pymcu-arm-toolchain

Vendored LLVM toolchain for the PyMCU ARM (Cortex-M) backend.

The ARM backend lowers PyMCU's architecture-agnostic IR to LLVM IR and then invokes five LLVM command-line tools to produce a flashable flat binary:

opt   llc   llvm-mc   ld.lld   llvm-objcopy

This package is the ARM counterpart of pymcu-avr-toolchain: a platform-specific wheel whose only job is to ship those tools so that pip install pymcu-compiler[arm] is fully self-contained — no brew install llvm or apt install lld step required on supported platforms.

License: MIT packaging + Apache-2.0 WITH LLVM-exception binaries

The Python packaging code in this repo is MIT.

The bundled LLVM binaries are distributed under the Apache License 2.0 with LLVM Exceptions — the official license of the LLVM Project. This is a permissive license with no copyleft. Specifically:

• LLVM is not GPL. There is no copyleft infection on this package, on PyMCU, or on firmware compiled with these tools.
• The LLVM Exception clause explicitly exempts compiled output: binary firmware produced by PyMCU through opt/llc carries no license obligation from LLVM.
• Contrast with AVR: pymcu-avr-toolchain bundles avr-gcc (derived from GCC), which is GPL-3.0. That package must therefore be GPL- licensed. The isolation mechanism is the same — a separate optional package — but the license differs because the upstream compilers differ.

See LICENSE for the complete text and a plain-language explanation.

Installation

pip install pymcu-arm-toolchain

libLLVM.{dylib,so} is 200–300 MB of code even in a release build — too large for PyPI's 100 MB per-file ceiling. The distribution is therefore split:

The LLVM tools are downloaded automatically. The first call to get_tool() (or any PyMCU ARM build) downloads the official LLVM release for your platform from GitHub and extracts the five required tools into the shared cache at ~/.pymcu/tools/. Subsequent calls are instant.

import pymcu_arm_toolchain
opt = pymcu_arm_toolchain.get_tool("opt")   # downloads on first call

For CI or air-gapped environments, override the download URL or install the binary wheel directly:

# Linux x86-64
pip install https://github.com/PyMCU/pymcu-arm-toolchain/releases/download/v22.1.7/pymcu_arm_toolchain-22.1.7-py3-none-manylinux_2_17_x86_64.whl

# Linux arm64
pip install https://github.com/PyMCU/pymcu-arm-toolchain/releases/download/v22.1.7/pymcu_arm_toolchain-22.1.7-py3-none-manylinux_2_17_aarch64.whl

# macOS Apple Silicon
pip install https://github.com/PyMCU/pymcu-arm-toolchain/releases/download/v22.1.7/pymcu_arm_toolchain-22.1.7-py3-none-macosx_14_0_arm64.whl

# Windows x86-64
pip install https://github.com/PyMCU/pymcu-arm-toolchain/releases/download/v22.1.7/pymcu_arm_toolchain-22.1.7-py3-none-win_amd64.whl

System LLVM (alternative)

If you already have LLVM 18+ installed (brew install llvm lld / apt install llvm lld), the Rp2040LlvmToolchain driver finds the tools via the cache or PATH automatically — no package needed.

How the driver resolves tools

Rp2040LlvmToolchain (in pymcu-arm) checks these sources in order:

A missing wheel never blocks a developer who already has LLVM installed.

Inspecting installed tools

pymcu-arm-toolchain-info          # entry-point
python -m pymcu_arm_toolchain status

Example output:

pymcu-arm-toolchain (LLVM 22.1.7, platform darwin-arm64)
  [ok] opt            ~/.pymcu/tools/darwin-arm64/llvm-arm/bin/opt
  [ok] llc            ~/.pymcu/tools/darwin-arm64/llvm-arm/bin/llc
  [ok] llvm-mc        ~/.pymcu/tools/darwin-arm64/llvm-arm/bin/llvm-mc
  [ok] ld.lld         ~/.pymcu/tools/darwin-arm64/llvm-arm/bin/ld.lld
  [ok] llvm-objcopy   ~/.pymcu/tools/darwin-arm64/llvm-arm/bin/llvm-objcopy

Seeding the cache manually

For platforms without a published wheel, or to use a locally installed LLVM:

# Download the pinned LLVM release automatically:
python -m pymcu_arm_toolchain fetch --cache

# Use a Homebrew LLVM (symlinked, dev convenience):
python -m pymcu_arm_toolchain fetch --cache \
    --from-dir /opt/homebrew/opt/llvm --link

# Copy from a system install:
python -m pymcu_arm_toolchain fetch --cache \
    --from-dir /usr/lib/llvm-19

For maintainers: publishing a new wheel

Release process

1. Update LLVM_VERSION in src/pymcu_arm_toolchain/__init__.py and version in pyproject.toml.
2. Tag and push:

   git tag v22.1.7
   git push origin v22.1.7
   

3. The build-wheels.yml workflow fires automatically:
   • Downloads the official LLVM .tar.xz, slims to the five tools + shared libs via scripts/stage-llvm.sh, builds a py3-none-<platform> wheel per platform.
   • Binary wheels → GitHub Releases (too large for PyPI's 100 MB limit).
   • PyPI receives only the pure-Python sdist stub.
   • publish-pypi uses OIDC trusted publishing (no stored token required).

Required GitHub configuration

Building a wheel locally

bash scripts/stage-llvm.sh /opt/homebrew/opt/llvm /tmp/staged-llvm

ARMT_TOOLCHAIN_DIR=/tmp/staged-llvm \
WHEEL_PLATFORM_TAG=macosx_14_0_arm64 \
uv build --wheel

Environment variables

Version history