HVM3 - Work In Progress

HVM3 will combine the strengths of HVM1 and HVM2 while addressing their limitations. It aims to be the long-term runtime for Bend. It has 2 modes:

• HVML: lazy mode. Pointers represent positive-to-negative ports in polarized nets, which coincides with the Interaction Calculus. Strengths: efficient lazy evaluation, β-optimality. Drawbacks: 1. whnf() may return a pending variable; 2.garbage collection is needed; 3. parallelism is less pervasive. It is based on HVM1.

• HVMS: strict mode. Pointers represent aux-to-main ports, resulting in a tree-like memory format. Strengths: efficient massively parallel evaluation and no garbage-collection. Drawbacks: not laziness and no optimal evaluation. It is based on HVM2.

HVM3 is a work-in-progress. Its features are being actively implemented.

Install

1. Install Haskell.

2. Clone this repository.

3. Install it with cabal install.

Usage

hvml run   file.hvml  # runs lazy-mode, interpreted
hvml run-c file.hvml  # runs lazy-mode, compiled
hvms run   file.hvms  # runs strict-mode, interpreted (TODO)
hvms run-c file.hvms  # runs strict-mode, compiled (TODO)

Note that run-c will also generate a standalone .main.c file, which you can compile and run independently. See examples on the book/ directory.

Performance

Benchmarks will be added later. In the few programs tested, HVM3 is up to 42x faster single-core than Bend, due to its compiler (Bend was interpreted). It is also 2x-3x faster than Node.js and Haskell in the first program I tested, but possibly slower in others. HVM3 is a work-in-progress. It is currently single threaded. Threading (both on CPU and GPU) will be added later.