Your algorithm is not a Fisher-Yates shuffle. shuffl chooses a random number between 1 and the number of elements (inclusive) each time. A Fisher-Yates shuffle chooses from a smaller set each iteration to avoid performing a biased shuffle.
Using the dataset A, B, C and D there are 24 possible outcomes (4!), each one with an equal probability of 1/24. The algorithm used for sub shuffl selects 4 random numbers each with a possible value of 1, 2, 3 or 4, or a total of 44 (256) possible outcomes, each with a probability of 1/256 (some of the final permutations are duplicated). 256 is not evenly divisible by 24, so some of the outcomes must be more likely than others.
Update: Oops. BrowserUK's shuffle is indeed a Fisher-Yates shuffle. $a = $_ + rand @{$r} - $_ is parsed as $a = $_ + (rand @{$r} - $_). I assumed it would be parsed as $a = $_ + (rand @{$r}) - $_