Zhongxing Yu

Zhen Yang, Fang Liu, Zhongxing Yu et al.

Code translation tools (transpilers) are developed for automatic source-to-source translation. Although learning-based transpilers have shown impressive enhancement against rule-based counterparts, owing to their task-specific pre-training on extensive monolingual corpora. Their current performance still remains unsatisfactory for practical deployment, and the associated training resources are also prohibitively expensive. LLMs pre-trained on huge amounts of human-written code/text have shown remarkable performance in many code intelligence tasks due to their powerful generality, even without task-specific training. Thus, LLMs can potentially circumvent the above limitations, but they have not been exhaustively explored yet. This paper investigates diverse LLMs and learning-based transpilers for automated code translation tasks, finding that: although certain LLMs have outperformed current transpilers, they still have some accuracy issues, where most of the failures are induced by a lack of comprehension of source programs, missing clear instructions on I/O types in translation, and ignoring discrepancies between source and target programs. Enlightened by the above findings, we further propose UniTrans, a Unified code Translation framework, applicable to various LLMs, for unleashing their power in this field. Specifically, UniTrans first crafts a series of test cases for target programs with the assistance of source programs. Next, it harnesses the above auto-generated test cases to augment the code translation and then evaluate their correctness via execution. Afterward, UniTrans further (iteratively) repairs incorrectly translated programs prompted by test case execution results. Extensive experiments are conducted on six settings of translation datasets between Python, Java, and C++. Three recent LLMs of diverse sizes are tested with UniTrans, and all achieve substantial improvements.

Simon Urli, Zhongxing Yu, Lionel Seinturier et al.

Program repair research has made tremendous progress over the last few years, and software development bots are now being invented to help developers gain productivity. In this paper, we investigate the concept of a " program repair bot " and present Repairnator. The Repairnator bot is an autonomous agent that constantly monitors test failures, reproduces bugs, and runs program repair tools against each reproduced bug. If a patch is found, Repairnator bot reports it to the developers. At the time of writing, Repairnator uses three different program repair systems and has been operating since February 2017. In total, it has studied 11 317 test failures over 1 609 open-source software projects hosted on GitHub, and has generated patches for 17 different bugs. Over months, we hit a number of hard technical challenges and had to make various design and engineering decisions. This gives us a unique experience in this area. In this paper, we reflect upon Repairnator in order to share this knowledge with the automatic program repair community.

Zhongxing Yu, Chenggang Bai, Lionel Seinturier et al.

Annotations have been formally introduced into Java since Java 5. Since then, annotations have been widely used by the Java community for different purposes, such as compiler guidance and runtime processing. Despite the ever-growing use, there is still limited empirical knowledge about the actual usage of annotations in practice, the changes made to annotations during software evolution, and the potential impact of annotations on code quality. To fill this gap, we perform the first large-scale empirical study about Java annotations on 1,094 notable open-source projects hosted on GitHub. Our study systematically investigates annotation usage, annotation evolution, and annotation impact, and generates 10 novel and important findings. We also present the implications of our findings, which shed light for developers, researchers, tool builders, and language or library designers in order to improve all facets of Java annotation engineering.

Thomas Durieux, Youssef Hamadi, Zhongxing Yu et al.

High-availability of software systems requires automated handling of crashes in presence of errors. Failure-oblivious computing is one technique that aims to achieve high availability. We note that failure-obliviousness has not been studied in depth yet, and there is very few study that helps understand why failure-oblivious techniques work. In order to make failure-oblivious computing to have an impact in practice, we need to deeply understand failure-oblivious behaviors in software. In this paper, we study, design and perform an experiment that analyzes the size and the diversity of the failure-oblivious behaviors. Our experiment consists of exhaustively computing the search space of 16 field failures of large-scale open-source Java software. The outcome of this experiment is a much better understanding of what really happens when failure-oblivious computing is used, and this opens new promising research directions.

Zhen Yang, Hongyi Lin, Yifan He et al.

In recent years, code intelligence has gained increasing importance in the field of automated software engineering. Meanwhile, the widespread adoption of Pretrained Language Models (PLMs) and Large Language Models (LLMs) has raised concerns regarding data contamination and its potential impact on model performance evaluation. This paper presents a systematic empirical study to investigate the fine-grained data contamination on code intelligence tasks. Our study involves diverse representative PLMs, namely RoBERTa and GPT-2, and LLMs, namely LLaMA and StarCoder, covering three major tasks: code translation, code generation, and code summarization. We categorize contamination scenarios into four types according to the code intelligence practice, namely input-only, output-only, unpaired, and paired contamination settings, and construct corresponding experimental and control groups for exploration. Experimental results show that, under the pre-training, fine-tuning, and inference paradigm adopted by PLMs, even deliberately injecting paired contamination does not lead to significant performance overestimation. But direct inference or small-scale fine-tuning uncovers the contamination effects. In contrast, LLMs with pre-training and inference paradigm are significantly affected by the paired contamination. Apart from the above, other contamination scenarios have no impact on both PLMs and LLMs. Our findings challenge the conventional belief that contamination inevitably leads to performance overestimation, providing new insights into the evaluation and deployment of code intelligence models.

Zhenyu Yang, Zhen Yang, Zhongxing Yu

With the advancement of deep learning techniques, the performance of Automatic Program Repair(APR) techniques has reached a new level. Previous deep learning-based APR techniques essentially modified program sentences in the Autoregressive(AR) manner, which predicts future values based on past values. Due to the manner of word-by-word generation, the AR-based APR technique has a huge time delay. This negative consequence overshadows the widespread adoption of APR techniques in real-life software development. To address the issue, we aim to apply the Non-Autoregressive(NAR) method to the APR task, which can output target code in a parallel manner to avoid huge inference delays. To effectively adapt the NAR manner for the APR task, we in this paper propose NARRepair, the first customized NAR code generation model for the APR task. The NARRepair features three major novelties, including 1) using repair actions to alleviate the over-correction issue, 2) extracting dependency information from AST to alleviate the issue of lacking inter-word dependency information, 3) employing two-stage decoding to alleviate the issue of lacking contextual information. We evaluated NARRepair on three widely used datasets in the APR community, and the results show that our technique can significantly improve the inference speed while maintaining high repair accuracy.

Martin Monperrus, Matias Martinez, He Ye et al.

This paper presents Megadiff, a dataset of source code diffs. It focuses on Java, with strict inclusion criteria based on commit message and diff size. Megadiff contains 663 029 Java diffs that can be used for research on commit comprehension, fault localization, automated program repair, and machine learning on code changes.

Benoit Baudry, Zimin Chen, Khashayar Etemadi et al.

Software bugs are common and correcting them accounts for a significant part of costs in the software development and maintenance process. This calls for automatic techniques to deal with them. One promising direction towards this goal is gaining repair knowledge from historical bug fixing examples. Retrieving insights from software development history is particularly appealing with the constant progress of machine learning paradigms and skyrocketing `big' bug fixing data generated through Continuous Integration (CI). In this paper, we present R-Hero, a novel software repair bot that applies continual learning to acquire bug fixing strategies from continuous streams of source code changes, implemented for the single development platform Github/Travis CI. We describe R-Hero, our novel system for learning how to fix bugs based on continual training, and we uncover initial successes as well as novel research challenges for the community.

Zhongxing Yu, Matias Martinez, Zimin Chen et al.

To effectively guide the exploration of the code transform space for automated code evolution techniques, we present in this paper the first approach for structurally predicting code transforms at the level of AST nodes using conditional random fields (CRFs). Our approach first learns offline a probabilistic model that captures how certain code transforms are applied to certain AST nodes, and then uses the learned model to predict transforms for arbitrary new, unseen code snippets. {Our approach involves a novel representation of both programs and code transforms. Specifically, we introduce the formal framework for defining the so-called AST-level code transforms and we demonstrate how the CRF model can be accordingly designed, learned, and used for prediction}. We instantiate our approach in the context of repair transform prediction for Java programs. Our instantiation contains a set of carefully designed code features, deals with the training data imbalance issue, and comprises transform constraints that are specific to code. We conduct a large-scale experimental evaluation based on a dataset of bug fixing commits from real-world Java projects. The results show that when the popular evaluation metric \emph{top-3} is used, our approach predicts the code transforms with an accuracy varying from 41\% to 53\% depending on the transforms. Our model outperforms two baselines based on history probability and neural machine translation (NMT), suggesting the importance of considering code structure in achieving good prediction accuracy. In addition, a proof-of-concept synthesizer is implemented to concretize some repair transforms to get the final patches. The evaluation of the synthesizer on the Defects4j benchmark confirms the usefulness of the predicted AST-level repair transforms in producing high-quality patches.

Zhongxing Yu, Matias Martinez, Benjamin Danglot et al.

Among the many different kinds of program repair techniques, one widely studied family of techniques is called test suite based repair. However, test suites are in essence input-output specifications and are thus typically inadequate for completely specifying the expected behavior of the program under repair. Consequently, the patches generated by test suite based repair techniques can just overfit to the used test suite, and fail to generalize to other tests. We deeply analyze the overfitting problem in program repair and give a classification of this problem. This classification will help the community to better understand and design techniques to defeat the overfitting problem. We further propose and evaluate an approach called UnsatGuided, which aims to alleviate the overfitting problem for synthesis-based repair techniques with automatic test case generation. The approach uses additional automatically generated tests to strengthen the repair constraint used by synthesis-based repair techniques. We analyze the effectiveness of UnsatGuided: 1) analytically with respect to alleviating two different kinds of overfitting issues; 2) empirically based on an experiment over the 224 bugs of the Defects4J repository. The main result is that automatic test generation is effective in alleviating one kind of overfitting issue--regression introduction, but due to oracle problem, has minimal positive impact on alleviating the other kind of overfitting issue--incomplete fixing.

Benjamin Danglot, Oscar Luis Vera-Pérez, Zhongxing Yu et al.

The adoption of agile development approaches has put an increased emphasis on developer testing, resulting in software projects with strong test suites. These suites include a large number of test cases, in which developers embed knowledge about meaningful input data and expected properties in the form of oracles. This article surveys various works that aim at exploiting this knowledge in order to enhance these manually written tests with respect to an engineering goal (e.g., improve coverage of changes or increase the accuracy of fault localization). While these works rely on various techniques and address various goals, we believe they form an emerging and coherent field of research, which we call `test amplification'. We devised a first set of papers from DBLP, looking for all papers containing `test' and `amplification' in their title. We reviewed the 70 papers in this set and selected the 4 papers that fit our definition of test amplification. We use these 4 papers as the seed for our snowballing study, and systematically followed the citation graph. This study is the first that draws a comprehensive picture of the different engineering goals proposed in the literature for test amplification. In particular, we note that the goal of test amplification goes far beyond maximizing coverage only. We believe that this survey will help researchers and practitioners entering this new field to understand more quickly and more deeply the intuitions, concepts and techniques used for test amplification.

Zhongxing Yu, Matias Martinez, Benjamin Danglot et al.

Among the many different kinds of program repair techniques, one widely studied family of techniques is called test suite based repair. Test-suites are in essence input-output specifications and are therefore typically inadequate for completely specifying the expected behavior of the program under repair. Consequently, the patches generated by test suite based program repair techniques pass the test suite, yet may be incorrect. Patches that are overly specific to the used test suite and fail to generalize to other test cases are called overfitting patches. In this paper, we investigate the feasibility and effectiveness of test case generation in alleviating the overfitting issue. We propose two approaches for using test case generation to improve test suite based repair, and perform an extensive evaluation of the effectiveness of the proposed approaches in enabling better test suite based repair on 224 bugs of the Defects4J repository. The results indicate that test case generation can change the resulting patch, but is not effective at turning incorrect patches into correct ones. We identify the problems related with the ineffectiveness, and anticipate that our results and findings will lead to future research to build test-case generation techniques that are tailored to automatic repair systems.