gcc_builtin_headers_ia32-8.h

Go to the documentation of this file.

// clang-format off
// avx512fp16
__gcc_v8hf __builtin_ia32_addsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
unsigned char __builtin_ia32_cmpsh_mask(__gcc_v8hf, __gcc_v8hf, int, unsigned char, int);
__gcc_v8hf __builtin_ia32_divsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
unsigned char __builtin_ia32_fpclasssh_mask(__gcc_v8hf, int, unsigned char);
__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_v8hf __builtin_ia32_maxsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
__gcc_v8hf __builtin_ia32_minsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
__gcc_v8hf __builtin_ia32_mulsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
__gcc_v8hf __builtin_ia32_rcpsh_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char);
__gcc_v8hf __builtin_ia32_reducesh_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int, int);
__gcc_v8hf __builtin_ia32_rndscalesh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int, int);
__gcc_v8hf __builtin_ia32_rsqrtsh_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char);
__gcc_v8hf __builtin_ia32_scalefsh_round_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
__gcc_v8hf __builtin_ia32_selectsh_128(unsigned char, __gcc_v8hf, __gcc_v8hf);
__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_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_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);
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_v8hf __builtin_ia32_vcvtusi2sh(__gcc_v8hf, unsigned int, int);
__gcc_v8hf __builtin_ia32_vcvtusi642sh(__gcc_v8hf, unsigned long long int, 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_v4sf __builtin_ia32_vfcmulcsh_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char, 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_v8hf __builtin_ia32_vfmaddsh3_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
__gcc_v8hf __builtin_ia32_vfmaddsh3_mask3(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
__gcc_v8hf __builtin_ia32_vfmaddsh3_maskz(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, int);
__gcc_v8hf __builtin_ia32_vfmsubsh3_mask3(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char, 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);
unsigned char __builtin_ia32_fpclassph128_mask(__gcc_v8hf, int, unsigned char);
unsigned short __builtin_ia32_fpclassph256_mask(__gcc_v16hf, int, unsigned short);
__gcc_v8hf __builtin_ia32_getexpph128_mask(__gcc_v8hf, __gcc_v8hf, unsigned char);
__gcc_v16hf __builtin_ia32_getexpph256_mask(__gcc_v16hf, __gcc_v16hf, unsigned short);
__gcc_v8hf __builtin_ia32_getmantph128_mask(__gcc_v8hf, int, __gcc_v8hf, unsigned char);
__gcc_v16hf __builtin_ia32_getmantph256_mask(__gcc_v16hf, int, __gcc_v16hf, 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);
__gcc_v8hf __builtin_ia32_rcpph128_mask(__gcc_v8hf, __gcc_v8hf, unsigned char);
__gcc_v16hf __builtin_ia32_rcpph256_mask(__gcc_v16hf, __gcc_v16hf, unsigned short);
_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_reduceph128_mask(__gcc_v8hf, int, __gcc_v8hf, unsigned char);
__gcc_v16hf __builtin_ia32_reduceph256_mask(__gcc_v16hf, int, __gcc_v16hf, unsigned short);
__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_rsqrtph128_mask(__gcc_v8hf, __gcc_v8hf, unsigned char);
__gcc_v16hf __builtin_ia32_rsqrtph256_mask(__gcc_v16hf, __gcc_v16hf, unsigned short);
__gcc_v8hf __builtin_ia32_scalefph128_mask(__gcc_v8hf, __gcc_v8hf, __gcc_v8hf, unsigned char);
__gcc_v16hf __builtin_ia32_scalefph256_mask(__gcc_v16hf, __gcc_v16hf, __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_vcvtdq2ph128_mask(__gcc_v4si, __gcc_v8hf, unsigned char);
__gcc_v8hf __builtin_ia32_vcvtdq2ph256_mask(__gcc_v8si, __gcc_v8hf, unsigned char);
__gcc_v8hf __builtin_ia32_vcvtpd2ph128_mask(__gcc_v2df, __gcc_v8hf, unsigned char);
__gcc_v8hf __builtin_ia32_vcvtpd2ph256_mask(__gcc_v4df, __gcc_v8hf, unsigned char);
__gcc_v4si __builtin_ia32_vcvtph2dq128_mask(__gcc_v8hf, __gcc_v4si, unsigned char);
__gcc_v8si __builtin_ia32_vcvtph2dq256_mask(__gcc_v8hf, __gcc_v8si, unsigned char);
__gcc_v2df __builtin_ia32_vcvtph2pd128_mask(__gcc_v8hf, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_vcvtph2pd256_mask(__gcc_v8hf, __gcc_v4df, unsigned char);
__gcc_v4sf __builtin_ia32_vcvtph2psx128_mask(__gcc_v8hf, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_vcvtph2psx256_mask(__gcc_v8hf, __gcc_v8sf, unsigned char);
__gcc_v2di __builtin_ia32_vcvtph2qq128_mask(__gcc_v8hf, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_vcvtph2qq256_mask(__gcc_v8hf, __gcc_v4di, unsigned char);
__gcc_v4usi __builtin_ia32_vcvtph2udq128_mask(__gcc_v8hf, __gcc_v4usi, unsigned char);
__gcc_v8usi __builtin_ia32_vcvtph2udq256_mask(__gcc_v8hf, __gcc_v8usi, unsigned char);
__gcc_v2udi __builtin_ia32_vcvtph2uqq128_mask(__gcc_v8hf, __gcc_v2udi, unsigned char);
__gcc_v4udi __builtin_ia32_vcvtph2uqq256_mask(__gcc_v8hf, __gcc_v4udi, unsigned char);
__gcc_v8uhi __builtin_ia32_vcvtph2uw128_mask(__gcc_v8hf, __gcc_v8uhi, unsigned char);
__gcc_v16uhi __builtin_ia32_vcvtph2uw256_mask(__gcc_v16hf, __gcc_v16uhi, unsigned short);
__gcc_v8hi __builtin_ia32_vcvtph2w128_mask(__gcc_v8hf, __gcc_v8hi, unsigned char);
__gcc_v16hi __builtin_ia32_vcvtph2w256_mask(__gcc_v16hf, __gcc_v16hi, unsigned short);
__gcc_v8hf __builtin_ia32_vcvtps2phx128_mask(__gcc_v4sf, __gcc_v8hf, unsigned char);
__gcc_v8hf __builtin_ia32_vcvtps2phx256_mask(__gcc_v8sf, __gcc_v8hf, unsigned char);
__gcc_v8hf __builtin_ia32_vcvtqq2ph128_mask(__gcc_v2di, __gcc_v8hf, unsigned char);
__gcc_v8hf __builtin_ia32_vcvtqq2ph256_mask(__gcc_v4di, __gcc_v8hf, unsigned char);
__gcc_v4si __builtin_ia32_vcvttph2dq128_mask(__gcc_v8hf, __gcc_v4si, unsigned char);
__gcc_v8si __builtin_ia32_vcvttph2dq256_mask(__gcc_v8hf, __gcc_v8si, unsigned char);
__gcc_v2di __builtin_ia32_vcvttph2qq128_mask(__gcc_v8hf, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_vcvttph2qq256_mask(__gcc_v8hf, __gcc_v4di, unsigned char);
__gcc_v4usi __builtin_ia32_vcvttph2udq128_mask(__gcc_v8hf, __gcc_v4usi, unsigned char);
__gcc_v8usi __builtin_ia32_vcvttph2udq256_mask(__gcc_v8hf, __gcc_v8usi, unsigned char);
__gcc_v2udi __builtin_ia32_vcvttph2uqq128_mask(__gcc_v8hf, __gcc_v2udi, unsigned char);
__gcc_v4udi __builtin_ia32_vcvttph2uqq256_mask(__gcc_v8hf, __gcc_v4udi, unsigned char);
__gcc_v8uhi __builtin_ia32_vcvttph2uw128_mask(__gcc_v8hf, __gcc_v8uhi, unsigned char);
__gcc_v16uhi __builtin_ia32_vcvttph2uw256_mask(__gcc_v16hf, __gcc_v16uhi, unsigned short);
__gcc_v8hi __builtin_ia32_vcvttph2w128_mask(__gcc_v8hf, __gcc_v8hi, unsigned char);
__gcc_v16hi __builtin_ia32_vcvttph2w256_mask(__gcc_v16hf, __gcc_v16hi, unsigned short);
__gcc_v8hf __builtin_ia32_vcvtudq2ph128_mask(__gcc_v4usi, __gcc_v8hf, unsigned char);
__gcc_v8hf __builtin_ia32_vcvtudq2ph256_mask(__gcc_v8usi, __gcc_v8hf, unsigned char);
__gcc_v8hf __builtin_ia32_vcvtuqq2ph128_mask(__gcc_v2udi, __gcc_v8hf, unsigned char);
__gcc_v8hf __builtin_ia32_vcvtuqq2ph256_mask(__gcc_v4udi, __gcc_v8hf, unsigned char);
__gcc_v8hf __builtin_ia32_vcvtuw2ph128_mask(__gcc_v8uhi, __gcc_v8hf, unsigned char);
__gcc_v16hf __builtin_ia32_vcvtuw2ph256_mask(__gcc_v16uhi, __gcc_v16hf, unsigned short);
__gcc_v8hf __builtin_ia32_vcvtw2ph128_mask(__gcc_v8hi, __gcc_v8hf, unsigned char);
__gcc_v16hf __builtin_ia32_vcvtw2ph256_mask(__gcc_v16hi, __gcc_v16hf, unsigned short);
__gcc_v4sf __builtin_ia32_vfcmaddcph128_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char);
__gcc_v4sf __builtin_ia32_vfcmaddcph128_maskz(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_vfcmaddcph256_mask(__gcc_v8sf, __gcc_v8sf, __gcc_v8sf, unsigned char);
__gcc_v8sf __builtin_ia32_vfcmaddcph256_maskz(__gcc_v8sf, __gcc_v8sf, __gcc_v8sf, unsigned char);
__gcc_v4sf __builtin_ia32_vfcmulcph128_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_vfcmulcph256_mask(__gcc_v8sf, __gcc_v8sf, __gcc_v8sf, unsigned char);
__gcc_v4sf __builtin_ia32_vfmaddcph128_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char);
__gcc_v4sf __builtin_ia32_vfmaddcph128_maskz(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_vfmaddcph256_mask(__gcc_v8sf, __gcc_v8sf, __gcc_v8sf, unsigned char);
__gcc_v8sf __builtin_ia32_vfmaddcph256_maskz(__gcc_v8sf, __gcc_v8sf, __gcc_v8sf, unsigned char);
__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);
__gcc_v4sf __builtin_ia32_vfmulcph128_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_vfmulcph256_mask(__gcc_v8sf, __gcc_v8sf, __gcc_v8sf, unsigned char);
// avx512fp16,evex512
__gcc_v32hf __builtin_ia32_addph512(__gcc_v32hf, __gcc_v32hf, int);
unsigned int __builtin_ia32_cmpph512_mask(__gcc_v32hf, __gcc_v32hf, int, unsigned int, int);
__gcc_v32hf __builtin_ia32_divph512(__gcc_v32hf, __gcc_v32hf, int);
unsigned int __builtin_ia32_fpclassph512_mask(__gcc_v32hf, int, unsigned int);
__gcc_v32hf __builtin_ia32_getexpph512_mask(__gcc_v32hf, __gcc_v32hf, unsigned int, int);
__gcc_v32hf __builtin_ia32_getmantph512_mask(__gcc_v32hf, int, __gcc_v32hf, unsigned int, int);
__gcc_v32hf __builtin_ia32_maxph512(__gcc_v32hf, __gcc_v32hf, int);
__gcc_v32hf __builtin_ia32_minph512(__gcc_v32hf, __gcc_v32hf, int);
__gcc_v32hf __builtin_ia32_mulph512(__gcc_v32hf, __gcc_v32hf, int);
__gcc_v32hf __builtin_ia32_rcpph512_mask(__gcc_v32hf, __gcc_v32hf, unsigned 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_v32hf __builtin_ia32_rndscaleph_mask(__gcc_v32hf, int, __gcc_v32hf, unsigned int, int);
__gcc_v32hf __builtin_ia32_rsqrtph512_mask(__gcc_v32hf, __gcc_v32hf, unsigned int);
__gcc_v32hf __builtin_ia32_scalefph512_mask(__gcc_v32hf, __gcc_v32hf, __gcc_v32hf, unsigned int, int);
__gcc_v32hf __builtin_ia32_selectph_512(unsigned int, __gcc_v32hf, __gcc_v32hf);
__gcc_v32hf __builtin_ia32_sqrtph512(__gcc_v32hf, int);
__gcc_v32hf __builtin_ia32_subph512(__gcc_v32hf, __gcc_v32hf, 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_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);
__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_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_v16sf __builtin_ia32_vfcmulcph512_mask(__gcc_v16sf, __gcc_v16sf, __gcc_v16sf, unsigned short, 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_v32hf __builtin_ia32_vfmaddph512_mask(__gcc_v32hf, __gcc_v32hf, __gcc_v32hf, unsigned int, int);
__gcc_v32hf __builtin_ia32_vfmaddph512_mask3(__gcc_v32hf, __gcc_v32hf, __gcc_v32hf, unsigned int, int);
__gcc_v32hf __builtin_ia32_vfmaddph512_maskz(__gcc_v32hf, __gcc_v32hf, __gcc_v32hf, unsigned int, int);
__gcc_v32hf __builtin_ia32_vfmaddsubph512_mask(__gcc_v32hf, __gcc_v32hf, __gcc_v32hf, unsigned int, int);
__gcc_v32hf __builtin_ia32_vfmaddsubph512_mask3(__gcc_v32hf, __gcc_v32hf, __gcc_v32hf, unsigned int, int);
__gcc_v32hf __builtin_ia32_vfmaddsubph512_maskz(__gcc_v32hf, __gcc_v32hf, __gcc_v32hf, unsigned int, int);
__gcc_v32hf __builtin_ia32_vfmsubaddph512_mask3(__gcc_v32hf, __gcc_v32hf, __gcc_v32hf, unsigned int, int);
__gcc_v32hf __builtin_ia32_vfmsubph512_mask3(__gcc_v32hf, __gcc_v32hf, __gcc_v32hf, unsigned int, int);
__gcc_v16sf __builtin_ia32_vfmulcph512_mask(__gcc_v16sf, __gcc_v16sf, __gcc_v16sf, unsigned short, int);
// 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);
// avx512ifma,evex512
__gcc_v8di __builtin_ia32_vpmadd52huq512(__gcc_v8di, __gcc_v8di, __gcc_v8di);
__gcc_v8di __builtin_ia32_vpmadd52luq512(__gcc_v8di, __gcc_v8di, __gcc_v8di);
// 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);
// avx512vbmi,evex512
__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);
// avx512vbmi2,evex512
__gcc_v32hi __builtin_ia32_compresshi512_mask(__gcc_v32hi, __gcc_v32hi, unsigned int);
__gcc_v64qi __builtin_ia32_compressqi512_mask(__gcc_v64qi, __gcc_v64qi, unsigned long long int);
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_v32hi __builtin_ia32_expandhi512_mask(__gcc_v32hi, __gcc_v32hi, unsigned int);
__gcc_v32hi __builtin_ia32_expandloadhi512_mask(const __gcc_v32hi *, __gcc_v32hi, unsigned int);
__gcc_v64qi __builtin_ia32_expandloadqi512_mask(const __gcc_v64qi *, __gcc_v64qi, unsigned long long int);
__gcc_v64qi __builtin_ia32_expandqi512_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);
unsigned char __builtin_ia32_cmpd128_mask(__gcc_v4si, __gcc_v4si, int, unsigned char);
unsigned char __builtin_ia32_cmpd256_mask(__gcc_v8si, __gcc_v8si, int, unsigned char);
unsigned char __builtin_ia32_cmppd128_mask(__gcc_v2df, __gcc_v2df, int, unsigned char);
unsigned char __builtin_ia32_cmppd256_mask(__gcc_v4df, __gcc_v4df, int, unsigned char);
unsigned char __builtin_ia32_cmpps128_mask(__gcc_v4sf, __gcc_v4sf, int, unsigned char);
unsigned char __builtin_ia32_cmpps256_mask(__gcc_v8sf, __gcc_v8sf, int, unsigned char);
unsigned char __builtin_ia32_cmpq128_mask(__gcc_v2di, __gcc_v2di, int, unsigned char);
unsigned char __builtin_ia32_cmpq256_mask(__gcc_v4di, __gcc_v4di, int, unsigned char);
__gcc_v2df __builtin_ia32_compressdf128_mask(__gcc_v2df, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_compressdf256_mask(__gcc_v4df, __gcc_v4df, unsigned char);
__gcc_v2di __builtin_ia32_compressdi128_mask(__gcc_v2di, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_compressdi256_mask(__gcc_v4di, __gcc_v4di, unsigned char);
__gcc_v4sf __builtin_ia32_compresssf128_mask(__gcc_v4sf, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_compresssf256_mask(__gcc_v8sf, __gcc_v8sf, unsigned char);
__gcc_v4si __builtin_ia32_compresssi128_mask(__gcc_v4si, __gcc_v4si, unsigned char);
__gcc_v8si __builtin_ia32_compresssi256_mask(__gcc_v8si, __gcc_v8si, unsigned char);
void __builtin_ia32_compressstoredf128_mask(__gcc_v2df *, __gcc_v2df, unsigned char);
void __builtin_ia32_compressstoredf256_mask(__gcc_v4df *, __gcc_v4df, unsigned char);
void __builtin_ia32_compressstoredi128_mask(__gcc_v2di *, __gcc_v2di, unsigned char);
void __builtin_ia32_compressstoredi256_mask(__gcc_v4di *, __gcc_v4di, unsigned char);
void __builtin_ia32_compressstoresf128_mask(__gcc_v4sf *, __gcc_v4sf, unsigned char);
void __builtin_ia32_compressstoresf256_mask(__gcc_v8sf *, __gcc_v8sf, unsigned char);
void __builtin_ia32_compressstoresi128_mask(__gcc_v4si *, __gcc_v4si, unsigned char);
void __builtin_ia32_compressstoresi256_mask(__gcc_v8si *, __gcc_v8si, unsigned char);
__gcc_v4si __builtin_ia32_cvtpd2dq128_mask(__gcc_v2df, __gcc_v4si, unsigned char);
__gcc_v4sf __builtin_ia32_cvtpd2ps_mask(__gcc_v2df, __gcc_v4sf, unsigned char);
__gcc_v4si __builtin_ia32_cvtpd2udq128_mask(__gcc_v2df, __gcc_v4si, unsigned char);
__gcc_v4si __builtin_ia32_cvtpd2udq256_mask(__gcc_v4df, __gcc_v4si, unsigned char);
__gcc_v4si __builtin_ia32_cvtps2udq128_mask(__gcc_v4sf, __gcc_v4si, unsigned char);
__gcc_v8si __builtin_ia32_cvtps2udq256_mask(__gcc_v8sf, __gcc_v8si, unsigned char);
__gcc_v4si __builtin_ia32_cvttpd2dq128_mask(__gcc_v2df, __gcc_v4si, unsigned char);
__gcc_v4si __builtin_ia32_cvttpd2udq128_mask(__gcc_v2df, __gcc_v4si, unsigned char);
__gcc_v4si __builtin_ia32_cvttpd2udq256_mask(__gcc_v4df, __gcc_v4si, unsigned char);
__gcc_v4si __builtin_ia32_cvttps2udq128_mask(__gcc_v4sf, __gcc_v4si, unsigned char);
__gcc_v8si __builtin_ia32_cvttps2udq256_mask(__gcc_v8sf, __gcc_v8si, unsigned char);
__gcc_v2df __builtin_ia32_expanddf128_mask(__gcc_v2df, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_expanddf256_mask(__gcc_v4df, __gcc_v4df, unsigned char);
__gcc_v2di __builtin_ia32_expanddi128_mask(__gcc_v2di, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_expanddi256_mask(__gcc_v4di, __gcc_v4di, unsigned char);
__gcc_v2df __builtin_ia32_expandloaddf128_mask(const __gcc_v2df *, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_expandloaddf256_mask(const __gcc_v4df *, __gcc_v4df, unsigned char);
__gcc_v4si __builtin_ia32_expandloaddi128_mask(const __gcc_v2di *, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_expandloaddi256_mask(const __gcc_v4di *, __gcc_v4di, unsigned char);
__gcc_v4sf __builtin_ia32_expandloadsf128_mask(const __gcc_v4sf *, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_expandloadsf256_mask(const __gcc_v8sf *, __gcc_v8sf, unsigned char);
__gcc_v4si __builtin_ia32_expandloadsi128_mask(const __gcc_v4si *, __gcc_v4si, unsigned char);
__gcc_v8si __builtin_ia32_expandloadsi256_mask(const __gcc_v8si *, __gcc_v8si, unsigned char);
__gcc_v4sf __builtin_ia32_expandsf128_mask(__gcc_v4sf, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_expandsf256_mask(__gcc_v8sf, __gcc_v8sf, unsigned char);
__gcc_v4si __builtin_ia32_expandsi128_mask(__gcc_v4si, __gcc_v4si, unsigned char);
__gcc_v8si __builtin_ia32_expandsi256_mask(__gcc_v8si, __gcc_v8si, unsigned char);
__gcc_v4sf __builtin_ia32_extractf32x4_256_mask(__gcc_v8sf, int, __gcc_v4sf, unsigned char);
__gcc_v4si __builtin_ia32_extracti32x4_256_mask(__gcc_v8si, int, __gcc_v4si, unsigned char);
__gcc_v2df __builtin_ia32_fixupimmpd128_mask(__gcc_v2df, __gcc_v2df, __gcc_v2di, int, unsigned char);
__gcc_v2df __builtin_ia32_fixupimmpd128_maskz(__gcc_v2df, __gcc_v2df, __gcc_v2di, int, unsigned char);
__gcc_v4df __builtin_ia32_fixupimmpd256_mask(__gcc_v4df, __gcc_v4df, __gcc_v4di, int, unsigned char);
__gcc_v4df __builtin_ia32_fixupimmpd256_maskz(__gcc_v4df, __gcc_v4df, __gcc_v4di, int, unsigned char);
__gcc_v4sf __builtin_ia32_fixupimmps128_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4si, int, unsigned char);
__gcc_v4sf __builtin_ia32_fixupimmps128_maskz(__gcc_v4sf, __gcc_v4sf, __gcc_v4si, int, unsigned char);
__gcc_v8sf __builtin_ia32_fixupimmps256_mask(__gcc_v8sf, __gcc_v8sf, __gcc_v8si, int, unsigned char);
__gcc_v8sf __builtin_ia32_fixupimmps256_maskz(__gcc_v8sf, __gcc_v8sf, __gcc_v8si, int, unsigned char);
__gcc_v2df __builtin_ia32_gather3div2df(__gcc_v2df, const void *, __gcc_v2di, unsigned char, int);
__gcc_v2di __builtin_ia32_gather3div2di(__gcc_v2di, const void *, __gcc_v2di, unsigned char, int);
__gcc_v4df __builtin_ia32_gather3div4df(__gcc_v4df, const void *, __gcc_v4di, unsigned char, int);
__gcc_v4di __builtin_ia32_gather3div4di(__gcc_v4di, const void *, __gcc_v4di, unsigned char, int);
__gcc_v4sf __builtin_ia32_gather3div4sf(__gcc_v4sf, const void *, __gcc_v2di, unsigned char, int);
__gcc_v4si __builtin_ia32_gather3div4si(__gcc_v4si, const void *, __gcc_v2di, unsigned char, int);
__gcc_v4sf __builtin_ia32_gather3div8sf(__gcc_v4sf, const void *, __gcc_v4di, unsigned char, int);
__gcc_v4si __builtin_ia32_gather3div8si(__gcc_v4si, const void *, __gcc_v4di, unsigned char, int);
__gcc_v2df __builtin_ia32_gather3siv2df(__gcc_v2df, const void *, __gcc_v4si, unsigned char, int);
__gcc_v2di __builtin_ia32_gather3siv2di(__gcc_v2di, const void *, __gcc_v4si, unsigned char, int);
__gcc_v4df __builtin_ia32_gather3siv4df(__gcc_v4df, const void *, __gcc_v4si, unsigned char, int);
__gcc_v4di __builtin_ia32_gather3siv4di(__gcc_v4di, const void *, __gcc_v4si, unsigned char, int);
__gcc_v4sf __builtin_ia32_gather3siv4sf(__gcc_v4sf, const void *, __gcc_v4si, unsigned char, int);
__gcc_v4si __builtin_ia32_gather3siv4si(__gcc_v4si, const void *, __gcc_v4si, unsigned char, int);
__gcc_v8sf __builtin_ia32_gather3siv8sf(__gcc_v8sf, const void *, __gcc_v8si, unsigned char, int);
__gcc_v8si __builtin_ia32_gather3siv8si(__gcc_v8si, const void *, __gcc_v8si, unsigned char, int);
__gcc_v2df __builtin_ia32_getexppd128_mask(__gcc_v2df, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_getexppd256_mask(__gcc_v4df, __gcc_v4df, unsigned char);
__gcc_v4sf __builtin_ia32_getexpps128_mask(__gcc_v4sf, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_getexpps256_mask(__gcc_v8sf, __gcc_v8sf, unsigned char);
__gcc_v2df __builtin_ia32_getmantpd128_mask(__gcc_v2df, int, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_getmantpd256_mask(__gcc_v4df, int, __gcc_v4df, unsigned char);
__gcc_v4sf __builtin_ia32_getmantps128_mask(__gcc_v4sf, int, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_getmantps256_mask(__gcc_v8sf, int, __gcc_v8sf, unsigned char);
__gcc_v8sf __builtin_ia32_insertf32x4_256(__gcc_v8sf, __gcc_v4sf, int);
__gcc_v8si __builtin_ia32_inserti32x4_256(__gcc_v8si, __gcc_v4si, int);
__gcc_v2df __builtin_ia32_loadapd128_mask(const __gcc_v2df *, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_loadapd256_mask(const __gcc_v4df *, __gcc_v4df, unsigned char);
__gcc_v4sf __builtin_ia32_loadaps128_mask(const __gcc_v4sf *, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_loadaps256_mask(const __gcc_v8sf *, __gcc_v8sf, unsigned char);
__gcc_v2di __builtin_ia32_loaddqudi128_mask(const __gcc_v2di *, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_loaddqudi256_mask(const __gcc_v4di *, __gcc_v4di, unsigned char);
__gcc_v4si __builtin_ia32_loaddqusi128_mask(const __gcc_v4si *, __gcc_v4si, unsigned char);
__gcc_v8si __builtin_ia32_loaddqusi256_mask(const __gcc_v8si *, __gcc_v8si, unsigned char);
__gcc_v2df __builtin_ia32_loadupd128_mask(const __gcc_v2df *, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_loadupd256_mask(const __gcc_v4df *, __gcc_v4df, unsigned char);
__gcc_v4sf __builtin_ia32_loadups128_mask(const __gcc_v4sf *, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_loadups256_mask(const __gcc_v8sf *, __gcc_v8sf, unsigned char);
__gcc_v4si __builtin_ia32_movdqa32load128_mask(const __gcc_v4si *, __gcc_v4si, unsigned char);
__gcc_v8si __builtin_ia32_movdqa32load256_mask(const __gcc_v8si *, __gcc_v8si, unsigned char);
void __builtin_ia32_movdqa32store128_mask(__gcc_v4si *, __gcc_v4si, unsigned char);
void __builtin_ia32_movdqa32store256_mask(__gcc_v8si *, __gcc_v8si, unsigned char);
__gcc_v2di __builtin_ia32_movdqa64load128_mask(const __gcc_v2di *, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_movdqa64load256_mask(const __gcc_v4di *, __gcc_v4di, unsigned char);
void __builtin_ia32_movdqa64store128_mask(__gcc_v2di *, __gcc_v2di, unsigned char);
void __builtin_ia32_movdqa64store256_mask(__gcc_v4di *, __gcc_v4di, unsigned char);
__gcc_v4df __builtin_ia32_permvardf256(__gcc_v4df, __gcc_v4di);
__gcc_v4di __builtin_ia32_permvardi256(__gcc_v4di, __gcc_v4di);
__gcc_v16qi __builtin_ia32_pmovdb128_mask(__gcc_v4si, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovdb128mem_mask(__gcc_v16qi *, __gcc_v4si, unsigned char);
__gcc_v16qi __builtin_ia32_pmovdb256_mask(__gcc_v8si, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovdb256mem_mask(__gcc_v16qi *, __gcc_v8si, unsigned char);
__gcc_v8hi __builtin_ia32_pmovdw128_mask(__gcc_v4si, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovdw128mem_mask(__gcc_v8hi *, __gcc_v4si, unsigned char);
__gcc_v8hi __builtin_ia32_pmovdw256_mask(__gcc_v8si, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovdw256mem_mask(__gcc_v8hi *, __gcc_v8si, unsigned char);
__gcc_v16qi __builtin_ia32_pmovqb128_mask(__gcc_v2di, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovqb128mem_mask(__gcc_v16qi *, __gcc_v2di, unsigned char);
__gcc_v16qi __builtin_ia32_pmovqb256_mask(__gcc_v4di, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovqb256mem_mask(__gcc_v16qi *, __gcc_v4di, unsigned char);
__gcc_v4si __builtin_ia32_pmovqd128_mask(__gcc_v2di, __gcc_v4si, unsigned char);
void __builtin_ia32_pmovqd128mem_mask(__gcc_v4si *, __gcc_v2di, unsigned char);
void __builtin_ia32_pmovqd256mem_mask(__gcc_v4si *, __gcc_v4di, unsigned char);
__gcc_v8hi __builtin_ia32_pmovqw128_mask(__gcc_v2di, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovqw128mem_mask(__gcc_v8hi *, __gcc_v2di, unsigned char);
__gcc_v8hi __builtin_ia32_pmovqw256_mask(__gcc_v4di, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovqw256mem_mask(__gcc_v8hi *, __gcc_v4di, unsigned char);
__gcc_v16qi __builtin_ia32_pmovsdb128_mask(__gcc_v4si, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovsdb128mem_mask(__gcc_v16qi *, __gcc_v4si, unsigned char);
__gcc_v16qi __builtin_ia32_pmovsdb256_mask(__gcc_v8si, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovsdb256mem_mask(__gcc_v16qi *, __gcc_v8si, unsigned char);
__gcc_v8hi __builtin_ia32_pmovsdw128_mask(__gcc_v4si, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovsdw128mem_mask(__gcc_v8hi *, __gcc_v4si, unsigned char);
__gcc_v8hi __builtin_ia32_pmovsdw256_mask(__gcc_v8si, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovsdw256mem_mask(__gcc_v8hi *, __gcc_v8si, unsigned char);
__gcc_v16qi __builtin_ia32_pmovsqb128_mask(__gcc_v2di, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovsqb128mem_mask(__gcc_v16qi *, __gcc_v2di, unsigned char);
__gcc_v16qi __builtin_ia32_pmovsqb256_mask(__gcc_v4di, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovsqb256mem_mask(__gcc_v16qi *, __gcc_v4di, unsigned char);
__gcc_v4si __builtin_ia32_pmovsqd128_mask(__gcc_v2di, __gcc_v4si, unsigned char);
void __builtin_ia32_pmovsqd128mem_mask(__gcc_v4si *, __gcc_v2di, unsigned char);
__gcc_v4si __builtin_ia32_pmovsqd256_mask(__gcc_v4di, __gcc_v4si, unsigned char);
void __builtin_ia32_pmovsqd256mem_mask(__gcc_v4si *, __gcc_v4di, unsigned char);
__gcc_v8hi __builtin_ia32_pmovsqw128_mask(__gcc_v2di, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovsqw128mem_mask(__gcc_v8hi *, __gcc_v2di, unsigned char);
__gcc_v8hi __builtin_ia32_pmovsqw256_mask(__gcc_v4di, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovsqw256mem_mask(__gcc_v8hi *, __gcc_v4di, unsigned char);
__gcc_v16qi __builtin_ia32_pmovusdb128_mask(__gcc_v4si, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovusdb128mem_mask(__gcc_v16qi *, __gcc_v4si, unsigned char);
__gcc_v16qi __builtin_ia32_pmovusdb256_mask(__gcc_v8si, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovusdb256mem_mask(__gcc_v16qi *, __gcc_v8si, unsigned char);
__gcc_v8hi __builtin_ia32_pmovusdw128_mask(__gcc_v4si, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovusdw128mem_mask(__gcc_v8hi *, __gcc_v4si, unsigned char);
__gcc_v8hi __builtin_ia32_pmovusdw256_mask(__gcc_v8si, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovusdw256mem_mask(__gcc_v8hi *, __gcc_v8si, unsigned char);
__gcc_v16qi __builtin_ia32_pmovusqb128_mask(__gcc_v2di, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovusqb128mem_mask(__gcc_v16qi *, __gcc_v2di, unsigned char);
__gcc_v16qi __builtin_ia32_pmovusqb256_mask(__gcc_v4di, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovusqb256mem_mask(__gcc_v16qi *, __gcc_v4di, unsigned char);
__gcc_v4si __builtin_ia32_pmovusqd128_mask(__gcc_v2di, __gcc_v4si, unsigned char);
void __builtin_ia32_pmovusqd128mem_mask(__gcc_v4si *, __gcc_v2di, unsigned char);
__gcc_v4si __builtin_ia32_pmovusqd256_mask(__gcc_v4di, __gcc_v4si, unsigned char);
void __builtin_ia32_pmovusqd256mem_mask(__gcc_v4si *, __gcc_v4di, unsigned char);
__gcc_v8hi __builtin_ia32_pmovusqw128_mask(__gcc_v2di, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovusqw128mem_mask(__gcc_v8hi *, __gcc_v2di, unsigned char);
__gcc_v8hi __builtin_ia32_pmovusqw256_mask(__gcc_v4di, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovusqw256mem_mask(__gcc_v8hi *, __gcc_v4di, unsigned char);
__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_v4si __builtin_ia32_pternlogd128_mask(__gcc_v4si, __gcc_v4si, __gcc_v4si, int, unsigned char);
__gcc_v4si __builtin_ia32_pternlogd128_maskz(__gcc_v4si, __gcc_v4si, __gcc_v4si, int, unsigned char);
__gcc_v8si __builtin_ia32_pternlogd256_mask(__gcc_v8si, __gcc_v8si, __gcc_v8si, int, unsigned char);
__gcc_v8si __builtin_ia32_pternlogd256_maskz(__gcc_v8si, __gcc_v8si, __gcc_v8si, int, unsigned char);
__gcc_v2di __builtin_ia32_pternlogq128_mask(__gcc_v2di, __gcc_v2di, __gcc_v2di, int, unsigned char);
__gcc_v2di __builtin_ia32_pternlogq128_maskz(__gcc_v2di, __gcc_v2di, __gcc_v2di, int, unsigned char);
__gcc_v4di __builtin_ia32_pternlogq256_mask(__gcc_v4di, __gcc_v4di, __gcc_v4di, int, unsigned char);
__gcc_v4di __builtin_ia32_pternlogq256_maskz(__gcc_v4di, __gcc_v4di, __gcc_v4di, int, unsigned char);
__gcc_v2df __builtin_ia32_rcp14pd128_mask(__gcc_v2df, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_rcp14pd256_mask(__gcc_v4df, __gcc_v4df, unsigned char);
__gcc_v4sf __builtin_ia32_rcp14ps128_mask(__gcc_v4sf, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_rcp14ps256_mask(__gcc_v8sf, __gcc_v8sf, unsigned char);
__gcc_v2df __builtin_ia32_rndscalepd_128_mask(__gcc_v2df, int, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_rndscalepd_256_mask(__gcc_v4df, int, __gcc_v4df, unsigned char);
__gcc_v4sf __builtin_ia32_rndscaleps_128_mask(__gcc_v4sf, int, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_rndscaleps_256_mask(__gcc_v8sf, int, __gcc_v8sf, unsigned char);
__gcc_v2df __builtin_ia32_rsqrt14pd128_mask(__gcc_v2df, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_rsqrt14pd256_mask(__gcc_v4df, __gcc_v4df, unsigned char);
__gcc_v4sf __builtin_ia32_rsqrt14ps128_mask(__gcc_v4sf, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_rsqrt14ps256_mask(__gcc_v8sf, __gcc_v8sf, unsigned char);
__gcc_v2df __builtin_ia32_scalefpd128_mask(__gcc_v2df, __gcc_v2df, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_scalefpd256_mask(__gcc_v4df, __gcc_v4df, __gcc_v4df, unsigned char);
__gcc_v4sf __builtin_ia32_scalefps128_mask(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_scalefps256_mask(__gcc_v8sf, __gcc_v8sf, __gcc_v8sf, unsigned char);
void __builtin_ia32_scatterdiv2df(void *, unsigned char, __gcc_v2di, __gcc_v2df, int);
void __builtin_ia32_scatterdiv2di(void *, unsigned char, __gcc_v2di, __gcc_v2di, int);
void __builtin_ia32_scatterdiv4df(void *, unsigned char, __gcc_v4di, __gcc_v4df, int);
void __builtin_ia32_scatterdiv4di(void *, unsigned char, __gcc_v4di, __gcc_v4di, int);
void __builtin_ia32_scatterdiv4sf(void *, unsigned char, __gcc_v2di, __gcc_v4sf, int);
void __builtin_ia32_scatterdiv4si(void *, unsigned char, __gcc_v2di, __gcc_v4si, int);
void __builtin_ia32_scatterdiv8sf(void *, unsigned char, __gcc_v4di, __gcc_v4sf, int);
void __builtin_ia32_scatterdiv8si(void *, unsigned char, __gcc_v4di, __gcc_v4si, int);
void __builtin_ia32_scattersiv2df(void *, unsigned char, __gcc_v4si, __gcc_v2df, int);
void __builtin_ia32_scattersiv2di(void *, unsigned char, __gcc_v4si, __gcc_v2di, int);
void __builtin_ia32_scattersiv4df(void *, unsigned char, __gcc_v4si, __gcc_v4df, int);
void __builtin_ia32_scattersiv4di(void *, unsigned char, __gcc_v4si, __gcc_v4di, int);
void __builtin_ia32_scattersiv4sf(void *, unsigned char, __gcc_v4si, __gcc_v4sf, int);
void __builtin_ia32_scattersiv4si(void *, unsigned char, __gcc_v4si, __gcc_v4si, int);
void __builtin_ia32_scattersiv8sf(void *, unsigned char, __gcc_v8si, __gcc_v8sf, int);
void __builtin_ia32_scattersiv8si(void *, unsigned char, __gcc_v8si, __gcc_v8si, int);
__gcc_v4si __builtin_ia32_selectd_128(unsigned char, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_selectd_256(unsigned char, __gcc_v8si, __gcc_v8si);
__gcc_v2df __builtin_ia32_selectpd_128(unsigned char, __gcc_v2df, __gcc_v2df);
__gcc_v4df __builtin_ia32_selectpd_256(unsigned char, __gcc_v4df, __gcc_v4df);
__gcc_v4sf __builtin_ia32_selectps_128(unsigned char, __gcc_v4sf, __gcc_v4sf);
__gcc_v8sf __builtin_ia32_selectps_256(unsigned char, __gcc_v8sf, __gcc_v8sf);
__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);
void __builtin_ia32_storeapd128_mask(__gcc_v2df *, __gcc_v2df, unsigned char);
void __builtin_ia32_storeapd256_mask(__gcc_v4df *, __gcc_v4df, unsigned char);
void __builtin_ia32_storeaps128_mask(__gcc_v4sf *, __gcc_v4sf, unsigned char);
void __builtin_ia32_storeaps256_mask(__gcc_v8sf *, __gcc_v8sf, unsigned char);
void __builtin_ia32_storedqudi128_mask(__gcc_v2di *, __gcc_v2di, unsigned char);
void __builtin_ia32_storedqudi256_mask(__gcc_v4di *, __gcc_v4di, unsigned char);
void __builtin_ia32_storedqusi128_mask(__gcc_v4si *, __gcc_v4si, unsigned char);
void __builtin_ia32_storedqusi256_mask(__gcc_v8si *, __gcc_v8si, unsigned char);
void __builtin_ia32_storeupd128_mask(__gcc_v2df *, __gcc_v2df, unsigned char);
void __builtin_ia32_storeupd256_mask(__gcc_v4df *, __gcc_v4df, unsigned char);
void __builtin_ia32_storeups128_mask(__gcc_v4sf *, __gcc_v4sf, unsigned char);
void __builtin_ia32_storeups256_mask(__gcc_v8sf *, __gcc_v8sf, unsigned char);
unsigned char __builtin_ia32_ucmpd128_mask(__gcc_v4si, __gcc_v4si, int, unsigned char);
unsigned char __builtin_ia32_ucmpd256_mask(__gcc_v8si, __gcc_v8si, int, unsigned char);
unsigned char __builtin_ia32_ucmpq128_mask(__gcc_v2di, __gcc_v2di, int, unsigned char);
unsigned char __builtin_ia32_ucmpq256_mask(__gcc_v4di, __gcc_v4di, int, unsigned char);
__gcc_v8sf __builtin_ia32_vcvtph2ps256_mask(__gcc_v8hi, __gcc_v8sf, unsigned char);
__gcc_v4sf __builtin_ia32_vcvtph2ps_mask(__gcc_v8hi, __gcc_v4sf, unsigned char);
__gcc_v8hi __builtin_ia32_vcvtps2ph256_mask(__gcc_v8sf, int, __gcc_v8hi, unsigned char);
__gcc_v8hi __builtin_ia32_vcvtps2ph_mask(__gcc_v4sf, int, __gcc_v8hi, unsigned char);
__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);
unsigned short __builtin_ia32_vpshufbitqmb128_mask(__gcc_v16qi, __gcc_v16qi, unsigned short);
unsigned int __builtin_ia32_vpshufbitqmb256_mask(__gcc_v32qi, __gcc_v32qi, unsigned int);
// avx512vl,avx512bw
unsigned short __builtin_ia32_cmpb128_mask(__gcc_v16qi, __gcc_v16qi, int, unsigned short);
unsigned int __builtin_ia32_cmpb256_mask(__gcc_v32qi, __gcc_v32qi, int, unsigned int);
unsigned char __builtin_ia32_cmpw128_mask(__gcc_v8hi, __gcc_v8hi, int, unsigned char);
unsigned short __builtin_ia32_cmpw256_mask(__gcc_v16hi, __gcc_v16hi, int, unsigned short);
__gcc_v16qi __builtin_ia32_pmovswb128_mask(__gcc_v8hi, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovswb128mem_mask(__gcc_v16qi *, __gcc_v8hi, unsigned char);
__gcc_v16qi __builtin_ia32_pmovswb256_mask(__gcc_v16hi, __gcc_v16qi, unsigned short);
void __builtin_ia32_pmovswb256mem_mask(__gcc_v16qi *, __gcc_v16hi, unsigned short);
__gcc_v16qi __builtin_ia32_pmovuswb128_mask(__gcc_v8hi, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovuswb128mem_mask(__gcc_v16qi *, __gcc_v8hi, unsigned char);
__gcc_v16qi __builtin_ia32_pmovuswb256_mask(__gcc_v16hi, __gcc_v16qi, unsigned short);
void __builtin_ia32_pmovuswb256mem_mask(__gcc_v16qi *, __gcc_v16hi, unsigned short);
__gcc_v16qi __builtin_ia32_pmovwb128_mask(__gcc_v8hi, __gcc_v16qi, unsigned char);
void __builtin_ia32_pmovwb128mem_mask(__gcc_v16qi *, __gcc_v8hi, unsigned char);
void __builtin_ia32_pmovwb256mem_mask(__gcc_v16qi *, __gcc_v16hi, unsigned short);
void __builtin_ia32_storedquhi128_mask(__gcc_v8hi *, __gcc_v8hi, unsigned char);
void __builtin_ia32_storedquhi256_mask(__gcc_v16hi *, __gcc_v16hi, unsigned short);
void __builtin_ia32_storedquqi128_mask(__gcc_v16qi *, __gcc_v16qi, unsigned short);
void __builtin_ia32_storedquqi256_mask(__gcc_v32qi *, __gcc_v32qi, unsigned int);
unsigned short __builtin_ia32_ucmpb128_mask(__gcc_v16qi, __gcc_v16qi, int, unsigned short);
unsigned int __builtin_ia32_ucmpb256_mask(__gcc_v32qi, __gcc_v32qi, int, unsigned int);
unsigned char __builtin_ia32_ucmpw128_mask(__gcc_v8hi, __gcc_v8hi, int, unsigned char);
unsigned short __builtin_ia32_ucmpw256_mask(__gcc_v16hi, __gcc_v16hi, int, unsigned short);
__gcc_v8hi __builtin_ia32_vpermi2varhi128(__gcc_v8hi, __gcc_v8hi, __gcc_v8hi);
__gcc_v16hi __builtin_ia32_vpermi2varhi256(__gcc_v16hi, __gcc_v16hi, __gcc_v16hi);
// avx512vl,avx512dq
__gcc_v2di __builtin_ia32_cvtpd2qq128_mask(__gcc_v2df, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_cvtpd2qq256_mask(__gcc_v4df, __gcc_v4di, unsigned char);
__gcc_v2di __builtin_ia32_cvtpd2uqq128_mask(__gcc_v2df, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_cvtpd2uqq256_mask(__gcc_v4df, __gcc_v4di, unsigned char);
__gcc_v2di __builtin_ia32_cvtps2qq128_mask(__gcc_v4sf, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_cvtps2qq256_mask(__gcc_v4sf, __gcc_v4di, unsigned char);
__gcc_v2di __builtin_ia32_cvtps2uqq128_mask(__gcc_v4sf, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_cvtps2uqq256_mask(__gcc_v4sf, __gcc_v4di, unsigned char);
__gcc_v4sf __builtin_ia32_cvtqq2ps128_mask(__gcc_v2di, __gcc_v4sf, unsigned char);
__gcc_v2di __builtin_ia32_cvttpd2qq128_mask(__gcc_v2df, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_cvttpd2qq256_mask(__gcc_v4df, __gcc_v4di, unsigned char);
__gcc_v2di __builtin_ia32_cvttpd2uqq128_mask(__gcc_v2df, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_cvttpd2uqq256_mask(__gcc_v4df, __gcc_v4di, unsigned char);
__gcc_v2di __builtin_ia32_cvttps2qq128_mask(__gcc_v4sf, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_cvttps2qq256_mask(__gcc_v4sf, __gcc_v4di, unsigned char);
__gcc_v2di __builtin_ia32_cvttps2uqq128_mask(__gcc_v4sf, __gcc_v2di, unsigned char);
__gcc_v4di __builtin_ia32_cvttps2uqq256_mask(__gcc_v4sf, __gcc_v4di, unsigned char);
__gcc_v4sf __builtin_ia32_cvtuqq2ps128_mask(__gcc_v2di, __gcc_v4sf, unsigned char);
__gcc_v2df __builtin_ia32_rangepd128_mask(__gcc_v2df, __gcc_v2df, int, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_rangepd256_mask(__gcc_v4df, __gcc_v4df, int, __gcc_v4df, unsigned char);
__gcc_v4sf __builtin_ia32_rangeps128_mask(__gcc_v4sf, __gcc_v4sf, int, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_rangeps256_mask(__gcc_v8sf, __gcc_v8sf, int, __gcc_v8sf, unsigned char);
__gcc_v2df __builtin_ia32_reducepd128_mask(__gcc_v2df, int, __gcc_v2df, unsigned char);
__gcc_v4df __builtin_ia32_reducepd256_mask(__gcc_v4df, int, __gcc_v4df, unsigned char);
__gcc_v4sf __builtin_ia32_reduceps128_mask(__gcc_v4sf, int, __gcc_v4sf, unsigned char);
__gcc_v8sf __builtin_ia32_reduceps256_mask(__gcc_v8sf, int, __gcc_v8sf, unsigned char);
// avx512vl,avx512vbmi2
__gcc_v8hi __builtin_ia32_compresshi128_mask(__gcc_v8hi, __gcc_v8hi, unsigned char);
__gcc_v16hi __builtin_ia32_compresshi256_mask(__gcc_v16hi, __gcc_v16hi, unsigned short);
__gcc_v16qi __builtin_ia32_compressqi128_mask(__gcc_v16qi, __gcc_v16qi, unsigned short);
__gcc_v32qi __builtin_ia32_compressqi256_mask(__gcc_v32qi, __gcc_v32qi, unsigned int);
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_v8hi __builtin_ia32_expandhi128_mask(__gcc_v8hi, __gcc_v8hi, unsigned char);
__gcc_v16hi __builtin_ia32_expandhi256_mask(__gcc_v16hi, __gcc_v16hi, unsigned short);
__gcc_v8hi __builtin_ia32_expandloadhi128_mask(const __gcc_v8hi *, __gcc_v8hi, unsigned char);
__gcc_v16hi __builtin_ia32_expandloadhi256_mask(const __gcc_v16hi *, __gcc_v16hi, unsigned short);
__gcc_v16qi __builtin_ia32_expandloadqi128_mask(const __gcc_v16qi *, __gcc_v16qi, unsigned short);
__gcc_v32qi __builtin_ia32_expandloadqi256_mask(const __gcc_v32qi *, __gcc_v32qi, unsigned int);
__gcc_v16qi __builtin_ia32_expandqi128_mask(__gcc_v16qi, __gcc_v16qi, unsigned short);
__gcc_v32qi __builtin_ia32_expandqi256_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,evex512
__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,avx512vl
void __builtin_ia32_vp2intersect_d_128(__gcc_v4si, __gcc_v4si, unsigned char *, unsigned char *);
void __builtin_ia32_vp2intersect_d_256(__gcc_v8si, __gcc_v8si, unsigned char *, unsigned char *);
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 *);
// avx512vp2intersect,evex512
void __builtin_ia32_vp2intersect_d_512(__gcc_v16si, __gcc_v16si, unsigned short *, unsigned short *);
void __builtin_ia32_vp2intersect_q_512(__gcc_v8di, __gcc_v8di, unsigned char *, unsigned char *);
// 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);
// avx512vpopcntdq,evex512
__gcc_v16si __builtin_ia32_vpopcntd_512(__gcc_v16si);
__gcc_v8di __builtin_ia32_vpopcntq_512(__gcc_v8di);
// avxneconvert
__gcc_v4sf __builtin_ia32_vbcstnebf162ps128(const __bf16 *);
__gcc_v8sf __builtin_ia32_vbcstnebf162ps256(const __bf16 *);
__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 *);
// avxvnniint16
__gcc_v4si __builtin_ia32_vpdpwsud128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpwsud256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpwsuds128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpwsuds256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpwusd128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpwusd256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpwusds128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpwusds256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpwuud128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpwuud256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
__gcc_v4si __builtin_ia32_vpdpwuuds128(__gcc_v4si, __gcc_v4si, __gcc_v4si);
__gcc_v8si __builtin_ia32_vpdpwuuds256(__gcc_v8si, __gcc_v8si, __gcc_v8si);
// 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);
// bmi
unsigned int __builtin_ia32_bextr_u32(unsigned int, unsigned int);
unsigned long long int __builtin_ia32_bextr_u64(unsigned long long int, unsigned long long int);
// bmi2
unsigned long long int __builtin_ia32_bzhi_di(unsigned long long int, unsigned long long int);
unsigned int __builtin_ia32_bzhi_si(unsigned int, unsigned int);
unsigned long long int __builtin_ia32_pdep_di(unsigned long long int, unsigned long long int);
unsigned int __builtin_ia32_pdep_si(unsigned int, unsigned int);
unsigned long long int __builtin_ia32_pext_di(unsigned long long int, unsigned long long int);
unsigned int __builtin_ia32_pext_si(unsigned int, unsigned int);
// clang
unsigned char __builtin_ia32_addcarryx_u32(unsigned char, unsigned int, unsigned int, unsigned int *);
unsigned char __builtin_ia32_addcarryx_u64(unsigned char, unsigned long long int, unsigned long long int, unsigned long long int *);
void __builtin_ia32_pause(void);
unsigned long long int __builtin_ia32_rdpmc(int);
unsigned long long int __builtin_ia32_rdtsc(void);
unsigned long long int __builtin_ia32_rdtscp(unsigned int *);
unsigned int __builtin_ia32_readeflags_u32(void);
unsigned long long int __builtin_ia32_readeflags_u64(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 *);
unsigned short __builtin_ia32_tzcnt_u16(unsigned short);
unsigned int __builtin_ia32_tzcnt_u32(unsigned int);
unsigned long long int __builtin_ia32_tzcnt_u64(unsigned long long int);
__gcc_v2df __builtin_ia32_undef128(void);
__gcc_v4df __builtin_ia32_undef256(void);
__gcc_v8df __builtin_ia32_undef512(void);
void __builtin_ia32_wbinvd(void);
void __builtin_ia32_writeeflags_u32(unsigned int);
void __builtin_ia32_writeeflags_u64(unsigned long long int);
unsigned long long int __rdtsc(void);
// cldemote
void __builtin_ia32_cldemote(const void *);
// clflushopt
void __builtin_ia32_clflushopt(const void *);
// clwb
void __builtin_ia32_clwb(const void *);
// clzero
void __builtin_ia32_clzero(void *);
// 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);
// crc32
unsigned long long int __builtin_ia32_crc32di(unsigned long long int, unsigned long long int);
unsigned int __builtin_ia32_crc32hi(unsigned int, unsigned short);
unsigned int __builtin_ia32_crc32qi(unsigned int, unsigned char);
unsigned int __builtin_ia32_crc32si(unsigned int, unsigned int);
// enqcmd
unsigned char __builtin_ia32_enqcmd(void *, const void *);
unsigned char __builtin_ia32_enqcmds(void *, const void *);
// f16c
__gcc_v4sf __builtin_ia32_vcvtph2ps(__gcc_v8hi);
__gcc_v8sf __builtin_ia32_vcvtph2ps256(__gcc_v8hi);
__gcc_v8hi __builtin_ia32_vcvtps2ph(__gcc_v4sf, int);
__gcc_v8hi __builtin_ia32_vcvtps2ph256(__gcc_v8sf, int);
// fma
__gcc_v2df __builtin_ia32_vfmaddsd3(__gcc_v2df, __gcc_v2df, __gcc_v2df);
__gcc_v4sf __builtin_ia32_vfmaddss3(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf);
// fma4
__gcc_v2df __builtin_ia32_vfmaddsd(__gcc_v2df, __gcc_v2df, __gcc_v2df);
__gcc_v4sf __builtin_ia32_vfmaddss(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf);
// fma|fma4
__gcc_v2df __builtin_ia32_vfmaddpd(__gcc_v2df, __gcc_v2df, __gcc_v2df);
__gcc_v4df __builtin_ia32_vfmaddpd256(__gcc_v4df, __gcc_v4df, __gcc_v4df);
__gcc_v4sf __builtin_ia32_vfmaddps(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf);
__gcc_v8sf __builtin_ia32_vfmaddps256(__gcc_v8sf, __gcc_v8sf, __gcc_v8sf);
__gcc_v2df __builtin_ia32_vfmaddsubpd(__gcc_v2df, __gcc_v2df, __gcc_v2df);
__gcc_v4df __builtin_ia32_vfmaddsubpd256(__gcc_v4df, __gcc_v4df, __gcc_v4df);
__gcc_v4sf __builtin_ia32_vfmaddsubps(__gcc_v4sf, __gcc_v4sf, __gcc_v4sf);
__gcc_v8sf __builtin_ia32_vfmaddsubps256(__gcc_v8sf, __gcc_v8sf, __gcc_v8sf);
// fsgsbase
unsigned int __builtin_ia32_rdfsbase32(void);
unsigned long long int __builtin_ia32_rdfsbase64(void);
unsigned int __builtin_ia32_rdgsbase32(void);
unsigned long long int __builtin_ia32_rdgsbase64(void);
void __builtin_ia32_wrfsbase32(unsigned int);
void __builtin_ia32_wrfsbase64(unsigned long long int);
void __builtin_ia32_wrgsbase32(unsigned int);
void __builtin_ia32_wrgsbase64(unsigned long long int);
// fxsr
void __builtin_ia32_fxrstor(void *);
void __builtin_ia32_fxrstor64(void *);
void __builtin_ia32_fxsave(void *);
void __builtin_ia32_fxsave64(void *);
// gfni
__gcc_v16qi __builtin_ia32_vgf2p8affineinvqb_v16qi(__gcc_v16qi, __gcc_v16qi, char);
__gcc_v16qi __builtin_ia32_vgf2p8affineqb_v16qi(__gcc_v16qi, __gcc_v16qi, char);
__gcc_v16qi __builtin_ia32_vgf2p8mulb_v16qi(__gcc_v16qi, __gcc_v16qi);
// invpcid
void __builtin_ia32_invpcid(unsigned int, void *);
// kl
unsigned char __builtin_ia32_aesdec128kl_u8(__gcc_v2di *, __gcc_v2di, const void *);
unsigned char __builtin_ia32_aesdec256kl_u8(__gcc_v2di *, __gcc_v2di, const void *);
unsigned char __builtin_ia32_aesenc128kl_u8(__gcc_v2di *, __gcc_v2di, const void *);
unsigned char __builtin_ia32_aesenc256kl_u8(__gcc_v2di *, __gcc_v2di, const void *);
unsigned int __builtin_ia32_encodekey128_u32(unsigned int, __gcc_v2di, void *);
unsigned int __builtin_ia32_encodekey256_u32(unsigned int, __gcc_v2di, __gcc_v2di, void *);
void __builtin_ia32_loadiwkey(__gcc_v2di, __gcc_v2di, __gcc_v2di, unsigned int);
// kl,widekl
unsigned char __builtin_ia32_aesdecwide128kl_u8(__gcc_v2di *, const __gcc_v2di *, const void *);
unsigned char __builtin_ia32_aesdecwide256kl_u8(__gcc_v2di *, const __gcc_v2di *, const void *);
unsigned char __builtin_ia32_aesencwide128kl_u8(__gcc_v2di *, const __gcc_v2di *, const void *);
unsigned char __builtin_ia32_aesencwide256kl_u8(__gcc_v2di *, const __gcc_v2di *, const void *);
// lwp
void __builtin_ia32_llwpcb(void *);
unsigned char __builtin_ia32_lwpins32(unsigned int, unsigned int, unsigned int);
unsigned char __builtin_ia32_lwpins64(unsigned long long int, unsigned int, unsigned int);
void __builtin_ia32_lwpval32(unsigned int, unsigned int, unsigned int);
void __builtin_ia32_lwpval64(unsigned long long int, unsigned int, unsigned int);
void * __builtin_ia32_slwpcb(void);
// lzcnt
unsigned short __builtin_ia32_lzcnt_u16(unsigned short);
unsigned int __builtin_ia32_lzcnt_u32(unsigned int);
unsigned long long int __builtin_ia32_lzcnt_u64(unsigned long long int);
// clang-format on