Add Numba

[Rupt]

• Add initial support for Numba compilation: asm, demangling, execution

Numba wraps Python functions in Dispatcher objects. Each dispatcher contains zero or more compiled argument-type-indexed overloads of its function. We import the user's code as a module, and emit the code from all overloads of all dispatchers that the module publicly exposes.

Name mangling is odd in Numba. It uses a similar syntax to C++, but also encodes non-symbol (r"[^a-z0-9_]") characters as _%02x. This encoding yields valid identifier names, so it is not strictly invertible. Here, I have hard-coded some replacements to decode some common cases.

To do

• Answer questions of [COMPILER REQUEST]: Numba #5591.

• Install in https://github.com/compiler-explorer/infra

Locally, I have installed a virtual environment with

python3.11 -m venv venv_numba
venv_numba/bin/python -m pip install numba==0.58.0 llvmlite==0.41.0 scipy==1.11.3 numpy==1.25.2
# (I recommend pinning versions because Python packages regularly violate SemVer, and CE wants long-term reproducibility.)

and configure its use with

# compiler-explorer/etc/config/numba.local.properties
compilers=&numba
group.numba.compilers=numba58
compiler.numba58.exe=/home/rupt/venv_numba/bin/python3.11

• Automatically run python tests?

Locally, I run this test suite with

/home/rupt/venv_numba/bin/python3.11 -m unittest etc/scripts/test_numba_wrapper.py

[mattgodbolt]

I like the idea of this! When we were using numba at work I hacked something similar. But, the main issue is the python installation, which a number of things are blocked on: virtual environments etc are troublesome to create and maintain over multiple OSes, and our installation system is not good with them: we install to a local directory and then relocate atomically to the "install" location at the moment and many python virtual environments don't work with this That's not to say we don't want this at all! I'm excited; but we have some infrastructural things to fix on our side!