/perfscope

A tool that statically evaluates source code commit's risk of introducing performance regression

NOTE: We are working on a follow-up of PerfScope. The source code will be only available on a per-request basis for now. Email ryanhuang@cs.ucsd.edu to request.

This is the source tree for PerfScope, a neat tool for analyzing given patch's performance impact to existing source repository.

PerfScope is a research project from the Opera Research Group in University of California, San Diego. It consists of several parts to operate:

  • Parser: parses the patch file for information like what file(s) are modified, which location does it lie, etc. Located: -- lib/parser/PatchCompiler, Debug+Asserts/bin/patch-c -- lib/parser/PatchIRDecoder, library

  • Mapper: maps the change location in source file to the corresponding programming constructs. Located: -- lib/mapper, library

  • Filter: prunes out insignificant changes such as stylish changes or renaming. Located: -- tools/PerfDiff, Debug+Asserts/bin/perfdiff

  • Profile Database: contains a list of frequent or expensive functions. White-list functions can also be added to the profile. Located: -- data

  • Analyzer: performs Performance Risk Analysis (PRA) on given commit. Located: -- tools/PerfScope, Debug+Asserts/bin/perfscope

  • TestDB: tracks performance test cases execution information to further recommend which test case may be useful to uncover the performance regression issues. (work in progress)

Two auxiliary tools:

  • ListFiles: list the file names and paths in a given module. Located: -- tools/ListFiles, Debug+Asserts/lib/LLVMListFiles.so
  • StaticProfiler: generate a list of expensive and frequent functions based on static analysis. Located: -- tools/StaticProfiler, Debug+Asserts/bin/staticprofiler

Most of the tools have a separate README in their location and its usages can be found using -h option.

For bug report or any other suggestion, please contact ryanhuang@cs.ucsd.edu.