https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118856
--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Oh, the lack of the allocator calls is due to constexpr and those doing nothing
at all.
I wonder why it isn't being optimized out for the pre- r15-7481 case. And why
do the ~allocator calls remain even if they are equally constexpr and do
nothing at all.
Anyway, here is a full self-contained testcase which shows the lack of proper
cleanup nesting:
int a, b, c, d;
struct A { A () { ++a; ++d; }; ~A () { --a; ++d; }; };
struct B { B (const A & = A {}) { ++b; ++d; }; ~B () { --b; ++d; }; };
struct C {
C (const B &, const A & = A {}) { ++c; ++d; throw 42; };
~C () { --c; ++d; };
int *begin () { return nullptr; };
int *end () { return nullptr; };
};
void
foo ()
{
for (auto &i : C { B {} })
;
}
int
main ()
{
try
{
foo ();
__builtin_abort ();
}
catch (int x)
{
if (x != 42 || a || b || c != 1 || d != 7)
__builtin_abort ();
}
}
This works with r15-7480 -O2 -std=c++20 and -O2 -std=c++23 and with current
trunk with -O2 -std=c++20 but aborts with -O2 -std=c++23.
