From migo@freeshell.org

The following test program demonstrates a regression in perl 5.18​:

  @​Class2​::ISA = qw(Class1);
  sub Class1​::AUTOLOAD {
  *{$Class1​::AUTOLOAD} = sub { print "defined\n" };
  print "autoloaded\n";
  }

  Class2->Class2​::SUPER​::a;
  Class2->Class2​::SUPER​::a;
  Class2->a;
  Class2->a;

In perl versions prior to 5.18.0 it prints​:

  autoloaded
  defined
  autoloaded
  defined

In 5.18.0 (and 5.18.2) it prints "autoloaded" 3 times in a row.
And Class2->can('Class2​::SUPER​::a') returns undef instead of CODE,
although the method gets re-defined every time, causing warnings
"Subroutine SubClass​::SUPER​::method redefined" in my real code.

Best regards,

Mikhael

From @tonycoz

On Tue Apr 29 15​:13​:01 2014, migo@​freeshell.org wrote​:

The following test program demonstrates a regression in perl 5.18​:

@​Class2​::ISA = qw\(Class1\);
sub Class1​::AUTOLOAD \{
    \*\{$Class1​::AUTOLOAD\} = sub \{ print "defined\\n" \};
    print "autoloaded\\n";
\}

Class2\->Class2​::SUPER​::a;
Class2\->Class2​::SUPER​::a;
Class2\->a;
Class2\->a;

In perl versions prior to 5.18.0 it prints​:

autoloaded
defined
autoloaded
defined

In 5.18.0 (and 5.18.2) it prints "autoloaded" 3 times in a row.
And Class2->can('Class2​::SUPER​::a') returns undef instead of CODE,
although the method gets re-defined every time, causing warnings
"Subroutine SubClass​::SUPER​::method redefined" in my real code.

Bisected to​:

aae4380 is the first bad commit
commit aae4380
Author​: Father Chrysostomos <sprout@​cpan.org>
Date​: Mon Sep 17 16​:24​:40 2012 -0700

  [perl #114924] Make method calls work with :​:SUPER packages
 
  Perl caches SUPER methods inside packages named Foo​::SUPER. But this
  interferes with actual method calls on those packages (SUPER->foo,
  foo​::SUPER->foo).
 
  The first time a package is looked up, it is vivified under the name
  with which it is looked up. So *SUPER​:: will cause that package
  to be called SUPER, and *main​::SUPER​:: will cause it to be named
  main​::SUPER.
 
  main->SUPER​::isa used to be very sensitive to the name of the
  main​::FOO package (where the cache is kept). If it happened to be
  called SUPER, that call would fail.
  ...
 
Tony

The RT System itself - Status changed from 'new' to 'open'

From @rjbs

This seems like a 5.22.0 blocker, and something we'd want to backport when fixed.

Agreed?

From @tonycoz

On Sun Jul 13 11​:27​:34 2014, rjbs wrote​:

This seems like a 5.22.0 blocker, and something we'd want to backport
when fixed.

I'm not sure it's fixable without reverting the patch that changed the behaviour.

Both before and after the patch, $Class2​::AUTOLOAD is set to "Class2​::SUPER​::a" for the SUPER call​:

  # system perl is 5.14.2
  tony@​mars​:.../git/perl2$ perl ../super-autoload.pl
  autoloaded Class2​::SUPER​::a
  defined
  autoloaded Class2​::a
  defined

  tony@​mars​:.../git/perl2$ ~/perl/5.20.0-dbg/bin/perl ../super-autoload.pl
  autoloaded Class2​::SUPER​::a
  autoloaded Class2​::SUPER​::a
  autoloaded Class2​::a
  defined

Before the patch, perl used *{"${package}​::SUPER"} as a cache for SUPER lookups, so the assignment to *{$Class1​::AUTOLOAD} updated the cache directly. The next call to SUPER​::a finds the entry in the cache and everything (sort of) works.

After the patch, perl no longer looks in *{"${package}​::SUPER"}, so the second call doesn't find the entry the first call added.

One solution might be to not paste "SUPER​::" into $AUTOLOAD​:

--- a/gv.c
+++ b/gv.c
@​@​ -1116,7 +1116,6 @​@​ Perl_gv_autoload_pvn(pTHX_ HV *stash, const char *name, ST
  }
  else
  packname = sv_2mortal(newSVhek(HvNAME_HEK(stash)));
- if (flags & GV_SUPER) sv_catpvs(packname, "​::SUPER");
  }
  if (!(gv = gv_fetchmeth_pvn(stash, S_autoload, S_autolen, FALSE,
  is_utf8 | (flags & GV_SUPER))))

but this doesn't restore the old behaviour​:

  tony@​mars​:.../git/perl2$ ./perl ../super-autoload.pl
  autoloaded Class2​::a
  autoloaded Class2​::a
  defined
  defined

It removes the ability to distinguish between SUPER and non-SUPER calls (if that's useful at all.)

Tony

From @rjbs

* Tony Cook via RT <perlbug-followup@​perl.org> [2014-07-14T21​:33​:11]

Both before and after the patch, $Class2​::AUTOLOAD is set to
"Class2​::SUPER​::a" for the SUPER call​:
[...]

Thanks, that's all useful information that I will have to digest.

What a mess. :( SUPER!

Anybody else want to throw opinions in here?

--
rjbs

From @ikegami

On Mon, Jul 14, 2014 at 9​:33 PM, Tony Cook via RT <perlbug-followup@​perl.org

wrote​:

One solution might be to not paste "SUPER​::" into $AUTOLOAD​:

[...]

but this doesn't restore the old behaviour​:

tony@​mars​:.../git/perl2$ ./perl ../super-autoload.pl
autoloaded Class2​::a
autoloaded Class2​::a
defined
defined

You're calling Class1​::a, not calling Class2​::a. That seems worse.

From @rjbs

I don't have any solution to offer to restore the old behavior without reintroducing old problems.

What would we want out of a solution to this moving forward? If we want to install the method that was autoloaded, there are two likely candidates for where to install it​:

1) the package where the called AUTOLOAD was called
2) the package where we first looked for the method

The first can be done, generally, by having AUTOLOAD use __PACKAGE___.

The second is more of a problem, here, because of the inherent ambiguity between SUPER and SUPER. We could provide (say) *AUTOLOAD to which a subroutine to be installed at the original call site could be assigned. We could provide a means to get the resolved package we're trying to look in (ParentClass instead of ChildClass​::SUPER for when AUTOLOAD is in GrandParentClass). It is unfortunate that this changed without our noticing or issuing a notice, but I'm not eager to vacillate between two buggy states if we can provide better.

--
rjbs