Indeed. There are trade-offs everywhere: by default, I always write

 sort {} grep {} @foo;
[download]

so that the grep happens first. The speed is more important than the memory for almost everything I do.

Here's a little bit of lies/damn-lies/statistics demonstrating the speed difference on an array in which most values appear twice.

#!/usr/local/bin/perl -w

use Benchmark;
use strict;

my @data;
my %seen;
foreach ( 1..100_000 ) { push @data, int($_/2) }
fisher_yates_shuffle ( \@data );

#my @gs = gs(); print join ( "\n", @gs );
#my @sg = sg(); print join ( "\n", @sg );

timethese ( 20, { sort_grep => \&sg,
                  grep_sort => \&gs,
} );

sub gs {
  return sort { $a <=> $b }
    grep { ! $seen{$_} ++ }@data;
}

sub sg {
  return grep { ! $seen{$_} ++ }
    sort { $a <=> $b } @data;
}

# fisher_yates_shuffle( \@array ) : from Perl Cookbook
sub fisher_yates_shuffle {
    my $array = shift;
    my $i;
    for ($i = @$array; --$i; ) {
        my $j = int rand ($i+1);
        next if $i == $j;
        @$array[$i,$j] = @$array[$j,$i];
    }
}
[download]

And the output I get:

Benchmark: timing 20 iterations of grep_sort, sort_grep...
 grep_sort:  5 wallclock secs ( 4.70 usr +  0.02 sys =  4.72 CPU) @  4
+.24/s (n=20)
 sort_grep: 16 wallclock secs (15.56 usr +  0.03 sys = 15.59 CPU) @  1
+.28/s (n=20)
[download]

As a side note, I regard this kind of optimization as "habitual", rather than of the premature kind. <laugh>