A performance comparison between an APL interpreter and compiler

Abstract

An analysis of the execution of several simple APL statements illustrates that interpretive overhead in the form of setup time is awesome, and is on the order of 100 times as much as the per-element time. Comparison with an experimental compiler for APL illustrates that the setup time can often be eliminated or reduced greatly, producing code fragments that are more than 300 times as fast in favorable cases. Further work on the compiler should make some compiled code fragments about 500 times as fast as interpreted programs in favorable cases. In less favorable cases, performance improvements of a factor of two are commonplace. Many opportunities remain to improve the performance in such cases. A production compiler based on this work might allow use of APL to solve problems that in the past have been prohibitively expensive for APL solution.