On Wed, Jan 15, 2025 at 10:27:56AM -0500, Jason Merrill wrote:
> > @@ -7432,12 +7426,18 @@ reshape_init_r (tree type, reshape_iter
> > {
> > vec<constructor_elt, va_gc> *v = 0;
> > CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, init);
> > - tree raw_init = cp_maybe_split_raw_data (d);
> > + bool inc_cur;
> > + reshape_iter dsave = *d;
> > + tree raw_init = cp_maybe_split_raw_data (d, &inc_cur);
> > CONSTRUCTOR_APPEND_ELT (v, NULL_TREE,
> > raw_init ? raw_init : d->cur->value);
> > if (has_designator_problem (d, complain))
> > - return error_mark_node;
> > - if (!raw_init)
> > + {
> > + if (!inc_cur)
> > + *d = dsave;
>
> Why restore *d on the error path? Won't the caller just return
> error_mark_node as well?
I vaguely remember from writing the original patch (which introduced
cp_maybe_split_raw_data) that here it actually mattered whether
in the error case we move the reshape_iter or not and that it broke
some tests, but I no longer remember what was it exactly.
Otherwise it would also just use consume_init like most of the other
cases.
I can surely try without that again and see what breaks.
Jakub
