Number of History Buffers

The number of history buffers determines the number of virtual method call sites that share their history. When n=1, all the virtual call sites share the same history buffer and the resulting predictor is called as a global history predictor. In contrast, a per-address history predictor keeps a separate history for all virtual method call sites. This is achieved when n = 2^w, where w is the word size. When the number of history buffers is between 1 and 2^w, a set of addresses use the same history buffer. The effect of the number of history buffers on misprediction rate was studied by using the most significant bits of the program counter to access the history buffers. To mask the effects of other parameters of the predictor, a fully associative target buffer of unlimited size was used. Further, all the bits of the target address were registered in the path history buffers and the hashing address was formed by concatenating the selected path history buffer with the program counter.

Figures 4 and 5 show the results of this investigation for javac and richards benchmarks respectively. The h most significant bits of the program counter select the history buffer corresponding to the call site. The global history predictor is simulated when h=0. In contrast, h=w corresponds to the per-address history predictor. It is observed from the figures that the global path history predictor performs better than those that use per-address or per-set history predictors. For example, the misprediction rates increase from 2.6% for global path history to 4.2% for the per-address scheme using javac with a path length of 2. This indicates that the correlation across call sites is more useful than the self history at a call site in predicting the targets. This can be ascribed to the execution of a series of virtual calls corresponding to the invocation of a single virtual call as shown earlier. A global path history can capture the effect of such constructs better than a per-address scheme. Thus, a global path history is used in refining the other parameters of the predictor.

  

Figure 4: Variation in misprediction rate with number of history buffer sets for javac. A fully associative target buffer of unlimited size was used.

  

Figure 5: Variation in misprediction rate with number of history buffer sets for richards. A fully associative target buffer of unlimited size was used.