Dealing with SVN keyword expansion with git-svn

Solution 1:

What's going on here: Git is optimized to switch between branches as quickly as possible. In particular, git checkout is designed to not touch any files that are identical in both branches.

Unfortunately, RCS keyword substitution breaks this. For example, using $Date$ would require git checkout to touch every file in the tree when switching branches. For a repository the size of the Linux kernel, this would bring everything to a screeching halt.

In general, your best bet is to tag at least one version:

$ git tag v0.5.whatever

...and then call the following command from your Makefile:

$ git describe --tags
v0.5.15.1-6-g61cde1d

Here, git is telling me that I'm working on an anonymous version 6 commits past v0.5.15.1, with an SHA1 hash beginning with g61cde1d. If you stick the output of this command into a *.h file somewhere, you're in business, and will have no problem linking the released software back to the source code. This is the preferred way of doing things.

If you can't possibly avoid using RCS keywords, you may want to start with this explanation by Lars Hjemli. Basically, $Id$ is pretty easy, and you if you're using git archive, you can also use $Format$.

But, if you absolutely cannot avoid RCS keywords, the following should get you started:

git config filter.rcs-keyword.clean 'perl -pe "s/\\\$Date[^\\\$]*\\\$/\\\$Date\\\$/"'
git config filter.rcs-keyword.smudge 'perl -pe "s/\\\$Date[^\\\$]*\\\$/\\\$Date: `date`\\\$/"'

echo '$Date$' > test.html
echo 'test.html filter=rcs-keyword' >> .gitattributes
git add test.html .gitattributes
git commit -m "Experimental RCS keyword support for git"

rm test.html
git checkout test.html
cat test.html

On my system, I get:

$Date: Tue Sep 16 10:15:02 EDT 2008$

If you have trouble getting the shell escapes in the smudge and clean commands to work, just write your own Perl scripts for expanding and removing RCS keywords, respectively, and use those scripts as your filter.

Note that you really don't want to do this for more files than absolutely necessary, or git will lose most of its speed.

Solution 2:

Unfortunately, RCS keyword substitution breaks this. For example, using $Date$ would require git checkout to touch every file in the tree when switching branches.

That is not true. $Date$ etc. expand to the value which holds at checkin time. That is much more useful anyway. So it doesn't change on other revisions or branches, unless the file is actually re-checked-in. From the RCS manual:

   $Date$ The  date  and  time the revision was checked in.  With -zzone a
          numeric time zone offset is appended;  otherwise,  the  date  is
          UTC.

This also means that the suggested answer above, with the rcs-keyword.smudge filter, is incorrect. It inserts the time/date of the checkout, or whatever it is that causes it to run.

Solution 3:

Here is a sample project containing the configuration and filter code needed for adding RCS keyword support to a git project:

https://github.com/turon/git-rcs-keywords

It's not as simple to setup as one would like, but it seems to work. It uses a smudge/clean filter pair written in perl (similar to what emk's answer described), and yes, it will touch all files with the extensions set in .gitattributes, generally slowing things down a bit.