Wouldn't we need to extend this to cherry-pick, too? Suppose I do this:
$ git log -2 --oneline --decorate foo
abcd123456 (foo) Revert 123456aaaa
123456aaaa Some useful commit for the future, but not now
$ git checkout bar
$ git cherry-pick foo^ foo
$ git log -2 --oneline --decorate
badc0ffee (bar) Revert 123456aaaa
babeface0 Some useful commit for the future, but not now
Now when I rebase bar, the revert appears to be untwinned.
Similar problems arise for other history modifying tools like
filter-branch, fast-export, reposurgeon, bfg, etc.
I guess we can use 'git patch-id' to see if the companion commit is
still in our history, but this seems tenuous. Can we make it work
anyway?
On Thu, Apr 18, 2019 at 10:33 AM Jakub Narebski <jnareb@xxxxxxxxx> wrote:
>
> Junio C Hamano <gitster@xxxxxxxxx> writes:
> > Ævar Arnfjörð Bjarmason <avarab@xxxxxxxxx> writes:
> >> On Wed, Apr 17 2019, Giuseppe Crinò wrote:
> >>
> >>> The feature I'm asking is to add an extra-step during rebasing,
> >>> checking whether there's a reference to a commit that's not going to
> >>> be included in history and asks the user whether the heuristics is
> >>> correct and if she wants to update those references.
> >>>
> >>> Scenario: it can happen for a commit message to contain the ID of an
> >>> ancestor commit. A typical example is a commit with the message
> >>> "revert 01a9fe8". If 01a9fe8 and the commit that reverts it are
> >>> involved in a rebase the message "revert 01a9fe8" is no longer valid
> >>> -- the old 01a9fe8 has now a different hash. This will most likely be
> >>> ignored by the person who's rebasing but will let the other people
> >>> reading history confused.
> >>
> >> This would be useful. Done properly we'd need some machinery/command to
> >> extract the commit id parts from the free-text of the commit
> >> message. That would be useful for other parts of git, e.g. as discussed
> >> here:
> >> https://public-inbox.org/git/xmqqvaxp9oyp.fsf@xxxxxxxxxxxxxxxxxxxxxxxxxxx/
> >
> > That's a helpful input.
> >
> > But in general we do not have an infrastructure to systematically
> > keep track of "this commit was rewritten to produce that other
> > commit", so even if a mention of an old/superseded commit can be
> > identified reliably, there is no reliable source to rewrite it to
> > the name of the corresponding commit in the new world.
> >
> > For that mapping, we'd need something like the "git change/evolve"
> > Stefan Xenos was working on, which hasn't materialized.
>
> Well, what about limiting changes and rewriting only to the commits
> being rewritten by [interactive] rebase? I mean that we would rewrite
> "revert 01a9fe8" only if:
>
> a.) the commit with this message is undergoing rewrite
> b.) the commit 01a9fe8 is undergoing rewrite in the same command
>
> We could use the infrastructure from git-filter-branch for this.
>
> It is serious limitation, but that might be good enough for Giuseppe
> Crinò use case. Though for example there is a question what to do if
> referred-to commit (01a9fe8 in the example) is simply dropped, or is
> gets split in two? Ask user?
>
>
> Another possibility would be to provide a command line option to rebase
> which would automatically generate replacements (in git-replace meaning)
> from old pre-rebase name to new post-rebase name (assuming no splitting,
> no dropping commits). This would make references just work... well, as
> long as refs/replace/* are in place (they are not copied by default).
>
> On the other hand some of our performance-improving features, like the
> commit-graph, do not work if there are replacements.
>
>
> Best,
> --
> Jakub Narębski
