Re: Fwd: Possibly nicer pathspec syntax?

Junio C Hamano <gitster@pobox.com> writes:

Answering my own questions, here are my findings so far and a tentative conclusion.

With or without the patch in this thread, parse_pathspec() behaves the same way for either CWD or FULL if you feed a non-empty pathspec with at least one positive element. IOW, if a caller feeds a non-empty pathspec and there is no "negative" element involved, it does not matter if we feed CWD or FULL.

There are only a handful of callers that pass neither preference bits to parse_pathspec(). Here are my observations on them that tells me that most of them are OK if we change them to prefer either CWD or FULL:

 - archive.c::path_exists() feeds a pathspec with a single element
   to see if read_tree_recursive() finds any matching paths, to
   allow its caller to iterate over the original pathspec and see
   if there is a typo (i.e. an element that matches nothing).  It
   should prefer FULL to match what parse_pathspec_arg(), its
   caller, uses.

   The caller probably should refrain from passing ones with
   negative magic.  I.e. "git archive -- t :\!t/perf" errors out
   because checking each element independently in the loop means
   that ":\!t/perf" is checked alone, triggering "there is nothing
   to exclude from".

 - blame.c::find_origin() feeds a pathspec with a single element,
   which is a path in the history and does so as a literal, hence
   no room for "negative" to kick in.

 - builtin/check-ignore.c::check_ignore(), when argc==0, does not
   call parse_pathspec().  It does not take any magic other than
   FROMTOP, so "negative" won't come into the picture.

 - builtin/checkout.c::cmd_checkout(), when argc==0, does not call
   parse_pathspec().  This codepath will get affected by Linus's
   change ("cd t && git checkout :\!perf" would try to check out
   everything except t/perf, but what is a reasonable definition of
   "everything" in the context of this command).  We need to
   decide, and I am leaning towards preferring CWD for this case.

 - revision.c::setup_revisions() calls parse_pathspec() only when
   the caller gave a non-empty pathspec.  This pathspec is used for
   pruning log traversal (e.g. "only show commits that touch these
   paths") and is affected by Linus's change.  It should favor
   FULL.

 - tree-diff.c::try_to_follow_renames() feeds a pathspec with a
   single element as a literal, hence no room for "negative" to
   kick in.

So, I am tempted to suggest us doing the following:

 * Leave a NEEDSWORK comment to archive.c::path_exists() that is
   used for checking the validation of pathspec elements.  To fix it
   properly, we need to be able to skip a negative pathspec to be
   passed to this function by the caller, and to do so, we need to
   expose a helper from the pathspec API that gets a single string
   and returns what magic it has, but that is of lower priority.

 * Retire the PATHSPEC_PREFER_CWD bit and replace its use with the
   lack of the PATHSPEC_PREFER_FULL bit.

 * Keep most of the above callsites that currently do not pass
   CWD/FULL as they are, except the ones that should take FULL (see
   above).

Comments?