Rather than drawing our own toplevel for error messages we really
should just use the the native tk_messageBox command to display
any error messages.

Major benefits for doing so are:
  - automatically centers over main window;
  - less code required on our part in git-gui;
  - includes a nifty error icon on most systems;
  - better fits the look-and-feel of the operating system.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

I found difffont to be a very awkward varible name, due to the use
of three f's in a row.  So use easier to read variable names.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

When we reshow the current diff file it can be faster to just fetch
the value from the file_states array than it is to ask for all paths
whose name exactly matches the one we want to show.  This is because
[array names -exact] is O(n) in the number of files.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

When we commit we know that whatever was in the index went as part
of the commit.  Since we generally assume that the user does not
update the index except through our user interface we can be reasonably
certain that any file which was marked as A/M/D in the index will have
had that A/M/D state changed to an _ (not different) by the commit.

We can use this knowledge to update the user interface post commit
by simply updating the index part of the file state of all files whose
index state was A/M/D to _ and then removing any file memory any which
wound up with a final state of __ (not different anywhere).  Finally we
redraw the file lists and update the diff view.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

If we are given a file whose path name contains an LF (\n) we now
escape it by inserting the common escape string \n instead of the
LF character whenever we display the name in the UI.  This way the
text fields don't start to span multiple lines just to display one
file, and it keeps the line numbers correct within the file lists.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

When we did a rescan to update the file lists we lost the tag which
indicated which file was currently in the diff viewer.  This occurs
because we delete every line from the two file list boxes (thus
removing the tag) and then redisplay the diff in the diff viewer
but then fail to restore the tag in the file list.

Now we restore that tag by searching for the file in the file lists
and adding the tag back when the diff viewer displays something.

We also no longer obtain the file path directly from the file list
text box.  Instead we now keep two Tcl lists, one for each file list,
holding the file names in sorted order.  These lists can be searched
with the native [lsearch -sorted] operation (which should be faster
than our crude bsearch) or can be quickly accessed by index to return
the file path.  This should help make things safer should we ever be
given a file name which contains an LF within it (as that would span
two lines in the file list, not one).
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

If we load a message file (e.g. MERGE_MSG) or we have just finished
making a commit and are clearing out the commit buffer we should also
clear out the undo/redo stack associated with that buffer.  The prior
undo/redo stack has no associated with the new content and therefore
is not useful to the user.

Also modified the sign-off operation to perform the entire update in
a single undo/redo operation, allowing the user to undo the signoff
in case they didn't actually want to do that.

I also noticed what may be a crash on Windows related to the up and
down arrow keys navigating within the diff viewer.  Since I got back
no stack trace (just an application exit with a loss of the commit
message) I suspect that the binding to scroll the text widget crashed
with an error and the wish process just terminated.  So now we catch
(and ignore) any sort of error related to the arrow keys in the diff
viewer.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Users have come to expect basic editing features within their
applications, such as cut/copy/paste/undo/redo/select-all.  I
found these features to be lacking in git-gui so now we have
them.

I also added basic keyboard bindings for the diff viewing area
so that the arrow keys move around single units (lines or columns)
and the M1-X/C keys will copy the selected text and the M1-A key
will select the entire diff.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Now that we call the update-index all files action "Include All" it
makes more sense to make this M1-I (so Control-I or Command-I depending
on platform) than M1-U, which stood for update but is somewhat confusing
to users.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

I started to find it very annoying that my test application kept
opening at the wrong location on my desktop, so now we save the
basic window geometry and sash positions into the config file as
gui.geometry.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Caching the Signed-Off-By line isn't very important (as its not
performance critical).  The major improvement here is that we
now report an error to the user if we can't obtain their name
from git-var.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

There's really no great reason to show the entire commit object id
within the GUI, especially if the user is unable to copy and paste
it into another interface such as gitk or a terminal window.  So
we'll just show them the first 8 digits and hope that is unique
within their repository.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

At least one user was confused by the term 'check-in'; he thought that
clicking that button would commit just that one file, but he wanted to
include all modified files into his next commit.

Since git doesn't really have a check-in concept this really was poor
language to use.  Git does have an update-index concept but that is a
little too low level to show to the user.

So instead we now talk about including files in a commit.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

When a file jumps between the file lists due to its state changing we
crashed thanks to a stale variable reference within the procedure as we
tried to setup the new icon.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

If there are modified files present in the working directory then we
should not let the user perform a pull as it may fail due to the modified
files being uncommitted but needing to be merged at the file level.

Yes there are many cases where a merge will complete successfully even
though there are modified or untracked files sitting in the working
directory.  But users generally shouldn't be attempting merges like that,
and if they are they probably are advanced enough to just use the command
line and bypass this little safety check.

We also no longer run a rescan after a successful pull has completed.
Usually this is unnecessary as a successful pull won't leave modified
files laying around.  Instead we just update our HEAD and PARENT values
with the new commit, if there is one.

Unfortunately this does let the user get into an insane state as there
are bugs in core Git's git-pull and git-merge programs where the exit
status is sent back as a 0 rather than non-0 when a failure is detected.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

If we have the index locked then no pull command is allowed to proceed
(as it would fail to get the index lock itself).  So disable the pull
menu items when we are doing any index based operations.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

The user must not modify the index while a git pull operation is running,
doing so might cause problems for the merge driver and specific strategy
being used.  Normally on the command line people are just really good and
don't try to run index altering operations while they are also running a
pull.  But in a slick GUI like git-gui we can't trust the user quite as
much.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

On very large projects (~1000 files) on Windows systems the update-index
--refresh stage of the rescan process takes a nontrival amount of time.
If the user is generally very careful with their file modification such
that the modification timestamp of the file differs only when the content
also differs then we can skip this somewhat expensive step and go right
to the diff-index and diff-files processes.

We save the user's prefernce in the current repository if they modify the
setting during a git-gui session, but we also load it through our dump of
repo-config --list so the user could move it to their ~/.gitconfig file
if they wanted it globally disabled.

We still keep update-index --refresh enabled by default however, as most
users will probably want it.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Most users of git-gui probably shouldn't be invoking git repack directly;
instead we should be looking at how many loose objects they have and
how many active packs they have and making the decision for them.  But
that's more work to code, and there's always going to be discussion about
what is the right default threshold and how do we know that the user is
willing to do the repack when we decide its time, etc.

So instead we'll just keep it simple and offer up the menu option.

Unfortunately right now we get now progress indication back from
git-pack-objects as its being invoked not through a tty, which makes
it disable progress output and the git-repack.sh wrapper won't let us
pass through --progress.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Since Tk will only supply new space gained from growing the top level to
the bottom/right most widget within a panedwindow and most users will be
growing a git-gui main window for the purposes of seeing more of the
currently shown diff, flipping the order around makes Tk do what the
user wants by default.

Of course because we also removed the paned window from the commit buffer
area it is now impossible to increase the visible space for the commit
message.  But I don't see this as a huge concern right now as its actually
very awkward to try and balance three paned window dividers within the
same top level window.  We could always add it back if users really want
to expand the commit buffer and see more.

I also corrected a number of bugs that I accidentally introduced in the
last commit.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Removed as much as possible from the merge_state proc, which is where
we spent most of our time before UI update.  This change makes our
running time match that of git status, except that we then need about
7 additional seconds to draw 6900 files on screen.

Apparently the [array names a -exact $v] operator in Tcl is O(n) rather
than O(1), which is really quite disappointing given that each array can
only have one entry for a given value.  Switching to a lookup with a
catch (whose error we ignore) runs in O(1) time and bought us most of
that improvement.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Loading 6900 newly added files required about 90 seconds on one system.
This is just far too long to perform a "status" type of operation.
git-status on the same system completes in just 8.2 seconds if it is
redirected to /dev/null.

Most of our performance improvement comes from moving all of the UI
updating out of the main fileevent handlers for the status process.
Instead we are only updating the file_states array and then only doing
the UI update when all states are known and have been finally determined.

The rescan execution is now down to almost 30 seconds for the same case,
a good (but not really all that impressive) improvement.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

When we were receiving a lot of output from diff-index we split records
at the wrong locations and started using the file status information
(mode and SHA1s) as path names, and then proceeded to try to use part
of the path names as status data.  This caused all sorts of havoc.

So I rewrote the parsing implementation to scan for the pair of null
bytes along the buffer and stop scanning (waiting for more data) if
both can't be found during this event.  This seems to work well under
high load (like when processing 6,983 added files).
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

We now create one menu entry per remote listing the first Pull: or fetch
entry associated with that remote as the branch to pull into the current
branch.

This is actually quite incorrect as we should be using the default
remote branch name listed in branch.<name>.merge for a new-style remote
described in the config file.  But its a good default to get started with.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

We're likely going to need key/value pairs from the repo-config beyond
just remote.*.url, so cache them all up front into a Tcl array where we
have fast access to them without needing to refork a repo-config --list
process.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

If the user closes a console and we get more ouptut for it then we
will get a Tcl error in the readable event handle for the file channel.
Since this loses the actual output and is quite unfriendly to the end
user instead reopen any console which the user closed prior to the
additional output arriving.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

The user might be using the new style config syntax remote.name.url
rather than the older standalone remote file.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Because the remote end is likely to send us progress meters by
resetting each line with a CR (and no LF) we should display those
meters by replacing the last line of text with the next line,
just like a normal xterm would do.

This makes the output of fetch look about the same as if we ran it
from within an xterm.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Apparently accelerators really only work correctly for function keys
(F1-F12) and "Cmd-q".  Apparently wish on Mac OS X reports itself
as unix and the OS is Darwin, this makes it a little difficult to
be sure we are running under Aqua.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Because we cd after getting the cdup value from Git we can't try
to get the gitdir until after we perform the cd, as usually the
gitdir is relative to the current working directory.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Apparently the Cygwin tclsh/wish executables don't pass the environment
that they inherited onto any children that they invoke.  This causes a
problem for some users during 'git fetch' or 'git push' as critical
environment variables like GIT_SSH and SSH_AUTH_SOCK aren't available
to the git processes.

So we work around this by forcing sh to start a login shell, thus
reloading the user's environment, then cd to the current directory,
and finally start the requested process.  Of course this won't
correctly handle any transient environment variables that were
inherited but were not supplied by the user's login shell.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

Signed-off-by: NShawn O. Pearce <spearce@spearce.org>

 * Make sure we are in the top level working directory.  This
   way we can access files using their repository path.

 * Reload the diff viewer if the current file's status has changed;
   as the diff may now be different.

 * Correctly handle the 'AD' file state: added but now gone from
   the working directory.
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>