PHP RFC: Normalize arrays' "auto-increment" value on copy on write


If two arrays are equal/identical, they should remain equal/identical arrays after the same array_push(..., $val) call is executed on both of them:

assert($array1 === $array2); // identical/equal
$array1[] = $array2[] = 123;
assert($array1 === $array2); // still identical/equal

This is currently not guaranteed, and because of arrays' all-doing nature, it is not possible to always enforce this property -- but it should be in some dangerous cases, namely when functions from (potential) different authors are interacting.

When an array is assigned to a new reference, and it is copied, before a modification, due to the copy-on-write behavior, it will result in an array that is identical in any way to the original one, inclusive of its “auto-increment” value:

$array1 = [0, 1, 2];
unset($array1[1], $array1[2]);
$array2 = $array1;
assert($array2 === [0]);
$array2[] = "push"; // triggers COW and then pushes the new entry
// Array
// (
//     [0] => 0
//     [3] => push
// )

This happens also between different function scopes. Our functions can receive “broken” array-lists from third-parties that only appear to be well-indexed, but that in reality are not, because they were misused during their lifetime (classic example, it was used unset($array[$lastIndex]) on them, instead of array_pop($array)).

As result of that, despite “copy on write”, the value-type semantics, and even a different scope, the following assertion can fail in some cases:

function test(array $array){
    if($array === [0, 1, 2]){
        $array[] = 3;
        assert($array === [0, 1, 2, 3]);
// For example:
$poison = [0, 1, 2, 3];


This RFC proposes to reset the “auto-increment” value in copies triggered by “copy on write”, in order to guarantee a deterministic behavior to foreign scopes especially. The “auto-increment” value of the new variable reference must be equivalent to the “auto-increment” value that the array would have if it was re-created entry by entry, as follows:

$array_copy = [];
foreach($array as $key => $value){
    $array_copy[$key] => $value;

The reset is not limited to new function scopes but any new by-value reference:

$array = [0, 1, 2, 3];
unset($array[3], $array[2]);
$arrayCopy = $array;
$arrayCopy[] = 2;
assert($arrayCopy === [0, 1, 2]); // this assertion must pass; it doesn't currently

Backward Incompatible Changes

This change is not backward compatible; code relying on the “auto-increment” value being remembered between copies of copy-on-write will break. However, the proposed change should be considered a bug-fix, rather than a behavior change; it offers protection against array-lists that were misused with unset() instead of array_pop/_splice/_shift and thus will only affect code that is already a candidate for improvements. Furthermore, the “auto-increment” value is copied inconsistently, when the array is empty:

$a = [0, 1];
$b = $a;
$b[] = 2;
// $b is [0 => 0, 2 => 2]
$a = [0, 1];
unset($a[0], $a[1]);
$b = $a;
$b[] = 2;
// $b is [0 => 2], rather than [2 => 2]

The proposed change would make the behavior consistent and safer.

Proposed PHP Version(s)


Proposed Voting Choices

Vote will require 2/3 majority


rfc/normalize-array-auto-increment-on-copy-on-write.txt · Last modified: 2019/06/28 21:13 by wesnetmo