"... In PAGE 4: ... For squash recovery, the wait bits provide a simple and efficient solu- tion to dependence prediction with a speedup of 7% on average. Table3 shows the percent of loads predicted and the mispre- diction rate for each of the predictors. Store Sets prediction rates are broken up into the loads that are predicted as independent of prior stores, and the loads that are predicted to be dependent upon a prior store.... ..."

"... In PAGE 4: ... For squash recovery, the wait bits provide a simple and efficient solu- tion to dependence prediction with a speedup of 7% on average. Table3 shows the percent of loads predicted and the mispre- diction rate for each of the predictors. Store Sets prediction rates are broken up into the loads that are predicted as independent of prior stores, and the loads that are predicted to be dependent upon a prior store.... ..."

"... In PAGE 9: ... For squash recovery, the wait bits provide a simple and efficient solution to dependence prediction with a speedup of 7% on average. Table3 shows the percent of loads predicted and the misprediction rate for each of the predictors. Store Sets prediction rates are broken up into the loads that are predicted as independent of prior stores, and the... ..."

"... In PAGE 11: ... For squash recovery, the wait bits provide a simple and e#0Ecient solution to dependence prediction with a speedup of 7#25 on average. Table3 shows the percent of loads predicted and the misprediction rate for each of the predictors. Store Sets prediction rates are broken up into the loads that are predicted as independent of prior stores, and the loads that are predicted to be dependent upon a prior store.... ..."

"... In PAGE 4: ... For squash recovery, the wait bits provide a simple and efficient solu- tion to dependence prediction with a speedup of 7% on average. Table3 shows the percent of loads predicted and the mispre- diction rate for each of the predictors. Store Sets prediction rates are broken up into the loads that are predicted as independent of prior stores, and the loads that are predicted to be dependent upon a prior store.... ..."

"... In PAGE 17: ... We report the mean squared value of the prediction error, MSE, the maximum absolute prediction error, MaxAE, and the number of samples for which the prediction error was greater than the desired tolerance of 10 microns. Table2 shows that M3 was much better than M1 or M2 for both periods, particularly the initial period. For the initial period, DMA with 3 models was slightly better than M3 on all three criteria we report, while for the stable period DMA was essentially the same as M3.... In PAGE 19: ... In the initial unstable period, the relatively simple model M15 had high weight, and in the later stable period, the more complex models M16 and M17 had more weight, as can be seen from Figure 5(b). Table2 shows, strikingly, that DMA with 17 models achieved signi cantly better per- formance in the initial unstable period than either M3 or DMA with 3 models, on all three criteria. This may be because it allows weight to be put on simple, parsimonious models in the early period before stable data has accumulated to estimate more complex models reliably.... ..."