Re: [PATCH] add -p: coalesce hunks before testing applicability

[Johannes Schindelin <Johannes.Schindelin@gmx.de>]

Hi,

On Thu, 13 Sep 2018, Phillip Wood wrote:

> On 03/09/2018 20:01, Jochen Sprickerhof wrote:
>
> > * Phillip Wood <phillip.wood@talktalk.net> [2018-08-30 14:47]:
> >> When $newhunk is created it is marked as dirty to prevent
> >> coalesce_overlapping_hunks() from coalescing it. This patch does not
> >> change that. What is happening is that by calling
> >> coalesce_overlapping_hunks() the hunks that are not currently selected
> >> are filtered out and any hunks that can be coalesced are (I think that
> >> in the test that starts passing with this patch the only change is the
> >> filtering as there's only a single hunk selected).
> >
> > Agreed here. It would be enough to include the first hunk in the test to
> > make it fail again. Still I would see the patch as going in the right
> > direction as we need something like coalesce_overlapping_hunks() to make
> > the hunks applicable after the edit.
>
> Yes in the long term we want to be able to coalesce edited hunks, but I
> think it is confusing to call coalesce_overlapping_hunks() at the moment
> as it will not coalesce the edited hunks.

Agreed. I actually have code to coalesce even edited hunks, but it is all
in C.

> >> This is a subtle change to the test for the applicability of an
> >> edited hunk. Previously when all the hunks were used to create the
> >> test patch we could be certain that if the test patch applied then if
> >> the user later selected any unselected hunk or deselected any
> >> selected hunk then that operation would succeed. I'm not sure that is
> >> true now (but I haven't thought about it for very long).
> >
> > I'm not sure here. If we use the same test from t3701, do s(plit),
> > y(es), e(dit), it would fail later on. Can you come up with an
> > example?
>
> I think that if you split a hunk, edit the first subhunk, transforming a
> trailing context line to a deletion then try if you try to stage the
> second subhunk it will fail. With your patch the edit will succeed as
> the second subhunk is skipped when testing the edited patch. Then when
> you try to stage the second subhunk it will fail as it's leading context
> will contradict the trailing lines of the edited subhunk. With the old
> method the edit failed but didn't store up trouble for the future.

Indeed, this is a problem I also stumbled over.

> >> We could restore the old test condition and coalesce the hunks by
> >> copying all the hunks and setting $hunk->{USE}=1 when creating the
> >> test patch if that turns out to be useful (it would be interesting to
> >> see if the test still passes with that change).
> >
> > We set USE=1 for $newhunk already, or where would you set it?
>
> To match the old test it needs to be set on the hunks we've skipped or
> haven't got to yet so they're all in the patch that's tested after
> editing a hunk.

The way I fixed this in the C code is by teaching the equivalent of the
`coalesce_overlapping_hunks()` function to simply ignore the equivalent of
`$hunk->{USE}`: the function signature takes an additional `use_all`
parameter, which will override the `use` field.

Furthermore, my C code actually does the coalescing as part of the
`reassemble_patch()` function, feeding the output directly into the
`stdin` of the `git apply` process (with, or without `--check`).

And crucially, my C code does not try to assemble a new `hunks` array, but
simply works in-place, reverting the changes if the hunk edits result in a
patch that does not apply. The Perl version probably does not need that
part, as it is pretty careless with memory (as Perl encourages to do).

See for yourself:
https://github.com/dscho/git/commit/6f8ac4809280f2cd018683ab5199b004ada2350e#diff-f58d2179be56b196b9f35c6d24799a8eR337

Ciao,
Dscho

P.S.: Yes, this is part of my work to complete Slavica's "`git add -i`
in C" project. There are quite a few loose ends to tie, but I can already
use it myself to call `git add -p`, which is what I care most about, as it
regularly takes more than one second to spin up Perl on Windows, which is
friggin' annoying, I tell ya.