Convert static types to heap types: use PyType_FromSpec()

[corona10]

BPO	40077
Nosy	@pitrou, @vstinner, @phsilva, @encukou, @skrah, @ericsnowcurrently, @serhiy-storchaka, @corona10, @pablogsal, @miss-islington, @shihai1991, @erlend-aasland, @h-vetinari, @nw0, @kumaraditya303, @CharlieZhao95
PRs	bpo-40077: Convert _jsonmodule to use PyType_FromSpec. #19177 bpo-40077: Convert _abc module to use PyType_FromSpec() #19202 [WIP] bpo-40077: Convert _bz2 extension module to use PyType_FromSpec() #19341 bpo-40077: Fix potential refleaks of _json #19344 bpo-40077: Remove redundant cast in _json extension module #19438 bpo-40077: Convert _bz2 module to use PyType_FromSpec #20960 bpo-40077: Convert _csv module to use PyType_FromSpec  #20974 bpo-40077: Convert _operaator to use PyType_FromSpec #21953 bpo-40077: Convert _operator to use PyType_FromSpec #21954 bpo-40077: Convert mmap to use heap types #23108 bpo-40077: Convert arraymodule to use heap types and establish module state #23124 bpo-40077: Convert _queuemodule to use heap types #23136 bpo-40077: Convert _elementtree types to heap types #23428 bpo-42327: Add PyModule_Add() (smaller). #23443 bpo-40077: Fix typo in simplequeue_get_state_by_type() #23975 bpo-40077: Convert itertools types to heap types #24065 bpo-40077: Add traverse/clear/free to arraymodule #24066 bpo-40077: Convert zoneinfo type to heap type #24481 bpo-40077: Convert _pickle module to use heap types. #30884
Files	bench_isinstance_check.py bench_subclass_check.py bench_isinstance_check.py

^{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 2020-03-26.16:04:44.633>
labels = ['extension-modules', 'type-feature', '3.10']
title = 'Convert static types to heap types: use PyType_FromSpec()'
updated_at = <Date 2022-02-02.05:50:10.693>
user = 'https://github.com/corona10'

bugs.python.org fields:

activity = <Date 2022-02-02.05:50:10.693>
actor = 'kumaraditya'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['Extension Modules']
creation = <Date 2020-03-26.16:04:44.633>
creator = 'corona10'
dependencies = []
files = ['49007', '49008', '49009']
hgrepos = []
issue_num = 40077
keywords = ['patch']
message_count = 47.0
messages = ['365087', '365088', '365117', '365125', '365142', '365145', '365149', '365154', '365205', '365220', '365241', '365242', '365255', '365257', '365317', '365558', '365773', '366024', '366061', '368669', '371120', '371813', '371890', '372073', '372074', '372076', '372326', '375943', '380109', '380267', '380516', '380517', '380518', '383848', '383908', '384228', '384236', '384271', '384272', '384293', '410810', '411630', '411903', '411904', '411907', '411964', '412338']
nosy_count = 16.0
nosy_names = ['pitrou', 'vstinner', 'phsilva', 'petr.viktorin', 'skrah', 'eric.snow', 'serhiy.storchaka', 'corona10', 'pablogsal', 'miss-islington', 'shihai1991', 'erlendaasland', 'h-vetinari', 'nw0', 'kumaraditya', 'CharlieZhao']
pr_nums = ['19177', '19202', '19341', '19344', '19438', '20960', '20974', '21953', '21954', '23108', '23124', '23136', '23428', '23443', '23975', '24065', '24066', '24481', '30884']
priority = 'normal'
resolution = None
stage = 'patch review'
status = 'open'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue40077'
versions = ['Python 3.10']

[corona10]

Some of modules is not using PyType_FromSpec.
We need to convert them.

This changes can bring

• allow to destroy types at exit!
• allow subinterpreters to have their own "isolated" typ

[vstinner]

We need to convert them.

Let me elaborate. Static types have multiple issues:

• Their lifetime is not well defined.
• It is not obvious when they are ready to be used.
• They are not destroyed at exit.
• They are incompatible with subinterpreters: each interpreter should have its own copy of a type, rather than static types are shared by all interpreters which cause problems with reference counting (require GIL or atomic operation).
• They are causing issues with stable ABI (limited C API): PEP-384.

[pablogsal]

Wouldn't having less static types slow down startup time?

[vstinner]

Pablo:

Wouldn't having less static types slow down startup time?

That's possible, even if I only expect a minor or non significant overhead.

But before starting talking about performances, we should focus on the correctness. Static types are causing issues with subinterpreters and the Python finalization.

[vstinner]

New changeset 33f15a1 by Dong-hee Na in branch 'master':
bpo-40077: Convert _json module to use PyType_FromSpec() (GH-19177)
33f15a1

[skrah]

Wouldn't having less static types slow down startup time?

Yes, and not only startup time:

https://bugs.python.org/issue15722

[vstinner]

In the _json module, PyModule_GetState() (called by get_json_state()) is only used by the garbage collector (traverse function) and to unload the module (clear and free functions). It's not used in "hot code" (let me consider that the GC is not part of the usual "hot code" :-)).

But maybe we should measure the overhead of future changes if PyModule_GetState() starts to be used in "hot code" by running a few microbenchmarks.

--

Stefan Krah:

Yes, and not only startup time:
https://bugs.python.org/issue15722

Aha, that's interesting. I didn't know that it could have an impact on runtime performance as well.

_decimal_pep3121-384_v1.patch attached to bpo-15722 seems to use:

#define _decimal_state_global ((_decimal_state *)PyModule_GetState(PyState_FindModule(&_decimal_module)))

whereas the commit 33f15a1 only uses:

static inline _jsonmodulestate*
get_json_state(PyObject *module)
{
    void *state = PyModule_GetState(module);
    assert(state != NULL);
    return (_jsonmodulestate *)state;
}

Maybe PyState_FindModule() adds a little overhead, even if this function is simple: in short, it gets the i-th item of a list (from PyInterpreterState.modules_by_index).

PyModule_GetState() function is really simple: it only reads PyModuleObject.md_state attribute.

Or maybe _decimal_state_global was used in "hot code".

If PyState_FindModule() or PyModule_GetState() is the source of the overhead, maybe we could try to optimise these functions, or pass directly the module state to inner functions.

--

PyState_FindModule() doesn't work for a module using PyModuleDef_Init(). The PEP-573 is going to give access to the module state in functions which didn't access to it previsouly.

The commit 33f15a1 removed a few assertions checking that "self" has the expected type. It wasn't possible to get the module state to get the types, because PEP-573 is not implemented yet and PyState_FindModule() doesn't work in _json (which uses PyModuleDef_Init()). I decided that it's ok to remove these assertions: it should not be possible to call these functions with another type in practice.

--

In his PR 19177, Dong-hee started by replacing direct access to PyTypeObject fields, like replacing:

type->free(self);

with:

freefunc free_func = PyType_GetSlot(tp, Py_tp_free);
free_func(self);

I asked him to revert these changes.

I'm interested to experiment a few C extension modules of the stdlib using the limited C API (PEP-384, stable ABI), but I don't think that it should done while converting modules to PyType_FromSpec().

We should separate the two changes. And I would prefer to first see the overhead of PyType_FromSpec(), and discuss the advantages and drawbacks.

[skrah]

Or maybe _decimal_state_global was used in "hot code".

Yes, _decimal has problems here that most modules don't have.
Modules like atexit are obviously fine. :)

I just posted it as an example why one should be a bit cautious.

The PEP-573 is going to give access to the module state in functions which didn't access to it previously.

That could help, we'll see.

[corona10]

And I would prefer to first see the overhead of PyType_FromSpec(), and discuss the advantages and drawbacks.

Should we stop the work until the overhead is measured?

[vstinner]

Should we stop the work until the overhead is measured?

Can you try to measure the _abc._abc_instancecheck() and _abc._abc_subclasscheck() functions performance before/after your change? Functions like register() are rarely call, so I don't care much of their performance (I expect a small overhead or no overhead).

[corona10]

It shown 1% slower for performance.

+--------------------------+-----------------+----------------------------+
| Benchmark | master-subclass | bpo-40077-subclass |
+==========================+=================+============================+
| bench _abc_subclasscheck | 295 ns | 300 ns: 1.01x slower (+1%) |
+--------------------------+-----------------+----------------------------+

+--------------------------+-------------------+----------------------------+
| Benchmark | master-isinstance | bpo-40077-isinstance |
+==========================+===================+============================+
| bench _abc_instancecheck | 229 ns | 232 ns: 1.01x slower (+1%) |
+--------------------------+-------------------+----------------------------+