Architecture¶

For contributors and anyone who wants to understand how the runtime is structured. If you just want to run Python in a browser tab, the Getting Started page is shorter. If you want to embed the native binary in a Linux/macOS scripting context, vs MicroPython covers the compatibility surface.

Overview¶

nanopython is a Python interpreter assembled from two cooperating Zig executables sharing one source tree:

• The compiler (src/compiler/): pure Zig. Lexes Python source, parses to a grammar tree, runs scope analysis, emits MicroPython .mpy bytecode. The parser is a Zig port of MicroPython's parse.c; the grammar it consumes is upstream's grammar.h patched at make port time (see port/grammar-patched/) to accept PEP 448 dict/list unpack in more positions and PEP 570 positional-only / markers. Ships as the standalone zigpy-compile binary AND links into the main interpreter.
• The VM (src/runtime/, src/objects/, src/modules/, src/memory/): Zig with one residual 86-line C tombstone. Loads bytecode, dispatches the bytecode interpreter, manages the object model, handles allocator + GC.

The main nanopython binary contains both halves; the runtime calls into the compiler at exec time to compile and run a string. The split matters because the compiler is fully testable in isolation (~93 unit tests across lexer.zig, parser.zig, scope.zig, emit.zig, compile.zig, qstr.zig, bigint.zig).

The full design story is in TR-001 (notes/tech-reports/TR-001.md), the methodology paper. This page covers the central patterns; the ADR series records the decisions.

The methodology in one paragraph¶

The codebase is the result of an incremental C→Zig strangler of MicroPython under a byte-exact differential test corpus. For each MicroPython subsystem, a Zig module was written matching the vendor C file's link surface (function signatures and struct layouts), so the C file could be dropped from the compile graph and the Zig module linked in its place. This is an internal link contract, used during the port and still relied on for the full config's residual vendor-C carveouts; it is not a public ABI that downstream code should link against. After every change, all 541 corpus tests ran against a frozen oracle (corpus/nano/basics/*.out); any byte drift failed the build. The strangler converged at the long-standing 541/541 invariant. Active work is now on user experience, feature reach, and size; see ADR 010.

The two configurations¶

port/mpconfigport.h has two arms selected via -Dconfig=...:

Why two configs¶

nano is the forcing function. It's deliberately minimal so the surface that has to stay byte-identical is small. The strangler froze nano-as-baseline early and uses byte-equality against those goldens as the unbreakable invariant.

full is the parity ceiling. Features like float (under nano too via the FLOAT carveout), complex, io, longint (MPZ), threading all live under full via vendor C carveouts that build.zig pulls in conditionally. The nano arm uses Zig replacements for almost everything; under full, the appropriate vendor C gets re-added so we match upstream MicroPython where the Zig port hasn't shipped yet.

Status¶

Per-feature build flags¶

Beyond -Dconfig, the build accepts orthogonal -Dfeature-* flags (ADR 002) so a downstream user can pick exactly the surface they need.

Flipping a flag away from its preset default emits a warning that the corpus invariant is not enforced for the resulting build. nano (all defaults) is the only frozen-output configuration.

Target architectures¶

The PyError convention¶

Every error in the runtime propagates via Zig's error union (replacing MicroPython's setjmp/longjmp-based NLR machinery):

pub const PyError = error{PyRaise};

// Internal fns return PyError!T
fn computeImpl(...) PyError!Obj {
    if (bad) try raiseTypeError(...);
    return result;
}

// C-ABI exports bridge via pyerr.bridge
pub export fn npy_compute(...) callconv(.c) Obj {
    return pyerr.bridge(Obj, computeImpl(...));
}

pyerr.bridge(T, r) catches error.PyRaise and returns std.mem.zeroes(T) (which is MP_OBJ_NULL for Obj); the raised exception is stashed in pyerr.pending_exc, a polled sentinel that vendor C callers (where any remain, primarily in the float and longint carveouts) check after the call.

Dropping the NLR machinery made wasm targeting straightforward. Wasm lacks a native setjmp, and the previous code path traded setjmp tricks for a runtime trap. See Internals: PyError Convention for the detailed walkthrough.

Type mirrors¶

src/mp_mirrors.zig (~1,200 lines, ~40 extern struct definitions) mirrors vendor MicroPython C structs field-for-field. Each mirror has @sizeOf / @offsetOf asserts in a comptime block, so a vendor git pull that shifts a struct layout fails the build at the right place instead of corrupting memory silently.

pub const MpRawCode = extern struct {
    proto_fun_indicator: [2]u8,
    kind: u8,
    is_generator: bool,
    fun_data: ?*const anyopaque,
    children: [*c]?*MpRawCode,
};

This lets Zig read/write the same memory the vendor C code did, by name (the type-mirror pattern described in TR-001 §3.1).

See Internals → Type Mirrors.

The 86-line tombstone¶

Under -Dconfig=nano the only .c file in the compile graph is src/objects/objtype/objmodule_shim.c (86 LOC). It holds MICROPY_REGISTERED_MODULES, the qstr-gen-emitted macro table that lists every available module. The macros are produced by tools/makeqstrdefs.py (vendor MicroPython tooling that scans .c files at build time, never .zig). Eliminating this tombstone needs a qstr-gen port; the shim is the architectural floor.

Under -Dconfig=full the build pulls in additional vendor C carveouts (MPZ, IO, builtin_help, warning, complex) that we have intentionally not ported. They're well-tuned vendor code, and the parity target benefits more from re-using them than from re-implementing them. The carveout list lives in build.zig; toggling between nano and full is a single -Dconfig= flip.

Test infrastructure¶

Four independent gates plus an aggregate. Each answers a different question. Read the "What 'green' means" section in CLAUDE.md for the full breakdown.

Beyond these: - Unit tests: make test-unit runs ~100 Zig unit tests across the compiler + VM tracks. - tools/compile-exec-test.py: secondary differential gate. Compiles each .py with zigpy-compile, runs the .mpy through vendor MicroPython, diffs the output. Catches "the Zig compiler produced subtly wrong bytecode" cases the byte-exact gate misses. - golden-programs/: 823 real Python programs sourced from MicroPython tests, CPython tests, PEP examples, and curated demos, bucketed by what they need (ADR 008). Drives the per-program coverage view in COMPATIBILITY.md. - tools/metrics.py: longitudinal recording of binary size + coverage + (opt-in) perf to metrics/history.ndjson. Each session records baseline at start and post-change at end.

See Build & Test for the full Makefile reference.

Layout reference¶

src/
├── main.zig                 — entry point + comptime force-analyze block
├── main_vm.zig              — VM-only entry point (for the slim build)
├── mp_mirrors.zig           — Zig mirrors of vendor MicroPython C structs (~40 extern structs)
├── mp_type_roms.zig         — ROM-resident type singletons
├── pyerr.zig                — PyError error union + polled-sentinel bridge
├── vendor_stubs.zig         — small C-ABI stubs for vendor symbols whose full implementation is deferred
├── compiler/                — pure-Zig compiler track
│   ├── lexer.zig            (with unit tests)
│   ├── parser.zig
│   ├── grammar_data.zig     (generated from the patched upstream grammar; see port/grammar-patched/)
│   ├── bigint.zig
│   ├── scope.zig
│   ├── emit.zig
│   ├── compile.zig
│   ├── qstr.zig
│   ├── mpy.zig              — .mpy v6 format
│   └── mpy_roundtrip.zig    — standalone roundtrip tool
├── runtime/                 — VM track
│   ├── runtime.zig          — binary_op, unary_op, call dispatch
│   ├── vm.zig               — bytecode interpreter
│   ├── mpstate.zig          — mp_state_ctx mirror
│   ├── persistentcode.zig   — .mpy loader
│   ├── zig_compile_bridge.zig — compile + load entry point
│   ├── raise.zig            — exception raise helpers
│   ├── argcheck.zig         — arg-checking helpers
│   ├── full_stubs.zig       — is_full-gated stubs for unported full features
│   └── ...
├── objects/                 — type implementations
│   ├── obj.zig              — base type dispatch
│   ├── objtype/             — type system (objtype, objmodule, objobject) + the 86-line tombstone
│   ├── primitives/          — int, cell
│   ├── strings/             — str, bytes, objstr, vstr, mpprint
│   ├── containers/          — list, dict, tuple, set, range, slice, array
│   ├── iters/               — filter, map, zip, deque, namedtuple, reversed, enumerate
│   └── callables/           — fun, generator, boundmeth, closure
├── modules/                 — builtin modules
│   ├── modbuiltins.zig      — print, len, range, etc.
│   ├── modsys.zig           — sys.path, sys.argv, etc.
│   ├── modmath.zig          — math.sqrt, math.pi, etc.
│   ├── modgc.zig            — gc.collect, gc.mem_free
│   ├── modstruct.zig        — struct.pack/unpack
│   ├── modarray.zig         — array.array
│   ├── modcollections.zig   — collections (namedtuple, OrderedDict, deque)
│   ├── moderrno.zig         — errno constants
│   ├── modmain.zig          — __main__ module ROM
│   ├── modmicropython.zig   — micropython builtin
│   └── file_io.zig          — file IO (is_full-gated)
├── memory/                  — allocator + GC
│   ├── alloc.zig            — m_malloc / m_realloc / m_new
│   ├── gc.zig               — minimal bump GC
│   └── ...
├── frontend/                — compiler-to-runtime bridge
│   ├── emitglue.zig
│   ├── parsenum.zig
│   ├── parsenumbase.zig
│   └── reader.zig
├── entry/
│   └── embed_exec.zig       — top-level script execution + REPL entry
├── support/
│   ├── smallint.zig         — Python-semantic divmod / floordiv
│   ├── unicode.zig          — unichar classification
│   ├── readline.zig         — REPL line editing (is_full-gated)
│   ├── mpprint.zig          — formatted printing
│   └── stream.zig           — stream protocol (is_full-gated)
└── tests.zig                — unit tests for shimless utility modules

See also¶

• TR-001 (notes/tech-reports/TR-001.md): methodology paper (scientific-paper format)
• ADR series: architecture decisions
• Internals: PyError Convention
• Internals: Type Mirrors
• Internals: qstr Generation
• Internals: Persistentcode Loader
• Internals: Wasm Composition
• Strangler Journey: the narrative (historical)