The "--parents" option did not appear until SVN 1.5.x
and is completely unnecessary in this case.
Reported-by: Jens Lehmann <Jens.Lehmann@web.de>
Signed-off-by: Eric Wong <normalperson@yhbt.net>
When mergetool is run without path limiters it loops
over each entry in 'git ls-files -u'. This includes
autoresolved paths.
Teach mergetool to only merge files listed in 'rerere status'
when rerere is enabled.
There are some subtle but harmless changes in behavior.
We now call cd_to_toplevel when no paths are given.
We do this because 'rerere status' paths are always relative
to the root. This is beneficial for the non-rerere use as
well in that mergetool now runs against all unmerged files
regardless of the current directory.
This also slightly tweaks the output when run without paths
to be more readable.
The old output:
Merging the files: foo
bar
baz
The new output:
Merging:
foo
bar
baz
Signed-off-by: David Aguilar <davvid@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add a regression test for the git log -M --follow $diff_option bug
introduced in v1.7.2-rc0~103^2~2, $diff_option being diff related
options like -p, --stat, --name-only etc.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Currently git-svn assumes that two tags created from the same
revision will have the same repo url, so it uses a ref to the
tag without checking that its url matches the current url.
This causes issues when fetching an svn repo where a tag was
created, deleted, and then recreated under the following
circumstances:
- Both tags were copied from the same revision.
- Both tags had the same name.
- Both tags had different repository paths.
- [Optional] Both tags have a file with the same name but
different content.
When all four conditions are met, a checksum mismatch error
occurs because the content of two files with the same path
differs (see t/t9155--git-svn-fetch-deleted-tag.sh):
Checksum mismatch: ChangeLog 065854....
expected: ce771b....
got: 9563fd....
When only the first three conditions are met, no error occurs
but the tag in git matches the first (deleted) tag instead of
the last (most recent) tag (see
t/t9156-git-svn-fetch-deleted-tag-2.sh).
The fix is to verify that the repo url for the ref matches the
current url. If the urls do not match, then a "tail" is grown
on the tag name by appending a dash and rechecking the new ref's
repo url until either a matching repo url is found or a new tag
is created.
Signed-off-by: David D. Kilzer <ddkilzer@kilzer.net>
Acked-by: Eric Wong <normalperson@yhbt.net>
The sed utilities on IRIX and Solaris do not interpret the sequence '\t'
to mean a tab character; they read a literal character 't'. So, use a
literal tab instead.
Signed-off-by: Brandon Casey <casey@nrlssc.navy.mil>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Without this, attempting to index a pack containing objects that have been
replaced results in a fatal error that looks like:
fatal: SHA1 COLLISION FOUND WITH <replaced-object> !
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Acked-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
dump_marks_helper() has a bug when dumping marks larger than 2^20-1,
i.e., when the sparse array has more than two levels. The bug was
that the 'base' counter was being shifted by 20 bits at level 3, and
then again by 10 bits at level 2, rather than a total shift of 20 bits
in this argument to the recursive call:
(base + k) << m->shift
There are two ways to fix this correctly, the elegant:
(base + k) << 10
and the one I chose due to edit distance:
base + (k << m->shift)
Signed-off-by: Raja R Harinath <harinath@hurrynot.org>
Acked-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Change the git-svn tests to use $PERL_PATH, not the "perl" in $PATH.
Using perl in $PATH was added by Sam Vilain in v1.6.6-rc0~95^2~3,
Philippe Bruhat introduced $PERL_PATH to the test suite in
v1.6.6-rc0~9^2, but the lib-git-svn.sh tests weren't updated to use
the new convention.
This resulted in the git-svn tests always being skipped on my
system. My /usr/bin/perl has access to SVN::Core and SVN::Repos, but
the perl in my $PATH does not.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are two ways a user might want to use "diff --relative":
1. For a file in a directory, like "subdir/file", the user
can use "--relative=subdir/" to strip the directory.
2. To strip part of a filename, like "foo-10", they can
use "--relative=foo-".
We currently handle both of those situations. However, if the user passes
"--relative=subdir" (without the trailing slash), we produce inconsistent
results. For the unified diff format, we collapse the double-slash of
"a//file" correctly into "a/file". But for other formats (raw, stat,
name-status), we end up with "/file".
We can do what the user means here and strip the extra "/" (and only a
slash). We are not hurting any existing users of (2) above with this
behavior change because the existing output for this case was nonsensical.
Patch by Jakub, tests and commit message by Jeff King.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Ensure that strcmp() isn't called when head is null.
Previously we were getting segfaults when checkout -B was done from a
detached HEAD.
Signed-off-by: Tay Ray Chuan <rctay89@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The test would not fail if the filtering failed to do anything, since
in
test -z "$(git diff HEAD directorymoved:newsubdir)"'
'directorymoved:newsubdir' is not valid, so git-diff fails without
printing anything on stdout. But then the exit status of git-diff is
lost, whereas test -z "" succeeds.
Use 'git diff --exit-code' instead, which does the right thing and has
the added bonus of showing the differences if there are any.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The .gitmodules file is parsed for "submodule.<name>.ignore" entries
before looking for them in .git/config. Thus settings found in .git/config
will override those from .gitmodules, thereby allowing the local developer
to ignore settings given by the remote side while also letting upstream
set defaults for those users who don't have special needs.
Signed-off-by: Jens Lehmann <Jens.Lehmann@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The new "ignore" config option controls the default behavior for "git
status" and the diff family. It specifies under what circumstances they
consider submodules as modified and can be set separately for each
submodule.
The command line option "--ignore-submodules=" has been extended to accept
the new parameter "none" for both status and diff.
Users that chose submodules to get rid of long work tree scanning times
might want to set the "dirty" option for those submodules. This brings
back the pre 1.7.0 behavior, where submodule work trees were never
scanned for modifications. By using "--ignore-submodules=none" on the
command line the status and diff commands can be told to do a full scan.
This option can be set to the following values (which have the same name
and meaning as for the "--ignore-submodules" option of status and diff):
"all": All changes to the submodule will be ignored.
"dirty": Only differences of the commit recorded in the superproject and
the submodules HEAD will be considered modifications, all changes
to the work tree of the submodule will be ignored. When using this
value, the submodule will not be scanned for work tree changes at
all, leading to a performance benefit on large submodules.
"untracked": Only untracked files in the submodules work tree are ignored,
a changed HEAD and/or modified files in the submodule will mark it
as modified.
"none" (which is the default): Either untracked or modified files in a
submodules work tree or a difference between the subdmodules HEAD
and the commit recorded in the superproject will make it show up
as changed. This value is added as a new parameter for the
"--ignore-submodules" option of the diff family and "git status"
so the user can override the settings in the configuration.
Signed-off-by: Jens Lehmann <Jens.Lehmann@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
check-ref-format --branch requires access to the repository
to resolve refs like @{-1}.
Noticed by Nguyễn Thái Ngọc Duy.
Cc: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
git-rebase calls out to merge strategies, but did not support merge
strategy options so far. Add this, in the same style used in
git-merge.
Sadly we have to do the full quoting/eval dance here, since
merge-recursive supports the --subtree=<path> option which potentially
contains whitespace.
This patch does not cover git rebase -i, which does not call any merge
strategy directly except in --preserve-merges, and even then only for
merges.
[jc: with a trivial fix-up for 'expr']
Signed-off-by: Mike Lundy <mike@fluffypenguin.org>
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When printing an error message saying a ref was requested that we do not
have, only print that ref, rather than the ref and everything sent to us
on the same packet line (e.g. protocol support specifications).
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The dcommit command fails if an otherwise unmodified file has
been touched in the working directory:
Cannot dcommit with a dirty index. Commit your changes
first, or stash them with `git stash'.
This happens because "git diff-index" reports a difference
between the index and the filesystem:
:100644 100644 d00491...... 000000...... M file
The fix is to run "git update-index --refresh" before
"git diff-index" as is done in git-rebase and
git-rebase--interactive before "git diff-files".
This changes dcommit to display a list of modified files before
exiting.
Also add a similar test case for "git svn rebase".
[ew: rearranged commit message subject]
Signed-off-by: David D. Kilzer <ddkilzer@kilzer.net>
Acked-by: Eric Wong <normalperson@yhbt.net>
The --no-python option was added to test-lib.sh by Johannes Schindelin
in early 2006 in abb7c7b3. It was later turned into a no-op by Junio C
Hamano in 7cdbff14 the same year.
Over three years is long enough before removing this old wart which
was retained for backwards compatibility. Our tests have been using
NO_PYTHON and "test_have_prereq PYTHON" for a long time now.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Running the tests with --quiet under a TAP harness will always fail,
since a TAP harness always needs actual test output to go along with
the plan that's being emitted.
Change the test-lib.sh to ignore the --quiet option under
HARNESS_ACTIVE to work around this. Then users that have --quiet in
their GIT_TEST_OPTS can run tests under prove(1) without everything
breaking.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When performing a non-interactive rebase, sometimes
"git rebase --continue" will fail if an unmodified file is
touched in the working directory:
You must edit all merge conflicts and then
mark them as resolved using git add
This is caused by "git diff-files" reporting a difference
between the index and the filesystem:
:100644 100644 d00491...... 000000...... M file
The fix is to run "git update-index --refresh" before
"git diff-files" as is done in git-rebase--interactive.
Signed-off-by: David D. Kilzer <ddkilzer@kilzer.net>
Acked-by: Johannes Schindelin <Johannes.Schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Given a file with:
(define archive-id "$Format:%ct|%h|a$")
and an export-subst attribute, the "%h" results in an full 40-digit
object name instead of the expected 7-digit one.
The export-subst feature requests unabbreviated object names because
that is the low-level default. The effect was not observable until
v1.7.1.1~17^2~3 (2010-05-03), which taught log --format=%h to respect
the --abbrev option.
Reported-by: Eli Barzilay <eli@barzilay.org>
Tested-by: Eli Barzilay <eli@barzilay.org>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If $HOME is unset (as in some automated build situations),
currently
git config --path path.home "~"
git config --path --get path.home
segfaults. Error out with
Failed to expand user dir in: '~/'
instead.
Reported-by: Julien Cristau <jcristau@debian.org>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We document how to run prove with the --state option in t/README. This
produces a .prove YAML file in the current directory. Change the t/
gitignore to ignore it, and clean it up on `make clean'.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
One test case checked the stdout and stderr of 'git add' by constructing a
single 'expect' file that contained both streams. But when the command
runs, the order of stdout and stderr output is unpredictable because it
depends on how the streams are buffered. At least on Windows, the buffering
is different from what the test case expected. Hence, check the two output
texts separately.
Signed-off-by: Johannes Sixt <j6t@kdbg.org>
Acked-by: Jens Lehmann <Jens.Lehmann@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
On systems with an echo which defaults to the XSI-conformant behavior
(Solaris, or others using Ksh), echo will interpret certain backslashed
characters as control sequences. This can cause a problem for interactive
rebase when it is used to rebase commits whose commit "subject" (the first
line) contains any of these backslashed sequences. In this case, echo will
substitute the control sequence for the backslashed characters and either
the rebased commit message will differ from the original, or the rebase
process will fail. Neither is desirable.
So work around this issue by replacing the echo statements used to print
out portions of the commit message, with printf.
Also, add a test to test for this breakage.
Signed-off-by: Brandon Casey <casey@nrlssc.navy.mil>
Acked-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A test case is added but the problem can only be seen when running
the test case with --valgrind.
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Acked-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We currently do not disable diff.renames configuration while rebase
internally runs "format-patch" to feed "am -3".
The end user configuration for "diff" should not affect the result
produced by the higher level command that is related to "diff" only
because internally it is implemented in terms of it.
For that matter, I have a feeling that format-patch should not even look
at diff.renames, but we seem to have been doing this for a long time so
there is no easy way to fix this thinko.
In any case, here is a much straightforward fix for "rebase".
[jn: with test case from David]
Reported-by: David D. Kilzer <ddkilzer@kilzer.net>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This test used 5-space indents since it was added in 2005, but
recently the temptation to use tabs to indent has been too
strong, resulting in uneven whitespace. Switch over completely
to tabs.
While at it, use a more modern style for consistency with other
tests:
- names of tests go on the same line as test_expect_success;
- extra whitespace after > redirection operators is removed.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
With v1.5.3.2~14 (apply --index-info: fall back to current index for
mode changes, 2007-09-17), git apply learned to stop worrying
about the lack of diff index line when a file already present in the
current index had no content change.
But it still worries too much: for rename patches, it is checking
that both the old and new filename are present in the current
index. This makes no sense, since a file rename generally
involves creating a file there was none before.
So just check the old filename.
Noticed while trying to use “git rebase” with diff.renames = copies.
[jn: add tests]
Reported-by: David D. Kilzer <ddkilzer@kilzer.net>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Most tests in t4150 begin by navigating to a sane state and
applying some patch:
git checkout first &&
git am patch1
If a previous test left behind unmerged files or a .git/rebase-apply
directory, they are untouched and the test fails, causing later tests
to fail, too. This is not a problem in practice because none of the
tests leave a mess behind.
But as a futureproofing measure, it is still best to avoid the problem
and clean up at the start of each test. In particular, this
simplifies the process of adding new tests that are known to fail.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Place setup commands in test_expect_success blocks. This makes the
rare event of the setup commands breaking on some platform easier to
diagnose, and more importantly, it visually distinguishes where
each test begins and ends.
Instead of running test -z against the result of "git diff" command
substitution, use "git diff --exit-code", to improve output when
running with the "-v" option.
Use test_cmp in place of "test $(foo) = $(bar)" for similar reasons.
Remove whitespace after the > and < redirection operators for
consistency with other tests.
The order of arguments to test_cmp is "test_cmp expected actual".
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
These two tests weren't about how "git reflog show <branch>" exits when
there is no reflog, but were about "checkout" and "branch" create or not
create reflog when creating a new <branch>. Update the tests to check
what we are interested in, using "git rev-parse --verify".
Also lose tests based on "test -f .git/logs/refs/heads/<branch>" from
nearby, to avoid exposing this particular implementation detail
unnecessarily.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The correct advice should have been taken from c289c31 (t/t7006: ignore
return status of shell's unset builtin, 2010-06-02). A real-life issue
we experienced was with "unset", not with "export" (exporting an
unset variable may have similar portability issues, though).
Signed-off-by: Junio C Hamano <gitster@pobox.com>
test_must_fail will account for segfaults in git, so it should be used
instead of "! git"
This patch does not change any of the commands that use pipes.
Signed-off-by: Jared Hance <jaredhance@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The '!()' notation is interpreted as a pattern-list on Ksh. The Ksh man
page describe it as follows:
!(pattern-list)
Matches anything except one of the given patterns.
Ksh performs a file glob using the pattern-list and then tries to execute
the first file in the list. If a space is added between the '!' and the
open parens, then Ksh will not interpret it as a pattern list, but in this
case, it is preferred to use test_must_fail, so lets do so.
Signed-off-by: Brandon Casey <casey@nrlssc.navy.mil>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
These two lines use the negation '!' operator to negate the result of a
simple command. Since these commands do not contain any pipes or other
complexities, the test_must_fail function can be used and is preferred
since it will additionally detect termination due to a signal.
This was noticed because the second use of '!' does not include a space
between the '!' and the opening parens. Ksh interprets this as follows:
!(pattern-list)
Matches anything except one of the given patterns.
Ksh performs a file glob using the pattern-list and then tries to execute
the first file in the list. If a space is added between the '!' and the
open parens, then Ksh will not interpret it as a pattern list, but in this
case, it is preferred to use test_must_fail, so lets do so.
Signed-off-by: Brandon Casey <casey@nrlssc.navy.mil>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some have found the wording of the description to be somewhat ambiguous
with respect to when it is desirable to use test_must_fail instead of
"! <git-command>". Tweak the wording somewhat to hopefully clarify that
it is _because_ test_must_fail can detect segmentation fault that it is
desirable to use it instead of "! <git-command>".
Signed-off-by: Brandon Casey <casey@nrlssc.navy.mil>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
To make the behavior of "git submodule add" more consistent with "git add"
ignored submodule paths should not be silently added when they match an
entry in a .gitignore file. To be able to override that default behavior
in the same way as we can do that for "git add", the new option "--force"
is introduced.
Signed-off-by: Jens Lehmann <Jens.Lehmann@web.de>
Acked-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The current make target 'aggregate-results' scanned all files matching
test-results/t*-*. Normally these are only the test counts (and the
exit values, which are ignored), but with --tee the suite also dumps
all output. Furthermore, with --verbose t1450 contains several lines
starting with "broken link from ..." which matches the criteria used
by aggregate-results.sh.
Rename the counts output files to *.counts, and only scan those.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Acked-by: Sverre Rabbelier <srabbelier@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
73e25e7c (git --paginate: do not commit pager choice too early,
2010-06-26) failed to take some cases into account.
1b. Builtins that do not use RUN_SETUP (like git config) do
not find GIT_DIR set correctly when the pager is launched
from run_builtin(). So the core.pager configuration is
not honored from subdirectories of the toplevel for them.
4a. External git commands (like git request-pull) relied on the
early pager launch to take care of handling the -p option.
Ever since 73e25e7c, they do not honor the -p option at all.
4b. Commands invoked through ! aliases (like ls) were also relying
on the early pager launch.
Fix (4a) by launching the pager (if requested) before running such a
“dashed external”. For simplicity, this still does not search for a
.git directory before running the external command; when run from a
subdirectory of the toplevel, therefore, the “[core] pager”
configuration is still not honored.
Fix (4b) by launching pager if requested before carrying out such an
alias. Actually doing this has no effect, since the pager (if any)
would have already been launched in a failed attempt to try a
dashed external first. The choice-of-pager-not-honored-from-
subdirectory bug still applies here, too.
(1b) is not a regression. There is no need to fix it yet.
Noticed by Junio.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
0af0ac7 (Move MERGE_RR from .git/rr-cache/ into .git/) moved the
location of MERGE_RR but I found a few references to the old
location.
Signed-off-by: Jay Soffian <jaysoffian@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
15b4f7a (merge-tree: use ll_merge() not xdl_merge(), 2010-01-16)
introduced a regression to merge-tree to cause it to segfault when merging
files which existed in one branch, but not in the other or in the
merge-base. This was caused by referencing entry->path at a time when
entry was known to be possibly-NULL.
To correct the problem, we save the path of the entry we came in with,
as the path should be the same among all the stages no matter which
sides are involved in the merge.
Signed-off-by: Will Palmer <wmpalmer@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
merge-tree had no test cases, so here we add some very basic tests for
it, including some known-breakages.
[jc: with obvious/trivial fixups]
Signed-off-by: Will Palmer <wmpalmer@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Sometimes it is useful to know if a file or directory will be ignored
before it is added to the work tree. An example is "git submodule add",
where it would be really nice to be able to fail with an appropriate
error message before the submodule is cloned and checked out.
Signed-off-by: Jens Lehmann <Jens.Lehmann@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Make the output TAP compliant for tests skipped on request (GIT_SKIP_TESTS).
Signed-off-by: Michael J Gruber <git@drmicha.warpmail.net>
Acked-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
04ece59 (GIT_SKIP_TESTS: allow users to omit tests that are known to break, 2006-12-28)
introduced GIT_SKIP_TESTS, and since then we have had two nested loops
iterating over GIT_SKIP_TESTS with the same loop variable.
Reduce this to one loop.
Signed-off-by: Michael J Gruber <git@drmicha.warpmail.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This script is part of the second batch of tests, from the same day
the test infrastructure was added to git. Update it to use a more
modern style in the spirit of v1.6.4-rc0~45^2~2 (2009-05-22).
In particular:
- Put setup code inside test assertions, to avoid unexpected
breakages and avoid stray output without -v (as t/README
recommends); and
- Put the test title on the same line as the "test_expect_success",
and end the line with a single-quote to begin the body of the test
which is one multi-line string.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>