Optimization of expressions in Fortran
A method of optimizing the computation of arithmetic and indexing expressions of a Fortran program is presented. The method is based on a linear analysis of the definition points of the variables and the branching and DO loop structure of the program.
The objectives of the processing are (1) to eliminate redundant calculations when references are made to common sub-expression values, (2) to remove invariant calculations from DO loops, (3) to efficiently compute subscripts containing DO iteration variables, and (4) to provide efficient index register usage.
The method presented requires at least a three-pass compiler, the second of which is scanned backward. It has been used in the development of several FORTRAN compilers that have proved to produce excellent object code without significantly reducing the compilation speed.