Model Checking of Software Systems under Weak Memory Models

When a program is compiled and run on a modern architecture, different optimizations may be applied to gain in efficiency. In particular, the access operations (e.g., read and write) to the shared memory may be performed in an out-of-order manner, i.e., in a different order than the order in which the operations have been issued by the program. The reordering of memory access operations leads to efficient use of instruction pipelines and thus an improvement in program execution times. However, the gain in this efficiency comes at a price. More precisely, programs running under modern architectures may exhibit unexpected behaviors by programmers. The out-of-order execution has led to the invention of new program semantics, called weak memory model (WMM). One crucial problem is to ensure the correctness of concurrent programs running under weak memory models. The thesis proposes three techniques for reasoning and analyzing concurrent programs running under WMMs. The first one is a sound and complete analysis technique for finite-state programs running under the TSO semantics (Paper II). This technique is based on a novel and equivalent semantics for TSO, called Dual TSO semantics, and on the use of well-structured transition framework. The second technique is an under-approximation technique that can be used to detect bugs under the POWER semantics (Paper III). This technique is based on bounding the number of contexts in an explored execution where, in each context, there is only one active process. The third technique is also an under-approximation technique based on systematic testing (a.k.a. stateless model checking). This approach has been used to develop an optimal and efficient systematic testing approach for concurrent programs running under the Release-Acquire semantics (Paper IV). The thesis also considers the problem of effectively finding a minimal set of fences that guarantees the correctness of a concurrent program running under WMMs (Paper I). A fence (a.k.a. barrier) is an operation that can be inserted in the program to prohibit certain reorderings between operations issued before and after the fence. Since fences are expensive, it is crucial to automatically find a minimal set of fences to ensure the program correctness. This thesis presents a method for automatic fence insertion in programs running under the TSO semantics that offers the best-known trade-off between the efficiency and optimality of the algorithm. The technique is based on a novel notion of correctness, called Persistence, that compares the behaviors of a program running under WMMs to that running under the SC semantics.UPMAR