buffer: add SIMD Neon optimization for byteLength

Author: anonrig

src/node_buffer.cc

@@ -38,6 +38,14 @@
 #include <cstring>
 #include <climits>
 
+#if defined(__aarch64__) || defined(_M_ARM64)
+#define NODE_HAS_SIMD_NEON 1
+#endif
+
+#if NODE_HAS_SIMD_NEON
+#include <arm_neon.h>
+#endif
+
 #define THROW_AND_RETURN_UNLESS_BUFFER(env, obj)                            \
   THROW_AND_RETURN_IF_NOT_BUFFER(env, obj, "argument")                      \
 
@@ -741,6 +749,38 @@ void SlowByteLengthUtf8(const FunctionCallbackInfo<Value>& args) {
   args.GetReturnValue().Set(args[0].As<String>()->Utf8Length(env->isolate()));
 }
 
+#if NODE_HAS_SIMD_NEON
+uint32_t FastByteLengthUtf8(Local<Value> receiver,
+                            const v8::FastOneByteString& source) {
+  uint32_t answer{0};
+  const auto data = reinterpret_cast<const uint8_t*>(source.data);
+  const uint8x16_t mask = vdupq_n_u8(0x80);
+  uint8x16_t result_vector = vdupq_n_u8(0);
+  size_t i = 0;
+
+  for(; i < source.length; i += 16) {
+    // load 16 bytes from data
+    uint8x16_t values = vld1q_u8(data + i);
+
+    // extract the high bits using 0x80 mask
+    uint8x16_t high_bits = vcgeq_u8(values, mask);
+
+    // accumulate the high bits to result_vector
+    result_vector = vqaddq_u8(result_vector, high_bits);
+  }
+
+  // sum the elements in the result_vector
+  uint64x2_t sum64 = vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(result_vector)));
+
+  answer = vgetq_lane_u64(sum64, 0) + vgetq_lane_u64(sum64, 1);
+
+  for (; i < source.length; ++i) {
+    answer += (data[i] >> 7);
+  }
+
+  return answer + source.length;
+}
+#else
 uint32_t FastByteLengthUtf8(Local<Value> receiver,
                             const v8::FastOneByteString& source) {
   uint32_t result = 0;
@@ -752,6 +792,7 @@ uint32_t FastByteLengthUtf8(Local<Value> receiver,
   result += length;
   return result;
 }
+#endif
 
 static v8::CFunction fast_byte_length_utf8(
     v8::CFunction::Make(FastByteLengthUtf8));