Re: [EGIT PATCH] Avoid breaking the walk prematurely

Robin Rosenberg [off-list ref] wrote:

The optimization broke the walk when all changes we adds, assuming
the parent was an empty tree. This assumption was wrong. The tree does
not have to bee empty, it can contain other trees and blobs.

http://code.google.com/p/egit/issues/detail?id=46

Signed-off-by: Robin Rosenberg <redacted>
---
 .../spearce/jgit/revwalk/RewriteTreeFilter.java    |    7 -------
 1 files changed, 0 insertions(+), 7 deletions(-)

We could introduce a fixed version later on if we supply an option to select
whether to optimize the walk this way. The problem with doing it without an
option is that imports from legacy SCMs do not necessarily record file moves
as add/removes in direct sucession, but other events may happen in beteens
for various reasons.

This is a fairly useful optimization in git.  Most people use it,
because most people don't pass "--full-history" when they run
"git log -- path".  So I think we really want to do it by default
in jgit.

Enabling this optimization again probably requires adding a new
method to TreeFilter which allows the filter graph to return only
nodes which select paths by name, and then do something like:

  else if (chgs == adds && isPathSetEmpty(c.getTree())) {
    c.parents = RevCommit.NO_PARENTS;
    return true;
  }

where isPathSetEmpty runs a second TreeWalk iteration but uses the
new (smaller) filter graph that only matches on paths.

I'm not going to write that out today.  But I thought I'd at least
write an email while the idea was fresh in my head.
 
-- 
Shawn.