Paul learned programming with 6502 in the 1980s, wrote drivers and eventually a full system emulator of Apple Newton Messagepads which was first released in 2004. He eventually graduated with a Ph.D. in Artficial Intelligence in 2006 and fell in love with Erlang when he introduced ejabberd to power Nabaztag wifi rabbits as the CTO of Violet in 2007. For a decade, he led Semiocast, a data science boutique consultancy using Erlang and Elixir to crawl and process the web. An open source advocate and hobbyist, he’s been a contributor to various Erlang projects and had recently a lot of fun working with AtomVM developers.
The just-in-time compiler was introduced at CodeBEAM V in 2020 by Lukas Larsson. It shipped with OTP-24 and succeeded HiPE in dramatically improving the speed of the BEAM.
In this talk we will discuss how native code compilation is brought to AtomVM, the virtual machine that runs on micro-controllers and in the browser, following a slightly different approach.
The talk will compare BEAM and AtomVM approaches to emulation. Typically, BeamASM, the JIT compiler of the BEAM, was designed with simplicity and maintenance in mind to cope with the size of OTP team. As AtomVM team is even smaller, simplicity and maintenance are also key requirements. AtomVM’s native compiler also needs to address AtomVM’s constraints such as portability to most important supported architectures (RISCV, ARM32, Xtensa, WASM) as well as compilation directly on the devices with very few amounts of RAM, as AtomVM can blink with 128 KB of RAM and run Distributed Erlang with as few as 400 KB.
Eventually, AtomVM JIT compiler, being a 2025 project, leverages newest software engineering technologies and the talk will also cover how an LLM eventually helped with such a complex task.
Key Takeaways:
Target Audience: