gcc/config/i386/avx512vlintrin.h

/* Copyright (C) 2014-2021 Free Software Foundation, Inc.

   This file is part of GCC.

   GCC is free software; you can redistribute it and/or modify
   it under the terms of the GNU General Public License as published by
   the Free Software Foundation; either version 3, or (at your option)
   any later version.

   GCC is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
   GNU General Public License for more details.

   Under Section 7 of GPL version 3, you are granted additional
   permissions described in the GCC Runtime Library Exception, version
   3.1, as published by the Free Software Foundation.

   You should have received a copy of the GNU General Public License and
   a copy of the GCC Runtime Library Exception along with this program;
   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
   <http://www.gnu.org/licenses/>.  */

#ifndef _IMMINTRIN_H_INCLUDED
#error "Never use <avx512vlintrin.h> directly; include <immintrin.h> instead."
#endif

#ifndef _AVX512VLINTRIN_H_INCLUDED
#define _AVX512VLINTRIN_H_INCLUDED

#ifndef __AVX512VL__
#pragma GCC push_options
#pragma GCC target("avx512vl")
#define __DISABLE_AVX512VL__
#endif /* __AVX512VL__ */

/* Internal data types for implementing the intrinsics.  */
typedef unsigned int __mmask32;
typedef int __v4si_u __attribute__ ((__vector_size__ (16),	\
				     __may_alias__, __aligned__ (1)));
typedef int __v8si_u __attribute__ ((__vector_size__ (32),	\
				     __may_alias__, __aligned__ (1)));
typedef long long __v2di_u __attribute__ ((__vector_size__ (16),	\
					   __may_alias__, __aligned__ (1)));
typedef long long __v4di_u __attribute__ ((__vector_size__ (32),	\
					   __may_alias__, __aligned__ (1)));

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_mov_pd (__m256d __W, __mmask8 __U, __m256d __A)
{
  return (__m256d) __builtin_ia32_movapd256_mask ((__v4df) __A,
						  (__v4df) __W,
						  (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_mov_pd (__mmask8 __U, __m256d __A)
{
  return (__m256d) __builtin_ia32_movapd256_mask ((__v4df) __A,
						  (__v4df)
						  _mm256_setzero_pd (),
						  (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_mov_pd (__m128d __W, __mmask8 __U, __m128d __A)
{
  return (__m128d) __builtin_ia32_movapd128_mask ((__v2df) __A,
						  (__v2df) __W,
						  (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_mov_pd (__mmask8 __U, __m128d __A)
{
  return (__m128d) __builtin_ia32_movapd128_mask ((__v2df) __A,
						  (__v2df)
						  _mm_setzero_pd (),
						  (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_load_pd (__m256d __W, __mmask8 __U, void const *__P)
{
  return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P,
						   (__v4df) __W,
						   (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_load_pd (__mmask8 __U, void const *__P)
{
  return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P,
						   (__v4df)
						   _mm256_setzero_pd (),
						   (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_load_pd (__m128d __W, __mmask8 __U, void const *__P)
{
  return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P,
						   (__v2df) __W,
						   (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_load_pd (__mmask8 __U, void const *__P)
{
  return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P,
						   (__v2df)
						   _mm_setzero_pd (),
						   (__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_store_pd (void *__P, __mmask8 __U, __m256d __A)
{
  __builtin_ia32_storeapd256_mask ((__v4df *) __P,
				   (__v4df) __A,
				   (__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_store_pd (void *__P, __mmask8 __U, __m128d __A)
{
  __builtin_ia32_storeapd128_mask ((__v2df *) __P,
				   (__v2df) __A,
				   (__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_mov_ps (__m256 __W, __mmask8 __U, __m256 __A)
{
  return (__m256) __builtin_ia32_movaps256_mask ((__v8sf) __A,
						 (__v8sf) __W,
						 (__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_mov_ps (__mmask8 __U, __m256 __A)
{
  return (__m256) __builtin_ia32_movaps256_mask ((__v8sf) __A,
						 (__v8sf)
						 _mm256_setzero_ps (),
						 (__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_mov_ps (__m128 __W, __mmask8 __U, __m128 __A)
{
  return (__m128) __builtin_ia32_movaps128_mask ((__v4sf) __A,
						 (__v4sf) __W,
						 (__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_mov_ps (__mmask8 __U, __m128 __A)
{
  return (__m128) __builtin_ia32_movaps128_mask ((__v4sf) __A,
						 (__v4sf)
						 _mm_setzero_ps (),
						 (__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_load_ps (__m256 __W, __mmask8 __U, void const *__P)
{
  return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P,
						  (__v8sf) __W,
						  (__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_load_ps (__mmask8 __U, void const *__P)
{
  return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P,
						  (__v8sf)
						  _mm256_setzero_ps (),
						  (__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_load_ps (__m128 __W, __mmask8 __U, void const *__P)
{
  return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P,
						  (__v4sf) __W,
						  (__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_load_ps (__mmask8 __U, void const *__P)
{
  return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P,
						  (__v4sf)
						  _mm_setzero_ps (),
						  (__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_store_ps (void *__P, __mmask8 __U, __m256 __A)
{
  __builtin_ia32_storeaps256_mask ((__v8sf *) __P,
				   (__v8sf) __A,
				   (__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_store_ps (void *__P, __mmask8 __U, __m128 __A)
{
  __builtin_ia32_storeaps128_mask ((__v4sf *) __P,
				   (__v4sf) __A,
				   (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_mov_epi64 (__m256i __W, __mmask8 __U, __m256i __A)
{
  return (__m256i) __builtin_ia32_movdqa64_256_mask ((__v4di) __A,
						     (__v4di) __W,
						     (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_mov_epi64 (__mmask8 __U, __m256i __A)
{
  return (__m256i) __builtin_ia32_movdqa64_256_mask ((__v4di) __A,
						     (__v4di)
						     _mm256_setzero_si256 (),
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_mov_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_movdqa64_128_mask ((__v2di) __A,
						     (__v2di) __W,
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_mov_epi64 (__mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_movdqa64_128_mask ((__v2di) __A,
						     (__v2di)
						     _mm_setzero_si128 (),
						     (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_load_epi64 (void const *__P)
{
  return (__m256i) (*(__v4di *) __P);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_load_epi64 (__m256i __W, __mmask8 __U, void const *__P)
{
  return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P,
							(__v4di) __W,
							(__mmask8)
							__U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_load_epi64 (__mmask8 __U, void const *__P)
{
  return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P,
							(__v4di)
							_mm256_setzero_si256 (),
							(__mmask8)
							__U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_load_epi64 (void const *__P)
{
  return (__m128i) (*(__v2di *) __P);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_load_epi64 (__m128i __W, __mmask8 __U, void const *__P)
{
  return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P,
							(__v2di) __W,
							(__mmask8)
							__U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_load_epi64 (__mmask8 __U, void const *__P)
{
  return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P,
							(__v2di)
							_mm_setzero_si128 (),
							(__mmask8)
							__U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_store_epi64 (void *__P, __mmask8 __U, __m256i __A)
{
  __builtin_ia32_movdqa64store256_mask ((__v4di *) __P,
					(__v4di) __A,
					(__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_store_epi64 (void *__P, __mmask8 __U, __m128i __A)
{
  __builtin_ia32_movdqa64store128_mask ((__v2di *) __P,
					(__v2di) __A,
					(__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_mov_epi32 (__m256i __W, __mmask8 __U, __m256i __A)
{
  return (__m256i) __builtin_ia32_movdqa32_256_mask ((__v8si) __A,
						     (__v8si) __W,
						     (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_mov_epi32 (__mmask8 __U, __m256i __A)
{
  return (__m256i) __builtin_ia32_movdqa32_256_mask ((__v8si) __A,
						     (__v8si)
						     _mm256_setzero_si256 (),
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_mov_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_movdqa32_128_mask ((__v4si) __A,
						     (__v4si) __W,
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_mov_epi32 (__mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_movdqa32_128_mask ((__v4si) __A,
						     (__v4si)
						     _mm_setzero_si128 (),
						     (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_load_epi32 (void const *__P)
{
  return (__m256i) (*(__v8si *) __P);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_load_epi32 (__m256i __W, __mmask8 __U, void const *__P)
{
  return (__m256i) __builtin_ia32_movdqa32load256_mask ((__v8si *) __P,
							(__v8si) __W,
							(__mmask8)
							__U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_load_epi32 (__mmask8 __U, void const *__P)
{
  return (__m256i) __builtin_ia32_movdqa32load256_mask ((__v8si *) __P,
							(__v8si)
							_mm256_setzero_si256 (),
							(__mmask8)
							__U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_load_epi32 (void const *__P)
{
  return (__m128i) (*(__v4si *) __P);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_load_epi32 (__m128i __W, __mmask8 __U, void const *__P)
{
  return (__m128i) __builtin_ia32_movdqa32load128_mask ((__v4si *) __P,
							(__v4si) __W,
							(__mmask8)
							__U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_load_epi32 (__mmask8 __U, void const *__P)
{
  return (__m128i) __builtin_ia32_movdqa32load128_mask ((__v4si *) __P,
							(__v4si)
							_mm_setzero_si128 (),
							(__mmask8)
							__U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_store_epi32 (void *__P, __m256i __A)
{
  *(__v8si *) __P = (__v8si) __A;
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_store_epi32 (void *__P, __mmask8 __U, __m256i __A)
{
  __builtin_ia32_movdqa32store256_mask ((__v8si *) __P,
					(__v8si) __A,
					(__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_store_epi32 (void *__P, __m128i __A)
{
  *(__v4si *) __P = (__v4si) __A;
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_store_epi32 (void *__P, __mmask8 __U, __m128i __A)
{
  __builtin_ia32_movdqa32store128_mask ((__v4si *) __P,
					(__v4si) __A,
					(__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_add_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
{
  return (__m128d) __builtin_ia32_addpd128_mask ((__v2df) __A,
						 (__v2df) __B,
						 (__v2df) __W,
						 (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_add_pd (__mmask8 __U, __m128d __A, __m128d __B)
{
  return (__m128d) __builtin_ia32_addpd128_mask ((__v2df) __A,
						 (__v2df) __B,
						 (__v2df)
						 _mm_setzero_pd (),
						 (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_add_pd (__m256d __W, __mmask8 __U, __m256d __A,
		    __m256d __B)
{
  return (__m256d) __builtin_ia32_addpd256_mask ((__v4df) __A,
						 (__v4df) __B,
						 (__v4df) __W,
						 (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_add_pd (__mmask8 __U, __m256d __A, __m256d __B)
{
  return (__m256d) __builtin_ia32_addpd256_mask ((__v4df) __A,
						 (__v4df) __B,
						 (__v4df)
						 _mm256_setzero_pd (),
						 (__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_add_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
{
  return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A,
						(__v4sf) __B,
						(__v4sf) __W,
						(__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_add_ps (__mmask8 __U, __m128 __A, __m128 __B)
{
  return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A,
						(__v4sf) __B,
						(__v4sf)
						_mm_setzero_ps (),
						(__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_add_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B)
{
  return (__m256) __builtin_ia32_addps256_mask ((__v8sf) __A,
						(__v8sf) __B,
						(__v8sf) __W,
						(__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_add_ps (__mmask8 __U, __m256 __A, __m256 __B)
{
  return (__m256) __builtin_ia32_addps256_mask ((__v8sf) __A,
						(__v8sf) __B,
						(__v8sf)
						_mm256_setzero_ps (),
						(__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_sub_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
{
  return (__m128d) __builtin_ia32_subpd128_mask ((__v2df) __A,
						 (__v2df) __B,
						 (__v2df) __W,
						 (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_sub_pd (__mmask8 __U, __m128d __A, __m128d __B)
{
  return (__m128d) __builtin_ia32_subpd128_mask ((__v2df) __A,
						 (__v2df) __B,
						 (__v2df)
						 _mm_setzero_pd (),
						 (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_sub_pd (__m256d __W, __mmask8 __U, __m256d __A,
		    __m256d __B)
{
  return (__m256d) __builtin_ia32_subpd256_mask ((__v4df) __A,
						 (__v4df) __B,
						 (__v4df) __W,
						 (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_sub_pd (__mmask8 __U, __m256d __A, __m256d __B)
{
  return (__m256d) __builtin_ia32_subpd256_mask ((__v4df) __A,
						 (__v4df) __B,
						 (__v4df)
						 _mm256_setzero_pd (),
						 (__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_sub_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
{
  return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A,
						(__v4sf) __B,
						(__v4sf) __W,
						(__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_sub_ps (__mmask8 __U, __m128 __A, __m128 __B)
{
  return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A,
						(__v4sf) __B,
						(__v4sf)
						_mm_setzero_ps (),
						(__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_sub_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B)
{
  return (__m256) __builtin_ia32_subps256_mask ((__v8sf) __A,
						(__v8sf) __B,
						(__v8sf) __W,
						(__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_sub_ps (__mmask8 __U, __m256 __A, __m256 __B)
{
  return (__m256) __builtin_ia32_subps256_mask ((__v8sf) __A,
						(__v8sf) __B,
						(__v8sf)
						_mm256_setzero_ps (),
						(__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_store_epi64 (void *__P, __m256i __A)
{
  *(__m256i *) __P = __A;
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_store_epi64 (void *__P, __m128i __A)
{
  *(__m128i *) __P = __A;
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_loadu_pd (__m256d __W, __mmask8 __U, void const *__P)
{
  return (__m256d) __builtin_ia32_loadupd256_mask ((const double *) __P,
						   (__v4df) __W,
						   (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_loadu_pd (__mmask8 __U, void const *__P)
{
  return (__m256d) __builtin_ia32_loadupd256_mask ((const double *) __P,
						   (__v4df)
						   _mm256_setzero_pd (),
						   (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_loadu_pd (__m128d __W, __mmask8 __U, void const *__P)
{
  return (__m128d) __builtin_ia32_loadupd128_mask ((const double *) __P,
						   (__v2df) __W,
						   (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_loadu_pd (__mmask8 __U, void const *__P)
{
  return (__m128d) __builtin_ia32_loadupd128_mask ((const double *) __P,
						   (__v2df)
						   _mm_setzero_pd (),
						   (__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_storeu_pd (void *__P, __mmask8 __U, __m256d __A)
{
  __builtin_ia32_storeupd256_mask ((double *) __P,
				   (__v4df) __A,
				   (__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_storeu_pd (void *__P, __mmask8 __U, __m128d __A)
{
  __builtin_ia32_storeupd128_mask ((double *) __P,
				   (__v2df) __A,
				   (__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_loadu_ps (__m256 __W, __mmask8 __U, void const *__P)
{
  return (__m256) __builtin_ia32_loadups256_mask ((const float *) __P,
						  (__v8sf) __W,
						  (__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_loadu_ps (__mmask8 __U, void const *__P)
{
  return (__m256) __builtin_ia32_loadups256_mask ((const float *) __P,
						  (__v8sf)
						  _mm256_setzero_ps (),
						  (__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_loadu_ps (__m128 __W, __mmask8 __U, void const *__P)
{
  return (__m128) __builtin_ia32_loadups128_mask ((const float *) __P,
						  (__v4sf) __W,
						  (__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_loadu_ps (__mmask8 __U, void const *__P)
{
  return (__m128) __builtin_ia32_loadups128_mask ((const float *) __P,
						  (__v4sf)
						  _mm_setzero_ps (),
						  (__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_storeu_ps (void *__P, __mmask8 __U, __m256 __A)
{
  __builtin_ia32_storeups256_mask ((float *) __P,
				   (__v8sf) __A,
				   (__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_storeu_ps (void *__P, __mmask8 __U, __m128 __A)
{
  __builtin_ia32_storeups128_mask ((float *) __P,
				   (__v4sf) __A,
				   (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_loadu_epi64 (void const *__P)
{
  return (__m256i) (*(__v4di_u *) __P);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_loadu_epi64 (__m256i __W, __mmask8 __U, void const *__P)
{
  return (__m256i) __builtin_ia32_loaddqudi256_mask ((const long long *) __P,
						     (__v4di) __W,
						     (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_loadu_epi64 (__mmask8 __U, void const *__P)
{
  return (__m256i) __builtin_ia32_loaddqudi256_mask ((const long long *) __P,
						     (__v4di)
						     _mm256_setzero_si256 (),
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_loadu_epi64 (void const *__P)
{
  return (__m128i) (*(__v2di_u *) __P);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_loadu_epi64 (__m128i __W, __mmask8 __U, void const *__P)
{
  return (__m128i) __builtin_ia32_loaddqudi128_mask ((const long long *) __P,
						     (__v2di) __W,
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_loadu_epi64 (__mmask8 __U, void const *__P)
{
  return (__m128i) __builtin_ia32_loaddqudi128_mask ((const long long *) __P,
						     (__v2di)
						     _mm_setzero_si128 (),
						     (__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_storeu_epi64 (void *__P, __m256i __A)
{
  *(__m256i_u *) __P = (__m256i_u) __A;
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_storeu_epi64 (void *__P, __mmask8 __U, __m256i __A)
{
  __builtin_ia32_storedqudi256_mask ((long long *) __P,
				     (__v4di) __A,
				     (__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_storeu_epi64 (void *__P, __m128i __A)
{
  *(__m128i_u *) __P = (__m128i_u) __A;
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_storeu_epi64 (void *__P, __mmask8 __U, __m128i __A)
{
  __builtin_ia32_storedqudi128_mask ((long long *) __P,
				     (__v2di) __A,
				     (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_loadu_epi32 (void const *__P)
{
  return (__m256i) (*(__v8si_u *) __P);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_loadu_epi32 (__m256i __W, __mmask8 __U, void const *__P)
{
  return (__m256i) __builtin_ia32_loaddqusi256_mask ((const int *) __P,
						     (__v8si) __W,
						     (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_loadu_epi32 (__mmask8 __U, void const *__P)
{
  return (__m256i) __builtin_ia32_loaddqusi256_mask ((const int *) __P,
						     (__v8si)
						     _mm256_setzero_si256 (),
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_loadu_epi32 (void const *__P)
{
  return (__m128i) (*(__v4si_u *) __P);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_loadu_epi32 (__m128i __W, __mmask8 __U, void const *__P)
{
  return (__m128i) __builtin_ia32_loaddqusi128_mask ((const int *) __P,
						     (__v4si) __W,
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_loadu_epi32 (__mmask8 __U, void const *__P)
{
  return (__m128i) __builtin_ia32_loaddqusi128_mask ((const int *) __P,
						     (__v4si)
						     _mm_setzero_si128 (),
						     (__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_storeu_epi32 (void *__P, __m256i __A)
{
  *(__m256i_u *) __P = (__m256i_u) __A;
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_storeu_epi32 (void *__P, __mmask8 __U, __m256i __A)
{
  __builtin_ia32_storedqusi256_mask ((int *) __P,
				     (__v8si) __A,
				     (__mmask8) __U);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_storeu_epi32 (void *__P, __m128i __A)
{
  *(__m128i_u *) __P = (__m128i_u) __A;
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_storeu_epi32 (void *__P, __mmask8 __U, __m128i __A)
{
  __builtin_ia32_storedqusi128_mask ((int *) __P,
				     (__v4si) __A,
				     (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_abs_epi32 (__m256i __W, __mmask8 __U, __m256i __A)
{
  return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A,
						 (__v8si) __W,
						 (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_abs_epi32 (__mmask8 __U, __m256i __A)
{
  return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A,
						 (__v8si)
						 _mm256_setzero_si256 (),
						 (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_abs_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A,
						 (__v4si) __W,
						 (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_abs_epi32 (__mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A,
						 (__v4si)
						 _mm_setzero_si128 (),
						 (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_abs_epi64 (__m256i __A)
{
  return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A,
						 (__v4di)
						 _mm256_setzero_si256 (),
						 (__mmask8) -1);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_abs_epi64 (__m256i __W, __mmask8 __U, __m256i __A)
{
  return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A,
						 (__v4di) __W,
						 (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_abs_epi64 (__mmask8 __U, __m256i __A)
{
  return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A,
						 (__v4di)
						 _mm256_setzero_si256 (),
						 (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_abs_epi64 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A,
						 (__v2di)
						 _mm_setzero_si128 (),
						 (__mmask8) -1);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_abs_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A,
						 (__v2di) __W,
						 (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_abs_epi64 (__mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A,
						 (__v2di)
						 _mm_setzero_si128 (),
						 (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtpd_epu32 (__m256d __A)
{
  return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A,
						     (__v4si)
						     _mm_setzero_si128 (),
						     (__mmask8) -1);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A)
{
  return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A,
						     (__v4si) __W,
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtpd_epu32 (__mmask8 __U, __m256d __A)
{
  return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A,
						     (__v4si)
						     _mm_setzero_si128 (),
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtpd_epu32 (__m128d __A)
{
  return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A,
						     (__v4si)
						     _mm_setzero_si128 (),
						     (__mmask8) -1);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A)
{
  return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A,
						     (__v4si) __W,
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtpd_epu32 (__mmask8 __U, __m128d __A)
{
  return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A,
						     (__v4si)
						     _mm_setzero_si128 (),
						     (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvttps_epi32 (__m256i __W, __mmask8 __U, __m256 __A)
{
  return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A,
						     (__v8si) __W,
						     (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvttps_epi32 (__mmask8 __U, __m256 __A)
{
  return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A,
						     (__v8si)
						     _mm256_setzero_si256 (),
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvttps_epi32 (__m128i __W, __mmask8 __U, __m128 __A)
{
  return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A,
						     (__v4si) __W,
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvttps_epi32 (__mmask8 __U, __m128 __A)
{
  return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A,
						     (__v4si)
						     _mm_setzero_si128 (),
						     (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvttps_epu32 (__m256 __A)
{
  return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A,
						      (__v8si)
						      _mm256_setzero_si256 (),
						      (__mmask8) -1);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvttps_epu32 (__m256i __W, __mmask8 __U, __m256 __A)
{
  return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A,
						      (__v8si) __W,
						      (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvttps_epu32 (__mmask8 __U, __m256 __A)
{
  return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A,
						      (__v8si)
						      _mm256_setzero_si256 (),
						      (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvttps_epu32 (__m128 __A)
{
  return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A,
						      (__v4si)
						      _mm_setzero_si128 (),
						      (__mmask8) -1);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvttps_epu32 (__m128i __W, __mmask8 __U, __m128 __A)
{
  return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A,
						      (__v4si) __W,
						      (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvttps_epu32 (__mmask8 __U, __m128 __A)
{
  return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A,
						      (__v4si)
						      _mm_setzero_si128 (),
						      (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A)
{
  return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A,
						     (__v4si) __W,
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvttpd_epi32 (__mmask8 __U, __m256d __A)
{
  return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A,
						     (__v4si)
						     _mm_setzero_si128 (),
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A)
{
  return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A,
						     (__v4si) __W,
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvttpd_epi32 (__mmask8 __U, __m128d __A)
{
  return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A,
						     (__v4si)
						     _mm_setzero_si128 (),
						     (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvttpd_epu32 (__m256d __A)
{
  return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A,
						      (__v4si)
						      _mm_setzero_si128 (),
						      (__mmask8) -1);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A)
{
  return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A,
						      (__v4si) __W,
						      (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvttpd_epu32 (__mmask8 __U, __m256d __A)
{
  return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A,
						      (__v4si)
						      _mm_setzero_si128 (),
						      (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvttpd_epu32 (__m128d __A)
{
  return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A,
						      (__v4si)
						      _mm_setzero_si128 (),
						      (__mmask8) -1);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A)
{
  return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A,
						      (__v4si) __W,
						      (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvttpd_epu32 (__mmask8 __U, __m128d __A)
{
  return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A,
						      (__v4si)
						      _mm_setzero_si128 (),
						      (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A)
{
  return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A,
						    (__v4si) __W,
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtpd_epi32 (__mmask8 __U, __m256d __A)
{
  return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A,
						    (__v4si)
						    _mm_setzero_si128 (),
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A)
{
  return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A,
						    (__v4si) __W,
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtpd_epi32 (__mmask8 __U, __m128d __A)
{
  return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A,
						    (__v4si)
						    _mm_setzero_si128 (),
						    (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi32_pd (__m256d __W, __mmask8 __U, __m128i __A)
{
  return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A,
						    (__v4df) __W,
						    (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A)
{
  return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A,
						    (__v4df)
						    _mm256_setzero_pd (),
						    (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi32_pd (__m128d __W, __mmask8 __U, __m128i __A)
{
  return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A,
						    (__v2df) __W,
						    (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A)
{
  return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A,
						    (__v2df)
						    _mm_setzero_pd (),
						    (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtepu32_pd (__m128i __A)
{
  return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A,
						     (__v4df)
						     _mm256_setzero_pd (),
						     (__mmask8) -1);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepu32_pd (__m256d __W, __mmask8 __U, __m128i __A)
{
  return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A,
						     (__v4df) __W,
						     (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A)
{
  return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A,
						     (__v4df)
						     _mm256_setzero_pd (),
						     (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtepu32_pd (__m128i __A)
{
  return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A,
						     (__v2df)
						     _mm_setzero_pd (),
						     (__mmask8) -1);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepu32_pd (__m128d __W, __mmask8 __U, __m128i __A)
{
  return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A,
						     (__v2df) __W,
						     (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A)
{
  return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A,
						     (__v2df)
						     _mm_setzero_pd (),
						     (__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi32_ps (__m256 __W, __mmask8 __U, __m256i __A)
{
  return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A,
						   (__v8sf) __W,
						   (__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepi32_ps (__mmask8 __U, __m256i __A)
{
  return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A,
						   (__v8sf)
						   _mm256_setzero_ps (),
						   (__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi32_ps (__m128 __W, __mmask8 __U, __m128i __A)
{
  return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A,
						   (__v4sf) __W,
						   (__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepi32_ps (__mmask8 __U, __m128i __A)
{
  return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A,
						   (__v4sf)
						   _mm_setzero_ps (),
						   (__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtepu32_ps (__m256i __A)
{
  return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A,
						    (__v8sf)
						    _mm256_setzero_ps (),
						    (__mmask8) -1);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepu32_ps (__m256 __W, __mmask8 __U, __m256i __A)
{
  return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A,
						    (__v8sf) __W,
						    (__mmask8) __U);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepu32_ps (__mmask8 __U, __m256i __A)
{
  return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A,
						    (__v8sf)
						    _mm256_setzero_ps (),
						    (__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtepu32_ps (__m128i __A)
{
  return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A,
						    (__v4sf)
						    _mm_setzero_ps (),
						    (__mmask8) -1);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepu32_ps (__m128 __W, __mmask8 __U, __m128i __A)
{
  return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A,
						    (__v4sf) __W,
						    (__mmask8) __U);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepu32_ps (__mmask8 __U, __m128i __A)
{
  return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A,
						    (__v4sf)
						    _mm_setzero_ps (),
						    (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtps_pd (__m256d __W, __mmask8 __U, __m128 __A)
{
  return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A,
						    (__v4df) __W,
						    (__mmask8) __U);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtps_pd (__mmask8 __U, __m128 __A)
{
  return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A,
						    (__v4df)
						    _mm256_setzero_pd (),
						    (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtps_pd (__m128d __W, __mmask8 __U, __m128 __A)
{
  return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A,
						    (__v2df) __W,
						    (__mmask8) __U);
}

extern __inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtps_pd (__mmask8 __U, __m128 __A)
{
  return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A,
						    (__v2df)
						    _mm_setzero_pd (),
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtepi32_epi8 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A,
						  (__v16qi)
						  _mm_undefined_si128 (),
						  (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovdb128mem_mask ((unsigned int *) __P, (__v4si) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A,
						  (__v16qi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepi32_epi8 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A,
						  (__v16qi)
						  _mm_setzero_si128 (),
						  __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtepi32_epi8 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
						  (__v16qi)
						  _mm_undefined_si128 (),
						  (__mmask8) -1);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
						  (__v16qi) __O, __M);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovdb256mem_mask ((unsigned long long *) __P, (__v8si) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepi32_epi8 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
						  (__v16qi)
						  _mm_setzero_si128 (),
						  __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtsepi32_epi8 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A,
						   (__v16qi)
						   _mm_undefined_si128 (),
						   (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovsdb128mem_mask ((unsigned int *) __P, (__v4si) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A,
						   (__v16qi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtsepi32_epi8 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A,
						   (__v16qi)
						   _mm_setzero_si128 (),
						   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtsepi32_epi8 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A,
						   (__v16qi)
						   _mm_undefined_si128 (),
						   (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovsdb256mem_mask ((unsigned long long *) __P, (__v8si) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A,
						   (__v16qi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtsepi32_epi8 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A,
						   (__v16qi)
						   _mm_setzero_si128 (),
						   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtusepi32_epi8 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A,
						    (__v16qi)
						    _mm_undefined_si128 (),
						    (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovusdb128mem_mask ((unsigned int *) __P, (__v4si) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A,
						    (__v16qi) __O,
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtusepi32_epi8 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A,
						    (__v16qi)
						    _mm_setzero_si128 (),
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtusepi32_epi8 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A,
						    (__v16qi)
						    _mm_undefined_si128 (),
						    (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovusdb256mem_mask ((unsigned long long *) __P, (__v8si) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A,
						    (__v16qi) __O,
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtusepi32_epi8 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A,
						    (__v16qi)
						    _mm_setzero_si128 (),
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtepi32_epi16 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A,
						  (__v8hi)
						  _mm_setzero_si128 (),
						  (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovdw128mem_mask ((unsigned long long *) __P, (__v4si) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A,
						  (__v8hi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepi32_epi16 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A,
						  (__v8hi)
						  _mm_setzero_si128 (),
						  __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtepi32_epi16 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A,
						  (__v8hi)
						  _mm_setzero_si128 (),
						  (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi32_storeu_epi16 (void *  __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A,
						  (__v8hi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepi32_epi16 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A,
						  (__v8hi)
						  _mm_setzero_si128 (),
						  __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtsepi32_epi16 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A,
						   (__v8hi)
						   _mm_setzero_si128 (),
						   (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovsdw128mem_mask ((unsigned long long *) __P, (__v4si) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A,
						   (__v8hi)__O,
						   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtsepi32_epi16 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A,
						   (__v8hi)
						   _mm_setzero_si128 (),
						   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtsepi32_epi16 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A,
						   (__v8hi)
						   _mm_undefined_si128 (),
						   (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovsdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A,
						   (__v8hi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtsepi32_epi16 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A,
						   (__v8hi)
						   _mm_setzero_si128 (),
						   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtusepi32_epi16 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A,
						    (__v8hi)
						    _mm_undefined_si128 (),
						    (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovusdw128mem_mask ((unsigned long long *) __P, (__v4si) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A,
						    (__v8hi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtusepi32_epi16 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A,
						    (__v8hi)
						    _mm_setzero_si128 (),
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtusepi32_epi16 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A,
						    (__v8hi)
						    _mm_undefined_si128 (),
						    (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovusdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A,
						    (__v8hi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtusepi32_epi16 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A,
						    (__v8hi)
						    _mm_setzero_si128 (),
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtepi64_epi8 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A,
						  (__v16qi)
						  _mm_undefined_si128 (),
						  (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovqb128mem_mask ((unsigned short *) __P, (__v2di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A,
						  (__v16qi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepi64_epi8 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A,
						  (__v16qi)
						  _mm_setzero_si128 (),
						  __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtepi64_epi8 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A,
						  (__v16qi)
						  _mm_undefined_si128 (),
						  (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovqb256mem_mask ((unsigned int *) __P, (__v4di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A,
						  (__v16qi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepi64_epi8 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A,
						  (__v16qi)
						  _mm_setzero_si128 (),
						  __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtsepi64_epi8 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A,
						   (__v16qi)
						   _mm_undefined_si128 (),
						   (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovsqb128mem_mask ((unsigned short *) __P, (__v2di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A,
						   (__v16qi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtsepi64_epi8 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A,
						   (__v16qi)
						   _mm_setzero_si128 (),
						   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtsepi64_epi8 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A,
						   (__v16qi)
						   _mm_undefined_si128 (),
						   (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovsqb256mem_mask ((unsigned int *) __P, (__v4di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A,
						   (__v16qi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtsepi64_epi8 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A,
						   (__v16qi)
						   _mm_setzero_si128 (),
						   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtusepi64_epi8 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A,
						    (__v16qi)
						    _mm_undefined_si128 (),
						    (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovusqb128mem_mask ((unsigned short *) __P, (__v2di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A,
						    (__v16qi) __O,
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtusepi64_epi8 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A,
						    (__v16qi)
						    _mm_setzero_si128 (),
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtusepi64_epi8 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A,
						    (__v16qi)
						    _mm_undefined_si128 (),
						    (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovusqb256mem_mask ((unsigned int *) __P, (__v4di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A,
						    (__v16qi) __O,
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtusepi64_epi8 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A,
						    (__v16qi)
						    _mm_setzero_si128 (),
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtepi64_epi16 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A,
						  (__v8hi)
						  _mm_undefined_si128 (),
						  (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovqw128mem_mask ((unsigned int *) __P, (__v2di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A,
						  (__v8hi)__O,
						  __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepi64_epi16 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A,
						  (__v8hi)
						  _mm_setzero_si128 (),
						  __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtepi64_epi16 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A,
						  (__v8hi)
						  _mm_undefined_si128 (),
						  (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovqw256mem_mask ((unsigned long long *) __P, (__v4di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A,
						  (__v8hi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepi64_epi16 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A,
						  (__v8hi)
						  _mm_setzero_si128 (),
						  __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtsepi64_epi16 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A,
						   (__v8hi)
						   _mm_undefined_si128 (),
						   (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovsqw128mem_mask ((unsigned int *) __P, (__v2di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A,
						   (__v8hi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtsepi64_epi16 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A,
						   (__v8hi)
						   _mm_setzero_si128 (),
						   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtsepi64_epi16 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A,
						   (__v8hi)
						   _mm_undefined_si128 (),
						   (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovsqw256mem_mask ((unsigned long long *) __P, (__v4di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A,
						   (__v8hi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtsepi64_epi16 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A,
						   (__v8hi)
						   _mm_setzero_si128 (),
						   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtusepi64_epi16 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A,
						    (__v8hi)
						    _mm_undefined_si128 (),
						    (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovusqw128mem_mask ((unsigned int *) __P, (__v2di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A,
						    (__v8hi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtusepi64_epi16 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A,
						    (__v8hi)
						    _mm_setzero_si128 (),
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtusepi64_epi16 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A,
						    (__v8hi)
						    _mm_undefined_si128 (),
						    (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovusqw256mem_mask ((unsigned long long *) __P, (__v4di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A,
						    (__v8hi) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtusepi64_epi16 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A,
						    (__v8hi)
						    _mm_setzero_si128 (),
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtepi64_epi32 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A,
						  (__v4si)
						  _mm_undefined_si128 (),
						  (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovqd128mem_mask ((unsigned long long *) __P,
				    (__v2di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A,
						  (__v4si) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepi64_epi32 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A,
						  (__v4si)
						  _mm_setzero_si128 (),
						  __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtepi64_epi32 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A,
						  (__v4si)
						  _mm_undefined_si128 (),
						  (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A,
						  (__v4si) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepi64_epi32 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A,
						  (__v4si)
						  _mm_setzero_si128 (),
						  __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtsepi64_epi32 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A,
						   (__v4si)
						   _mm_undefined_si128 (),
						   (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovsqd128mem_mask ((unsigned long long *) __P, (__v2di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A,
						   (__v4si) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtsepi64_epi32 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A,
						   (__v4si)
						   _mm_setzero_si128 (),
						   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtsepi64_epi32 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A,
						   (__v4si)
						   _mm_undefined_si128 (),
						   (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovsqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A,
						   (__v4si)__O,
						   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtsepi64_epi32 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A,
						   (__v4si)
						   _mm_setzero_si128 (),
						   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_cvtusepi64_epi32 (__m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A,
						    (__v4si)
						    _mm_undefined_si128 (),
						    (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A)
{
  __builtin_ia32_pmovusqd128mem_mask ((unsigned long long *) __P, (__v2di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A,
						    (__v4si) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtusepi64_epi32 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A,
						    (__v4si)
						    _mm_setzero_si128 (),
						    __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_cvtusepi64_epi32 (__m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A,
						    (__v4si)
						    _mm_undefined_si128 (),
						    (__mmask8) -1);
}

extern __inline void
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
{
  __builtin_ia32_pmovusqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A,
						    (__v4si) __O, __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtusepi64_epi32 (__mmask8 __M, __m256i __A)
{
  return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A,
						    (__v4si)
						    _mm_setzero_si128 (),
						    __M);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_broadcastss_ps (__m256 __O, __mmask8 __M, __m128 __A)
{
  return (__m256) __builtin_ia32_broadcastss256_mask ((__v4sf) __A,
						      (__v8sf) __O,
						      __M);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_broadcastss_ps (__mmask8 __M, __m128 __A)
{
  return (__m256) __builtin_ia32_broadcastss256_mask ((__v4sf) __A,
						      (__v8sf)
						      _mm256_setzero_ps (),
						      __M);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_broadcastss_ps (__m128 __O, __mmask8 __M, __m128 __A)
{
  return (__m128) __builtin_ia32_broadcastss128_mask ((__v4sf) __A,
						      (__v4sf) __O,
						      __M);
}

extern __inline __m128
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_broadcastss_ps (__mmask8 __M, __m128 __A)
{
  return (__m128) __builtin_ia32_broadcastss128_mask ((__v4sf) __A,
						      (__v4sf)
						      _mm_setzero_ps (),
						      __M);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_broadcastsd_pd (__m256d __O, __mmask8 __M, __m128d __A)
{
  return (__m256d) __builtin_ia32_broadcastsd256_mask ((__v2df) __A,
						       (__v4df) __O,
						       __M);
}

extern __inline __m256d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_broadcastsd_pd (__mmask8 __M, __m128d __A)
{
  return (__m256d) __builtin_ia32_broadcastsd256_mask ((__v2df) __A,
						       (__v4df)
						       _mm256_setzero_pd (),
						       __M);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_broadcastd_epi32 (__m256i __O, __mmask8 __M, __m128i __A)
{
  return (__m256i) __builtin_ia32_pbroadcastd256_mask ((__v4si) __A,
						       (__v8si) __O,
						       __M);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A)
{
  return (__m256i) __builtin_ia32_pbroadcastd256_mask ((__v4si) __A,
						       (__v8si)
						       _mm256_setzero_si256 (),
						       __M);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_set1_epi32 (__m256i __O, __mmask8 __M, int __A)
{
  return (__m256i) __builtin_ia32_pbroadcastd256_gpr_mask (__A, (__v8si) __O,
							   __M);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_set1_epi32 (__mmask8 __M, int __A)
{
  return (__m256i) __builtin_ia32_pbroadcastd256_gpr_mask (__A,
							   (__v8si)
							   _mm256_setzero_si256 (),
							   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_broadcastd_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pbroadcastd128_mask ((__v4si) __A,
						       (__v4si) __O,
						       __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pbroadcastd128_mask ((__v4si) __A,
						       (__v4si)
						       _mm_setzero_si128 (),
						       __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_set1_epi32 (__m128i __O, __mmask8 __M, int __A)
{
  return (__m128i) __builtin_ia32_pbroadcastd128_gpr_mask (__A, (__v4si) __O,
							   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_set1_epi32 (__mmask8 __M, int __A)
{
  return (__m128i)
	 __builtin_ia32_pbroadcastd128_gpr_mask (__A,
						 (__v4si) _mm_setzero_si128 (),
						 __M);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_broadcastq_epi64 (__m256i __O, __mmask8 __M, __m128i __A)
{
  return (__m256i) __builtin_ia32_pbroadcastq256_mask ((__v2di) __A,
						       (__v4di) __O,
						       __M);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A)
{
  return (__m256i) __builtin_ia32_pbroadcastq256_mask ((__v2di) __A,
						       (__v4di)
						       _mm256_setzero_si256 (),
						       __M);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_set1_epi64 (__m256i __O, __mmask8 __M, long long __A)
{
  return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A, (__v4di) __O,
							   __M);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_set1_epi64 (__mmask8 __M, long long __A)
{
  return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A,
							   (__v4di)
							   _mm256_setzero_si256 (),
							   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_broadcastq_epi64 (__m128i __O, __mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pbroadcastq128_mask ((__v2di) __A,
						       (__v2di) __O,
						       __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A)
{
  return (__m128i) __builtin_ia32_pbroadcastq128_mask ((__v2di) __A,
						       (__v2di)
						       _mm_setzero_si128 (),
						       __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_set1_epi64 (__m128i __O, __mmask8 __M, long long __A)
{
  return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A, (__v2di) __O,
							   __M);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_set1_epi64 (__mmask8 __M, long long __A)
{
  return (__m128i)
	 __builtin_ia32_pbroadcastq128_gpr_mask (__A,
						 (__v2di) _mm_setzero_si128 (),
						 __M);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_broadcast_f32x4 (__m128 __A)
{
  return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A,
						          (__v8sf)_mm256_undefined_pd (),
							  (__mmask8) -1);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_broadcast_f32x4 (__m256 __O, __mmask8 __M, __m128 __A)
{
  return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A,
							  (__v8sf) __O,
							  __M);
}

extern __inline __m256
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_broadcast_f32x4 (__mmask8 __M, __m128 __A)
{
  return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A,
							  (__v8sf)
							  _mm256_setzero_ps (),
							  __M);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_broadcast_i32x4 (__m128i __A)
{
  return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si)
							   __A,
						           (__v8si)_mm256_undefined_si256 (),
							   (__mmask8) -1);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_broadcast_i32x4 (__m256i __O, __mmask8 __M, __m128i __A)
{
  return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si)
							   __A,
							   (__v8si)
							   __O, __M);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_broadcast_i32x4 (__mmask8 __M, __m128i __A)
{
  return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si)
							   __A,
							   (__v8si)
							   _mm256_setzero_si256 (),
							   __M);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi8_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A,
						    (__v8si) __W,
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A,
						    (__v8si)
						    _mm256_setzero_si256 (),
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi8_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A,
						    (__v4si) __W,
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A,
						    (__v4si)
						    _mm_setzero_si128 (),
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi8_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A,
						    (__v4di) __W,
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A,
						    (__v4di)
						    _mm256_setzero_si256 (),
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi8_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A,
						    (__v2di) __W,
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A,
						    (__v2di)
						    _mm_setzero_si128 (),
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi16_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A,
						    (__v8si) __W,
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A,
						    (__v8si)
						    _mm256_setzero_si256 (),
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi16_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A,
						    (__v4si) __W,
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A,
						    (__v4si)
						    _mm_setzero_si128 (),
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi16_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A,
						    (__v4di) __W,
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A,
						    (__v4di)
						    _mm256_setzero_si256 (),
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi16_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A,
						    (__v2di) __W,
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A,
						    (__v2di)
						    _mm_setzero_si128 (),
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepi32_epi64 (__m256i __W, __mmask8 __U, __m128i __X)
{
  return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X,
						    (__v4di) __W,
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X)
{
  return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X,
						    (__v4di)
						    _mm256_setzero_si256 (),
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepi32_epi64 (__m128i __W, __mmask8 __U, __m128i __X)
{
  return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X,
						    (__v2di) __W,
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X)
{
  return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X,
						    (__v2di)
						    _mm_setzero_si128 (),
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepu8_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A,
						    (__v8si) __W,
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A,
						    (__v8si)
						    _mm256_setzero_si256 (),
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepu8_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A,
						    (__v4si) __W,
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A,
						    (__v4si)
						    _mm_setzero_si128 (),
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepu8_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A,
						    (__v4di) __W,
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A,
						    (__v4di)
						    _mm256_setzero_si256 (),
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_mask_cvtepu8_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A,
						    (__v2di) __W,
						    (__mmask8) __U);
}

extern __inline __m128i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A)
{
  return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A,
						    (__v2di)
						    _mm_setzero_si128 (),
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_mask_cvtepu16_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A,
						    (__v8si) __W,
						    (__mmask8) __U);
}

extern __inline __m256i
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm256_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A)
{
  return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A,
						    (__v8si)
						    _mm256_setzero_si256 (),
						    (__mmask8) __U);
}

extern __inline __m128i