Sanusi B. A., Ogunshile E. K., Aydın M. E.
International Journal of Innovative Science and Research Technology, cilt.11, sa.9, ss.657-669, 2026 (Hakemli Dergi)
Özet
The increasing complexity of modern software systems demands more intelligent and reliable compiler designs
capable of detecting code anomalies efficiently. Traditional compilers, while effective in syntax and semantic analysis, lack
adaptive capabilities for identifying non-trivial errors. This paper presents a performance-aware hybrid compiler that
integrates the Communicating Stream X-Machine (CSXM) formal framework with Machine Learning (ML) techniques for
intelligent code analysis. The approach combines formal verification with predictive models, including Random Forest,
Convolutional Neural Networks (CNN), and CNN-LSTM architectures. Experimental evaluation was conducted using both
real-world (DeepFix) and synthetically generated datasets, with over 150 training and testing iterations. Performance was
assessed using accuracy, precision, recall, and F1-score. In addition, this study provides a detailed analysis of compilation
latency and ML inference overhead, highlighting the trade-off between detection accuracy and execution efficiency. Results
show that ML integration significantly improves anomaly detection but introduces measurable latency. The paper also
discusses practical challenges, including dataset labelling effort and model generalisability. These findings demonstrate that
a performance-aware hybrid CSXM–ML compiler offers a promising approach for developing intelligent, adaptive
compilation systems that balance correctness, efficiency, and scalability.