Large Language Models (LLMs), such as transformer-based neural networks trained on billions of parameters, have become increasingly prevalent in software engineering (SE). These models, trained on extensive datasets that include code repositories, exhibit remarkable capabilities for SE tasks. However, evaluating their effectiveness poses significant challenges, primarily due to the potential overlap between the datasets used for training and those employed for evaluation. To address this issue, we introduce SnipGen, a comprehensive repository mining framework designed to leverage prompt engineering across various downstream tasks for code generation. SnipGen aims to mitigate data contamination by generating robust testbeds and crafting tailored data points to assist researchers and practitioners in evaluating LLMs for code-related tasks. In our exploratory study, SnipGen mined approximately 227K data points from 338K recent code changes in GitHub commits, focusing on method-level granularity. SnipGen features a collection of prompt templates that can be combined to create a Chain-of-Thought-like sequence of prompts, enabling a nuanced assessment of LLMs’ code generation quality. By providing the mining tool, the methodology, and the dataset, SnipGen empowers researchers and practitioners to rigorously evaluate and interpret LLMs’ performance in software engineering contexts.

Code completion is one of the main features of modern Integrated Development Environments (IDEs). Its objective is to speed up code writing by predicting the next code token(s) the developer is likely to write. Research in this area has substantially bolstered the predictive performance of these techniques. However, the support to developers is still limited to the prediction of the next few tokens to type. In this work, we take a step further in this direction by presenting a large-scale empirical study aimed at exploring the capabilities of state-of-the-art deep learning (DL) models in supporting code completion at different granularity levels, including single tokens, one or multiple entire statements, up to entire code blocks (e.g., the iterated block of a for loop). To this aim, we train and test several adapted variants of the recently proposed RoBERTa model, and evaluate its predictions from several perspectives, including: (i) metrics usually adopted when assessing DL generative models (i.e., BLEU score and Levenshtein distance); (ii) the percentage of perfect predictions (i.e., the predicted code snippets that match those written by developers); and (iii) the "semantic" equivalence of the generated code as compared to the one written by developers. The achieved results show that BERT models represent a viable solution for code completion, with perfect predictions ranging from ~7%, obtained when asking the model to guess entire blocks, up to ~58%, reached in the simpler scenario of few tokens masked from the same code statement.

Assessing the similarity between code components plays a pivotal role in a number of Software Engineering (SE) tasks, such as clone detection, impact analysis, refactoring, etc. Code similarity is generally measured by relying on manually defined or hand-crafted features, e.g., by analyzing the overlap among identifiers or comparing the Abstract Syntax Trees of two code components. These features represent a best guess at what SE researchers can utilize to exploit and reliably assess code similarity for a given task. Recent work has shown, when using a stream of identifiers to represent the code, that Deep Learning (DL) can effectively replace manual feature engineering for the task of clone detection. However, source code can be represented at different levels of abstraction: identifiers, Abstract Syntax Trees, Control Flow Graphs, and Bytecode. We conjecture that each code representation can provide a different, yet orthogonal view of the same code fragment, thus, enabling a more reliable detection of similarities in code. In this paper, we demonstrate how SE tasks can benefit from a DL-based approach, which can automatically learn code similarities from different representations.

During software evolution, the source code of a system frequently changes due to bug fixes or new feature requests. Some of these changes may accidentally degrade performance of a newly released software version. A notable problem of regression testing is how to find problematic changes (out of a large number of committed changes) that may be responsible for performance regressions under certain test inputs.

Code smells are symptoms of poor design and implementation choices that may hinder code comprehension and possibly increase change- and fault-proneness of source code. Several techniques have been proposed in the literature for detecting code smells. These techniques are generally evaluated by comparing their accuracy on a set of detected candidate code smells against a manually-produced oracle. Unfortunately, such comprehensive sets of annotated code smells are not available in the literature with only few exceptions. In this paper we contribute (i) a dataset of 243 instances of five types of code smells identified from 20 open source software projects, (ii) a systematic procedure for validating code smell datasets, (iii) LANDFILL, a Web-based platform for sharing code smell datasets, and (iv) a set of APIs for programmatically accessing LANDFILL's contents. Anyone can contribute to Landfill by (i) improving existing datasets (e.g., Adding missing instances of code smells, flagging possibly incorrectly classified instances), and (ii) sharing and posting new datasets. Landfill is available at www.sesa.unisa.it/landfill/, while the video demonstrating its features in action is available at http://www.sesa.unisa.it/tools/landfill.jsp.

GUI-based models extracted from Android app execution traces, events, or source code can be extremely useful for challenging tasks such as the generation of scenarios or test cases. However, extracting effective models can be an expensive process. Moreover, existing approaches for automatically deriving GUI-based models are not able to generate scenarios that include events which were not observed in execution (nor event) traces. In this paper, we address these and other major challenges in our novel hybrid approach, coined as MONKEYLAB. Our approach is based on the Record→Mine→Generate→Validate framework, which relies on recording app usages that yield execution (event) traces, mining those event traces and generating execution scenarios using statistical language modeling, static and dynamic analyses, and validating the resulting scenarios using an interactive execution of the app on a real device. The framework aims at mining models capable of generating feasible and fully replayable (i.e., Actionable) scenarios reflecting either natural user behavior or uncommon usages (e.g., Corner cases) for a given app. We evaluated MONKEYLAB in a case study involving several medium-to-large open-source Android apps. Our results demonstrate that MONKEYLAB is able to mine GUI-based models that can be used to generate actionable execution scenarios for both natural and unnatural sequences of events on Google Nexus 7 tablets.

Deep learning subsumes algorithms that automatically learn compositional representations. The ability of these models to generalize well has ushered in tremendous advances in many fields such as natural language processing (NLP). Recent research in the software engineering (SE) community has demonstrated the usefulness of applying NLP techniques to software corpora. Hence, we motivate deep learning for software language modeling, highlighting fundamental differences between state-of-the-practice software language models and connectionist models. Our deep learning models are applicable to source code files (since they only require lexically analyzed source code written in any programming language) and other types of artifacts. We show how a particular deep learning model can remember its state to effectively model sequential data, e.g., Streaming software tokens, and the state is shown to be much more expressive than discrete tokens in a prefix. Then we instantiate deep learning models and show that deep learning induces high-quality models compared to n-grams and cache-based n-grams on a corpus of Java projects. We experiment with two of the models' hyper parameters, which govern their capacity and the amount of context they use to inform predictions, before building several committees of software language models to aid generalization. Then we apply the deep learning models to code suggestion and demonstrate their effectiveness at a real SE task compared to state-of-the-practice models. Finally, we propose avenues for future work, where deep learning can be brought to bear to support model-based testing, improve software lexicons, and conceptualize software artifacts. Thus, our work serves as the first step toward deep learning software repositories.

Energy consumption of mobile applications is nowadays a hot topic, given the widespread use of mobile devices. The high demand for features and improved user experience, given the available powerful hardware, tend to increase the apps’ energy consumption. However, excessive energy consumption in mobile apps could also be a consequence of energy greedy hardware, bad programming practices, or particular API usage patterns. We present the largest to date quantitative and qualitative empirical investigation into the categories of API calls and usage patterns that—in the context of the Android development framework—exhibit particularly high energy consumption profiles. By using a hardware power monitor, we measure energy consumption of method calls when executing typical usage scenarios in 55 mobile apps from different domains. Based on the collected data, we mine and analyze energy-greedy APIs and usage patterns. We zoom in and discuss the cases where either the anomalous energy consumption is unavoidable or where it is due to suboptimal usage or choice of APIs. Finally, we synthesize our findings into actionable knowledge and recipes for developers on how to reduce energy consumption while using certain categories of Android APIs and patterns

In the recent years, studies of design and programming practices in mobile development are gaining more attention from researchers. Several such empirical studies used Android applications (paid, free, and open source) to analyze factors such as size, quality, dependencies, reuse, and cloning. Most of the studies use executable files of the apps (APK files), instead of source code because of availability issues (most of free apps available at the Android official market are not open-source, but still can be downloaded and analyzed in APK format). However, using only APK files in empirical studies comes with some threats to the validity of the results. In this paper, we analyze some of these pertinent threats. In particular, we analyzed the impact of third-party libraries and code obfuscation practices on estimating the amount of reuse by class cloning in Android apps. When including and excluding third-party libraries from the analysis, we found statistically significant differences in the amount of class cloning 24,379 free Android apps. Also, we found some evidence that obfuscation is responsible for increasing a number of false positives when detecting class clones. Finally, based on our findings, we provide a list of actionable guidelines for mining and analyzing large repositories of Android applications and minimizing these threats to validity

Approaches that support software maintenance need to be evaluated and compared against existing ones, in order to demonstrate their usefulness in practice. However, oftentimes the lack of well-established sets of benchmarks leads to situations where these approaches are evaluated using different datasets, which results in biased comparisons. In this data paper we describe and make publicly available a set of benchmarks from six Java applications, which can be used in the evaluation of various software engineering (SE) tasks, such as feature location and impact analysis. These datasets consist of textual description of change requests, the locations in the source code where they were implemented, and execution traces. Four of the benchmarks were already used in several SE research papers, and two of them are new. In addition, we describe in detail the methodology used for generating these benchmarks and provide a suite of tools in order to encourage other researchers to validate our datasets and generate new benchmarks for other subject software systems. Our online appendix: http://www.cs.wm.edu/semeru/data/msr13/.

Question & answer (Q&A) websites, such as Stack Overflow (SO), are widely used by developers to find and provide answers to technical issues and concerns in software development. Mobile development is not an exception to the rule. In the latest SO dump, more than 400K questions were labeled with tags related to mobile technologies. Although, previous works have analyzed the main topics and trends in SO threads, there are no studies devoted specifically to mobile development. In this paper we used topic modeling techniques to extract hot-topics from mobile-development related questions. Our findings suggest that most of the questions include topics related to general questions and compatibility issues, and the most specific topics, such as crash reports and database connection, are present in a reduced set of questions.

In this paper, we propose a technique called LACT for automatically categorizing software systems in open-source repositories. LACT is based on latent Dirichlet Allocation, an information retrieval method which is used to index and analyze source code documents as mixtures of probabilistic topics. For an initial evaluation, we performed two studies. In the first study, LACT was compared against an existing tool, MUDABlue, for classifying 41 software systems written in C into problem domain categories. The results indicate that LACT can automatically produce meaningful category names and yield classification results comparable to MUDABlue. In the second study, we applied LACT to 43 software systems written in different programming languages such as C/C++, Java, C#, PHP, and Perl. The results indicate that LACT can be used effectively for the automatic categorization of software systems regardless of the underlying programming language or paradigm. Moreover, both studies indicate that LACT can identify several new categories that are based on libraries, architectures, or programming languages, which is a promising improvement as compared to manual categorization and existing techniques.