From: "Carsten Fuchs" <>
Sent: Monday, January 07, 2013 9:21 AM
Hi Philip,

Am 2013-01-06 18:21, schrieb Philip Oakley:
Your issue [my mistake] is that the (gits's) merge process is a three way merge, so you have the two commits F and N to merge, but git will also locate the merge-base at M (which has the old directory structure), and compares the diffs between them [M-F] and [M-N] (AFAIKI), so you will get these hundreds (thousands) of renames on that basis, and
a great difficulty in (git) trying to decide what to do.

Thank you very much for that explanation, it helps me a lot with understanding this!

So I'm thinking that it would be useful to have a merge commit, if possible, immediately before the two flag day change commits, and then adjust the level of rename detection (--rename-threshold) on the subsequent merge. (can't remember the default threshold)

I had this (a helper merge commit), indeed not strictly immediately before the flag day change commit, but close enough so that I should have recognized if the affected files from the few intermediate commits (between the last merge commit and D) were involved in or responsible for the conflicts.

However, it rather looked as if a main source of trouble were a large number of index.html "sentinel" files: As they all have the exact same contents, it seemed that the rename detection started to associate files at completely different, unrelated paths with each other.

They do sound like they would give some hassle to rename detection!

Also you could simply try an Ours/Theirs strategy (as appropriate) which would stop git trying to do more than it needs to, given that you will already have carefully made the two tree 'compatible' ;-) which will establish a new merge base for future merges.

Ah!!  :-)

I really should have thought of trying this myself. Using
    git merge -s ours master
worked quickly and without any problems, and created the new merge commit G just as expected.

However, I'm unsure if this is the proper solution:
Of course, logically I expected that commits F and G have the same tree (as G's only purpose is to serve as a new merge basis), even if G was created with the default merge strategy. The "ours" strategy does exactly that (refer to same tree in G) quasi on the direct route, per definition.

But I wonder if this argument is enough?

There are two separate issues, one is to create a merge-base (any base) with the new layout (i.e. structure effects), and the second is to ensure you have the right (good-enough) basis (i.e. content) in your merge.

Between the two you will then have easy rename detection and easy content merging.

Normally structure variations are small, so the normal rename detection heuristic is OK, but in your case that promise wasn't kept.

That is, do I understand correctly that if I had used the default merge strategy, and somehow solved all the conflicts (so that none of the files had been changed from F), the result would have technically been exactly the same?

Obviously/hopefully your solution to any conflicts would have ended up being an "ours" choice anyway... Given that you already had a recent merge before the restructuring I would expect that it would be exactl;y the same!

Best regards,


Reply via email to