When "git diff --no-index" is given an absolute pathname, it
would generate a diff header with the absolute path
prepended by the prefix, like:

  diff --git a/dev/null b/foo

Not only is this nonsensical, and not only does it violate
the description of diffs given in git-diff(1), but it would
produce broken binary diffs. Unlike text diffs, the binary
diffs don't contain the filenames anywhere else, and so "git
apply" relies on this header to figure out the filename.

This patch just refuses to use an invalid name for anything
visible in the diff.

Now, this fixes the "git diff --no-index --binary a
/dev/null" kind of case (and we'll end up using "a" as the
basename), but some other insane cases are impossible to
handle. If you do

	git diff --no-index --binary a /bin/echo

you'll still get a patch like

	diff --git a/a b/bin/echo
	old mode 100644
	new mode 100755
	index ...

and "git apply" will refuse to apply it for a couple of
reasons, and the diff is simply bogus.

And that, btw, is no longer a bug, I think. It's impossible
to know whethe the user meant for the patch to be a rename
or not. And as such, refusing to apply it because you don't
know what name you should use is probably _exactly_ the
right thing to do!

Original problem reported by Imre Deak. Test script and problem
description by Jeff King.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: NShawn O. Pearce <spearce@spearce.org>