Re: [PATCH 0/6] Optionally restrict range-diff output to "left" or "right" range only

From: Johannes Schindelin <johannes.schindelin@gmx.de>
Date: Feb 8, 2021, 22:36 UTC
Message-ID: <nycvar.QRO.7.76.6.2102082329460.29765@tvgsbejvaqbjf.bet>
In-Reply-To: <YB0Z5JXbe22ko2Mj@coredump.intra.peff.net>

Hi Peff,

On Fri, 5 Feb 2021, Jeff King wrote:

Show 21 quoted lines

> On Thu, Feb 04, 2021 at 04:56:16PM -0800, Junio C Hamano wrote:
>
> > > As an aside: I am curious if I'm missing something when you say the
> > > "only way" is to ask for a 'git range-diff ...@{u}'. IIUC what you're
> > > describing, I often resort to using 'git cherry' for that exact thing.
> > > But, I may not be quite understanding your use-case (and why git-cherry
> > > doesn't do what you want already).
> > >
> > > My latter question is purely for satisfying my own curiosity; I don't
> > > have any problem with a '--{left,right}-only' option in range-diff. From
> > > my quick read of the patches, it all looks pretty sane to me.
> >
> > The question is addressed to Dscho, and I am also somewhat curious.
> > Perhaps the reason would be that the output from cherry is not as
> > easy to read as range-diff, without any post-processing.
>
> I had the same curiosity; I'd use git-cherry (or rev-list --cherry) for
> this.
>
> I suspect the big difference is the quality of the matching. git-cherry
> is purely looking at patch-ids.

Indeed. Whenever I had tried `git cherry` in the past (which, admittedly, has been with geometrically decreasing frequency given the results), it completely failed to help me. And it's not only its reliance on perfect matches of the diff _with context lines_, it is also that the commit messages are completely ignored.

`git cherry`'s track record with me is so perfect that I want to put this line into all my Bash profiles:

	eval "$(set | sed -n '/^__git_main /,/^}$/{s/--list-cmds=list-mainporcelain[^)]*/& | grep -v ^cherry\$/;p}')"

Show 16 quoted lines

> So it is quite likely to say "this was not applied upstream" when what
> got applied differed slightly (e.g., fixups upstream, applied to a
> different base, etc). Whereas range-diff has some cost heuristics for
> deciding that two patches are basically the same thing.  So it would
> find more cases (and as a bonus, give you the diff to see what tweaks
> were made upstream).
>
> It does make me wonder if it would be useful for rev-list, etc to have
> an option to make "--cherry" use the more clever heuristics instead of
> just a patch-id. It would never show the same diff output as range-diff,
> but maybe more scripts would find the advanced heuristic useful.
>
> I know it would probably make rebase's "ignore if in upstream" feature
> less clunky when I rebase topics. But it would also make it more
> dangerous! E.g., right now I see any upstream tweaks as potential
> conflicts when I rebase, and I manually review them for sanity.

Yeah, I thought the same when I read the paragraphs before this one. It might sound convenient, but there _are_ false positives in `git range-diff`'s output, therefore I would recommend never using `git range-diff --left-only` or `[...] --right-only` with `-s`. IOW _always_ inspect the differences.

Ciao, Dscho