Here's another way that avoids a lot of fuss. It's rather like quent's but avoids
unneeded churning on the counter.

It also works with multicharacter substitutions and, with some odd results unless
the start count < 1, the target can be a regex.

#!/usr/bin/perl -w # -*-Perl-*-

use strict;

sub zaxo {
    my ($str,$chr,$new,$nbr)=@_;
    $str=~s/$chr/(--$nbr>0)?$chr:$new/geo;
    $str;
}

my $s = "Terence and Philip are sweet\n";
my $c = 'e';
my $r = 1;
my $n = 3;
print zaxo( $s,$c,$r,$n);
exit;
[download]

Against the bbfu benchmark posted above (corrected &mine output) I get:

Benchmark: timing 500000 iterations of BBFU'S, MINE, Quent's, Zaxo's..
+.<BR/>
    BBFU'S: 12 wallclock secs (12.18 usr +  0.05 sys = 12.23 CPU)
      MINE: 13 wallclock secs (12.32 usr +  0.06 sys = 12.38 CPU)
   Quent's: 13 wallclock secs (13.40 usr +  0.03 sys = 13.43 CPU)
    Zaxo's: 10 wallclock secs (10.59 usr +  0.00 sys = 10.59 CPU)
[download]

I must to agree with MeowChow that this is a litmussy kind of question.

After Compline,
Zaxo