Improving Syntactical Clone Detection Methods through the Use of an Intermediate Representation

Abstract

Detection of type-3 and type-4 clones remains a difficult task. Current methods are complex, both on a conceptual and computational level. Similarly, their usage requires substantial implementation efforts. Instead of creating yet another method, it might be more productive to combine the simplicity of syntactic approaches with the abstractions granted by intermediate representations (IR). To this end, we devised a c-like IR based on LLVM and ran NiCad on it (LLNiCad). To establish whether the clone detection capabilities of syntactic approaches can be improved through an IR, we compared NiCad and LLNiCad on three open source projects taken from Krutz's benchmark and a subset of Google code jam solutions. In our results, the f1-score of LLNiCad consistently outperforms NiCad. Indeed, for all clone types in Krutz's benchmark, LLNiCad has a f1-score that is 37% higher than NiCad; with both better precision and recall. For type-4 clones in our GCJ benchmark, the f1-score of LLNiCad also outperforms CCCD (a semantic clone detector) by 44%. These findings suggest that IRs are beneficial for improving clone detection and that they have a larger impact on type-3 and type-4 clones.