Re: [RFC PATCH 06/15] Automatically reuse sparse limiting arguments in revision walking

On Sat, Sep 4, 2010 at 7:58 PM, Nguyen Thai Ngoc Duy [off-list ref] wrote:

On Sun, Sep 5, 2010 at 10:13 AM, Elijah Newren [off-list ref] wrote:

quoted

In a sparse repository, by automatically making use of sparse limits
specified at clone time, we can avoid walking uninteresting commits and

I have a nicer approach here. Instead of modifying setup_revisions()
and similar functions, I update get_pathspec() to rewrite the
pathspecs from command line in narrow/shallow repos. get_pathspec()
currently does some form of rewriting already (prepending prefix).

So if you do "git log", get_pathspec() would return "git log -- PATH1
PATH2...". Will repost my series soon, or I can send that particular
patch to you.

Ooh, that's clever.  I like it.  Do you also do verification that any
paths specified by the user are a subset of the paths the sparse clone
is limited to?  That'd be really nice.  Much better than my simple
dumb "just append" logic.

How do you reconcile conflicting needs, though?  For most cases,
whenever the user specifies paths, get_pathspec should make sure those
paths are a subset of the sparse paths (throwing an error if they're
not) and then just use the user-specified ones.  However, I think
doing this in all cases would break ls-files/cat-file -p, as the
pathspecs passed to those are not necessarily rooted at the toplevel
of the repository (and the prefix doesn't correct for that either).

Also, when no paths are specified by the user, the correct answer can
still vary by command.  ls-tree should be given no pathspecs in such a
case, even in a sparse clone.  Nor should commit (it wouldn't hurt too
bad to send it the sparse pathspecs, though it'd send commit through
the "partial commit" section of the code base each time rather than
the "as-is commit" section, which seems suboptimal).  And on the other
end of the spectrum, log/rev-list/diff should all take the sparse
pathspecs in such a case.