Nathan R. Tallent

Performance Analysis of Optimized Code: Binary Analysis for Performance Insight Nathan R. Tallent

Modern applications frequently employ sophisticatedobject-oriented design. In these codes, deep loopnests are often spread across multiple routines. Toachieve high performance, such codes rely oncompilers to inline routines and optimize loops. Consequently, to effectively interpret performance, transformed loops must be understood in the callingcontext of transformed routines. 314. xml alex a. out bod. sh booksurge. cmd booksurge. sh bs. sh config config. test ftpbooksurge. sh ftp. sh html2utf8. c knv. sh libri. sh ls. sh nemesis. list nemesisxml. php nemesisxml. php. old onix onix212. xml onix21. list onix21. xml onixnew. list reha. sh sdl_old. sh sdl. sh test. sh text. xml this validate. sh VDM-ACS-280. xml To understand the performance of optimizedobject-oriented code, we describe how to analyzeoptimized object code and its debugging sections torecover its program structure and reconstruct amapping back to its source code. Using this mapping, we combine the recovered static program structurewith dynamic call path profiles to expose inlinedframes and loop nests. Experiments show thatperformance visualizations based on this informationprovide unique insight into the performance ofcomplex object-oriented codes written in C++.***This work should be of interest to performance toolsdevelopers and, more broadly, application developerswho care about performance. It is implemented inRice University's HPCToolkit, a performance analysistoolkit.