Branch prediction for enhancing fine-grained parallelism in Prolog

Abstract

Branch instructions create barriers to instruction fetching, thus greatly reducing the fine-grained parallelism of programs. One common method for solving this problem is branch prediction. We first present four lemmas to clarify the relationship between the branch prediction hit rate and system performance, hardware efficiency, and branch prediction overhead. We then propose a new branch prediction method called PAM (Period Adaptive Method). An abstract model and detailed implementation of PAM are described. The prediction hit rate of this method was measured using ten Prolog benchmark programs and found to be 97%. When implemented in a superscalar Prolog system, PAM enhances the degree of system parallelism by 80%.