Code clones are similar code fragments that often arise from copy-and-paste programming. Neural networks can classify pairs of code fragments as clone/not-clone with high accuracy. However, finding clones in industrial-scale code needs a more scalable approach than pairwise comparison. We extend existing neural network-based clone detection schemes to handle codebases that far exceed available memory, using indexing and search methods for external storage such as disks and solid-state drives. We generate a high-dimensional vector embedding for each code fragment using a transformer-based neural network. We then find similar embeddings using efficient multidimensional nearest neighbor search algorithms on external storage to find similar embeddings without pairwise comparison. We identify specific problems with industrial-scale code bases, such as large sets of almost identical code fragments that interact poorly with k-nearest neighbour search algorithms, and provide an effective solution. We demonstrate that our disk-based clone search approach achieves similar clone detection accuracy as an equivalent in-memory technique. Using a solid-state drive as external storage, our approach is around 2 x slower than the in-memory approach for a problem size that can fit within memory. We further demonstrate that our approach can scale to over a billion lines of code, providing valuable insights into the trade-offs between indexing speed, query performance, and storage efficiency for industrial-scale code clone detection.

Abstract Hidden code clones negatively impact software maintenance, but manually detecting them in large codebases is impractical. Additionally, automated approaches find detection of syntactically‐divergent clones very challenging. While recent deep neural networks (for example BERT‐based artificial neural networks) seem more effective in detecting such clones, their pairwise comparison of every code pair in the target system(s) is inefficient and scales poorly on large codebases. We present SSCD, a BERT‐based clone detection approach that targets high recall of Type 3 and Type 4 clones at a very large scale (in line with our industrial partner's requirements). It computes a representative embedding for each code fragment and finds similar fragments using a nearest neighbor search. Thus, SSCD avoids the pairwise‐comparison bottleneck of other neural network approaches, while also using a parallel, GPU‐accelerated search to tackle scalability. This article describes the approach, proposing and evaluating several refinements to improve Type 3/4 clone detection at scale. It provides a substantial empirical evaluation of the technique, including a speed/efficacy comparison of the approach against SourcererCC and Oreo, the only other neural‐network approach currently capable of scaling to hundreds of millions of LOC. It also includes a large in‐situ evaluation on our industrial collaborator's code base that assesses the original technique, the impact of the proposed refinements and illustrates the impact of incremental, active learning on its efficacy. We find that SSCD is significantly faster and more accurate than SourcererCC and Oreo. SAGA, a GPU‐accelerated traditional clone detection approach, is a little better than SSCD for T1/T2 clones, but substantially worse for T3/T4 clones. Thus, SSCD is both scalable to industrial code sizes, and comparatively more accurate than existing approaches for difficult T3/T4 clone searching. In‐situ evaluation on company datasets shows that SSCD outperforms the baseline approach (CCFinderX) for T3/T4 clones. Whitespace removal and active learning further improve SSCD effectiveness.

Code clones can detrimentally impact software maintenance and manually detecting them in very large code-bases is impractical. Additionally, automated approaches find detection of Type 3 and Type 4 (inexact) clones very challenging. While the most recent artificial deep neural networks (for ex-ample BERT-based artificial neural networks) seem to be highly effective in detecting such clones, their pairwise comparison of every code pair in the target system(s) is inefficient and scales poorly on large codebases.We therefore introduce SSCD, a BERT-based clone detection approach that targets high recall of Type 3 and Type 4 clones at scale (in line with our industrial partner’s requirements). It does so by computing a representative embedding for each code fragment and finding similar fragments using a nearest neighbour search. SSCD thus avoids the pairwise-comparison bottleneck of other Neural Network approaches while also using parallel, GPU-accelerated search to tackle scalability.This paper details the approach and an empirical assessment towards configuring and evaluating that approach in industrial setting. The configuration analysis suggests that shorter input lengths and text-only based neural network models demonstrate better efficiency in SSCD, while only slightly decreasing effectiveness. The evaluation results suggest that SSCD is more effective than state-of-the-art approaches like SAGA and SourcererCC. It is also highly efficient: in its optimal setting, SSCD effectively locates clones in the entire 320 million LOC BigCloneBench (a standard clone detection benchmark) in just under three hours.

Locating bugs is an important, but effort-intensive and time-consuming task, when dealing with large-scale systems. To address this, Information Retrieval (IR) techniques are increasingly being used to suggest potential buggy source code locations, for given bug reports. While IR techniques are very scalable, in practice their effectiveness in accurately localizing bugs in a software system remains low. Results of empirical studies suggest that the effectiveness of bug localization techniques can be augmented by the configuration of queries used to locate buggy code. However, in most IR-based bug localization techniques, presented by researchers, the impact of the queries’ configurations is not fully considered. In a similar vein, techniques consider all code elements as equally suspicious of being buggy while localizing bugs, but this is not always the case either.In this paper, we present a new method-level, information-retrieval-based bug localization technique called "BoostNSift". BoostNSift exploits the important information in queries by ‘boost’ing that information, and then ‘sift’s the identified code elements, based on a novel technique that emphasizes the code elements’ specific relatedness to a bug report over its generic relatedness to all bug reports. To evaluate the performance of BoostNSift, we employed a state-of-the-art empirical design that has been commonly used for evaluating file level IR-based bug localization techniques: 6851 bugs are selected from commonly used Eclipse, AspectJ, SWT, and ZXing benchmarks and made openly available for method-level analyses. The performance of BoostNSift is compared with the openly-available state-of-the-art IR-based BugLocator, BLUiR, and BLIA techniques. Experiments show that BoostNSift improves on BLUiR by up to 324%, on BugLocator by up to 297%, and on BLIA up to 120%, in terms of Mean Reciprocal Rank (MRR). Similar improvements are observed in terms of Mean Average Precision (MAP) and Top-N evaluation measures.