Export models with __slots__ on the generated Space classes

[fumitoh]

Generated Space classes now declare __slots__. CPython keeps slotted attributes in a fixed-size array instead of an instance dictionary and specialises the bytecode that reads them, so an exported model runs faster and takes less memory. A new use_slots keyword on export_model(), Model.export() and Exporter() turns the declarations off.

The motivation is a CPython 3.14 change (5d3201fe / gh-123219) that stopped specialising LOAD_ATTR for instance attributes outside a type's 30-entry shared-key table. _c_Projection of lifelib's BasicTerm_SC has 73 attributes, so all 39 _v_* memo caches fell off the fast path. __slots__ is not just a 3.14 workaround, though — it is a solid win on 3.13 too.

Measurements

lifelib basiclife/BasicTerm_SC, 3,000 model points per round, best of three, identical checksums both ways:

export style	Python 3.13.9	Python 3.14.6	ItemSpace
as generated today	3.68 s	3.22 s	1,632 bytes
__slots__	2.90 s	2.31 s	616 bytes

What is in the diff

• use_slots=True on export_model(), Model.export() and Exporter.__init__, threaded to SpaceTranslator.
• __slots__ = () on BaseMxObject / BaseParent / BaseSpace in the _mx_sys.py template, in both modes. A subclass that declares no __slots__ of its own still gets a __dict__, so use_slots=False is unaffected. BaseModel and the generated model class are left alone deliberately — the model class carries arbitrary Reference names and there is one instance.
• A {slots_decl} placeholder in SpaceTranslator.class_template, expanding to the empty string under use_slots=False.
• No __weakref__ slot, by decision.

A name missing from __slots__ is an AttributeError at run time, not a compile-time error, so each slot name is collected next to the statement that assigns the attribute. ref_names() and space_names() are now the single source of truth for the References and the child Spaces, shared by the assignment emitters and the slot list.

The propagation rule

__call__ assigns a Space's own parameters onto every Space in its subtree through _mx_assign_params, and replays every enclosing ItemSpace root's _mx_copy_params as well. A Space class therefore declares the parameters of every parameterized ancestor, not only its own — _c_SubChild of the NestedParams sample needs slots for Parent's x and Child's y although it has no parameters of its own.

Names __slots__ cannot express

A Space parameter is the only name that reaches __slots__ with a leading underscore — modelx rejects one in the name of a Reference or a Space, and a Cells name only ever arrives behind a _v_ or _has_ prefix. Three cases needed handling:

• A name the class body binds (a Cells sharing its name with a parameter, or with a Reference including a model-level global) makes type.__new__ raise, so the whole _mx_classes.py fails to import. The exporter now raises ValueError at export time instead, naming the Space and the name and pointing at use_slots=False.
• A name inherited from the _mx_sys template (_cells, _mx_walk, …) gets no error from CPython at all: the slot silently shadows the member. A parameter named _cells would take over the property that populates _mx_cells. The reserved set is read from _mx_sys.BaseSpace.__mro__ rather than listed, so it cannot fall behind the template.
• A private name (__x) is mangled with the class body it is written from, and a __slots__ string is mangled with the class that declares it, so a descendant must declare _c_<Owner>__x. Without that, creating any ItemSpace of such a Space raised AttributeError although it works today.
• A non-NFKC name. Python normalises every identifier in a source file; a __slots__ string is only mangled, never normalised. str.isidentifier() accepts a fullwidth letter and modelx keeps the name verbatim, so a Reference named Ｒate was assigned as self.Rate and declared as 'Ｒate', and the exported package did not import at all. Slot names, the guarded names and the mangling class name are all normalised now.

Verification

• use_slots=False is byte-identical to v0.32.0 across all 44 generated modules of 10 sample models. Only _mx_sys.py differs, in both modes, because it is copied verbatim and now carries __slots__ = () on the base classes.
• __slots__ covers every attribute a use_slots=False export actually assigns at run time, checked over 12 sample models and 6 lifelib models including TradLife_A (14 classes) and IntegratedLife (9), with the models exercised until ItemSpaces are created three levels deep.
• Full suite green on Python 3.13 (1447 passed, 6 skipped) and the export suite on 3.12. Docs build adds no warnings.
• New modelx/tests/export/test_slots.py: the slots-cover-__dict__ equivalence test over every sample plus a model with nested Spaces under a parameterized Space, multiple parameters, cacheless Cells, model-level globals, macros and every Reference kind ref_value() handles; a static per-class assignment scan; byte-identity of the two modes; no __dict__ under slots; and one test per rejected name.

For the reviewer

Two consequences are documented rather than guarded, both loud and both with an obvious workaround:

• The attributes of a Space class are fixed at export time, so a macro or formula that assigns a Reference the model does not already have raises AttributeError in the exported model. The name is not knowable when the class is generated. The task memo's "macros are not affected" is true about where macros live, not about what they do.
• Exported Space objects are not weak-referenceable, have no __dict__, and cannot be pickled with pickle protocol 0 or 1.

Two things worth an explicit decision:

1. The new ValueError refuses one model that exports correctly today. m.alias = m.S.cells — a model-level Reference aliasing a Cells — generates self.rate = self.rate, which resolves to the same bound method and is a no-op, so v0.32.0 exports it correctly. It cannot be slotted at all while the class body defines the method, so the export is refused; the message no longer claims the export was already wrong. Falling back to an unslotted class for that one Space would be the alternative.
2. modelx-cython breaks on use_slots=True output — tracer.py reads the traced Space's __dict__ to discover its refs. That single line is the only blocker; compiled models were never affected by the CPython regression in the first place, since @cy.cclass turns attributes into C struct fields. Fixing it is out of scope here and is recorded in devnotes/DependentPackages.md §2.2, but with use_slots=True as the default the two releases need coordinating.

🤖 Generated with Claude Code