array_key_exists on $GLOBALS takes linear time in PHP 8.1.0+

Description

Compare the following in PHP 8.0.22 vs 8.1.9 (https://3v4l.org/91CFl)

<?php

for ($i = 0; $i < 100000; $i++) {
    $GLOBALS["a$i"] = rand(0, 1000000);
}

$start = microtime(true);
for ($runs = 1000; $runs > 0; $runs--) {
    array_key_exists('a1111111', $GLOBALS);
}
$end = microtime(true);
printf("[GLOBALS]array_key_exists() took %.5f seconds\n", $end - $start);

It takes 0.00002 seconds in 8.0.22, and 1.9 seconds in 8.1.9.

Further testing makes it clear that the speed of array_key_exists on $GLOBALS increases linearly with the size of $GLOBALS.

This regression doesn't affect array_key_exists on non-$GLOBALS arrays, and doesn't affect the speed of isset($GLOBALS[$foo]).

The commit that introduced the slowdown was https://github.com/php/php-src/commit/3c68f38fdaec7a6f81f986a639bdd29716cf8f22, which restricted usage of $GLOBALS in various ways. I'm not sure if this particular side effect was known or intended.

array_key_exists is used somewhat often on $GLOBALS in MediaWiki, most prominently here.

The behavior can be mostly replaced by calling isset($GLOBALS[$foo]), which remains fast, although that has different semantics when the value is null. I'm not aware of a better way to get the full behavior of array_key_exists($foo, $GLOBALS) in constant-time in 8.1+.

PHP Version

PHP 8.1.9

Operating System

Ubuntu 20.04

Reproduziere den PHP 8.0.22/8.1.9-Benchmark aus dem Issue und untersuche anschließend den Commit 3c68f38fdaec7a6f81f986a639bdd29716cf8f22, der die Verwendung von $GLOBALS eingeschränkt hat. Die Arbeit ist abgeschlossen, wenn array_key_exists($foo, $GLOBALS) nicht mehr linear mit der Größe von $GLOBALS skaliert und dabei seine Semantik für Nullwerte beibehält; vergleiche das Ergebnis mit dem Verhalten von isset.