Feature-specific profiling

Abstract. High-level languages come with significant readability and maintain-ability benefits. Their performance costs, however, are usually not predictable, at least not easily. Programmers may accidentally use high-level features in ways that compiler writers could not anticipate, and they may thus produce underper-forming programs as a result. This paper introduces feature-specific profiling, a profiling technique that re-ports performance costs in terms of linguistic constructs. With a feature-specific profiler, a programmer can identify specific instances of language features that are responsible for performance problems. After explaining the architecture of our feature-specific profiler, the paper presents the evidence in support of adding feature-specific profiling to the programmer’s toolset. 1 Weighing Language Features Many linguistic features,1 come with difficult-to-predict performance costs. First, the cost of a specific use of a feature depends on its context. For instance, use of reflection may not observably impact the execution time of some programs but may have disas