Well, it all depends on what you do with the results. Consider the following benchmark code:

use strict;
use warnings;
use Benchmark qw(cmpthese);

my @source = 1 .. 1000000;
my @result = @source;

cmpthese (-3,
    {
        for  => '@result = test_for ()',
        map  => '@result = test_map ()',
        mapf => '@result = test_mapf ()',
        forv => 'test_for ()',
        mapv => 'test_mapf ()',
    });


sub test_for {
    my @sqrt_results;
    for my $result (@source) {
        push @sqrt_results , sqrt($result);
    }

    return @sqrt_results;
}


sub test_map {
    return my @sqrt_results = map { sqrt $_ } @source;
}


sub test_mapf {
    return map { sqrt $_ } @source;
}
[download]

Prints:

       Rate  map  for mapf forv mapv
map  2.24/s   --  -3% -31% -49% -78%
for  2.31/s   3%   -- -29% -48% -77%
mapf 3.27/s  46%  41%   -- -26% -68%
forv 4.43/s  98%  92%  36%   -- -56%
mapv 10.1/s 353% 339% 210% 129%   --
[download]

Update: and for about 100,000 or fewer elements percentages are:

         Rate  for  map mapf forv mapv
for   23849/s   --  -3% -29% -46% -76%
map   24506/s   3%   -- -27% -45% -76%
mapf  33666/s  41%  37%   -- -24% -66%
forv  44571/s  87%  82%  32%   -- -56%
mapv 100159/s 320% 309% 198% 125%   --
[download]

In this case for my @source = 1 .. 100;, but the percentages are very much the same over a wide range of elements.