Re: Bug? git commit fileA tries to delete fileB and other oddities

[Mikael Simonsson <m@mikaelsimonsson.com>]

git bisect made that pretty easy, thanks. It looks like it started
with this commit:

"unpack-trees: optimize walking same trees with cache-tree"
https://github.com/git/git/commit/b4da37380b7774248086f42bcd59397a44e1ac79

(I ran git bisect between tags v2.19.2 and v2.20.0.)


Testing the commit in question:
% git checkout b4da37380b7774248086f42bcd59397a44e1ac79
% gmake -j 6 NO_OPENSSL=YesPlease NO_PERL=YesPlease NO_TCLTK=YesPlease
NO_GETTEXT=YesPlease NO_EXPAT=YesPlease all

% ~/tmp/git-source/git version
git version 2.18.0.751.gb4da37380b

% ~/tmp/git-source/git commit null_terminated.hh

# Please enter the commit message for your changes. Lines starting
# with '#' will be ignored, and an empty message aborts the commit.
#
# On branch master
# Your branch is up to date with 'origin/master'.
#
# Changes to be committed:
#       deleted:    algo/find.hh
#       new file:   app/.reserved
#       deleted:    ascii/contains_icase.hh
#       deleted:    ascii/contains_icase.test.cc
#       modified:   null_terminated.hh
#
# Changes not staged for commit:
Aborting commit due to empty commit message.


Note that the latest version behaves differently (it tries to delete
another file):

% git version
git version 2.24.1

% git commit null_terminated.hh

# Please enter the commit message for your changes. Lines starting
# with '#' will be ignored, and an empty message aborts the commit.
#
# On branch master
# Your branch is up to date with 'origin/master'.
#
# Changes to be committed:
#       deleted:    defer.hh
#       modified:   null_terminated.hh
Aborting commit due to empty commit message.


The commit before "unpack-trees: optimize walking same trees with
cache-tree" works as it should:
% git checkout b4da37380b7774248086f42bcd59397a44e1ac79^
% gmake -j 6 NO_OPENSSL=YesPlease NO_PERL=YesPlease NO_TCLTK=YesPlease
NO_GETTEXT=YesPlease NO_EXPAT=YesPlease all

% ~/tmp/git-source/git version
git version 2.18.0.750.g0d1ed5963d

% ~/tmp/git-source/git commit null_terminated.hh

# Please enter the commit message for your changes. Lines starting
# with '#' will be ignored, and an empty message aborts the commit.
#
# On branch master
# Your branch is up to date with 'origin/master'.
#
# Changes to be committed:
#       modified:   null_terminated.hh
#
# Changes not staged for commit:
Aborting commit due to empty commit message.


Here's my local commit from about a week ago when the problems started.

I'm sure I ran "git commit readme.md" because I had several files
staged. The commit pulled in 3 staged files (I think there were more)
and deleted 3 files that 100% weren't staged.

% git show --numstat 7d5d43d849f1391b3cc11f12e037b8cb5174e3a3
commit 7d5d43d849f1391b3cc11f12e037b8cb5174e3a3
Author: Mikael Simonsson <m@mikaelsimonsson.com>
Date:   Sun Dec 8 10:38:38 2019 +0000

    Add short info and requirements for ranges

56      0       algo/find.hh
0       1       app/.reserved
16      0       ascii/contains_icase.hh
28      0       ascii/contains_icase.test.cc
0       41      bench/bench.hh
0       41      debug.hh
76      0       readme.md

-Mikael



On Tue, Dec 17, 2019 at 7:33 AM Jeff King <peff@peff.net> wrote:
>
> On Mon, Dec 16, 2019 at 09:50:33AM +0000, Mikael Simonsson wrote:
>
> > I will try to recreate the problem with a new repository.
> >
> > So far I think I've narrowed it down to a bug introduced in git 2.20.0.
> >
> > Not buggy:
> > [...]
>
> If you can build Git from source, you might try using "git bisect" to
> find the exact commit where the problem starts.
>
> > The buggy versions all try to delete fileX when running "git commit fileA":
> >
> > % git commit fileA
> >
> > # Please enter the commit message for your changes. Lines starting
> > # with '#' will be ignored, and an empty message aborts the commit.
> > #
> > # On branch master
> > # Your branch is up to date with 'origin/master'.
> > #
> > # Changes to be committed:
> > #       deleted:    fileX
> > #       modified:   fileA
> > #
> > # Changes not staged for commit:
>
> Is there anything about "fileX" and "fileA"'s names that might be
> relevant? E.g., might they case-fold to the same name or something?
>
> -Peff