On Wed, Dec 05, 2018 at 03:49:26PM -0500, Jason Merrill wrote:
> On 11/28/18 3:42 AM, Jakub Jelinek wrote:
> > Whenever we need to clone a cdtor (either because the target doesn't support
> > aliases the way we need, e.g. initlist105.C testcase on darwin, where it has
> > been originally reported, or when it has virtual bases, like the made up
> > initlist106.C on x86_64-linux), we rely on DECL_INITIAL of the local
> > variables being unshared, because the tree unsharing gimplify.c performs
> > doesn't unshare DECL_INITIAL. clone_body has some code to recurse on the
> > DECL_INITIAL, but it handles just decls FOR_EACH_LOCAL_DECL walks. I
> > believe it is generally ok that not all temporaries are covered in local
> > decls, the gimplifier should take care of those fine if we don't need
> > debug info for them.
>
> I think any temporaries that have DECL_INITIAL should be pushed so that they
> end up in local_decls. set_up_extended_ref_temp already adds a DECL_EXPR
> for it, I guess we need a pushdecl as well. Though the comment for
> get_temp_regvar suggests that this is problematic somehow.
Ok, will play with it tomorrow.
Jakub
