Fix a pathological case in git detecting proper renames
commit9ae8fcb36ac9fde8e048a304cc3717f2c7914e78
authorLinus Torvalds <torvalds@linux-foundation.org>
Fri, 30 Nov 2007 00:41:09 +0000 (29 16:41 -0800)
committerJunio C Hamano <gitster@pobox.com>
Fri, 30 Nov 2007 23:49:17 +0000 (30 15:49 -0800)
tree2a53af966304453022426a88d43c9c419079182c
parent32d75d29f99cca8e0874b1bdf94ded48b576c906
Fix a pathological case in git detecting proper renames

On Thu, 29 Nov 2007, Jeff King wrote:
>
> I think it will get worse, because you are simultaneously calculating
> all of the similarity scores bit by bit rather than doing a loop. Though
> perhaps you mean at the end you will end up with a list of src/dst pairs
> sorted by score, and you can loop over that.

Well, after thinking about this a bit, I think there's a solution that may
work well with the current thing too: instead of looping just *once* over
the list of rename pairs, loop twice - and simply refuse to do copies on
the first loop.

This trivial patch does that, and turns Kumar's test-case into a perfect
rename list.

It's not pretty, it's not smart, but it seems to work. There's something
to be said for keeping it simple and stupid.

And it should not be nearly as expensive as it may _look_. Yes, the loop
is "(i = 0; i < num_create * num_src; i++)", but the important part is
that the whole array is sorted by rename score, and we have a

if (mx[i].score < minimum_score)
break;

in it, so uthe loop actually would tend to terminate rather quickly.

Anyway, Kumar, the thing to take away from this is:

 - git really doesn't even *care* about the whole "rename detection"
   internally, and any commits you have done with renames are totally
   independent of the heuristics we then use to *show* the renames.

 - the rename detection really is for just two reasons: (a) keep humans
   happy, and keep the diffs small and (b) help automatic merging across
   renames. So getting renames right is certainly good, but it's more of a
   "politeness" issue than a "correctness" issue, although the merge
   portion of it does matter a lot sometimes.

 - the important thing here is that you can commit your changes and not
   worry about them being somehow "corrupted" by lack of rename detection,
   even if you commit them with a version of git that doesn't do rename
   detection the way you expected it. The rename detection is an
   "after-the-fact" thing, not something that actually gets saved in the
   repository, which is why we can change the heuristics _after_ seeing
   examples, and the examples magically correct themselves!

 - try out the two patches I've posted, and see if they work for you. They
   pass the test-suite, and the output for your example commit looks sane,
   but hey, if you have other test-cases, try them out.

Here's Kumar's pretty diffstat with both my patches:

 Makefile                                         |    6 +++---
 board/{cds => freescale}/common/cadmus.c         |    0
 board/{cds => freescale}/common/cadmus.h         |    0
 board/{cds => freescale}/common/eeprom.c         |    0
 board/{cds => freescale}/common/eeprom.h         |    0
 board/{cds => freescale}/common/ft_board.c       |    0
 board/{cds => freescale}/common/via.c            |    0
 board/{cds => freescale}/common/via.h            |    0
 board/{cds => freescale}/mpc8541cds/Makefile     |    0
 board/{cds => freescale}/mpc8541cds/config.mk    |    0
 board/{cds => freescale}/mpc8541cds/init.S       |    0
 board/{cds => freescale}/mpc8541cds/mpc8541cds.c |    0
 board/{cds => freescale}/mpc8541cds/u-boot.lds   |    4 ++--
 board/{cds => freescale}/mpc8548cds/Makefile     |    0
 board/{cds => freescale}/mpc8548cds/config.mk    |    0
 board/{cds => freescale}/mpc8548cds/init.S       |    0
 board/{cds => freescale}/mpc8548cds/mpc8548cds.c |    0
 board/{cds => freescale}/mpc8548cds/u-boot.lds   |    4 ++--
 board/{cds => freescale}/mpc8555cds/Makefile     |    0
 board/{cds => freescale}/mpc8555cds/config.mk    |    0
 board/{cds => freescale}/mpc8555cds/init.S       |    0
 board/{cds => freescale}/mpc8555cds/mpc8555cds.c |    0
 board/{cds => freescale}/mpc8555cds/u-boot.lds   |    4 ++--
 23 files changed, 9 insertions(+), 9 deletions(-)

and here it is before:

 Makefile                                           |    6 +-
 board/cds/mpc8548cds/Makefile                      |   60 -----
 board/cds/mpc8555cds/Makefile                      |   60 -----
 board/cds/mpc8555cds/init.S                        |  255 --------------------
 board/cds/mpc8555cds/u-boot.lds                    |  150 ------------
 board/{cds => freescale}/common/cadmus.c           |    0
 board/{cds => freescale}/common/cadmus.h           |    0
 board/{cds => freescale}/common/eeprom.c           |    0
 board/{cds => freescale}/common/eeprom.h           |    0
 board/{cds => freescale}/common/ft_board.c         |    0
 board/{cds => freescale}/common/via.c              |    0
 board/{cds => freescale}/common/via.h              |    0
 board/{cds => freescale}/mpc8541cds/Makefile       |    0
 board/{cds => freescale}/mpc8541cds/config.mk      |    0
 board/{cds => freescale}/mpc8541cds/init.S         |    0
 board/{cds => freescale}/mpc8541cds/mpc8541cds.c   |    0
 board/{cds => freescale}/mpc8541cds/u-boot.lds     |    4 +-
 .../mpc8541cds => freescale/mpc8548cds}/Makefile   |    0
 board/{cds => freescale}/mpc8548cds/config.mk      |    0
 board/{cds => freescale}/mpc8548cds/init.S         |    0
 board/{cds => freescale}/mpc8548cds/mpc8548cds.c   |    0
 board/{cds => freescale}/mpc8548cds/u-boot.lds     |    4 +-
 .../mpc8541cds => freescale/mpc8555cds}/Makefile   |    0
 board/{cds => freescale}/mpc8555cds/config.mk      |    0
 .../mpc8541cds => freescale/mpc8555cds}/init.S     |    0
 board/{cds => freescale}/mpc8555cds/mpc8555cds.c   |    0
 .../mpc8541cds => freescale/mpc8555cds}/u-boot.lds |    4 +-
 27 files changed, 9 insertions(+), 534 deletions(-)

so it certainly makes the diffs prettier.

Linus

Signed-off-by: Junio C Hamano <gitster@pobox.com>
diffcore-rename.c