Reduce array_diff() long fast path overhead

Check the first value's type inline before calling php_array_diff_long(),
and scan the exclude arrays before the first array so sparse or non-int
exclusions fall back without a full scan of the first array. Allocate the
bitset with ZEND_BITSET_ALLOCA like the other bitsets in this file. Add a
test building a bitset of more than one word, a NEWS entry, and clarify
that only the values to exclude need to be dense.

Suggested by @ArtUkrainskiy.

Author: mehmetcansahin

NEWS

@@ -125,6 +125,8 @@ PHP                                                                        NEWS
     in a php://filter URL. (Sjoerd Langkemper)
   . Improved performance of array_chunk() when not preserving keys.
     (mehmetcansahin)
+  . Improved performance of array_diff() for integer-only arrays when the
+    values to exclude are densely distributed. (mehmetcansahin)
 
 - URI:
   . Fix casing of enum cases in UriHostType and UrlHostType to match the RFC

UPGRADING

@@ -107,8 +107,8 @@ PHP 8.7 UPGRADE NOTES
   . Improved performance of strcspn() and strpbrk() when searching longer strings
     with a single-byte character mask.
   . Improved performance of array_chunk() when not preserving keys.
-  . Improved performance of array_diff() for integer-only arrays with densely
-    distributed values.
+  . Improved performance of array_diff() for integer-only arrays when the
+    values to exclude are densely distributed.
   . Improved performance of array_splice() when inserting without removing
     elements.
   . Improved performance of str_rot13().

ext/standard/array.c

@@ -6072,6 +6072,18 @@ static zend_always_inline bool php_array_values_are_all_long(HashTable *ht)
 	return true;
 }
 
+static zend_always_inline bool php_array_first_value_is_long(HashTable *ht)
+{
+	zval *value;
+
+	ZEND_HASH_FOREACH_VAL(ht, value) {
+		ZVAL_DEREF(value);
+		return Z_TYPE_P(value) == IS_LONG;
+	} ZEND_HASH_FOREACH_END();
+
+	return false;
+}
+
 static zend_never_inline bool php_array_diff_long(zval *args, uint32_t argc, uint32_t exclude_count, zval *return_value)
 {
 	zend_bitset exclude;
@@ -6082,12 +6094,7 @@ static zend_never_inline bool php_array_diff_long(zval *args, uint32_t argc, uin
 	const uint64_t max_range = (uint64_t) exclude_count * 8;
 	uint32_t range, bitset_len;
 	uint32_t i;
-
-	/* The copy loop below cannot cheaply undo a partially built result, so
-	 * the first array must be checked upfront. */
-	if (!php_array_values_are_all_long(Z_ARRVAL(args[0]))) {
-		return false;
-	}
+	ALLOCA_FLAG(use_heap);
 
 	/* Check the value types and determine whether the values are dense enough
 	 * for a bitset. Limiting the range to eight bits per exclude element keeps
@@ -6111,9 +6118,16 @@ static zend_never_inline bool php_array_diff_long(zval *args, uint32_t argc, uin
 		} ZEND_HASH_FOREACH_END();
 	}
 
+	/* Checked after the exclude arrays, whose type and range checks can rule
+	 * the bitset out after a few values. The copy loop below cannot cheaply undo
+	 * a partially built result, so the first array must be checked before it. */
+	if (!php_array_values_are_all_long(Z_ARRVAL(args[0]))) {
+		return false;
+	}
+
 	range = (uint32_t) span + 1;
 	bitset_len = zend_bitset_len(range);
-	exclude = safe_emalloc(bitset_len, ZEND_BITSET_ELM_SIZE, 0);
+	exclude = ZEND_BITSET_ALLOCA(bitset_len, use_heap);
 	zend_bitset_clear(exclude, bitset_len);
 
 	for (i = 1; i < argc; i++) {
@@ -6140,7 +6154,7 @@ static zend_never_inline bool php_array_diff_long(zval *args, uint32_t argc, uin
 		}
 	} ZEND_HASH_FOREACH_END();
 
-	efree(exclude);
+	free_alloca(exclude, use_heap);
 	return true;
 }
 
@@ -6244,7 +6258,8 @@ PHP_FUNCTION(array_diff)
 		RETURN_THROWS();
 	}
 
-	if (php_array_diff_long(args, argc, (uint32_t)num, return_value)) {
+	if (php_array_first_value_is_long(Z_ARRVAL(args[0]))
+			&& php_array_diff_long(args, argc, (uint32_t)num, return_value)) {
 		return;
 	}
 

ext/standard/tests/array/array_diff_long_values.phpt

@@ -43,6 +43,16 @@ for ($i = 1; $i <= 256; $i++) {
     $sparseValues[] = $i * 65536;
 }
 var_dump(array_diff([1, 2], $sparseValues) === [1, 2]);
+
+$wideExclude = [];
+for ($i = 0; $i < 100; $i++) {
+    $wideExclude[] = $i * 7;
+}
+$wideInput = range(-1, 700);
+var_dump(array_diff($wideInput, $wideExclude) === array_filter(
+    $wideInput,
+    fn($v) => $v < 0 || $v > 693 || $v % 7 !== 0,
+));
 ?>
 --EXPECT--
 {"0":1,"2":3}
@@ -61,3 +71,4 @@ bool(true)
 {"0":1,"2":3}
 int(9)
 bool(true)
+bool(true)