tgcrypto2 (Zig backend)

Fast and Portable Cryptography Extension Library for Pyrogram — reimplemented in Zig 0.16, drop-in compatible with the original C extension, plus a native xor(a, b) helper.

Build & publish to PyPI

python -m pip install --upgrade build twine
python -m build          # produces sdist + wheel in dist/
twine upload dist/*

ziglang==0.16.0 is declared as a [build-system] requires dependency in pyproject.toml, so pip/build fetches it automatically into an isolated build environment — no system Zig install needed, and no .so is committed to the repo. When someone runs pip install tgcrypto2, setup.py's custom build_ext step shells out to python -m ziglang build so to compile src/tgcrypto.zig fresh for their platform/Python version, then places the resulting _tgcrypto extension where Python expects it.

Local development build

pip install -e .
python -m pytest

What's inside

• ige256_encrypt, ige256_decrypt
• ctr256_encrypt, ctr256_decrypt

All functions match the reference C implementation's behavior and signatures exactly (verified against the full upstream NIST/random test suite).

Free-threading (Python 3.13+ / PEP 703)

_tgcrypto opts into running under a free-threaded (Py_GIL_DISABLED, python3.1xt) interpreter via the Py_mod_gil module slot. This matters because a C extension that does not declare support gets silently downgraded by CPython: the GIL is re-enabled for the entire process for as long as that module stays imported, not just for that module — so an unaware dependency can quietly take away the benefit of free-threading from everything else running in the same interpreter.

This module is safe to opt in because every exported function is stateless per call: each allocates and returns a fresh output buffer, and the AES context (std.crypto.core.aes) lives on the stack for the duration of a single call with no shared mutable state. The one thing to be aware of: ctr256_*/cbc256_* mutate the caller-supplied iv/state buffer in place (matching the reference C behavior and this project's own upstream tests, which pass plain bytes). If the same bytearray is handed to two threads concurrently, that's a data race in the caller's Python code — no different from any two threads mutating a shared mutable object without a lock, and not something the GIL was protecting either, since the GIL is already released around the actual compute.

On Python < 3.13 (no Py_mod_gil slot exists) this is a no-op — the module builds and behaves identically to a normal GIL-bound extension.