Compiling for time-predictability and performance

Abstract

Single-path is a code-generation technique to disconnect a program’s execution time from run-time conditions. Full time-predictability means programs have a constant execution time, which, when coupled with sufficient performance, can replace static worst-case execution time analysis in the design process for a real-time system. This cumulative dissertation collates six works improving the time-predictability and performance of single- path code. Constant execution times are achieved by compensating for varying access times to main memory. Performance is improved by optimizing code that is executed a fixed number of times, by more efficient use of dual-issue processor pipelines, and by enabling the use of generic register allocators for allocating predicate registers. This dissertation also discusses how the challenges of correctness and practicality require this type of research to use the scientific method to advance the state of the art. The selected papers are put into context before evaluating the end-to-end cumulative performance impacts. In addition to achieving constant execution times on a time-predictable processor, the results show varying but significant improvements of up to 145% in performance and a reduced code size of up to 28%.