unset is not setting the magic

[Ovid]

On my Mac, I'm getting the following error when I run make test:

PERL_DL_NONLAZY=1 "/Users/ovid/perl5/perlbrew/perls/perl-5.38.2/bin/perl" "-MExtUtils::Command::MM" "-MTest::Harness" "-e" "undef *Test::Harness::Switches; test_harness(0, 'blib/lib', 'blib/arch')" t/*.t
t/unset.t .. 1/?
#   Failed test 'var should start as unset'
#   at t/unset.t line 9.
# Looks like you failed 1 test of 4.
t/unset.t .. Dubious, test returned 1 (wstat 256, 0x100)
Failed 1/4 subtests

Test Summary Report
-------------------
t/unset.t (Wstat: 256 (exited 1) Tests: 4 Failed: 1)
  Failed test:  1
  Non-zero exit status: 1
Files=1, Tests=4,  0 wallclock secs ( 0.01 usr  0.01 sys +  0.02 cusr  0.01 csys =  0.05 CPU)
Result: FAIL
Failed 1/1 test programs. 1/4 subtests failed.
make: *** [test_dynamic] Error 1

Somehow, the magic is not getting set when I do this:

my $var = unset;

[haarg]

Outside of some special cases, magic is not copied across assignments.

[Ovid]

Oh! Duh. Don't I feel stupid. Working on a project with magic now and that should have been dead obvious to me.

Originally I had unset($var) and that worked, but it doesn't work for this use case:

sub something ($self, $new = unset) {
    if ( is_set($new) ) {
        $self->{something} = $new;
    }
    else {
        return $self->{something};
    }
}

Any pointers on where I can read those special cases?

[Ovid]

Or is there some way the code can know what it's been assigned to and attach magic to that? (I assume the answer is "no")

[Leont]

Any pointers on where I can read those special cases?

The special case is tainting, and it's very special and widely hated.

[haarg]

vstring magic is the only one that "copies" across assignment, and it has special handling in sv_setsv_flags. tainting is another special case, but it's not even really doing copying. It sets a global flag that impacts every assignment in a given statement.

[Leont]

I almost got it to work using the vstring approach, but the magic vtable doesn't get copied along so it doesn't quite work. A one-line fix in core probably would make this possible (but should it?)

[Ovid]

@Leont What is the one-line fix, what does it do, and what would the impact be? Would it allow for more things like this to be possible?

[Leont]

@Leont What is the one-line fix, what does it do, and what would the impact be? Would it allow for more things like this to be possible?

this sv_magic call would need to become a sv_magicext call with smg->mg_vtbl as vtable argument.

Though in reality it's probably a bit more complicated as this path is only taken by SVp_POK values, and you probably want unset values to be undefined.

[Ovid]

@Leont Yeah, they would need to be undefined. Setting them to any value, even undef, would make them "set."