argp mishandles domains

[Colin Watson]

On Sun, Mar 09, 2008 at 02:47:55PM +0100, Robert Luberda wrote:
> On the other hand while testing updated message catalogue translations I
> noticed, that man seems not to use translations from the
> man-db-gnulib.mo file, for example it shows untranslated messages
> 
> [17]/tmp> LC_ALL=pl_PL man --help | grep help
>   -?, --help                 give this help list
> 
> even though translation is available:
> 
> [19]/usr/share/locale/pl/LC_MESSAGES> msgunfmt man-db-gnulib.mo | grep
> -1  'give this help'
> 
> msgid "give this help list"
> msgstr "wyświetlenie tego tekstu pomocy"

This appears to be a Gnulib bug.

man-db sets its default text domain to "man-db", and keeps all Gnulib's
strings in a separate "man-db-gnulib" domain, using 'gnulib-tool
--po-domain=man-db' which results in
-DDEFAULT_TEXT_DOMAIN=\"man-db-gnulib\". (I checked this general
approach with Bruno a while back and he said it was fine.)

However, Gnulib's argp implementation has three bugs which cause this
not to work properly:

  * It hardcodes "libc" as a domain in two places. This is obviously
    wrong in Gnulib.

  * It uses argp_domain as both the domain used to translate its own
    strings (i.e. string literals in lib/argp-*.c) and the domain used
    to translate strings provided by the user, which normally have to
    use gettext_noop so that they can be used as 'struct argp_option'
    initialisers. This is dreadfully inconvenient because you have to
    copy strings about all over the place and keep your POT file up to
    date as the argp implementation changes. If argp is in libc then
    this is obviously impossible.

    Instead, argp should use DEFAULT_TEXT_DOMAIN to translate its own
    string literals (falling back to the default program domain if that
    is not set), and should reserve argp_domain for translating strings
    that appear in that argp structure.

  * In a number of places, argp uses the domain of the root argp
    structure when translating text from a child argp structure. This is
    the direct cause of Robert's bug, because the standard --help and
    --version options are implemented as a child argp structure with its
    own domain.

    My patch changes this to use argp_domain from the child instead.
    However, on reflection, I'm not certain that this is correct;
    arguably it needs to walk up the tree until it finds a domain to
    use. This depends on whether you think that argp_domain == NULL
    means "use default program domain" or "use same domain as parent
    argp structure".

Could somebody review the attached patch and see if it looks right?

2008-03-10  Colin Watson  <address@hidden>

        * lib/argp-help.c: Define ARGP_TEXT_DOMAIN to
        DEFAULT_TEXT_DOMAIN if set.
        (validate_uparams, fill_in_uparams, hol_help, _help)
        (__argp_failure): Use our default domain rather than the
        provided argp_domain to translate our own strings.
        (hol_entry_help): Translate user strings using the appropriate
        child argp_domain, rather than that from the root argp.

        * lib/argp-parse.c: Define ARGP_TEXT_DOMAIN to
        DEFAULT_TEXT_DOMAIN if set.
        (argp_default_argp, argp_version_argp): Set argp_domain to
        ARGP_TEXT_DOMAIN rather than "libc".
        (argp_version_parser, parser_finalize, parser_parse_opt): Use
        our default domain rather than the provided argp_domain to
        translate our own strings.

Thanks,

-- 
Colin Watson                                       address@hidden

gnulib.argp-domain.diff
Description: Text Data