>I've gotten significant speed gains in certain projects by simply replacing certain hand-optimized assembly in libraries (ie not my code) with the plain C code

>I've gotten significant speed gains in certain projects by simply replacing certain hand-optimized assembly in libraries (ie not my code) with the plain C code equivalent. That's an interesting point, plus there's the portability issue. My own breadcrumbs of legacy code for this kind of innerloopish stuff has been to write a straightforward 'C' implementation (and time it), an optimized 'C' version (which itself can depend on the processor used), and a handtuned assembly version where really needed. It allows you to back out of the tricky stuff plus acts as a form of documentation.