See the diagrams for HVs at the bottom of PerlGuts Illustrated.

• MAX is the number of buckets;
• FILL is the number of those buckets in use;
• KEYS the number of keys in the hash.
• In the empty case ARRAY will be null (even though MAX will be minimum 7), hence the 0 returned for the empty case.

I think that the differences in moritz' benchmark come about because die is an uncommonly slow opcode--essentially a long jump--which completely swamps the conditional test and so skews the timing beyond recognition.

[0] Perl> for(2..6){ 
    %a =(); %b = 1.."1e$_"; 
    cmpthese -1, { 
        A=>q[ for(1 .. 10000){     if( %a ){ 1; } } ], 
        B=>q[ for(1 .. 10000){ unless( %b ){ 1; } } ] 
    } 
};;
    Rate    B    A
B 72.0/s   -- -90%
A  726/s 908%   --
    Rate    B    A
B 72.6/s   -- -90%
A  736/s 914%   --
    Rate    B    A
B 70.5/s   -- -91%
A  751/s 965%   --
    Rate    B    A
B 69.5/s   -- -91%
A  759/s 992%   --
    Rate    B    A
B 71.5/s   -- -90%
A  715/s 900%   --

[0] Perl> for(2..6){ 
    %a =(); %b = 1.."1e$_"; 
    cmpthese -1, { 
        A=>q[ for(1 .. 10000){ unless( %a ){ 1; } } ], 
        B=>q[ for(1 .. 10000){     if( %b ){ 1; } } ] 
    } 
};;
    Rate    B    A
B 75.9/s   -- -89%
A  709/s 834%   --
    Rate    B    A
B 73.7/s   -- -90%
A  730/s 890%   --
    Rate    B    A
B 72.7/s   -- -90%
A  747/s 928%   --
    Rate    B    A
B 71.5/s   -- -90%
A  725/s 914%   --
    Rate    B    A
B 70.6/s   -- -90%
A  708/s 903%   --
[download]

The difference between empty and full hashes is remarkably consistant--if surprisingly high--regardless of the magnitude of the hash.