205 #define vreinterpretq_u8_m128i(x) vreinterpretq_u8_s64(x)  macro
1614     uint8x16_t idx = vreinterpretq_u8_m128i(b);  // input b  in _mm_shuffle_epi8()
1863     uint8x16_t a = vreinterpretq_u8_m128i(_a);  in _mm_blendv_epi8()
1864     uint8x16_t b = vreinterpretq_u8_m128i(_b);  in _mm_blendv_epi8()
2199     uint8x16_t input = vreinterpretq_u8_m128i(a);  in _mm_movemask_epi8()
2220     uint8x16_t input = vreinterpretq_u8_m128i(a);  in _mm_movemask_epi8()
2456         vqsubq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b)));  in _mm_subs_epu8()
2796         vrhaddq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b)));  in _mm_avg_epu8()
2945         vqaddq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b)));  in _mm_adds_epu8()
3113     uint8x16_t a = vreinterpretq_u8_m128i(_a);  in _mm_maddubs_epi16()
3479         vmaxq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b)));  in _mm_max_epu8()
3488         vminq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b)));  in _mm_min_epu8()
4455     uint8x16_t u8x16 = vreinterpretq_u8_m128i(a);    /* xxxx xxxx xxxx DCBA */  in _mm_cvtepu8_epi16()
4465     uint8x16_t u8x16 = vreinterpretq_u8_m128i(a);      /* xxxx xxxx xxxx DCBA */  in _mm_cvtepu8_epi32()
4475     uint8x16_t u8x16 = vreinterpretq_u8_m128i(a);      /* xxxx xxxx xxxx xxBA */  in _mm_cvtepu8_epi64()
5304 #define _mm_extract_epi8(a, imm) vgetq_lane_u8(vreinterpretq_u8_m128i(a), (imm))
5647     uint8x16_t w = vreinterpretq_u8_m128i(EncBlock);  in _mm_aesenc_si128()
5769         vaesmcq_u8(vaeseq_u8(vreinterpretq_u8_m128i(a), vdupq_n_u8(0))) ^  in _mm_aesenc_si128()
5770         vreinterpretq_u8_m128i(b));  in _mm_aesenc_si128()
5777                              vreinterpretq_u8_m128i(a), vdupq_n_u8(0))),  in _mm_aesenclast_si128()
5784     uint8x16_t u8 = vaeseq_u8(vreinterpretq_u8_m128i(a), vdupq_n_u8(0));  in _mm_aeskeygenassist_si128()