I don't know how accurate this is or exactly what it's measuring but the results suggest something useful is happening.

#!perl

use Benchmark;

$count = shift || die "Need a count!\n";

sub one { eval("use CGI;")}
sub two { eval("use CGI ':cgi';")}
sub three { eval("use CGI ':all';")}

timethese (
$count,{
'Method 1' => '&one',
'Method 2' => '&two',
'Method 3' => '&three',
}
);
exit;
[download]

Results:

Benchmark: timing 500 iterations of Method 0, Method 1, Method 2, Method 3...

Method 1: 0 wallclock secs (0.33 usr + 0.00 sys = 0.33 CPU) @ 1515.15/s (n=500)
Method 2: 2 wallclock secs (1.87 usr + 0.00 sys = 1.87 CPU) @ 267.38/s (n=500)
Method 3: 6 wallclock secs (5.55 usr + 0.00 sys = 5.55 CPU) @ 90.09/s (n=500)

Reference: Benchmarking Your Code.