Error flow in computer programs

White box program analysis has been applied to program testing for some time, but this analysis is primarily grounded in program syntax, while errors arise from incorrect program semantics. We introduce a semantically-based technique called error flow analysis, which is used to investigate the behavior of a program at the level of data state transitions. Error flow analysis is based on a model of program execution as a composition of functions that each map a prior data state into a subsequent data state. According to the fault/failure model, failure occurs when a fault causes an infection in the data state which then propagates to output. A faulty program may also produce coincidentally correct output for a given input if the fault resists infection, or an infection is cancelled by subsequent computation. We investigate this phenomenon using dynamic error flow analysis to track the infection and propagation of errors in the data states of programs with seeded faults. This information is gathered for a particular fault over many inputs on a path-by-path basis to estimate execution, infection, and failure rates as well as characteristics of error flow behavior for the fault. Those paths that exhibit high failure rates would be more desirable to test for this fault than those with lower failure rates, and we look for error flow characteristics that correlate with high failure rate. We present the results of dynamic error flow experiments on several programs, and suggest ways in which error flow information can be used in program analysis and testing