zitcompiler

JIT-like native extension compilation for Python using the Zig compiler. Write a Zig module, compile it at runtime, and load the result as a callable Python object — with optional comptime parameter injection so the same Zig source can be specialized differently on each call.

Requires the ziglang package (pip install ziglang).

Usage

Basic: compile and load a function

Write a Zig module that exports a standard CPython PyCFunction:

// my_ext.zig
extern fn Py_IncRef(obj: ?*PyObject) void;
extern var _Py_NoneStruct: PyObject;
const PyObject = extern struct { ob_refcnt: i64, ob_type: ?*anyopaque };

export fn greet(self: ?*PyObject, args: ?*PyObject) callconv(.c) ?*PyObject {
    _ = self; _ = args;
    @import("std").debug.print("Hello from Zig!\n", .{});
    Py_IncRef(&_Py_NoneStruct);
    return &_Py_NoneStruct;
}

Load it from Python with the high-level helper:

from pathlib import Path
from zitcompiler import zitcompiled

greet = zitcompiled(Path("my_ext.zig"), "greet")
greet()  # Hello from Zig!

Comptime parameters: specialize Zig code from Python

ZigModuleDef lets you inject Python dataclass values as Zig comptime constants. The Zig module imports them as a named module and the compiler eliminates all dead branches at build time.

from dataclasses import dataclass
from pathlib import Path
from zitcompiler import ZigModuleDef, zitcompiled

@dataclass
class Params:
    multiplier: int = 7

module_def = ZigModuleDef(top_level=Params, structs=[], module_name="params")
get_multiplier = zitcompiled(Path("my_ext.zig"), "get_multiplier", module_def=module_def)

// my_ext.zig
const params = @import("params");  // injected at compile time
const PyObject = extern struct { ob_refcnt: i64, ob_type: ?*anyopaque };
extern fn PyLong_FromLong(v: c_long) ?*PyObject;

export fn get_multiplier(self: ?*PyObject, args: ?*PyObject) callconv(.c) ?*PyObject {
    _ = self; _ = args;
    return PyLong_FromLong(params.multiplier);  // 7, resolved at compile time
}

Structs defined in the dataclass are emitted as Zig struct types in the injected module, enabling comptime reflection (e.g. @typeInfo(params.Point).@"struct".fields.len).

Low-level API

For finer control, use zig_build_lib directly and then load_function / load_class:

import asyncio
from pathlib import Path
from zitcompiler import BuildLibOptions, load_function, zig_build_lib

opts = BuildLibOptions(
    module_path=Path("my_ext.zig"),
    link_python=True,
    output_path=Path("/tmp/my_ext.so"),
)
so_path = asyncio.run(zig_build_lib(opts))
greet = load_function(so_path, "greet")
greet()

load_class works the same way for exported PyTypeObject symbols.

Zetaclasses

zetaclass is a drop-in replacement for @dataclass that compiles the class to a native Zig struct at decoration time. The resulting type behaves like a regular Python class but __init__ and __eq__ run as compiled C slots — no Python interpreter overhead.

Usage

from zitcompiler.zetaclasses import zetaclass

@zetaclass
class Point:
    x: int
    y: int

p1 = Point(1, 2)
p2 = Point(x=1, y=2)
assert p1 == p2

Supported field types: int (i64), float (f64), str (Python string object).

Default values work the same as with @dataclass:

@zetaclass
class Config:
    host: str = "localhost"
    port: int = 8080
    timeout: float = 30.0

cfg = Config()            # all defaults
cfg2 = Config(port=9090)  # keyword override
assert cfg != cfg2

Positional, keyword, and mixed argument styles are all supported:

Config("example.com", 443)       # positional
Config(host="example.com")       # keyword only
Config("example.com", timeout=5) # mixed

Attribute access works normally — fields are readable and writable:

print(cfg.host)   # "localhost"
cfg.port = 9090

Current limitations

• Supported field types: int, float, str. Other types raise TypeError at decoration time.
• __repr__, __hash__, ordering operators, and frozen are not yet implemented.
• Compilation runs synchronously at decoration time (same as other zitcompiler calls).

Internal implementation

When @zetaclass is applied, the decorator:

1. Reads field names, types, and default values from the class annotations (following the MRO).
2. Generates a params.zig source file containing an extern struct with ob_base: PyObject as its first field — the layout CPython requires for all heap-allocated objects — followed by one field per annotation.
3. Generates a Defaults struct holding comptime constants for each field that has a default value. String defaults are stored as [:0]const u8 (null-terminated slice); numeric defaults as i64/f64.
4. Compiles params.zig against core.zig (the static Zig library shipped with the package) using zig build-lib.
5. Loads the exported PyTypeObject symbol via load_class, which calls PyType_Ready to finalise the type.

core.zig provides the comptime slot generators:

Slot	Generator	What it does
tp_init	initFn(T, Defaults)	Iterates struct fields at comptime; reads positional args, then kwargs, then comptime defaults. Stores values with ref-counting for ?*PyObject fields.
tp_richcompare	richCompareFn(T)	Field-by-field equality via != for numeric fields and PyObject_RichCompareBool for string fields. Returns Py_NotImplemented for non-EQ/NE ops or mismatched types.
tp_members	membersArray(T)	Comptime-generates a null-terminated PyMemberDef[] using @offsetOf for each field. Python's built-in member descriptor machinery handles all get/set at runtime — no Zig code runs on attribute access.
tp_dealloc	deallocFn(T)	Decrefs all ?*PyObject fields then calls PyObject_Free. Only wired in when the struct contains object fields; otherwise null and CPython inherits the default from object.

The Defaults struct approach avoids any Python-level wrapper class: defaults are resolved entirely in the compiled tp_init slot, so the loaded PyTypeObject is the final Python type with no subclassing or runtime indirection.

JSON schema → native class

zitcompiler.jsonschemas can compile a native class directly from a JSON schema definition, without writing any Python class by hand.

Quick start

import json
from pathlib import Path
from zitcompiler.jsonschemas import zetaify

schema = json.loads(Path("person.json").read_text())
Person = zetaify(schema)

p = Person(name="Alice", age=30)
print(p)  # Person(name='Alice', age=30)

zetaify accepts a raw JSON-decoded dict or a pre-parsed SchemaDef. It returns a fully compiled @zetaclass type.

Saving a type stub

Pass stub_path to write a .pyi file alongside your schema. The stub includes the @zetaclass decorator so it reads as a valid, annotated class definition:

Person = zetaify(schema, stub_path=Path("person.pyi"))

Generated person.pyi:

from zitcompiler.zetaclasses import zetaclass

@zetaclass
class Person:
    name: str
    age: int
    height: float | None
    active: bool | None

Forwarding zetaclass options

Any keyword accepted by @zetaclass can be forwarded through zetaify:

Person = zetaify(schema, frozen=True, kw_only=True)

Lower-level API

Use parse_schema and generate_stub directly if you need the intermediate representations:

from zitcompiler.jsonschemas import generate_stub, parse_schema

schema_def = parse_schema(raw)          # → SchemaDef
print(generate_stub(schema_def))        # plain class body
print(generate_stub(schema_def, with_decorator=True))  # with @zetaclass header

Current limitations

• Only flat (non-nested) schemas are supported. Nested object properties are not yet handled.
• Optional fields (T | None, i.e. properties absent from required) are present in the written stub but excluded from the compiled native struct, since the Zig backend has no nullable scalar type mapping yet. Only required fields with types string, integer, or number become native struct fields.
• boolean and null JSON schema types appear in stubs as bool / None but are not currently supported as zetaclass field types; including them as required fields will raise a TypeError at compilation time.

Tests

Unit tests

uv run pytest tests/

Covers Zig module compilation (zig_build_lib), comptime parameter injection, and load_function / load_class.

Build backend integration tests

nox -f tests/build_backend/noxfile.py

Tests the hatch build hook end-to-end. The session:

1. Installs hatchling and a local editable copy of zitcompiler into a fresh virtual environment.
2. Builds the greetings test package (tests/build_backend/test_pkg/) with uv build --no-build-isolation. During the wheel build, the hatch hook imports greetings, which triggers a zitcompiled() call, and the resulting .so is bundled into the wheel.
3. Installs the built wheel and runs pytest tests/build_backend/test_aot.py, which verifies that the pre-compiled .so is present inside the installed package and that all exported symbols (function and zetaclasses) work correctly.

The test package (greetings) contains one zitcompiled() Zig function (hello_world) and three @zetaclass types (Greeter, Point, Color). The hello_world function is the AoT target; zetaclasses compile at import time as usual.

Known limitations

Incremental compilation on Linux (ELF targets)

BuildLibOptions.incremental = True passes -fincremental to the Zig compiler, which activates the elf2 linker backend. As of Zig 0.16, elf2 does not implement saving linker state to disk, so incremental build-lib always fails on ELF targets with:

error(compilation): TODO implement saving linker state for elf2

Why this exists: true incremental linking requires persisting the linker's internal data structures (symbol tables, section allocations, relocation records, virtual address assignments) between builds so subsequent builds can restore and patch only what changed. The elf2 linker tracks dirty sections in memory via a ZigObject structure but cannot yet serialize that state to disk.

This is independent of what you link against — the error occurs even for a minimal Zig module with no external dependencies.

Roadmap: tracked in ziglang/zig#21165. Incremental compilation works today for pure Zig executables and (as of April 2026) the LLVM backend; build-lib on ELF is the remaining gap. Once Zig lands linker state serialization, incremental = True will work transparently.

zitcompiler emits a logging.WARNING when incremental = True is requested on a non-Windows, non-macOS platform.

Examples

hello_world

zig build-lib examples/hello_world.zig

hello_world_ext (Python C extension, module-level function)

zig build-lib -dynamic -lc examples/hello_world_ext.zig -femit-bin=hello_world.so $(python3-config --ldflags --embed)

import sys; sys.path.insert(0, ".")
import hello_world
hello_world.hello_world()  # prints: Hello from zig!

greeter_ext (Python C extension, class with method)

zig build-lib -dynamic -lc examples/greeter_ext.zig -femit-bin=greeter.so $(python3-config --ldflags --embed)

import sys; sys.path.insert(0, ".")
import greeter
g = greeter.Greeter()
g.hello_world()  # prints: Hello from zig!