Path Splitting: A Technique for Improving Data-Flow Analysis

The eeectiveness of traditional global optimizations is limited by the quality of data-ow information available to the compiler. The rich control structure of many programs hinders global optimizations by polluting data-ow information and forcing the compiler to make worst-case assumptions during data-ow analysis. Path splitting is a new technique for improving the amount of data-ow information statically available to the compiler. It replicates code in order to provide optimal reaching deenitions information within given regions of the control-ow graph. This improved information is used to extend the applicability of various classical code optimizations, including copy propagation, common subexpression elimination, dead code elimination, and code hoisting. An experimental implementation of path splitting in the SUIF C compiler indicates that path splitting eeectively restructures loops, reducing the cycle count within them by up to a factor of two, and resulting in over 7% run time performance improvements for C programs. Although path splitting can cause exponential growth in code size, applying it selectively limits code growth to below 5% for realistic C programs, and does not noticeably degrade cache performance.