Each CV usually has a pointer, CvGV(cv), back to the GV that corresponds
to the CV's name (or to *foo::__ANON__ for anon CVs). This pointer wasn't
reference counted, to avoid loops. This could leave it dangling if the GV
is deleted.
We fix this by:
For named subs, adding backref magic to the GV, so that when the GV is
freed, it can trigger processing the CV's CvGV field. This processing
consists of: if it looks like the freeing of the GV is about to trigger
freeing of the CV too, set it to NULL; otherwise make it point to
*foo::__ANON__ (and set CvAONON(cv)).
For anon subs, make CvGV a strong reference, i.e. increment the refcnt of
*foo::__ANON__. This doesn't cause a loop, since in this case the
__ANON__ glob doesn't point to the CV. This also avoids dangling pointers
if someone does an explicit 'delete $foo::{__ANON__}'.
Note that there was already some partial protection for CvGV with
commit
f1c32fec87699aee2eeb638f44135f21217d2127. This worked by
anonymising any corresponding CV when freeing a stash or stash entry.
This had two drawbacks. First it didn't fix CVs that were anonmous or that
weren't currently pointed to by the GV (e.g. after local *foo), and
second, it caused *all* CVs to get anonymised during cleanup, even the
ones that would have been deleted shortly afterwards anyway. This commit
effectively removes that former commit, while reusing a bit of the
actual anonymising code.