#define TOB(x) ((x)*4*0x01010101 + 0x03020100) /* ..or not TOB */

{

const __m128i sixteen = _mm_set1_epi8(16);

const __m128i t0 = _mm_shuffle_epi8(src1, sel);

const __m128i s1 = _mm_shuffle_epi8(src2, _mm_sub_epi8(sel, sixteen));

const __m128i mask = _mm_or_si128(_mm_cmpeq_epi8(sel, sixteen),

_mm_cmpgt_epi8(sel, sixteen)); /* (>=16) = 0xff : 00 */

return _mm_blendv_epi8(t0, s1, mask);

#define LOAD_MSG_0_1(buf) \

buf = _mm_perm_epi8(m0, m1, _mm_set_epi32(TOB(6),TOB(4),TOB(2),TOB(0)) );