Web lists-archives.com

rebase preserve-merges: incorrect merge commits




Hello,

I think that rebase preserve-merges algorithm needs further
improvements. Probably, you already know it.

I was playing with rebasing linux kernel tree and found that rebase
preserve-merges works in counterintuitive way while handling
merge-commits. I don't want to discuss arising merge conflicts which
can be illuminated with help of rerere trained at original branch. An
major issue here with silently auto-merging, some merge commits were
silently rebased with different content and it breaks the build :)

I crafted toy example to demonstrate an issue:
https://github.com/matwey/example-git-rebase-pm
Here I have commit range v0.1..v0.2, then I return to v0.1 and
introduce new commit abc-0.1 with new file on v0.1. Then I want to
rebase v0.1..v0.2 onto new abc-0.1. It is obvious that as soon as I
introduced single new file the difference between old v0.2 and new
v0.2 should have only this new file. However it actually contains
other changes.

Reproduce as the following:

git checkout v0.2
git branch abc-0.2
git checkout v0.1
git branch abc-0.1
git checkout abc-0.1
vim LICENSE
git add LICENSE
git commit -a
git rebase --preserve-merges --onto abc-0.1 v0.1 abc-0.2

Actual result:

diff --git a/LICENSE b/LICENSE
deleted file mode 100644
index 297edb3..0000000
--- a/LICENSE
+++ /dev/null
@@ -1,2 +0,0 @@
-Hello, world!
-
diff --git a/mod.c b/mod.c
index e6fa107..e339597 100644
--- a/mod.c
+++ b/mod.c
@@ -2,5 +2,5 @@
 #include "mod.h"

 int mod_fun(char* buf, int len) {
-       return -calc_fun(buf, len);
+       return -calc_fun(buf, len, 0);
 }

Expected result:

diff --git a/LICENSE b/LICENSE
deleted file mode 100644
index 297edb3..0000000
--- a/LICENSE
+++ /dev/null
@@ -1,2 +0,0 @@
-Hello, world!
-


As far as I understand the root cause of this that when new merge
commit is created by rebase it is done simply by git merge
$new_parents without taking into account any actual state of the
initial merge commit. This is why all manually introduced changes in
merge-commits are lost at rebase stage. I think that the following
improvement of merge-commit recreating stage would be possible:

1) Let we have merge commit in the source tree

-- A --\
        M--
-- B --/

And we want to rebase M to the new tree where new parents A' and B'
already existing.

-- A'

-- B'

2) Then I suggest to perform it in three steps. First, lets take all
possible differences between A and M, B and M and so on... and then
try to cherry-pick the diffs on A' and B' respectively. Here AM' and
BM' are temporary commits which is very close in their states (in
ideal case, equal).

-- A' -- (AM)'

-- B' -- (BM)'

3) Then we do git merge (AM)' (BM)' and ask user to resolve the
conflicts if any. Here we obtain the following tree. Where M" is
produced by git merge.

-- A' -- (AM)' --\
                  M"
-- B' -- (BM)' --/

4) Then we do merge between A' and B' but reset the state to be the
same as in M" before new merge commit is created:

-- A' -- (AM)' --\
     \            \
       M' (==M")    M"
     /            /
-- B' -- (BM)' --/

5) Then (AM)' (BM)' and M" are discarded from the tree.


Probably, I missed something important or incorrectly read
git-rebase--interactive.sh

-- 
With best regards,
Matwey V. Kornilov