On Wed, 2004-08-11 at 14:07, Patrick Smears wrote:
> I _have_ seen instances in the past where RCS merge genuinely performed
> 3-way merges incorrectly, in a way that could have led to lost changes, if
> it had not been spotted. (A colleague produced some small recipes that
> demonstrated the problem, and was going to submit them to the RCS
> developers; I don't know what became of that. The problem was intermittent,
> in that merging identical files more than once could give different results
> - suggesting some sort of uninitialised memory problem...). But if
> Subversion's implementation is completely new, that's probably not
> relevant...
Yes, it's a new implementation. Sander wrote libsvn_diff, which
performs diff2 and diff3 algorithms on 'tokens'. We were forced to take
that route rather than re-use the GNU diff/diff3 code (as CVS does),
because the GNU code is GPL'd.
An interesting experiment would be for Kevin to repeat the update using
'-diff3-cmd' argument: see if GNU diff3 gives significantly different
results.
---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscribe@subversion.tigris.org
For additional commands, e-mail: users-help@subversion.tigris.org
Received on Wed Aug 11 21:16:13 2004