Well, eval is clearly the looser, chr wins:

use Benchmark qw(cmpthese);

my $S = 'Compruebe si las direcciones URL que encontr\u00e9 en el arch
+ivo de configuraci\u00f3n son v\u00e1lidos';

cmpthese(1e6,
    {
        eval => sub { $_ = $S; s/\\u([0-9a-fA-F]{4})/eval "\"\\x{$1}\"
+"/ge; },
        chr  => sub { $_ = $S; s/\\u([0-9a-fA-F]{4})/chr hex $1/ge; },
        pack => sub { $_ = $S; s/\\u([0-9a-fA-F]{4})/pack 'U', hex $1/
+ge; }
    }
);
__END__
         Rate eval pack  chr
eval  38865/s   -- -84% -88%
pack 242131/s 523%   -- -28%
chr  336700/s 766%  39%   --

# without eval

         Rate pack  chr
pack 242131/s   -- -27%
chr  330033/s  36%   --
[download]

which is reasonable, since chr is more specialized than pack and should have less overhead (otherwise chr would have been implemented in terms of pack).