Re: [PATCH v2 5/7] merge-recursive: add merge function arg to 'merge_recursive_generic'

[Elijah Newren <newren@gmail.com>]

On Mon, May 9, 2022 at 12:24 PM Victoria Dye <vdye@github.com> wrote:
>
> Elijah Newren wrote:
> > On Wed, Apr 27, 2022 at 11:16 AM Victoria Dye via GitGitGadget
> > <gitgitgadget@gmail.com> wrote:
> >>
> >> From: Victoria Dye <vdye@github.com>
> >>
> >> Replace the hardcoded 'merge_recursive()' function used by the
> >> 'merge_recursive_generic()' with a caller-specific merge function. This will
> >> allow us to use 'merge_ort_recursive()' (and therefore avoid the index
> >> expansion of 'merge_recursive()') in commands that perform merges with
> >> 'merge_recursive_generic()', such as 'git stash pop'.
> >>
> >> Note that this patch is strictly a refactor; all callers still use
> >> 'merge_recursive()', and any changing to 'merge_ort_recursive()' will be
> >> done in a later commit.
> >
> > I'm not sure if we can gut merge_recursive_generic(), but I don't
> > think stash should use it...
> >
> >> Signed-off-by: Victoria Dye <vdye@github.com>
> >> ---
> >>  builtin/am.c              | 2 +-
> >>  builtin/merge-recursive.c | 2 +-
> >>  builtin/stash.c           | 2 +-
> >>  merge-ort.c               | 3 ++-
> >>  merge-recursive.c         | 4 ++--
> >>  merge-recursive.h         | 9 ++++++++-
> >>  6 files changed, 15 insertions(+), 7 deletions(-)
> >>
> >> diff --git a/builtin/am.c b/builtin/am.c
> >> index 0f4111bafa0..6d01185d122 100644
> >> --- a/builtin/am.c
> >> +++ b/builtin/am.c
> >> @@ -1614,7 +1614,7 @@ static int fall_back_threeway(const struct am_state *state, const char *index_pa
> >>         if (state->quiet)
> >>                 o.verbosity = 0;
> >>
> >> -       if (merge_recursive_generic(&o, &our_tree, &their_tree, 1, bases, &result)) {
> >> +       if (merge_recursive_generic(&o, &our_tree, &their_tree, 1, bases, merge_recursive, &result)) {
> >>                 repo_rerere(the_repository, state->allow_rerere_autoupdate);
> >>                 free(their_tree_name);
> >>                 return error(_("Failed to merge in the changes."));
> >> diff --git a/builtin/merge-recursive.c b/builtin/merge-recursive.c
> >> index b9acbf5d342..687ed1e527b 100644
> >> --- a/builtin/merge-recursive.c
> >> +++ b/builtin/merge-recursive.c
> >> @@ -81,7 +81,7 @@ int cmd_merge_recursive(int argc, const char **argv, const char *prefix)
> >>         if (o.verbosity >= 3)
> >>                 printf(_("Merging %s with %s\n"), o.branch1, o.branch2);
> >>
> >> -       failed = merge_recursive_generic(&o, &h1, &h2, bases_count, bases, &result);
> >> +       failed = merge_recursive_generic(&o, &h1, &h2, bases_count, bases, merge_recursive, &result);
> >>
> >>         free(better1);
> >>         free(better2);
> >> diff --git a/builtin/stash.c b/builtin/stash.c
> >> index 1bfba532044..16171eb1dab 100644
> >> --- a/builtin/stash.c
> >> +++ b/builtin/stash.c
> >> @@ -554,7 +554,7 @@ static int do_apply_stash(const char *prefix, struct stash_info *info,
> >>         bases[0] = &info->b_tree;
> >>
> >>         ret = merge_recursive_generic(&o, &c_tree, &info->w_tree, 1, bases,
> >> -                                     &result);
> >> +                                     merge_recursive, &result);
> >>         if (ret) {
> >>                 rerere(0);
> >>
> >> diff --git a/merge-ort.c b/merge-ort.c
> >> index 8545354dafd..4bccdfcf355 100644
> >> --- a/merge-ort.c
> >> +++ b/merge-ort.c
> >> @@ -4737,7 +4737,8 @@ void merge_incore_recursive(struct merge_options *opt,
> >>         trace2_region_enter("merge", "incore_recursive", opt->repo);
> >>
> >>         /* We set the ancestor label based on the merge_bases */
> >> -       assert(opt->ancestor == NULL);
> >> +       assert(opt->ancestor == NULL ||
> >> +              !strcmp(opt->ancestor, "constructed merge base"));
> >
> > ...and here's one of the reasons why.  The fact that
> > merge_recursive_generic() uses this string when exactly one merge base
> > is passed is something that is only correct for git-am; it is wrong
> > and actively misleading for git-stash since it has a real merge base
> > that is not internally constructed by the operation using the merge
> > machinery.  (The merge base it uses is something like $STASH^1, IIRC.)
> >
> > In fact, this was half the coin around why merge_recursive_generic()
> > wasn't converted when merge-ort was written; see
> > https://lore.kernel.org/git/CABPp-BHW61zA+MefvWK47iVZKY97rxc2XZ-NjXzuJxEhgSLqUw@mail.gmail.com/
> > and https://lore.kernel.org/git/CABPp-BFr=1iVY739cfh-1Hp82x-Mny-k4y0f3zZ_YuP3PxiGfQ@mail.gmail.com/
> > for more details.
> >
>
> All of that makes sense, thanks for the context!
>
> > The use of merge_recursive_generic() by stash is also a bit weird;
> > most of the time, stash is going to have actual commits instead of
> > just trees.  But stash dereferences those commits to trees, passes
> > them to merge_recursive_generic(), and then merge_recursive_generic()
> > has to create fake commits containing those trees, because the merge
> > machinery wants commits.  It feels a bit like a Rube Goldberg machine.
> > Also, stash also always calls merge_recursive_generic() with exactly
> > one merge base, which together with having real commits both kind of
> > defeat the need for "generic".    I think stash should just use
> > merge_trees()/merge_incore_nonrecursive() directly (much as
> > sequencer.c does).  The only special case to worry about with stash is
> > when c_tree != HEAD^{tree}, i.e. when applying changes on top of
> > already present changes instead of just on top of HEAD.  But in that
> > case, I think stash should be the thing to create a fake commit rather
> > than invoking some wrapper that will create fake commits for all three
> > trees.
> >
>
> I'm a bit confused about this. The non-recursive merge functions
> ('merge_trees()' & 'merge_ort_nonrecursive()' or the lower-level
> 'merge_incore_nonrecursive()') merge trees, not commits, so performing a
> non-recursive merge requires dereferencing commits to trees anyway. I think
> I agree with your other message [1] that the 'stash' merge doesn't need to
> merge recursively, but that would mean it also doesn't use the commits
> *directly* (i.e., as arguments in the merge).
>
> Apologies if I'm missing something obvious, but I want to make sure I
> understand your suggestion.
>
> [1] https://lore.kernel.org/git/CABPp-BFANwZn73w8zrVySB7mh0bQQBdGJjBuSJy50UOeyYT6aA@mail.gmail.com/

Oh, right, it's only the recursive merge that needs commits (so that
it can find ancestors and ancestors of ancestors, etc.).  So, ignore
my comments about making fake commits; that's not needed.