Re: [PATCH 5/7] shortlog: implement `--group=author` in terms of `--group=<format>`

From: Jeff King <peff@peff.net>
Date: Oct 11, 2022, 02:15 UTC
Message-ID: <Y0TRxDRFRgFonhtH@coredump.intra.peff.net>
In-Reply-To: <Y0TLf/J22ioQ5UCt@nand.local>

On Mon, Oct 10, 2022 at 09:48:47PM -0400, Taylor Blau wrote:

Show 8 quoted lines

> I think that shortlog_finish_setup() is probably the most reasonable
> choice (and I changed it to that locally). I spent a day or so thinking
> on and off about this while writing the series, and didn't come up with
> any satisfying names.
> 
> I think it points to something deeper about the API that doesn't quite
> sit right with me. But shortlog_finish_setup() is the least-unsatisfying
> name so far, so let's go with that ;-).

Yeah, touching that block in make_cover_letter() definitely tickled my memory that there was some awkwardness there. That's when I added shortlog_init() at all (which is good, because otherwise you'd have had an uninitialized log.format string-list).

I think the general pattern of:

  foo_init();
  foo.option = whatever;
  foo_finish_setup();

  foo_do_the_thing();

is OK. It's a little complicated, but it gives callers the freedom to tweak options with a lot of flexibility (e.g., based on command line config, command line options, etc). We have a similar pattern with diff_setup_done().

The other option is for any option-tweaking to go through methods which maintain invariants (like if GROUP_AUTHOR is set, then the "%an" format has been added). That's usually the "right" object-oriented way to do it. But I think even in this simple case it gets awkward, because the correct format can't be known until you see whether log->email is set. So it really has to be a "finalize" step after both log->email and log->group are set.

Show 14 quoted lines

> > This loses the HAS_MULTI_BITS() optimization. The idea there is that if
> > you have a single group-by that can't produce multiple outputs, then
> > there's no need to do duplicate detection.
> >
> > The equivalent in an all-formats world is something like:
> >
> >   log.format.nr > 1 && !log.trailers.nr
> >
> > (because trailers are special in that one trailer key can produce
> > multiple idents for a single commit).
> 
> Hmm. I suspect that this is going to become more complicated by the time
> that you read the final patch ;-). Let's wait until then and see what
> you think.

Yes, it's pretty gross with the trailer util thing. You'd probably want to do something like:

  static int needs_dedup(const struct string_list *formats)
  {
	struct string_list_item *item;
	if (formats->nr > 1)
		return 1;
	for_each_string_list_item(item, formats) {
		if (item->util)
			return 1;
	}
	return 0;
  }

and perhaps call it only once and cache the result, rather than iterating for each commit/format.

But if we leave trailers as is, then the logic is much easier. And implementing the optimization for --group=format should fall out pretty naturally (and that both preserves it for --group=author, and extends it to --group=format, which I should have noticed when reviewing patch 4).

Show 16 quoted lines

> > > @@ -439,8 +440,8 @@ int cmd_shortlog(int argc, const char **argv, const char *prefix)
> > >  	log.file = rev.diffopt.file;
> > >  	log.date_mode = rev.date_mode;
> > >
> > > -	if (!log.groups)
> > > -		log.groups = SHORTLOG_GROUP_AUTHOR;
> > > +	shortlog_init_group(&log);
> > > +
> > >  	string_list_sort(&log.trailers);
> >
> > Now that we have a "finish the shortlog init" function, probably this
> > trailer sort should go into it, too. The current caller doesn't need it,
> > but it removes on more gotcha from using the shortlog API.
> 
> We'll drop this list by the end of the series, too, so it probably isn't
> worth moving it into shortlog_finish_setup() in the interim.

Ah, right. Well, see my other comments. :)

-Peff