Splat Provides Programmers a Fast and Accurate Study of Memory Behavior without the Necessity of a Costly Memory Simulator. the Tool Is Suitable for Use as a Step in an Iterative Optimization

Memory performance is becoming a major bottleneck in current microprocessors. A great deal of research has aimed at developing techniques for improving memory performance. Some of these techniques rely on hardware alone, but many require programmer or compiler support. Examples of the latter are software prefetching, blocking, and copying. To use these techniques effectively, the programmer must have some knowledge of the program’s behavior. For instance, prefetching is useful only if it is limited to instructions that frequently produce cache misses. Adding a prefetch instruction to every memory instruction could result in significant performance degradation. These techniques might also require quantification of the different types of cache misses (see sidebar, page 60). For instance, microprocessors can avoid compulsory misses through both hardware and software prefetching. Blocking, or tiling, is a method of avoiding capacity misses; copying and padding are techniques for reducing the effect of conflict misses. Many processors provide hints in their memory instructions that the compiler can use for optimizing memory performance. Examples of such hints are the PowerPC’s cache bypass facility and the hints incorporated by the IA-64 instruction set. Effective use of these hints requires information about the program’s locality behavior. The process of obtaining information about a program’s locality characteristics is data locality analysis. Traditionally, this analysis takes place either at compile time or at runtime. The former approach incurs low overhead but is relatively inaccurate because the compiler lacks some information. The runtime approach usually takes the form of a memory hierarchy simulation, which is quite accurate but very slow. In this article, we introduce SPLAT (Static and Profiled Data Locality Analysis Tool). The tool’s purpose is to provide a fast study of memory behavior without the necessity of a costly memory simulator. SPLAT consists of a static locality analysis enhanced by simple profiling data. Its overhead is low because it performs most of the analysis at compile time, and because the required profiling support is just a basic-block-execution count. Many commercial compilers support this profiling option. Compared with simulation techniques, SPLAT’s estimation technique is highly accurate for numeric codes. The tool is useful not only for compilers but also for programmers. To tune a program, programmers should know its performance, the Jesús Sánchez Antonio González