Static analyses over weak memory

Writing concurrent programs with shared memory is often not trivial. Correctly synchronising the threads and handling the non-determinism of executions require a good understanding of the interleaving semantics. Yet, interleavings are not sufficient to model correctly the executions of modern, multicore processors. These executions follow rules that are weaker than those observed by the interleavings, often leading to reorderings in the sequence of updates and readings from memory; the executions are subject to a weaker memory consistency. Reorderings can produce executions that would not be observable with interleavings, and these possible executions also depend on the architecture that the processors implement. It is therefore necessary to locate and understand these reorderings in the context of a program running, or to prevent them in an automated way. In this dissertation, we aim to automate the reasoning behind weak memory consistency and perform transformations over the code so that developers need not to consider all the specifics of the processors when writing concurrent programs. We claim that we can do automatic static analysis for axiomatically-defined weak memory models. The method that we designed also allows re-use of automated verification tools like model checkers or abstract interpreters that were not designed for weak memory consistency, by modification of the input programs. We define an abstraction in detail that allows us to reason statically about weak memory models over programs. We locate the parts of the code where the semantics could be affected by the weak memory consistency. We then provide a method to explicitly reveal the resulting reorderings so that usual verification techniques can handle the program semantics under a weaker memory consistency. We finally provide a technique that synthesises synchronisations so that the program would behave as if only interleavings were allowed. We finally test these approaches on artificial and real software. We justify our choice of an axiomatic model with the scalability of the approach and the runtime performance of the programs modified by our method.