I prefer to let the C part of Perl to do the chore of the work. Especially, if you can chuck work down to the regex engine. Or, even better, if you can pretend to use the regex engine but don't fire it up but get a simple character scan to work for you:

I split $s1 on \0 and then join it back with the corresponding substrings taken from $s2. That reduces the number of comparisons and substrings done in Perl.

sub map_split {
    my $ofs = 0;
    join "", map { $ofs += length; $_ => substr $s2, $ofs++, 1 } split
+ /\0/, $s1, -1;
};
[download]

On this machine I get:

                           Rate  split1 substr1 using_str_bit_ops_and_
+tr map_split
split1                   1.85/s      --    -94%                    -10
+0%     -100%
substr1                  32.9/s   1676%      --                     -9
+5%      -98%
using_str_bit_ops_and_tr  714/s  38390%   2067%                       
+--      -62%
map_split                1898/s 102263%   5663%                     16
+6%       --
[download]

Update: Weird - I would have imagined ikegami's method of using the binary AND of strings to be way faster than mine. And it's even faster than avar's, even though I think that avar's method does far less copying of data (and the use of @+ is far more elegant than my approach).