Jeff King <> writes:

> When showing a combined diff, we are comparing two (or more)
> parents to a final state, and some of these states may be
> the same.  Here's a table of the possible contents for a
> given path (for two parents, but we will generalize to more
> in a moment; we also omit isomorphic cases where the parents
> are swapped):
>   case | M | P1 | P2
>   -------------------
>   1    | A | A  | A
>   2    | A | A  | B
>   3    | A | B  | B
>   4    | A | B  | C
> In case 1, the path was not relevant to the merge at all,
> and we omit it as uninteresting. In case 2, we did resolve
> the path, but in favor of one side. We also consider this
> uninteresting and do not show the diff.
> In case 4, we had a real content-level merge, and the
> combined diff is interesting. We show it.
> That leaves case 3, in which both parents are the same, but
> the merge picks a new content. This should be rare in
> normal merges, though it could happen if you updated an
> unrelated file due to a resolution elsewhere (i.e., an evil
> merge that crosses a file boundary). But it happens
> frequently in the fake merge commits we create for stashes,
> in which one parent is the base of the stash and the other
> is the index (in which case it simply means that the index
> entry for the path was not touched).
> Right now, we treat it the same as case 4, and show a normal
> combined diff. However, the result is harder to read, and
> the combined nature of the diff gives no extra information;
> every marker in the combined diff will be identical for both
> parents.
> This patch adds a new option, "--simplify-combined-diff",
> which converts this case into a normal, non-combined diff.
> It would not be correct to simply omit it, because there
> really is an interesting change going from B..A. It's just
> that there are not two interesting changes, which the
> combined diff would show.
> When generalizing this to more than two parents, we have two
> options:
>   1. Either simply to a single parent content, or not at all.
>   2. Omit parents whose contents are duplicates of other
>      parents.
> For a case like "A B B C", option (2) would still result in
> a combined diff, but one with fewer sources. However, it
> would also be ambiguous. The parents in a combined diff are
> marked only by their position, so omitting a position means
> that a reader can no longer tell which line goes with which
> parent.
> Instead, we choose option (1). Either you get the full
> combined diff, or you get a normal non-combined diff.

Very nicely analyzed.  The changes to the code looked also good from
a cursory read.

To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to
More majordomo info at

Reply via email to