Your benchmark code is wrong. You call each of your coderefs instead of passing them. You meant instead something like this:

cmpthese( 500_000, {
  statreuse => $coderef,
  nostatreuse => $coderef1,
});
[download]

With working benchmark code I see the _ version running 78% faster than the separate calls.

Update: Meh, typo'd the percentage faster. Actual results running 500,000 tests with test.lnk being a plain file:

                Rate nostatreuse   statreuse
nostatreuse  65189/s          --        -44%
statreuse   116279/s         78%          --

Results from stock OS X perl v5.8.6 on a dual 2.7 G5.