The advanced array analyses in SUIF are important to the successful parallelization of the NAS benchmarks, as can be seen in Figure 5(B2)-(D2). Comparing SUIF with the baseline system, we observe that the array analyses have two important effects. They enable the compiler to locate significantly more parallelism in two of the programs, cgm and embar. They also increase the granularity of parallelism in appbt and appsp by parallelizing an outer loop instead of inner loops nested inside it. Observe that what seems like a moderate improvement of coverage in appbt---from 85% to nearly 100%---is significant. This difference corresponds to a change in ideal speedup from 2.75 to 4 on 4 processors.
The improvements in coverage and granularity in NAS translate to good speedup results. Six of the eight programs yield a speedup. Of the other two, buk's low coverage is not surprising as it implements a bucket sort algorithm. Applu, although it has high coverage, is too fine-grained to yield any speedup. Overall, the advanced array analyses are important for NAS; half of the benchmark suite would not speed up without these techniques.