A Retargetable Framework for Automated Discovery of Custom Instructions

The problem of efficiently mapping a software application onto an extensible processor has received considerable attention. However, except for specialized kinds of computation accelerators, end-to-end studies of the problems are hard to find in the literature. We propose a classification of previous work on the mapping problem; we then frame previous results into this classification, and propose a new framework for solving this problem. By dividing the problem into several parts-some of them solved exactly, some of them relying on greedy algorithms-we provide a generic scheme that can be adapted to different kinds of hardware accelerators. We implemented our approach on top of a GCC-based compiler toolchain for extensible processors. Benchmarks taken from MiBench show a speedups up to 6.74 x using the SimpleScalar/ARM cycle-exact simulator.