Method for vectorizing and executing on an SIMD machine outer loops in the presence of recurrent inner loops

摘要

A method is disclosed for increasing the number of procedural language scalar instructions which may be executed on a vector SIMD machine. The method, operative in the compiling of the source-to-object code, utilizes a mapping representation of the source code statements onto a dependence graph and noting that the innermost sensitive level n and loop-independent edges of the program mapped onto the graph can be vectorized if every level n dependence can be reordered to operate as an innermost loop and if there are no cycles after the reordering. This indicates that the code can be vectorized at level n.