499 #define _mm256_shrdi_epi16(A, B, C) \  argument
501 	(__v16hi)(__m256i)(B),(int)(C))
502 #define _mm256_mask_shrdi_epi16(A, B, C, D, E) \  argument
504 	(__v16hi)(__m256i)(D), (int)(E), (__v16hi)(__m256i)(A),(__mmask16)(B))
505 #define _mm256_maskz_shrdi_epi16(A, B, C, D) \  argument
506   ((__m256i) __builtin_ia32_vpshrd_v16hi_mask ((__v16hi)(__m256i)(B), \
509 #define _mm256_shrdi_epi32(A, B, C) \  argument
511 	(__v8si)(__m256i)(B),(int)(C))
512 #define _mm256_mask_shrdi_epi32(A, B, C, D, E) \  argument
514 	(__v8si)(__m256i)(D), (int)(E), (__v8si)(__m256i)(A),(__mmask8)(B))
515 #define _mm256_maskz_shrdi_epi32(A, B, C, D) \  argument
516   ((__m256i) __builtin_ia32_vpshrd_v8si_mask ((__v8si)(__m256i)(B), \
519 #define _mm256_shrdi_epi64(A, B, C) \  argument
521 	(__v4di)(__m256i)(B),(int)(C))
522 #define _mm256_mask_shrdi_epi64(A, B, C, D, E) \  argument
524 	(__v4di)(__m256i)(D), (int)(E), (__v4di)(__m256i)(A),(__mmask8)(B))
525 #define _mm256_maskz_shrdi_epi64(A, B, C, D) \  argument
526   ((__m256i) __builtin_ia32_vpshrd_v4di_mask ((__v4di)(__m256i)(B), \
529 #define _mm_shrdi_epi16(A, B, C) \  argument
531 	(__v8hi)(__m128i)(B),(int)(C))
532 #define _mm_mask_shrdi_epi16(A, B, C, D, E) \  argument
534 	(__v8hi)(__m128i)(D), (int)(E), (__v8hi)(__m128i)(A),(__mmask8)(B))
535 #define _mm_maskz_shrdi_epi16(A, B, C, D) \  argument
536   ((__m128i) __builtin_ia32_vpshrd_v8hi_mask ((__v8hi)(__m128i)(B), \
539 #define _mm_shrdi_epi32(A, B, C) \  argument
541 	(__v4si)(__m128i)(B),(int)(C))
542 #define _mm_mask_shrdi_epi32(A, B, C, D, E) \  argument
544 	(__v4si)(__m128i)(D), (int)(E), (__v4si)(__m128i)(A),(__mmask8)(B))
545 #define _mm_maskz_shrdi_epi32(A, B, C, D) \  argument
546   ((__m128i) __builtin_ia32_vpshrd_v4si_mask ((__v4si)(__m128i)(B), \
549 #define _mm_shrdi_epi64(A, B, C) \  argument
551 	(__v2di)(__m128i)(B),(int)(C))
552 #define _mm_mask_shrdi_epi64(A, B, C, D, E) \  argument
554 	(__v2di)(__m128i)(D), (int)(E), (__v2di)(__m128i)(A),(__mmask8)(B))
555 #define _mm_maskz_shrdi_epi64(A, B, C, D) \  argument
556   ((__m128i) __builtin_ia32_vpshrd_v2di_mask ((__v2di)(__m128i)(B), \
559 #define _mm256_shldi_epi16(A, B, C) \  argument
561 						(__v16hi)(__m256i)(B),(int)(C))
562 #define _mm256_mask_shldi_epi16(A, B, C, D, E) \  argument
564 	(__v16hi)(__m256i)(D), (int)(E), (__v16hi)(__m256i)(A),(__mmask16)(B))
565 #define _mm256_maskz_shldi_epi16(A, B, C, D) \  argument
566   ((__m256i) __builtin_ia32_vpshld_v16hi_mask ((__v16hi)(__m256i)(B), \
569 #define _mm256_shldi_epi32(A, B, C) \  argument
571 	(__v8si)(__m256i)(B),(int)(C))
572 #define _mm256_mask_shldi_epi32(A, B, C, D, E) \  argument
574 	(__v8si)(__m256i)(D), (int)(E), (__v8si)(__m256i)(A),(__mmask8)(B))
575 #define _mm256_maskz_shldi_epi32(A, B, C, D) \  argument
576   ((__m256i) __builtin_ia32_vpshld_v8si_mask ((__v8si)(__m256i)(B), \
579 #define _mm256_shldi_epi64(A, B, C) \  argument
581 	(__v4di)(__m256i)(B),(int)(C))
582 #define _mm256_mask_shldi_epi64(A, B, C, D, E) \  argument
584 	(__v4di)(__m256i)(D), (int)(E), (__v4di)(__m256i)(A),(__mmask8)(B))
585 #define _mm256_maskz_shldi_epi64(A, B, C, D) \  argument
586   ((__m256i) __builtin_ia32_vpshld_v4di_mask ((__v4di)(__m256i)(B), \
589 #define _mm_shldi_epi16(A, B, C) \  argument
591 	(__v8hi)(__m128i)(B),(int)(C))
592 #define _mm_mask_shldi_epi16(A, B, C, D, E) \  argument
594 	(__v8hi)(__m128i)(D), (int)(E), (__v8hi)(__m128i)(A),(__mmask8)(B))
595 #define _mm_maskz_shldi_epi16(A, B, C, D) \  argument
596   ((__m128i) __builtin_ia32_vpshld_v8hi_mask ((__v8hi)(__m128i)(B), \
599 #define _mm_shldi_epi32(A, B, C) \  argument
601 	(__v4si)(__m128i)(B),(int)(C))
602 #define _mm_mask_shldi_epi32(A, B, C, D, E) \  argument
604 	(__v4si)(__m128i)(D), (int)(E), (__v4si)(__m128i)(A),(__mmask8)(B))
605 #define _mm_maskz_shldi_epi32(A, B, C, D) \  argument
606   ((__m128i) __builtin_ia32_vpshld_v4si_mask ((__v4si)(__m128i)(B), \
609 #define _mm_shldi_epi64(A, B, C) \  argument
611 	(__v2di)(__m128i)(B),(int)(C))
612 #define _mm_mask_shldi_epi64(A, B, C, D, E) \  argument
614 	(__v2di)(__m128i)(D), (int)(E), (__v2di)(__m128i)(A),(__mmask8)(B))
615 #define _mm_maskz_shldi_epi64(A, B, C, D) \  argument
616   ((__m128i) __builtin_ia32_vpshld_v2di_mask ((__v2di)(__m128i)(B), \