Automatic checking of the usage of the C++ move semantics

Abstract

The C++ programming language is a favorable choice when implementing high performance applications, like real-time and embedded programming, large telecommunication systems, financial simulations, as well as a wide range of other speed sensitive programs. While C++ has all the facilities to handle the computer hardware without compromises, the copy based value semantics of assignment is a common source of performance degradation. New language features, like the move semantics were introduced recently to serve an instrument to avoid unnecessary copies. Unfortunately, correct usage of move semantics is not trivial, and unintentional expensive copies of C++ objects - like copying containers instead of using move semantics - may determine the main (worst-case) time characteristics of the programs. In this paper we introduce a new approach of investigating performance bottlenecks for C++ programs, which operates at language source level and targets the move semantics of the C++ programming language. We detect copies occurring in operations marked as move operations, i.e. intended not containing expensive copy actions. Move operations are marked with generalized attributes - a new language feature introduced to C++11 standard. We implemented a tool prototype to detect such copy/move semantic errors in C++ programs. Our prototype is using the open source LLVM/Clang parser infrastructure, therefore highly portable.