Practical Assignment Sinking for Dynamic Compilers

Assignment sinking improves the efficiency of programs by avoiding variable assignments on paths where the variable isn't used. A single assignment is partially live if its value is used on one execution path and not on another path. By sinking the assignment onto the path where the value is used, the other path is made more efficient. In many cases, the path to which the assignment is sunk is also made more efficient because the generated code quality can be improved. Assignment sinking has traditionally been accomplished using the classic Partial Redundancy Elimination (PRE) framework. But PRE involves several dataflow analyses to identify the optimal program points for each assignment which limits its usefulness in scenarios where compile time is important, such as in a dynamic compiler. In this talk, we describe a more efficient method to sink assignments using fewer dataflow analyses. This algorithm has been implemented in the TestaRossa Just-In-Time compiler in the IBM Developer Kit for Java

Greg Steffan

Last modified: Fri Aug 31 10:23:47 EDT 2007