ext/opcache: Use NEON for zend_adler32() on AArch64

UPGRADING

@@ -173,6 +173,9 @@ PHP 8.7 UPGRADE NOTES
     target does not keep is now released at that assignment instead of after
     the last one.
 
+- Opcache:
+  . Improved performance of file cache checksum verification on AArch64.
+
 - Standard:
   . Improved performance of strcspn() and strpbrk() when searching longer strings
     with a single-byte character mask.

ext/opcache/tests/file_cache_checksum.phpt

@@ -0,0 +1,66 @@
+--TEST--
+File cache checksum is the Adler-32 of the cached script
+--SKIPIF--
+<?php
+if (PHP_INT_SIZE != 8) die('skip 64-bit only');
+@mkdir(__DIR__ . '/file_cache_checksum_cache', 0777, true);
+?>
+--INI--
+opcache.enable=1
+opcache.enable_cli=1
+opcache.jit=disable
+opcache.file_cache="{PWD}/file_cache_checksum_cache"
+opcache.file_cache_only=1
+opcache.file_update_protection=0
+--EXTENSIONS--
+opcache
+--FILE--
+<?php
+
+/* Spans several Adler-32 blocks of 5552 bytes, with every byte value and a
+ * long run of 0xff, the worst case for the sums. */
+$bytes = str_repeat(implode('', array_map('chr', range(0, 255))), 64) . str_repeat("\xff", 20011);
+$file = __DIR__ . '/file_cache_checksum.inc';
+file_put_contents($file, "<?php\nreturn '" . addcslashes($bytes, "\\'") . "';\n");
+var_dump(opcache_compile_file($file));
+
+$iterator = new RecursiveIteratorIterator(
+    new RecursiveDirectoryIterator(__DIR__ . '/file_cache_checksum_cache', RecursiveDirectoryIterator::SKIP_DOTS)
+);
+foreach ($iterator as $fileinfo) {
+    if (str_ends_with($fileinfo->getFilename(), 'file_cache_checksum.inc.bin')) {
+        $bin = file_get_contents($fileinfo->getPathname());
+    }
+}
+
+/* The header is zend_file_cache_metainfo (zend_file_cache.c): 80 bytes, with
+ * mem_size and str_size at offset 40 and the checksum of the rest at 72. */
+['mem' => $mem, 'str' => $str] = unpack('Qmem/Qstr', $bin, 40);
+var_dump(strlen($bin) === 80 + $mem + $str && $mem + $str > 3 * 5552);
+var_dump(sprintf('%08x', unpack('L', $bin, 72)[1]) === hash('adler32', substr($bin, 80)));
+
+?>
+--CLEAN--
+<?php
+function removeDirRecursive($dir) {
+    if (!is_dir($dir)) return;
+    $iterator = new RecursiveIteratorIterator(
+        new RecursiveDirectoryIterator($dir, RecursiveDirectoryIterator::SKIP_DOTS),
+        RecursiveIteratorIterator::CHILD_FIRST
+    );
+    foreach ($iterator as $fileinfo) {
+        if ($fileinfo->isDir()) {
+            @rmdir($fileinfo->getRealPath());
+        } else {
+            @unlink($fileinfo->getRealPath());
+        }
+    }
+    @rmdir($dir);
+}
+removeDirRecursive(__DIR__ . '/file_cache_checksum_cache');
+@unlink(__DIR__ . '/file_cache_checksum.inc');
+?>
+--EXPECT--
+bool(true)
+bool(true)
+bool(true)

ext/opcache/zend_accelerator_util_funcs.c

@@ -457,7 +457,34 @@ static zend_always_inline void adler32_do16_loop(unsigned char *buf, const unsig
 	unsigned int s1 = *s1_out;
 	unsigned int s2 = *s2_out;
 
-#ifdef XSSE2
+#if defined(__aarch64__) || defined(_M_ARM64)
+	/* zend_simd.h maps the SSE2 loop below to NEON too, but its _mm_madd_epi16()
+	 * and _mm_sad_epu8() are slow there, so use NEON directly.
+	 * For each 16-byte block: s2 += 16 * s1 + sum((16 - i) * buf[i]), then
+	 * s1 += sum(buf[i]). The weighted products fit in 16 bits
+	 * (255 * 16 + 255 * 8), and all 32-bit sums wrap like the scalar code's, so
+	 * the result is identical for any input. */
+	static const uint8_t weights[16] = {16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1};
+	const uint8x8_t weights_lo = vld1_u8(weights);
+	const uint8x8_t weights_hi = vld1_u8(weights + 8);
+	const unsigned int blocks = (unsigned int) ((end - buf) / 16);
+	uint32x4_t weighted_sum = vdupq_n_u32(0);
+	uint32x4_t byte_sum = vdupq_n_u32(0);
+	uint32x4_t byte_sum_at_block_start = vdupq_n_u32(0);
+
+	do {
+		const uint8x16_t read = vld1q_u8(buf);
+		byte_sum_at_block_start = vaddq_u32(byte_sum_at_block_start, byte_sum);
+		uint16x8_t products = vmull_u8(vget_low_u8(read), weights_lo);
+		products = vmlal_u8(products, vget_high_u8(read), weights_hi);
+		weighted_sum = vpadalq_u16(weighted_sum, products);
+		byte_sum = vpadalq_u16(byte_sum, vpaddlq_u8(read));
+		buf += 16;
+	} while (buf != end);
+
+	s2 += 16 * (blocks * s1 + vaddvq_u32(byte_sum_at_block_start)) + vaddvq_u32(weighted_sum);
+	s1 += vaddvq_u32(byte_sum);
+#elif defined(XSSE2)
 	const __m128i zero = _mm_setzero_si128();
 
 	__m128i accumulate_s2 = zero;