A lightweight LLVM-Python binding for writing JIT compilers
Llvmpy became hard to maintain because:
- It has a cumbersome architecture.
- The C++11 requirement of recent LLVM versions does not go well with the compiler and runtime ABI requirements of some Python versions, especially under Windows.
Llvmpy also proved to be responsible for a sizable part of Numba’s compilation times, because of its inefficient layering and object encapsulation. Fixing this issue inside the llvmpy codebase seemed a time-consuming and uncertain task.
The Numba developers decided to start a new binding from scratch, with an entirely different architecture, centered around the specific requirements of a JIT compiler.
While llvmpy exposed large parts of the LLVM C++ API for direct calls into the LLVM library, llvmlite takes an entirely different approach. Llvmlite starts from the needs of a JIT compiler and splits them into two decoupled tasks:
- Construction of a Module, function by function, Instruction by instruction.
- Compilation and optimization of the module into machine code.
The compilation of an LLVM module takes the IR in textual form and feeds it into LLVM’s parsing API. It then returns a thin wrapper around LLVM’s C++ module object. This is the role of the binding layer.
Once parsed, the module’s source code cannot be modified, which loses the flexibility of the direct mapping of C++ APIs into Python that was provided by llvmpy but saves a great deal of maintenance.
Despite minimizing the API surface with LLVM, llvmlite is impacted by changes to LLVM’s C++ API, which can occur at every feature release. Therefore, each llvmlite version is targeted to a specific LLVM feature version and works across all given bugfix releases of that version.
EXAMPLE: Llvmlite 0.12.0 works with LLVM 3.8.0 and 3.8.1, but it does not work with LLVM 3.7.0 or 3.9.0.
Numba’s requirements determine which LLVM version is supported.
At this time, we reserve the possibility of slightly breaking the llvmlite API at each release, for the following reasons:
- Changes in LLVM behaviour, such as differences in the IR across versions.
- As a young library, llvmlite has room for improvement or fixes to the existing APIs.