I think I may have found a way that could be more efficient:

sub grab {
  my $class = shift;
  my $dice  = shift;
  my $self  = ref($class) || $class;

  unless ($dice =~ m/^(\d+)[Dd](\d+)([Ss])?$/ ) {
    croak "Invalid dice format to $self->grab: $dice";
  }

  #$1 - Quantity
  #$2 - Type
  #$3 - Face  
  my @dice = map { Die->get( $2, $3 ? 1 : 0 ) } 1..$1;
  
  return bless { dice => \@dice }, $self;
}
[download]

Here's a summary of what I did:

I changed the code to be more in-line, as there is no need for a logic branch if the regex fails to return a match. Inside the croak command, I figured it was better to report back the passed in $dice, rather than $!, which is generaly set by system-level errors.

The regex was optimized. The /i modifier wasn't necessary because you were wanting to match d and D, using a character class is faster anyhow, as in: [Dd]. Same with the [Ss] on the end, which was made an optional match, rather than use the (.*) or (.*s). The Mastering Regex book cautions against using the /i when a character class will do just as well.

Instead of creating temporary variables, I used the $1, $2, $3 variables directly. Of course, I compensated for the loss in readability, by using comments =).

I could have pushed onto an array inside a loop, but map was used and the values were copied directly onto the array. I benchmarked using for, foreach and map, with map coming out as a winner in my tests. The others won when I experimented with making $1 equal to 10000 or so, but this will most likely not get this high, in the real-world.

Also, you'll notice that $3 ? 1 : 0 was used, it's not possible to do something like $3 &&= 1, since $3 is a read-only variable.