Research Of Automatic Compiler Tuning Base On Machine Learning

Compiler optimization used manual tuning technique in the past several decades. The rapid development of computer architecture makes the compiler optimization more and more complicated. The compiler developer must pay more time on manual performance tuning, and the tuning result usually can not catch up with the development of the computer architecture. Based on Open64 compiler, the thesis researches an automatic compiler tuning technique that using a method of machine learning.The thesis improves the existed ICI and FCO iterative compilation tool, and uses it to collect the machine learning samples. Instance based learning and decision tree learning are used to automatic tune the software pipelining, loop unrolling and the region formation phases in Open64 compiler. After that, genetic programming is use for automatic performance tuning of the if conversion phase.The thesis compares the result of the three machine learning methods at last. The experimental results show that, automatic compiler tuning can get better performance than the manual tuning in the training cases and in the new cases, they get close performance.