Re: [PATCH on master v2] revision: use commit graph in get_reference()

[Jonathan Tan <jonathantanmy@google.com>]

> On Sun, Dec 09, 2018 at 09:51:28AM +0900, Junio C Hamano wrote:
> 
> > > -static int parse_commit_in_graph_one(struct commit_graph *g, struct commit *item)
> > > +static struct commit *parse_commit_in_graph_one(struct repository *r,
> > > +						struct commit_graph *g,
> > > +						struct commit *shell,
> > > +						const struct object_id *oid)
> > 
> > Now the complexity of the behaviour of this function deserves to be
> > documented in a comment in front.  Let me see if I can get it
> > correctly without such a comment by explaining the function aloud.
> > 
> > The caller may or may not have already obtained an in-core commit
> > object for a given object name, so shell could be NULL but otherwise
> > it could be used for optimization.  When shell==NULL, the function
> > looks up the commit object using the oid parameter instead.  The
> > returned in-core commit has the parents etc. filled as if we ran
> > parse_commit() on it.  If the commit is not yet in the graph, the
> > caller may get a NULL even if the commit exists.

In the next revision, I'll unify parse_commit_in_graph_one() (quoted
above) with parse_commit_in_graph(), so that the comment I wrote for the
latter can cover the entire functionality. I think the comment covers
the details that you outline here.

> Yeah, this was the part that took me a bit to figure out, as well. The
> optimization here is really just avoiding a call to lookup_commit(),
> which will do a single hash-table lookup. I wonder if that's actually
> worth this more complex interface (as opposed to just always taking an
> oid and then always returning a "struct commit", which could be old or
> new).

Avoidance of lookup_commit() is more important than an optimization, I
think. Here, we call lookup_commit() only when we know that that object
is a commit (by its presence in a commit graph). If we just called it
blindly, we might mistakenly create a commit for that hash when it is
actually an object of another type. (We could inline lookup_commit() in
parse_commit_in_graph_one(), removing the object creation part, but that
adds complexity as well.)