Run-Time Data-Flow Analysis

Parallelizing compilers have made greatprogress in recent years. However, there still remains a gap between thecurrent ability of parallelizing compilers and their final goals. Inorder to achieve the maximum parallelism, run-time techniques were usedin parallelizing compilers during last few years. First, this paperpresents a basic run-time privatization method. The definition ofrun-time dead code is given and its side effect is discussed. Toeliminate the imprecision caused by the run-time dead code, backwarddata-flow information must be used. Proteus Test, which can usebackward information in run-time, is then presented to exploit moredynamic parallelism. Also, a variation of Proteus Test, the Advanced ProteusTest, is offered to achieve partial parallelism. Proteus Testwas implemented on the parallelizing compiler AFT. In the end of thispaper the program fpppp.f of Spec95fp Benchmark is taken as an example,to show the effectiveness of Proteus Test.