I've found that this is a good way to introduce bespoke code into the codebase whose perf gains don't generalize.

Unless its an easy memory/parallel/algorithmic win, its not worth it.