Compiling Multi-Threaded Object-Oriented Programs

A compiler must take special care when translating and transforming a multi-threaded program: The presence of synchronization operations and the potential of access conflicts have implications on register allocation and instruction scheduling. Moreover, a compiler might want to inhibit hardware assisted instruction reordering through memory fences, to enable a strong memory model on a platform with weakly ordered memory semantics. Unless the language or user asserts that an input program is free from ’access conflicts’, the compiler must be aware of the concurrency and the data sharing in a program, to guarantee that the output program complies with a certain (well-defined) program semantics. This paper reports on an approach to compile multithreaded object-oriented programs sharing data on a global heap. We developed a whole-program analysis that determines compile-time abstractions for threads and objects. Beyond escape information, the analysis determines the use and update actions of threads to objects and their approximated happens-before ordering. This information is used to implement two variants of sparse instrumentation, addressing the problem of access conflict detection in concurrent programs: (1) Object Race Detection checks if accesses to objects follow a locking discipline at runtime. (2) Object Consistency verifies if threads behave such that accesses to objects happen in a serializable order. Overlapping object accesses, which entail the possibility of data corruption, are detected. Supported, in part, by a grant from Intel Corp., Microprocessor Research Lab (MRL).