The GRIN project: A highly optimising back end for lazy functional languages

Abstract

Low level optimisations from conventional compiler technology often give very poor results when applied to code from lazy functional languages, mainly because of the completely different structure of the code, unknown control flow, etc. A novel approach to compiling laziness is needed.

We describe a complete back end for lazy functional languages, which uses various interprocedural optimisations to produce highly optimised code. The main features of our new back end are the following. It uses a monadic intermediate code, called GRIN (Graph Reduction Intermediate Notation). This code has a very “functional flavour”, making it well suited for analysis and program transformations, but at the same time provides the “low level” machinery needed to express many concrete implementation concerns. Using a heap points-to analysis, we are able to eliminate most unknown control flow due to evals (i.e., forcing of closures) and applications of higher order functions, in the program. A transformation machinery uses many, each very simple, GRIN program transformations to optimise the intermediate code. Eventually, the GRIN code is translated into RISC machine code, and we apply an interprocedural register allocation algorithm, followed by many other low level optimisations. The elimination of unknown control flow, made earlier, will help a lot in making the low level optimisations work well.

Preliminary measurements look very promising: we are currently twice as fast as the Glasgow Haskell Compiler for some small programs. Our approach still gives us many opportunities for further optimisations (though yet unexplored).