Empirical evaluation of some features of instruction set processor architectures

Abstract

This paper presents methods for empirical evaluation of features of Instruction Set Processors (ISPs). ISP features are evaluated in terms of the time used or saved by having or not having the feature. The methods are based on analysis of traces of program executions. The concept of a register life is introduced, and used to answer questions like: How many registers are used simultaneously? How many would be sufficient all of the time? Most of the time? What would the overhead be if the number of registers were reduced? What are registers used for during their lives? The paper also discusses the problem of detecting desirable but non-existing instructions. Other problems are briefly discussed. Experimental results are presented, obtained by analyzing 41 programs running on the DECsystem10 ISP.