Using Control Flow Analysis for Space and Time Efficient Stub Generation

A stub generator is a tool that automatically produces the communication software required for presentation conversion. An open problem in stub generation is finding the right trade-off between code size and execution speed of the generated code. The generation of tables that are interpreted at run-time leads to small but slow object code. The direct generation of program code leads to fast but big object code. This trade-off can be solved by using a mixture of the two code generation techniques that uses direct generation of program code only for types of an interface specification that are expected to be used frequently at run-time. However, this approach requires that the frequency distribution of type usage at run-time exhibits sufficient locality. In this paper, we give experimental evidence that this is indeed the case. A second experiment indicates that relatively simple heuristics exist to predict the most frequently used types by compile-time analysis of the interface specification. The experiment was conducted by correlating a trace of X.400 e-mail messages with the X.400 interface specification. We found that by using a very simple heuristic we could correctly predict 8 of the 10 most frequently used types in the trace. This prediction is a prerequisite for implementing stub generators that produce space and time efficient conversion routines.