Why are you penalizing the one-regex methods by requiring a separate assignment to $_? You should perhaps do 2 separate benchmarks; one assuming $_ is to be changed, and the other assuming a lexical $f is to be changed.

Update: I'm just saying that benchmarks for performing an operation on $_ won't necessarily carry over to benchmarks for performing an operation on a lexical. In fact, with a lexical, the two-regex method outshines the others even more:

use Benchmark 'cmpthese';
use Regexp::Common 'whitespace';

my $f = '   this is a string with  spaces  to    remove      ';
cmpthese(-3, {
  'Regexp-Common' => sub { my $g=$f; $g=~s/$RE{ws}{crop}//g;    },
  'two-s///'      => sub { my $g=$f; $g=~s/^\s+//; $g=~s/\s+$//;    },
  'one-s///'      => sub { my $g=$f; $g=~s/^\s+|\s+$//g;        },
  's-capture'     => sub { my $g=$f; $g=~s/^\s*(.*?)\s*$/$1/;   },
  's-capture2'    => sub { my $g=$f; $g=~s/^\s*(\S+(?:\s+\S+)?)?\s*$/$
+1/; },
  'm-capture'     => sub { my $g=$f; ($g) = $g=~/(\S+(?:\s+\S+)?)/;   
+ },
});
[download]

shows "two-s" gaining more than the others when working on lexicals.