Newsgroups: php.internals
Path: news.php.net
Xref: news.php.net php.internals:84766
Mailing-List: contact internals-help@lists.php.net; run by ezmlm
Received-SPF: pass (pb1.pair.com: domain gmail.com designates 209.85.212.169 as permitted sender)
MIME-Version: 1.0
In-Reply-To: <CABDNdAHd2WzYMGLTb3byhP6_WWwbzw1CVFbVs+5yu-UJ0KVkqg@mail.gmail.com>
References: <CABDNdAHd2WzYMGLTb3byhP6_WWwbzw1CVFbVs+5yu-UJ0KVkqg@mail.gmail.com>
Date: Sat, 14 Mar 2015 11:33:30 +0100
Message-ID: <CAF+90c8bXm_+_kCwX-NAon6o0xfJjD2crDaUfdxs0=mpk0-fbg@mail.gmail.com>
To: Philip Sturgeon <pjsturgeon@gmail.com>
Cc: "internals@lists.php.net" <internals@lists.php.net>
Content-Type: multipart/alternative; boundary=089e010d89f2ecf72a05113d2571
Subject: Re: [PHP-DEV] [VOTE] [RFC] Anonymous Classes
From: nikita.ppv@gmail.com (Nikita Popov)

--089e010d89f2ecf72a05113d2571
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: quoted-printable

On Fri, Mar 13, 2015 at 8:33 PM, Philip Sturgeon <pjsturgeon@gmail.com>
wrote:

> A two week discussion period has been held and there are no outstanding
> issues.
>
> Serialization has been disabled, and generated names have been
> explained better in the newest version of the RFC
>
> https://wiki.php.net/rfc/anonymous_classes
>
> The implementation needs to be updated with changes from master, but
> that can be done at a later point and should not be used as a reason
> to vote against.
>

I'm not yet sure which way to vote on this RFC. I don't think there's
anything principally wrong with anonymous classes, but the current RFC
seems incomplete to me with regard to scoping (and as future changes in
this area are not necessarily backwards-compatible, I'd rather solve this
now than later). Lets start off with an example from the RFC:

class Outside {
    protected $data;

    public function __construct($data) {
        $this->data =3D $data;
    }

    public function getArrayAccess() {
        return new class($this->data) extends Outside implements ArrayAcces=
s {
            public function offsetGet($offset) { return $this->data[$offset=
]; }
            public function offsetSet($offset, $data) { return
($this->data[$offset] =3D $data); }
            public function offsetUnset($offset) { unset
<http://www.php.net/unset>($this->data[$offset]); }
            public function offsetExists($offset) { return isset
<http://www.php.net/isset>($this->data[$offset]); }
        };
    }}

So ... WTF. It probably took me 10 minutes yesterday to finally understand
the evil, evil things this code is doing and why it is doing them. This is
what happens: The new anon class extends Outside (the wrapping class) and
as such also inherits the parent constructor. The new class($this->data)
bit passes $this->data to the constructor (the one inherited from Outside),
as such $this->data will be assigned in the inner class as well. Because
the anonymous class extended from Outside it is allowed to access the
protected $data member.

Why is this (imho very weird and unintuitive) approach used? Quoting from
the RFC: "[E]xtending Outer [sic] allows the nested class implementing
ArrayAccess permission to execute protected methods, declared in the Outer
[sic] class, on the same $this=E2=86=92data".

So the reason behind this is that anonymous classes as implemented by this
RFC - and much unlike the anonymous class implementations you will find in
other languages like Java or D - are considered to be totally unrelated to
the wrapping class and have only "public" access to its scope. I don't
think this is good. The solution ("extends Outside") to work around this
that is presented in the RFC has a number of problems:

 * As PHP does not support multiple inheritance, extending Outside for
scope access means no other class may be extended.
 * Extending the wrapping doesn't just give you access to restricted
methods/properties of the wrapping class - it will also import everything
into the inner class. This means that the anonymous class will have a bunch
of additional methods and properties (potentially public ones) which have
nothing to do with whatever the class is actually for.
 * The approach used to pass the data into the inner class by using the
constructor of the wrapping class assumes that the constructor a) only
performs simple assignments and b) accepts the entire state of the wrapping
class as parameters - if this is not the case you will likely not be able
to duplicate the state of the wrapping class in the inner anonymous class.

To solve this without such hacks, two things are necessary:

a) Assuming that the anonymous class somehow got hold of an instance of the
wrapping class (i.e. it has an object $obj instanceof Outside), it should
be able to access private and protected properties and methods of that
object. This is consistent with the general rule that any code within a
class body can access it's privates, unless explicitly rebound. For
example, the following code should work:

    class Outside {
        private $private;
        public function getAnon() {
            return new class($this) {
                private $outside;
                public function __construct($outside) { $this->outside =3D
$outside; }
                public function getPrivateOfOutside() { return
$this->outside->private; }
            };
        }
    }

b) Providing some easy way to access the instance of the wrapping class.
The previous example passed it in as a constructor, which is somewhat
verbose. A possible syntax for accessing the wrapping instance would be
Outside::$this (to borrow Java syntax):

    class Outside {
        private $private;
        public function getAnon() {
            return new class {
                public function getPrivateOfOutside() { return
Outside::$this->private; }
            };
        }
    }

Point a) is the more fundamental issue, which I think should be addressed
from the start. It's also not backwards compatible to introduce it as a
later point. Point b) is pretty much just syntax sugar once a) stands. (The
example of a) can be further simplified by using a trait that automatically
implements that constructor and property, to avoid the boilerplate.)

Nikita

--089e010d89f2ecf72a05113d2571--