Function Call Overhead

The measurements on this page are based on tests from the C++ Performance TR1. This page specificallly contains tests similar to those from section 5.3 (Classes and Interfaces) measuring the overhead of different approaches to call functions.

The tests are not identical to those used by the performance TR, primarily to address some slightly questionable pratices (e.g., not initializing variables) and to fit them into the commonly used framework. Also, compilers have become a lot smarter and actually see through a number of the uses and entirely optimize away what is intended to be measured. Thus, the tests use xor instead of addition because compilers see through the use addition and optimize the entire loop out while they don't optimize this use of xor.

What's Being Tested

All the tests do the moral equivalent of this loop

unsigned int x(0u);
unsigned int n(10000000u);
for (unsigned int i(0u); i != n; ++i) {
    x ^= i;
}

with the core of the loop (x ^= i;) replaced by a function call using an object suitable for the function call to be worked on.

Report Conventions

Object Names

The naming convention for the names of objects is as follows:

x	an object of base type `X`
xp	a pointer to an object of type `X`
y	an object of derived type `Y` using `X` as a base
yp	a pointer of derived type `Y` using `X` as a base
o	an object of type `S`

Function Names

The components of the function names being called is as follows:

nm	non-member
s	static
nv	non-virtual
v	virtual
o	defined in other source file
i	defined inline
c	defined in the class body or upon its first declaration
p	a pointer to function or pointer to member function

Function Call Overhead

What's Being Tested

Report Conventions

Object Names

Function Names

Results