Re: diff.renames not working?

On Fri, Sep 13, 2019 at 03:24:06PM -0500, Robert Dailey wrote:

This is due to the way that rename detection works. When looking for renames, the tree diff gives us a series of candidate "sources" that were deleted and candidates "dests" that were added. And then we try to match them up.

Copy detection differs in that it uses any file touched in the diff as a source, not just deletions.

And then "--find-copies-harder" uses even unmodified files as sources (but see below).

So here's a simplified setup:

  git init repo
  cd repo

  seq 100 >a
  git add a
  git commit -m a

  cp a b
  seq 99 >a
  git add a b
  git commit -m b

and then we can try these commands:

  # this won't find a rename, because "a" was not deleted
  git diff-tree --name-status -M HEAD^ HEAD

  # this will find the copy, because now we consider "a" a source
  git diff-tree --name-status -C HEAD^ HEAD

  # this _won't_ find the copy, because we limited our tree diff to
  # just look at "b"; hence we don't even consider "a" a source
  git diff-tree --name-status -C HEAD^ HEAD -- b

And that last one is the one that confused you. Naively it seems like doing this would work (two "-C" are the same as "--find-copies-harder"):

  git diff-tree --name-status -C -C HEAD^ HEAD -- b

but it doesn't. That's because we're still using the tree diff to find sources, and just adding unmodified entries to the source list. But our pathspec prevents the diff from even considering "a". While this might seem useless at first, you can imagine something like:

  git diff-tree -C -C HEAD^ HEAD -- subdir/

which would consider all files in subdir as sources, but not those outside (which may be especially important for performance).

But there's no (clean) way to expand the set of paths that we consider as sources without also showing them in the output. There are two useful variants I could imagine, though:

  - a way to consider _all_ paths in the repository, not just those in
    the pathspec, as sources, but show only the entries from the
    pathspec. This could probably be a "harder" version of
    "--find-copies-harder", something like "-C -C -C <revs> -- b".
    Naturally this would be even more expensive in a big repo.

  - a way to independently specify the source pathspec and the
    output-limiting pathspec. This is a cheaper version of the one
    above, where you could look at a subset of the tree a sources, but
    limit the set of shown paths even further. It's not conceptually
    that difficult, but syntactically it gets weird since you have two
    lists of pathspecs on the command-line.

I think the first one wouldn't be _too_ hard if somebody is interested in getting their feet wet with the diffcore-rename.c code. The second is probably not worth the effort.

So yes, that does work, and is why I added the "(clean)" qualifier above. It behaves like the "-C -C -C" I proposed. But the fact that it does so is entirely accidental. What happens is this:

  - we're a little sloppy about what constitutes a traversal option and
    what is a diff option. Many diff commands rely on setup_revisions(),
    which parses both. So "diff --follow" probably _should_ be flagged
    as an error, but isn't.

  - the implementation of "--follow" works by doing a separate,
    from-scratch tree-level diff on each commit (it _has_ to ignore your
    pathspec, since by definition it allows only a single file to be in
    the pathspec). And then rather than throwing away that result, it
    feeds it to the rest of the diff pipeline, which then shows the
    output you expected.

So it does do what you want, but only for the single-file case. And certainly it was never intended to, and that might change in the future.

Accident. :) See above.

It doesn't work because this is a copy, not a rename.

It does, but the pathspec prevents it from finding a source candidate.

-Peff