In this benchmark, we explore the importance of keeping data within the CPU cache to avoid expensive retrieval from RAM. By simply ensuring linear data access and accessing by blocks that fit in L1 and L2, we can achieve massive performance gains without changing the underlying algorithm.
Returning a moved big object will actually create extra overhead for the compiler, as a zero-cost operation is traded for a cheap move operation.
const T& binds to everything but it is dangerous
Ideally, we should never throw on the move Constructor of an object. If a move constructor throws an exception while a vector is resizing, the vector is in a “broken” state: half of the elements are in the new memory, and half are in the old memory. Move operations often modify the source object (leaving it empty or null), making it dangerous to undo the changes. Therefore, the compiler wants to avoid this risk and uses a copy rather than a movment. We want to avoid this, and the way to do this is to declare the move constructor as noexcept. Like this, we give the promise that we do not throw, and the compiler can safely use it. ...