Re: [PATCH 1/5] merge-ort: replace string_list_df_name_compare with faster alternative

[Elijah Newren <newren@gmail.com>]

On Thu, May 27, 2021 at 2:00 PM René Scharfe <l.s.r@web.de> wrote:
>
> Am 27.05.21 um 10:37 schrieb Elijah Newren via GitGitGadget:
> > From: Elijah Newren <newren@gmail.com>
> >
> > Gathering accumulated times from trace2 output on the mega-renames
> > testcase, I saw the following timings (where I'm only showing a few
> > lines to highlight the portions of interest):
> >
> >     10.120 : label:incore_nonrecursive
> >         4.462 : ..label:process_entries
> >            3.143 : ....label:process_entries setup
> >               2.988 : ......label:plist special sort
> >            1.305 : ....label:processing
> >         2.604 : ..label:collect_merge_info
> >         2.018 : ..label:merge_start
> >         1.018 : ..label:renames
> >
> > In the above output, note that the 4.462 seconds for process_entries was
> > split as 3.143 seconds for "process_entries setup" and 1.305 seconds for
> > "processing" (and a little time for other stuff removed from the
> > highlight).  Most of the "process_entries setup" time was spent on
> > "plist special sort" which corresponds to the following code:
> >
> >     trace2_region_enter("merge", "plist special sort", opt->repo);
> >     plist.cmp = string_list_df_name_compare;
> >     string_list_sort(&plist);
> >     trace2_region_leave("merge", "plist special sort", opt->repo);
> >
> > In other words, in a merge strategy that would be invoked by passing
> > "-sort" to either rebase or merge, sorting an array takes more time than
> > anything else.  Serves me right for naming my merge strategy this way.
> >
> > Rewrite the comparison function and remove as many levels of indirection
> > as possible (e.g. the old code had
> >     cmp_items() ->
> >       string_list_df_name_compare() ->
> >         df_name_compare()
> > now we just have sort_dirs_next_to_their_children()), and tweak it to be
> > as optimized as possible for our specific case.  These changes reduced
> > the time spent in "plist special sort" by ~25% in the mega-renames case.
> >
> > For the testcases mentioned in commit 557ac0350d ("merge-ort: begin
> > performance work; instrument with trace2_region_* calls", 2020-10-28),
> > this change improves the performance as follows:
> >
> >                             Before                  After
> >     no-renames:        5.622 s ±  0.059 s     5.235 s ±  0.042 s
> >     mega-renames:     10.127 s ±  0.073 s     9.419 s ±  0.107 s
> >     just-one-mega:   500.3  ms ±  3.8  ms   480.1  ms ±  3.9  ms
>
> Interesting.
>
> >
> > Signed-off-by: Elijah Newren <newren@gmail.com>
> > ---
> >  merge-ort.c | 64 ++++++++++++++++++++++++++++++++++-------------------
> >  1 file changed, 41 insertions(+), 23 deletions(-)
> >
> > diff --git a/merge-ort.c b/merge-ort.c
> > index 142d44d74d63..367aec4b7def 100644
> > --- a/merge-ort.c
> > +++ b/merge-ort.c
> > @@ -2746,31 +2746,50 @@ static int detect_and_process_renames(struct merge_options *opt,
> >
> >  /*** Function Grouping: functions related to process_entries() ***/
> >
> > -static int string_list_df_name_compare(const char *one, const char *two)
> > +static int sort_dirs_next_to_their_children(const void *a, const void *b)
> >  {
> > -     int onelen = strlen(one);
> > -     int twolen = strlen(two);
>
> The old code scans both strings fully, while the new one stops when it
> reaches a difference and doesn't look at any further characters.  How
> much does that contribute to the speedup?  (I suspect a lot.)

Oh, indeed, good catch.  It appears to be responsible for essentially all of it.

> >       /*
> > -      * Here we only care that entries for D/F conflicts are
> > -      * adjacent, in particular with the file of the D/F conflict
> > -      * appearing before files below the corresponding directory.
> > -      * The order of the rest of the list is irrelevant for us.
> > +      * Here we only care that entries for directories appear adjacent
> > +      * to and before files underneath the directory.  In other words,
> > +      * we do not want the natural sorting of
> > +      *     foo
> > +      *     foo.txt
> > +      *     foo/bar
> > +      * Instead, we want "foo" to sort as though it were "foo/", so that
> > +      * we instead get
> > +      *     foo.txt
> > +      *     foo
> > +      *     foo/bar
> > +      * To achieve this, we basically implement our own strcmp, except that
> > +      * if we get to the end of either string instead of comparing NUL to
> > +      * another character, we compare '/' to it.
> >        *
> > -      * To achieve this, we sort with df_name_compare and provide
> > -      * the mode S_IFDIR so that D/F conflicts will sort correctly.
> > -      * We use the mode S_IFDIR for everything else for simplicity,
> > -      * since in other cases any changes in their order due to
> > -      * sorting cause no problems for us.
> > +      * The reason to not use df_name_compare directly was that it was
> > +      * just too expensive, so I had to reimplement it.
> >        */
> > -     int cmp = df_name_compare(one, onelen, S_IFDIR,
> > -                               two, twolen, S_IFDIR);
> > -     /*
> > -      * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
> > -      * that 'foo' comes before 'foo/bar'.
> > -      */
> > -     if (cmp)
> > -             return cmp;
> > -     return onelen - twolen;
> > +     const char *one = ((struct string_list_item *)a)->string;
> > +     const char *two = ((struct string_list_item *)b)->string;
>
> Casting away const, hmm. :-/  Harmless because no actual write is
> attempted, but still looks needlessly scary to me.

Right, that should have been
+     const char *one = ((const struct string_list_item *)a)->string;
+     const char *two = ((const struct string_list_item *)b)->string;
but since I was just assigning to a const char * on those lines, I'm
not sure why it'd qualify as scary.  Regardless, I'm happy to put
these consts back in.

> > +     unsigned char c1, c2;
> > +
> > +     while (*one && (*one == *two)) {
> > +             one++;
> > +             two++;
> > +     }
> > +
> > +     c1 = *one;
> > +     if (!c1)
> > +             c1 = '/';
> > +
> > +     c2 = *two;
> > +     if (!c2)
> > +             c2 = '/';
> > +
> > +     if (c1 == c2) {
> > +             /* Getting here means one is a leading directory of the other */
> > +             return (*one) ? 1 : -1;
> > +     }
> > +     else
> > +             return c1-c2;
> >  }
> >
> >  static int read_oid_strbuf(struct merge_options *opt,
> > @@ -3481,8 +3500,7 @@ static void process_entries(struct merge_options *opt,
> >       trace2_region_leave("merge", "plist copy", opt->repo);
> >
> >       trace2_region_enter("merge", "plist special sort", opt->repo);
> > -     plist.cmp = string_list_df_name_compare;
> > -     string_list_sort(&plist);
> > +     QSORT(plist.items, plist.nr, sort_dirs_next_to_their_children);
>
> How much does the direct use of QSORT instead of string_list_sort()
> contribute to the speedup?  (I suspect only a bit.)

Yep, I'll fix up the commit message.