THE EFFECTS OF EXPLOITING PARALLEL ARCHITECTURES DURING LOCAL TRANSFORMATIONS ON COMPILER PERFORMANCE

Abstract

Multiprocessor personal computers and distributed computer systems are widely available. Exploiting these parallel and distributed architectures can enhance software performance. Compiler technology is the starting point for researchers with the goal of faster execution through parallelism. Compilers are well-understood programs. Traditionally, designers divide compilers into five or six phases executing sequentially. This paper outlines previous work in parallel compilers. In this thesis I explored the possibility of performing local transformations in parallel. I developed an optimizing program to perform these transformations. The optimizer reads an intermediate-code file, partitions the input into basic blocks, and optimizes the blocks on a varying number of processors. The intermediate-code file is in the form of Register Transfer Language (RTL). This optimizer prop-am executes on an Ethernet-based network of personal computers. The experiment gathered optimization times sequentially and in parallel, with between 2 and 15 computers. The optimizer performs redundancy elimination and partial evaluation transformation. The specific redundancy elimination transformations are common subexpression elimination and dead variable elimination. The specific partial evaluation transformations are algebraic simplification and constant folding. The input into the optimizer program consisted of three source files from the GNU gcc compiler. The optimizer gathered optimization times for a 526-basic-block file, a 1598-basic-block file, and a 4242-basic-block file. The experiment gathered optimization times for each file sequentially, and with 2, 3, 4, 5, 10, and 15 processors. The experiment Exploiting Parallel Architectures also measured execution times for the execution of redundancy elimination alone, and for all transformations combined. Results showed the number of transformations applied for these data sets had no significant impact on execution time. For all tests the sequential executions were the fastest because of the high overhead associated with distributed computing. The four-processor tests showed improvement over the two-processor tests. For the smallest file, the four-processor case showed further improvement. At four processors the medium and large files showed a slight slow down. These time increases are possibly due to file management issues.