On Thu, Feb 13, 2025 at 10:38:51PM +0100, René Scharfe wrote:

Am 12.02.25 um 08:05 schrieb Patrick Steinhardt:

quoted

And this is fixed by immediately processing all commits that we
currently have in the list. As `clear_commit_marks_1()` only processes
immediate children of the handed-in commit we know that it will have
processed the first parent, and `list` will contain remaining commits,
if any.

clear_commit_marks_1() processes the whole ancestral chain of first
parents down to the root or first clean ancestor.

quoted

We also end up adding grandchildren to the list, so this change
essentially converts the algorithm from depth-first to breadth-first.

It's still depth-first, but all second and higher numbered parents added
to the list are cleaned before starting to clean the next commit from
the input list.  So we go from "clean straight down for each input
commit first and clean their side branches later" to "clean all
ancestors for each input commit".

Ah, fair.

quoted

I bet we can construct cases where this will perform _worse_ than the
current algorithm, e.g. when you have branch thickets where every commit
is a merge: But I assume that for the most common cases this might be an
improvement indeed.

I won't bet, but I'd like to see such a case.  Can't imagine one.

quoted

The question to me is: does this actually matter in the real world? It
would be nice to maybe get some numbers that demonstrate the improvement
in a repository.

Well, the maximum list length for clear_commit_marks_many() calls with
nr > 1 in the test suite goes from 12 in t6600 to 4 with the patch.  Not
that exciting.  The question to me is: Why pile up parents in the list
when we can clean them earlier with no downside?  Or is there any?

If it really is without downsides then yes, it's a nice improvement. I
was mostly wondering whether you're fixing something where the old way
of doing things performs _significantly_ worse and where the change
could lead to a user-visible improvement. Like, requiring us to store
orders of magnitudes less commits at the same time.

Patrick