Abstract

A dynamic software update is an update to program code that is applied at runtime. While the mechanics of altering the image of a running program are straightforward, at least on a von Neumann machine, what constitutes a valid update and when it may be applied is up for debate. At the same time, anecdotal evidence suggests that dynamic updates coexist poorly with program optimization. This is hardly surprising—-a dynamic update makes implicit assumptions about the intermediate state of the program at the time it is applied, which in turn limit the compiler’s scope for optimization. The greater the dependence of an update on the original program, the more disruptive it is. I will describe an approach to optimizing dynamically updateable programs that involves restricting the kinds of updates that may be applied at each update point. Restrictions that are too onerous can be relaxed with dynamic deoptimization, a technique due to Hölzle et al. for source-level debugging of optimized programs. This will be a half-hour talk.