Browse Books

Reviewer: Frank Lawrence Friedman

This book represents the publication of Ellis's PhD dissertation which received the 1985 ACM Doctoral Dissertation Award (the fourth such award in the series initiated by the ACM in 1982). The importance of the dissertation is perhaps best summarized by John White, the Chairman of the 1985 ACM Doctoral Dissertation Award Subcommittee, in the Series Foreword presented at the beginning of the book book: John Ellis was judged to have made an outstanding contribution to the problem of exploiting the parallelism available in emerging multiprocessor architectures. Ellis's thesis is that high-quality compilers can be written for VLIW (Very Long Instruction Word) computers. These machines drive multiple parallel RISCs (Reduced Instruction Set Computers) with a single instruction stream. Each instruction, however, is long enough to command all of the RISCs at once. Exploiting the parallelism available in VLIW machines can be very difficult if done by hand. Thus, Ellis developed a compiler that incorporates a number of optimizations based on trace scheduling and memory reference and memory bank disambiguation. Trace scheduling increases parallelism by scheduling several basic blocks at once. The compiler increases parallelism further by unrolling loops and by using disambiguation algorithms to tell at least some of the time when vector and bank references cannot collide. Very Long Instruction Word (VLIW) computers are reduced-instruction-set machines with a large number of parallel, pipelined functional units, but a single thread of control. They offer much promise for order-of-magnitude increases in processing speed, but are impossible to program without a high-level language compiler. By building a working compiler for such languages for use in compiling ordinary scientific programs, Ellis has demonstrated the practicality of combining old and new compiling technology to successfully address the problem of efficient use of parallel architectures. Ellis's compiler uses several new compilation techniques, such as trace scheduling (to identify more parallelism), memory-reference and memory-bank disambiguation (to increase memory bandwith), and new code-generation algorithms to build his compiler. The dissertation includes an in-depth discussion of these techniques and algorithms and a discussion of the results of preliminary experiments testing the compiler and various aspects of VLIW architectures. The results show that for many scientific applications, VLIW architectures may be built which can achieve an order-of-mag- nitude improvement in efficiency over current machines. The ability to build compiler systems that produce code for the efficient use of parallel architectures is a critical and largely missing link in the quest to successfully use these architectures in the variety of application domains for which they are intended. The book does indeed make an outstanding contribution, and is must reading for computer researchers and practitioners involved in the compiling and parallel architecture areas. In fact, owing largely to the beauty of the prose, the painstaking attention to clear, concise statements of motivation, goals, methods, and results, this thesis belongs in every computer scientist's library.