Those gains are often caused by globally applied optimizations like inlining and devirtualization. These then can enable actual stream processing and loop chunking optimizations which are of important but lesser effect. Perhaps dead code and constant elimination of you leave in flags.

In other words, if you rewrite the algorithms from object oriented into stream processing, you would get most of the gains.

This is caused by all CPU, FPU and GPUs actually internally being simple numerical stream machines.