_mm_loadu_epi8
q[0] = _mm_loadu_epi8(in);
cv = _mm_loadu_epi8(iv);
q[0] = cv ^ _mm_loadu_epi8(in);
iv = _mm_loadu_epi8(ivp);
cv = _mm_loadu_epi8(in + nbytes - 16);
q[i] = _mm_loadu_epi8(in + nbytes - 16*n + 16*i);
cv = _mm_loadu_epi8(in + nbytes - 16);
q[6] = _mm_loadu_epi8(in + nbytes + 16*6);
q[5] = _mm_loadu_epi8(in + nbytes + 16*5);
q[4] = _mm_loadu_epi8(in + nbytes + 16*4);
q[3] = _mm_loadu_epi8(in + nbytes + 16*3);
q[2] = _mm_loadu_epi8(in + nbytes + 16*2);
q[1] = _mm_loadu_epi8(in + nbytes + 16*1);
q[0] = _mm_loadu_epi8(in + nbytes + 16*0);
cv = _mm_loadu_epi8(in + nbytes + 16*6);
cv = _mm_loadu_epi8(in + nbytes + 16*5);
cv = _mm_loadu_epi8(in + nbytes + 16*4);
cv = _mm_loadu_epi8(in + nbytes + 16*3);
cv = _mm_loadu_epi8(in + nbytes + 16*2);
cv = _mm_loadu_epi8(in + nbytes + 16*1);
cv = _mm_loadu_epi8(in + nbytes + 16*0);
cv = _mm_loadu_epi8(in + nbytes - 16);
_mm_storeu_epi8(t, aes_sse2_4x32_xts_update(_mm_loadu_epi8(t)));
t[0] = _mm_loadu_epi8(tweak);
q[i] = _mm_loadu_epi8(in + 16*i) ^ t[i];
q[i] = _mm_loadu_epi8(in + 16*i) ^ t[i];
t[0] = _mm_loadu_epi8(tweak);
q[i] = _mm_loadu_epi8(in + 16*i) ^ t[i];
q[i] = _mm_loadu_epi8(in + 16*i) ^ t[i];
q[0] = _mm_loadu_epi8(auth);
q[0] ^= _mm_loadu_epi8(in);
q[0] = _mm_loadu_epi8(authctr);
q[0] ^= _mm_loadu_epi8(in);
_mm_storeu_epi8(out, _mm_loadu_epi8(in) ^ q[2]);
q[2] = _mm_loadu_epi8(authctr);
b = _mm_loadu_epi8(in) ^ q[0];
q[0] = _mm_loadu_epi8(in);
rk = aes_schedule_transform(_mm_loadu_epi8(key), ipt);
rk = aes_schedule_transform(_mm_loadu_epi8(key + 8), ipt);
rk = aes_schedule_transform(_mm_loadu_epi8(key + 16), ipt);
ork = _mm_loadu_epi8(key);
rk = aes_schedule_transform(_mm_loadu_epi8(key + 8), ipt);
rk = aes_schedule_transform(_mm_loadu_epi8(key + 16), ipt);
return _mm_loadu_epi8(in);