Compiler Analyses for Improved Return Value Prediction

Speculative method-level parallelism has been shown to benefit from return value prediction. In this paper we propose and analyse two compiler analyses designed to improve the cost and performance of a hybrid return value predictor in a Java virtual machine setting. A return value use analysis determines which return values are consumed, and enables us to eliminate 2.6% of all non-void dynamic method calls in the SPEC JVM98 benchmarks as prediction candidates. An extension of this analysis detects return values used only inside boolean and branch expressions, and allows us to safely substitute incorrect predictions for 14.1% of return values at runtime, provided all future use expressions are satisfied correctly. We find an average 3.2% reduction in memory and up to 7% increase in accuracy. A second, interprocedural parameter dependence analysis reduces the number of inputs to a memoization-based sub-predictor for 10.2% of all method calls, and the combination of both analyses reduces overall memory requirements by 5.3%.