There appears to be a bug in your code which is masked by the data you chose and the test you made. Your search only matches the start of the key, when it should match the end as well. For example with your data, a search for the key 12345 will look in bin 1234 for a string matching $/.'12345', but it could find a string like 1234567 and still return a match.

Also, for interest sake, you can reduce the memory footprint further by stripping out the bucket index from each of the keys. Using the above example again, if we wanted to store 1234567, we can just store 567 in the bucket keyed by 1234. The trivial case of storing 1234 will be stored as an empty string, which will be found correctly by searching for $/$/.

The following code reduced the footprint by another couple of megs. This reduction should also speed up the searching somewhat.

#!/usr/bin/perl

$hash{ substr $_, 0, 4  } .= $/ . substr($_, 4) for 1 .. 1000000;
$hash{$_} .= $/ foreach keys %hash;

$n=0;
print time, $/;
for (1 .. 1000000) {
  next unless exists $hash{ substr $_, 0, 4 }
  and 1+index( $hash{ substr $_, 0, 4 }, $/.substr($_, 4).$/ );
  $n++;
}
print time, $/, $n;
[download]

Other than that, it is an interesting concept that I can't see myself ever using :)