WCET Squeezing: On-Demand Feasibility Refinement for Proven Precise WCET-Bounds

The Worst-Case Execution Time (WCET) computed by a WCET analyzer is usually not tight, leaving a gap between the actual and the computed WCET of a program. In this article we present a novel on-demand WCET feasibility refinement technique, called WCET Squeezing, for minimizing this gap.WCET Squeezing provides conceptually new means for addressing the classical problem of WCET computation, by deriving a WCET bound that comes as close as possible to the actual one. WCET Squeezing is an anytime algorithm, that is, it can be stopped at any time without violating the soundness of its results. This anytime property allows to apply WCET Squeezing not only for deriving precise WCET bounds but to also prove additional timing constraints over the program. Namely, WCET Squeezing can be used to guarantee that a program is fast enough by ensuring that the WCET of the program is below some required limit. If the initially computed WCET of the program is above this limit, WCET Squeezing can be stopped as soon as the squeezed WCET of the program is below the limit (proving the program meets the required timing constraint), or if the squeezed WCET is tight but above the given limit (proving the program cannot meet the timing constraint). WCET Squeezing can also be used until a given time budget is exhausted to compute a tight(er) WCET bound for a program. These new applications of WCET Squeezing are out of the scope of traditional WCET analyzers.WCET Squeezing combines symbolic program execution with the Implicit Path Enumeration Technique (IPET) for computing a precise WCET bound. WCET Squeezing is applicable as a post-process to any WCET analyzer which encodes the IPET problem as an Integer Linear Program (ILP). We implemented our method in the r-TuBound toolchain and evaluated our implementation on a set examples taken from the M\ue4lardalen WCET benchmark suite. Our experiments demonstrate that WCET Squeezing can significantly tighten the WCET bounds of programs. Moreover, the derived WCET bounds are proven to be precise at a moderate computational cost