Automatic Loop Invariant Generation for Data Dependence Analysis

Abstract

Parallelization of programs relies on sound and precise analysis of data dependences in the code, specifically, when dealing with loops. State-of-art tools are based on dynamic profiling and static analysis. They tend to over- and, occasionally, to under-approximate dependences. The former misses parallelization opportunities, the latter can change the behavior of the parallelized program. In this paper we present a sound and highly precise approach to generate data dependences based on deductive verification. The central technique is to infer a specific form of loop invariant tailored to express dependences. To achieve full automation, we adapt predicate abstraction in a suitable manner. To retain as much precision as possible, we generalized logic-based symbolic execution to compute abstract dependence predicates. We implemented our approach for Java on top of a deductive verification tool. The evaluation shows that our approach can generate highly precise data dependences for representative code taken from HPC applications.