Automatic Loop Invariant Generation for Data Dependence Analysis

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 pa...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:2022 IEEE/ACM 10th International Conference on Formal Methods in Software Engineering (FormaliSE) S. 34 - 45
Hauptverfasser: Tabar, Asmae Heydari, Bubel, Richard, Hahnle, Reiner
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: ACM 01.05.2022
Schlagworte:
Abstraction; Invariants
Automation
Behavioral sciences
Codes
Computing methodologies→Massively parallel algorithms
Formal software verification
Java
Parallel algorithms
Software
Software and its engineering→Software verification
Static analysis
Theory of computation→Programming logic
ISSN:2575-5099
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung: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.
ISSN:2575-5099
DOI:10.1145/3524482.3527649