Quick test

src/WebSocketProtocol.h

@@ -25,6 +25,9 @@
 #include <cstdlib>
 #include <string_view>
 
+// Quick test
+#include <immintrin.h>
+
 namespace uWS {
 
 /* We should not overcomplicate these */
@@ -114,24 +117,59 @@ T cond_byte_swap(T value) {
 // Optimized for predominantly 7-bit content by Alex Hultman, 2016
 // Licensed as Zlib, like the rest of this project
 // This runs about 40% faster than simdutf with g++ -mavx
-static bool isValidUtf8(unsigned char *s, size_t length)
-{
-    for (unsigned char *e = s + length; s != e; ) {
-        if (s + 16 <= e) {
-            uint64_t tmp[2];
-            memcpy(tmp, s, 16);
-            if (((tmp[0] & 0x8080808080808080) | (tmp[1] & 0x8080808080808080)) == 0) {
-                s += 16;
-                continue;
+static bool isValidUtf8(unsigned char *s, size_t length) {
+    auto firstUtf8EscapeByte = [](unsigned char *s, unsigned char *e) {
+        // Align
+        if (s + 32 <= e) {
+            int mask = _mm256_movemask_epi8(_mm256_loadu_si256((const __m256i*)s));
+            if (mask) {
+                return s + __builtin_ctz(mask);
+            }
+            s += 32 - ((uintptr_t)s % 32);
+        } else {
+            // Worst path
+            while (s < e) {
+                if (*s & 0x80) {
+                    return s;
+                }
+                s++;
             }
+            return e;
         }
 
-        while (!(*s & 0x80)) {
-            if (++s == e) {
-                return true;
+        while (s + 128 <= e) {
+            // Aligned
+            int mask = _mm256_movemask_epi8(_mm256_load_si256((const __m256i*)s));
+            if (mask) {
+                return s + __builtin_ctz(mask);
+            }
+            s += 32;
+            mask = _mm256_movemask_epi8(_mm256_load_si256((const __m256i*)s));
+            if (mask) {
+                return s + __builtin_ctz(mask);
             }
+            s += 32;
+            mask = _mm256_movemask_epi8(_mm256_load_si256((const __m256i*)s));
+            if (mask) {
+                return s + __builtin_ctz(mask);
+            }
+            s += 32;
+            mask = _mm256_movemask_epi8(_mm256_load_si256((const __m256i*)s));
+            if (mask) {
+                return s + __builtin_ctz(mask);
+            }
+            s += 32;
         }
 
+        // Exit
+        while ((*s & 0x80) == 0 && s < e) {
+            s++;
+        }
+        return s;
+    };
+
+    for (unsigned char *e = s + length; (s = (unsigned char *) firstUtf8EscapeByte(s, e)) != e; ) {
+
         if ((s[0] & 0x60) == 0x40) {
             if (s + 1 >= e || (s[1] & 0xc0) != 0x80 || (s[0] & 0xfe) == 0xc0) {
                 return false;