GFTrans: an on-the-fly static analysis framework for code performance profiling

Area of Science:

• Computer Science
• Software Engineering
• Artificial Intelligence

Background:

• Software efficiency is critical for system maintenance and performance.
• Identifying runtime bottlenecks in large, complex systems is challenging.
• Traditional dynamic profiling methods introduce significant latency, hindering agile development.

Purpose of the Study:

• To introduce GFTrans, a novel static analysis framework for predicting C program performance.
• To enable developers to identify performance bottlenecks early in the development cycle without code execution.

Main Methods:

• Developed GFTrans, a static analysis framework utilizing a Transformer architecture.
• Implemented an "anchor-based embedding" technique to integrate control flow and data dependencies.
• Incorporated a dynamic gating mechanism to fuse semantic representations with handcrafted statistical features.

Main Results:

• GFTrans achieved 78.64% accuracy on a dataset of real-world C functions.
• Outperformed baseline models such as Random Forest and Code2Vec.
• Identified potential performance bottlenecks in milliseconds.

Conclusions:

• GFTrans offers an effective solution for predicting C program performance through static analysis.
• The framework significantly reduces the latency associated with bottleneck identification.
• Enables developers to optimize code proactively during the development phase.