A lightweight LLVM python binding for writing JIT compilers
The old llvmpy binding exposes a lot of LLVM APIs but the mapping of C++-style memory management to Python is error prone. Numba and many JIT compilers do not need a full LLVM API. Only the IR builder, optimizer, and JIT compiler APIs are necessary.
llvmlite is a project originally tailored for Numba's needs, using the following approach:
- A small C wrapper around the parts of the LLVM C++ API we need that are not already exposed by the LLVM C API.
- A ctypes Python wrapper around the C API.
- A pure Python implementation of the subset of the LLVM IR builder that we need for Numba.
- The IR builder is pure Python code and decoupled from LLVM's frequently-changing C++ APIs.
- Materializing a LLVM module calls LLVM's IR parser which provides better error messages than step-by-step IR building through the C++ API (no more segfaults or process aborts).
- Most of llvmlite uses the LLVM C API which is small but very stable (low maintenance when changing LLVM version).
- The binding is not a Python C-extension, but a plain DLL accessed using ctypes (no need to wrestle with Python's compiler requirements and C++ 11 compatibility).
- The Python binding layer has sane memory management.
- llvmlite is quite faster than llvmpy's thanks to a much simpler architeture (the Numba test suite is twice faster than it was).
The llvmlite.llvmpy namespace provides a minimal llvmpy compatibility layer.
llvmlite works with Python 2.7 and Python 3.4 or greater.
As of version 0.29.0, llvmlite requires LLVM 7.0.x or later
Historical compatibility table:
| llvmlite versions | compatible LLVM versions |
|---|---|
| 0.29.0 - ... | 7.0.x, 7.1.x, 8.0.x |
| 0.27.0 - 0.28.0 | 7.0.x |
| 0.23.0 - 0.26.0 | 6.0.x |
| 0.21.0 - 0.22.0 | 5.0.x |
| 0.17.0 - 0.20.0 | 4.0.x |
| 0.16.0 - 0.17.0 | 3.9.x |
| 0.13.0 - 0.15.0 | 3.8.x |
| 0.9.0 - 0.12.1 | 3.7.x |
| 0.6.0 - 0.8.0 | 3.6.x |
| 0.1.0 - 0.5.1 | 3.5.x |
You'll find the documentation at http://llvmlite.pydata.org
We recommend you use the binaries provided by the Numba team for the Conda package manager. You can find them in Numba's anaconda.org channel. For example:
$ conda install --channel=numba llvmlite(or, simply, the official llvmlite package provided in the Anaconda distribution)
If you don't want to use our pre-built packages, you can compile and install llvmlite yourself. The documentation will teach you how: http://llvmlite.pydata.org/en/latest/install/index.html