GenIE: Reuse-Oriented Generation of Domain-Specific Instruction Extensions

Abstract

Custom instruction set extensions (ISEs) can boost efficiency in embedded systems but are often narrowly tailored to specific workloads. We propose GenIE (Generation of Instruction Extensions), a fully automated framework for generating, selecting, and reusing microarchitecture-aware ISEs. By introducing a specialization graph, GenIE structures the candidate space for efficient filtering, merging, and reuse across applications. Our end-to-end RISC-V flow integrates profiling, candidate extraction, and retargeting. Experiments on CVA5 with Embench-IoT benchmarks show runtime reductions of up to 42.9 % (avg. 14.3%), while preserving low hardware and encoding overhead. GenIE enables compact, reusable ISEs with strong cross-workload generalization.