8214     __m128i a128, r128;  variable
8216     r128 = _mm_slli_epi16 (a128, b);
8217     r128 = _mm_packs_epi16 (r128,r128); //saturated s8, use 64 low bits only
8218     return64(r128);
8226     __m128i a128, r128;  variable
8228     r128 = _mm_slli_epi32 (a128, b); //shift_res
8229     r128 = _mm_packs_epi32 (r128,r128); //saturated s16, use 64 low bits only
8230     return64(r128);
8264     __m128i a128, r128;  variable
8266     r128 = _mm_slli_epi16 (a128, b); //shift_res
8267     r128 = _mm_packus_epi16 (r128,r128); //saturated u8, use 64 low bits only
8268     return64(r128);
8276     __m128i a128, r128;  variable
8278     r128 = _mm_slli_epi32 (a128, b); //shift_res
8279     r128 = _MM_PACKUS1_EPI32 (r128); //saturated s16
8280     return64(r128);
8435     __m128i a128, r128;  variable
8437     r128 = _mm_slli_epi16 (a128, b);
8438     r128 = _mm_packus_epi16 (r128,r128); //saturated u8, use 64 low bits only
8439     return64(r128);
8446     __m128i a128, r128;  variable
8448     r128 = _mm_slli_epi32 (a128, b); //shift_res
8449     r128 = _MM_PACKUS1_EPI32 (r128); //saturated s16, use 64 low bits only
8450     return64(r128);