patchFixes git-cherry algorithmic flaws

Old algorithm:
        - printed IDs of identical patches with minus (-) sign; they
	  should not be printed at all
        - did not print anything from the changes in the upstream

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
---
 git-cherry.sh |   26 +++++++++++++++++++++++++-
 1 files changed, 25 insertions(+), 1 deletions(-)

Dear diary, on Mon, Aug 07, 2006 at 12:30:13PM CEST, I got a letter where Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> said that...

Ping? Is this patch bogus or was it just forgotten?

Petr Baudis <pasky@suse.cz> writes:

These are not fixes to "algorithmic flaws". It's more like that Ilpo is writing a different program to fill different needs, and I did not see what workflow wanted to have the list of changes that were in the upstream and our changes. Maybe what Ilpo wanted to see was something like "git log upstream...mine" (three-dots not two to mean symmetric difference). I dunno. That operation certainly did not exist when we did git-cherry originally.

The original purpose of git-cherry (which probably is different from what Ilpo wanted to have, and that is why Ilpo modified it into a different program) is for a developer in the contributor role to see which ones of local patches have been accepted upstream and which ones still remain unapplied -- the intent is to help rebase only the latter and keep trying to convince upstream that these remaining ones are also worth applying.

So minus (-) lines are very much needed to if you want to see which ones have been accepted. Plus lines are used to pick which ones to rebase by older version of git-rebase, but I do not think we do that anymore. And in any case we are _not_ interested in whatever happened in the upstream that did not come from the branch we are looking at.

I suspect we do not use it anywhere anymore. Maybe we can remove it?

	... goes and looks ...
	git grep -e git.cherry --and --not -e git.cherry-pick

Nah, no such luck. One of the documentation suggests that you drive cvsexportcommit using its output, like this:

	git cherry cvs mine | sed -n -e 's/^\+ //p' |
        xargs -L 1 git-cvsexportcommit -c -p -v

and I can see why cherry is (perhaps slightly) more desirable than "git rev-list cvs..mine"

So unless we come up with an alternative way to do this, we cannot change it or drop it. Not yet.

Dear diary, on Sun, Sep 24, 2006 at 03:49:22AM CEST, I got a letter where Junio C Hamano <junkio@cox.net> said that...

Hmm, well, what's curious is that the documentation says

	Every commit with a changeset that doesn't exist in the other branch
	has its id (sha1) reported, prefixed by a symbol.  Those existing only
	in the <upstream> branch are prefixed with a minus (-) sign, and those
	that only exist in the <head> branch are prefixed with a plus (+)
	symbol.

which is in contradiction of Ilpo's description of the old algorithm (and also your description of it). It would seem he just wants to fix it according to the documented behaviour.

I guess the documentation is what's broken then?

Petr Baudis <pasky@suse.cz> writes:

Ah I did not realize that, but yes the documentation is incorrect.

I wonder if we can kill it by introducing a new rev notation and using regular rev-list family of commands instead.

What we want here is a way to say "give me commits that are in B but not in A, but before returning a commit see if there is an equivalent change in the set of commits that are in A but not in B, and filter it out".

Time for "rev-list A....B"? ;-)

On Sun, 24 Sep 2006, Junio C Hamano wrote:

I was going to do a same conclusion but didn't send it just yet... :-) I found out what the documentation says when looking a tool to do a job. Then I wonder how such obvious bug could have passed unnoticed... Of course I have no clue what the "original purpose" is supposed to be... ;-) Then I "fixed" it and as it is _so easy_ to send patches with git I thought I could contribute the "fix"... I was a bit turned down though from not receiving any reply or so, well, until now... :-) Though I remember now that I was wandering whether the tool was correct and that documentation is not... But since I thought that when I'm cherry-picking (and, e.g., cleaning up log messages) between topic-old and topic-cleaner, the patch id based _difference_ seems to be the most useful one...

I think that your formalization is very close to what I was expecting to 
get (sort of one-way definition)... However, my git-cherry way produces 
"difference" but on a higher level (than git-diff) since it includes both 
+ and - "changes". Of course, when I have then modified one of the 
changesets slightly, I have different patch id, and thus + and - with same 
log message (with verbose), which IMHO is a good thing to notice, 
especially if I return to the work after 2 weeks or so :-).  

A real life example: In a branch, I have changed tcp_packets_in_flight (~10 callers) to input sk instead of tp in a single changeset and >10 minor changesets. I would love to see tcp_packets_in_fligth change information just once when doing diffing topic-old topic-new during cherry picking, instead of a lengthy diff full of search-and-replace "noise", which increases possiblity of an human error...

But anyway, I'm not claiming that your approach is less useful...

Ilpo Järvinen wrote:

+1 on an idea that git-cherry is "diff of logs". Perhaps to add some header. Of course use patch_ids _and_ the commit message to compare (we can have both match, patch id match only, commit message match only, and commit title (first line) match only), but that can be selected using options.