Zend: add_function_array() fast paths for an empty operand.

[devnexen]

Skip the duplicate and merge when either operand is empty, unless sharing would change the internal pointer or next free index.

[devnexen]

Some numbers, release build without opcache, best of 7 runs, in ms.

8 elements, 1M iterations:

	hash before	hash after	list before	list after
$a + []	36.3	6.2	26.8	6.2
[] + $a	50.2	7.2	46.5	6.6
$x = []; $x += $a	51.5	10.2	47.4	9.7
$a + $b	86.0	89.1	44.3	44.7

100 elements, 100k iterations:

	hash before	hash after	list before	list after
$a + []	24.8	0.6	9.9	0.6
[] + $a	66.9	0.8	42.6	0.7
$x = []; $x += $a	63.6	1.0	42.6	1.0
$a + $b	83.4	84.7	26.1	26.4

$a + $b is there as a control, the difference is noise on my machine.

[staabm]

changed code uses a mix of tabs and spaces

[ndossche]

This looks so complicated and I'm still not convinced it's correct; I can't follow the soundness of the second case.

The first case definitely isn't correct. zend_array_dup() deals with references. So this behaviour changes:

$x = 1;
$b = ['a' => &$x];
$r = [] + $b;
unset($x);
$b['a'] = 5;
echo $r['a'];

[ndossche]

These comments make it sound like the internal pointer is the only thing that changes, which is misleading

[devnexen]

This looks so complicated and I'm still not convinced it's correct; I can't follow the soundness of the second case.

The first case definitely isn't correct. zend_array_dup() deals with references. So this behaviour changes:

$x = 1;
$b = ['a' => &$x];
$r = [] + $b;
unset($x);
$b['a'] = 5;
echo $r['a'];

ah good point, it is ... I ll see about these.