I believe the qr// case is much faster in your benchmark because it does not involve dereferencing a coderef and then invoking a subroutine. In any case, I like the idea of using qr// so I do not need to use string eval. I thought that since qr// precompiles regexes, the subroutine using it would be just as fast as the subroutine with eval. I found that isn't the case:

#!/usr/bin/perl -w
#
# compiled regex with qr vs. eval and /o

use strict;
use Benchmark qw(timethese cmpthese);

use vars qw( @words $foo $bar );

@words = qw(heaven evan seven);

sub anon_sub_eval {
  my $pat = shift;
  eval 'sub { grep /$pat/o, @_ }';
}

sub anon_sub_qr {
  my $p   = shift;
  my $pat = qr/$p/;
  eval 'sub { grep /$pat/, @_ }';
}

$foo = anon_sub_eval( q/evan/);
$bar = anon_sub_qr  ( q/evan/);

my $results = timethese( -10,
 {
    anon_sub_eval => sub { $foo->(@words) },
    anon_sub_qr   => sub { $bar->(@words) },
 }
);

cmpthese($results);

__END__
Benchmark: running anon_sub_eval, anon_sub_qr, each for at least 10 CP
+U seconds...
anon_sub_eval: 13 wallclock secs (10.73 usr + -0.01 sys = 10.72 CPU) @
+ 156307.84/s (n=1675620)
anon_sub_qr: 12 wallclock secs (10.40 usr +  0.01 sys = 10.41 CPU) @ 1
+36413.54/s (n=1420065)
                  Rate   anon_sub_qr anon_sub_eval
anon_sub_qr   136414/s            --          -13%
anon_sub_eval 156308/s           15%            --
[download]

Does anyone know why the qr// version is slower?

UPDATE: Thanks to a tip from suaveant, I changed the sub with qr// to not use slashes around the compiled regex:

sub anon_sub_qr {
  my $p   = shift;
  my $pat = qr/$p/;
  eval 'sub { grep $pat, @_ }';
}

__END__
Benchmark: running anon_sub_eval, anon_sub_qr, each for at least 10 CP
+U seconds...
anon_sub_eval: 11 wallclock secs (10.66 usr +  0.00 sys = 10.66 CPU) @
+ 158363.41/s (n=1688154)
anon_sub_qr: 11 wallclock secs (10.39 usr +  0.00 sys = 10.39 CPU) @ 2
+25783.45/s (n=2345890)
                  Rate anon_sub_eval   anon_sub_qr
anon_sub_eval 158363/s            --          -30%
anon_sub_qr   225783/s           43%            --
[download]

Lesson learned: do not use slashes around a regex precompiled with qr// in a grep expression. (But the difference is nominal in a regular match expression: $f =~ /$re/ vs. $f =~ $re)

--sacked