perl -e '"01234" =~ /^(.+)(.+)((?:.z?)+)$(?{ print "$1 $2 $3\n" })(*FAIL)/'
... program outputs only five variants and "0 1 234" disappears.

I see the same behavior on Strawberries 5.10.1.5, 5.12.3.0, 5.14.4.1 and ActiveState 5.8.9. (Have to use  (?!) in place of  (*FAIL) previous to 5.10, but these patterns are exactly equivalent.)

Why is this behavior seen? Dunno. Will have to think about this a bit.

Also changing the first capture to non-greedy should not affect the number of results (regexp is anchored on both sides), yet it does (for z? case, things work OK without it):

perl -e '"01234" =~ /^(.+?)(.+)((?:.z?)+)$(?{ print "$1 $2 $3\n" })(*F
+AIL)/'
0 123 4
0 12 34
0 1 234
01 23 4
[download]

It looks like a bug. It smells like a bug. It ate my whole project. But... Is it a bug?

Is it a bug?

If you enable re 'debug', then you'll get to see what's going on in detail. The compiled regexes differ in much the way you'd expect:

Final program:                               Final program:
   1: BOL (2)                                   1: BOL (2)
   2: OPEN1 (4)                                 2: OPEN1 (4)
   4:   PLUS (6)                                4:   PLUS (6)
   5:     REG_ANY (0)                           5:     REG_ANY (0)
   6: CLOSE1 (8)                                6: CLOSE1 (8)
   8: OPEN2 (10)                                8: OPEN2 (10)
  10:   PLUS (12)                              10:   PLUS (12)
  11:     REG_ANY (0)                          11:     REG_ANY (0)
  12: CLOSE2 (14)                              12: CLOSE2 (14)
  14: OPEN3 (16)                               14: OPEN3 (16)
  16:   PLUS (18)                              16:   CURLYX[2] {1,3276
+7} (24)
  17:     REG_ANY (0)                          18:     REG_ANY (19)
                                               19:     CURLY {0,1} (23
+)
                                               21:       EXACT <z> (0)
                                               23:   WHILEM[1/1] (0)
                                               24:   NOTHING (25)
  18: CLOSE3 (20)                              25: CLOSE3 (27)
  20: EOL (21)                                 27: EOL (28)
  21: EVAL (23)                                28: EVAL (30)
  23: OPFAIL (24)                              30: OPFAIL (31)
  24: END (0)                                  31: END (0)
[download]

Comparing the traces is harder.

For

perl -e 'use re "debug"; "01234" =~ /^(.+)(.+)((?:.z?)+)$(?{ print "$1
+ $2 $3\n" })(*FAIL)/' 2>&1 | less
[download]

for the missing "0 1 234" the trace has

whilem: (cache) already tried at this position...
failed...
[download]

and my wild guess is that it cached the effect of (*FAIL) and not the real failure of a regexp, hence EVAL is skipped and last result is missed. I didn't find the way to disable the cache with "use re".

To add to the previous post: For

perl -e 'use re "debug"; "01234" =~ /^(.+?)(.+)((?:.z?)+)$(?{ print "$
+1 $2 $3\n" })(*FAIL)/' 2>&1 | less
[download]

two results are missing and there are two entries

whilem: (cache) already tried at this position...
failed...
[download]

It figures.

(had to register to see past Anonymous Monk's default depth limit)

For the record: I'm convinced it's not a bug but a feature, and asked to close my bug report (which is rejected now).

I also figured out that it's possible to disable caching by making (*FAIL) conditional:

perl -e '"01234" =~ /^(.+?)(.+)((?:.z?)+)$(?(?{ print "$1 $2 $3\n" })(
+*FAIL)|(*ACCEPT))/'
[download]

(?(?{ print "$1 $2 $3\n" })(*FAIL)|(*ACCEPT))
[download]

I don't understand the purpose of the  (*ACCEPT) false clause in the quoted code fragment. Because the print will always return true (unless there's some terrible I/O failure :), this clause will never be executed. The following versions of the code (with and without (*ACCEPT)) test the same in all 5.10+ Perl versions I have in captivity (see this):

use 5.010;  # need (?(?{ code }) pattern)

use Test::More 'no_plan';
use Test::NoWarnings;

note 'perl version ', $];

for my $rw ('.+', '.+ w?') {
    for my $rx ('.+', '.+ x?') {
        for my $ry ('.+', '(?: . y?)+') {

            my $captures = qr{ ($rw) ($rx) ($ry) }xms;

            local our @ra;
            use re 'eval';
            '01234' =~ m{
              \A $captures \z
            # (?(?{ push @ra, [ $1, $2, $3 ] }) (*F) | (*ACCEPT))
              (?(?{ push @ra, [ $1, $2, $3 ] }) (*F))
              }xms;

            is_deeply \@ra,
              [ [ qw(012  3  4) ],
                [ qw(01  23  4) ],
                [ qw(01  2  34) ],
                [ qw(0  123  4) ],
                [ qw(0  12  34) ],
                [ qw(0  1  234) ],
                ], $captures;

            }  # end for $ry
        }  # end for $rx
    }  # end for $rw

done_testing;
[download]

(Of course, the push statement always returns true.) The true magick seems to reside in the use of the regex conditional expression.

Output:

c:\@Work\Perl\monks\kroki>perl permute_via_regex_1.pl
# perl version 5.014004
ok 1 - (?^msx: (.+) (.+) (.+) )
ok 2 - (?^msx: (.+) (.+) ((?: . y?)+) )
ok 3 - (?^msx: (.+) (.+ x?) (.+) )
ok 4 - (?^msx: (.+) (.+ x?) ((?: . y?)+) )
ok 5 - (?^msx: (.+ w?) (.+) (.+) )
ok 6 - (?^msx: (.+ w?) (.+) ((?: . y?)+) )
ok 7 - (?^msx: (.+ w?) (.+ x?) (.+) )
ok 8 - (?^msx: (.+ w?) (.+ x?) ((?: . y?)+) )
1..8
ok 9 - no warnings
1..9
[download]

Of course, your final thought still holds true: none of this is guaranteed against future regex engine optimizations and other "improvements"!

---

Give a man a fish:  <%-{-{-{-<