update fix, and fix other bugs:
```
	var_push_dtor_no_addref(var_hash, rval);
	}
	*rval = *rval_ref;
+	Z_ADDREF_PP(rval_ref);
+	if (Z_REFCOUNT_PP(rval) > 1) {
+		SEPARATE_ZVAL_IF_NOT_REF(rval);
+	}
	Z_ADDREF_PP(rval);
	Z_SET_ISREF_PP(rval);
```

I don't think this qualifies as security issue - since you need specially crafter destructor.

In PHP 7.0+ this results in:

array(1) {
  [0]=>
  object(ryat)#1 (2) {
    ["ryat"]=>
    array(1) {
      [0]=>
      *RECURSION*
    }
    ["chtg"]=>
    &array(1) {
      [0]=>
      *RECURSION*
    }
  }
}

I believe this output is correct, as the following code (to which the unserialize should roughly correspond) has the same output:

$a = [];
$a[0] = 1;
$o = new ryat;
$o->ryat =& $a[1];
$o->chtg = 2;
$a[1] = [$o];
unset($a[1]);
$a[1] = 3;
$a[2] =& $o->chtg;
unset($o);
gc_collect_cycles();
var_dump($a[2]);

This also shows that the issue is not related to unserialize(), but rather to GC with certain destructors. The GC problem has been fixed in 7.0 (and won't be fixed in 5.6), so closing here.

@nikic

Unfortunately, the bug was still unfixed in PHP 7.x, you can test the following code:
```
$poc = 'a:4:{i:0;i:1;i:1;a:1:{i:0;O:4:"ryat":2:{s:4:"ryat";R:3;s:4:"chtg";i:2;}}i:1;i:3;i:2;R:5;}';
$out = unserialize($poc);
gc_collect_cycles();
for ($i = 0; $i < 5; $i++) {
	$v[$i] = 'hi'.$i;
}
var_dump($out[2]);

class ryat
{
	var $ryat;
	var $chtg;
	
	function __destruct()
	{
		$this->chtg = $this->ryat;
		$this->ryat = 1;
	}
}
```

The bug is related to GC with certain destructors, but trigged easily via unserialize().

Right, issue not fully fixed :( Here's a reduced testcase without unserialize (generates memory errors under valgrind):

<?php

class ryat {
    var $ryat;
    var $chtg;

    function __destruct() {
        $this->chtg = $this->ryat;
        $this->ryat = 1;
    }
}

$o = new ryat;
$o->ryat = $o;
$x =& $o->chtg;

unset($o);
gc_collect_cycles();
var_dump($x);

The problem is that the GC uses refcount to check whether a new reference has been added during destruction. However, in this case refcount of object(ryat) stays the same, while a reference to it still becomes externally visible.

@dmitry: Can you please take a look at this issue? The GC code for dealing with references added during __destruct() is not entirely correct, even in PHP 7, because the refcount check is insufficient to detect new outside reference. I don't see any simple fix for this problem :/