You are still testing subroutine call speed rather than the snippets you purport to be testing.

Whatever code snippets you supply to benchmark, get wrapped internally into subs (See Benchmark::runloop.)

By using strings instead of subs, you have removed one layer of indirection, but you are still swamping the time taken for the code under test, by the time taken to invoke the subroutine that gets wrapped around it.

The only way to get anything like an accurate measurement for this type of micro-benchmark, is to add a multiplier loop inside the subroutine Benchmark constructs, so as to amortise the costs of calling that sub over a large number of iterations, to give a+(k/1e4) ~= b/(k/1e4). (* or whatever multiplier is appropriate.)

Also, I'm not sure what the cost of use strict; and use warnings is, when they have already been loaded, but there must be some if only to discover they are already loaded plus the calling of (or attempted call of) their import subs.

As Benchmark already adds use strict to the subs it constructs, that's pure duplication. And as it already has use warnings in force internally, when it eval's the subs into existance, I don't think you are gaining anything by adding it to the code that gets eval'd. You are simply mudding the waters further by adding another fixed cost to the tests.