Returning references is fast, but making use of the reference later is not always very practical, as you now must write $r->{$key} where otherwise you could have written $r{$key}.

So often, I let the user decide, and base my return value on wantarray: If a sub is called in scalar context, I return a ref, but in list context, I return a flat list.

sub foo {
    my %hash = ( a => 1, b => 2 );
    return wantarray ? %hash : \%hash;
}
[download]

Call as either of:

my $hashref = foo();
[download]

or

my %hash = foo();
[download]