Re: [PATCH] last-modified: implement faster algorithm

> > Regardless of how you handle the above, I think that the commit slab
> > name here is a little generic. I guess it's OK since this is only
> > visible within this compilation unit, but perhaps something like
> > "active_paths_bitmap" would be more descriptive.
>
> I struggled a lot naming this thing, so I'm open to suggestions.

> > Likewise, I wonder if we should have elemtype here be just 'struct
> > bitmap'. Unfortunately I don't think the EWAH code has a function like:
> >
> >     void bitmap_init(struct bitmap *);
> >
> > and only has ones that allocate for us. So we may consider adding one,
> > or creating a dummy bitmap and copying its contents, or otherwise.
> >
> >>  struct last_modified {
> >>  	struct hashmap paths;
> >>  	struct rev_info rev;
> >>  	bool recursive;
> >>  	bool show_trees;
> >> +
> >> +	const char **all_paths;
> >> +	size_t all_paths_nr;
> >
> > I wonder if all_paths should be a strvec here? I think that this code
> > was all written when the type was called argv_array (hilariously, that
> > change took place towards the end of July, 2020, and the --go-faster
> > code where this patch came from was written just a couple of weeks
> > earlier.)
>
> Ahha, that might be a good idea. This might allow us to get rid of the
> hashmap, which stops us from storing the paths twice. Not sure what the
> impact on the performance would be, because the hashmap now is valuable
> for path_idx() lookups.

> > In the GitHub version of this patch, we pass all active paths to the
> > parent, assign the PARENT1 flag if it doesn't already have it, and put
> > it in the queue as well.
> >
> > In your version, we'd skip past the next for-loop, and do the same
> > pass-to-parent dance below, along with inserting the parent into the
> > prio queue.
>
> This is the "shortcut" I'm mentioning in my cover letter. In my testing
> it seemed it didn't provide any performance gains to keep it. I consider
> less code better code, so I left it out.

> > So I think that this is all functionally equivalent, but I had to work
> > through a little bit of the details here, mostly since I haven't looked
> > at or thought about this code in many years ;-).
> >
> >>  static int last_modified_run(struct last_modified *lm)
> >>  {
> >> +	int max_count, queue_popped = 0;
> >> +	struct prio_queue queue = { compare_commits_by_gen_then_commit_date };
> >> +	struct prio_queue not_queue = { compare_commits_by_gen_then_commit_date };
> >> +	struct commit_list *list;
> >>  	struct last_modified_callback_data data = { .lm = lm };
> >>
> >>  	lm->rev.diffopt.output_format = DIFF_FORMAT_CALLBACK;
> >>  	lm->rev.diffopt.format_callback = last_modified_diff;
> >>  	lm->rev.diffopt.format_callback_data = &data;
> >> +	lm->rev.no_walk = 1;
> >
> > This one is new relative to the original patch. Why set no_walk here?
>
> This comes from
> https://github.com/ttaylorr/git/commit/e8ea49705873d28f64b815bd00d14bdf6d48ca4d
>
> Well, it basically squashes various commits together. There are various
> commits doing different things here. I don't think it's valuable for
> anyone to see the full history of the iterations at GitHub, that's why I
> squashed it in.
>
> Would you consider it better to not set `no_walk`?