Possible bug with argument parsing in git blame

debayan@deep-blur:~/testrepo$ git blame 22621a32..8486802f -- dir/subdir/newsubdir/b

81531975 b (debayan 2010-05-13 15:32:17 +0530 1) four 81531975 b (debayan 2010-05-13 15:32:17 +0530 2) five d700ac9c b (debayan 2010-05-13 15:33:27 +0530 3) blah 2656ab56 b (debayan 2010-05-13 15:33:34 +0530 4) bleh d0c6e851 b (debayan 2010-05-13 15:42:21 +0530 5) foo 7bf5510e b (debayan 2010-05-13 15:42:53 +0530 6) yo c4515289 b (debayan 2010-05-13 15:43:10 +0530 7) lala a257c7e4 b (debayan 2010-05-13 16:03:46 +0530 8) yes ed8e4601 b (debayan 2010-05-13 16:03:55 +0530 9) np e96d14bf dir/subdir/newsubdir/b (debayan 2010-05-13 16:05:05 +0530 10) yo dace00d1 dir/subdir/newsubdir/b (debayan 2010-05-13 16:05:11 +0530 11) boye 8486802f dir/subdir/newsubdir/b (debayan 2010-05-13 16:07:51 +0530 12) what?

debayan@deep-blur:~/testrepo$ git blame 22621a32..8486802f dir/subdir/newsubdir/b fatal: cannot stat path 22621a32..8486802f: No such file or directory

The only difference between the first and the second command is the "--" separator. Is this normal behaviour?

On Mon, May 17, 2010 at 03:51:18PM +0530, Debayan Banerjee wrote:

I can't reproduce the problem here. For example, in git.git, both of the following produce the same output:

  git blame HEAD~5..HEAD~3 alloc.c
  git blame HEAD~5..HEAD~3 -- alloc.c

Can you provide the exact set of commands to replicate your repository state and show the failure?

Also, I tested with the current 'master'. Are you using an older version of git?

-Peff

On Sun, May 23, 2010 at 07:22:41PM +0530, Debayan Banerjee wrote:

Sorry, I still can't reproduce with:

  mkdir repo && cd repo && git init
  mkdir subdir

  commit() {
    echo content >>subdir/file
    git add subdir/file
    git commit -m foo
  }

  commit; commit; commit; commit; commit

  git blame HEAD~3..HEAD~1 subdir/file >no-dash &&
  git blame HEAD~3..HEAD~1 -- subdir/file >dash &&
  diff no-dash dash &&
  echo ok

So there must be some difference between your setup and my test case. Can you give more details?

-Peff

PS I tested on a variety of git versions back to v1.5.5, and I couldn't trigger the problem, so I don't think it is a version issue.

On 24 May 2010 01:38, Jeff King <peff@peff.net> wrote:

I used master too. The test cases you provided work fine for me too. It also seems that '/' is not a problem. I tested myself with some combinations and i could not reproduced the bug. But the strange behaviour i reported with my particular case still exists. My repo structure can be cloned from http://github.com/debayan/git-blame-test .

On Mon, May 24, 2010 at 11:47:02AM +0530, Debayan Banerjee wrote:

OK, I see what is happening. The issue is that you are blaming a file from some revision, but that file doesn't exist in the working tree. You can recreate it with:

  git init
  echo contents >>foo && git add foo && git commit -m one
  echo contents >>foo && git add foo && git commit -m two
  echo contents >>foo && git add foo && git commit -m three
  git mv foo bar && git commit -m moved

Now "git blame HEAD~2..HEAD~1 -- foo" will work, but "git blame HEAD~2..HEAD~1 foo" will not work.

This is as expected. When parsing revision/filename combinations, if there is no explicit "--" separator, then we only allow filenames if the file exists in the working tree. The idea is to prevent typos. For example, in the same repo:

  git bar ;# works, we have bar in the working tree
  git log foo ;# complains, foo is not a revision and not a file, so
               # maybe a typo
  git log -- foo ;# works, filters on foo. The "--" makes it explicit

Things are a little bit funny with blame because:

  1. It should take exactly one filename, so it doesn't have this
     ambiguity.

  2. The error message isn't very helpful. I think this is because blame
     actually accepts several different combinations (like "<path>
     <rev>", and it ends up confused.

I would be tempted to improve this, but I think (1) may no longer be the case soon. As I understand it, some of Bo Yang's work in this area may mean that blame will soon accept multiple files, and the argument parsing may change.

-Peff