Compiler Support for Value-Based Indirect Branch Prediction

Indirect branch targets are hard to predict as there may be multiple targets corresponding to a single indirect branch instruction. Value Based BTB Indexing (VBBI), a recently proposed indirect branch prediction technique, utilizes the compiler to identify a ‘hint instruction', whose output value strongly correlates with the target address of an indirect branch. At run time, multiple targets are stored at different branch target buffer (BTB) locations indexed using the branch PC and the hint instruction output value. In this paper, we present compiler support for the VBBI prediction scheme. We also propose compiler and run time optimizations to increase the dynamic instruction count between the indirect branch and its corresponding hint instruction. The more the dynamic instructions between the hint-jump instruction pair, the more likely that the hint value will be available when making the prediction. Our evaluation shows that the proposed compiler and run time optimizations improve the VBBI prediction accuracy from 66% to 80%. This translates into performance improvement from 17.2% (baseline VBBI) to 24.8% (optimized VBBI) over the traditional BTB design and from 11% (baseline VBBI) to 17.3% (optimized VBBI) over the best previously proposed indirect branch prediction scheme.