There are several computer system architectures which have the goal of exploiting parallelism-multiprocessors, vector machines and array processors. For each of these architectures there have been attempts to design compilers to optimize programs written in conventional languages (e.g. "vectorizing" compilers for the FORTRAN language). There have also been new language designs to facilitate using these systems, such as Concurrent PASCAL for multiprocessors,6 and languages that utilize the features of such systems directly, such as GLYPNIR for the Illiac IV array processor19 and various "vectorizing" dialects of FORTRAN. These languages almost always make the multiprocessor, vector, or array properties of the computer visible to the programmer-that is, they are actually vehicles whereby the programmer helps the compiler uncover parallelism. Many of these languages or dialects are "unnatural" in that they closely reflect the behavior of the system for which they were designed, rather than reflecting the way programmers think about problem solutions.