Inductive Logic Programming for Compiler Tuning

  • Craig Blackmore

Student thesis: Doctoral ThesisDoctor of Philosophy (PhD)


Software performance is heavily dependent on the compiler settings used to generate executables from source code. These settings have a significant impact on execution time, energy consumption and code size, and they can be modified easily without the need to alter the underlying source code nor the compiler itself.

Automated Compiler Tuning (ACT) is the practice of selecting compiler settings to improve the performance of computer software. Existing work shows significant potential gains beyond industry standard GCC -O3, but the search space is very large. This motivated the development of predictive machine learning approaches, most of which rely on feature vectors of simple statistics that summarise program characteristics. Inductive Logic Programming (ILP) can potentially exploit full program structure that existing methods do not consider.

This thesis applies ILP to ACT to investigate the hypothesis that exploiting full program structure increases the accuracy and effectiveness of predictive compiler tuning. It explores the design, feasibility and benefits of applying ILP to discover meaningful rules that relate effective compiler settings to program structure. During the process, assumptions in existing work are identified and challenged to increase understanding of the problem and give a robust evaluation of compiler tuning techniques.

The research is conducted on the industry standard GCC compiler and a state-of-the-art benchmark suite targeting ARM Cortex-M3 (CM3) and Cortex-A8 (CA8) processors. The ILP method outperforms two state-of-the-art approaches and also identifies settings that should always be disabled to improve performance on the target platform. This leads to the construction of two new program-agnostic optimisation levels that outperform GCC’s -O3 and are competitive with program-specific predictive approaches. Finally, a case study on an image classification application from the ARM CMSIS Neural Network library finds an 8% improvement over -O3 on the CM3.
Date of Award6 Jun 2019
Original languageEnglish
Awarding Institution
  • The University of Bristol
SupervisorOliver Ray (Supervisor) & Kerstin I Eder (Supervisor)

Cite this