Control flow graph based code optimization using graph neural networks.

Area of Science:

• Computer Science
• Software Engineering
• Artificial Intelligence

Background:

• Selecting effective compiler optimization flags is complex and requires significant expertise.
• Previous methods relied on static, spatial, or dynamic code features.
• Recent advancements explore deep neural networks applied directly to source code.

Purpose of the Study:

• To develop an automated method for identifying optimal compiler optimization flags.
• To combine static, spatial, and deep learning features for improved flag selection.
• To leverage Graph Neural Networks (GNNs) for source code representation and analysis.

Main Methods:

• Source code was represented as graphs to facilitate GNN analysis.
• A dataset of 12,000 graphs was created using 256 optimization flag combinations across 47 benchmarks.
• A Graph Neural Network model was trained and evaluated on this dataset.

Main Results:

• The GNN model successfully learned to select effective optimization flags.
• A maximum speed-up of 48.6% was achieved compared to enabling all optimization flags.
• This demonstrates the potential of GNNs in optimizing software performance.

Conclusions:

• Graph Neural Networks offer a promising approach for automating compiler optimization flag selection.
• The proposed method significantly enhances code execution speed.
• This research advances the field of automated software optimization through deep learning.