Objects with both `__slots__` and `__dict__` have much larger size than needed (up to 4x) on Python 3.13

[ariebovenberg]

Bug report

Bug description:

On Python 3.13, instances that have both __slots__ and __dict__ defined (e.g. through inheritance)
can have a size that is more than four times of that in previous Python versions (e.g. 3.12).

import tracemalloc
import gc


class _Point2D:
    __slots__ = ("x", "y")


class Point3D_OnlySlots(_Point2D):
    __slots__ = ("z",)

    def __init__(self, x, y, z):
        self.x, self.y, self.z = x, y, z


class Point3D_DictAndSlots(_Point2D):
    def __init__(self, x, y, z):
        self.x, self.y, self.z = x, y, z


class Point3D_OnlyDict:
    def __init__(self, x, y, z):
        self.x, self.y, self.z = x, y, z


gc.collect()  # clear freelists
tracemalloc.start()
_ = [Point3D_OnlySlots(1, 2, 3) for _ in range(1_000_000)]
print(
    f"1M Point3D (only __slots__) instances:         {tracemalloc.get_traced_memory()[0]:_} bytes"
)

gc.collect()  # clear freelists
tracemalloc.start()
_ = [Point3D_OnlyDict(1, 2, 3) for _ in range(1_000_000)]
print(
    f"1M Point3D (only __dict__) instances:          {tracemalloc.get_traced_memory()[0]:_} bytes"
)

gc.collect()  # clear freelists
tracemalloc.start()
_ = [Point3D_DictAndSlots(1, 2, 3) for _ in range(1_000_000)]
print(
    f"1M Point3D (__dict__ and __slots__) instances: {tracemalloc.get_traced_memory()[0]:_} bytes"
)

On python 3.13.3, this prints:

1M Point3D (only __slots__) instances:         64_448_792 bytes
1M Point3D (only __dict__) instances:          104_451_928 bytes
1M Point3D (__dict__ and __slots__) instances: 416_448_848 bytes

On python 3.12.11, this prints:

1M Point3D (only __slots__) instances:         64_448_792 bytes
1M Point3D (only __dict__) instances:          96_451_808 bytes
1M Point3D (__dict__ and __slots__) instances: 96_452_232 bytes

The apparent cause seems to be the object layout changes (see here)
which introduced inline values. It appears that these don't mesh well with __slots__.
Could this be because the inline values need to have a fixed offset?
What appears to cause the dramatic increase above is that having (nonempty) __slots__ will
trigger __dict__ materialization (size 30) as soon as a non-slot attribute is set. In contrast to the optimized "no slots" case, this dict doesn't shrink as more instances are created.

Of course, mixing __slots__ and __dict__ is not recommended, but it's a trap that can be easily fallen into.
For example, if forgetting to define __slots__ anywhere in a complex class hierarchy.

I'm unsure if I'm understanding exactly what's going on though. @markshannon perhaps you can shed some light on this?

Related: #115776, #115822

CPython versions tested on:

3.13

Operating systems tested on:

macOS

Linked PRs

• gh-135385: Fix memory regression for classes with both __slots__ and __dict__ #135389
• gh-135385: Do not reserve inline values for attributes stored in slots #156656

[Eclips4]

Hello! Thanks for the report.
Results on current main:

1M Point3D (only __slots__) instances:         64_448_952 bytes
1M Point3D (only __dict__) instances:          104_452_152 bytes
1M Point3D (__dict__ and __slots__) instances: 120_452_152 bytes

[ariebovenberg]

Ah, silly of me not to check the current main as well. Is this something that will be backported into 3.13?

[shuimu5418]

File：Objects/typeobject.c
Func：type_ready_managed_dict()
Lines：9060-9062
Cond：if (type->tp_itemsize == 0)

---

EDIT: Apologies, I posted my analysis before seeing Eclips4's comment above. I have since confirmed locally by compiling and testing the main branch that this issue is indeed fixed there.

It is too difficult for me. But I think i can link these issues:

• Issue: Use static analysis to assist inlining values and compact object layout. #115775

• Issue: Inline values array into the object #115776

[serhiy-storchaka]

The 3.13 regression is fixed, but such instances still reserve inline values for names which are stored in slots.

__static_attributes__ contains every name assigned as self.X in the class body, and _PyDict_NewKeysForClass() inserts all of them into the shared keys, even when X is a slot. A member descriptor is a data descriptor, so such an attribute is never stored in the instance dict, and the reserved space is never used.

Skipping them in _PyDict_NewKeysForClass() (a _PyType_Lookup() per name, at class creation only) gives for the example above:

only __slots__              64_449_113 bytes
only __dict__              104_452_257 bytes
__dict__ and __slots__     104_452_697 bytes   (120_452_257 before)

xml.dom.minidom, where all 11 names in Element.__static_attributes__ are slots, goes from 328 to 232 bytes per node. The whole test suite passes and I see no difference in benchmarks.