Yhc/Erlang/Proof of concept

Introduction

This Wiki article describes an experiment targeting execution of Haskell programs on top of the Erlang Virtual Machine (BEAM). Haskell source code is compiled to Yhc Core with York Haskell Compiler (Yhc), then the program further discussed converts Yhc Core to Core Erlang; finally Erlang Compiler (erlc) compiles Core Erlang to the BEAM file format which can be loaded and executed by the Erlang VM.

There have been numerous discussions about Haskell (mainly GHC) runtime lacking some properties that are available in Erlang environment, as well as about possible improvements in Erlang language syntax and type system to bring some elements available in Haskell.

This experiment is an attempt to answer the critics from both sides. Once it becomes possible to execute Haskell programs in Erlang environment, Haskell users get access to the robust concurrency-oriented runtime, still being able to use Haskell native syntax. Erlang users get possibility to develop some algorithms with regard to the Haskell strong type system, while still being able to code directly in Erlang, where it seems more appropriate implementation-wise.

Implementation details

This section discusses in deep the approach taken in this experiment. It is good to remember that nothing yet is final; some of the techniques described may possibly make it into the mainstream code, while others may not. This is only the beginning.

Core Erlang overview

The Core Erlang initiative project is a "collaboration between the High-Performance Erlang (HiPE) project at the Department of Information Technology of Uppsala University, and Ericsson's OTP/Erlang developers".

Core Erlang is an intermediate form of Erlang source compilation. It provides a desugared (compared to Erlang) syntax of a strict functional language. Erlang source may be compiled to Core Erlang, and Core Erlang may be compiled to BEAM bytecode.

Core Erlang plays the same role in the Erlang compilation process as Yhc Core in the Haskell (Yhc) compilation process. So, it turns out to be the most convenient to do the conversion between these formats rather than between e. g. Haskell source and Erlang source.

There was an attempt made earlier to do similar things, only converting from Haskell source (indeed, a subset of Haskell syntax) to Core Erlang. This project is called Haskerl, developed by Torbjörn Törnkvist, (not to be confused with Will Partain's Haskerl. Some source code from Haskerl was used in this experiment, in particular, the algebraic data type to represent Core Erlang internally, and the pretty printer module for Core Erlang, both with some necessary extensions.

The whole compilation chain looks like this:

1. Haskell source modules are compiled into Yhc Core and linked;
2. Some overall program optimizations (functionality provided with the Yhc Core library) are performed on the linked Yhc Core;
3. Yhc Core is converted to Core Erlang;
4. The Erlang compiler erlc produces a BEAM file.

Haskell on BEAMs ;)

From the Erlang VM standpoint, a Haskell program is just a large(ish) Erlang module. To run the program, certain function exported from that module (usually, main) needs to be called with arguments as necessary. Often the special force function has to be applied to values returned from Haskell-originated module: otherwise some unfinished computation may be returned instead of the expected result.

From Haskell program standpoint, Erlang VM is just an execution environment providing system calls that are strict on all their arguments, and may have variable number of arguments. Haskell program may spawn concurrent processes able to receive messages of certain types (the idea of typed processes was borrowed from this Livejournal article (in Russian). The message distribution/transport mechanism is entirely provided by the Erlang VM runtime.

Lazy computations

Erlang is a strict language. This means that functions always evaluate their agruments, and values get passed around already computed. In Haskell, due to its lazy/non-strict nature, some values are passed around un-evaluated, and may be evaluated later as needed (or never at all). "Traditional" implementation of Haskell runtime often combine non-strict evaluation with memoization which serves to avoid redundant evaluation of the same value several times.

However memoization is not possible when compiling Haskell into (Core) Erlang because of Erlang's single assignment nature. Once created, objects in Erlang are immutable (with few exceptions that are of no value in this experiment). As alternative to memoization, the following approach, based on strictness analysis is used.

The Yhc Core Strictness Analyzer is able to determine whether a function is strict on some (or none) of its arguments. This is done via analysis from the bottom up, with OS/platform primitives (usually assumed strict on all arguments, and the same applies to Erlang primitives). If a function passes its argument to another function which will evaluate it, the former function is also strict on that argument. If a function unconditionally evaluates its argument as a case statement scrutinee, it is also strict on that argument, and this strictness propagates to other functions which call it.

After Yhc Core is linked and optimized, strictness analysis is run on it. All Erlang primitives (see below about possible ways to call Erlang functions from Haskell) are considered strict on all arguments (and they naturally are). If a function is determined to be strict on some of its arguments, for each such argument a code is inserted into the function's body to make sure these arguments will be evaluated as early as possible, and will be passed around evaluated. While this does not replace memoization, it is expected that such approach will at least eliminate some redundant computations.

Another consequence, for functions with side effects involved in sequential computations, the runtime implementation must carefully observe that already computed value is passed to the continuation, rather than an unevaluated thunk, since repeated evaluation of the thunk results in repeated side effect.

Examples

Sample code to demonstrate results of the experiment was checked into the Yhc Darcs repo. Haskell and Erlang sources as well as compiled BEAM files are located at http://darcs.haskell.org/yhc/src/translator/erlang/00proof/ .