15720     __m128i a_sh, b_sh;  in vtrnq_s8()  local
15721 …a_sh = _mm_shuffle_epi8 (a, *(__m128i*)mask8_16_even_odd); //a0, a2, a4, a6, a8, a10, a12, a14, a1…  in vtrnq_s8()
15724 …r8x16.val[0] =  _mm_unpacklo_epi8(a_sh, b_sh); //(a0, b0, a2, b2, a4, b4, a6, b6, a8,b8, a10,b10, …  in vtrnq_s8()
15725 …r8x16.val[1] =  _mm_unpackhi_epi8(a_sh, b_sh); // (a1, b1, a3, b3, a5, b5, a7, b7, a9,b9, a11,b11,…  in vtrnq_s8()
15733     __m128i a_sh, b_sh;  in vtrnq_s16()  local
15734     a_sh = _mm_shuffle_epi8 (a, *(__m128i*) mask8_32_even_odd); //a0, a2, a4, a6,  a1, a3, a5, a7  in vtrnq_s16()
15736     v16x8.val[0] = _mm_unpacklo_epi16(a_sh, b_sh); //a0, b0, a2, b2, a4, b4, a6, b6  in vtrnq_s16()
15737     v16x8.val[1] = _mm_unpackhi_epi16(a_sh, b_sh); //a1, b1, a3, b3, a5, b5, a7, b7  in vtrnq_s16()
15746     __m128i a_sh, b_sh;  in vtrnq_s32()  local
15747     a_sh = _mm_shuffle_epi32 (a, 216); //a0, a2, a1, a3  in vtrnq_s32()
15750     v32x4.val[0] = _mm_unpacklo_epi32(a_sh, b_sh); //a0, b0, a2, b2  in vtrnq_s32()
15751     v32x4.val[1] = _mm_unpackhi_epi32(a_sh, b_sh); //a1, b1, a3,  b3  in vtrnq_s32()
15769     __m128 a_sh, b_sh;  in vtrnq_f32()  local
15770     a_sh = _mm_shuffle_ps (a, a, _MM_SHUFFLE(3,1, 2, 0)); //a0, a2, a1, a3, need to check endiness  in vtrnq_f32()
15773     f32x4.val[0] = _mm_unpacklo_ps(a_sh, b_sh); //a0, b0, a2, b2  in vtrnq_f32()
15774     f32x4.val[1] = _mm_unpackhi_ps(a_sh, b_sh); //a1, b1, a3,  b3  in vtrnq_f32()
15939     __m128i a_sh, b_sh;  in vuzpq_s8()  local
15940 …a_sh = _mm_shuffle_epi8 (a, *(__m128i*)mask8_16_even_odd); //a0, a2, a4, a6, a8, a10, a12, a14, a1…  in vuzpq_s8()
15943 …v8x16.val[0] = _mm_unpacklo_epi64(a_sh, b_sh); ///a0, a2, a4, a6, a8, a10, a12, a14,  b0, b2, b4, …  in vuzpq_s8()
15944 …v8x16.val[1] = _mm_unpackhi_epi64(a_sh, b_sh); //a1, a3, a5, a7, a9, a11, a13, a15,  b1, b3, b5, b…  in vuzpq_s8()
15952     __m128i a_sh, b_sh;  in vuzpq_s16()  local
15953      a_sh = _mm_shuffle_epi8 (a, *(__m128i*)mask8_32_even_odd); //a0, a2, a4, a6,  a1, a3, a5, a7  in vuzpq_s16()
15955     v16x8.val[0] = _mm_unpacklo_epi64(a_sh, b_sh); //a0, a2, a4, a6, b0, b2, b4, b6  in vuzpq_s16()
15956     v16x8.val[1] = _mm_unpackhi_epi64(a_sh, b_sh); //a1, a3, a5, a7, b1, b3, b5, b7  in vuzpq_s16()
15965     __m128i a_sh, b_sh;  in vuzpq_s32()  local
15966     a_sh = _mm_shuffle_epi32 (a, 216); //a0, a2, a1, a3  in vuzpq_s32()
15969     v32x4.val[0] = _mm_unpacklo_epi64(a_sh, b_sh); //a0, a2, b0, b2  in vuzpq_s32()
15970     v32x4.val[1] = _mm_unpackhi_epi64(a_sh, b_sh); //a1, a3, b1, b3  in vuzpq_s32()