Seems to me that if your code is so speed-critical that you're willing to muck around with benchmarking $_ and a lexical that you should be looking at a much higher level speedup, such as writing the routine in C.

Prob'ly the best discussion of this sort of analysis is in Jon Bentley's Programming Pearls. It's well worth checking out.

xoxo,
Andy
--
<megaphone> Throw down the gun and tiara and come out of the float! </megaphone>