Single brackets matching in .gitignore rules

Hi,

I'm trying to make JGit .gitignore parser compatible to Git behavior and need help to understand some corner cases.

Git seem to fail to match file names if the ignore rule contain a single bracket '[' and seem to have inconsistent behavior for a single ']'. For all experiments below I was using git 2.1.0 from Fedora 21 repositories.

Example table for '[': ---------------------- rule | file | match? ---------------------- [ [ false [* [ false *[ [ false *[ a[ false ----------------------

I would expect that in all cases above Git ignore rules must match, since there is no valid character group defined and so the pattern should be interpreted literally. However, it looks like Git simply gives up on parsing the rule, probably because it contains unmatched '[' character starting a (broken) character group.

Next, the surprising table for ']': ---------------------- rule | file | match? ---------------------- ] ] true ]* ] true *] ] true *] a] true ----------------------

Here Git does not give up on parsing, treats unmatched ']' character literally, and doesn't dislike that it is an "unmatched" end of a broken character group. Why?

If Git interprets [ character only as part of the glob rules (character groups), it should interpret unmatched ] character equally. Also if Git interprets unmatched ] character literally, why not same behavior for the [ character? IMHO the first table for single '[' character is just a bug in Git, and it should be consistent to the ']' table.

Anyway, it would be nice to hear what should be the "right" way to interpret the tables above.

BTW the only official documentation I found about ignore rules:

https://www.kernel.org/pub/software/scm/git/docs/gitignore.html http://man7.org/linux/man-pages/man3/fnmatch.3.html http://man7.org/linux/man-pages/man7/glob.7.html

Andrey Loskutov <loskutov@gmx.de> writes:

] by itself is not special in a glob.

Andreas.

On Sunday 27 September 2015 10:58 Andreas Schwab wrote:

OK, this would explain the current state: once Git sees opened bracket, it always tries to parse the glob pattern and if this fails (if the bracket is not closed), the ignore rule does not match anything. This does not happen for closing bracket and therefore the difference in behavior.

Thanks!

P.S. In case anyone is interested, here is the path for JGit: https://git.eclipse.org/r/#/c/56773/4

On Sun, Sep 27, 2015 at 5:01 AM, Andrey Loskutov <loskutov@gmx.de> wrote:

This is already answered. I just want to add that C Git has stopped using system fnmatch for some time (part of the reason is system fnmatch behaves differently in corner cases). If you don't mind C, have a look at dowild() in wildmatch.c, or t/t3070-wildmatch.sh for some corner cases (but your cases aren't there, may be worth adding too) -- Duy

Duy Nguyen <pclouds@gmail.com> writes:

What exactly are you suggesting to "fix", though? Barf when an unmatched ] appears as a part of the pattern, instead of taking it literally?

On Monday 28 September 2015 10:40 Junio C Hamano wrote:

I would rather treat the opened but unmatched [ bracket literally, so that it works similar like ]. Originally I've implemented this "symmetrical" behavior in JGit without to know that unmatched [ and ] have different behavior in C Git.

Andrey Loskutov <loskutov@gmx.de> writes:

Ah, OK. As long as the (broken) pattern '[' has never been accepted, it is safe to assume that no working .gitignore file people have should contain such a string and relying that it does not match anything, I guess.