1.0.19.22: fix bug #425

[sbcl/tcr.git] / doc / GIT-FOR-SBCL-HACKERS.txt

Git For SBCL Hackers
====================

Git is structured as a ton of programs called git-foo, accessible
either directly, or as "git foo". Commands "git help" and "git help
foo" provide the documentation -- which is there also as the manual
pages.

Make sure you have Git 1.5.something.

When running on tty the commands pipe their output through less
automatically, so you don't need to mentally add "| less" anywhere.

Let's get started. First off, we need a gitified SBCL repository to
clone. At the time of writing there are two Git mirrors for the CVS
history:

 git://sbcl.boinkor.net/sbcl.git
 
and

 git://repo.or.cz/sbcl.git

but the latter is actually a mirror of the first one, so it doesn't
matter which one you use, really.

The command

 git clone git://sbcl.boinkor.net/sbcl.git sbcl-git

clones the SBCL Git mirror into the directory sbcl-git (there's a
naming convention in play here, but ignore that for now.) The clone
contains full history, and is an independent repository on its own
right. Doing the clone takes as long as downloading 25Mb takes on your
line, but after that things are very fast.

To update your clone from the original source at a later date, use
git-pull. Go ahead and try out now, to see that it works, and see how
snazzy a no-op pull is.

 git pull

The directory .git inside the clone contains Git's view of the
repository -- no other Git files or directories are in the tree. Just
for kicks:

 cd sbcl-git
 rm -rf *                # leaves .git alone
 time git reset --hard   # this takes < 2 seconds, restores everything

So, two commands so far:

 git-clone
   One-time operation to clone a repository.

 git-reset
   Used to restore state: using the --hard flag resets the entire tree
   to the state of the last commit but doesn't delete new files which
   may be in the tree. (It doesn't actually "restore state", or rather
   does it as a side, effect, but you can learn the details on your
   own.)

Let's hack a bit.

 git branch hack-a-bit   # creates a new branch called "hack-a-bit"
 git branch              # shows all branches and tells which one is active
 git checkout hack-a-bit # switches to the "hack-a-bit" branch
 git branch              # shows that we're now on "hack-a-bit"

Once you get your bearings you can do the create-branch-and-checkout
in one step, which you will be doing a lot, since branches are really
nice in Git! The magic is: "git checkout -b hack-a-bit-more" -- but
leave that for another time.

Let's do something. Edit BUGS, and maybe give SBCL some love.

 git diff                # shows you the changes in your tree
 git status              # shows you the state of files in the tree

git-status will tell you that some files have been modified, but it
has "no changes added to commit". If you tried git-commit right now,
nothing would happen. What's going on?

Git has a notion of a separate "staging area", from which commits are
made. You can add content to the it by using git-add:

 git add BUGS
 git status
 git diff

Now git-status shows changes as part of a pending commit, but git-diff
is silent!

By default git-diff shows the differences between the working tree and
the staging area (which starts out identical to HEAD after a checkout).

Edit BUGS again. Now you have three versions of it (ignoring all the
historical versions for a second) to compare:

 git diff               # between tree and staging area
 git diff HEAD          # between tree and last commit
 git diff --cached      # between staging area and last commit

If we were to do a git-commit now, it would commit the version in the
staging area, not the one in the tree.

We like our latest version, so do

 git add BUGS

again. Now the latest version is in the staging area, and version that
used to be there is gone. You don't need to worry about the staging
area. Either you will find out that it's pretty neat (consider it a
temporary commit), or you can mostly ignore it. Do

 git commit

now, and we'll move on in a second. Just to call spade a spade, the
staging area is really called "index".

Now, our changes are in the repository, and git-status should show a
clean tree. (By the way, can you imagine how long all the diffs and
statuses you've done during this tutorial would have taken with CVS?)

To get the differences between your current branch, and the master
branch, do

 git diff master

To view the last commit on master, do

 git diff master^..master

To view the commit logs, do

 git log

You'll see long hex-string at the head of each commit. This is the
commit id, which is a SHA1 hash based on the content of the tree, so
you can share these with other hackers, and they can use the same
commit id to poke at their own repositories. Locally any unique prefix
of the hash is enough to identify a commit, so you don't need to use
the full hashes.

This is where the fun starts. Pick an interesting looking commit a
while back, copy the commit id, and do

 git diff <<paste commit id here>>

Git will show you all the changes between the commit you selected and
current version as a single diff. Similarly,

 git diff -w <<commit id 1>> <<commit id 2>>

can be used to compare two arbitrary versions. The -w switch tells Git
to ignore whitespace changes -- you can usually leave it out, but it's
nice when diffing across the great whitespacification patches.

Onwards: just so that we have a bit more history on the branch, edit
BUGS again, and git-commit again. You can use

 git commit -a

to automatically add all the changed files to the staging area to
speed things up. Repeat. Now git-log should show a few local changes.

Let's see how merging works. To create a new branch based on master
and switch to it do

 git checkout -b merge-experiment master

Merge the changes we did on the hack-a-bit branch.

 git merge hack-a-bit

Bing-bada-bing! Done. Have a look at git-log. You see that we have
full branch history merged. If there had been conflicts, the merge
would not have been automatic, but you would have been called to
resolve conflicts, etc.

This is very nice for merging short-lived local branches, but not
always so good for merging things back onto master, or into "mainline"
history since you get all the commits that happened on the branch:

  "first cut at x"
  "first cut at y"
  "oops, x was wrong"
  "implemented z"
  "aargh, x was still wrong"
  ...

When merging a branch like this, with oopses in the history, it is far
better if we are able to merge it as a few logical patches -- the
resulting history will be easier to understand.

First option is to use --squash to squash all the commits into a
single one.

 git checkout -b merge-experiment-2 master
 git merge --squash hack-a-bit

This has the side-effect of not committing the changes immediately.

This is in effect similar to the usual way of merging back changes
from a CVS branch: if we're told that the patch came from a branch, we
can then go and look at the branch history, but we don't see the
evolution of the branch in the mainline history.

If the changeset is small, --squash is exactly what we want, but for
long-lived branches that are best merged in a few steps we can use

 git merge --squash <<commit id>>

to merge the changes upto a certain commit. Repeat a few times, and
you have the whole branch merged. Other Git commands provide more
advanced options. See eg. git-cherry-pick, which you will end up
using sooner or later.

Now, let's assume we have a private Git repository in ~/sbcl-git, and
we want to publish it to the world. The easiest way is to fire up a
browser, and go to

 http://repo.or.cz/w/sbcl.git

There, click on the "fork" link, and set up your own SBCL repository,
called sbcl/yourname. Select the "push" mode, and write something
along the lines of "My private SBCL tree. Pulls from main sbcl.git
repository, and trickles changes back to upstream CVS manually --
from where they end up in sbcl.git. Turtles, you see." in the comment
box. Then you will be directed to set up an account, which you will
then have to add as a "pusher" to your SBCL fork. 

Finally, add the following snippet (adjusting for your own name) in
~/sbcl-git/.git/config

 [remote "public"]
        url = git+ssh://repo.or.cz/srv/git/sbcl/yourname.git

After that is done, you're ready to publish your tree.

 git push --all public

Now anyone can get at your repository at

 git://repo.or.cz/sbcl/yourname.git

and you can publish your own hacks on it via git-push.

 git push public <<name of branch to update>>

Since we're not (yet?) officially using Git, we want to get our
changes back into the CVS. git-cvsexport is the perfect tool for this.
This assumes that you have a developer CVS checkout in ~/sbcl-cvs, and
wish to commit the changes you have wrought on branch foo-hacks
(compared to master):

 git checkout -b foo-hacks-to-cvs master
 git merge --squash foo-hacks
 edit version.lisp-expr to be "CVS ready"
 git commit -a               # edit the message to be "CVS ready"
 cd ~/sbcl-cvs
 GIT_DIR=~/sbcl-git/.git git cvsexportcommit -v foo-hacks-to-cvs
 review, fix any problems
 cvs commit -F .msg

To make things a bit easier, add eg. this stanza to ~/sbcl-git/.git/config:

 [alias]
    sbcl-export = ! cd ~/sbcl-cvs && GIT_DIR=~/sbcl-git/.git git-cvsexportcommit -v

Then you can just run

 git sbcl-export foo-hacks-to-cvs

from inside ~/sbcl-git/, and have it prepare your CVS tree for commit.

git-cvsexportcommit is fairly conservative by default, and will fail
if the patch doens't apply cleanly. If that happens, you can fix the
issues manually:

 .cvsexportcommit.diff   -- holds the patch
 .msg                    -- holds the commit message

Finally, delete the foo-hacks-to-cvs branch after you've committed
code to CVS. Of course, instead of using git-cvexportcommit you can
also manually make and apply patches, etc. For hairier cases it may
even be easier in the end:

  git format-patch -p master..foo-hacks

Generates a patch for each commit between master and the HEAD of
foo-hacks, naming them

  0001-first-line-of-commit-message.patch
  0002-and-so-and-so-forth.patch
  ...

Due to, among other things, the cvs->git synchronization lag it is
easy to get conflicts on version.lisp-expr so you may want to delay
finalizing version.lisp-expr and the .msg file that cvsexportcommit
produced by bumping the version and adding it to the commit message.

To get latest changes from the CVS Git mirror you originally cloned
from, do

  git pull

on the branch you want to update on your private repository.

This completes our whirlwind tour. I'm not saying this makes you
proficient in using Git, but at least you should be up and walking.
Reading

 http://eagain.net/articles/git-for-computer-scientists/
 http://www.kernel.org/pub/software/scm/git/docs/everyday.html
 http://www.kernel.org/pub/software/scm/git/docs/user-manual.html
 
and various Git manual pages is a good idea.

One command I can in particular recommend getting familiar with is
git-rebase. git-gui, git-citool, and gitk provide graphical interfaces
for working with Git -- particularly gitk is an invaluable tool for
visualizing history.