Lvalue subroutines should have a single point of return – and not use the return keyword. So something like

my @x;
sub foo :lvalue {
    my ($cond, $i) = @_;
    if ($cond) {
        return $x[$i];
    } else {
        return $x[0];
    }
}
[download]

should instead be written as

my @x;
sub foo :lvalue {
    my ($cond, $i) = @_;
    my $lvalue;
    if ($cond) {
        $lvalue = \$x[$i];
    } else {
        $lvalue = \$x[0];
    }
    $$lvalue;
}
[download]

which is a general solution, or

my @x;
sub foo :lvalue {
    my ($cond, $i) = @_;
    my $lvalue;
    ($cond) ? $x[$i] : $x[0];
}
[download]

Edit: further notes I see you are using the mem module, which is basically useless. If you need to perform an assignment at compile time, just use a BEGIN block:

our @fields;
BEGIN { @fields=qw(one two three) }
[download]

While this is a little more verbose, you now don't need any non-standard modules. A similar rationale applies to the P package. It attempts to be very DWIM-y, but I find it's design distasteful. Here, explicitly using the builtin printf rather than the P function would have been clearer.

When posting minimal code snippets for others to reproduce a problem, then it's customary to remove such irrelevant dependencies – I don't shove my similar Util::Underscore in everyone's face.