Incremental program analysis (IPA) has gained increasing attention as an effective approach for maintaining up-to-date analysis results by leveraging previously computed results in response to program changes. Consequently, a variety of IPA algorithms and tools have been proposed. However, their empirical performance in practical, real-world scenarios remains insufficiently investigated. To address this gap, this study presents a comprehensive examination of the current state-of-the-art in IPA evaluation. Specifically, we identify two key limitations: (1) the lack of standardized benchmarks reflecting real-world program changes, and (2) the inadequacy and imbalanced distribution of evaluation metrics.To overcome these challenges, we propose an automated pipeline for constructing real-world program change benchmarks and develop a unified incremental evaluation framework for systematically evaluating IPA tools. Using the proposed evaluation pipeline, we constructed large-scale benchmarks of real-world program changes—sourced from 4,084 commits across 20 Java projects—and systematically evaluated two IPA tools for Java. The results demonstrate that, although incremental analysis substantially improves efficiency compared to exhaustive analysis, existing IPA tools exhibit inconsistencies and markedly higher peak memory consumption. Finally, we distill practical insights from our findings to inform future research and development in the field of incremental program analysis.

The IFDS framework supports interprocedural dataflow analysis with distributive flow functions over finite domains. A large class of interprocedural dataflow analysis problems can be formulated as IFDS problems and thus can be solved with the IFDS framework precisely. Unfortunately, scaling IFDS analysis to large-scale programs is challenging in terms of both massive memory consumption and low analysis efficiency. This paper presents DStream, a scalable system dedicated to precise and highly parallel IFDS analysis for large-scale programs. DStream leverages a streaming-based out-of-core computation model to reduce memory footprint significantly and adopts fine-grained data parallelism to achieve efficiency. We implemented a taint analysis as a DStream instance analysis and compared DStream with three state-of-the-art tools. Our exper-iments validate that DStream outperforms all other tools with average speedups from 4.37x to 14.46x on a commodity PC with limited available memory. Meanwhile, the experiments confirm that DStream successfully scales to large-scale programs which the state-of-the-art tools (e.g., FlowDroid and/or DiskDroid) fail to analyze.