Improve merge performance by avoiding in-index merges.
authorShawn O. Pearce <spearce@spearce.org>
Thu, 28 Dec 2006 07:35:34 +0000 (02:35 -0500)
committerJunio C Hamano <junkio@cox.net>
Wed, 10 Jan 2007 23:57:44 +0000 (15:57 -0800)
In the early days of Git we performed a 3-way read-tree based merge
before attempting any specific merge strategy, as our core merge
strategies of merge-one-file and merge-recursive were slower script
based programs which took far longer to execute.  This was a good
performance optimization in the past, as most merges were able to
be handled strictly by `read-tree -m -u`.

However now that merge-recursive is a C based program which performs
a full 3-way read-tree before it starts running we need to pay the
cost of the 3-way read-tree twice if we have to do any sort of file
level merging.  This slows down some classes of simple merges which
`read-tree -m -u` could not handle but which merge-recursive does
automatically.

For a really trivial merge which can be handled entirely by
`read-tree -m -u`, skipping the read-tree and just going directly
into merge-recursive saves on average 50 ms on my PowerPC G4 system.
May sound odd, but it does appear to be true.

In a really simple merge which needs to use merge-recursive to handle
a file that was modified on both branches, skipping the read-tree
in git-merge saves on average almost 100 ms (on the same PowerPC G4)
as we avoid doing some work twice.

We only avoid `read-tree -m -u` if the only strategy to use is
merge-recursive, as not all merge strategies perform as well as
merge-recursive does.

Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
git-merge.sh

index 477002910ede7f6bc6bcb9a79f6ddebd248f6bd2..1c4f6693f5b6919da2a867a02465a382477cfcac 100755 (executable)
@@ -298,24 +298,30 @@ f,*)
        ;;
 ?,1,*,)
        # We are not doing octopus, not fast forward, and have only
-       # one common.  See if it is really trivial.
-       git var GIT_COMMITTER_IDENT >/dev/null || exit
-
-       echo "Trying really trivial in-index merge..."
+       # one common.
        git-update-index --refresh 2>/dev/null
-       if git-read-tree --trivial -m -u -v $common $head "$1" &&
-          result_tree=$(git-write-tree)
-       then
-           echo "Wonderful."
-           result_commit=$(
-               echo "$merge_msg" |
-               git-commit-tree $result_tree -p HEAD -p "$1"
-           ) || exit
-           finish "$result_commit" "In-index merge"
-           dropsave
-           exit 0
-       fi
-       echo "Nope."
+       case " $use_strategies " in
+       *' recursive '*|*' recur '*)
+               : run merge later
+               ;;
+       *)
+               # See if it is really trivial.
+               git var GIT_COMMITTER_IDENT >/dev/null || exit
+               echo "Trying really trivial in-index merge..."
+               if git-read-tree --trivial -m -u -v $common $head "$1" &&
+                  result_tree=$(git-write-tree)
+               then
+                       echo "Wonderful."
+                       result_commit=$(
+                               echo "$merge_msg" |
+                               git-commit-tree $result_tree -p HEAD -p "$1"
+                       ) || exit
+                       finish "$result_commit" "In-index merge"
+                       dropsave
+                       exit 0
+               fi
+               echo "Nope."
+       esac
        ;;
 *)
        # An octopus.  If we can reach all the remote we are up to date.