Mapping Modern JVM Language Code to Analysis-Friendly Graphs: A Study with Kotlin

Abstract

Kotlin is a modern JVM language, gaining adoption rapidly and becoming Android official programming language. With its wide usage, the need for code analysis of Kotlin is increasing. Exposing code semantics explicitly with a properly structured format is the first step in code analysis and the construction of such representation is the foundation for downstream tasks. Recently, graph-based approaches became a promising way of encoding source code semantics. However, this work mainly focuses on representation learning with limited interpretability and shallow domain knowledge. The known evolvements of code semantics in new-generation programming languages have been overlooked. How to establish an effective mapping between naturally concise Kotlin source code with graph-based representation needs to be studied by analyzing known language features. Moreover, the feasibility of enhancing the mapping with code semantics automatically learned from the program needs to be explored. In this paper, we first propose a first-sight, rule-based mapping method, using composite representation with AST, CFG, DFG and language features. To examine the possibility of exposing code semantics in the mapped graph, we use Latent Semantic Indexing-based source code summarization to learn more features of each method, and then enrich the attributes of the corresponding node in the graph. We evaluate these mapping strategies with comparative experiments by simulating a code search solution as a downstream task. The experiment result shows that the graph-based method with built-in language features outperforms the text-based way without introducing greater complexity. Comparative experiments also prove that adding code semantics to the graph benefits the capacity of downstream tasks. When exploring the whole mapping process, our study explicitly revealed the practical barriers to extracting and exposing the hidden semantics from Kotlin source code, which may help enlighten source code representations for other modern languages.