I doubted that repeated splice on a long array is efficient (because arrays are stored as continuous memory in perl, and splice requires copying the remainder of the array), so wrote a small benchmark:

use 5.010;

my @a = map int(10 * rand), 1..1000;

sub d_grep {
    my @copy = @a;
    @copy = grep { $_ % 2 == 0 } @copy;
}

sub d_splice {
    my @copy = @a;
    my $i = 0;
    while ($i < @copy) {
        if ( $copy[$i] % 2 == 1 ) {
            splice @copy, $i, 1
        } else {
            $i++;
        }
    }
}
use Benchmark qw(cmpthese);

cmpthese -2, {
    grep   => \&d_grep,
    splice => \&d_splice,
};
__END__
         Rate splice   grep
splice 2003/s     --   -49%
grep   3900/s    95%     --
[download]

I intentionally used in-place edits in both cases, so that the two implementations remain comparable. If you find some flaws in the benchmark, please let me known - writing benchmarks isn't my particular strength :-)

For longer arrays the difference in speed grows.