On Tue, Mar 21, 2023 at 11:31 AM Jeff King [off-list ref] wrote:

On Mon, Mar 20, 2023 at 06:57:21PM -0700, Elijah Newren wrote:

quoted

Honestly, looking back at those two patches of mine, I think both were
rather suboptimal.  A better solution that would handle both F->D and
D->F would be having fast-export sort the diff_filepairs such that it
processes the deletes before the modifies.  Another improved solution
would be having fast-import sort the files given to it and handling
deletes first.  Either should fix this.

Might be a good task for a new contributor.  Any takers?  (Tagging as
#leftoverbits.)

There was a patch a while ago, but it didn't get applied:

  https://lore.kernel.org/git/1493079137-1838-1-git-send-email-miguel.torroja@gmail.com/ (local)

It got hung up on the fact that fast-export can also generate renames,
and ordering there is tricky. I stand by the sentiment from back then
that it is still worth it to order things to make the no-rename case
work, even if there are still corner cases with renames (since you can
have a cycle of renames, you should not use them if you want to be
robust against this kind of ordering dependency).

Ah, thanks for the heads up.  Stinkin' renames.  ;-)  I tend to forget
those in the fast-export/fast-import context since fast-export doesn't
detect renames by default (and e.g. filter-repo would turn them off if
they were on by default).

So, I think we probably need to have the fix in fast-import.
Currently, parse_new_commit() calls the various file_change_*()
functions immediately, which in turn immediately call the various
tree_content_*() functions (e.g. file_change_cr() for a rename will
call both tree_content_remove() and tree_content_set()).  Instead of
immediately handling all these, I think that queueing all the
file_change_*() calls up, splitting the renames into delete + modify,
and sorting all the deletes first should fix things for both F->D and
D->F, even when cycles of renames are present as per your example in
that thread.