Higher Abstractions for Dynamic Analysis

The developers of tools for dynamic analysis are faced with choosing from the many approaches to gathering runtime data. Typically, dynamic analysis involves instrumenting the program under investigation to record its runtime behavior. Current approaches for byte-code based systems like Java and Smalltalk rely often on inserting byte-code into the program under analysis. However, detailed knowledge of the target programming language or virtual machine is required to implement dynamic analysis tools. Obtaining and exploiting this knowledge to build better analysis tools is cumbersome and often distracts the tool builder from the actual goal, which is the analysis of the runtime behavior of a system. In this paper, we argue that we need to adopt a higher level view of a software system when considering the task of abstracting runtime information. We focus on objectoriented virtual machine based languages. We want to be able to deal with the runtime system as a collection of reified first-class entities. We propose to achieve this by introducing a layer of abstraction, i.e., a behavioral middle layer. This has the advantage that the task of collecting dynamic information is not concerned with low level details of a specific language or virtual machine. The positive eect of such a behavioral middle layer is twofold: on the one hand it provides us with a standard API for all dynamic analysis based tools to use, on the other hand it allows the tool developer to abstract from the actual implementation technique.