HOWEVER, scrutinizing the Benchmark code reveals that *it* in fact string evals.

I don't see how it can possibly get something useful out of a sub reference's stringification, but I'm too lazy to look at Benchmark's source. Let's try something different:

#!/usr/bin/perl -w
use Benchmark qw(cmpthese);

sub single { $foo = 'foo' }
sub double { $foo = "foo" }
cmpthese (10000000, {
        single => sub { single() },
        double => sub { double() }
});
[download]

- Yes, I reinvent wheels.
- Spam: Visit eurotraQ.

It yields a negligible difference 1-3% which I would attribute to system noise. Interestingly(?) it takes much no longer to run, or at least reports more wallclock seconds (15-19 vs 2 or 3 of the other (although it was obviously longer than that)).

--
perl -pew "s/\b;([mnst])/'$1/g"