Improving the First-Miss Computation in Set-Associative Instruction Caches

Abstract

The methods for worst case execution time (WCET) computation need to analyse both the control flow of the task, and the architecture effects involved by the hosting architecture. An important architectural effect that needs to be predicted is the instruction cache behavior. This prediction is commonly performed by assigning to each program instruction a category that describes its behavior. One of these categories, first miss, means that the first reference is a cache miss, while the subsequent references give hits. Yet, there is variations in the meanings of this category according to the used methods, capturing overlapping but not equivalents sets of cache behaviors. In this paper, we have analysed the shortcomings of the First-Miss computation methods, and we have deduced an improved first miss computation approach which captures a maximum of cache behaviors while eliminating some of the most time-consuming processing. We have implemented our method in the frame of C. Ferdinand's categorization method, enhancing his approach for first miss handling, and compared it with the non-enhanced versions. The results shows a tighter WCET, and a greatly reduced computation time.