Global analysis for partitioning non-strict programs into sequential threads

In this paper we present a new solution to the problem ofcompiling aneager, non-strict language into multiple se-quential threads. The solution is described using an in-termediate program form developed for the programminglanguage Id [Nik90], a functional language extended withI-structures [ANP86] and M-structures [BNA91]. A similarintermediate form has also been suggested for imperativelanguages [BP89], as a way of exposing parallelism in thoselanguages. With suitable restrictions, we believe our methodis also appropriate for lazy, purely functional languages suchas Haskell [HWe90].Throughout this paper, a thread will mean a subset ofthe instructions comprising a procedure body, such that:1. A compile-time instruction ordering can be determinedfor the thread which is valid for all contexts in whichthe containing procedure can be invoked.2. Once the first instruction in a thread is executed, itis always possible to execute each of the remaining in-structions, in the compile-time ordering, without pause,interruption, or execution of instructions from otherthreads. [Tra91b]Threads of this form are required to implement a languageon parallel multithreaded hardware (e.g., Monsoon [PT91,PC90] and *T [N PA92]), and they yield efficient implemen-tations on conventional architectures when combined with asuit able abstract machine such as TAM [CSS+ 91],Our algorithms are greedy, and the following commentsfrom [SCVE91] apply:Partitioning decisions imply trade-offs betweenpa~allelism, synchronization cost, and sequentialefficiency. However, given the limits on threadsize imposed by the language model, the use ofsplit-phase accesses, and the control paradigm,we simply attempt to make partitions as large aspossible . . .]