You can squeeze out some micro-optimisations. In general, ++$x is slightly faster than $x++; playing with skeeve3_i, I also found that /.../ was faster than /.{3}/ (to an extent that surprised me).

          Rate l3epost  l3epre l3ipost  l3ipre
l3epost 42.2/s      --     -3%     -8%    -10%
l3epre  43.4/s      3%      --     -5%     -8%
l3ipost 45.8/s      9%      6%      --     -3%
l3ipre  47.1/s     12%      9%      3%      --
[download]

is the output from this code:

use strict;
use Benchmark qw/ cmpthese /;

our $a = "abcdefghij" x 1000;
cmpthese(-1, {
  l3epost => q{ my %a; $a{$1}++ while $a =~ /(?=(.{3}))/g },
  l3epre  => q{ my %a; ++$a{$1} while $a =~ /(?=(.{3}))/g },
  l3ipost => q{ my %a; $a{$1}++ while $a =~ /(?=(...))/g },
  l3ipre  => q{ my %a; ++$a{$1} while $a =~ /(?=(...))/g },
});
[download]

It shouldn't be hard to improve perl to compile $x++ to a preincrement in void context; fixing the regexp engine to make an explicit fixed count as fast as unrolling is likely to be harder.

Update: Now I'm confused: I checked the source, and void postincrement should be getting replaced with preincrement at compile time; I definitely don't understand now why explicit preinc is showing consistently faster times.

Hugo