fresh patch :)
1. It's easier that just stopping at dot, and i think this should be commented in the wiki or somewhere.
2. I read your first part of reply for 20mins, but im not sure; If i understand right, you mean, let the __locale_map* and strcut binding* be the id-card for msgcat list instead of the long name string, not only faster, but also more easy to construct pathname string. But there's some questions:
+ I removed name from msgcat, i can't find its use there, is it safe?
+ gettextdir() is replaced by a new loop, since i need the pointer of struct binding not only the dirname, but then, gettextdir() is only called by bindtextdomain(), is there a need to keep it? Or we have a better way to get the pointer of struct binding?
+ you said msgcat's indexed by  ( struct __locale_map *, struct binding *, category ), but i found lm(locale_map) is located by category, so if category is different, then we can't get the same lm, so we can just compare lm, right?

2017-02-11 10:36 GMT+08:00 Rich Felker <dalias@libc.org>:
On Thu, Feb 09, 2017 at 05:49:13PM +0800, He X wrote:
> sry!
>
> 2017-02-08 22:31 GMT+08:00 Rich Felker <dalias@libc.org>:
>
> > On Wed, Feb 08, 2017 at 06:13:30PM +0800, He X wrote:
> > > here the patch is: http://paste.ubuntu.com/23953329/
> > > The code tested, but maybe it sucks.
> >
> > Patches need to be attached and sent to the list, not pastebins that
> > might disappear. The latter don't work for discussing and preserving
> > discussion of the patch.

> --- a/src/locale/dcngettext.c 2017-02-06 14:39:17.860482624 +0000
> +++ b/src/locale/dcngettext.c 2017-02-06 14:39:17.860482624 +0000
> @@ -19,6 +19,7 @@
>  };
>
>  static void *volatile bindings;
> +char *__strchrnul(const char *, int);
>
>  static char *gettextdir(const char *domainname, size_t *dirlen)
>  {
> @@ -143,7 +143,7 @@
>
>       catname = catnames[category];
>       catlen = catlens[category];
> -     loclen = strlen(locname);
> +     loclen = __strchrnul(locname, '.') - locname;
>
>       size_t namelen = dirlen+1 + loclen+1 + catlen+1 + domlen+3;
>       char name[namelen+1], *s = name;
> @@ -157,6 +157,8 @@
> +rewrite_loc:
>       memcpy(s, locname, loclen);
>       s[loclen] = '/';
>       s += loclen + 1;
> +skip_loc:
>       memcpy(s, catname, catlen);
>       s[catlen] = '/';
>       s += catlen + 1;
> @@ -174,7 +175,22 @@
>               void *old_cats;
>               size_t map_size;
>               const void *map = __map_file(name, &map_size);
> -             if (!map) goto notrans;
> +             if (!map) {
> +                     if (s = strchr(name + dirlen + 1, '@')) {
> +                             *s++ = '/';
> +                             goto skip_loc;
> +                     }
> +                     if (locname && (s = strchr(name + dirlen + 1, '_')) && (strchr(name + dirlen +1, '/') > s) ) {
> +                             if (locname = strchr(locname, '@')) {
> +                                     loclen = __strchrnul(lm->name, '.') - locname;
> +                                     goto rewrite_loc;
> +                             } else {
> +                                     *s++ = '/';
> +                                     goto skip_loc;
> +                             }
> +                     }
> +                     goto notrans;
> +             }

This doesn't work because it changes both the key used for the lookup
and the filename mapped. If you try this code with a translation that
requires a fallback, and run it under strace, you'll see that _every_
call to gettext will try again to find the nonexistent files.

It could be fixed, but I think the code should be refactored so that,
rather than the msgcat list being indexed by pathname strings, it's
indexed by tuples of:

        ( struct __locale_map *, struct binding *, category )

These are all integers/pointers and thus compare very fast versus the
current strcmp operation, and it's very quick to look them up. Then we
only have to construct the pathname string when a new file needs to be
loaded, not on every call, and you're free to clobber the pathname
string while doing fallbacks.

>               p = calloc(sizeof *p + namelen + 1, 1);
>               if (!p) {
>                       __munmap((void *)map, map_size);
> --- a/src/locale/locale_map.c 2017-02-06 14:39:17.797148750 +0000
> +++ b/src/locale/locale_map.c 2017-02-06 14:39:17.797148750 +0000
> @@ -32,6 +32,7 @@
>       struct __locale_map *new = 0;
>       const char *path = 0, *z;
>       char buf[256];
> +     char *dotp;
>       size_t l, n;
>
>       if (!*val) {
> @@ -40,6 +41,12 @@
>               (val = getenv("LANG")) && *val ||
>               (val = "C.UTF-8");
>       }
> +     if (dotp = strchr(val, '.')) {
> +             char part[256];
> +             memcpy(part, val, dotp - val);
> +             memcpy(&part[dotp - val], ".UTF-8\0", 7);
> +             val = part;
> +     }
>
>       /* Limit name length and forbid leading dot or any slashes. */
>       for (n=0; n<LOCALE_NAME_MAX && val[n] && val[n]!='/'; n++);

I don't think this part is desirable, but if it were, it would need to
be done differently. As-is, it has serious UB, use of part[] after the
end of its lifetime. It also seems to have no check to see that
dotp-val is less than 256-7 or even that it's bounded, whereas the
code that immediately follows checks the length of the string pointed
to by val.

I think what it should be doing is the opposite, stopping when hitting
a dot in the name and only using the part up to the dot, except in the
one special case "C.UTF-8". The subsequent path search for the locale
file should probably then be repeated with combinations of dropping
@mod and _CC suffixes, but this dropping should _not_ affect the name
that's saved and reported back. (That is, if LC_TIME=fr_CA but only a
"fr" locale file exists, the "fr" file should get mapped but the name
returned by setlocale, and saved for use by gettext, should still be
the full "fr_CA" in case applications have "fr_CA" translations.)

Rich