zend_memnstr: Use a compact Sunday shift table for short needles

Zend/zend_operators.c

@@ -3853,7 +3853,7 @@ static zend_always_inline void zend_memnstr_ex_pre(unsigned int td[], const char
 }
 /* }}} */
 
-ZEND_API const char* ZEND_FASTCALL zend_memnstr_ex(const char *haystack, const char *needle, size_t needle_len, const char *end) /* {{{ */
+static zend_never_inline const char* ZEND_FASTCALL zend_memnstr_sunday(const char *haystack, const char *needle, size_t needle_len, const char *end)
 {
 	unsigned int td[256];
 	size_t i;
@@ -3885,7 +3885,52 @@ ZEND_API const char* ZEND_FASTCALL zend_memnstr_ex(const char *haystack, const c
 
 	return NULL;
 }
-/* }}} */
+
+static zend_never_inline const char* ZEND_FASTCALL zend_memnstr_sunday_short(const char *haystack, const char *needle, size_t needle_len, const char *end)
+{
+	uint8_t td[256];
+	size_t i;
+	const char *p;
+
+	if (needle_len == 0 || (end - haystack) < needle_len) {
+		return NULL;
+	}
+
+	/* The dispatcher only sends short needles here, so every
+	 * Sunday shift fits in one byte. Avoid filling a 1 KiB table per word. */
+	memset(td, (unsigned char)(needle_len + 1), sizeof(td));
+	for (i = 0; i < needle_len; i++) {
+		td[(unsigned char)needle[i]] = (unsigned char)(needle_len - i);
+	}
+
+	p = haystack;
+	end -= needle_len;
+
+	while (p <= end) {
+		for (i = 0; i < needle_len; i++) {
+			if (needle[i] != p[i]) {
+				break;
+			}
+		}
+		if (i == needle_len) {
+			return p;
+		}
+		if (UNEXPECTED(p == end)) {
+			return NULL;
+		}
+		p += td[(unsigned char)(p[needle_len])];
+	}
+
+	return NULL;
+}
+
+ZEND_API const char* ZEND_FASTCALL zend_memnstr_ex(const char *haystack, const char *needle, size_t needle_len, const char *end)
+{
+	if (EXPECTED(needle_len < 32)) {
+		return zend_memnstr_sunday_short(haystack, needle, needle_len, end);
+	}
+	return zend_memnstr_sunday(haystack, needle, needle_len, end);
+}
 
 ZEND_API const char* ZEND_FASTCALL zend_memnrstr_ex(const char *haystack, const char *needle, size_t needle_len, const char *end) /* {{{ */
 {

ext/standard/tests/strings/str_search_short_sunday.phpt

@@ -0,0 +1,35 @@
+--TEST--
+Forward substring search with binary needles across the compact Sunday table boundary
+--FILE--
+<?php
+$checked = 0;
+$failed = 0;
+foreach ([9, 15, 16, 30, 31, 32, 33, 254, 255, 256] as $length) {
+    // 0xfe occurs only at the end of the planted needle. The other bytes
+    // exercise unsigned shift-table indexing, embedded NULs and dense prefixes.
+    $needle = substr(str_repeat("\0\x80\xffa", $length), 0, $length - 1) . "\xfe";
+    foreach ([1023, 1024, 1025, 8191, 8192, 8193] as $size) {
+        $base = substr(str_repeat("\0\x80\xffa", $size), 0, $size);
+        foreach ([
+            [$base, false, [$base]],
+            // Bytes absent from the needle use the default needle_len + 1 shift.
+            [str_repeat('z', $size), false, [str_repeat('z', $size)]],
+            [$needle . $base, 0, ['', $base]],
+            [$base . $needle, $size, [$base, '']],
+            // After the first match, the next successful search uses a short remainder.
+            [$base . $needle . 'gap' . $needle . 'tail', $size, [$base, 'gap', 'tail']],
+        ] as [$haystack, $position, $parts]) {
+            $failed += strpos($haystack, $needle) !== $position;
+            $failed += str_contains($haystack, $needle) !== ($position !== false);
+            $failed += substr_count($haystack, $needle) !== (count($parts) - 1);
+            $failed += explode($needle, $haystack) !== $parts;
+            $checked++;
+        }
+    }
+}
+echo "checked $checked\n";
+echo $failed ? "FAILED ($failed)\n" : "ALL OK\n";
+?>
+--EXPECT--
+checked 300
+ALL OK