TIMTOWTDI (and likely a lot more than just these three)

use 5.018003;
use warnings;

use DP;
use Test::More tests => 3;
use Benchmark "cmpthese";

my $n = 100;
my $s = "AAAAAXXXXXXXAAAXXXXXAXXXXAAAAA" x 100;
my $A = "A" x 10000;

sub s_ger      { $s =~ s{A\K([^A]+)(?=A)}        {"A" x         length
+ $1 }ger }
sub substr_ger { $s =~ s{A\K([^A]{1,10000})(?=A)}{substr $A, 0, length
+ $1 }ger }
sub split_join {
    my @s = split m/A\K([^A]+)(?=A)/ => $s;
    for ($_ = 1; $_ < $#s; $_ += 2) {
        $s[$_] = "A" x length $s[$_];
        }
    join "" => @s;
    }

is (s_ger (),      "A" x 3000, "s//x/ger");
is (substr_ger (), "A" x 3000, "s//substr/ger");
is (split_join (), "A" x 3000, "split/join");

cmpthese (-2, {
    s_ger => \&s_ger,
    subst => \&substr_ger,
    spljn => \&split_join,
    });
[download]

->

$ perl test.pl
1..3
ok 1 - s//x/ger
ok 2 - s//substr/ger
ok 3 - split/join
        Rate spljn s_ger subst
spljn 3330/s    --   -2%   -6%
s_ger 3405/s    2%    --   -4%
subst 3550/s    7%    4%    --
[download]

YMMV