Evaluating the Design of a VLIW Processor for Real-Time Systems

Abstract

Nowadays, many real-time applications are very complex and as the complexity and the requirements of those systems become more demanding, more hardware processing capacity is necessary. Unfortunately, the correct functioning of real-time systems depends not only on the logically correct response but also on the time when it is produced. General-purpose processor design fails to deliver analyzability due to their nondeterministic behavior caused by the use of cache memories, dynamic branch prediction, speculative execution, and out-of-order pipelines. In this article, we investigate the pipeline performance of Very Long Instruction Word (VLIW) architectures for real-time systems with an in-order pipeline considering Worst-Case Execution Time (WCET) performance. Techniques on obtaining the WCET of VLIW machines are also considered and we make a quantification on how important are hardware techniques such as static branch prediction, predication, and pipeline speed of complex operations such as memory access and multiplication for high-performance real-time systems. The memory hierarchy is out of the scope of this article and we used a classic deterministic structure formed by a direct mapped instruction cache and a data scratchpad memory. A VLIW prototype was implemented in VHDL from scratch considering the HP VLIW ST231 ISA. We also show some compiler insights and we use a representative subset of the Mälardalen’s WCET benchmarks for validation and performance quantification.

Patmos: a time-predictable microprocessor

“…Starke, Carminati, and Oliveira present a VLIW processor for real-time systems [65,64]. The processor implements a four-issue pipeline, supports static branch prediction, and predication.…”