Worst-Case Execution Time Analysis for Optimized Code

In the eld of real-time systems, accurate estimates of the worst-case execution time of programs are required for real-time modelling and veri cation, scheduling analysis, feasibility analysis, and dimensioning of the system hardware. At present, such estimates are produced by measuring program runs on inputs which are believed to produce long execution times. Unfortunately, such measurements provide no guarantee for nding the worst execution time, thereby producing unsafe systems. To obtain safe estimates (estimates which do not underestimate the worst-case time), we have to use formal static analysis. Because the market demands high performance, low cost products, we must use optimizing compilers when developing software for real-time systems. An optimizing compiler improves the performance of a program and reduces the size of the code, both important factors to save costs while maintaining performance. An optimizing compiler makes static analysis of a program more di cult by complicating the relationship between the program source code (where we can obtain knowledge about possible program executions) and the object code (where we can nd concrete execution times). In this thesis, we present (1) a new framework for static timing analysis of optimized programs and (2) co-transformation, a new approach to keep track of timing information while performing the various transforming and optimizing steps in a compiler. The co-transformer is based on a close cooperation between compiler and timing analyzer. We have implemented a prototype co-transformer, and evaluated it using a representative set of compiler optimizations. We show that the idea is feasible, but that much more work is needed to design a datastructure powerful enough to capture all necessary information about program execution. In addition, the thesis points out some directions for future research in the eld of worst-case execution time analysis in particular and the real-time programming eld in general.