Your benchmark is utterly flawed. You aren't measuring what you think you are measuring.

• The arrays @empty and @prealloc are lexical. This means, they are known in the file, and only in the file. They cannot be accessed from Benchmark.pm. After all, you are passing in a string, not a coderef. The @empty and @prealloc in the passed strings will be evalled (in the main package) and will hence be package variables (as they aren't my'ed).
• @::empty just grows and grows. Meaning you need the allocate more and more memory. Of course it will be slower! It will be enlightening to print the sizes of @::empty and @::prealloc after the benchmark was run.
• You aren't measuring the influence of the pre-allocation at all. Except for the very first run of "assign", the @::empty array will contain at least $::size elements.
• (q{data} x $::size) returns a single string, even in list context. This explains why "slice" appears to be much faster with strings than with numerals. The slice shouldn't be a win as you state, because that forces Perl to build two relatively large lists.

The following benchmark suggest that preallocating is actually a tad bit slower, and that using slices loses. Assigning is a little faster than pushing.

#!/usr/bin/perl
 
use strict;
use warnings;

use Benchmark qw /cmpthese timethese/;
    
our $size = 100_000;
    
cmpthese timethese (-10 =>  {
    push      =>  'my @arr; push @arr => $_ for 0 .. $::size - 1',
    assign    =>  'my @arr; $arr [$_] =  $_ for 0 .. $::size - 1',
    assignpre =>  'my @arr; $#arr = $::size - 1;
                            $arr [$_] =  $_ for 0 .. $::size - 1', 
    slice     =>  'my @arr; @arr [0 .. $::size - 1] = (0 .. $::size - 
+1)',
    slicepre  =>  'my @arr; $#arr = $::size - 1;
                            @arr [0 .. $::size - 1] = (0 .. $::size - 
+1)',
} => 'none');
__END__
            Rate  slicepre     slice      push assignpre    assign
slicepre  3.82/s        --       -1%      -37%      -40%      -41%   
slice     3.87/s        1%        --      -36%      -39%      -40%
push      6.08/s       59%       57%        --       -4%       -5%  
assignpre 6.36/s       66%       64%        5%        --       -1%
assign    6.44/s       68%       66%        6%        1%        --
[download]

Abigail