Recent compilers provided many of optimization passes, thus even for expert
programmers it is really hard to know which compiler optimization among several
optimizations passes can improve program performance. Moreover, the search space for
optimization sequences is very huge. Therefore, finding an optimal sequence for even a
simple code is not an easy task. The goal of this paper is to find a set of a good
optimization sequences using parallel genetic algorithm. The method firstly classifies the
programs into three clusters then applying three versions of genetic algorithms each one
to cluster in parallel. In order to enhance the result, the migration strategy between these
three algorithms is applied. Three optimal sequences at the same time are obtained from
this method. However, the proposed method improved the execution time on average by
87% compared with the O2 optimization flag. This method also outperforms the
sequential version of genetic algorithm on average of the execution time by 74.8% in
case of using Tournament selection and 72.5% in case of multi-selection method. LLVM
framework is used to validate and execute the proposed method. In addition, Polybench,
Standerford, Shootout benchmarks are used as case study to verify the effectiveness of
the proposed method