Efficient machine-code test-program induction

Abstract

Technology advances allow integrating an entire system on a single chip, including memories and peripherals. The testing of these devices is becoming a major issue for chip manufacturing industries. This paper presents a methodology, similar to genetic programming, for inducing test programs. However, it includes the ability to explicitly specify registers and resorts to directed acyclic graphs instead of trees. Moreover, it exploits a database containing the assembly-level semantics associated with each graph node. This approach is extremely efficient and versatile: candidate solutions are translated into source-code programs allowing millions of evaluations per second. The proposed approach is extremely versatile: the macro library allows the target processor and the environment to be changed easily. The approach was verified on three processors with different instruction sets, different formalisms and different conventions. A complete set of experiments on a test function is also reported for the SPARC processor.