Optimizing Interrupt-Driven Embedded Software

Software for embedded microcontroller units (MCUs) represents both an interesting opportunity and a difficult challenge for compiler optimization. Since these systems tend to be small—often limited to a few KB of on-chip RAM—highly aggressive techniques are feasible and worthwhile. On the other hand, the effectiveness of traditional dataflow analyses is limited by their inability to cope with interrupt-driven concurrency and the direct interaction of embedded software with hardware devices. We present an integrated collection of static analysis techniques that leads to effective, wholeprogram optimization of C code running on MCUs. Our first main contribution is a technique supporting whole-program dataflow analysis in the presence of interrupt-driven concurrency. This model is based on an automatic classification of each data element in a system based on its role in concurrent execution, followed by an analysis that efficiently approximates the effects of preemptions by interrupt handlers. Our second contribution is a technique for flowing data through volatile-qualified objects in MCU software, which can be backed by either RAM or device registers. We implemented our dataflow framework in a tool called cXprop. As a standalone optimizer, cXprop can reduce code size of sensor network programs based on TinyOS by 12%, while reducing their CPU usage by 8.3%. We have also used cXprop as a crucial enabler for two other projects. First, we used its analysis to drive offline RAM compression, reducing the RAM requirements of TinyOS applications by an average of 22%. Second, we used cXprop’s analysis of concurrency to greatly increase the efficiency of Safe TinyOS, our project to provide memory safety to sensor net applications.