Abstract: Iterative search of compiler optimization parameters is a popular and modern method to increase performance of compiled code relatively to default compilation. Most of the existing papers apply optimization parameters to the whole application, while our paper describes results got by the specially modified compiler capable to parameterize each function independently. Six well-known search algorithms were implemented and used to search the best parameters in the space of all possible combinations of options and their values. Standard SPEC CPU2000 benchmarks were compiled using internal version of Intel® Compiler 10.1. Applications were executed on Intel® Core 2TM with Windows* 2003 OS platforms. Six compiler options having the most impact on performance were selected.
As a result we got the maximum performance increase of 14.5% for a 178.facere benchmark and 2.2% performance increase of total SPEC CPU2000 execution time. Search algorithms comparison shows that most algorithms get very close results varying within 2% from each other. Some algorithms could get their best results after 20-40 iterations while others could get the same results only after 100 iterations which was the limit of a number of iterations.