A Study of the E ects of Compiler-Controlled Speculation on Instruction and Data Caches

Compiler-controlled speculation has been shown to be e ective in increasing instruction level parallelism (ILP) found in non-numeric programs. However, it is not clear the extent to which speculatively scheduled code may a ect the instruction and data caches. In particular, the amount of time spent resolving cache misses may be signi cant enough to prevent the more aggressive speculation models from attaining their best potential performance results. The objective of this paper is to quantify these e ects using aggressive speculation models. Index terms instruction cache e ects, data cache e ects, compiler-controlled speculation, safe speculation, VLIW, superscalar