Stop gv_try_downgrade from anonymising referenced CVs
I keep discovering ways in which gv_try_downgrade, which is supposed
to be an optimisation, changes observable behaviour even without look-
ing at the stash.
This one had me confused at first:
$ ./perl -Ilib -e 'use constant foo=>1; BEGIN { $x = \&foo } undef &$x; $x->()'
Undefined subroutine called at -e line 1.
$ ./perl -Ilib -e 'use constant foo=>1; BEGIN { $x = \&{"foo"} } undef &$x; $x->()'
Undefined subroutine &main::foo called at -e line 1.
Notice how the first example (where gv_try_downgrade kicks in)
shows no name in the error message. This only happens on non-
threaded builds.
What’s happening is that, when the BEGIN block is freed, the GV op
corresponding to &foo get freed, triggering gv_try_downgrade, which
checks to see whether it can downgrade the GV to a simple reference
to a constant (the way constants are stored by default). It then pro-
ceeds to do that, so the GV qua GV ceases to exist, and the CV gets
automatically anonymised as a result (the same thing happens with
‘$x = \&{"foo"}; Dump $x; delete $::{foo}’, but legitimately in
that case).
The solution here is to check the reference count on the CV before
downgrading the GV. If the CV’s reference count > 1, then we should
leave it alone.