The --no-ff option instructs git-rebase to always recreate commits as
they are being replayed, even if fast-forwards are possible.

However, if git-rebase is asked to recreate merge commits (via the -p
option), it suddenly ignores the --no-ff option and fast-forwards
both normal and merge commits whenever possible.

git-rebase--interactive, which is responsible for recreating merge
commits during a rebase, maintains a variable fast_forward to decide
whether the current replay should be tried as a fast-forward.
Previously, fast_forward was on by default and would get toggled only
if a parent was rewritten or a squash was in effect. Also turn
fast_forward off if --no-ff is in use, which is signalled by
git-rebase through the variable force_rebase.

If --no-ff is not in use, try to fast-forward HEAD using git-reset as
before. In contrast, if --no-ff is in use, replay normal commits
using git-cherry-pick and merge commits using git-merge. Note that
git-rebase--interactive already provides this machinery for enabling
and disabling fast-forwards, controlled by fast_forward being
assigned either t (for boolean true) or f (for boolean false).

As mentioned above, git-rebase--interactive needs to detect when a
squash is in effect. If several commits are squashed into one, each
of them is picked using the git-cherry-pick option -n and they get
all rewritten to the same commit, the squash commit. Previously,
fast_forward was assigned f if and only if -n was specified. This no
longer holds for fast_forward might be turned off due to a use of
--no-ff. To correctly notice squashes, explicitly check for -n.

Add test.

Signed-off-by: Fabian Ruch <>

The code checking force_rebase is copied from pick_one, although
using a ternary operator to initialise fast_forward might be more
readable. Moreover, the code snippet used to detect squash mode is
copied from the f arm of the fast_forward case switch, although the
code base prefers to spell out test(1).

The test recreates a topic branch that merged a second topic branch.
Therefore, the test case tests the recreation of both normal and
merge commits.

Commit b499549 first introduced the --no-ff option to git-rebase and
since then force_rebase seems to respected only by pick_one but not
by its sibling pick_one_preserving_merges. I couldn't find a reason
why. Was pick_one_preserving_merges merely overlooked?

Is it a usability issue that conflicting merges will have to be
resolved again when being replayed now? The same applies to -p and
the replay of merges with rewritten parents. Should the possibly
required resolution be mentioned alongside git-rerere in the
git-rebase manual page?

   Fabian        |  3 ++-
 t/ | 12 ++++++++++++
 2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/ b/
index f267d8b..264a768 100644
--- a/
+++ b/
@@ -266,10 +266,11 @@ pick_one_preserving_merges () {
        sha1=$(git rev-parse $sha1)
+       case "$force_rebase" in '') ;; ?*) fast_forward=f ;; esac
        if test -f "$state_dir"/current-commit
-               if test "$fast_forward" = t
+               if [ "$1" != "-n" ]
                        while read current_commit
diff --git a/t/ 
index 8c251c5..838937b 100755
--- a/t/
+++ b/t/
@@ -81,6 +81,18 @@ test_expect_success 'setup for merge-preserving rebase' \
        git commit -a -m "Modify B2"
+test_expect_success '--no-ff records new commits' '
+       (
+       cd clone3 &&
+       test_when_finished 'cd clone3 && git checkout topic' &&
+       git checkout -b recreated-topic &&
+       # recreate topic with merged topic2 (branching-off point A1)
+       git rebase -p --no-ff HEAD~2 &&
+       test $(git rev-parse new-topic^) != $(git rev-parse topic^) &&
+       test $(git rev-parse new-topic) != $(git rev-parse topic)
+       )
 test_expect_success '--continue works after a conflict' '
        cd clone2 &&

To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to
More majordomo info at

Reply via email to