Contrary to what you might expect, llvmlite does not use any LLVM shared libraries that may be present on the system, or in the conda environment. The parts of LLVM required by llvmlite are statically linked at build time. As a result, installing llvmlite from a binary package does not also require the end user to install LLVM. (For more details on the reasoning behind this, see: :ref:`why-static`)
Building LLVM for llvmlite is challenging, so we strongly recommend installing a binary package where we have built and tested everything for you. Official conda packages are available in the Anaconda distribution:
conda install llvmlite
conda install --channel numba llvmlite
Binary wheels are also available for installation from PyPI:
pip install llvmlite
Building llvmlite requires first building LLVM. Do not use prebuilt LLVM binaries from your OS distribution or the LLVM website! There will likely be a mismatch in version or build options, and LLVM will be missing certain patches that are critical for llvmlite operation.
Before building, you must have the following:
- On Windows:
- Visual Studio 2015 (Update 3) or later, to compile LLVM and llvmlite. The free Express edition is acceptable.
- CMake installed.
- On Linux:
- g++ (>= 4.8) and CMake
- If building LLVM on Ubuntu, the linker may report an error
if the development version of
libeditis not installed. If you run into this problem, install
- On Mac:
- Xcode for the compiler tools, and CMake
If you can build llvmlite inside a conda environment, you can install a prebuilt LLVM binary package and skip this step:
conda install -c numba llvmdev
The manual instructions below describe the main steps, but refer to the recipe for details:
#. Download the LLVM 6.0.0 source code. (Note that PPC64LE requires LLVM 6.0.1 for specific bug fixes.)
Download or git checkout the llvmlite source code.
Decompress the LLVM tar file and apply the following patches from the
llvm-lto-static.patch: Fix issue with LTO shared library on Windows
D47188-svml.patch: Add support for vectorized math functions via Intel SVML
0001-Transforms-Add-missing-header-for-InstructionCombini.patch: Fix release bug with LLVM 6.0.0, skip on LLVM 6.0.1.
twine_cfg_undefined_behavior.patch: Fix obscure memory corruption bug in LLVM that hasn't been fixed in master yet
- For Linux/macOS:
export PREFIX=desired_install_location CPU_COUNT=N(
Nis number of parallel compile tasks)
- Run the build.sh script in the llvmdev conda recipe from the LLVM source directory
- For Windows:
- Run the bld.bat script in the llvmdev conda recipe from the LLVM source directory.
To build the llvmlite C wrapper, which embeds a statically linked copy of the required subset of LLVM, run the following from the llvmlite source directory:
python setup.py build
If your LLVM is installed in a nonstandard location, set the
LLVM_CONFIGenvironment variable to the location of the corresponding
llvm-config.exeexecutable. This variable must persist into the installation of llvmlite---for example, into a Python environment.
EXAMPLE: If LLVM is installed in
llvm-configbinary located at
To validate your build, run the test suite by running:
python -m llvmlite.tests
If the validation is successful, install by running:
python setup.py install
Why Static Linking to LLVM?
The llvmlite package uses LLVM via ctypes calls to a C wrapper that is statically linked to LLVM. Some people are surprised that llvmlite uses static linkage to LLVM, but there are several important reasons for this:
- The LLVM API has not historically been stable across releases - Although things have improved since LLVM 4.0, there are still enough changes between LLVM releases to cause compilation issues if the right version is not matched with llvmlite.
- The LLVM shipped by most Linux distributions is not the version llvmlite needs - The release cycles of Linux distributions will never line up with LLVM or llvmlite releases.
- We need to patch LLVM - The binary packages of llvmlite are built against LLVM with a handful of patches to either fix bugs or to add features that have not yet been merged upstream. In some cases, we've had to carry patches for several releases before they make it into LLVM.
- We don't need most of LLVM - We are sensitive to the install size of llvmlite, and a full build of LLVM is quite large. We can dramatically reduce the total disk needed by an llvmlite user (who typically doesn't need the rest of LLVM, ignoring the version matching issue) by statically linking to the library and pruning the symbols we do not need.
- Numba can use multiple LLVM builds at once - Some Numba targets (AMD GPU, for example) may require different LLVM versions or non-mainline forks of LLVM to work. These other LLVMs can be wrapped in a similar fashion as llvmlite, and will stay isolated.
- We need to support Windows + Python 2.7 - Python 2.7 extensions on Windows needs to be built with Visual Studio 2008 for ABI compatibility reasons. This presents a serious issue as VS2008 can no longer build LLVM. The best workaround we have found (until the sunset of Python 2.7) is to build LLVM and the llvmlite C wrapper with VS2015 and call it through ctypes. This is not ideal, but experience has shown it seems to work.
Static linkage of LLVM was definitely not our goal early in Numba development, but seems to have become the only workable solution given our constraints.