Interestingly, benchmarking shows that the 2nd form of the code you show here is actually *slower* than the 1st.

I don't know if foreach(array_keys(...) as ...) has already been optimized internally, but in any case, it doesn't seem that this suggested optimization is necessary. Recommend this ticket can be closed.

As long as iterate thru the whole dataset, otherwise it can be orders of magnitude slower.

$c = 20000;

// create test array
$arr = [];
for ($i = 0; $i < 1000; $i++) {
    $arr[$i . 'x'] = $i . 'y'; // use sting keys, they may be slower than numeric ones
}

$t = microtime(true);
for ($i = 0; $i < $c; $i++) {
    foreach ($arr as $k => $ignore) {
        $u = $k . '.'; // use key
        break;
    }
}
var_dump(round(microtime(true) - $t, 6));

$t = microtime(true);
for ($i = 0; $i < $c; $i++) {
    foreach (array_keys($arr) as $k) {
        $u = $k . '.'; // use key
        break;
    }
}
var_dump(round(microtime(true) - $t, 6));