I was curious - my "goal" when jotting down the code was to use the minimal number of Perl ops and to let Perl do as much work as possible in the C parts of Perl. So I returned a list from map instead of returning a string:

This returns the list and joins all the fragments later:

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

This does a partial fragment concatenation in the map and returns a single string, but still does the concatenation later, but of half as many elements - one additional op in the map call:

sub map_split_join {

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

I didn't expect the difference to be that big:

                 Rate map_split_join          subst      map_split
map_split_join 1697/s             --            -7%           -13%
subst          1828/s             8%             --            -6%
map_split      1949/s            15%             7%             --
[download]

So, I can only recommend reading/listening to When Perl Is Not Quite Fast Enough. Undoubtedly, there is still lots of potential in optimizing this.