Once the command git diff *.cpp just stopped printing anything in one repository, despite git diff *.h works fine. git diff works fine for *.cpp files.
What could it be?
The fix is to be careful to run, e.g.:
git diff '*.cpp'
or:
git diff \*.cpp
or more formally:
git diff -- '*.cpp'
or similar.
What's going on here?
Missing from the question: the desired output was a diff for subdir/subfile.cpp, but the top level of the work-tree contained one file named file.cpp or similar.
The problem stems from the fact that you're using a Unix/Linux-style shell, which expands * and other wildcard or glob characters before running the command you enter at the command line. But there is a bit of subtlety here as well.
Because there exists a file named file.cpp in the current directory, when you run:
git diff *.cpp
the shell replaces *.cpp with the names of all the files whose name ends with .cpp, and therefore runs:
git diff file.cpp
Git then dutifully produces the diff—or in this case, no diff since there is no difference—for the one file named file.cpp.
When there are no files named file.cpp or zorg.cpp or similar in the top level directory, however, this shell simply invokes git with arguments diff and *.cpp, as if you had quoted the asterisk. This gives Git the chance to expand the *.cpp argument, and when Git expands it, it does so in a different way than the shell.
Why use --?
The git diff command takes a number of options, such as -s, -p, -w, --name-status, --name-only, and so on.
Depending on what files you have, suppose you want a diff listing for the file named -z in the current directory. If you then run:
git diff -z
Git thinks you mean to supply the -z option, rather than to get a diff listing for the file named -z. A similar problem applies if you want diffs for -z* and the like.
In general, you can work around this problem by using the file name ./-z instead of just -z. Since ./-z does not start with -, Git is not fooled into thinking it's an option. But this problem is more general, and strikes in other cases (other commands) as well. For instance, suppose you have a file named develop and you run:
git checkout develop
Git will think you mean to check out the branch named develop.
All Git commands accept -- as a separator, generally meaning there no more options: anything after this point is an argument instead. For git diff, anything after -- is treated as a pathspec, which includes doing glob expansion, provided the glob characters made it past the shell.
This is what the syntax description in the SYNOPSIS section of the git diff documentation means:
git diff [<options>] [<commit>] [--] [<path>...]
git diff [<options>] --cached [<commit>] [--] [<path>...]
git diff [<options>] <commit> <commit> [--] [<path>...]
git diff [<options>] <blob> <blob>
git diff [<options>] --no-index [--] <path> <path>
The square brackets indicate that something is optional, so all the options are optional. The angle brackets indicate that some argument should be replaced with a string that meets the requirements of the type inside the brackets. The literal -- or other literal options imply that you should type those characters literally—so git diff --cached requires the literal string --cached, for instance. Last, the ... means "repeat the previous as often as you like".
Since the literal string -- is optional, you don't have to enter it—but if you do, everything after it must have the form of a <path>. That form is quite general: almost any character is valid. The documentation is missing a cross-reference to the definition of a pathspec, though (and probably should use <pathspec>, not just <path>, here). The full description of pathspecs is in the gitglossary.
So I have one .cpp file in the repository root directory, and so git diff *.cpp now prints changes only for this file, while git diff '*.cpp' works fine for all files in subdirectories.
I had an awful list of old stashes
I have first removed the very old ones
git reflog expire --expire-unreachable=7.days refs/stash
I have one huge stash left, which contains many stashed changes. Some are to keep some other would damage my production system. I went through the diff
git diff stash#{0}^1 stash#{0}
and I know which ones are to keep
I could do
git checkout --patch stash#{0} -- myfilename
to unstash changes on myfilename and is working fine.
However, I have a large folder with many files with stashed changes inside. I would like to apply all of them but only within that subfolder.
I have tried to approach it using wildcards in ksh but I does not work
git checkout --patch stash#{0} -- myfolder/*
results in
error pathspec [...] did not match any files known to git
The solution does not need to be git based, can be a shell script to wrap git calls
Have you tried :
git checkout --patch stash#{0} -- myfolder
without the ending * ?
Chances are your shell expands myfolder/* before executing the git command, and lists the elements which currently exist on disk, which is probably not what you want.
1.I'm new to git and would like to know what happen if I've a file which was modified and already staged in the past(but now modified again), and I want to commit the file using a command such as :
git commit -m "yada yada" ~/home/Dan/project/file_to_commit.py
Is this equivalent to doing:
a.git add ~/home/dan/project/file_to_commit.py
b.git commit -m "yada yada"
If not please explain.
2.How can I see the remotes branch commits logs?(pushes) without doing git pull?
Thanks
This might be better as two separate questions, and the second question is already answered correctly, but I'll take a stab at answering both. Before I start, though, I want to say that the path ~/home/Dan/project/file_to_commit.py is pretty suspect: it suggests that your git directory is /.git, which is not a good idea. I'm going to assume, below, that the .git directory is much further down and you're just adding project/file or file (and I'll trim the paths in the question).
Note that the TL;DR version of the first answer is that they're almost the same, and you only need to know about the difference in some edge cases. (Hence the existing answer from eleventhend is probably good enough for most purposes.)
Q1: Add and commit vs git commit path
... I've a file which was modified and already staged in the past (but now modified again), and I want to commit the file using a command such as:
git commit -m "yada yada" file_to_commit.py
Is this equivalent to doing:
git add file_to_commit.py
git commit -m "yada yada"
If not please explain.
No, it's not exactly equivalent. This is a little bit tricky and it will help a lot if we define some terms and get a few things pinned down a bit more.
Also, you say "already staged in the past (but now modified again)", which leaves a bit of ambiguity: did you do a git commit in between these operations? I'll address both the "yes" and "no" cases by describing the full, general case.
The index
First, we need to define git's index or staging area (it has even a few more names, e.g., cache as in git diff --cached, but "index" and "staging area" are the two most common terms). Git has a (one, single) standard index—let's call this "the" index, and when we want to refer to another, different index, we'll spell out which other one we mean. When you run most normal git commands, including git add, git updates this index.1
Next, we need to take some notes on what's actually in the index. Aside from some interesting but not relevant here cases like merges, and one thing I'm leaving out on purpose, in essence, the index has one entry per file that will be in the next commit.2 That is, suppose you've made a commit, or checked out some branch, so that your current commit, which is now in your work tree, has 100 files in it. (Your work tree may have additional untracked and/or ignored files, as long as it also has those 100 files.)
Using git add
When you run git add, git stores a new copy of each of the files being added into the repository, which it calls blob objects. It calculates a hash value for each blob as it adds it to the repository, then puts the new hash into the index.
When you run git commit—at least, without either paths or -a—git reads the index and uses that to form the new commit. If the new commit would have the same tree as the previous commit,3 git requires that you add the --allow-empty flag. (This doesn't mean that the index is empty, but rather that the index matches the index you'd get by re-checking-out the current commit. So --allow-empty might be the wrong name for this flag; it maybe should have been called --allow-same or allow-unchanged or some such.)
Hence, if you do git add path and then git commit -m message, you'll get a commit that uses the index as updated by the git add, which may have additional updates from before that git add. Since there's just the one entry per path, though, if you do:
... hack on foo.py ...
$ git add foo.py
$ echo '# add a final comment' >> foo.py
$ git add foo.py
$ git commit -m 'update foo'
there will only be one update to foo.py in the new commit.
So what's the difference?
I claimed above that git commit path (and git commit -a) is not exactly the same as doing the git add and then git commit. Let's look at the precise difference.
When you give path names (or -a) to git commit, git uses a new, different, temporary index. It starts by copying something—exactly what gets a bit complicated—to the temporary index, then it adds each file that is to be committed, then it makes a commit from the temporary index. Once the commit is done, git goes back to using the normal, ordinary index, and it updates the index. That is, after adding stuff to the temporary index and committing, it also adds to the regular index.
To see how this really works we need some examples. Suppose we have two files and we git add a change to one of them:
# assume file1 and file2 are in the HEAD commit
echo add stuff to file1 >> file1
git add file1
echo add stuff to file2 too >> file2
At this point, git status will tell us that we have changes to file1 that are staged to be committed, and changes to file2 that are not staged to be committed.
If we run git add file2; git commit, we'll get both updates in the new commit. Once the commit is done, git status will show there is nothing to commit. But if, instead, we do:
git commit -m message file2
and then run git status, we'll see that file1 is still staged and ready to commit. The commit we just made has only the change to file2.
This is because the git commit file2 command started by making a temporary index using the HEAD commit, adding file2, making the commit, and then updating the normal index with the new file2. This last bit is important: if git didn't copy the change back into the (regular) index, the index would still have the old version of file2, and the next commit would undo the change we just committed.
This copy-back step has an important side effect. Suppose that we have a complicated change to foo.py. For instance, suppose we went to fix a bug, and along the way we refactored a few functions. We've tested the fix and it works, so we did git add foo.py and were about to commit it:
... hack on foo.py ...
... test, hack more, test, until fixed ...
git add foo.py
git commit -m 'refactor foo.py and then fix a bug'^C
At this point we realized that we shouldn't commit both changes together: we should commit the refactored code first, and then commit the bug fix.4
Well, we've already git add-ed the refactored and fixed code, so it's safely stashed away in the index, right? (No, WRONG! Danger Will Robinson! But let's proceed, since this is an example.) So we can just undo the fix part, leaving only the refactoring, and commit that first:
... edit foo.py to remove just the fix ...
git commit -m 'refactor code to prep for bug fix' foo.py
Once that commit is done, we can commit the staged version:
git commit -m 'fix bug #12345' foo.py
Alas, git now tells us that there's nothing to commit. What happened to the carefully staged full-fix version of foo.py?
The answer is, git commit foo.py wiped it out. Git first added the refactor-only foo.py to a temporary index and committed that; but then it copied the refactor-only foo.py back to the regular index, losing our carefully staged full-fix version. We can either re-create it, or fish around in the repository for the "dangling blob" that is left behind because of this.
(This should probably be considered a bug in git, although it's not clear what to do with the staged but uncommitted version.)
git commit -a and/or paths: --only vs --include
Here I need to quote myself from just a moment ago. When using -a or paths, git commit starts by copying something—exactly what gets a bit complicated. If you look closely at the git commit documentation, you will find the -i or --include option (and a corresponding, but default, -o / --only option). These control what goes into the temporary index.
When using --include, git commit creates its temporary index from the current index. When using the default --only mode, git commit creates its temporary index from the HEAD commit.
(Because of the copy-back step at the end, the only way to see that both commands are in fact using a temporary index is to view the index in the middle of the commit operation. If we use --include and check after the commit is done, the regular index will match the new HEAD commit, as if git commit were adding to the regular index rather than to the temporary index. Fortunately it's very easy to view the regular index "in the middle", by not supplying the -m flag, so that git commit fires up the editor. While that's going on, run git status in another window, or using job control. Here's an example:
# At this point I've modified both a.py and mxgroup.py
# but have not `git add`ed either one.
$ git add a.py
$ git status --short
M a.py
M mxgroup.py
# We see that a plain "git commit" would commit a.py,
# but not mxgroup.py.
$ git commit -i mxgroup.py
# now waiting in the editor
# Now, in another window:
$ git status -s
M a.py
M mxgroup.py
This shows that the (regular) index is still set up the way we had it. Once we write the message out and exit the editor, the commit process will update the regular index for the new mxgroup.py entry, and the now-committed a.py change is also in the new HEAD commit, so git status will show neither file.)
Q2: Viewing logs
How can I see the remotes branch commits logs?(pushes) without doing git pull?
Git itself operates almost entirely locally. You may be able to do this directly with web viewers, but it's pretty convenient to just do locally, by first running git fetch.
The git pull command is actually meant as a convenience. There are two steps needed to incorporate other people's commits:
obtain those commits so that you have them locally; and
merge or rebase using those commits.
These two steps are handled by different git commands: git fetch does step 1, and git merge and git rebase—you must pick one of the two—does whichever version of step 2 you want.
The git pull command simply does step 1, then does step 2. It chooses merge unless you instruct it otherwise. For historical reasons, it has multiple ways of choosing which operation to run in step 2.
My recommendation is that as a newbie to git, you avoid git pull entirely. There are two reasons for this, only one of which is valid today (unless you're stuck with very old versions of git):
Traditionally, git pull has had various bugs, some of which can even lose work. (As far as I know these are all fixed since git 2.0.)
While it is convenient, git pull obscures what's happening and makes you choose merge-vs-rebase too early. It is true that rebase is almost always the right answer, but git pull defaults to doing merge, which means that its default is wrong for new users. Plus, of course, there's that "obscures what's happening" issue. If you knew about fetch and then rebase as separate steps, this question probably would not even have come up.
1The index is just a file, and you can find it in .git/index. You can make git use a different index by setting GIT_INDEX_FILE in the environment, but this is mainly meant for use by scripts like git stash.
2The cases I'm leaving out are:
Conflicted merges, which record up to three entries per path, using non-zero stage numbers. Once your resolve the conflict and git add the result, the nonzero stages are cleaned out and the normal stage-0 result is stored instead, and we're back to the normal, ready-to-commit case for that index intry.
Removing a file that is in the current commit (git rm, with or without --cached). This writes a special stage-0 entry marking the file as to-be-omitted, rather than simply removing the entry.
3If you're committing a merge, git allows the tree to match those of any or all parents, since the merge commit needs to record the multiple parents. The "empty" test is thus applied only to non-merge, single-parent commits. This is documented much better in modern git than it was in old versions of git, but it still has the wrong name.
4This has nothing to do with git itself. The idea here is to commit small, readable, understandable, and most importantly testable changes. Any time you find yourself writing up a commit as "do A and B, and fix C, and add D and E" it's an indication that you should probably split this into one commit per thing—in this case, about 5 separate commits.
[updated]
It is actually equivalent. When you commit a file directly, using git commit <filepath>, it stages the current file before committing. You do have to stage the file the first time the file is added before committing it (tell the repository to start tracking the file) using git add <file>.
Sample workflow:
git add <file>
Make some changes, yada yada
git commit -m "yada yada" <file>
Make some more changes, blah blah
git commit -m "blah blah" <file>
2.
To see the commit logs of a remote git repository, first use git fetch on the repository, then run git log <path/branch> to view the log.
See here: https://github.com/abhikp/git-test/wiki/View-the-commit-log-of-a-remote-branch
I far too frequently use the mouse to do things like this:
/home/me-$ git log --name-status -1
commit a10e63af1f4b1b2c28055fed55d4f2bb3225a541
Author: Me <me#me.com>
Date: Tue Aug 18 13:04:04 2015 -0400
XYZ-376 make ctors public
M x/y/z/Class1.java
M x/y/z/Class2.java
/home/me-$ git checkout -- x/y/z/Class2.java # <-- copy/paste with the mouse
I know that some git commands accept wildcards, and this mitigates this problem somewhat, but I'm wondering if there is a way do specifically reference pathspecs, etc. from previous commands.
How can I run commands like this without using the mouse, and without retyping long paths by hand?
I typically use a subshell ($(<command in subshell here...>)) for this.
For example, sometimes I had many files deleted and I had to git rm every one of them.
There's the command git ls-files --deleted that returns the names of all the missing files. I can combine it with git rm like this:
git rm $(git ls-files --deleted)
This is somewhat a bad example, because (as I discovered later), this operation can be achieved much easier with git add --all. But I think it illustrates the point.
In your case, if you wanted to checkout all files that have been changed in the previous commit, it would be hard to parse the output of git log --name-status, because it contains additional information, but you could use something like git diff HEAD^ --name-only instead.
So:
git checkout $(git diff HEAD^ --name-only)
will do it in your example.
One nice thing that I noticed using the $(...) syntax is that it works both in Bash and in PowerShell.
This'd be the kind of thing you run a shell under emacs for, run all your shells in it and have a command to walk back through the buffer looking for patterns in the output.
For retrieving output from a previous command that you didn't capture inside the shell session, you're going to have to get it from your terminal emulator's buffers somehow. The xterm family has a configurable "copy the whole scrollback buffer" thingy, then xclip -o will print the selection and you can pipe it through an extraction filter.
But it's either capture the output within the session or scrape it from the output buffers afterwards, that's everywhere the data's ever been.
On OS X, "Mouseless Copy" supported by iTerm2 (and probably some other terminal emulators) is a workable solution: https://www.iterm2.com/features.html
search for some portion of the string (⌘F)
expand selection right (tab) or left (shift-tab)
paste selection with (option-enter) or copy/paste in the usual way
I'm trying to run this command:
git filter-branch --force --index-filter 'git rm --cached --ignore-unmatch filename.js' --prune-empty --tag-name-filter cat -- --all
but I keep getting this error:
fatal: ambiguous argument 'rm': unknown revision or path not in the working tree
.
Use '--' to separate paths from revisions, like this:
'git <command> [<revision>...] -- [<file>...]'
It depends on the shell you are using.
On Windows, with msysgit for instance, see issue 477:
Single quotes do not have a special meaning with CMD. Do not expect that they work
the same as with a POSIX shell. Call filter-branch like this:
git filter-branch --commit-filter "GIT_COMMITTER_NAME=void GIT_AUTHOR_NAME=void GIT_COMMITTER_EMAIL=just.a.test#kernel.org GIT_AUTHOR_EMAIL=just.a.test#kernel.org; git commit-tree \"$#\"" HEAD
Multiple lines:
git filter-branch --commit-filter "GIT_COMMITTER_NAME=void \
GIT_AUTHOR_NAME=void \
GIT_COMMITTER_EMAIL=just.a.test#kernel.org \
GIT_AUTHOR_EMAIL=just.a.test#kernel.org; \
git commit-tree \"$#\"" HEAD
As mentioned in "How to pass a programmatically generated list of files to git filter-branch?"
Each argument to the various ...-filters needs to be a single string. That string is saved as a shell variable.
So make sure 'git rm --cached --ignore-unmatch filename.js' is considered a string in the shell you are in.
As Constantine Ketskalo points out in the comments:
Windows 10, PyCharm, GitPython, same command as in question.
Simply changed ' to " inside the string and it worked!
On windows, you have to use double quote " instead of single '