Re: [PATCH v2 4/4] log: --author-date-order

On Mon, Jun 10, 2013 at 12:39:24AM -0700, Junio C Hamano wrote:

quoted

I'm not excited about introducing yet another place that parses commit
objects (mostly not for correctness, but because we have had
inconsistency in how malformed objects are treated). It is at least
using split_ident_line which covers the hard bits. I wonder how much
slower it would be to simply call format_commit_message to do the
parsing.

The thought certainly crossed my mind, not exactly in that form but
more about splitting the machinery used in pretty.c into a more
reusable form.

The result of my attempt however did not become all that reusable
(admittedly I didn't spend too much brain cycles on it), so I punted
;-).

Yes, I feel like it has been tried before. The problem is that a clean
interface would let you get individual pieces of information with a
single call. But an efficient interface will utilize the same parsing
pass to get multiple items out, and stop parsing when we have gotten all
required items (but leave the parser in a consistent state so that we
can pick it up later).

The format_commit_one parser does that, but the "format_commit_context"
it holds is a bit bulky. I think it might be possible to pull out the
parsing bits into a separate struct, and you could call it something
like:

  struct commit_parser parser;
  unsigned long authordate;
  const char *authorname;
  int authorlen;

  commit_parser_init(&parser, commit);
  authordate = commit_parse_authordate(&parser);
  authorname = commit_parse_authorname(&parser, &authorlen);

where the second parse call is basically "free", because we've already
done (and cached) the hard work in the first call.

So they might look like:

  static void parse_author_ident(struct commit_parser *parser)
  {
          if (!parser->author.name_begin) {
                  if (!parser->authorline.start)
                          parse_commit_header(parser);
                  split_ident_line(&parser->author,
                                   parser->authorline.start,
                                   parser->authorline.len);
          }
  }

  unsigned long commit_parse_authordate(struct commit_parser *parser)
  {
          parse_author_ident(parser);
          /* XXX should check for malformedness here */
          return strtoul(ident.date_begin, NULL, 10);
  }

  const char *commit_parse_authorname(struct commit_parser *parser,
                                      unsigned long *len)
  {
          parse_author_ident(parser);
          *len = parser.author.name_end - parser.author.name_begin;
          return parser.author.name_begin;
  }

and so forth. It would be easy (and have the same efficiency) for
format_commit_message to build on that, and it calling it from regular
code is not too bad.

But you are right.  The commit->buffer may no longer be there, and
the --author-date-order option needs to read the object again
in this codepath.  That would be in line with what --pretty/format
would do, I guess.

Or we could extend parse_commit() API to take an optional commit
info slab to store not just author date but other non-essential
stuff like people's names, and we arrange that extended API to be
triggered when we know --author-date-order is in effect?

I like the latter option. It takes a non-trivial amount of time to load
the commits from disk, and now we are potentially doing it 2 or 3 times
for a run (once to parse, once to get the author info for topo-sort, and
possibly later to show it if --pretty is given; though I did not check
and maybe we turn off save_commit_buffer with --pretty). It would be
nice to have an extended parse_object that handled that. I'm not sure of
the interface. Maybe variadic with pairs of type/slab, like:

  parse_commit_extended(commit,
                        PARSE_COMMIT_AUTHORDATE, &authordate_slab,
                        PARSE_COMMIT_DONE);

?

-Peff