Your algorithm looks perfectly fine to me. random_perm is optimal in the limited sense that it achieves a random permutation in place with the miniminum number of swaps. (You can't do this problem in place without swapping.) It is also nice in that its implementation varies little from language to language, and in that respect it is pretty general. In contrast, to implement the algorithm like the one you posted in C is not straightforward; Perl does a lot behind the scenes to allow one to write splice(@in, rand @in, 1) while @in.

That said, I confess that I am quite surprised by the benchmarks. Then again, this is not by any means the first time that intuitions from C programming as to the efficiencies of various operations turn out to be waaaay off when the problem is translated to Perl. Your post is a good reminder to benchmark, benchmark, benchmark. And just to act on this reminder, I decided to benchmark against your original a version of shuffle that omits the last call to splice:

sub shuffle_2 {
  my @in = @_;
  my @out;
  push @out, splice(@in, rand @in, 1) while @in > 1;
  push @out, @in;
  @out;
}
[download]

            Rate shuffle_2   shuffle
shuffle_2 5856/s        --       -5%
shuffle   6140/s        5%        --
[download]

BTW, I should clarify that I don't recall the real name of the algorithm used by random_perm, if it actually has a name.

Update: As pointed out by several others, the algorithm used by random_perm goes by the name of Fisher-Yates. I've also found it referred to as "Knuth's shuffle."

the lowliest monk