A Compilation Framework for the Automatic Restructuring of Pointer-Linked Data Structures

Abstract

Memory access patterns are a key factor in the performance of data-intensive applications. Unfortunately, changing the layout in memory of pointer-linked data structures is not a trivial task; certainly not if this is to be done by compiler techniques. The presence of type-unsafe constructs complicates this problem even further. In this chapter, we describe a new, generic restructuring framework for the optimization of data layout of pointer-linked data structures. Our techniques are based on two compiler techniques, pool allocation and structure splitting. By determining a type-safe subset of the data structures of the application, addressing can be done in a logical way (by pool, object identifier and field) instead of traditional pointers. This enables tracing and restructuring per data structure. Further, we describe and evaluate our restructuring methodology, which involves compile-time analysis, run-time rewriting of memory regions and updating referring pointers on both the heap and the stack. Our experiments show that restructuring of pointer-linked data structures can significantly improve performance, while the overhead incurred by the tracing and rewriting is worth paying for.