Treffer: Safe code transfromations for speculative execution in real-time systems

Although compiler optimization techniques are standard and successful in non-real-time systems, if naively applied, they can destroy safety guarantees and deadlines in hard real-time systems. For this reason, real-time systems developers have tended to avoid automatic compiler optimization of their code. However, real-time applications in several areas have been growing substantially in size and complexity in recent years. This size and complexity makes it impossible for real-time programmers to write optimal code, and consequently indicates a need for compiler optimization. Recently researchers have developed or modified analyses and transformations to improve performance without degrading worst-case execution times. Moreover, these optimization techniques can sometimes transform programs which may not meet constraints/deadlines, or which result in timeouts, into deadline-satisfying programs. One such technique, speculative execution, also used for example in parallel computing and databases, can enhance performance by executing parts of the code whose execution may or may not be needed. In some cases, rollback is necessary if the computation turns out to be invalid. However, speculative execution must be applied carefully to real-time systems so that the worst-case execution path is not extended. Deterministic worst-case execution for satisfying hard real-time constraints, and speculative execution with rollback for improving average-case throughput, appear to lie on opposite ends of a spectrum of performance requirements and strategies. Deterministic worst-case execution for satisfying hard real-time constraints, and speculative execution with rollback for improving average-case throughput, appear to lie on opposite ends of a spectrum of performance requirements and strategies. Nonetheless, this thesis shows that there are situations in which speculative execution can improve the performance of a hard real-time system, either by enhancing average performance while not affecting the worst-case, or by actually ...