Controlling speculative execution through a virtually ordered memory system

Abstract

Processors which extract parallelism through speculative execution must be able to identify when mis-speculation has occurred. The three places where mis-speculation can occur are register accesses, control flow prediction and memory accesses. Controlling register and control flow speculation has been well studied, but no scalable techniques for identifying memory dependence violations have been identified. Since speculative execution occurs out of order this requires tracking the causal order, as well as the addresses of memory accesses.

This thesis uses simulations to investigate tracking the causal order of memory accesses using explicit tags known as virtual timestamps, a distributed and scalable method. Realizable virtual timestamps are necessarily restricted in length and it is demonstrated that naive allocation schemes seriously constrain execution by inefficiently allocating virtual timestamps. Efficiently allocating virtual timestamps requires analysis of the number required by each section of code. Basic statically and dynamically evaluated analysis methods are established to avoid virtual timestamp allocation becoming a resource bottleneck.

The same analysis is also used to efficiently allocate state saving resources in a fixed hardware order. The hardware order provides an alternative way of maintaining the causal order using a simple hardware organization. The ability to predict the resources required by regions of code is used as a way of selecting instructions to execute speculatively. This enables resources to be allocated efficiently and is shown to allow large amounts of parallelism to be extracted. It also promotes the effectiveness of speculative execution by issuing less instructions that will ultimately be rolled back.

Using a hierarchy of hardware ordering modules, themselves ordered by explicit virtual timestamps, a scalable ordering system is proposed. This hierarchy forms the basis of a twisted memory system, a multiple version memory system capable of identifying speculative memory dependence violations. The preliminary investigations presented here show that twisted memory has the potential to support aggressive speculative parallel execution. Particular attention is paid to memory bandwidth requirements.