To avoid this problem, you should either declare @a and @b as global variables or pass the code snippets as anonymous subroutine references.

Here's an improved benchmark:

#!/usr/bin/perl -w

use 5.006;

use strict;
use Benchmark qw/ cmpthese /;

use vars qw/ @a @b /;

@a = ('a'..'z');

cmpthese(-3, {
   '1' => '@b = map uc,@a',
   '2' => '@b = map uc $_,@a',
   '3' => '@b = map uc(),@a',
   '4' => '@b = map uc($_),@a',
   '5' => '@b = map {uc} @a',
   '6' => '@b = map {uc $_} @a',
   '7' => '@b = map {uc()} @a',
   '8' => '@b = map {uc($_)} @a',
});

__END__

Benchmark: running 1, 2, 3, 4, 5, 6, 7, 8, each for at least 3 CPU sec
+onds...
         1:  4 wallclock secs ( 3.17 usr +  0.01 sys =  3.18 CPU) @ 97
+06.92/s (n=30868)
         2:  4 wallclock secs ( 3.18 usr +  0.02 sys =  3.20 CPU) @ 97
+10.31/s (n=31073)
         3:  4 wallclock secs ( 3.13 usr +  0.02 sys =  3.15 CPU) @ 97
+73.02/s (n=30785)
         4:  4 wallclock secs ( 3.13 usr +  0.01 sys =  3.14 CPU) @ 98
+98.09/s (n=31080)
         5:  3 wallclock secs ( 3.17 usr +  0.00 sys =  3.17 CPU) @ 96
+11.67/s (n=30469)
         6:  3 wallclock secs ( 3.18 usr +  0.01 sys =  3.19 CPU) @ 96
+50.47/s (n=30785)
         7:  4 wallclock secs ( 3.04 usr +  0.02 sys =  3.06 CPU) @ 95
+82.68/s (n=29323)
         8:  4 wallclock secs ( 3.32 usr +  0.01 sys =  3.33 CPU) @ 95
+33.33/s (n=31746)
    Rate   8   7   5   6   1   2   3   4
8 9533/s  -- -1% -1% -1% -2% -2% -2% -4%
7 9583/s  1%  -- -0% -1% -1% -1% -2% -3%
5 9612/s  1%  0%  -- -0% -1% -1% -2% -3%
6 9650/s  1%  1%  0%  -- -1% -1% -1% -3%
1 9707/s  2%  1%  1%  1%  -- -0% -1% -2%
2 9710/s  2%  1%  1%  1%  0%  -- -1% -2%
3 9773/s  3%  2%  2%  1%  1%  1%  -- -1%
4 9898/s  4%  3%  3%  3%  2%  2%  1%  --
[download]

Update
This node was written partially because I didn't grok the point that chipmunk was making, but I believe it still serves as a useful clarification of the quote below. The important point here is that chipmunk is talking about the Benchmark _file_ and not package. See the follow ups for some interesting implications of this subtle difference. Also I have removed the later portion of the quoted sentence 'where the lexical @a and @b are not in scope ' because it appears I am discussing the lexical variables, whereas I am not. My apologies.
End Update

The eval of the code snippets occurs within Benchmark,

Actually, this isn't correct. Benchmark goes through some interesting (not necessarily completely correct either)contortions to execute code from the package it was called from, in this case main.

This can be seen from this snippet from Benchmark::runloop which is the primary timing routine in Benchmark. (BTW, i only know this cause ive been working on a OO version of Benchmark, coming Real Soon Now to the Code Catacombs :-)

   # find package of caller so we can execute code there
    my($curpack) = caller(0);
    my($i, $pack)= 0;
    while (($pack) = caller(++$i)) {
    last if $pack ne $curpack;
    }

    my ($subcode, $subref);
    if (ref $c eq 'CODE') {
    $subcode = "sub { for (1 .. $n) { local \$_; package $pack; &\$c; 
+} }";
        $subref  = eval $subcode;
    }
    else {
    $subcode = "sub { for (1 .. $n) { local \$_; package $pack; $c;} }
+";
        $subref  = _doeval($subcode);
    }
[download]

Oh and a note to andreychek I dont think that including print,warn statements in a benchmark (unless warn() is what you are benchmarking) is a good idea. The overhead of them executing probably drowns out the difference between the functions. Yves / DeMerphq
--
Have you registered your Name Space?

Regardless of any contortions Benchmark goes through to execute code in the calling package, lexical variables are not in any package. It is impossible for code to access lexical variables that are not in scope.

This can be seen with a very simple Benchmark:

#!perl

use Benchmark;

$main::variable = 'global';
my $variable = 'lexical';

timethese(1,
          {
           anon   => sub { print "$variable\n"; },
           string => 'print "$variable\n";',
          }
         );
__END__
Benchmark: timing 1 iterations of anon, string...
lexical
      anon:  0 wallclock secs ( 0.00 usr +  0.00 sys =  0.00 CPU)
            (warning: too few iterations for a reliable count)
global
    string:  0 wallclock secs ( 0.00 usr +  0.00 sys =  0.00 CPU)
            (warning: too few iterations for a reliable count)
[download]

The anonymous sub snippet accesses the lexical $variable, because the snippet is compiled in the same scope as the lexical variable. The string snippet accesses the global $variable, because the snippet is compiled in a completely separate scope, when the eval is executed.