>> >> When setting a localized signal handler, the system signal handler
>> >> is set to SIG_DFL then back to perl's signal handler. This briefly
>> >> exposes SIG_DFL when switching between alternate non-default signal
>> >> handlers.
> >
> > The below patch fixes this bug.  In the process it also partially fixes a
> > magic bug of long standing (probably since 5.000).
> >
> > When localizing a magical scalar for assignment, Perl has until now done an
> > extra store of undef before storing the actual desired value.  To
> > illustrate, given this source code:
> >
> >     { package Foo;
> >       sub TIEHASH { bless {}, 'Foo' }
> >       sub FETCH { print "Fetch $_[1]\n"; $_[0]->{$_[1]} }
> >       sub STORE { print "Store $_[1] = $_[2]\n"; $_[0]->{$_[1]} = $_[2] }
> >     }
> >
> >     tie %x, 'Foo';
> >     $x{plugh} = "dick";
> >     { local $x{plugh} = "jane" }
> >
> > Released perls and blead do this:
> >
> >  $ perl foo
> >  Store plugh = dick
> >  Fetch plugh
> >  Store plugh =
> >  Store plugh = jane
> >  Store plugh = dick
> >
> > Whereas blead with the below patch does this:
> >
> >  $ ./perl foo
> >  Store plugh = dick
> >  Fetch plugh
> >  Store plugh = jane
> >  Store plugh = dick
> >
> > The below patch fixes this problem for hash elements and slices.  However,
> > due to the OPf_SPECIAL flag not meaning the same thing in the AELEM opcodes,
> > let alone all the opcodes that can extract scalar values, this fix is not
> > entirely applicable to those cases; that will require deeper hacking.  At
> > least this patch fixes hashes, which are the most common case.

Great fix! Thanks, applied as #34819, except the perlapi.c part, which
strips off a Tolkien quote. I'll fix that separately.