On Tue, Jan 12, 2021 at 09:17:37AM -0500, Jeff King wrote:
> It looks like this was broken in v2.10.0, via dfb7a1b4d0 (patch-ids:
> stop using a hand-rolled hashmap implementation, 2016-07-29).
>
> I think the issue is that it is allowing duplicate entries in the
> hashmap. The algorithm is something like:
>
> - iterate over left-hand commits, inserting patch-id for each into
> hashmap
>
> - iterate over right-hand commits, seeing if any are present in
> hashmap. If so, we exclude the commit _and_ mark the patch-id as
> "seen"
>
> - iterate again over left-hand commits, omitting any whose patch-ids
> are marked as "seen"
>
> So if two commits on the left-hand side have the same patch-id, if we
> insert two entries into the hashmap, then only one of them is going to
> get its "seen" flag marked in the middle step.
Yeah, that's definitely it. Here's what the fix would look like directly
on top of dfb7a1b4d0:
diff --git a/patch-ids.c b/patch-ids.c
index db31fa647a..a8ed4f0872 100644
--- a/patch-ids.c
+++ b/patch-ids.c
@@ -66,12 +66,24 @@ struct patch_id *add_commit_patch_id(struct commit *commit,
struct patch_ids *ids)
{
struct patch_id *key = xcalloc(1, sizeof(*key));
+ struct patch_id *old;
if (init_patch_id_entry(key, commit, ids)) {
free(key);
return NULL;
}
+ /*
+ * Don't allow duplicates. Note that we can't use hashmap_put here; we
+ * must retain the _old_ struct, because some commit->util is pointing
+ * to it.
+ */
+ old = hashmap_get(&ids->patches, key, ids);
+ if (old) {
+ free(key);
+ return old;
+ }
+
hashmap_add(&ids->patches, key);
return key;
}
diff --git a/t/t6007-rev-list-cherry-pick-file.sh b/t/t6007-rev-list-cherry-pick-file.sh
index 28d4f6b259..0c5f1dcc54 100755
--- a/t/t6007-rev-list-cherry-pick-file.sh
+++ b/t/t6007-rev-list-cherry-pick-file.sh
@@ -207,4 +207,16 @@ test_expect_success '--count --left-right' '
test_cmp expect actual
'
+test_expect_success '--cherry-pick with duplicates on each side' '
+ git checkout -b dup-orig &&
+ test_commit dup-base &&
+ git revert dup-base &&
+ git cherry-pick dup-base &&
+ git checkout -b dup-side HEAD~3 &&
+ test_tick &&
+ git cherry-pick -3 dup-orig &&
+ git rev-list --cherry-pick dup-orig...dup-side >actual &&
+ test_must_be_empty actual
+'
+
test_done
Unfortunately we can't do the same thing now. The "seen" flag went away
in 683f17ec44 (patch-ids: replace the seen indicator with a commit
pointer, 2016-07-29), in favor of having each patch_id struct point to a
"struct commit". We could revert that, or turn that pointer into a list,
but it gets worse. Later, b3dfeebb92 (rebase: avoid computing
unnecessary patch IDs, 2016-07-29) built on that by lazily computing the
patch-ids. So each patch_id struct must point to exactly one commit!
We'll have to either:
- split them apart into two structs at the time of the lazy computation
or
- accept having duplicate patch_ids, but on lookup make sure we find
_all_ matches, not just the first one.
Hmm, that last one might not be too bad, since we have
hashmap_get_next() (it does mean iterating over a chained bucket
linearly, but that's already the worst-case for a hash lookup).
-Peff
