PHP RFC: Final anonymous classes

• Date: 2023-11-15
• Author: Daniil Gentili daniil@daniil.it
• Status: Declined

This RFC proposes to add support for final anonymous classes.

This RFC proposes to add support for final anonymous classes.

This should also allow some additional opcache optimizations, such as any JIT logic gated behind a check on ZEND_ACC_FINAL, i.e. https://github.com/php/php-src/blob/master/ext/opcache/jit/zend_jit_trace.c#L4507.

Example syntax:

$x = new final class {};

Extending a final anonymous class throws an error:

$x = new final class {};
class_alias($x::class, 'alias');
class aliasExtends extends alias {}

Fatal error: Class aliasExtends cannot extend final class class@anonymous in %s on line %d

None.

Next PHP 8.4

See Backward Incompatible Changes.

2/3 required to accept. Voting started on 2023-12-03 and will end on 2023-12-18 00:00 GMT.

Add support for final anonymous classes?
Real name	Yes	No
alcaeus (alcaeus)
alec (alec)
ashnazg (ashnazg)
brzuchal (brzuchal)
crell (crell)
cschneid (cschneid)
derick (derick)
devnexen (devnexen)
dharman (dharman)
galvao (galvao)
geekcom (geekcom)
heiglandreas (heiglandreas)
kalle (kalle)
kguest (kguest)
levim (levim)
mbeccati (mbeccati)
nicolasgrekas (nicolasgrekas)
nielsdos (nielsdos)
nikic (nikic)
ocramius (ocramius)
ramsey (ramsey)
sebastian (sebastian)
stas (stas)
weierophinney (weierophinney)
Final result:	8	16
This poll has been closed.

Final anonymous classes implementation: https://github.com/php/php-src/pull/11126

• Pull request discussion: https://github.com/php/php-src/pull/11126
• Internals discussion: https://externals.io/message/121356, https://externals.io/message/121685

After feedback received from Nikolas Grekas in the last [RFC] discussion thread (https://externals.io/message/121685), I moved here a large chunk of the rationale and removed basically all the alternative polls I had initially planned to propose:

Hi Daniil,

  >> While I'm open to Proposal 1, which introduces final anonymous classes
  >> without breaking BC, Proposals 2 and 3 are a different story.
  >> In summary, I advocate for the RFC to focus on the non-BC-breaking option.
  >> Let's maintain our commitment to stability and gradual evolution in PHP.
  >> Cheers,
  >> Nicolas

  > Agree with your points, just adding final anonymous classes seems the best solution to me, but given the interest in alternative solutions both in the pull request discussion, and in the previous mailing list thread, I think I'll leave the other options in the RFC, to see how the votes will go (I'm actually curious myself :).
  > Regards,
  > Daniil Gentili

I think this is a dangerous game. Breaking BC shouldn't be proposed unless absolutely needed IMHO.

Nicolas

To be entirely honest, I'm a bit on the fence: on one hand, looking at code like new class {}, you would assume that since the class apparently has no name, it should not be impossible to extend it, but on the other hand, there are valid usecases for extending even anonymous classes (comments in the PR (https://github.com/php/php-src/pull/11126) referenced proxying, I can think of phpunit mocking to a much, much lesser extent given that you should realistically (hopefully) never have to mock an anonymous class that does not already implement an interface), and completely precluding the possibility of extending a class that

1. Has a name (even if it's not immediately obvious)
2. Can be referenced to using its name (class_exists, new ReflectionClass, new $clazz, and yes, even class_alias)

seems a tad bit too restrictive...

On the other hand, I also really don't like non-final classes, in all of my projects, I use CS rules that force all classes to either be abstract or final, because unfortunately, I've had to work with a lot of code that very frequently violates encapsulation by using inheritance.

Still, there are some useful patterns, mainly regarding testing and mocking, for example I use https://github.com/dg/bypass-finals as a dev dependency to make all final classes non-final at runtime to allow mocking in phpunit, but it works by installing a custom default stream contexts that intercepts requires, tokenizes files and removes final keywords from classes; this approach would break for anonymous classes if they were rendered final by default without an option to make them non-final.

This is why I remain ambivalent about the options, as seen both in my emails and in the original text of the RFC:

Personally, I would have instead preferred the much cleaner approach of making all anonymous classes final by default, (preferrably) without offering the option to make them non-final.

However, I understand that this might be a little bit too restrictive for something that may have some valid usecases, even if extending anonymous classes currently requires some hack-ish workarounds with class_alias.

Thus, this RFC initially included three mutually exclusive proposals:

1. Add support for final anonymous classes (new final class {} syntax, no breaking changes)
2. OR Make all anonymous classes final by default, without the option to make them final (breaking change)
3. OR Make all anonymous classes final by default, provide an optional open keyword to make them non-final (like in Kotlin, new open class {}, breaking changes).

As an extra proposal related to the last two options, possibly requiring a separate RFC, in the last two cases, it might be a good idea to also disallow the use of class_alias altogether for final anonymous classes (suggested by nikolas-grekas in https://github.com/php/php-src/pull/11126#issuecomment-1522042841).