_mm_loadu_si128
r0 = _mm_loadu_si128((const __m128i *)c);
r1 = _mm_loadu_si128((const __m128i *)k);
r2 = _mm_loadu_si128((const __m128i *)k + 1);
r3 = _mm_loadu_si128((const __m128i *)in);
return _mm_loadu_si128(p);
in0 = _mm_loadu_si128((const __m128i *)chacha_const32);
in1 = _mm_loadu_si128((const __m128i *)k);
in2 = _mm_loadu_si128((const __m128i *)k + 1);
in0 = _mm_loadu_si128((const __m128i *)chacha_const32);
in1 = _mm_loadu_si128((const __m128i *)k);
in2 = _mm_loadu_si128((const __m128i *)k + 1);
r0 ^= _mm_loadu_si128((const __m128i *)p + 0);
r1 ^= _mm_loadu_si128((const __m128i *)p + 1);
r2 ^= _mm_loadu_si128((const __m128i *)p + 2);
r3 ^= _mm_loadu_si128((const __m128i *)p + 3);
r0 = in0 = _mm_loadu_si128((const __m128i *)c);
r1 = in1 = _mm_loadu_si128((const __m128i *)k);
r2 = in2 = _mm_loadu_si128((const __m128i *)k + 1);
r3 = in3 = _mm_loadu_si128((const __m128i *)in);