[nicolas-grekas]

zend_hash_iterator_add() looks for a free slot in EG(ht_iterators) starting from slot 0 every time, so a foreach by reference, a foreach over object properties, an ArrayIterator or array_walk() pays one probe per iterator that's already alive.

This adds EG(ht_iterators_free_hint): all slots below it are in use, zend_hash_iterator_add() starts looking from there and zend_hash_iterator_del() lowers it. The returned slot is still the lowest free one, so slot indexes and EG(ht_iterators_used) don't change. 51451ab tried scanning backwards instead, and was reverted the next day by ceb77b6; keeping first-fit avoids changing which slot gets reused.

Found while profiling #23945: a recursive walk of 17.6k Twig nodes, each iterated through new ArrayIterator($this->nodes), spends 2.4M instructions in zend_hash_iterator_add()/del(). That's 1.5M with this patch, -1.2% on the whole walk (callgrind).

Instructions per foreach ($a as &$v) {} over a 1-element array, nested in d-1 other loops by reference: 364 -> 365 at d=1, 389 -> 370 at d=5, 464 -> 369 at d=20. Same savings per loop for ArrayIterator, foreach over properties and array_walk().

[nicolas-grekas]

Rebased. I also removed the NEWS and UPGRADING entries, since they were the only cause of the conflicts.

@ndossche @iluuu1994 would one of you have time for a look? Nothing changes when a single iterator is alive. When loops are nested, zend_hash_iterator_add() doesn't scan the slots in use anymore. An empty foreach by reference then needs 20% fewer instructions at depth 20.