Re: [PATCH v4 12/12] t3070: disable two fnmatch tests that have different results on different libc

[Nguyen Thai Ngoc Duy <pclouds@gmail.com>]

On Fri, Oct 12, 2012 at 2:22 PM, Johannes Sixt <j.sixt@viscovery.net> wrote:
> Am 10/10/2012 12:40, schrieb Nguyễn Thái Ngọc Duy:
>> fnmatch on glibc-2.12.1 returns no match. glibc-2.15 returns ok.
>
> There are many more cases that fail with the fnmatch() that we ship in
> compat/fnmatch. To test this on Linux, you have to remove the "#if defined
> _LIBC || !defined __GNU_LIBRARY__" brackets from compat/fnmatch/fnmatch.c
> and build with NO_FNMATCH.

Thanks. The point of checking against fnmatch is to make sure that
wildmatch does not divert in behavior from fnmatch. For some corner
cases that behavior is undefined, I think it is ok for different
fnmatch versions to behave differently.

But some of failed tests make me worry about fnmatch. compat/fnmatch
for example does not match '**/foo' with '/foo' (it does '*/foo' with
'/foo'). 'A' is matched with '[[:digit:][:space:][:upper:]]' but not
'[[:digit:][:upper:][:space:]]'. Perhaps we better off convert git to
use wildmatch only to keep matching behavior more reliable. wildmatch
does not support non-FNM_PATHNAME mode, but that should be easy to
fix. The only downside I see is fnmatch may support locale while
wildmatch does not.
-- 
Duy