Frequently Asked Questions
for GNU gettext
Questions
General
#general_mailinglist
Where is the mailing list?
#general_source
Where is the newest gettext source?
#general_announce
I want to be notified of new gettext
releases.
Problems building GNU gettext
#building_solaris_libasprintf
On Solaris, I get a build
error ???text relocations remain??? in the
libasprintf
subdirectory
#building_rpath_check
During ???make check???, some tests
named
rpath-
Nxyz
fail: ???ld: fatal error ... -lrpathz???
#building_install
???make install??? fails
Problems integrating GNU gettext
#integrating_howto
How do I make use of
gettext()
in my package?
#integrating_undefined
I get a linker error ???undefined
reference to libintl_gettext???
#integrating_abuse_gettextize
gettextize adds multiple
references to the same directories/files
to
Makefile.am
and
#integrating_abuse_gettextize
configure.ac
#integrating_noop
My program compiles and links fine,
but doesn't output translated
strings.
GNU gettext on Windows
#windows_woe32
What does Woe32 mean?
#windows_howto
How do I compile, link and run a program
that uses the gettext()
function?
#windows_setenv
Setting the
LANG
environment variable doesn't have any effect
Other
#newline
What does this mean: ???`msgid' and `msgstr'
entries do not both
end with '\n'???
#translit
German umlauts are displayed like ???ge"andert???
instead of
???ge??ndert???
#localename
The
LANGUAGE
environment variable is ignored after I set
LANG=en
#nonascii_strings
I use accented characters in my
source code. How do I tell the
C/C++ compiler in which encoding it is (like
xgettext
's
--from-code
option)?
Answers
General
Where is the mailing list?
Three mailing lists are available:
bug-gnu-gettext@gnu.org
This mailing list is for discussion of features and bugs of the GNU
gettext
software
, including
libintl, the gettext-tools, and its autoconf macros.
translation-i18n@lists.sourceforge.net
This mailing list is for methodology questions around
internationalization, and for discussions of translator tools,
including but not limited to GNU gettext.
translation@iro.umontreal.ca
This is the email address of the
http://www.iro.umontreal.ca/contrib/po/HTML/
Free Translation
Project
, that is the project which manages the translated message
catalogs for many free software packages. Note that KDE and GNOME
packages are not part of this project; they have their own translation
projects:
http://i18n.kde.org/
i18n.kde.org
and
http://developer.gnome.org/projects/gtp/
gtp
.
The
bug-gnu-gettext
list
is archived as part of the
http://mail.gnu.org/archive/html/bug-gnu-utils/
bug-gnu-utils
archives.
bug-gnu-gettext
cannot be
subscribed on its own; to receive its contents by mail, subscribe to
bug-gnu-utils
.
Where is the newest gettext source?
The newest gettext release is available on
ftp.gnu.org
and its mirrors, in
http://ftp.gnu.org/gnu/gettext/
http://ftp.gnu.org/gnu/gettext/
.
Prereleases are announced on the
http://mail.gnu.org/pipermail/autotools-announce
autotools-announce
mailing list
.
Note that prereleases are meant for testing and not meant for use in
production environments. Please don't use the ???gettextize??? program of a
prerelease on projects which you share with other programmers via CVS.
If you want to live on the bleeding edge, you can also use the
development sources. Instructions for retrieving the gettext CVS are
found
http://savannah.gnu.org/projects/gettext
here
.
Note that building from CVS requires special tools (autoconf, automake,
m4, groff, bison, etc.) and requires that you pay attention to the
README-alpha
and
autogen.sh
files in the CVS.
I want to be notified of new gettext
releases.
If you are interested in stable gettext releases, you can follow the
http://mail.gnu.org/pipermail/info-gnu
info-gnu
mailing list
. It
is also available as a newsgroup
nntp://news.gmane.org/gmane.org.fsf.announce
gmane.org.fsf.announce
through
http://www.gmane.org/
gmane.org
.
You can also periodically check the download location.
If you are interested in testing prereleases as well, you can subscribe
to the
http://mail.gnu.org/pipermail/autotools-announce
autotools-announce
mailing
list
.
Problems building GNU gettext
On Solaris, I get a
build error ???text relocations remain??? in the
libasprintf
subdirectory
libtool (or more precisely, the version of libtool that was available
at the time the gettext release waas made) doesn't support linking C++
libraries with some versions of GCC. As a workaround, you can configure
gettext with the option
--disable-libasprintf
.
During ???make check???, some tests
named
rpath-
Nxyz
fail: ???ld: fatal error ... -lrpathz???
If only a few among the many rpath tests fail, you can probably ignore
the problem. The rpath tests are sensitive to incomplete shared library
support in the system, and to bugs in libtool that creates the shared
libraries. Some known failures are listed in
autoconf-lib-link/tests/rpath.README
.
To ignore the problem, just proceed with
cd gettext-tools
make check
cd ..
???make install??? fails
???
make install DESTDIR=
/some/tempdir
??? can fail with
an error message relating to
libgettextlib
or
libgettextsrc
, or can
silently fail to install
libgettextsrc
.
On some platforms, this is due to limitations of libtool regarding
DESTDIR
. On other platforms, it
is due to the way the system handles shared libraries, and libtool
cannot work around it. Fortunately, on Linux and other glibc based
systems,
DESTDIR
is
supported if no different version of gettext is already installed (i.e.
it works if you uninstall the older gettext before building and
installing the newer one, or if you do a plain ???
make install
??? before ???
make install DESTDIR=
/some/tempdir
???). On other
systems, when
DESTDIR
does not work, you can still do ???
make
install
??? and copy the installed files to
/some/tempdir
afterwards.
If ???
make install
??? without
DESTDIR
fails, it's a bug which
you are welcome to report to the usual bug report address.
Problems integrating GNU gettext
How do I make use of
gettext()
in my package?
It's not as difficult as it sounds. Here's the recipe for C or C++
based packages.
Add an invocation of
AM_GNU_GETTEXT([external])
to the package's
configure.{ac,in}
file.
Invoke ???
gettextize --copy
???.
It will do most of the autoconf/automake related work for you.
Add the
gettext.h
file to the package's source directory, and include it in all source
files that contain translatable strings or do output via
printf
or
fprintf
.
In the source file defining the main() function of the program,
add these lines to the header
#include <locale.h>
#include "gettext.h"
and these lines near the beginning of the main() function:
setlocale (LC_ALL, "");
bindtextdomain (PACKAGE,
LOCALEDIR);
textdomain (PACKAGE);
Mark all strings that should be translated with _(), like this:
_("No errors found.")
. While
doing this, try to turn the strings into good English, one entire
sentence per string, not more than one paragraph per string, and use
format strings instead of string concatenation. This is needed so that
the translators can provide accurate translations.
In every source file containing translatable strings, add these lines
to the header:
#include "gettext.h"
#define _(string) gettext (string)
In the freshly created
po/
directory, set up the
POTFILES.in
file, and do a ???
make update-po
???.
Then distribute the generated
.pot
file to your nearest translation project.
Shortly before a release, integrate the translators'
.po
files into the
po/
directory and do ???
make update-po
??? again.
You find detailed descriptions of how this all works in the GNU gettext
manual, chapters ???The Maintainer's View??? and ???Preparing Program
Sources???.
I get a linker error ???undefined
reference to libintl_gettext???
This error means that the program uses the
gettext()
function after having
included the
<libintl.h>
file from GNU gettext (which remaps it to
libintl_gettext()
), however at
link time a function of this name could not be linked in. (It is
expected to come from the
libintl
library, installed by GNU gettext.)
There are many possible reasons for this error, but in any case you
should consider the
-I
,
-L
and
-l
options passed to the
compiler. In packages using
autoconf
generated configure scripts,
-I
options come from the
CFLAGS
and
CPPFLAGS
variables
(in Makefiles also
DEFS
and
INCLUDES
),
-L
options come from the
LDFLAGS
variable, and
-l
options come from the
LIBS
variable. The first thing
you should check are the values of these variables in your environment
and in the  package's
config.status
autoconfiguration result.
To find the cause of the error, a little analysis is needed. Does the
program's final link command contains the option ???-lintl????
If yes:
Find out where the
libintl
comes from. To do this, you have to check for
libintl.a
and
libintl.so*
(
libintl.dylib
on MacOS X) in
each directory given as a -L option, as well as in the compiler's
implicit search directories. (You get these implicit search directories
for gcc by using ???
gcc -v
???
instead of ???
gcc
??? in the
final link command line; compilers other than GCC usually look in
/usr/lib
and
/lib
.) A shell command like
$ for d in /usr/local/lib
/usr/lib /lib; do ls -l $d/libintl.*; done
will show where the
libintl
comes from. By looking at the dates and whether each library defines
libintl_gettext
(via ???
nm
path
/libintl.so
| grep libintl_gettext
???) you can now distinguish three possible
causes of the error:
Some older libintl is used instead of the newer one. The fix
is to remove the old library or to reorganize your -L options.
The used libintl is the new one, and it doesn't contain
libintl_gettext. This would be a bug in gettext. If this is the case,
please report it to the usual bug report address.
The used libintl is a static library (libintl.a), there are
no uses of gettext in .o files before the ???-lintl??? but there are some
after the ???-lintl???. In this case the fix is to move the ???-lintl??? to the
end or near the end of the link command line. The only libintl
dependency that needs to be mentioned after ???-lintl??? is ???-liconv???.
If no:
In this case it's likely a bug in the package you are building: The
package's Makefiles should make sure that ???-lintl??? is used where needed.
Test whether libintl was found by configure. You can check this by doing
$ grep
'\(INTLLIBS\|LIBINTL\)' config.status
and looking whether the value of this autoconf variable is non-empty.
If yes: It should be the responsibility of the Makefile to
use the value of this variable in the link command line. Does the
Makefile.in rule for linking the program use
@INTLLIBS@
or
@LIBINTL@
?
If no: It's a Makefile.am/in bug.
If yes: Something strange is going on. You need to dig
deeper.
Note that
@INTLLIBS@
is
for
gettext.m4
versions
<= 0.10.40 and
@LIBINTL@
is for
gettext.m4
versions >= 0.11, depending on which
gettext.m4
was used to build
the package's
configure
-
regardless of which gettext you have now installed.
If no: So libintl was not found.
Take a look at the package's
configure.in/ac
.
Does it invoke AM_GNU_GETTEXT?
If no: The gettext maintainers take no responsibilities for
lookalikes named CY_GNU_GETTEXT, AM_GLIB_GNU_GETTEXT, AM_GNOME_GETTEXT
and similar, or for homebrewn autoconf checks. Complain to the package
maintainer.
If yes: It looks like the
-I
and
-L
options were inconsistent.
You should have a
-I
somedir
/include
in the
CFLAGS
or
CPPFLAGS
if and only if you
also have a
-L
somedir
/lib
in the
LDFLAGS
. And
somedir
/include
should contain a
libintl.h
if and only if
somedir
/lib
contains
libintl.{a,so}
.
This case can also happen if you have configured a GCC < 3.2 with
the same
--prefix
option
as you used for GNU libiconv or GNU gettext. This is fatal, because
these versions of GCC implicitly use
-L
prefix
/lib
but
not
-I
prefix
/include
. The
workaround is to use a different
--prefix
for GCC.
gettextize adds multiple
references to the same directories/files
to
Makefile.am
and
configure.ac
If
gettextize
is used on
a package, then the
po/
,
intl/
,
m4/
directories of the package
are removed, and then
gettextize
is invoked on the package again, it will re-add the
po/
,
intl/
,
m4/
directories and change
Makefile.am
,
configure.ac
and
ChangeLog
accordingly. This is
normal. The second use of
gettextize
here is an abuse of the program.
gettextize
is a wizard intended to transform a
working
source package
into a
working
source package
that uses the newest version of gettext. If you
start out from a nonfunctional source package (it is nonfunctional
since you have omitted some directories), you cannot expect that
gettextize
corrects it.
Often this question arises in packages that use CVS. See the section
???CVS Issues / Integrating with CVS??? of the GNU gettext documentation.
This section mentions a program
autopoint
which is designed to reconstruct those files and directories created by
gettextize
that can be
omitted from a CVS repository.
My program compiles and links fine,
but doesn't output translated
strings.
There are several possible reasons. Here is a checklist that allows you
to determine the cause.
Check that the environment variables LC_ALL, LC_MESSAGES,
LC_CTYPE, LANG, LANGUAGE together specify a valid locale and language.
To check this, run the commands
$ gettext --version
$ gettext --help
You should see at least some output in your desired language. If not,
either
You have chosen a too exotic language.
gettext
is localized to 33
languages. Choose a less exotic language, such as Galician or
Ukrainian. Or
There is a problem with your environment variables. Possibly
LC_ALL points to a locale that is not installed, or LC_MESSAGES and
LC_CTYPE are inconsistent.
Check that your program contains a
setlocale
call.
To check this, run your program under ltrace. For example,
$ ltrace ./myprog
...
setlocale(6,
"")
= "de_DE.UTF-8"
If you have no ltrace, you can also do this check by running your
program under the debugger. For example,
$ gdb ./myprog
(gdb) break main
(gdb) run
Breakpoint 1, main ()
(gdb) break setlocale
(gdb) continue
Breakpoint 2, setlocale ()
;; OK, the breakpoint has been hit, setlocale() is being
called.
Either way, check that the return value of
setlocale()
is non-NULL. A NULL
return value indicates a failure.
Check that your program contains a
textdomain
call, a
bindtextdomain
call referring
to the same message domain, and then really calls the
gettext
,
dgettext
or
dcgettext
function.
To check this, run the program under ltrace. For example,
$ ltrace ./myprog
...
textdomain("hello-c")
= "hello-c"
bindtextdomain("hello-c", "/opt/share"...) = "/opt/share"...
dcgettext(0, 0x08048691, 5, 0x0804a200, 0x08048689) =
0x4001721f
If you have no ltrace, you can also do this check by running your
program under the debugger. For example,
$ gdb ./myprog
(gdb) break main
(gdb) run
Breakpoint 1, main ()
(gdb) break textdomain
(gdb) break bindtextdomain
(gdb) break gettext
(gdb) break dgettext
(gdb) break dcgettext
(gdb) continue
Breakpoint 2, textdomain ()
(gdb) continue
Breakpoint 3, bindtextdomain ()
(gdb) continue
Breakpoint 6, dcgettext ()
Note that here
dcgettext()
is called instead of the
gettext()
function mentioned in the source code; this is due to an optimization
in
<libintl.h>
.
When using libintl on a non-glibc system, you have to add a prefix ???
libintl_
??? to all the function
names mentioned here, because that's what the functions are really
named, under the hood.
If
gettext
/
dgettext
/
dcgettext
is not called at all,
the possible cause might be that some autoconf or Makefile macrology
has turned off internationalization entirely (like the
--disable-nls
configuration
option usually does).
Check that the
.mo
file that contains the translation is really there where the program
expects it.
To check this, run the program under strace and look at the
open()
calls. For example,
$ strace ./myprog 2>&1
| grep '^open('
open("/etc/ld.so.preload", O_RDONLY)    = -1
ENOENT (No such file or directory)
open("/etc/ld.so.cache",
O_RDONLY)      = 5
open("/lib/libc.so.6",
O_RDONLY)        = 5
open("/usr/lib/locale/locale-archive", O_RDONLY|O_LARGEFILE)
= 5
open("/usr/share/locale/locale.alias", O_RDONLY) = 5
open("/opt/share/locale/de/LC_MESSAGES/hello-c.mo", O_RDONLY)
= 5
...
A nonnegative
open()
return value means that the file has been found.
If you have no strace, you can also guess the
.mo
file's location: it is
localedir
/
lang
/LC_MESSAGES/
domain
.mo
where
domain
is the argument
passed to
textdomain()
,
localedir
is the second argument
passed to
bindtextdomain()
,
and
lang
is the language (
LL
) or language and territory (
LL
_
CC
),
depending on the environment variables checked in step 1.
Check that the .mo file contains a translation for the string
that is being asked for.
To do this, you need to convert the .mo file back to PO file format,
through the command
$ msgunfmt
localedir
/
lang
/LC_MESSAGES/
domain
.mo
and look for an
msgid
that matches the given string.
GNU gettext on Windows
What does Woe32 mean?
???Woe32??? denotes the Windows 32-bit operating systems for x86: Windows
NT/2000/XP and Windows 95/98/ME. Microsoft uses the term ???Win32??? to
denote these; this is a psychological trick in order to make everyone
believe that these OSes are a ???win??? for the user. However, for most
users and developers, they are a source of woes, which is why I call
them ???Woe32???.
How do I compile, link and run a
program that uses the gettext()
function?
When you use RedHat's cygwin environment, it's as on Unix:
You need to add an
-I
option to the compilation command line, so that the compiler finds the
libintl.h
include file, and
You need to add an
-L
option to the link command line, so that the linker finds the
libintl
library.
When you use the Mingw environment (either from within cygwin, with
CC="gcc -mno-cygwin"
, or from
MSYS, with
CC="gcc"
), I
don't know the details.
When you use the Microsoft Visual C/C++ (MSVC) compiler, you will
likely use the precompiled Woe32 binaries. For running a program that
uses gettext(), one needs the
.bin.woe32.zip
packages of
gettext-runtime
and
libiconv
. As a
developer, you'll also need the
xgettext
and
msgfmt
programs that
are contained in the
.bin.woe32.zip
package of
gettext-tools
.
Then
You need to add an
-MD
option to all compilation and link command lines. MSVC has three
different, mutually incompatible, compilation models (
-ML
,
-MT
,
-MD
); the default is
-ML
.
intl.dll
uses the
-MD
model, therefore the rest
of the program must use
-MD
as well.
You need to add an
-I
option to the compilation command line, so that the compiler finds the
libintl.h
include file.
You need to add an
-L
option to the link command line, so that the linker finds the
intl.lib
library.
You need to copy the
intl.dll
and
iconv.dll
to the
directory where your
.exe
files are created, so that they will be found at runtime.
Setting the
LANG
environment variable doesn't have any effect
If neither LC_ALL, LC_MESSAGES nor LANGUAGES is set, it's the LANG
environment variable which determines the language into which gettext()
translates the messages.
You can test your program by setting the LANG environment variable from
outside the program. In a Windows command interpreter:
set LANG=de_DE
.\myprog.exe
Or in a Cygwin shell:
$ env LANG=de_DE ./myprog.exe
If this test fails, look at the question ???My program compiles and links
fine, but doesn't output translated
strings.??? above.
If this test succeeds, the problem is related in the way you set the
environment variable. Here is a checklist:
Check that you are using the
-MD
option in all compilation
and link command lines. Otherwise you might end up calling the
putenv()
function from
Microsoft's
libc.lib
,
whereas
intl.dll
is using
the
getenv()
function
from Mictosoft's
msvcrt.lib
.
Check that you set the environment variable using
both
SetEnvironmentVariable()
and
putenv()
. A convenient way to
do so, and to deal with the fact that some Unix systems have
setenv()
and some don't, is the
following function.
#include <string.h>
#include <stdlib.h>
#if defined _WIN32
# include <windows.h>
#endif
int my_setenv (const char * name, const char * value) {
size_t namelen = strlen(name);
size_t valuelen = (value==NULL ? 0 : strlen(value));
#if defined _WIN32
/* On Woe32, each process has two copies of the
environment variables,
one managed by the OS and one
managed by the C library. We set
the value in both locations, so that
other software that looks in
one place or the other is guaranteed
to see the value. Even if it's
a bit slow. See also
<
http://article.gmane.org/gmane.comp.gnu.mingw.user/8272
http://article.gmane.org/gmane.comp.gnu.mingw.user/8272
>
<
http://article.gmane.org/gmane.comp.gnu.mingw.user/8273
http://article.gmane.org/gmane.comp.gnu.mingw.user/8273
>
<
http://www.cygwin.com/ml/cygwin/1999-04/msg00478.html
http://www.cygwin.com/ml/cygwin/1999-04/msg00478.html
>
*/
if (!SetEnvironmentVariableA(name,value))
return -1;
#endif
#if defined(HAVE_PUTENV)
char* buffer = (char*)malloc(namelen+1+valuelen+1);
if (!buffer)
return -1; /* no need to set errno =
ENOMEM */
memcpy(buffer,name,namelen);
if (value != NULL) {
buffer[namelen] = '=';
memcpy(buffer+namelen+1,value,valuelen);
buffer[namelen+1+valuelen] = 0;
} else
buffer[namelen] = 0;
return putenv(buffer);
#elif defined(HAVE_SETENV)
return setenv(name,value,1);
#else
/* Uh oh, neither putenv() nor setenv() ... */
return -1;
#endif
}
Other
What does this mean: ???`msgid' and `msgstr'
entries do not both end
with '\n'???
It means that when the original string ends in a newline, your
translation must also end in a newline. And if the original string does
not end in a newline, then your translation should likewise not have a
newline at the end.
German umlauts are displayed like
???ge"andert??? instead of ???ge??ndert???
This symptom occurs when the
LC_CTYPE
facet of the locale is not set; then gettext() doesn't know which
character set to use, and converts all messages to ASCII, as far as
possible.
If the program is doing
setlocale (LC_MESSAGES, "");
then change it to
setlocale (LC_CTYPE, "");
setlocale (LC_MESSAGES, "");
or do both of these in a single call:
setlocale (LC_ALL, "");
If the program is already doing
setlocale (LC_ALL, "");
then the symptom can still occur if the user has not set
LANG
, but instead has set
LC_MESSAGES
to a valid locale
and has set
LC_CTYPE
to
nothing or an invalid locale. The fix for the user is then to set
LANG
instead of
LC_MESSAGES
.
The
LANGUAGE
environment variable is ignored after I set
LANG=en
This is because ???en??? is a language name, but not a valid locale name.
The
ABOUT-NLS
file
says:
In the
LANGUAGE
environment variable, but not in the
LANG
environment variable,
LL
_
CC
combinations can be
abbreviated as
LL
to
denote the language's main dialect.
Why is
LANG=en
not
allowed? Because
LANG
is
a setting for the entire locale, including monetary information, and
this depends on the country: en_GB, en_AU, en_ZA all have different
currencies.
I use accented characters in my
source code. How do I tell the
C/C++ compiler in which encoding it is (like
xgettext
's
--from-code
option)?
Short answer: If you want your program to be useful to other people,
then
don't use accented characters
(or other non-ASCII characters) in string literals
in the source code
. Instead, use
only ASCII for string literals, and use
gettext()
to retrieve their
display-ready form.
Long explanation:
The reason is that the ISO C standard specifies that the character set
at compilation time can be different from the character set at
execution time.
The character encoding at compilation time is the one which determines
how the source files are interpreted and also how string literals are
stored in the compiled code. This character encoding is generally
unspecified; for recent versions of GCC, it depends on the LC_CTYPE
locale in effect during the compilation process.
The character encoding at execution time is the one which determines
how standard functions like
isprint()
,
wcwidth()
etc. work and
how strings written to standard output should be encoded. This
character encoding is specified by POSIX to depend on the LC_CTYPE
locale in effect when the program is executed; see also the description
in the
ABOUT-NLS
file.
Strings in the compiled code are not magically converted between the
time the program is compiled and the time it is run.
Therefore what could you do to get accented characters to work?
Can you ensure that the execution character set is the same as the
compilation character set? Even if your program is to be used only in a
single country, this is not realistically possible. For example, in
Germany there are currently three character encodings in use: UTF-8,
ISO-8859-15 and ISO-8859-1. Therefore you would have to explicitly
convert the accented strings from the compilation character set to the
execution character set at runtime, for example through iconv().
Can you ensure that the compilation character set is the one in which
your source files are stored? This is not realistically possible
either: For compilers other than GCC, there is no way to specify the
compilation character set. So let's assume for a moment that everyone
uses GCC; then you will specify the LC_CTYPE or LC_ALL environment
variable in the Makefile. But for this you have to assume that everyone
has a locale in a given encoding. Be it UTF-8 or ISO-8859-1 - this is
not realistic. People often have no locale installed besides the one
they use.
Use of wide strings
L"..."
doesn't help solving the problem, because on systems like FreeBSD or
Solaris, the way how wide string literals are stored in compiled code
depends on the compilation  character set, just as it does for
narrow strings
"..."
.
Moreover, wide strings have problems of their own.
Use of ISO C 99 Unicode escapes "\u
xxxx
"
doesn't help either because these characters are converted to the
compilation character set at compile time; so again, since you can't
guarantee that the compilation character set is not ASCII, you're
risking compilation errors just as if the real character had been used
in the source instead of the Unicode escape.
So, in summary, there is no way to make accented characters in string
literals work in C/C++.
You might then wonder what
xgettext
's
--from-code
option is good
for. The answer is
For the comments in C/C++ source code. The compiler ignores them.
For other programming languages like Java, for which the compiler
converts all string literals to UTF-8.
GNU gettext FAQ
Bruno Haible <
mailto:bruno@clisp.org
bruno@clisp.org
>
Last modified: 24 February 2004
