Using the entropy in the SPARC instruction set

Abstract

The paper analyses entropy in an instruction set used to implement software on the SPARC architecture. Statistics about static and dynamic characteristics of programs are used to get information about entropy, redundancy and instruction dependence of instruction sets. Instruction streams are shown to contain on average less than one bit information when treated as high order Markov source instructions. With such low entropies one can achieve much smaller static and dynamic code, due to more compact encoding. Arithmetic coding, a coding technique which can get as close as desired to the entropy, is used to encode and decode the instruction stream. Smaller code size means smaller memory requirements, less swapping, and lower memory/CPU bus traffic. >