+ size_t i = 0;
+#if __SSE2__
+ const uint8_t c_expanded[16] = {
+ c[0], c[1], c[0], c[1], c[0], c[1], c[0], c[1],
+ c[0], c[1], c[0], c[1], c[0], c[1], c[0], c[1]
+ };
+ __m128i cc = *(__m128i *)c_expanded;
+ __m128i *out = (__m128i *)s;
+
+ for ( ; i < (n & ~15); i += 16) {
+ _mm_storeu_si128(out++, cc);
+ _mm_storeu_si128(out++, cc);
+ }
+
+ s = (uint8_t *)out;
+#endif
+ for ( ; i < n; ++i) {