gcc_builtin_headers_ia32-6.h

Go to the documentation of this file.

// clang-format off
// amx-fp16
void __builtin_ia32_tdpfp16ps(unsigned char, unsigned char, unsigned char);
__gcc_v256si __builtin_ia32_tdpfp16ps_internal(unsigned short, unsigned short, unsigned short, __gcc_v256si, __gcc_v256si, __gcc_v256si);
// avx
short __builtin_ia32_vec_ext_v16hi(__gcc_v16hi, int);
char __builtin_ia32_vec_ext_v32qi(__gcc_v32qi, int);
long long int __builtin_ia32_vec_ext_v4di(__gcc_v4di, int);
int __builtin_ia32_vec_ext_v8si(__gcc_v8si, int);
__gcc_v16hi __builtin_ia32_vec_set_v16hi(__gcc_v16hi, short, int);
__gcc_v32qi __builtin_ia32_vec_set_v32qi(__gcc_v32qi, char, int);
__gcc_v4di __builtin_ia32_vec_set_v4di(__gcc_v4di, long long int, int);
__gcc_v8si __builtin_ia32_vec_set_v8si(__gcc_v8si, int, int);
// avx2
__gcc_v4si __builtin_ia32_gatherd_d(__gcc_v4si, const int *, __gcc_v4si, __gcc_v4si, char);
__gcc_v8si __builtin_ia32_gatherd_d256(__gcc_v8si, const int *, __gcc_v8si, __gcc_v8si, char);
__gcc_v2df __builtin_ia32_gatherd_pd(__gcc_v2df, const double *, __gcc_v4si, __gcc_v2df, char);
__gcc_v4df __builtin_ia32_gatherd_pd256(__gcc_v4df, const double *, __gcc_v4si, __gcc_v4df, char);
__gcc_v4sf __builtin_ia32_gatherd_ps(__gcc_v4sf, const float *, __gcc_v4si, __gcc_v4sf, char);
__gcc_v8sf __builtin_ia32_gatherd_ps256(__gcc_v8sf, const float *, __gcc_v8si, __gcc_v8sf, char);
__gcc_v2di __builtin_ia32_gatherd_q(__gcc_v2di, const long long int *, __gcc_v4si, __gcc_v2di, char);
__gcc_v4di __builtin_ia32_gatherd_q256(__gcc_v4di, const long long int *, __gcc_v4si, __gcc_v4di, char);
__gcc_v4si __builtin_ia32_gatherq_d(__gcc_v4si, const int *, __gcc_v2di, __gcc_v4si, char);
__gcc_v4si __builtin_ia32_gatherq_d256(__gcc_v4si, const int *, __gcc_v4di, __gcc_v4si, char);
__gcc_v2df __builtin_ia32_gatherq_pd(__gcc_v2df, const double *, __gcc_v2di, __gcc_v2df, char);
__gcc_v4df __builtin_ia32_gatherq_pd256(__gcc_v4df, const double *, __gcc_v4di, __gcc_v4df, char);
__gcc_v4sf __builtin_ia32_gatherq_ps(__gcc_v4sf, const float *, __gcc_v2di, __gcc_v4sf, char);
__gcc_v4sf __builtin_ia32_gatherq_ps256(__gcc_v4sf, const float *, __gcc_v4di, __gcc_v4sf, char);
__gcc_v2di __builtin_ia32_gatherq_q(__gcc_v2di, const long long int *, __gcc_v2di, __gcc_v2di, char);
__gcc_v4di __builtin_ia32_gatherq_q256(__gcc_v4di, const long long int *, __gcc_v4di, __gcc_v4di, char);
__gcc_v4di __builtin_ia32_pslldqi256_byteshift(__gcc_v4di, int);
__gcc_v4di __builtin_ia32_psrldqi256_byteshift(__gcc_v4di, int);
// avx512bf16
__gcc_v32hf __builtin_ia32_selectpbf_512(unsigned int, __gcc_v32hf, __gcc_v32hf);
__gcc_v8hf __builtin_ia32_selectsbf_128(unsigned char, __gcc_v8hf, __gcc_v8hf);
// avx512bf16,avx512vl
__gcc_v8hf __builtin_ia32_selectpbf_128(unsigned char, __gcc_v8hf, __gcc_v8hf);
__gcc_v16hf __builtin_ia32_selectpbf_256(unsigned short, __gcc_v16hf, __gcc_v16hf);
// avx512bf16,avx512vl|avxneconvert
__gcc_v8hf __builtin_ia32_vcvtneps2bf16128(__gcc_v4sf);
__gcc_v8hf __builtin_ia32_vcvtneps2bf16256(__gcc_v8sf);
// avx512bitalg
__gcc_v64qi __builtin_ia32_vpopcntb_512(__gcc_v64qi);
__gcc_v32hi __builtin_ia32_vpopcntw_512(__gcc_v32hi);
// avx512bw
__gcc_v32hi __builtin_ia32_dbpsadbw512(__gcc_v64qi, __gcc_v64qi, int);
__gcc_v32hi __builtin_ia32_packssdw512(__gcc_v16si, __gcc_v16si);
__gcc_v64qi __builtin_ia32_packsswb512(__gcc_v32hi, __gcc_v32hi);
__gcc_v32hi __builtin_ia32_packusdw512(__gcc_v16si, __gcc_v16si);
__gcc_v64qi __builtin_ia32_packuswb512(__gcc_v32hi, __gcc_v32hi);
__gcc_v64qi __builtin_ia32_pavgb512(__gcc_v64qi, __gcc_v64qi);
__gcc_v32hi __builtin_ia32_pavgw512(__gcc_v32hi, __gcc_v32hi);
__gcc_v32hi __builtin_ia32_permvarhi512(__gcc_v32hi, __gcc_v32hi);
__gcc_v32hi __builtin_ia32_pmaddubsw512(__gcc_v64qi, __gcc_v64qi);
__gcc_v16si __builtin_ia32_pmaddwd512(__gcc_v32hi, __gcc_v32hi);
__gcc_v32hi __builtin_ia32_pmulhrsw512(__gcc_v32hi, __gcc_v32hi);
__gcc_v32hi __builtin_ia32_pmulhuw512(__gcc_v32hi, __gcc_v32hi);
__gcc_v32hi __builtin_ia32_pmulhw512(__gcc_v32hi, __gcc_v32hi);
__gcc_v64qi __builtin_ia32_pshufb512(__gcc_v64qi, __gcc_v64qi);
__gcc_v32hi __builtin_ia32_pshufhw512(__gcc_v32hi, int);
__gcc_v32hi __builtin_ia32_pshuflw512(__gcc_v32hi, int);
__gcc_v8di __builtin_ia32_pslldqi512_byteshift(__gcc_v8di, int);
__gcc_v32hi __builtin_ia32_psllv32hi(__gcc_v32hi, __gcc_v32hi);
__gcc_v32hi __builtin_ia32_psllw512(__gcc_v32hi, __gcc_v8hi);
__gcc_v32hi __builtin_ia32_psllwi512(__gcc_v32hi, int);
__gcc_v32hi __builtin_ia32_psrav32hi(__gcc_v32hi, __gcc_v32hi);
__gcc_v32hi __builtin_ia32_psraw512(__gcc_v32hi, __gcc_v8hi);
__gcc_v32hi __builtin_ia32_psrawi512(__gcc_v32hi, int);
__gcc_v8di __builtin_ia32_psrldqi512_byteshift(__gcc_v8di, int);
__gcc_v32hi __builtin_ia32_psrlv32hi(__gcc_v32hi, __gcc_v32hi);
__gcc_v32hi __builtin_ia32_psrlw512(__gcc_v32hi, __gcc_v8hi);
__gcc_v32hi __builtin_ia32_psrlwi512(__gcc_v32hi, int);
__gcc_v32hi __builtin_ia32_vpermi2varhi512(__gcc_v32hi, __gcc_v32hi, __gcc_v32hi);
// avx512bw,avx512vl
__gcc_v8hi __builtin_ia32_dbpsadbw128(__gcc_v16qi, __gcc_v16qi, int);
__gcc_v16hi __builtin_ia32_dbpsadbw256(__gcc_v32qi, __gcc_v32qi, int);
__gcc_v8hi __builtin_ia32_permvarhi128(__gcc_v8hi, __gcc_v8hi);
__gcc_v16hi __builtin_ia32_permvarhi256(__gcc_v16hi, __gcc_v16hi);
__gcc_v16hi __builtin_ia32_psllv16hi(__gcc_v16hi, __gcc_v16hi);
__gcc_v8hi __builtin_ia32_psllv8hi(__gcc_v8hi, __gcc_v8hi);
__gcc_v16hi __builtin_ia32_psrav16hi(__gcc_v16hi, __gcc_v16hi);
__gcc_v8hi __builtin_ia32_psrav8hi(__gcc_v8hi, __gcc_v8hi);
__gcc_v16hi __builtin_ia32_psrlv16hi(__gcc_v16hi, __gcc_v16hi);
__gcc_v8hi __builtin_ia32_psrlv8hi(__gcc_v8hi, __gcc_v8hi);
// avx512cd
__gcc_v8di __builtin_ia32_vpconflictdi_512(__gcc_v8di);
__gcc_v16si __builtin_ia32_vpconflictsi_512(__gcc_v16si);
__gcc_v16si __builtin_ia32_vplzcntd_512(__gcc_v16si);
__gcc_v8di __builtin_ia32_vplzcntq_512(__gcc_v8di);
// avx512cd,avx512vl
__gcc_v2di __builtin_ia32_vpconflictdi_128(__gcc_v2di);
__gcc_v4di __builtin_ia32_vpconflictdi_256(__gcc_v4di);
__gcc_v4si __builtin_ia32_vpconflictsi_128(__gcc_v4si);
__gcc_v8si __builtin_ia32_vpconflictsi_256(__gcc_v8si);
__gcc_v4si __builtin_ia32_vplzcntd_128(__gcc_v4si);
__gcc_v8si __builtin_ia32_vplzcntd_256(__gcc_v8si);
__gcc_v2di __builtin_ia32_vplzcntq_128(__gcc_v2di);
__gcc_v4di __builtin_ia32_vplzcntq_256(__gcc_v4di);
// avx512dq
__gcc_v16sf __builtin_ia32_insertf32x8(__gcc_v16sf, __gcc_v8sf, int);
__gcc_v8df __builtin_ia32_insertf64x2_512(__gcc_v8df, __gcc_v2df, int);
__gcc_v16si __builtin_ia32_inserti32x8(__gcc_v16si, __gcc_v8si, int);
__gcc_v8di __builtin_ia32_inserti64x2_512(__gcc_v8di, __gcc_v2di, int);
__gcc_v2df __builtin_ia32_rangesd128_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int, int);
__gcc_v4sf __builtin_ia32_rangess128_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int, int);
// avx512dq,avx512vl
__gcc_v4df __builtin_ia32_insertf64x2_256(__gcc_v4df, __gcc_v2df, int);
__gcc_v4di __builtin_ia32_inserti64x2_256(__gcc_v4di, __gcc_v2di, int);
// avx512er
__gcc_v2df __builtin_ia32_rcp28sd_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int);
__gcc_v4sf __builtin_ia32_rcp28ss_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v2df __builtin_ia32_rsqrt28sd_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int);
__gcc_v4sf __builtin_ia32_rsqrt28ss_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
// avx512f
__gcc_v8df __builtin_ia32_addpd512(__gcc_v8df, __gcc_v8df, int);
__gcc_v16sf __builtin_ia32_addps512(__gcc_v16sf, __gcc_v16sf, int);
__gcc_v2df __builtin_ia32_addsd_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int);
__gcc_v4sf __builtin_ia32_addss_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v16si __builtin_ia32_alignd512(__gcc_v16si, __gcc_v16si, int);
__gcc_v8di __builtin_ia32_alignq512(__gcc_v8di, __gcc_v8di, int);
__gcc_v8df __builtin_ia32_divpd512(__gcc_v8df, __gcc_v8df, int);
__gcc_v16sf __builtin_ia32_divps512(__gcc_v16sf, __gcc_v16sf, int);
__gcc_v2df __builtin_ia32_divsd_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int);
__gcc_v4sf __builtin_ia32_divss_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v2df __builtin_ia32_getexpsd128_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int);
__gcc_v4sf __builtin_ia32_getexpss128_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v2df __builtin_ia32_getmantsd_round_mask(__gcc_v2df, __gcc_v2df, int, __gcc_v2df, unsigned char, int);
__gcc_v4sf __builtin_ia32_getmantss_round_mask(__gcc_v4sf, __gcc_v4sf, int, __gcc_v4sf, unsigned char, int);
__gcc_v16sf __builtin_ia32_insertf32x4(__gcc_v16sf, __gcc_v4sf, int);
__gcc_v8df __builtin_ia32_insertf64x4(__gcc_v8df, __gcc_v4df, int);
__gcc_v16si __builtin_ia32_inserti32x4(__gcc_v16si, __gcc_v4si, int);
__gcc_v8di __builtin_ia32_inserti64x4(__gcc_v8di, __gcc_v4di, int);
__gcc_v2df __builtin_ia32_loadsd128_mask(const __gcc_v2df *, __gcc_v2df, unsigned char);
__gcc_v4sf __builtin_ia32_loadss128_mask(const __gcc_v4sf *, __gcc_v4sf, unsigned char);
__gcc_v8df __builtin_ia32_maxpd512(__gcc_v8df, __gcc_v8df, int);
__gcc_v16sf __builtin_ia32_maxps512(__gcc_v16sf, __gcc_v16sf, int);
__gcc_v2df __builtin_ia32_maxsd_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int);
__gcc_v4sf __builtin_ia32_maxss_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v8df __builtin_ia32_minpd512(__gcc_v8df, __gcc_v8df, int);
__gcc_v16sf __builtin_ia32_minps512(__gcc_v16sf, __gcc_v16sf, int);
__gcc_v2df __builtin_ia32_minsd_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int);
__gcc_v4sf __builtin_ia32_minss_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v8df __builtin_ia32_mulpd512(__gcc_v8df, __gcc_v8df, int);
__gcc_v16sf __builtin_ia32_mulps512(__gcc_v16sf, __gcc_v16sf, int);
__gcc_v2df __builtin_ia32_mulsd_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int);
__gcc_v4sf __builtin_ia32_mulss_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v8df __builtin_ia32_permdf512(__gcc_v8df, int);
__gcc_v8di __builtin_ia32_permdi512(__gcc_v8di, int);
__gcc_v8df __builtin_ia32_permvardf512(__gcc_v8df, __gcc_v8di);
__gcc_v8di __builtin_ia32_permvardi512(__gcc_v8di, __gcc_v8di);
__gcc_v16sf __builtin_ia32_permvarsf512(__gcc_v16sf, __gcc_v16si);
__gcc_v16si __builtin_ia32_permvarsi512(__gcc_v16si, __gcc_v16si);
__gcc_v8di __builtin_ia32_pmuldq512(__gcc_v16si, __gcc_v16si);
__gcc_v8di __builtin_ia32_pmuludq512(__gcc_v16si, __gcc_v16si);
__gcc_v16si __builtin_ia32_prold512(__gcc_v16si, int);
__gcc_v8di __builtin_ia32_prolq512(__gcc_v8di, int);
__gcc_v16si __builtin_ia32_prolvd512(__gcc_v16si, __gcc_v16si);
__gcc_v8di __builtin_ia32_prolvq512(__gcc_v8di, __gcc_v8di);
__gcc_v16si __builtin_ia32_prord512(__gcc_v16si, int);
__gcc_v8di __builtin_ia32_prorq512(__gcc_v8di, int);
__gcc_v16si __builtin_ia32_prorvd512(__gcc_v16si, __gcc_v16si);
__gcc_v8di __builtin_ia32_prorvq512(__gcc_v8di, __gcc_v8di);
__gcc_v16si __builtin_ia32_pshufd512(__gcc_v16si, int);
__gcc_v16si __builtin_ia32_pslld512(__gcc_v16si, __gcc_v4si);
__gcc_v16si __builtin_ia32_pslldi512(__gcc_v16si, int);
__gcc_v8di __builtin_ia32_psllq512(__gcc_v8di, __gcc_v2di);
__gcc_v8di __builtin_ia32_psllqi512(__gcc_v8di, int);
__gcc_v16si __builtin_ia32_psllv16si(__gcc_v16si, __gcc_v16si);
__gcc_v8di __builtin_ia32_psllv8di(__gcc_v8di, __gcc_v8di);
__gcc_v16si __builtin_ia32_psrad512(__gcc_v16si, __gcc_v4si);
__gcc_v16si __builtin_ia32_psradi512(__gcc_v16si, int);
__gcc_v8di __builtin_ia32_psraq512(__gcc_v8di, __gcc_v2di);
__gcc_v8di __builtin_ia32_psraqi512(__gcc_v8di, int);
__gcc_v16si __builtin_ia32_psrav16si(__gcc_v16si, __gcc_v16si);
__gcc_v8di __builtin_ia32_psrav8di(__gcc_v8di, __gcc_v8di);
__gcc_v16si __builtin_ia32_psrld512(__gcc_v16si, __gcc_v4si);
__gcc_v16si __builtin_ia32_psrldi512(__gcc_v16si, int);
__gcc_v8di __builtin_ia32_psrlq512(__gcc_v8di, __gcc_v2di);
__gcc_v8di __builtin_ia32_psrlqi512(__gcc_v8di, int);
__gcc_v16si __builtin_ia32_psrlv16si(__gcc_v16si, __gcc_v16si);
__gcc_v8di __builtin_ia32_psrlv8di(__gcc_v8di, __gcc_v8di);
double __builtin_ia32_reduce_fadd_pd512(double, __gcc_v8df);
float __builtin_ia32_reduce_fadd_ps512(float, __gcc_v16sf);
double __builtin_ia32_reduce_fmax_pd512(__gcc_v8df);
float __builtin_ia32_reduce_fmax_ps512(__gcc_v16sf);
double __builtin_ia32_reduce_fmin_pd512(__gcc_v8df);
float __builtin_ia32_reduce_fmin_ps512(__gcc_v16sf);
double __builtin_ia32_reduce_fmul_pd512(double, __gcc_v8df);
float __builtin_ia32_reduce_fmul_ps512(float, __gcc_v16sf);
__gcc_v2df __builtin_ia32_rndscalesd_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int, int);
__gcc_v4sf __builtin_ia32_rndscaless_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int, int);
__gcc_v2df __builtin_ia32_scalefsd_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int);
__gcc_v4sf __builtin_ia32_scalefss_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v8di __builtin_ia32_selectq_512(unsigned char, __gcc_v8di, __gcc_v8di);
__gcc_v16sf __builtin_ia32_shuf_f32x4(__gcc_v16sf, __gcc_v16sf, int);
__gcc_v8df __builtin_ia32_shuf_f64x2(__gcc_v8df, __gcc_v8df, int);
__gcc_v16si __builtin_ia32_shuf_i32x4(__gcc_v16si, __gcc_v16si, int);
__gcc_v8di __builtin_ia32_shuf_i64x2(__gcc_v8di, __gcc_v8di, int);
__gcc_v8df __builtin_ia32_shufpd512(__gcc_v8df, __gcc_v8df, int);
__gcc_v16sf __builtin_ia32_shufps512(__gcc_v16sf, __gcc_v16sf, int);
__gcc_v2df __builtin_ia32_sqrtsd_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int);
__gcc_v4sf __builtin_ia32_sqrtss_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
void __builtin_ia32_storesd128_mask(__gcc_v2df *, __gcc_v2df, unsigned char);
void __builtin_ia32_storess128_mask(__gcc_v4sf *, __gcc_v4sf, unsigned char);
__gcc_v8df __builtin_ia32_subpd512(__gcc_v8df, __gcc_v8df, int);
__gcc_v16sf __builtin_ia32_subps512(__gcc_v16sf, __gcc_v16sf, int);
__gcc_v2df __builtin_ia32_subsd_round_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char, int);
__gcc_v4sf __builtin_ia32_subss_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v16si __builtin_ia32_vpermi2vard512(__gcc_v16si, __gcc_v16si, __gcc_v16si);
__gcc_v8df __builtin_ia32_vpermi2varpd512(__gcc_v8df, __gcc_v8di, __gcc_v8df);
__gcc_v16sf __builtin_ia32_vpermi2varps512(__gcc_v16sf, __gcc_v16si, __gcc_v16sf);
__gcc_v8di __builtin_ia32_vpermi2varq512(__gcc_v8di, __gcc_v8di, __gcc_v8di);
__gcc_v8df __builtin_ia32_vpermilpd512(__gcc_v8df, int);
__gcc_v16sf __builtin_ia32_vpermilps512(__gcc_v16sf, int);
__gcc_v8df __builtin_ia32_vpermilvarpd512(__gcc_v8df, __gcc_v8di);
__gcc_v16sf __builtin_ia32_vpermilvarps512(__gcc_v16sf, __gcc_v16si);
// avx512f,vaes
__gcc_v8di __builtin_ia32_aesdec512(__gcc_v8di, __gcc_v8di);
__gcc_v8di __builtin_ia32_aesdeclast512(__gcc_v8di, __gcc_v8di);
__gcc_v8di __builtin_ia32_aesenc512(__gcc_v8di, __gcc_v8di);
__gcc_v8di __builtin_ia32_aesenclast512(__gcc_v8di, __gcc_v8di);
// avx512f,vpclmulqdq
__gcc_v8di __builtin_ia32_pclmulqdq512(__gcc_v8di, __gcc_v8di, char);
// avx512fp16
__gcc_v32hf __builtin_ia32_addph512(__gcc_v32hf, __gcc_v32hf, int);
__gcc_v8hf __builtin_ia32_addsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
unsigned int __builtin_ia32_cmpph512_mask(__gcc_v32hf, __gcc_v32hf, int, unsigned int, int);
unsigned char __builtin_ia32_cmpsh_mask(__gcc_v8hf, __gcc_v8hf, int, unsigned char, int);
__gcc_v32hf __builtin_ia32_divph512(__gcc_v32hf, __gcc_v32hf, int);
__gcc_v8hf __builtin_ia32_divsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
__gcc_v8hf __builtin_ia32_getexpsh128_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
__gcc_v8hf __builtin_ia32_getmantsh_round_mask(__gcc_v8hf, __gcc_v8hf, int, __gcc_v8hf, unsigned char, int);
__gcc_v8hf __builtin_ia32_loadsh128_mask(const __gcc_v8hf *, __gcc_v8hf, unsigned char);
__gcc_v32hf __builtin_ia32_maxph512(__gcc_v32hf, __gcc_v32hf, int);
__gcc_v8hf __builtin_ia32_maxsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
__gcc_v32hf __builtin_ia32_minph512(__gcc_v32hf, __gcc_v32hf, int);
__gcc_v8hf __builtin_ia32_minsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
__gcc_v32hf __builtin_ia32_mulph512(__gcc_v32hf, __gcc_v32hf, int);
__gcc_v8hf __builtin_ia32_mulsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
_Float16 __builtin_ia32_reduce_fadd_ph512(_Float16, __gcc_v32hf);
_Float16 __builtin_ia32_reduce_fmax_ph512(__gcc_v32hf);
_Float16 __builtin_ia32_reduce_fmin_ph512(__gcc_v32hf);
_Float16 __builtin_ia32_reduce_fmul_ph512(_Float16, __gcc_v32hf);
__gcc_v32hf __builtin_ia32_reduceph512_mask(__gcc_v32hf, int, __gcc_v32hf, unsigned int, int);
__gcc_v8hf __builtin_ia32_reducesh_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int, int);
__gcc_v32hf __builtin_ia32_rndscaleph_mask(__gcc_v32hf, int, __gcc_v32hf, unsigned int, int);
__gcc_v8hf __builtin_ia32_rndscalesh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int, int);
__gcc_v32hf __builtin_ia32_scalefph512_mask(__gcc_v32hf, __gcc_v32hf, __gcc_v32hf, unsigned int, int);
__gcc_v8hf __builtin_ia32_scalefsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
__gcc_v32hf __builtin_ia32_selectph_512(unsigned int, __gcc_v32hf, __gcc_v32hf);
__gcc_v8hf __builtin_ia32_selectsh_128(unsigned char, __gcc_v8hf, __gcc_v8hf);
__gcc_v32hf __builtin_ia32_sqrtph512(__gcc_v32hf, int);
__gcc_v8hf __builtin_ia32_sqrtsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
void __builtin_ia32_storesh128_mask(__gcc_v8hf *, __gcc_v8hf, unsigned char);
__gcc_v32hf __builtin_ia32_subph512(__gcc_v32hf, __gcc_v32hf, int);
__gcc_v8hf __builtin_ia32_subsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
int __builtin_ia32_vcomish(__gcc_v8hf, __gcc_v8hf, int, int);
__gcc_v16hf __builtin_ia32_vcvtdq2ph512_mask(__gcc_v16si, __gcc_v16hf, unsigned short, int);
__gcc_v8hf __builtin_ia32_vcvtpd2ph512_mask(__gcc_v8df, __gcc_v8hf, unsigned char, int);
__gcc_v16si __builtin_ia32_vcvtph2dq512_mask(__gcc_v16hf, __gcc_v16si, unsigned short, int);
__gcc_v8df __builtin_ia32_vcvtph2pd512_mask(__gcc_v8hf, __gcc_v8df, unsigned char, int);
__gcc_v16sf __builtin_ia32_vcvtph2psx512_mask(__gcc_v16hf, __gcc_v16sf, unsigned short, int);
__gcc_v8di __builtin_ia32_vcvtph2qq512_mask(__gcc_v8hf, __gcc_v8di, unsigned char, int);
__gcc_v16usi __builtin_ia32_vcvtph2udq512_mask(__gcc_v16hf, __gcc_v16usi, unsigned short, int);
__gcc_v8udi __builtin_ia32_vcvtph2uqq512_mask(__gcc_v8hf, __gcc_v8udi, unsigned char, int);
__gcc_v32uhi __builtin_ia32_vcvtph2uw512_mask(__gcc_v32hf, __gcc_v32uhi, unsigned int, int);
__gcc_v32hi __builtin_ia32_vcvtph2w512_mask(__gcc_v32hf, __gcc_v32hi, unsigned int, int);
__gcc_v16hf __builtin_ia32_vcvtps2phx512_mask(__gcc_v16sf, __gcc_v16hf, unsigned short, int);
__gcc_v8hf __builtin_ia32_vcvtqq2ph512_mask(__gcc_v8di, __gcc_v8hf, unsigned char, int);
__gcc_v8hf __builtin_ia32_vcvtsd2sh_round_mask(__gcc_v8hf, __gcc_v2df, __gcc_v8hf, unsigned char, int);
__gcc_v2df __builtin_ia32_vcvtsh2sd_round_mask(__gcc_v2df, __gcc_v8hf, __gcc_v2df, unsigned char, int);
int __builtin_ia32_vcvtsh2si32(__gcc_v8hf, int);
long long int __builtin_ia32_vcvtsh2si64(__gcc_v8hf, int);
__gcc_v4sf __builtin_ia32_vcvtsh2ss_round_mask(__gcc_v4sf, __gcc_v8hf, __gcc_v4sf, unsigned char, int);
unsigned int __builtin_ia32_vcvtsh2usi32(__gcc_v8hf, int);
unsigned long long int __builtin_ia32_vcvtsh2usi64(__gcc_v8hf, int);
__gcc_v8hf __builtin_ia32_vcvtsi2sh(__gcc_v8hf, int, int);
__gcc_v8hf __builtin_ia32_vcvtsi642sh(__gcc_v8hf, long long int, int);
__gcc_v8hf __builtin_ia32_vcvtss2sh_round_mask(__gcc_v8hf, __gcc_v4sf, __gcc_v8hf, unsigned char, int);
__gcc_v16si __builtin_ia32_vcvttph2dq512_mask(__gcc_v16hf, __gcc_v16si, unsigned short, int);
__gcc_v8di __builtin_ia32_vcvttph2qq512_mask(__gcc_v8hf, __gcc_v8di, unsigned char, int);
__gcc_v16usi __builtin_ia32_vcvttph2udq512_mask(__gcc_v16hf, __gcc_v16usi, unsigned short, int);
__gcc_v8udi __builtin_ia32_vcvttph2uqq512_mask(__gcc_v8hf, __gcc_v8udi, unsigned char, int);
__gcc_v32uhi __builtin_ia32_vcvttph2uw512_mask(__gcc_v32hf, __gcc_v32uhi, unsigned int, int);
__gcc_v32hi __builtin_ia32_vcvttph2w512_mask(__gcc_v32hf, __gcc_v32hi, unsigned int, int);
int __builtin_ia32_vcvttsh2si32(__gcc_v8hf, int);
long long int __builtin_ia32_vcvttsh2si64(__gcc_v8hf, int);
unsigned int __builtin_ia32_vcvttsh2usi32(__gcc_v8hf, int);
unsigned long long int __builtin_ia32_vcvttsh2usi64(__gcc_v8hf, int);
__gcc_v16hf __builtin_ia32_vcvtudq2ph512_mask(__gcc_v16usi, __gcc_v16hf, unsigned short, int);
__gcc_v8hf __builtin_ia32_vcvtuqq2ph512_mask(__gcc_v8udi, __gcc_v8hf, unsigned char, int);
__gcc_v8hf __builtin_ia32_vcvtusi2sh(__gcc_v8hf, unsigned int, int);
__gcc_v8hf __builtin_ia32_vcvtusi642sh(__gcc_v8hf, unsigned long long int, int);
__gcc_v32hf __builtin_ia32_vcvtuw2ph512_mask(__gcc_v32uhi, __gcc_v32hf, unsigned int, int);
__gcc_v32hf __builtin_ia32_vcvtw2ph512_mask(__gcc_v32hi, __gcc_v32hf, unsigned int, int);
__gcc_v16sf __builtin_ia32_vfcmaddcph512_mask(__gcc_v16sf, __gcc_v16sf, __gcc_v16sf, unsigned short, int);
__gcc_v16sf __builtin_ia32_vfcmaddcph512_mask3(__gcc_v16sf, __gcc_v16sf, __gcc_v16sf, unsigned short, int);
__gcc_v16sf __builtin_ia32_vfcmaddcph512_maskz(__gcc_v16sf, __gcc_v16sf, __gcc_v16sf, unsigned short, int);
__gcc_v4sf __builtin_ia32_vfcmaddcsh_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v4sf __builtin_ia32_vfcmaddcsh_maskz(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v4sf __builtin_ia32_vfcmaddcsh_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v4sf __builtin_ia32_vfcmaddcsh_round_mask3(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v16sf __builtin_ia32_vfcmulcph512_mask(__gcc_v16sf, __gcc_v16sf, __gcc_v16sf, unsigned short, int);
__gcc_v4sf __builtin_ia32_vfcmulcsh_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v16sf __builtin_ia32_vfmaddcph512_mask(__gcc_v16sf, __gcc_v16sf, __gcc_v16sf, unsigned short, int);
__gcc_v16sf __builtin_ia32_vfmaddcph512_mask3(__gcc_v16sf, __gcc_v16sf, __gcc_v16sf, unsigned short, int);
__gcc_v16sf __builtin_ia32_vfmaddcph512_maskz(__gcc_v16sf, __gcc_v16sf, __gcc_v16sf, unsigned short, int);
__gcc_v4sf __builtin_ia32_vfmaddcsh_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v4sf __builtin_ia32_vfmaddcsh_maskz(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v4sf __builtin_ia32_vfmaddcsh_round_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v4sf __builtin_ia32_vfmaddcsh_round_mask3(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
__gcc_v16sf __builtin_ia32_vfmulcph512_mask(__gcc_v16sf, __gcc_v16sf, __gcc_v16sf, unsigned short, int);
__gcc_v4sf __builtin_ia32_vfmulcsh_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, int);
// avx512fp16,avx512vl
unsigned char __builtin_ia32_cmpph128_mask(__gcc_v8hf, __gcc_v8hf, int, unsigned char);
unsigned short __builtin_ia32_cmpph256_mask(__gcc_v16hf, __gcc_v16hf, int, unsigned short);
__gcc_v8hf __builtin_ia32_maxph128(__gcc_v8hf, __gcc_v8hf);
__gcc_v16hf __builtin_ia32_maxph256(__gcc_v16hf, __gcc_v16hf);
__gcc_v8hf __builtin_ia32_minph128(__gcc_v8hf, __gcc_v8hf);
__gcc_v16hf __builtin_ia32_minph256(__gcc_v16hf, __gcc_v16hf);
_Float16 __builtin_ia32_reduce_fadd_ph128(_Float16, __gcc_v8hf);
_Float16 __builtin_ia32_reduce_fadd_ph256(_Float16, __gcc_v16hf);
_Float16 __builtin_ia32_reduce_fmax_ph128(__gcc_v8hf);
_Float16 __builtin_ia32_reduce_fmax_ph256(__gcc_v16hf);
_Float16 __builtin_ia32_reduce_fmin_ph128(__gcc_v8hf);
_Float16 __builtin_ia32_reduce_fmin_ph256(__gcc_v16hf);
_Float16 __builtin_ia32_reduce_fmul_ph128(_Float16, __gcc_v8hf);
_Float16 __builtin_ia32_reduce_fmul_ph256(_Float16, __gcc_v16hf);
__gcc_v8hf __builtin_ia32_rndscaleph_128_mask(__gcc_v8hf, int, __gcc_v8hf, unsigned char);
__gcc_v16hf __builtin_ia32_rndscaleph_256_mask(__gcc_v16hf, int, __gcc_v16hf, unsigned short);
__gcc_v8hf __builtin_ia32_selectph_128(unsigned char, __gcc_v8hf, __gcc_v8hf);
__gcc_v16hf __builtin_ia32_selectph_256(unsigned short, __gcc_v16hf, __gcc_v16hf);
__gcc_v8hf __builtin_ia32_sqrtph(__gcc_v8hf);
__gcc_v16hf __builtin_ia32_sqrtph256(__gcc_v16hf);
__gcc_v8hf __builtin_ia32_vfmaddph(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf);
__gcc_v16hf __builtin_ia32_vfmaddph256(__gcc_v16hf, __gcc_v16hf, __gcc_v16hf);
__gcc_v8hf __builtin_ia32_vfmaddsubph(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf);
__gcc_v16hf __builtin_ia32_vfmaddsubph256(__gcc_v16hf, __gcc_v16hf, __gcc_v16hf);
// avx512ifma
__gcc_v8di __builtin_ia32_vpmadd52huq512(__gcc_v8di, __gcc_v8di, __gcc_v8di);
__gcc_v8di __builtin_ia32_vpmadd52luq512(__gcc_v8di, __gcc_v8di, __gcc_v8di);
// avx512ifma,avx512vl|avxifma
__gcc_v2di __builtin_ia32_vpmadd52huq128(__gcc_v2di, __gcc_v2di, __gcc_v2di);
__gcc_v4di __builtin_ia32_vpmadd52huq256(__gcc_v4di, __gcc_v4di, __gcc_v4di);
__gcc_v2di __builtin_ia32_vpmadd52luq128(__gcc_v2di, __gcc_v2di, __gcc_v2di);
__gcc_v4di __builtin_ia32_vpmadd52luq256(__gcc_v4di, __gcc_v4di, __gcc_v4di);
// avx512vbmi
__gcc_v64qi __builtin_ia32_permvarqi512(__gcc_v64qi, __gcc_v64qi);
__gcc_v64qi __builtin_ia32_vpermi2varqi512(__gcc_v64qi, __gcc_v64qi, __gcc_v64qi);
__gcc_v64qi __builtin_ia32_vpmultishiftqb512(__gcc_v64qi, __gcc_v64qi);
// avx512vbmi,avx512vl
__gcc_v16qi __builtin_ia32_permvarqi128(__gcc_v16qi, __gcc_v16qi);
__gcc_v32qi __builtin_ia32_permvarqi256(__gcc_v32qi, __gcc_v32qi);
__gcc_v16qi __builtin_ia32_vpermi2varqi128(__gcc_v16qi, __gcc_v16qi, __gcc_v16qi);
__gcc_v32qi __builtin_ia32_vpermi2varqi256(__gcc_v32qi, __gcc_v32qi, __gcc_v32qi);
__gcc_v16qi __builtin_ia32_vpmultishiftqb128(__gcc_v16qi, __gcc_v16qi);
__gcc_v32qi __builtin_ia32_vpmultishiftqb256(__gcc_v32qi, __gcc_v32qi);
// avx512vbmi2
void __builtin_ia32_compressstorehi512_mask(__gcc_v32hi *, __gcc_v32hi, unsigned int);
void __builtin_ia32_compressstoreqi512_mask(__gcc_v64qi *, __gcc_v64qi, unsigned long long int);
__gcc_v16si __builtin_ia32_vpshldd512(__gcc_v16si, __gcc_v16si, int);
__gcc_v8di __builtin_ia32_vpshldq512(__gcc_v8di, __gcc_v8di, int);
__gcc_v16si __builtin_ia32_vpshldvd512(__gcc_v16si, __gcc_v16si, __gcc_v16si);
__gcc_v8di __builtin_ia32_vpshldvq512(__gcc_v8di, __gcc_v8di, __gcc_v8di);
__gcc_v32hi __builtin_ia32_vpshldvw512(__gcc_v32hi, __gcc_v32hi, __gcc_v32hi);
__gcc_v32hi __builtin_ia32_vpshldw512(__gcc_v32hi, __gcc_v32hi, int);
__gcc_v16si __builtin_ia32_vpshrdd512(__gcc_v16si, __gcc_v16si, int);
__gcc_v8di __builtin_ia32_vpshrdq512(__gcc_v8di, __gcc_v8di, int);
__gcc_v16si __builtin_ia32_vpshrdvd512(__gcc_v16si, __gcc_v16si, __gcc_v16si);
__gcc_v8di __builtin_ia32_vpshrdvq512(__gcc_v8di, __gcc_v8di, __gcc_v8di);
__gcc_v32hi __builtin_ia32_vpshrdvw512(__gcc_v32hi, __gcc_v32hi, __gcc_v32hi);
__gcc_v32hi __builtin_ia32_vpshrdw512(__gcc_v32hi, __gcc_v32hi, int);
// avx512vl
__gcc_v4si __builtin_ia32_alignd128(__gcc_v4si, __gcc_v4si, int);
__gcc_v8si __builtin_ia32_alignd256(__gcc_v8si, __gcc_v8si, int);
__gcc_v2di __builtin_ia32_alignq128(__gcc_v2di, __gcc_v2di, int);
__gcc_v4di __builtin_ia32_alignq256(__gcc_v4di, __gcc_v4di, int);
__gcc_v8sf __builtin_ia32_insertf32x4_256(__gcc_v8sf, __gcc_v4sf, int);
__gcc_v8si __builtin_ia32_inserti32x4_256(__gcc_v8si, __gcc_v4si, int);
__gcc_v4df __builtin_ia32_permvardf256(__gcc_v4df, __gcc_v4di);
__gcc_v4di __builtin_ia32_permvardi256(__gcc_v4di, __gcc_v4di);
__gcc_v4si __builtin_ia32_prold128(__gcc_v4si, int);
__gcc_v8si __builtin_ia32_prold256(__gcc_v8si, int);
__gcc_v2di __builtin_ia32_prolq128(__gcc_v2di, int);
__gcc_v4di __builtin_ia32_prolq256(__gcc_v4di, int);
__gcc_v4si __builtin_ia32_prolvd128(__gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_prolvd256(__gcc_v8si, __gcc_v8si);
__gcc_v2di __builtin_ia32_prolvq128(__gcc_v2di, __gcc_v2di);
__gcc_v4di __builtin_ia32_prolvq256(__gcc_v4di, __gcc_v4di);
__gcc_v4si __builtin_ia32_prord128(__gcc_v4si, int);
__gcc_v8si __builtin_ia32_prord256(__gcc_v8si, int);
__gcc_v2di __builtin_ia32_prorq128(__gcc_v2di, int);
__gcc_v4di __builtin_ia32_prorq256(__gcc_v4di, int);
__gcc_v4si __builtin_ia32_prorvd128(__gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_prorvd256(__gcc_v8si, __gcc_v8si);
__gcc_v2di __builtin_ia32_prorvq128(__gcc_v2di, __gcc_v2di);
__gcc_v4di __builtin_ia32_prorvq256(__gcc_v4di, __gcc_v4di);
__gcc_v2di __builtin_ia32_psraq128(__gcc_v2di, __gcc_v2di);
__gcc_v4di __builtin_ia32_psraq256(__gcc_v4di, __gcc_v2di);
__gcc_v2di __builtin_ia32_psraqi128(__gcc_v2di, int);
__gcc_v4di __builtin_ia32_psraqi256(__gcc_v4di, int);
__gcc_v2di __builtin_ia32_psravq128(__gcc_v2di, __gcc_v2di);
__gcc_v4di __builtin_ia32_psravq256(__gcc_v4di, __gcc_v4di);
__gcc_v2di __builtin_ia32_selectq_128(unsigned char, __gcc_v2di, __gcc_v2di);
__gcc_v4di __builtin_ia32_selectq_256(unsigned char, __gcc_v4di, __gcc_v4di);
__gcc_v8sf __builtin_ia32_shuf_f32x4_256(__gcc_v8sf, __gcc_v8sf, int);
__gcc_v4df __builtin_ia32_shuf_f64x2_256(__gcc_v4df, __gcc_v4df, int);
__gcc_v8si __builtin_ia32_shuf_i32x4_256(__gcc_v8si, __gcc_v8si, int);
__gcc_v4di __builtin_ia32_shuf_i64x2_256(__gcc_v4di, __gcc_v4di, int);
__gcc_v4si __builtin_ia32_vpermi2vard128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpermi2vard256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v2df __builtin_ia32_vpermi2varpd128(__gcc_v2df, __gcc_v2di, __gcc_v2df);
__gcc_v4df __builtin_ia32_vpermi2varpd256(__gcc_v4df, __gcc_v4di, __gcc_v4df);
__gcc_v4sf __builtin_ia32_vpermi2varps128(__gcc_v4sf, __gcc_v4si, __gcc_v4sf);
__gcc_v8sf __builtin_ia32_vpermi2varps256(__gcc_v8sf, __gcc_v8si, __gcc_v8sf);
__gcc_v2di __builtin_ia32_vpermi2varq128(__gcc_v2di, __gcc_v2di, __gcc_v2di);
__gcc_v4di __builtin_ia32_vpermi2varq256(__gcc_v4di, __gcc_v4di, __gcc_v4di);
// avx512vl,avx512bitalg
__gcc_v16qi __builtin_ia32_vpopcntb_128(__gcc_v16qi);
__gcc_v32qi __builtin_ia32_vpopcntb_256(__gcc_v32qi);
__gcc_v8hi __builtin_ia32_vpopcntw_128(__gcc_v8hi);
__gcc_v16hi __builtin_ia32_vpopcntw_256(__gcc_v16hi);
// avx512vl,avx512bw
__gcc_v8hi __builtin_ia32_vpermi2varhi128(__gcc_v8hi, __gcc_v8hi, __gcc_v8hi);
__gcc_v16hi __builtin_ia32_vpermi2varhi256(__gcc_v16hi, __gcc_v16hi, __gcc_v16hi);
// avx512vl,avx512vbmi2
void __builtin_ia32_compressstorehi128_mask(__gcc_v8hi *, __gcc_v8hi, unsigned char);
void __builtin_ia32_compressstorehi256_mask(__gcc_v16hi *, __gcc_v16hi, unsigned short);
void __builtin_ia32_compressstoreqi128_mask(__gcc_v16qi *, __gcc_v16qi, unsigned short);
void __builtin_ia32_compressstoreqi256_mask(__gcc_v32qi *, __gcc_v32qi, unsigned int);
__gcc_v4si __builtin_ia32_vpshldd128(__gcc_v4si, __gcc_v4si, int);
__gcc_v8si __builtin_ia32_vpshldd256(__gcc_v8si, __gcc_v8si, int);
__gcc_v2di __builtin_ia32_vpshldq128(__gcc_v2di, __gcc_v2di, int);
__gcc_v4di __builtin_ia32_vpshldq256(__gcc_v4di, __gcc_v4di, int);
__gcc_v4si __builtin_ia32_vpshldvd128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpshldvd256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v2di __builtin_ia32_vpshldvq128(__gcc_v2di, __gcc_v2di, __gcc_v2di);
__gcc_v4di __builtin_ia32_vpshldvq256(__gcc_v4di, __gcc_v4di, __gcc_v4di);
__gcc_v8hi __builtin_ia32_vpshldvw128(__gcc_v8hi, __gcc_v8hi, __gcc_v8hi);
__gcc_v16hi __builtin_ia32_vpshldvw256(__gcc_v16hi, __gcc_v16hi, __gcc_v16hi);
__gcc_v8hi __builtin_ia32_vpshldw128(__gcc_v8hi, __gcc_v8hi, int);
__gcc_v16hi __builtin_ia32_vpshldw256(__gcc_v16hi, __gcc_v16hi, int);
__gcc_v4si __builtin_ia32_vpshrdd128(__gcc_v4si, __gcc_v4si, int);
__gcc_v8si __builtin_ia32_vpshrdd256(__gcc_v8si, __gcc_v8si, int);
__gcc_v2di __builtin_ia32_vpshrdq128(__gcc_v2di, __gcc_v2di, int);
__gcc_v4di __builtin_ia32_vpshrdq256(__gcc_v4di, __gcc_v4di, int);
__gcc_v4si __builtin_ia32_vpshrdvd128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpshrdvd256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v2di __builtin_ia32_vpshrdvq128(__gcc_v2di, __gcc_v2di, __gcc_v2di);
__gcc_v4di __builtin_ia32_vpshrdvq256(__gcc_v4di, __gcc_v4di, __gcc_v4di);
__gcc_v8hi __builtin_ia32_vpshrdvw128(__gcc_v8hi, __gcc_v8hi, __gcc_v8hi);
__gcc_v16hi __builtin_ia32_vpshrdvw256(__gcc_v16hi, __gcc_v16hi, __gcc_v16hi);
__gcc_v8hi __builtin_ia32_vpshrdw128(__gcc_v8hi, __gcc_v8hi, int);
__gcc_v16hi __builtin_ia32_vpshrdw256(__gcc_v16hi, __gcc_v16hi, int);
// avx512vl,avx512vnni|avxvnni
__gcc_v4si __builtin_ia32_vpdpbusd128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpbusd256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpbusds128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpbusds256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpwssd128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpwssd256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpwssds128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpwssds256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
// avx512vnni
__gcc_v16si __builtin_ia32_vpdpbusd512(__gcc_v16si, __gcc_v16si, __gcc_v16si);
__gcc_v16si __builtin_ia32_vpdpbusds512(__gcc_v16si, __gcc_v16si, __gcc_v16si);
__gcc_v16si __builtin_ia32_vpdpwssd512(__gcc_v16si, __gcc_v16si, __gcc_v16si);
__gcc_v16si __builtin_ia32_vpdpwssds512(__gcc_v16si, __gcc_v16si, __gcc_v16si);
// avx512vp2intersect
void __builtin_ia32_vp2intersect_q_512(__gcc_v8di, __gcc_v8di, unsigned char *, unsigned char *);
// avx512vp2intersect,avx512vl
void __builtin_ia32_vp2intersect_q_128(__gcc_v2di, __gcc_v2di, unsigned char *, unsigned char *);
void __builtin_ia32_vp2intersect_q_256(__gcc_v4di, __gcc_v4di, unsigned char *, unsigned char *);
// avx512vpopcntdq
__gcc_v16si __builtin_ia32_vpopcntd_512(__gcc_v16si);
__gcc_v8di __builtin_ia32_vpopcntq_512(__gcc_v8di);
// avx512vpopcntdq,avx512vl
__gcc_v4si __builtin_ia32_vpopcntd_128(__gcc_v4si);
__gcc_v8si __builtin_ia32_vpopcntd_256(__gcc_v8si);
__gcc_v2di __builtin_ia32_vpopcntq_128(__gcc_v2di);
__gcc_v4di __builtin_ia32_vpopcntq_256(__gcc_v4di);
// avxneconvert
__gcc_v4sf __builtin_ia32_vbcstnebf162ps128(const _Float16 *);
__gcc_v8sf __builtin_ia32_vbcstnebf162ps256(const _Float16 *);
__gcc_v4sf __builtin_ia32_vbcstnesh2ps128(const _Float16 *);
__gcc_v8sf __builtin_ia32_vbcstnesh2ps256(const _Float16 *);
__gcc_v4sf __builtin_ia32_vcvtneebf162ps128(const __gcc_v8hf *);
__gcc_v8sf __builtin_ia32_vcvtneebf162ps256(const __gcc_v16hf *);
__gcc_v4sf __builtin_ia32_vcvtneeph2ps128(const __gcc_v8hf *);
__gcc_v8sf __builtin_ia32_vcvtneeph2ps256(const __gcc_v16hf *);
__gcc_v4sf __builtin_ia32_vcvtneobf162ps128(const __gcc_v8hf *);
__gcc_v8sf __builtin_ia32_vcvtneobf162ps256(const __gcc_v16hf *);
__gcc_v4sf __builtin_ia32_vcvtneoph2ps128(const __gcc_v8hf *);
__gcc_v8sf __builtin_ia32_vcvtneoph2ps256(const __gcc_v16hf *);
// avxvnniint8
__gcc_v4si __builtin_ia32_vpdpbssd128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpbssd256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpbssds128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpbssds256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpbsud128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpbsud256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpbsuds128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpbsuds256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpbuud128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpbuud256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpbuuds128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpbuuds256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
// clang
unsigned int __builtin_ia32_readeflags_u32(void);
unsigned char __builtin_ia32_subborrow_u32(unsigned char, unsigned int, unsigned int, unsigned int *);
unsigned char __builtin_ia32_subborrow_u64(unsigned char, unsigned long long int, unsigned long long int, unsigned long long int *);
void __builtin_ia32_writeeflags_u32(unsigned int);
// cmpccxadd
signed int __builtin_ia32_cmpccxadd32(void *, signed int, signed int, int);
signed long long int __builtin_ia32_cmpccxadd64(void *, signed long long int, signed long long int, int);
// fma4
__gcc_v2df __builtin_ia32_vfmaddsd(__gcc_v2df, __gcc_v2df, __gcc_v2df);
__gcc_v4sf __builtin_ia32_vfmaddss(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf);
// invpcid
void __builtin_ia32_invpcid(unsigned int, void *);
void __builtin_ia32_directstore_u32(unsigned int *, unsigned int);
void __builtin_ia32_directstore_u64(unsigned long int *, unsigned long int);
void __builtin_ia32_prefetchi(const void *, unsigned int);
void __builtin_ia32_aadd32(void *, signed int);
void __builtin_ia32_aadd64(void *, signed long long int);
void __builtin_ia32_aand32(void *, signed int);
void __builtin_ia32_aand64(void *, signed long long int);
void __builtin_ia32_aor32(void *, signed int);
void __builtin_ia32_aor64(void *, signed long long int);
void __builtin_ia32_axor32(void *, signed int);
void __builtin_ia32_axor64(void *, signed long long int);
unsigned long long int __builtin_ia32_rdpru(int);
// rdseed
unsigned int __builtin_ia32_rdseed16_step(unsigned short *);
unsigned int __builtin_ia32_rdseed32_step(unsigned int *);
unsigned int __builtin_ia32_rdseed64_step(unsigned long long int *);
// sse2
__gcc_v2di __builtin_ia32_pslldqi128_byteshift(__gcc_v2di, int);
__gcc_v2di __builtin_ia32_psrldqi128_byteshift(__gcc_v2di, int);
__gcc_v4di __builtin_ia32_aesdec256(__gcc_v4di, __gcc_v4di);
__gcc_v4di __builtin_ia32_aesdeclast256(__gcc_v4di, __gcc_v4di);
__gcc_v4di __builtin_ia32_aesenc256(__gcc_v4di, __gcc_v4di);
__gcc_v4di __builtin_ia32_aesenclast256(__gcc_v4di, __gcc_v4di);
// vpclmulqdq
__gcc_v4di __builtin_ia32_pclmulqdq256(__gcc_v4di, __gcc_v4di, char);
// xop
__gcc_v16qi __builtin_ia32_vpcomb(__gcc_v16qi, __gcc_v16qi, char);
__gcc_v4si __builtin_ia32_vpcomd(__gcc_v4si, __gcc_v4si, char);
__gcc_v2di __builtin_ia32_vpcomq(__gcc_v2di, __gcc_v2di, char);
__gcc_v16qi __builtin_ia32_vpcomub(__gcc_v16qi, __gcc_v16qi, char);
__gcc_v4si __builtin_ia32_vpcomud(__gcc_v4si, __gcc_v4si, char);
__gcc_v2di __builtin_ia32_vpcomuq(__gcc_v2di, __gcc_v2di, char);
__gcc_v8hi __builtin_ia32_vpcomuw(__gcc_v8hi, __gcc_v8hi, char);
__gcc_v8hi __builtin_ia32_vpcomw(__gcc_v8hi, __gcc_v8hi, char);
__gcc_v16qi __builtin_ia32_vprotbi(__gcc_v16qi, char);
__gcc_v4si __builtin_ia32_vprotdi(__gcc_v4si, char);
__gcc_v2di __builtin_ia32_vprotqi(__gcc_v2di, char);
__gcc_v8hi __builtin_ia32_vprotwi(__gcc_v8hi, char);
// clang-format on