Measurement based automatic analysis of FORTRAN programs

Abstract

A graph model of computer programs has been developed in a series of studies directed toward improving analysis of the structure of programs executed on different computer configurations. One inherent weakness of the model has been the need for estimates of the mean number of times a program would cycle around its loop structures and estimates of branching probabilities. Extensive improvements were made in the model on the assumption that good estimates would be inserted during a manual transformation of a given program into a computer processable graph representation. The combination of improved tools for measurement of program activities and recently developed analysis programs now permit automatic analysis of source programs. The automatic analysis is based on more reliable measured a priori statistics. This paper discusses a valuable by-product of this measurement and analysis which directs attention toward those parts of a program which are leading candidates for application of optimization techniques. In particular we present an example of the automatic analysis of programs written in the FORTRAN IV language. FORTRAN was selected as a first target for analysis because there exists a large number of time-consuming programs written entirely in FORTRAN.