Locality considerations in exploring custom instruction selection algorithms

This paper explores custom instruction identifying methodologies for critical code segments of embedded applications, considering the locality in selection algorithms. We propose two policies for selection of custom instructions. In the first approach, called local selection, the structurally equal custom instructions are classified into the same group, called template, and the selection of custom instructions is done locally in each template. On the other hand, in the second approach no classification is made for the selection and the exploration is made globally on all the enumerated matches. We describe a design flow to establish the desired performance. We study the effects of locality on the overall speed and area of the system. Our experiments show that, locally selected custom instructions give better results in terms of both performance and performance per area.