Analyse de dépendances ML pour les évaluateurs de logiciels critiques

Critical software needs to obtain an assessment before commissioning in order to ensure compliance with standards. This assessment is given after a long task of software analysis performed by assessors. They may be helped by tools, used interactively, to build models using information-flow analyses. Tools like SPARK-Ada exist for Ada subsets used for critical software. But some emergent languages such as those of the ML family lack such adapted tools. Providing similar tools for ML languages requires special attention on specific features such as higher-order functions and pattern-matching. This work presents an information-flow analysis for such a language specifically designed according to the needs of assessors. This analysis is built as an abstract interpretation of the operational semantics enriched with dependency information. It is proved correct according to a formal definition of the notion of dependency using the Coq proof assistant. This work gives a strong theoretical basis for building an efficient tool for fault tolerance analysis.