Re: [PATCH 5/5] blame: use xdi_diff_hunks(), get rid of struct patch

[Junio C Hamano <gitster@pobox.com>]

René Scharfe <rene.scharfe@lsrfire.ath.cx> writes:

> Based on a patch by Brian Downing, this replaces the struct patch based
> code for blame passing with calls to xdi_diff_hunks().  This way we
> avoid generating and then parsing patches; we only let the interesting
> infos be passed to our callbacks instead.  This makes blame a bit faster:
>
>    $ blame="./git blame -M -C -C -p --incremental v1.6.0"
>
>    # master
>    $ /usr/bin/time $blame Makefile >/dev/null
>    1.38user 0.14system 0:01.52elapsed 100%CPU (0avgtext+0avgdata 0maxresident)k
>    0inputs+0outputs (0major+12226minor)pagefaults 0swaps
>    $ /usr/bin/time $blame cache.h >/dev/null
>    1.66user 0.13system 0:01.80elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
>    0inputs+0outputs (0major+12262minor)pagefaults 0swaps
>
>    # this patch series
>    $ /usr/bin/time $blame Makefile >/dev/null
>    1.27user 0.12system 0:01.40elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
>    0inputs+0outputs (0major+11836minor)pagefaults 0swaps
>    $ /usr/bin/time $blame cache.h >/dev/null
>    1.52user 0.12system 0:01.70elapsed 97%CPU (0avgtext+0avgdata 0maxresident)k
>    0inputs+0outputs (0major+12052minor)pagefaults 0swaps
>
> Signed-off-by: Rene Scharfe <rene.scharfe@lsrfire.ath.cx>

The resulting series reads quite clean.  I like it.

> Brian, your numbers looked much more impressive.  Could you please clock
> this code with your repository and the file server.c?  I wonder if this
> callback mechanism is just too complicated or if your case simply benefits
> lots more than the two files from git mentioned above.
>
> The patch series ends here without adding xdiff caching, for two reasons.
> It's quite easy to add it; patch 4 from your series applies unchanged and
> patch 5 is just needs a few small changes to account for the absence of
> compare_buffer().  More importantly, speed actually went down with caching
> for the test case.  The common tail optimization (xdi_diff() vs. xdl_diff())
> seems to beat caching for cache.h and Makefile..

Perhaps revision.c in our history would be more interesting than cache.h
or Makefile, as there are more line migrations from different places to
that file.