Re: Feature request: rebase -i inside of rebase -i

[Phillip Wood <phillip.wood123@gmail.com>]

Hi Philip

On 01/04/2020 17:43, Philip Oakley wrote:
> Hi Phillip,
> 
> On 31/03/2020 14:36, Phillip Wood wrote:
>> Hi Philip, George and Johannes
>>
>> I really like the idea of being able to extend or rewind an existing
>> rebase to reedit commits.
>>
>> On 31/03/2020 11:57, Philip Oakley wrote:
>>> Hi george,
>>>
>>> On 31/03/2020 01:00, George Spelvin wrote:
>>>> Thinking about Philip Oakley's suggestion, it dawned on me that
>>>> we can *already* do a nested rebase manually, and having a less
>>>> manual alias for the procedure would be reasonable.
>>>>
>>>> Suppose the last four commits are O-A-B-C, and whether they were
>>>> created
>>>> by this rebase or existed before is irrelevant.
>>>>
>>>> If I want to rebase --nested -i O, then I --edit-todo and
>>>
>>> Maybe `--rework` as a possible alternate option name, if the concept of
>>> it being truly nested process does not work out?
>>
>> or `--rewind` ?
> 
> A possibility, though it feels a bit narrow in suggesting the capabilities
>>
>>>> prepend the following four lines:
>>>> reset O
>>>> pick A
>>>> pick B
>>>> pick C
>>>>
>>>> If a nested rebase command does just that, I think it would cover my
>>>> use case.  If it adds a comment saying "nested rebase ends here",
>>>> it's easy to cancel the nested rebase if there was a mistake.
>>>>
>>>> A slightly fancier thing a nestrd rebase could do is see if any of the
>>>> newly created picks are also used in merges that were already in the
>>>> todo
>>>> list.  In that case, follow the pick by a label command and update the
>>>> later merge to refer to the label.
>>
>> If we're going to support rewinding a rebase that creates merges then
>> this is a prerequisite otherwise it wont work properly. It will also
>> need to update any existing labels that refer to a commits that get
>> rewritten when rewinding.
> 
> I would agree that the `label` instruction would need expanding to allow
> arbitrary refs (e.g. specific oids and other branches) to be labelled
> rather than just a presumed 'HEAD' ref. I did notice the man page
> doesn't actually define the format of the extra instructions (there was
> fun with awkward chars in label names on Windows).
> 
> I'm of the opinion that we don't re-label/rename the previous labels -
> they are what they are, but we do offer the ability to provide new
> labels. Hence the extension to the label format to allow the labelling
> of arbitrary refs, not just HEAD, along with showing existing progress,
> so folks can _see_ the new oids etc.

I'm not sure I understand how this would work. Say in the example below 
I decide I want to rework abc while editing fgh and run 'git rebase 
--rework abc^'. If we don't update the labels automatically how do I 
ensure that the merge will pick up the reworked abc as it's parent.

#pick abc
#label topic
#reset cde
edit fgh
merge topic

>>
>> When cancelling the nested rebase we need to take care to restore any
>> labels to their previous value if they have been updated by the nested
>> rebase. We also need to restore the list or rewritten commits so that
>> we don't report that we've rewritten the commits from the nested
>> rebase that we're aborting. These two requirements unfortunately make
>> it difficult to implement the nested rebase just by updating the todo
>> list.
> 
> I'm against the original conceptual idea of 'nesting' (or rewinding).
> What's done is done, it's in the object store, as is all the old
> original work, so we have a wider foundation to build from.  Clearly we
> are not in a clean work-state, with the half complete jobs, so it's more
> of a recovery activity than clean coding (from the user perspective).

I don't like the idea of nesting either but I think it would be useful 
to be able to get back to a known state when the changes I make with 
'rebase --rework' turn out to be a bad idea so I don't throw away all 
the work that I did before running 'rebase --rework'. I've been using a 
script that rewinds and I've got myself into situations where a 
speculative change made while rewinding turned out to be a bad idea and 
I want to get back to the state before the last rewind. When working on 
a patch series I can rewind multiple times before completing the rebase 
as I jump back and forth modifying and rewording each patch.

Best Wishes

Phillip

>> It needs to save the current labels (and reference the commits
>> somewhere so they don't get gc'd) and the rewritten-list. `git rebase
>> --abort` (or whatever we end up using to abort the nested part of the
>> rebase) needs to restore the labels and rewritten-list. I think it
>> should probably restore the todo list as well - if the original part
>> of the todo list gets edited during the nested rebase should we drop
>> those changes to the list or keep them when the nested rebase is aborted?
> 
> If we haven't aborted, then we just have the actual sequence of work,
> with some of the commits, ultimately, being left as unconnected stubs
> (once their temporary tips have gone upon completion of the rebase -
> could they be explicitly dropped?).  On completion, those abandoned
> commits could be explicitly marked as having been removed in the various
> book-keeping lists and logs (if it was useful and helpful). I think this
> is a different conceptual view of the work of the rebase - resolve process.
>>
>> Best Wishes
>>
>> Phillip
> --
> Philip
>