Re: Nonexistent changes appear rebasing but only with rebase.backend=apply
Hi Marco,
On Fri, Jun 25, 2021 at 12:15 AM Marco Giuliano <marco.giuliano@tesisquare.com> wrote:
Show 7 quoted lines
>
> Hi Philip,
> oh, so I guessed correctly...
> Could you be more specific about the copy/rename issue of the apply backend ?
> Is there any bug report I can refer to ?
>
> Thank you very much for your support
Could you do us a favor? Please avoid top-posting and instead answer after the quoted original text, much like I'll do here...
Show 35 quoted lines
> On Thu, Jun 24, 2021 at 8:39 PM Phillip Wood <phillip.wood123@gmail.com> wrote:
> >
> > Hi Marco
> >
> > On 24/06/2021 17:23, Marco Giuliano wrote:
> > > Thanks Felipe and Philip for your answers.
> > >
> > > Let's proceed in order:
> > > @Felipe: I tried rebasing with --no-fork-point but the problem remains the same
> > >
> > > @Philip:
> > > I'm a basic git user, so bear with me if I say silly things...
> > > I tried to search for rebased-patches in .git folder when rebase
> > > stopped waiting for
> > > conflict resolution, but I didn't find any file named like that.
> > > There's a folder named rebase-apply though did you mean that ?
> >
> > Looking at the source I thought they ended up just in .git but I haven't
> > checked again, as you seem to have found the source of the problem below
> > lets not worry about that.
> >
> > > Anyway, looking at the conflict file of "fileA" directly (not behind a
> > > visual diff tool) I noticed that the marker line >>>>>>>> COMMIT
> > > DESCR: FILENAME indicates a different file name then the current
> > > conflicted file.
> > > That reminded me that those two files A & B, were actually copies
> > > (real copy, not symlink) of other two files inside the same repo.
> > > Is it somehow possible that auto-detected-renaming is involved in this
> > > (since the files are identical but in two different locations) ?
> > > Trying to give you some hints, maybe it is totally unrelated...
> >
> > I meant to ask if anything had been copied or renamed but forgot. The
> > merge backend detects copies and renames and handles them correctly but
> > the apply backend does not so I think this is the source of the discrepancy.
> >
So I've got a hunch here. First, Marco, from your example, can you run
git diff --name-status -C Y...X
and report whether the files involved were either renamed or copied on the upstream side of history. (Also, please note that there are three dots between Y and X) Also, can you report whether a message of the form "inexact rename detection was skipped due to too many files" when using the merge backend and that files you copied from end up in modify/delete conflicts when using the merge backend?
My suspicion is:
* You both renamed some files and copied them on the upstream side of history
* You made further changes to the renamed files, more so than to the "copies"
* You have enough added/removed/renamed files that with the merge
backend you are triggering merge.renameLimit and not getting rename
detection, causing various reports of modify/delete conflicts
* The apply backend's use of fake trees can be viewed as a poor-man's
version of the irrelevant rename detection optimization I added to
merge-ort, but it works well enough to get you under merge.renameLimit
so that renames are detected for the apply backend
* The fact that you modified what you view as renames more than what
you viewed as copies has the perverse effect of causing git's rename
detection to pair up files with the copy rather than the rename, and
apply changes to it
The net result is that the 'merge' backend isn't apply your changes from various files to their renamed counterpart, while the 'apply' backend is almost doing that but applying them to what you think of as the copy rather than the intended renamed file. The questions I asked above may be able to help us either confirm or rule out this theory.
