Re: [Gnu-arch-users] [arch advocacy] Games from Within: litmus test for revision control

[Tom Lord]

   From: Robert Anderson <address@hidden>


   > I have not tried, but I think this would not work out:

   > 1) Have a project with 'foo' directory.
   > 2) Branch it
   > 3) On one branch rename the 'foo' directory to 'bar'
   > 4) On another branch create file 'foo/baz'
   > 5) Merge.

   > Now correct behaviour would be to place the 'baz' file in the renamed
   > directory.

   I think your definition of "correct behavior" is potentially
   debatable, but I agree that this is probably the best definition and I
   did confirm that it does not work like that.

   The changeset records the addition of the file using named path
   elements, not the logical identities of the path elements.

   Tom, can you comment on this?  This seems like a pretty fundamental
   design point.

When a file or directory is added, the changeset indexes are
*supposed* to include paths and ids for the parent dirs.  `dopatch'
is *supposed* to use that information to add the file in the correct
place in the inexact patching case described in the quoted stuff 
above.

Initially, `larch' botched that case.   Then I fixed that, mostly.
I *thought* I had gotten this case right in `tla' to start with but
either I didn't, or it got broke between the early `tla' and the last
couple of releases.

Either way: the changeset is supposed to have enough information for this case;
the design leaves plenty of room for that information;  `dopatch' is 
supposed to figure this out --- one of those `supposed to'-s isn't actually
the case.  That's an inexact patching bug.  It's not a design bug.

There's also a project bug: this bug got lost in the shuffle over the past
several *months*, it seems.

It's not lost now and will be fixed.

-t