On Mon, Jan 28, 2019 at 8:52 AM Jonathan Wakely <jwak...@redhat.com> wrote:
>
> On 28/01/19 14:29 +0100, Jakub Jelinek wrote:
> >On Mon, Jan 28, 2019 at 01:55:38PM +0100, Richard Biener wrote:
> >> So I guess in the end we're being lucky. Somehow. I've played with
> >>
> >> __attribute__((noipa)) void
> >> qux ()
> >> {
> >> S buf1;
> >> foo ((char *)&buf1);
> >> S *p = new (&buf1) (S);
> >> bar (p);
> >> p->~S ();
> >> {
> >> char buf2[128];
> >> baz (buf2);
> >> }
> >> }
> >
> >I'd think the above is already invalid, by doing a placement new into
> >a variable with non-trivial ctor and dtor while it is still constructed,
> >then destruct the placement new created var in there and after a while
> >destruct the original variable doesn't feel right to me, but I'm not a C++
>
> Right. When the second object is constructed in that location, the
> lifetime of the first one ends. When the destructor is automatically
> run at the end of the scope you're destroying something that is no
> longer alive, so undefined.
Indeed.
> >language lawyer. I'd expect that usually either the whole var has
> >char/std::byte etc. array type, or the placement new is into a field inside
> >of some class (again char/std::byte etc. array type).
> >Would could be valid is:
>
> Yeah I think the one below is OK. I'm still looking at the original
> testcase at the top of the thread.
The original testcase looks good to me. And I agree with Jakub's
point, that destroying an object created in a buffer is necessarily
different from destroying the buffer itself.
Jason
