[Svnmerge] [PATCH] Performance patch for merging

[Giovanni Bajo]

Raman Gupta <rocketraman at fastmail.fm> wrote:

>> The attached patch fixes a performance issue with the merge command. In
>> the current version, the merge command does not remove phantom revisions
>> from the list of revs that are used for 1) merging itself, and 2)
>> obtaining the log messages for each merged revision to build the commit
>> log message.

Alas, this patch is wrong.

You must not remove the phantom revisions from 1). In fact, it's done on
purpose: with phantom revisions, you get more compact ranges and you call
svn merge less times. For instance, if my property is "1-100", HEAD is 200,
and I simply call "svnmerge merge" I want to invoke "svn merge" exactly
*once*: to merge the interval 100-200. It doesn't matter if there are
phantom revisions in that range: svn merge will ignore those. By calling
svnmerge merge just once you get much faster results, and many less
conflicts.

In fact, I would *really* appreciate if you covered this usage with a
testcase, so that somebody else does not try again to skip phantom revisions
from merging.

As for 2), I know that log construction is very slow, but the problem should
be fixed otherwise. You should use minimal_merge_intervals() in
construct_merged_log_message(), so to call "svn log" for ranges of revisions
(instead of one revision at a time). Even here, the idea is to call "svn
log" less times as possible, so it doesn't really matter if we call it once
for a solid range like 100-200 including phantom revisions: those phantom
revisions will not appear in the "svn log" output and thus will not be part
of the commit message.
-- 
Giovanni Bajo