I'd call it "abuse". My bet is this pattern of application is well-known and tolerated for the sake of critical mass of existing "cool examples of (ab)using re-engine", and therefore safe to use in the future :). Stand-alone (*F) is guaranteed to fail, there's no need to "force to backtrack" while staying in the same branch; and as there are no other branches in your example, the whole matching must have been optimized away. On the other hand, something like (?(?{CODE})(*F)), with CODE result depending on sub-matches so far, is legitimate use and another matter entirely, but not the case here.

The impression is, aforementioned tolerance goes as far as injection of (*F) makes (but not always) engine fail to fail early, which is funny.

my $match = qr[([ab]+)([ab]+)];
my $str = 'aba';

$str =~ /^ $match $ (?{  print "1: $1-$2\n" }) a /x;
$str =~ /^ $match $ (?{  print "2: $1-$2\n" }) b /x;
$str =~ /^ $match $ (?{  print "3: $1-$2\n" }) (*F) b /x;
$str =~ /^ $match $ (?{  print "4: $1-$2\n" }) (*F) .. /x;

__END__

1: ab-a
1: a-ba
3: ab-a
3: a-ba
[download]