A Characterization of Prolog Execution

We analyze the execution of a new suite of medium-sized and realistic benchmarks through simulations on ageneral-purpose, register-oriented architecture at the abstract machine and architectural levels to identify the mostcritical characteristics of Prolog for efficient execution. We propose improvements through modest enhancements tothe architecture and to the Warren abstract machine to add significant support for the identified issues. Architecturalsupport for tag-handling operations through an architecture which distinguishes between the tag and value fields ofan object leads to a twelve percent decrease in static code and a seven percent decrease in executed instructions.The addition of orthogonal tag instructions reduces code size by an additional nine percent and reduces dynamic in-structions by eleven percent. The introduction of push and pop instructions decreases code size by seven percentand the number of instructions executed by eight percent. Static code is reduced by forty percent and the number ofinstructions executed decreases by nine percent by utilizing procedure argument mode information at compile timewithin compiled unification operations. Knowledge of dereferencing characteristics reduces static code by six per-cent and reduces instructions executed by six percent by eliminating dereferencing for objects which are directly ref-erenced and when previously dereferenced values may be substituted for non-dereferenced objects. Optimization ofarithmetic expression evaluation eliminates six percent of the executed instructions which create arithmetic expres-sions and eliminates seven percent of the executed instructions which evaluate these structures and decreases thestatic code size by ten percent. A modified WAM-database execution model is proposed which reduces static codeby twenty percent with a modest four percent increase in executed instructions. We find Prolog to attain only mod-est speedups through simple pipelined and multiple-operation-issue machine implementations and propose new di-rections to explore to allow Prolog to better exploit the increasing levels of instruction-level parallelism attainable inmodern architectures.