A good description, tachyon++ - you are correct that the lookarounds offer more flexibility but let me point out the differences in terms of benchmarks:

#!/usr/bin/perl
use Benchmark qw/cmpthese/;

$defaulttext = q/foo / x 30;
# $defaulttext = q/foobar / x 30;

cmpthese( 100_000, {
    slash_b => q{$text=$defaulttext; $text =~ s/\bfoo\b//g;},
    neg_look=> q{$text=$defaulttext; $text =~ s/(?<!\w)foo(?!\w)//g;},
    pos_look=> q{$text=$defaulttext; $text =~ s/(?<=[^\w])foo(?=[^\w])
+//g;},
});
[download]

With $defaulttext being 'foo foo ...' all three methods take approx. the same time, the changing of $text takes a decisive amount of time.

With $defaulttext being 'foobar foobar ...' - i.e. no replacements are done - I get the following results:

          Rate    pos_look  neg_look  slash_b
pos_look 27894/s     --        -8%     -35%
neg_look 30441/s     9%         --     -29%
slash_b  42662/s    53%        40%      --
[download]

This shows that the \b variant is about 50% quicker and the negative lookaround is better than the negated character class.

But the most important difference can be seen from the following code

$text= q/foo bar foo/;
($tmp = $text) =~ s/\bfoo\b//g;
print $tmp,"\n";

($tmp = $text) =~ s/(?<!\w)foo(?!\w)//g;
print $tmp,"\n";

($tmp = $text) =~ s/(?<=[^\w])foo(?=[^\w])//g;
print $tmp,"\n";

# which prints:
 bar
 bar
foo bar foo
[download]

The positive lookaround does not behave like the others at the boundaries of the string. This is because the positive lookaround looks for a character (class) but - as there is no character before the beginning of the string or after the end - it fails. The negative lookaround works even if no character is there.

-- Hofmator