You're using a constant input which starts with a "3" though, which unfairly penalizes test1 and test2 (it's the final situation they check for). For inputs starting with a "1", test3 is still the fastest, but the difference between it and the other tests is much smaller.

Also, I'd recommend running your benchmarks like this:

cmpthese(-1, {
    test_1 => q{ test1("3.01.000") },
    test_2 => q{ test2("3.01.000") },
    test_3 => q{ test3("3.01.000") },
});
[download]

... using q{ ... } instead of sub { ... }. If you use sub { ... } you're wrapping each iteration in an extra sub call layer. For micro-optimization benchmarks like this, that extra layer can make a significant difference to the results.