Re: [PATCH v2 6/6] exclude: filter patterns by directory level

On Sun, Mar 10, 2013 at 5:58 PM, Junio C Hamano [off-list ref] wrote:

Junio C Hamano [off-list ref] writes:

quoted

Nguyễn Thái Ngọc Duy  [off-list ref] writes:

quoted

A non-basename pattern that does not contain /**/ can't match anything
outside the attached directory. Record its directory level and avoid
matching unless the pathname is also at the same directory level.

Without defining what a "directory level" is, the above is a bit
hard to grok, but I think you mean an entry "b/c/*.c" that appears
in "a/.gitignore" file will want to match a path that is directly
in "a/b/c" directory (and not in its subdirectories),
"a/b/x.c" at the two levels deep subdirectory or "a/b/c/d/x.c" that is
four levels deep will never match the pattern.

The logic feels sound.

Actually, I think you may be able to do a lot more with a simpler
change.  If your top-level .gitignore has "a/b/c/*.c" in it, you
certainly want to mark it not to be applied when you are looking at
paths directly in directory a/b/ because they will never match, but
you also know that nothing will match when you are inside a/b/d/,
even though the pattern and the path you are checking are at the
same levels.  Your dirlen approach will fail for that case, no?

The idea behind prep_exclude() that organizes the exclode patterns
into a stack structure and update the groups near the leaves by
popping those for the old directory we were in and pushing those for
the new directory we are going into is to give us a place to tweak
the elements on the whole stack for optimization when we notice that
we are looking at paths in different directories.  Instead of giving
a "dirlen" member to each element, you could give a "do not look at
me" flag to it, and when you notice that you were in a/b/c/ and now
you are going to look at paths in a/b/d/, you can look at the group
that was read from the .gitignore from the top-level, and mark
entries that cannot be relevant (e.g. "a/b/c/*.c") as such.

The mark does not have to be a boolean.  "a/b/*.c" when you are in
"a/b/c/" can be marked as "This never matches, and I do not have to
re-check until I pop one level".  When digging deeper to "a/b/c/d",
you add one to that.  When switching to "a/b/e", you would first pop
twice ("d" and then "c"), each time decrementing the "I do not have
to re-check" counter by one, and then when pushing "e" down, you
notice that you need to re-check, and mark it again as "no need to
re-check for one pop".  So it is not like you have to re-scan all
entries textually every time you switch directories. Most entries
that are level-limited you would increment or decrement its counter
and only the ones at the level boundary need to be re-checked.

A bit confused by "dirlen" (what is it?). I think what you're trying
to say is "mark whether a pattern is applicable for entries in this
directory in prep_exclude, update the marks as we push and pop
directories". It does not sound simpler (and it's actually more
powerful, as you said it could avoid checking "a/b/c/*.c" when
standing in a/b/d). I'll give it a try.
-- 
Duy