PyType_FromSpec API fails to use metaclass of bases

[seberg]

BPO	45383
Nosy	@encukou, @seberg
PRs	bpo-45383: Inherit MetaClass from bases in FromSpec API #28748

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = None
created_at = <Date 2021-10-05.21:51:26.452>
labels = ['interpreter-core', 'expert-C-API', 'type-crash', '3.11']
title = 'PyType_FromSpec API fails to use metaclass of bases'
updated_at = <Date 2021-12-17.17:19:49.187>
user = 'https://github.com/seberg'

bugs.python.org fields:

activity = <Date 2021-12-17.17:19:49.187>
actor = 'seberg'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['Interpreter Core', 'C API']
creation = <Date 2021-10-05.21:51:26.452>
creator = 'seberg'
dependencies = []
files = []
hgrepos = []
issue_num = 45383
keywords = []
message_count = 9.0
messages = ['403273', '408532', '408720', '408723', '408790', '408791', '408793', '408797', '408801']
nosy_count = 2.0
nosy_names = ['petr.viktorin', 'seberg']
pr_nums = ['28748']
priority = 'normal'
resolution = None
stage = None
status = 'open'
superseder = None
type = 'crash'
url = 'https://bugs.python.org/issue45383'
versions = ['Python 3.11']

[seberg]

The PyType_FromSpec fails to take care about MetaClasses.

 https://bugs.python.org/issue15870

Asks to create a new API to pass in the MetaClass. This issue is only about "inheriting" the metaclass of the bases correctly. Currently, Python fails to take into account that the bases may be MetaClass and not PyType_Type instances.

[encukou]

I haven't forgotten this issue, but when I get to it it always leads to a rabbit hole. Sometimes just chasing refleaks, but there are deeper issues as well.

AFAICS, there's no way to call metatype.tp_new for such a class. I guess the safest option is to fail when the metaclass has a custom tp_new -- because that means the metaclass is requesting that it wants to allocate/initialize its types itself.

We can call metatype.tp_init, and probably should. And __init_subclass__ too, I suppose.

But at that point, this is duplicating a lot of existing functionality, and I'm starting to wonder if this wouldn't all be better with calling the metaclass instead.
What's missing?

• basicsize/itemsize could be allowed with __basicsize__/__itemsize__ in the dict.
• flags: we could add mechanisms to set individual flags after the type is created, as needed.
• slots can usually be applied after the class is created; maybe there should be a public function for this.
• members could theoretically be copied to individual descriptors; there doesn't seem much need for keeping tp_members around.

Does that seem like a more reasonable direction to explore?

[seberg]

Sorry, I need some time to dive back into this, so some things might be garbled :). Yes, I do agree supporting a custom tp_new here seems incredibly tricky. I have not thought about the implications of this, though.

guess the safest option is to fail when the metaclass has a custom tp_new

That would seem OK, but I can't quite judge it. It may mean that I have to do a custom factory to create new metaclass instances from Python, but that is probably for the better anyway.

Now, calling tp_new is a bit useless, since from C we don't have a dict anyway (at least not really). So yeah, this does not make sense at all for most Python defined metaclasses... (they may want to inspect/mutate the dict)

But at that point, this is duplicating a lot of existing functionality, and I'm starting to wonder if this wouldn't all be better with calling the metaclass instead.

I do not think I am following :(. My worry is that I want people to create a MetaClass instance through C (but not be locked into static types forever).

My current thought is that I would like it be possible to do something like:

    /* Create a new type object */
    type_spec = {stuff};
    newtype = PyType_FromSpec(&type_spec);
    /* Finalize the MetaClass */
    metatype_spec = {more_stuff};
    PyArray_InitDTypeFromSpec(newtype, &metatype_spec);

Of course I can move this into a single function and create the class for the user. But I am uncertain that piping everything through tp_new will help? At some point I thought that the user should create a subclass, and I create another class inheriting it. But it also seems convoluted and complicated.

I have no idea right now, but maybe there could also be a way to make creating a metaclass-factory in C easier, rather than supporting PyType_FromSpec for metaclasses.
(I.e. an PyType_InitFromSpec() doing most of what PyType_FromSpec does, but really meant to be only used be such metaclass factories.)

• basicsize/itemsize could be allowed with __basicsize__/__itemsize__ in the dict.

Do we need this? I need the basicsize of the metaclass, but that of the class seems fixed/OK?

• flags: we could add mechanisms to set individual flags after the type is created, as needed.

Seems fine, yeah.

• slots can usually be applied after the class is created; maybe there should be a public function for this.
• members could theoretically be copied to individual descriptors; there doesn't seem much need for keeping tp_members around.

But a Python MetaClass (that the author may not even realize about) might need access to these to work properly?
A bit far fetched, but...

[seberg]

It is probably early, but I am starting to like the idea of a "C MetaClass factory" helper/indicator.

It seems to me that yes, at least tp_new cannot be called reasonable for a class that is created in C, it is just too confusing/awkward to try to push the C stuff through the Python API. (And the Python API is the typical one that should not be inconvenienced)

Which gives two possibilities if I want this capability?:

1. Force use of a custom class factory in Python (i.e. not through __new__). But that seems hard, since I need to refuse __new__()!).
2. Use a class factor in C which never calls __new__ and knows that this is OK.

This is not my turf, so I like unholy, but maybe pragmatic things :). Would a slot/flag indicating Py_using_metaclass_cinit_pretty_promise_please "solve" these issues?

I mean, we have two ways to create classes (C and Python), I am not sure it is plausible to untangle this on the MetaClass level, so maybe the only way forward is to embrace it: Some Python MetaClasses just can't be instantiated from C, because we don't know it will work. If we want to allow instantiating the MetaClass from C, we need some way to set this up. And either we cannot use PyType_FromSpec then, or we need to tell it that we know what we are doing.

[encukou]

I don't see how instantiating a metaclass with non-default tp_new would work if you don't know some details about the specific metaclass. So IMO we can we limit ourselves to scenarios where either:

1. the metaclass uses default tp_new, or
2. the code that creates the class knows about the metaclass

For case 2), we could leave allocation to the calling code, and add an API function that does the rest of the work of applying the spec. Something like:

    /* Create a metaclass */
    metatype_spec = {stuff};
    metatype = PyType_FromSpecAndBases(&metatype_spec, &PyType_Type);
    /* Create a type */
    type_spec = {other_stuff};
    newtype = alloc_metatype();
    PyType_ApplySpec(newtype, &type_spec);

PyType_ApplySpec would assert PyType_Ready wasn't called before, and call it after filling in the name, slots, etc.
The metatype could disable its tp_new to disallow PyType_FromSpec (and Python __new__), effectively enforcing "using_metaclass_cinit_pretty_promise_please".

There could be a slot for code to run at the end of PyType_ApplySpec -- the "PyArray_InitDTypeFromSpec" in your pseudocode.

That seems better than calling the metaclass, but to answer questions on that idea:

> - basicsize/itemsize could be allowed with __basicsize__/__itemsize__ in the dict.
Do we need this? I need the basicsize of the metaclass, but that of the class seems fixed/OK?

That's the size of the instance. One more level down :)

> - members could theoretically be copied to individual descriptors; there doesn't seem much need for keeping tp_members around.
But a Python MetaClass (that the author may not even realize about) might need access to these to work properly?
A bit far fetched, but...

Seems very far-fetched to me. IMO these, like e.g. tp_methods, should only be used as arguments for the code that puts the descriptors in __dict__. How the descriptors got there is a detail, the class doesn't need to use tp_members (nor advertise that it does).

[seberg]

Fully, agree! In the end, PyType_FromSpec replaces type.__new__() (and init I guess) when working in C. In Python, we would call type.__new__ (maybe via super) from the metatype.__new__, but right now, in C, the metatype cannot reliably use PyType_FromSpec in its own metatype.__new__ to do the same.

I agree with the scenarios:

• If we do not have a custom metatype.__new__ (init?) then PyType_FromSpec should have no reason to refuse doing the work, because nothing can go wrong.
• If we do have a custom tp_new the user has to provide C API to create the metaclass instance. But they still need a way to call type.__new__ in C (i.e. get what PyType_FromSpec does, and promising to do the rest).

PyType_ApplySpec would provide that way to create a custom metatype.__new__ in C when PyType_FromSpec() would otherwise reject it to make the first scenario safe.
A flag probably can do the same. I have no preference, ApplySpec seems great to me.

[encukou]

Nice! It's starting to look reasonable, I'll try an implementation when I get some focus time. (Sadly I can't promise it'll be this year.)

Just one detail:

A flag probably can do the same. I have no preference, ApplySpec seems great to me.

I didn't get what you mean here.