|Version 24 (modified by nominolo, 8 years ago) (diff)|
- Git for Darcs Users
- Git Concepts
- General Settings
- Git Overview
- darcs init
- darcs get
- darcs put
- darcs add
- darcs record
- darcs pull
- darcs push
- darcs changes
- darcs tag
- darcs whatsnew
- darcs diff
- darcs revert
- darcs unrecord
- darcs unpull
- darcs amend-record
- darcs rollback
- darcs annotate
- General Notes
Git for Darcs Users
Just like Darcs, every Git command comes with a --help option. For example git add --help. You can also check out the official Git documentation.
Also see "General Notes" below for features present in Git but not in Darcs.
In order to understand some commands it is important to compare both Git's and Darcs' internal model. A Darcs repository is a collection of patches with some dependencies between them. The working directory is what you get when you apply all those patches in a valid order.
Git on the other hand tracks states of the working directory. Each commit refers to a particular "version" of the working tree. So far, this is just a different view of the same thing, in fact, Git internally stores some sort of diffs for space efficiency. The important difference is that a Git commit also remembers which version we make our patch against, i.e., the parent of the new commit. Several commits can have the same parent or multiple parents, in which case they are merges. As a result, a Git repository forms a directed acyclic graph. These are often depicted in ASCII art like this:
o---o---o---o / \ o---o---A---o---o---o---B---o
After commit "A" two developers (or the same developer in different branches) performed different commits on top of the same original version. The "B" commit is a merge. If the two branches contained conflicting changes "B" would contain its resolution.
A branch in Git is now merely a pointer to a commit. For example, a typical situation is this:
o---o---o <-- feature1 ; o---o <-- develop ; o---o---A---o---o---o <-- master
Here we have three branches "feature1", "develop", and "master". "master" is the default branch that is created automatically, when you initialise a new Git repository. Some commands also default to the master branch for certain actions, but otherwise it is not special in any way.
Note that Git branches all exist in the same repository. You can have several physical clones of the same repository like in Darcs, but it is often more convenient to work with multiple branches in the same directory. The most important commands for working with branch are git checkout and git branch, see below.
The typical Git workflow is to use one branch per feature and merge into their original branch once they are considered ready. For example, if we think ready that "feature1" is ready, we can git merge it into master:
$ git checkout develop # switch to develop branch $ git merge feature1 # merge feature1 into current branch o---o---o---o---o <-- develop, feature1 ; o---o---A---o---o---o <-- master
The above merge is actually a fast-forward, meaning that no new commit is necessary, since "develop" didn't contain any other changes--the pointer for the "develop" branch is merely set to point to the same commit that feature1 pointed to. On the other hand, if we now merge "devel" into "master" we get a now commit:
$ git checkout master $ git merge devel o---o---o---o---o <-- develop, feature1 / \ o---o---A---o---o---o-------o <-- master
XXX: talk about remote/origin/master, fetch, and how it's mostly just already known merge
First of all, git rebase is a very dangerous feature, it should never be done in shared repositories (this is like darcs amend-record and darcs unpull, but even harder to fix.)
git rebase is a way to "rewrite history". You can use this to remove bad patches or to move a branch. For example, consider the following history:
A---B---C---D <-- feature1 ; o---o---o---o---o <-- master
Now, it turns out that B was not a good patch. With git rebase, you can choose to just remove B completely (similar to darcs unpull)
A---C'---D' <-- feature1 ; o---o---o---o---o <-- master
or you could edit B and replace it with another change B' (similar to darcs amend-record).
A---B'---C'---D' <-- feature1 ; o---o---o---o---o <-- master
The easiest way is to use rebase for these use cases is via git rebase -i (see its documentation).
Note in the above examples how patches C and D got renamed to C', D'. This indicates that they now have different commit ids because they are now are the same patches but based upon a different history. That means that another user that already has the original patches C and D and tries to update from the rebased branch will get conflicts! (This is comparable to how an amended patch in darcs will lead to conflicts, if someone else already has the original patch in the repository.)
Therefore: Never rebase a published branch.
Another use case is to move a branch. In the above example we can rebase feature1 onto master, resulting in this history:
A'---B'---C'---D' <-- feature1 ; o---o---o---o---o <-- master
You can do this if the changes in the feature branch semantically aren't really a branch, but make equally sense on top of the master's history. In the above example, merging feature1 into master will now be a simple fast-forward and will not introduce a merge commit.
Finally, rebase can be used to move a feature branch onto another branch. For example, say we have a big feature and we want to implement a smaller feature:
o---o---o <-- small_feature ; o---o <-- big_feature ; o---o---A---o---o---o <-- master
Now we realise that small_feature is already useful and want to merge it into master without the commits that were made on the big_feature branch. We can therefore rebase small_feature onto master:
$ git rebase --onto master big_feature small_feature o---o <-- big_feature ; | o'--o'--o' <-- small_feature | / o---o---A---o---o---o <-- master
See git rebase --help for more usage information and more examples.
Uh-oh, I trashed my repo with a rebase, what do I do?
- Recover from the backup copy (you did make backup copy before the rebase, didn't you?)
- In case you were in a hurry and option 1 is not an option for you, there may be some hope. Git actually has a bit of a functional philosophy in that it doesn't immediately throw away the orignal commits. Git uses garbage collection for this, which is called automatically every once in a while.
Suppose we started with this repository state
A---B---C---D <-- feature1 ; o---o---o---o---o <-- master
and decided to remove B from our history. The actual repository now looks like this:
C'--D' <-- feature1 ; A---B---C---D ; o---o---o---o---o <-- master
If you happen to have the commit id of D (maybe in your console backlog) you can create a branch feature1_old that points to D via
$ git checkout -b feature1_old <commit-id-of-D>
Now you have a handle to both D' and D.
Just like Darcs, Git has global and per-repository configuration options. To globally set your committer name and email use
git config --global user.name "Haskell Curry" git config --global user.email email@example.com
Since Git keeps several branches in the same repository it is very useful to show the current branch in your shell prompt.
For an overview of what repositories (or parts of repositories) are modified by various git commands:
The git "local repository" corresponds to darcs internal store of patches. The git "workspace" corresponds to your normal code tree in darcs. The git "index" is a staging area within the current repository, for storing partially-committed hunks. It has no equivalent in darcs.
git clone <repo-url> [<local-name>]
Possible repo URLs look like this:
git clone http://darcs.haskell.org/ghc.git # via HTTP (slowest) git clone git://darcs.haskell.org/ghc.git # git's protocol (fast, read-only) git clone [username@]darcs.haskell.org:ghc.git # via SSH
There's no default command to do that, but the following should work:
ssh me@remote cd /some/directory.git git init --bare exit cd my/local/repo git push me@remote:/some/directory
The --bare option disables a checkout of the working copy, i.e., only the contents of the .git directory are stored.
Note: If the repository is supposed to be shared by several users, it's best to init it with either of these commands:
git init --bare --shared=group # everyone in the same group can read and write git init --bare --shared=all # same group can read/write, others can read
You can also set this after the fact, by setting the configuration variable core.sharedRepository. See git config --help for more information.
git add <dir-or-file>
Git supports interactive recording very similar to darcs.
git add -p
git add -i
The main difference is that Git does not automatically commit the changes. You have to do that manually using
git commit [-m "commit message"]
If you do not supply a commit message, it will open your default editor. If you want to abort the commit, use an empty commit message.
To see what will be committed, use
git diff --cached
Tip: If you want to see the diff when you edit the commit message, use
git commit -v
darcs record -a
git commit -a
This will add and commit all (not ignored) files. It will not add newly created files. (To do this call git add . before in the repo root directory.)
There is no direct mapping for the interactive darcs pull. Cherry-picking is not as streamlined as in Darcs.
darcs pull -a
Here is how you update (everything) from the source repo:
If all you want to do is to keep updated then this is fine. The above is actually a shortcut for
git pull origin
where origin is the name of your default remote branch. (You can name it as you like, but certain Git commands will use origin as the default if no argument is specified.)
XXX: will this pull into the current branch, or always into master? (Websites suggest always into master, so you likely need to follow git pull with git rebase <branch-name>)
Like in Darcs, you may get conflicts. To resolve conflicts, edit the conflicting file, git add it, and git commit it.
If you want to see whether you get conflicts before pulling git pull is actually ... XXX
Selectively pushing patches is not available directly in Git. git push does the same as darcs push -a.
A comparable interactive workflow is to merge a selection of patches from a local branch into the local master branch and then git push that.
In general, even though a central repository is possible, Git promotes a pull model. That is, to work on a project you typically "fork" (git clone) the source repository, add your changes, publish your repository, and send a pull-request to the upstream maintainer. The reasoning behind that is that you don't have something akin to a list of committers, but rather the maintainer has a set of trusted peers. This model is very different than what seems to be common among darcs users, but it has its advantages.
Obviously, this requires that it's made easy to publish your version of the repository easily. This is where websites like GitHub come into play. GitHub is free for open source projects (it offers a paid service with private repos), and makes it particularly easy to share with Git. GitHub automates things like forking and sending pull requests. GitHub has a quota of 100 MB, but a forked repository will not count on your quota. This is particularly useful for large code bases like GHC. (The GitHub quota isn't always correct; so if it seems wrong check again the next day.)
darcs push is also used to exchange patches between local repositories. See "Local Branches" below for how to work with branches in Git.
Of course, you need to be able to publish your local changes to a remote repo (even if it's not the main repo). This is done using git push which is largely equivalent to darcs push -a
darcs push -a
git push [<repo-url-or-alias>]
Without argument this will push to your origin.
git log git log <file-or-directory>
darcs changes --last <N>
git log -n <N>
darcs changes --summary
git log --stat
darcs changes --match
git log --grep="something"
(the =-sign is important)
Other useful variants
git log -p
Shows the patch for each commit.
git grep <text>
Look for something anywhere in the repository's history (tag names, commit messages, file contents).
git show <commit-id>
Show the changes by the given patch
git log v2.5..v2.6 # commits between v2.5 and v2.6 git log v2.5.. # commits since v2.5 git log --since="2 weeks ago" # commits from the last 2 weeks git log v2.5.. Makefile # commits since v2.5 which modify # Makefile
See git log --help for a lot of extra options, to refine the output.
git tag <tagname>
This will fail if the tag already exists. If you want to move an existing tag use git tag -f <tagname>, but never move a tag in a public repo/branch. Use this only on local branches, and only if the tag exists nowhere else. git tag --help contains a discussion of this.
git diff <commit1>..<commit2> # show diff between two commits
Not sure if this gives you fine-grained reversion of individual hunks:
git reset --hard
Note: git reset only resets the staged files, i.e., the things added with git add.
Not sure if this will unrecord any patches except the most recent???
git reset --soft HEAD^
It is not easy to amend any patch except the last one committed.
If the change to be amended is the latest commit
git commit --amend
TODO describe workflow if amended patch is not the current HEAD.
TODO add note for merge commits
git revert <commit-id>
Working directory must be clean. (You can use git stash to save local changes).
git branch git branch <name> git branch -b <name> git checkout git branch -d <name> git branch -D <name> git stash git show-branch
git pull git fetch git merge
- feature branches
- git rerere
Fix a bug
git pull <upstream> # get latest changes git checkout -b fix_bug # start a branch to fix the bug # ... hack ... git add -i # select the proper changes git diff --cached # verify what will be committed git commit # ... test ... oops, forgot something git add -i # add the new patches git commit --amend # add them to previous commit # ... test ... looks fine now git checkout master # back to main branch git pull # make sure it's up to date git merge fix_bug # merge in our local changes # ... if we get a conflict here, edit the file then # ... add the changed files with git add, then git commit git push # push changes to personal public repo # or directly to <upstream> git branch -d fix_foo # delete the branch we no longer need
Yes, it is a bit more complicated than using darcs to do the same thing:
darcs pull <upstream> # get latest changes # ... hack ... darcs record # select the proper changes # ... test ... oops, forgot something darcs amend-record # add the new patches # ... test ... looks fine now darcs pull # make sure it's up to date # ... if we get a conflict here, edit the file then # ... darcs amend-record darcs push # push changes <upstream>