Use AVX2 when unmasking websockets frames

This already used SSE2, but let's use a wider vector if available to
get this over more quickly.

src/lib/lwan-websocket.c

@@ -26,7 +26,7 @@
 #include <sys/socket.h>
 
 #if defined(__x86_64__)
-#include <emmintrin.h>
+#include <immintrin.h>
 #endif
 
 #include "lwan-io-wrappers.h"
@@ -156,19 +156,39 @@ static void unmask(char *msg, size_t msg_len, char mask[static 4])
     if (sizeof(void *) == 8) {
         const uint64_t mask64 = (uint64_t)mask32 << 32 | mask32;
 
-#if defined(__x86_64__)
+#if defined(__AVX2__)
+        const size_t len256 = msg_len / 32;
+        if (len256) {
+            const __m256i mask256 =
+                _mm256_setr_epi64x((int64_t)mask64, (int64_t)mask64,
+                                   (int64_t)mask64, (int64_t)mask64);
+            for (size_t i = 0; i < len256; i++) {
+                __m256i v = _mm256_loadu_si256((__m256i *)msg);
+                _mm256_storeu_si256((__m256i *)msg,
+                                    _mm256_xor_si256(v, mask256));
+                msg += 32;
+            }
+
+            msg_len = (size_t)(msg_end - msg);
+        }
+#endif
+
+#if defined(__SSE2__)
         const size_t len128 = msg_len / 16;
         if (len128) {
-            const __m128i mask128 = _mm_setr_epi64((__m64)mask64, (__m64)mask64);
+            const __m128i mask128 =
+                _mm_setr_epi64((__m64)mask64, (__m64)mask64);
             for (size_t i = 0; i < len128; i++) {
                 __m128i v = _mm_loadu_si128((__m128i *)msg);
                 _mm_storeu_si128((__m128i *)msg, _mm_xor_si128(v, mask128));
                 msg += 16;
             }
+
+            msg_len = (size_t)(msg_end - msg);
         }
 #endif
 
-        const size_t len64 = (size_t)((msg_end - msg) / 8);
+        const size_t len64 = msg_len / 8;
         for (size_t i = 0; i < len64; i++) {
             uint64_t v = string_as_uint64(msg);
             v ^= mask64;