Work around sed and make interactions on the backslash at the end of line.
Traditionally 'i' and 'a' commands to sed have been unfriendly
with make, primarily because different make implementations did
unexpected things to backslashes at the end of lines. So work
it around by not using 'i' command.
Use $GITPERLLIB instead of $RUNNING_GIT_TESTS and centralize @INC munging
This makes the Git perl scripts check $GITPERLLIB instead of
$RUNNING_GIT_TESTS, which makes more sense if you are setting up your shell
environment to use a non-installed Git instance.
It also weeds out the @INC munging from the individual scripts and makes
Makefile add it during the .perl files processing, so that we can change
just a single place when we modify this shared logic. It looks ugly in the
scripts, too. ;-)
And instead of doing arcane things with the @INC array, we just do 'use lib'
instead, which is essentialy the same thing anyway.
I first want to do three separate patches but it turned out that it's quite
a lot neater when bundled together, so I hope it's ok.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
Perly Git: make sure we do test the freshly built one.
We could BEGIN { push @INC, '@@INSTLIBDIR@@'; } but that is not
a good idea for normal execution. The would prevent a
workaround for a user who is trying to override an old, faulty
Git.pm installed on the system path with a newer version
installed under $HOME/.
Back in the old days, we called Git's die() from the .xs code, but we had to
hijack Perl's die() for that. Now we don't call Git's die() so no need to do
the hijacking and it silences a compiler warning.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
This makes us not include ppport.h which seems not to give us anything
real anyway; it is useful for checking for portability warts but since
Devel::PPPort is a portability wart itself, we shouldn't require it
for build. You can check for portability problems by calling make check
in perl/.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
PerlIO_*() is not portable before 5.7.3, according to ppport.h, and it's
more clear what is going on when we do it in the Perl part of the Git module
anyway.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
The syntax for 'require' was wrong, and it was always failing, which
resulted in installing our own version of Error.pm anyways.
Now we used to ship our own Error.pm in the same directory, so after
fixing the syntax, 'require' always succeeds, but it does not test if
the platform has Error.pm module installed anymore. So rename the
source we ship to private-Error.pm, and install that as Error.pm when
the platform does not have one already.
Signed-off-by: Pavel Roskin <proski@gnu.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
On some platforms, Git.xs refuses to link with the rest of git
unless the latter is compiled with -fPIC, and we have USE_PIC
control in the Makefile for the user to set it. At least we
know x86-64 is such, so set it in the Makefile.
The original suggestion by Marco Roeland conservatively did this
only for Linux x86-64, but let's keep the Makefile simple and if
it breaks somebody let them holler.
Perly Git: arrange include path settings properly.
Before "use Git" takes effect, we would need to set up the Perl
library path to point at the local installation location. So
that instruction needs to be in BEGIN{} block.
Git.pm: Support for perl/ being built by a different compiler
dst_ on #git reported that on Solaris 9, Perl was built by Sun CC
and perl/ is therefore being built with it as well, while the rest
of Git is built with gcc. The problem (the first one visible, anyway)
is that we passed perl/ even various gcc-specific options. This
separates those to a special variable.
This is not really meant for an application yet since it's not clear
if it will alone help anything.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
I'm about to introduce get_object() and it will be better for consistency
if the object type always goes first. And writing 'blob' there explicitly
is not much bother.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
The code is stolen from git-annotate and completely untested since
I don't have access to any Microsoft operating system now. Someone
ActiveState-savvy should look at it anyway and try to implement
the input pipe as well, if it is possible at all; also, the implementation
seems to be horribly whitespace-unsafe.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
- We passed our own *.a archives as LIBS to the submake that runs
in perl/; separate LIBS and EXTLIBS and pass the latter which
tells what the system libraries are used.
- The quoting of preprocesor symbol definitions passed down to
perl/ submake was loose and we lost double quotes around
include directives. Use *_SQ to quote them properly.
- The installation location of perl/ submake is not
architecture neutral anymore, so use SITEARCH instead of
SITELIB.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
When perl/Makefile is stale with respect to perl/Makefile.PL, it
prevents "make clean" from completing which is quite irritating.
Fix it by calling subdirectory make clean twice as needed.
Git.pm: Add support for subdirectories inside of working copies
This patch adds support for subdirectories inside of working copies;
you can specify them in the constructor either as the Directory
option (it will just get autodetected using rev-parse) or explicitly
using the WorkingSubdir option. This makes Git->repository() do the
exact same path setup and repository lookup as the Git porcelain
does.
This patch also introduces repo_path(), wc_path() and wc_subdir()
accessor methods and wc_chdir() mutator.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
Git.pm: Implement options for the command interface
This gives the user a way to easily pass options to the command routines.
Currently only the STDERR option is implemented and can be used to adjust
what shall be done with error output of the called command (most usefully,
it can be used to silence it).
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
Currently if an external command returns error exit code, a generic exception
is thrown and there is no chance for the caller to retrieve the command's
output.
This patch introduces a Git::Error::Command exception class which is thrown
in this case and contains both the error code and the captured command output.
You can use the new git_cmd_try statement to fatally catch the exception
while producing a user-friendly message.
It also adds command_close_pipe() for easier checking of exit status of
a command we have just a pipe handle of. It has partial forward dependency
on the next patch, but basically only in the area of documentation.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
So far, errors just killed the whole program and in case of an error
inside of libgit it would be totally uncatchable. This patch makes
Git.pm throw standard Perl exceptions instead. In the future we might
subclass Error to Git::Error or something but for now Error::Simple
is more than enough.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
I have been thinking about how to do the error reporting the best
way and after scraping various overcomplicated concepts, I have
decided that by far the most elegant way is to throw Error exceptions;
the closest sane alternative is to catch the dies in Git.pm by
enclosing the calls in eval{}s and that's really _quite_ ugly.
The only "small" trouble is that Error.pm turns out sadly not to be
part of the standard distribution, and installation from CPAN is
a bother, especially if you can't install it system-wide. But since
it is very small, I've decided to just bundle it.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
Git.pm: Call external commands using execv_git_cmd()
Instead of explicitly using the git wrapper to call external commands,
use the execv_git_cmd() function which will directly call whatever
needs to be called. GitBin option becomes useless so drop it.
This actually means the exec_path() thing I planned to use worthless
internally, but Jakub wants it in anyway and I don't mind, so...
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
This patch introduces a very basic and barebone Git.pm module
with a sketch of how the generic interface would look like;
most functions are missing, but this should give some good base.
I will continue expanding it.
Most desirable now is more careful error reporting, generic_in() for feeding
input to Git commands and the repository() constructor doing some poking
with git-rev-parse to get the git directory and subdirectory prefix.
Those three are basically the prerequisities for converting git-mv.
I will send them as follow-ups to this patch.
Currently Git.pm just wraps up exec()s of Git commands, but even that
is not trivial to get right and various Git perl scripts do it in
various inconsistent ways. In addition to Git.pm, there is now also
Git.xs which provides barebone Git.xs for directly interfacing with
libgit.a, and as an example providing the hash_object() function using
libgit.
This adds the Git module, integrates it to the build system and as
an example converts the git-fmt-merge-msg.perl script to it (the result
is not very impressive since its advantage is not quite apparent in this
one, but I just picked up the simplest Git user around).
Compared to v3, only very minor things were fixed in this patch (some
whitespaces, a missing export, tiny bug in git-fmt-merge-msg.perl);
at first I wanted to post them as a separate patch but since this
is still only in pu, I decided that it will be cleaner to just resend
the patch.
My current working state is available all the time at
http://pasky.or.cz/~xpasky/git-perl/Git.pm
and an irregularily updated API documentation is at
http://pasky.or.cz/~xpasky/git-perl/Git.html
Many thanks to Jakub Narebski, Junio and others for their feedback.
Signed-off-by: Petr Baudis <pasky@suse.cz> Signed-off-by: Junio C Hamano <junkio@cox.net>
It had the wrong test for whether a commit was a merge. What it did was to
say that a non-merge has exactly one parent (which sounds almost right),
but the fact is, initial trees have no parent at all, but they're
obviously not merges.
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
The hash expansion is separated out into a function of its own, the hash
array (and size) names are made more obvious, and the code is generally
made to look a bit more like the object-ref hashing.
It also gets rid of "find_object()" returning an index (or negative
position if no object is found), since that is made redundant by the
simplified object rehashing. The basic operation is now "lookup_object()"
which just returns the object itself.
There's an almost unmeasurable speed increase, but more importantly, I
think the end result is more readable.
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
With history simplification, we still show merges that are required
to make the history _complete_, i.e. say that you had:
a
|
b
/ \
c d
| |
and neither "a" nor "b" actually changed the file, but both "c" and "d"
did: in this case we have to leave "b" around just because otherwise there
would be no way to show the _relationship_, even if "b" itself doesn't
actually change the tree in any way what-so-ever.
It would make sense to make that further simplification if the
"--parents" flag wasn't present. In that case the user is
literally asking for a list of commits and is not interested in
the relationship between them.
This patch also fixes a real bug. Without this patch, the
"--parents --full-history" combination (which you'd get if you
do something like
gitk --full-history Makefile
or similar) will actually _drop_ merges where all children are identical.
That's wrong in the --full-history case, because it means that the graph
ends up missing lots of entries.
In the process, this also should make
git-rev-list --full-history Makefile
give just the _true_ list of all commits that changed Makefile (and
properly ignore merges that were identical in one parent), because now
we're not asking for "--parent", so we don't need the unnecessary merge
commits to keep the history together.
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
If no delta is attempted on some objects then it is useless to load them
in memory, neither create any delta index for them. The best thing to
do is therefore to load and index them only when really needed.
Signed-off-by: Nicolas Pitre <nico@cam.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
upload-pack.c: <sys/poll.h> includes <ctype.h> on OpenBSD 3.8
Merlyn reports that <sys/poll.h> on OpenBSD 3.8 includes <ctype.h>
and having our custom ctype (done in git-compat-util.h which is
included via cache.h) makes upload-pack.c uncompilable. Try to
work it around by including the system headers first.
There are a few special places where some programs accessed the object
hash array directly, which bothered me because I wanted to play with some
simple re-organizations.
So this patch makes the object hash array data structures all entirely
local to object.c, and the few users who wanted to look at it now get to
use a function to query how many object index entries there can be, and to
actually access the array.
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
In the repacking window, if both objects we are looking at already came
from the same (old) pack-file, don't bother delta'ing them against each
other.
That means that we'll still always check for better deltas for (and
against!) _unpacked_ objects, but assuming incremental repacks, you'll
avoid the delta creation 99% of the time.
Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
Commit 29e4d3635709778bcc808dbad0477efad82f8d7e fixed the
underlying update-index races but git-commit was not careful
enough to preserve the index file timestamp when copying the
index file. This caused t3402 test to occasionally fail.
* js/patch:
diff.c: fix get_patch_id()
t4014: fix test commit labels.
format-patch: use clear_commit_marks() instead of some ad-hockery
t4014: fix for whitespace from "wc -l"
t4014: add format-patch --ignore-if-in-upstream test
format-patch: introduce "--ignore-if-in-upstream"
add diff_flush_patch_id() to calculate the patch id
The commit tag and commit comments used in the test claimed that
the #1 commit was merged upstream where the test actually let the
upstream merge #2 commit. Fix them.
git.c:main() relies on the value of errno being set by the last attempt to
execute the command. However, if something goes awry in handle_alias(),
that assumption is wrong. So restore errno before returning from
handle_alias().
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <junkio@cox.net>
commit does not always succeed, so we'll have to check for
it in the absence of set -e. This fixes a regression
introduced in 9e4bc7dd1bb9d92491c475cec55147fa0b3f954d
Signed-off-by: Eric Wong <normalperson@yhbt.net> Signed-off-by: Junio C Hamano <junkio@cox.net>
Indexes are only needed when we are about preparing to commit. Prime them
inside commit() when we have all the info we need, and remove all the
redundant index setups.
While we are at it, make sure that index handling is correct when opening
new branches, and on initial import.
Signed-off-by: Martin Langhoff <martin@catalyst.net.nz> Signed-off-by: Junio C Hamano <junkio@cox.net>
git wrapper: fix command name in an error message.
When the command execution by execv_git_cmd() fails with an errno
other than ENOENT, we used an uninitialized variable instead of
the string that holds the command name to report what failed.
git-svn: be verbose by default on fetch/commit, add -q/--quiet option
Slower connections can make git-svn look as if it's doing
nothing for a long time; leaving the user wondering if we're
actually doing anything. Now we print some file progress just
to assure the user that something is going on while they're
waiting.
Added the -q/--quiet option to users to revert to the old method
if they preferred it.
Signed-off-by: Eric Wong <normalperson@yhbt.net> Signed-off-by: Junio C Hamano <junkio@cox.net>
git-svn: add --follow-parent and --no-metadata options to fetch
--follow-parent:
This is especially helpful when we're tracking a directory
that has been moved around within the repository, or if we
started tracking a branch and never tracked the trunk it was
descended from.
This relies on the SVN::* libraries to work. We can't
reliably parse path info from the svn command-line client
without relying on XML, so it's better just to have the SVN::*
libs installed.
This also removes oldvalue verification when calling update-ref
In SVN, branches can be deleted, and then recreated under the
same path as the original one with different ancestry
information, causing parent information to be mismatched /
misordered.
Also force the current ref, if existing, to be a parent,
regardless of whether or not it was specified.
--no-metadata:
This gets rid of the git-svn-id: lines at the end of every commit.
With this, you lose the ability to use the rebuild command. If
you ever lose your .git/svn/git-svn/.rev_db file, you won't be
able to fetch again, either. This is fine for one-shot imports.
Also fix some issues with multi-fetch --follow-parent that were
exposed while testing this. Additionally, repack checking is
simplified greatly.
git-svn log will not work on repositories using this, either.
Signed-off-by: Eric Wong <normalperson@yhbt.net> Signed-off-by: Junio C Hamano <junkio@cox.net>
The 'graft-branches' command can now analyze tree matches for
merge detection after commits are done, when --branch or
--branch-all-refs options are used.
We ensure that tree joins (--branch and --branch-all-refs
options) during commit time only add SVN parents that occurred
before the commit we're importing
Also fixed branch detection via merge messages, this manner of
merge detection (a la git-svnimport) is really all fuzzy, but at
least it actually works now :)
Add some new tests to go along with these fixes, too.
Signed-off-by: Eric Wong <normalperson@yhbt.net> Signed-off-by: Junio C Hamano <junkio@cox.net>
Tested on a plain Ubuntu Hoary installation
using subversion 1.1.1-2ubuntu3
1.1.x issues I had to deal with:
* Avoid the noisy command-line client compatibility check if we
use the libraries.
* get_log() arguments differ (now using a nice wrapper from
Junio's suggestion)
* get_file() is picky about what kind of file handles it gets,
so I ended up redirecting STDOUT. I'm probably overflushing
my file handles, but that's the safest thing to do...
* BDB kept segfaulting on me during tests, so svnadmin will use FSFS
whenever we can.
* If somebody used an expanded CVS $Id$ line inside a file, then
propsetting it to use svn:keywords will cause the original CVS
$Id$ to be retained when asked for the original file. As far as
I can see, this is a server-side issue. We won't care in the
test anymore, as long as it's not expanded by SVN, a static
CVS $Id$ line is fine.
While we're at making ourselves more compatible, avoid grep
along with the -q flag, which is GNU-specific. (grep avoidance
tip from Junio, too)
Signed-off-by: Eric Wong <normalperson@yhbt.net> Signed-off-by: Junio C Hamano <junkio@cox.net>
- combine_diff() took cnt (count) which is unsigned in nature but the
parameter type was declared as "int";
- find_next() took "uninteresting" parameter, which masked a static
function of the same name;
- show_parent_lno() took an unused parameter "cnt";
- show_patch_diff() used a local variable in nested inner scope with
the same name with different type, masking the one in the outer scope;
- the last loop in show_patch_diff iterated over lines so it should use
the local variable "lno"
* ew/rebase:
rebase: allow --skip to work with --merge
rebase: cleanup rebasing with --merge
rebase: allow --merge option to handle patches merged upstream
pkt-line.h uses GCC's __attribute__ extension but does not include
git-compat-util.h. So it will not compile with a compiler that does
not support this extension.
git-repack: Be careful when updating the same pack as an existing one.
After a clone, packfiles are read-only by default and "mv" to
replace the pack with a new one goes interactive, asking if the
user wants to replace it. If one is successfully moved and the
other is not, the pack and its idx would become out-of-sync and
corrupts the repository.
Recovering is straightforward -- it is just the matter of
finding the remaining .tmp-pack-* and make sure they are both
moved -- but we should be extra careful not to do something so
alarming to the users.
This lets you use something like this in your $GIT_DIR/config
file.
[diff]
color = auto
[diff.color]
new = blue
old = yellow
frag = reverse
When diff.color is set to "auto", colored diff is enabled when
the standard output is the terminal. Other choices are "always",
and "never". Usual boolean true/false can also be used.
The colormap entries can specify colors for the following slots:
plain - lines that appear in both old and new file (context)
meta - diff --git header and extended git diff headers
frag - @@ -n,m +l,k @@ lines (hunk header)
old - lines deleted from old file
new - lines added to new file
We no longer have to recommit each patch to remove the parent
information we're rebasing since we're using the low-level merge
strategies directly instead of git-merge.
Signed-off-by: Eric Wong <normalperson@yhbt.net> Signed-off-by: Junio C Hamano <junkio@cox.net>
git-commit: filter out log message lines only when editor was run.
The current behaviour strips out lines starting with a # even when fed
through stdin or -m. This is particularly bad when importing history from
another SCM (tailor 0.9.23 uses git-commit). In the best cases all lines
are stripped and the commit fails with a confusing "empty log message"
error, but in many cases the commit is done, with loss of information.
Note that it is quite peculiar to just have "#" handled as a leading
comment char here. One commonly meet CVS: or CG: or STG: as prefixes, and
using GIT: would be more robust as well as consistent with other commit
tools. However, that would break any tool relying on the # (if any).
Signed-off-by: Yann Dirson <ydirson@altern.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
This cleans up the use of safe_strncpy() even more. Since it has the
same semantics as strlcpy() use this name instead. Also move the
definition from inside path.c to its own file compat/strlcpy.c, and use
it conditionally at compile time, since some platforms already has
strlcpy(). It's included in the same way as compat/setenv.c.
Signed-off-by: Peter Eriksen <s022018@student.dtu.dk> Signed-off-by: Junio C Hamano <junkio@cox.net>
apply: replace NO_ACCURATE_DIFF with --inaccurate-eof runtime flag.
It does not make much sense to build git whose behaviour is
different depending on the brokenness of diff implementation of
the platform because the brokenness of the patch that is applied
with the tool depends on brokenness of the diff the person who
generates the patch uses. So we do not use NO_ACCURATE_DIFF
anymore, but help people to apply patches that do not record
incomplete lines correctly with a runtime flag.
Signed-off-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Junio C Hamano <junkio@cox.net>