Re: [BUG REPORT] Git does not correctly replay bisect log

From: Christian Couder <christian.couder@gmail.com>
Date: Dec 6, 2018, 23:02 UTC
Message-ID: <CAP8UFD2+Kk84J=xwGo2bRGUeOXaRCKbLhZSS769qvDMFN8Pg3g@mail.gmail.com>
In-Reply-To: <5ab4e3fb2e22097753bde7702e67d6ce924283a2.camel@gmail.com>

On Thu, Dec 6, 2018 at 6:30 PM Lukáš Krejčí <lskrejci@gmail.com> wrote:

Show 5 quoted lines

>
> I am talking about `git bisect replay`. The shell script, as far as I
> can see, only updates the references (ref/bisect/*) and never checks if
> the revisions marked as 'good' are ancestors of the 'bad' one.
> Therefore, $GIT_DIR/BISECT_ANCESTORS_OK file is never created.

Indeed `git bisect replay` first only updates the references (ref/bisect/*) according to all the "git bisect {good,bad}" instructions it finds in the log it is passed. After doing that though, before exiting, it calls `bisect_auto_next` which calls `bisect_next` which calls `git bisect--helper --next-all` which checks the merge bases.

I think it is a bug.

`git bisect replay` is right to only update the references (ref/bisect/*) and not to compute and checkout the best commit to test at each step except at the end, but it should probably just create the $GIT_DIR/BISECT_ANCESTORS_OK file if more than one bisection step has been performed (because the merge bases are checked as part of the first bisection step only).

> The first time the ancestors are checked is in the helper (`git-bisect-
> -help --next-all`) that has only limited information from refs/bisect*.

`git-bisect--helper --next-all` knows how to get refs/bisect* information, otherwise it couldn't decide which is the next best commit to test.

Thanks for your help in debugging this, Christian.