Re: [PATCH] bisect: error out when given any good rev that is not an ancestor of the bad rev

[Junio C Hamano <gitster@pobox.com>]

Junio C Hamano <gitster@pobox.com> writes:

> Christian Couder <chriscool@tuxfamily.org> writes:
>
>> Before this patch "git bisect" doesn't really work when it is given
>> some good revs that are siblings of the bad rev.
>>
>> For example if there is the following history:
>>
>> A-B-C-D
>>    \E-F
>>
>> and we launch "git bisect start D F" then only C and D will be
>> considered as possible first bad commit. This is wrong because A, B and
>> E may be bad too if the bug exists everywhere except in F that fixes it.
>
> Please don't.
>
> bisect is about finding a single regression by partitioning the graph into
> older good section and newer bad section with a *single* "first bad
> commit".
>
> Your "this could also be possible" scenario is already outside the
> realm.  You are assuming A, B and F is good, and D is bad.  But if E is
> bad, then that breakage cannot possibly affect the transition between B
> and D from good to bad (E cannot break D), so C must *also* be bad.

... which means you are dealing with *two* breakages.  That's outside what
bisect deals with.

And this does not need to have forked development.  If the graph were like
this:

  A-B-C-D-E-F

and if F is bad and B is good, with your logic, after checking that D is
already bad, we cannot discount E --- after somehow fixing D, we _might_
also be introducing another breakage with E.  You cannot even check for
that anyway, but the logic is the same.