Ondřej Čertík - LPython: Novel, Fast, Retargetable Python Compiler | SciPy 2024

LPython is a novel, fast, retargetable Python compiler that aims to fix common Python issues like slow import time, low runtime performance, and packaging challenges

Key features:

• Strict subset of CPython - anything that runs in LPython will run unmodified in CPython
• Requires type annotations for compilation
• Compiles to fast, lean binaries without Python dependencies
• Interactive usage and binary compilation supported
• Runs in Jupyter notebooks
• WebAssembly backend available for browser execution

Architecture:

• Front-ends: LPython (Python), LFortran (Fortran), LC (C++)
• Uses Abstract Semantic Representation (ASR) as intermediate representation
• Multiple backends including LLVM, WebAssembly, and high-level code generation (C, C++, Python, Fortran, Julia)

Performance characteristics:

• Fast compilation times (30 seconds on Mac laptop)
• Runtime performance competitive with or faster than C in some cases
• Small binary sizes (~16KB)
• No Python runtime dependency in compiled code

Development aspects:

• Written in C++, compiles quickly
• Small team (5-10 people)
• Easy to extend with new frontends (2 months to add C++) and backends (2 weeks to 2 months)
• Currently alpha stage but approaching beta
• Clear error messages with GitHub issue reporting links

Integration capabilities:

• Can call into C libraries
• Python decorator interface similar to Numba
• SymPy support for symbolic computation
• NumPy array operations supported
• OpenMP support for loop parallelization