Depends on what we are talking:

Assignment: (reference appears to be consistantly slower than regular)

use Benchmark qw (timethese);

timethese(10, {"regular" => \&regular, "reference" => \&reference});

sub regular {
    my %hash;
    for my $i (1 .. 100000) {
        $hash{$i} = $i;    
    }
}

sub reference {
    my $hash;
    for my $i (1 .. 100000) {
        $hash->{$i} = $i;    
    }
}
[download]

Benchmark: timing 10 iterations of reference, regular...
 reference:  8 wallclock secs ( 7.76 usr +  0.16 sys =  7.92 CPU) @  1
+.26/s (n=1
0)
   regular:  8 wallclock secs ( 7.20 usr +  0.14 sys =  7.34 CPU) @  1
+.36/s (n=1
0)
Benchmark: timing 10 iterations of reference, regular...
 reference:  9 wallclock secs ( 7.67 usr +  0.17 sys =  7.84 CPU) @  1
+.27/s (n=1
0)
   regular:  8 wallclock secs ( 7.23 usr +  0.08 sys =  7.31 CPU) @  1
+.37/s (n=1
0)
[download]

Assignment + Passing as ref: (reference appears to be faster than regular)

use Benchmark qw (timethese);

timethese(10, {"regular" => \&regular, "reference" => \&reference});

sub regular {
    my %hash;
    for my $i (1 .. 100000) {
        $hash{$i} = $i;    
    }
    helper(\%hash);
}

sub reference {
    my $hash;
    for my $i (1 .. 100000) {
        $hash->{$i} = $i;    
    }
    helper($hash);
}

sub helper {
    my $hash = shift;
}
[download]

Benchmark: timing 10 iterations of reference, regular...
 reference:  8 wallclock secs ( 7.72 usr +  0.14 sys =  7.86 CPU) @  1
+.27/s (n=1
0)
   regular:  9 wallclock secs ( 7.91 usr +  0.08 sys =  7.98 CPU) @  1
+.25/s (n=1
0)
Benchmark: timing 10 iterations of reference, regular...
 reference:  8 wallclock secs ( 7.70 usr +  0.11 sys =  7.81 CPU) @  1
+.28/s (n=1
0)
   regular:  9 wallclock secs ( 7.95 usr +  0.06 sys =  8.02 CPU) @  1
+.25/s (n=1
0)
[download]