Carl Federl and I did some pretty good testing a couple of years back and it very much supports what Matt says... it depends on the function, how it's written, how it's used, where it's used, etc. Carl and I found out that some functions actually operated quicker than the equivalent inline code while others cause huge drags... and everything in between. It also depends on whether or not parallelism is spawned and for what reason.
--Jeff Moden
Change is inevitable... Change for the better is not.