On Sat, Jan 20, 2018 at 7:00 PM, Josh Bleecher Snyder <josharian@xxxxxxxxx> wrote: > Hi, all. > > I'm experimenting with some new porcelain for interactive rebase. One > goal is to leave the work tree untouched for most operations. It looks > to me like 'git merge-tree' may be the right plumbing command for > doing the merge part of the pick work of the todo list, one commit at > a time. If I'm wrong about this, I'd love pointers; what follows may > still be interesting anyway. I don't have a concrete alternative (yet?) but here are some pointers to two alternate merge-without-touching-working-tree possibilities, if your current route doesn't pan out as well as you like: I posted some patches last year to make merge-recursive.c be able to do merges without touching the working tree. Adding a few flags would then enable it for any of 'merge', 'cherry-pick', 'am', or 'rebase'...though for unsuccessful merges, there's a clear question of what/how conflicts should be reported to the user. That probably depends a fair amount on the precise use-case. Although that series was placed on the backburner due to the immediate driver of the feature going away, I'm still interested in such a change, though I think it would fall out as a nice side effect of implementing Junio's proposed ideal-world-merge-recursive rewrite[1]. I have started looking into that[2], but no guarantees about how quickly I'll find time to finish or even whether I will. [1] https://public-inbox.org/git/xmqqd147kpdm.fsf@xxxxxxxxxxxxxxxxxxxxxxxxxxx [2] https://github.com/newren/git/blob/ort/ort-cover-letter contains overview of ideas and notes to myself about what I was hoping to accomplish; currently it doesn't even compile or do anything > 4. API suggestion > > Here's what I really want 'git merge-tree' to output. :) ... > If the merge had conflicts, write the "as merged as possible" tree to You'd need to define "as merged as possible" more carefully, because I thought you meant a tree containing all the three-way merge conflict markers and such being present in the "resolved" file, but from your parenthetical note below it appears you think that is a different tree that would also be useful to diff against the first one. That leaves me wondering what the first tree is. (Is it just the tree where for each path, if that path had no conflicts associated with it then it's the merge-resolved-file, and otherwise it's the file contents from the merge-base?). Both of these trees are actually rather non-trivial to define. The wording above isn't actually sufficient, because content conflicts aren't the only kind of conflict. More on that below. There is already a bunch of code in merge-recursive.c to create a forcibly-merged-accepting-conflict-markers-in-the-resolution and record it as a tree (this is used for creating virtual merge bases in the recursive case, namely when there isn't a single merge-base for the two branches you are merging). It might be reusable for what you want here, but it's not immediately clear whether all the things it does are appropriate; someone would have to consider the non-content (path-based) conflicts carefully. > the object database and give me its sha, and then also give me the > three-way merge diff output for all conflicts, as a regular patch > against that tree, using full path names and shas. (Alternatively, > maybe better, give me a second sha for a tree containing all the > three-way merge diff patches applied, which I can diff against the > first tree to find the conflict patches.) As far as I can tell, you're assuming that it's possible with two trees that are crafted "just right", that you can tell where the merge conflicts are, with binary files being your only difficulty. Content conflicts aren't the only type that exist; there are also path-based conflicts. These type of conflicts also make it difficult to know how the two trees you are requesting should even be created. For example, if there is a modify/delete conflict, how can that be determined from just two trees? If the first tree has the base version of the file, then the second tree either has a file at the same position or it doesn't. Neither case looks like a conflict, but the original merge had one. You need more information. The exact same thing can be said for rename/delete conflicts. Similarly, rename/add (one side renames an existing file to some new path (say, "new_path"), and the other adds a brand new file at "new_path), or rename/rename(2to1) (each side renames a different file to the same location), won't be detectable just by diffing two trees. These are often handled by moving both files to some other location, so there's no way to record in a tree that there was a conflict. rename/rename(1to2) is similar, but instead of two different original files being renamed to the same thing, this is one file being renamed differently on different sides of history. I know that several of the examples above involved rename detection, which git-merge-trees won't even do, but that means you're even more likely to face the modify/delete conflict cases. And our list still isn't done, either: Directory/file conflicts (one side puts a directory of the same name that the other side adds as a file) will also cause problems. Finally, directory rename detection (currently in pu under review) adds a few "implicit dir rename" conflict types (renames of multiple directories would cause multiple files to be renamed to the same location, or an existing file/dir being in the way of one or more path(s) getting implicitly renamed). This means that the number of types of non-textual conflicts might also grow in the future so it may be unwise to try to special case existing exceptions with a bag of clever workarounds. Hope that helps, Elijah
