StGIT and rerere

Hi,

It seems stgit does not use git-rerere, so why not? Any reason other than it hasn't been done yet?

I abuse stgit heavily, by frequently reording patches, which for some patches result in re-occuring conflicts. git-rerere seems to be the solution.

What's the "rules" for when to invoke rerere? It seems it is mostly automatic in git, but since only the porcelainish commands use it, that means StGIT doesn't.

So here is what I *think* needs to be done. Seems simple enough.

stg push, stg pick, stg import, stg goto, stg fold, stg float
	do what push does and invoke git-rerere at the end whether the push ends with 
conflicts or not

stg pop
	nothing, or do I need to remove rr-cache/MERGE_RR, like git-reset does?

stg status --reset, stg push --undo
	remove rr-cache/MERGE_RR ?

stg refresh
	do what stgit does normally and then invoke git-rerere

stg resolved:
	do what stgit does normally and then invoke git-rerere

stg clean, stge delete:
	remove rr-cache/MERGE_RR ?

Anything else or comments on this list?

-- robin

Robin Rosenberg <robin.rosenberg.lists@dewire.com> wrote:

I didn't know it exists. I've been thinking at a way to avoid duplicating the conflict fixing but haven't got to any results. This looks like a good idea.

My problem was with maintaining a public branch where re-basing patches is not welcomed but I would still like to use StGIT in my development branch. When pulling from upstream in my devel branch, I get conflicts in some patches. The problem is that I get the same conflicts in the patches already merged in the public branch where the patches were previously checked in.

Another case is several branches with common patches that generate conflicts.

It could probably be invoked by stgit.git.merge() if the git-read-tree (and maybe the diff3 merge) failed (BTW, I replaced git-read-tree with git-recursive-merge in my local tree and seems to detect renames properly; I'll push it once I'm convinced there are no problems).

Note, however, that I haven't looked at how git-rerere works and I might have misunderstood its functionality.

the git.merge() function handles all the merges.

I think pop shouldn't do anything.

Yes (not sure for push --undo).

Why should it invoke git-rerere? This just creates a commit. Or is it needed for storing rerere ids?

No need to call rerere here since resolved is an StGIT-only function, it doesn't affect the repository (it just unmarks the conflict files so that stgit allows you to refresh).

That's not needed. Delete can act like pop for the top patch.

Catalin Marinas <catalin.marinas@arm.com> writes:

I think Documentation/git-rerere.txt describes its operation in enough details for the end user, but if there is anything unclear please ask away.

There are two details that are not mentoined in the end-user documentation.

* Identifying "same" conflicts.

This is done by hashing the text inside the conflict markers:

	<<<<<<<
        a
        =======
	b
        >>>>>>>

Each conflicted section like the above are first canonicalized so that section "a" sorts alphabetically earlier than section "b" (in other words, if the conflict were <<< b === a >>> then they are swapped to read <<< a === b >>>), concatenated and fed into SHA-1 digest. The final digest after conflicted file is read as a whole becomes the name of the conflict, and naming is done per file.

The reason for this canonicalization is because the order in the conflicted section depends on the direction of the merge (if you pull test into master you get the conflict marked the other way from the case you pulled master into test).

* rerere's resolution can affect outside conflict markers.

If an earlier conflict read like this:

	...
	z
	<<<<<<<
        a
        =======
        b
        >>>>>>>
        c
	d
	e
        f
        g
	...
        
and your earlier resolution was like this:

	...
	w
        b
        dd
        e
        f
        g
	...

That is what rerere records (i.e. including removal of c and change from z to w).

Then, when you see the same conflict like this (notice that g is now h):

	...
	z
	<<<<<<<
        a
        =======
        b
        >>>>>>>
        c
	d
	e
        f
        h
	...

it is resolved like this:

	...
	w
        b
        d
        e
        f
        h
	...

This is often useful when e.g. a <<< === >>> section changes function signature of a (static) function; parts outside the conflicted section needs matching changes to adjust to it and that is recorded and replayed by rerere.