Well, one of the great things about Perl is that there's always a way of cheating. In this case, the cheat is to note that s/// returns the number of replacements (and gets rid of the letters we've already looked at). So:

{
 local $/ = undef;                # Read everything in one go

 $_ = lc(<STDIN>);
 s/[^a-z]//g;                     # Get rid of nonalphabetic character
+s

 for my $letter ('a'..'z')
   {
     my $count = s/$letter//g;    # Magic!
     print "$letter = $count\n";
   }
}
[download]

Now, that's many times faster than your original, and a fair bit shorter. I suspect it can be made faster still, though (is the slowdown produced by using s/// offset by the speedup on the next iteration, for instance?)

perl /home/ahunter/original.pl 154.02s user 2.39s system 88% cpu 2:56.29 total
perl /home/ahunter/flib.pl < xlib.ps 4.48s user 0.14s system 93% cpu 4.942 total

Well, that *was* impressive...
-- Andrew