I didn't know that! Is the logic behind replacing the sub with a string expression, to fool the cache?

use Benchmark 'cmpthese';

cmpthese(-2, {
    predecl => '
        my $y;
        my $x;
        for $x (1..10000) { $y+=$x }
    ',
    lexical => '
        my $y;
        for my $x (1..10000) { $y+=$x }
    ',
});

__END__
[download]

          Rate lexical predecl
lexical 3996/s      --     -0%
predecl 4001/s      0%      --
[download]

vs

use Benchmark 'cmpthese';

cmpthese(-2, {
    predecl => sub {
        my $y;
        my $x;
        for $x (1..10000) { $y+=$x }
    },
    lexical => sub {
        my $y;
        for my $x (1..10000) { $y+=$x }
    },
});
[download]

          Rate predecl lexical
predecl 4011/s      --     -0%
lexical 4015/s      0%      --
[download]

Which is more or less what haukex and choroba demonstrated.