I think we (or at least I) may be chasing noise. I can't replicate the larger 20% number. Additionally I get noise in both directions on multiple runs. My guess would be over enough iterations we'd see a slightly slower /g.

use Benchmark qw[ cmpthese ];;

my $str = '123456789';
cmpthese -1, {
 a=>q[ my ($a,$b) = $str =~  m/(23)[^8]+(8)/g; ],
 b=>q[ my ($a,$b) = $str =~  m/(23)[^8]+(8)/; ],
 c=>q[ my ($a) = $str =~  m/(23)/g ],
 d=>q[ my ($a) = $str =~  m/(23)/; ],
};

        Rate    a    b    c    d
 a 7047422/s   --  -2% -26% -29%
 b 7218432/s   2%   -- -25% -28%
 c 9578119/s  36%  33%   --  -4%
 d 9960542/s  41%  38%   4%   --

        Rate    a    b    c    d
 a 7143583/s   --   2% -24% -24%
 b 7005183/s  -2%   -- -25% -25%
 c 9378794/s  31%  34%   --  -0%
 d 9387510/s  31%  34%   0%   --
[download]