values of the pointer variables.}  */
      const char *inptr = *inbuf;
      char *outbuf = data->__outbuf;
      char *outend = data->__outbufend;
      char *outptr;

      do
        @{
          /* @r{Remember the start value for this round.}  */
          inptr = *inbuf;
          /* @r{The outbuf buffer is empty.}  */
          outptr = outbuf;

          /* @r{For stateful encodings the state must be safe here.}  */

          /* @r{Run the conversion loop.  @code{status} is set}
             @r{appropriately afterwards.}  */

          /* @r{If this is the last step, leave the loop. There is}
             @r{nothing we can do.}  */
          if (data->__is_last)
            @{
              /* @r{Store information about how many bytes are}
                 @r{available.}  */
              data->__outbuf = outbuf;

             /* @r{If any non-reversible conversions were performed,}
                @r{add the number to @code{*written}.}  */

             break;
           @}

          /* @r{Write out all output that was produced.}  */
          if (outbuf > outptr)
            @{
              const char *outerr = data->__outbuf;
              int result;

              result = fct (next_step, next_data, &outerr,
                            outbuf, written, 0);

              if (result != __GCONV_EMPTY_INPUT)
                @{
                  if (outerr != outbuf)
                    @{
                      /* @r{Reset the input buffer pointer.  We}
                         @r{document here the complex case.}  */
                      size_t nstatus;

                      /* @r{Reload the pointers.}  */
                      *inbuf = inptr;
                      outbuf = outptr;

                      /* @r{Possibly reset the state.}  */

                      /* @r{Redo the conversion, but this time}
                         @r{the end of the output buffer is at}
                         @r{@code{outerr}.}  */
                    @}

                  /* @r{Change the status.}  */
                  status = result;
                @}
              else
                /* @r{All the output is consumed, we can make}
                   @r{ another run if everything was ok.}  */
                if (status == __GCONV_FULL_OUTPUT)
                  status = __GCONV_OK;
           @}
        @}
      while (status == __GCONV_OK);

      /* @r{We finished one use of this step.}  */
      ++data->__invocation_counter;
    @}

  return status;
@}
@end smallexample
@end deftypevr

This information should be sufficient to write new modules.  Anybody
doing so should also take a look at the available source code in the GNU
C library sources.  It contains many examples of working and optimized
modules.

@c File charset.texi edited October 2001 by Dennis Grace, IBM Corporation