This solves the problem of rebasing local commits against an
upstream that has renamed files.
Signed-off-by: Eric Wong <redacted>
---
Documentation/git-rebase.txt | 20 +++++
git-rebase.sh | 176 ++++++++++++++++++++++++++++++++++++++++--
2 files changed, 186 insertions(+), 10 deletions(-)
diff --git a/Documentation/git-rebase.txt b/Documentation/git-rebase.txt
index 08ee4aa..c339c45 100644
--- a/Documentation/git-rebase.txt
+++ b/Documentation/git-rebase.txt
@@ -7,7 +7,7 @@ git-rebase - Rebase local commits to a n
SYNOPSIS
--------
-'git-rebase' [--onto <newbase>] <upstream> [<branch>]
+'git-rebase' [--merge] [--onto <newbase>] <upstream> [<branch>]
'git-rebase' --continue | --skip | --abort
@@ -106,6 +106,24 @@ OPTIONS
--abort::
Restore the original branch and abort the rebase operation.
+--skip::
+ Restart the rebasing process by skipping the current patch.
+ This does not work with the --merge option.
+
+--merge::
+ Use merging strategies to rebase. When the recursive (default) merge
+ strategy is used, this allows rebase to be aware of renames on the
+ upstream side.
+
+-s <strategy>, \--strategy=<strategy>::
+ Use the given merge strategy; can be supplied more than
+ once to specify them in the order they should be tried.
+ If there is no `-s` option, a built-in list of strategies
+ is used instead (`git-merge-recursive` when merging a single
+ head, `git-merge-octopus` otherwise). This implies --merge.
+
+include::merge-strategies.txt[]
+
NOTES
-----
When you rebase a branch, you are changing its history in a way that
diff --git a/git-rebase.sh b/git-rebase.sh
index e6b57b8..68bddfb 100755
--- a/git-rebase.sh
+++ b/git-rebase.sh
@@ -34,7 +34,81 @@ When you have resolved this problem run
If you would prefer to skip this patch, instead run \"git rebase --skip\".
To restore the original branch and stop rebasing run \"git rebase --abort\".
"
+
+MRESOLVEMSG="
+When you have resolved this problem run \"git rebase --continue\".
+To restore the original branch and stop rebasing run \"git rebase --abort\".
+"
unset newbase
+strategy_args=
+do_merge=
+dotest=$GIT_DIR/.dotest-merge
+prec=4
+
+continue_merge () {
+ test -n "$prev_head" || die "prev_head must be defined"
+ test -d "$dotest" || die "$dotest directory does not exist"
+
+ unmerged=$(git-ls-files -u)
+ if test -n "$unmerged"
+ then
+ echo "You still have unmerged paths in your index"
+ echo "did you forget update-index?"
+ die "$MRESOLVEMSG"
+ fi
+
+ mh="$GIT_DIR/MERGE_HEAD"
+ mm="$GIT_DIR/MERGE_MSG"
+ if test -f "$mh" && test -f "$mm"
+ then
+ git-commit -F "$mm" || die "commit failed: $MRESOLVEMSG"
+ else
+ echo "Previous merge succeeded automatically"
+ fi
+
+ prev_head=`git-rev-parse HEAD^0`
+
+ # save the resulting commit so we can read-tree on it later
+ echo "$prev_head" > "$dotest/`printf %0${prec}d $msgnum`.result"
+ echo "$prev_head" > "$dotest/prev_head"
+
+ # onto the next patch:
+ msgnum=$(($msgnum + 1))
+ printf "%0${prec}d" "$msgnum" > "$dotest/msgnum"
+}
+
+call_merge () {
+ cmt="$(cat $dotest/`printf %0${prec}d $1`)"
+ echo "$cmt" > "$dotest/current"
+ git-merge $strategy_args "rebase-merge: $cmt" HEAD "$cmt" \
+ || die "$MRESOLVEMSG"
+}
+
+finish_rb_merge () {
+ set -e
+
+ msgnum=1
+ echo "Finalizing rebased commits..."
+ git-reset --hard "`cat $dotest/upstream`"
+ end="`cat $dotest/end`"
+ while test "$msgnum" -le "$end"
+ do
+ msgnum=`printf "%0${prec}d" "$msgnum"`
+ printf "%0${prec}d" "$msgnum" > "$dotest/msgnum"
+
+ git-read-tree `cat "$dotest/$msgnum.result"`
+ git-checkout-index -q -f -u -a
+ git-commit -C "`cat $dotest/$msgnum`"
+
+ echo "Committed $msgnum"
+ echo ' '`git-rev-list --pretty=oneline -1 HEAD | \
+ sed 's/^[a-f0-9]\+ //'`
+ msgnum=$(($msgnum + 1))
+ done
+ rm -r "$dotest"
+ echo "All done."
+}
+
while case "$#" in 0) break ;; esac
do
case "$1" in@@ -46,17 +120,42 @@ do
exit 1
;;
esac
+ if test -d "$dotest"
+ then
+ prev_head="`cat $dotest/prev_head`"
+ end="`cat $dotest/end`"
+ msgnum="`cat $dotest/msgnum`"
+ continue_merge
+ while test "$msgnum" -le "$end"
+ do
+ call_merge "$msgnum"
+ continue_merge
+ done
+ finish_rb_merge
+ exit
+ fi
git am --resolved --3way --resolvemsg="$RESOLVEMSG"
exit
;;
--skip)
+ if test -d "$dotest"
+ then
+ die "--skip is not supported when using --merge"
+ fi
git am -3 --skip --resolvemsg="$RESOLVEMSG"
exit
;;
--abort)
- [ -d .dotest ] || die "No rebase in progress?"
+ if test -d "$dotest"
+ then
+ rm -r "$dotest"
+ elif test -d .dotest
+ then
+ rm -r .dotest
+ else
+ die "No rebase in progress?"
+ fi
git reset --hard ORIG_HEAD
- rm -r .dotest
exit
;;
--onto)
@@ -64,6 +163,24 @@ do
newbase="$2"
shift
;;
+ -M|-m|--m|--me|--mer|--merg|--merge)
+ do_merge=t
+ ;;
+ -s=*|--s=*|--st=*|--str=*|--stra=*|--strat=*|--strate=*|\
+ --strateg=*|--strategy=*|\
+ -s|--s|--st|--str|--stra|--strat|--strate|--strateg|--strategy)
+ case "$#,$1" in
+ *,*=*)
+ strategy=`expr "$1" : '-[^=]*=\(.*\)'` ;;
+ 1,*)
+ usage ;;
+ *)
+ strategy="$2"
+ shift ;;
+ esac
+ do_merge=t
+ strategy_args="${strategy_args}-s $strategy "
+ ;;
-*)
usage
;;@@ -75,16 +192,25 @@ do
done
# Make sure we do not have .dotest
-if mkdir .dotest
+if test -z "$do_merge"
then
- rmdir .dotest
-else
- echo >&2 '
+ if mkdir .dotest
+ then
+ rmdir .dotest
+ else
+ echo >&2 '
It seems that I cannot create a .dotest directory, and I wonder if you
are in the middle of patch application or another rebase. If that is not
the case, please rm -fr .dotest and run me again. I am stopping in case
you still have something valuable there.'
- exit 1
+ exit 1
+ fi
+else
+ if test -d "$dotest"
+ then
+ die "previous dotest directory $dotest still exists." \
+ 'try git-rebase < --continue | --abort >'
+ fi
fi
# The tree must be really really clean.
@@ -152,6 +278,38 @@ then
exit 0
fi
-git-format-patch -k --stdout --full-index "$upstream"..ORIG_HEAD |
-git am --binary -3 -k --resolvemsg="$RESOLVEMSG"
+if test -z "$do_merge"
+then
+ git-format-patch -k --stdout --full-index "$upstream"..ORIG_HEAD |
+ git am --binary -3 -k --resolvemsg="$RESOLVEMSG"
+ exit 0
+fi
+
+# start doing a rebase with git-merge
+# this is rename-aware if the recursive (default) strategy is used
+
+mkdir -p "$dotest"
+echo "$upstream" > "$dotest/upstream"
+prev_head=`git-rev-parse HEAD^0`
+echo "$prev_head" > "$dotest/prev_head"
+
+msgnum=0
+for cmt in `git-rev-list "$upstream"..ORIG_HEAD | tac`
+do
+ msgnum=$(($msgnum + 1))
+ echo "$cmt" > "$dotest/`printf "%0${prec}d" $msgnum`"
+done
+
+printf "%0${prec}d" 1 > "$dotest/msgnum"
+printf "%0${prec}d" "$msgnum" > "$dotest/end"
+
+end=$msgnum
+msgnum=1
+
+while test "$msgnum" -le "$end"
+do
+ call_merge "$msgnum"
+ continue_merge
+done
+finish_rb_merge--
1.4.0.ge24c
Eric Wong [off-list ref] writes:
This solves the problem of rebasing local commits against an
upstream that has renamed files.
I think leveraging the merge strategy to perform rebase is
sound, but the selection of merge base for this purpose is quite
different from the regular merge, and I think unfortunately this
patch is probably wrong in letting git-merge choose the merge
base.
But let's mention other things as well.
- You kept the original "format-patch piped to am" workflow
optionally working.
- You check if merge or patch was used for failed rebase and
follow the appropriate codepath while resuming, which is
good.
- The list of commits you generate with tac seem to include
merge commit -- you may want to give --no-merges to
rev-list.
- I do not think we use "tac" elsewhere -- is it portable
enough?
- Exiting with success unconditionally after "git am" feels
wrong. I would do "exit $?" instead of "exit 0" there.
Suppose you have this commit ancestry graph:
----------------------------------------------------------------
Example: git-rebase --onto master A topic
A---B---C topic B'--C' topic
/ --> /
D---E---F---G master D---E---F---G master
----------------------------------------------------------------
This is slightly different from the one at the beginning of the
script. The idea is A turned out to be not so cool, and we
would want to drop it.
+call_merge () {
+ cmt="$(cat $dotest/`printf %0${prec}d $1`)"
+ echo "$cmt" > "$dotest/current"
+ git-merge $strategy_args "rebase-merge: $cmt" HEAD "$cmt" \
+ || die "$MRESOLVEMSG"
+}
call_merge is first called with B in cmt, and HEAD is pointing
at G. But the merge in this function makes a merge between B
and G, taking the effect of E->A.
I think the three-way merge you would want here is not between B
and G using E as the pivot, but between B and G using A as the
pivot. That's how cherry-pick and revert works. I would
leverage the interface that is one level lower for this -- the
strategy modules themselves.
git-merge-$strategy $cmt^ -- HEAD $cmt
The strategy modules take merge base(s), double-dash as the
separator, our head and the other head. They do not make commit
themselves (instead they leave working tree and index in
committable state) and signal the results with their exit
status:
0 -- success
1 -- conflicts
2 -- did not handle the merge at all
This tests Eric's "renaming rebase" patch. It tests only very
basic cases and most of the tests except the last one passes.
Signed-off-by: Junio C Hamano <redacted>
---
t/t3402-rebase-merge.sh | 96 +++++++++++++++++++++++++++++++++++++++++++++++
1 files changed, 96 insertions(+), 0 deletions(-)
diff --git a/t/t3402-rebase-merge.sh b/t/t3402-rebase-merge.sh
new file mode 100755
index 0000000..8c7a519
--- /dev/null
+++ b/t/t3402-rebase-merge.sh
@@ -0,0 +1,96 @@
+#!/bin/sh
+#
+# Copyright (c) 2006 Junio C Hamano
+#
+
+test_description='git rebase --merge test'
+
+. ./test-lib.sh
+
+T="A quick brown fox
+jumps over the lazy dog."
+for i in 1 2 3 4 5 6 7 8 9 10
+do
+ echo "$i $T"
+done >original
+
+test_expect_success setup '
+ git add original &&
+ git commit -m"initial" &&
+ git branch side &&
+ echo "11 $T" >>original &&
+ git commit -a -m"master updates a bit." &&
+
+ echo "12 $T" >>original &&
+ git commit -a -m"master updates a bit more." &&
+
+ git checkout side &&
+ (echo "0 $T" ; cat original) >renamed &&
+ git add renamed &&
+ git update-index --force-remove original &&
+ git commit -a -m"side renames and edits." &&
+
+ tr "[a-z]" "[A-Z]" <original >newfile &&
+ git add newfile &&
+ git commit -a -m"side edits further." &&
+
+ tr "[a-m]" "[A-M]" <original >newfile &&
+ rm -f original &&
+ git commit -a -m"side edits once again." &&
+
+ git branch test-rebase side &&
+ git branch test-rebase-pick side &&
+ git branch test-reference-pick side &&
+ git checkout -b test-merge side
+'
+
+test_expect_success 'reference merge' '
+ git merge -s recursive "reference merge" HEAD master
+'
+
+test_expect_success rebase '
+ git checkout test-rebase &&
+ git rebase --merge master
+'
+
+test_expect_success 'merge and rebase should match' '
+ git diff-tree -r test-rebase test-merge >difference &&
+ if test -s difference
+ then
+ cat difference
+ (exit 1)
+ else
+ echo happy
+ fi
+'
+
+test_expect_success 'rebase the other way' '
+ git reset --hard master &&
+ git rebase --merge side
+'
+
+test_expect_success 'merge and rebase should match' '
+ git diff-tree -r test-rebase test-merge >difference &&
+ if test -s difference
+ then
+ cat difference
+ (exit 1)
+ else
+ echo happy
+ fi
+'
+
+test_expect_success 'picking rebase' '
+ git reset --hard side &&
+ git rebase --merge --onto master side^^ &&
+ mb=$(git merge-base master HEAD) &&
+ if test "$mb" = "$(git rev-parse master)"
+ then
+ echo happy
+ else
+ git show-branch
+ (exit 1)
+ fi
+'
+
+test_done
--
1.4.0.g1910f
Junio C Hamano [off-list ref] wrote:
Eric Wong [off-list ref] writes:
quoted
This solves the problem of rebasing local commits against an
upstream that has renamed files.
I think leveraging the merge strategy to perform rebase is
sound, but the selection of merge base for this purpose is quite
different from the regular merge, and I think unfortunately this
patch is probably wrong in letting git-merge choose the merge
base.
But let's mention other things as well.
- You kept the original "format-patch piped to am" workflow
optionally working.
I left it as the default, too. I figured that it's best not
to change the default (and most likely faster) behavior of
something people rely on.
- You check if merge or patch was used for failed rebase and
follow the appropriate codepath while resuming, which is
good.
- The list of commits you generate with tac seem to include
merge commit -- you may want to give --no-merges to
rev-list.
Good point, I'll change it.
- I do not think we use "tac" elsewhere -- is it portable
enough?
Nope. perl -e 'print reverse <>' is equivalent and we already use
plenty of perl.
- Exiting with success unconditionally after "git am" feels
wrong. I would do "exit $?" instead of "exit 0" there.
Oops, I'll change that, too.
Suppose you have this commit ancestry graph:
----------------------------------------------------------------
Example: git-rebase --onto master A topic
A---B---C topic B'--C' topic
/ --> /
D---E---F---G master D---E---F---G master
----------------------------------------------------------------
This is slightly different from the one at the beginning of the
script. The idea is A turned out to be not so cool, and we
would want to drop it.
quoted
+call_merge () {
+ cmt="$(cat $dotest/`printf %0${prec}d $1`)"
+ echo "$cmt" > "$dotest/current"
+ git-merge $strategy_args "rebase-merge: $cmt" HEAD "$cmt" \
+ || die "$MRESOLVEMSG"
+}
call_merge is first called with B in cmt, and HEAD is pointing
at G. But the merge in this function makes a merge between B
and G, taking the effect of E->A.
I think the three-way merge you would want here is not between B
and G using E as the pivot, but between B and G using A as the
pivot. That's how cherry-pick and revert works. I would
leverage the interface that is one level lower for this -- the
strategy modules themselves.
git-merge-$strategy $cmt^ -- HEAD $cmt
Changing the 'git-merge $strategy_args "rebase-merge: $cmt" HEAD "$cmt"'
line in call_merge() to this seems to have broken more tests.
I'm not an expert at merging strategies by any measure, I've just
trusted merge-recursive to Do The Right Thing(TM) more often than not,
and use rerere to avoid repeating work.
I'm not entirely sure I want (or fully understand how) to support
cherry-picking/revert with this, as I've already dropped --skip when
--merge was in use.
I'll think about this some more when I'm less distracted.
The strategy modules take merge base(s), double-dash as the
separator, our head and the other head. They do not make commit
themselves (instead they leave working tree and index in
committable state) and signal the results with their exit
status:
0 -- success
1 -- conflicts
2 -- did not handle the merge at all
Cool, that's good.
--
Eric Wong