Code mapping algorithm for custom instructions on reconfigurable instruction set processors

Abstract

Instruction set extension with custom instructions has become common to speed up execution of applications. And it is crucial of the problem to generate new optimised code with custom instructions through retargetable tool suites. This article proposes a code mapping algorithm, which maps static code of programs to dynamic hot paths corresponding to custom instructions. The algorithm uses parameterised polynomial representations and invokes some string matching method. It can play both fine-grained mapping and coarse-grained mapping and find nested matching optimisation in representations. Correctness of the algorithm is formally proved, and time complexity is also analysed from it. Furthermore, some experiments are designed with some programs of NetBench and MiBench. The results show that the algorithm can achieve high accuracy and have notable performance improvement.