234   return __builtin_shufflevector(__a, (__v8hf)_mm_setzero_ph(), 0, 1, 2, 3, 4,  in _mm256_zextph128_ph256()
254       __a, (__v8hf)_mm_setzero_ph(), 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,  in _mm512_zextph128_ph512()
280   __builtin_ia32_vcomish((__v8hf)A, (__v8hf)B, (int)(P), (int)(R))
287   return __builtin_ia32_vcomish((__v8hf)__A, (__v8hf)__B, _CMP_EQ_OS,  in _mm_comieq_sh()
293   return __builtin_ia32_vcomish((__v8hf)__A, (__v8hf)__B, _CMP_LT_OS,  in _mm_comilt_sh()
299   return __builtin_ia32_vcomish((__v8hf)__A, (__v8hf)__B, _CMP_LE_OS,  in _mm_comile_sh()
305   return __builtin_ia32_vcomish((__v8hf)__A, (__v8hf)__B, _CMP_GT_OS,  in _mm_comigt_sh()
311   return __builtin_ia32_vcomish((__v8hf)__A, (__v8hf)__B, _CMP_GE_OS,  in _mm_comige_sh()
317   return __builtin_ia32_vcomish((__v8hf)__A, (__v8hf)__B, _CMP_NEQ_US,  in _mm_comineq_sh()
323   return __builtin_ia32_vcomish((__v8hf)__A, (__v8hf)__B, _CMP_EQ_OQ,  in _mm_ucomieq_sh()
329   return __builtin_ia32_vcomish((__v8hf)__A, (__v8hf)__B, _CMP_LT_OQ,  in _mm_ucomilt_sh()
335   return __builtin_ia32_vcomish((__v8hf)__A, (__v8hf)__B, _CMP_LE_OQ,  in _mm_ucomile_sh()
341   return __builtin_ia32_vcomish((__v8hf)__A, (__v8hf)__B, _CMP_GT_OQ,  in _mm_ucomigt_sh()
347   return __builtin_ia32_vcomish((__v8hf)__A, (__v8hf)__B, _CMP_GE_OQ,  in _mm_ucomige_sh()
353   return __builtin_ia32_vcomish((__v8hf)__A, (__v8hf)__B, _CMP_NEQ_UQ,  in _mm_ucomineq_sh()
595       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
600       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
605       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
631       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
636       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
641       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
667       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
672       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
677       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
703       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
708       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
713       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
719       (__v8hf)__A, (__v8hf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)-1,  in _mm_min_sh()
727   return (__m128h)__builtin_ia32_minsh_round_mask((__v8hf)__A, (__v8hf)__B,  in _mm_mask_min_sh()
728                                                   (__v8hf)__W, (__mmask8)__U,  in _mm_mask_min_sh()
736       (__v8hf)__A, (__v8hf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)__U,  in _mm_maskz_min_sh()
742       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
747       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
752       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
758       (__v8hf)__A, (__v8hf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)-1,  in _mm_max_sh()
766   return (__m128h)__builtin_ia32_maxsh_round_mask((__v8hf)__A, (__v8hf)__B,  in _mm_mask_max_sh()
767                                                   (__v8hf)__W, (__mmask8)__U,  in _mm_mask_max_sh()
775       (__v8hf)__A, (__v8hf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)__U,  in _mm_maskz_max_sh()
781       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
786       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
791       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
811   ((__mmask8)__builtin_ia32_cmpsh_mask((__v8hf)(__m128h)(X),                   \
812                                        (__v8hf)(__m128h)(Y), (int)(P),         \
816   ((__mmask8)__builtin_ia32_cmpsh_mask((__v8hf)(__m128h)(X),                   \
817                                        (__v8hf)(__m128h)(Y), (int)(P),         \
822       (__v8hf)(__m128h)(X), (__v8hf)(__m128h)(Y), (int)(P), (__mmask8)-1,      \
827       (__v8hf)(__m128h)(X), (__v8hf)(__m128h)(Y), (int)(P), (__mmask8)(M),     \
840   __m128h src = (__v8hf)__builtin_shufflevector(  in _mm_mask_load_sh()
841       (__v8hf)__W, (__v8hf)_mm_setzero_ph(), 0, 8, 8, 8, 8, 8, 8, 8);  in _mm_mask_load_sh()
843   return (__m128h)__builtin_ia32_loadsh128_mask((const __v8hf *)__A, src, __U & 1);  in _mm_mask_load_sh()
849       (const __v8hf *)__A, (__v8hf)_mm_setzero_ph(), __U & 1);  in _mm_maskz_load_sh()
901   __builtin_ia32_storesh128_mask((__v8hf *)__W, __A, __U & 1);  in _mm_mask_store_sh()
1171       (__v8hf)__A, (__v8hf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)-1);  in _mm_rcp_sh()
1178   return (__m128h)__builtin_ia32_rcpsh_mask((__v8hf)__A, (__v8hf)__B,  in _mm_mask_rcp_sh()
1179                                             (__v8hf)__W, (__mmask8)__U);  in _mm_mask_rcp_sh()
1186       (__v8hf)__A, (__v8hf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)__U);  in _mm_maskz_rcp_sh()
1192       (__v8hf)__A, (__v8hf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)-1);  in _mm_rsqrt_sh()
1199   return (__m128h)__builtin_ia32_rsqrtsh_mask((__v8hf)__A, (__v8hf)__B,  in _mm_mask_rsqrt_sh()
1200                                               (__v8hf)__W, (__mmask8)__U);  in _mm_mask_rsqrt_sh()
1206       (__v8hf)__A, (__v8hf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)__U);  in _mm_maskz_rsqrt_sh()
1211       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (int)(((D) << 2) | (C)),     \
1212       (__v8hf)_mm_setzero_ph(), (__mmask8)-1, (int)(R)))
1216       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (int)(((D) << 2) | (C)),     \
1217       (__v8hf)_mm_setzero_ph(), (__mmask8)-1, _MM_FROUND_CUR_DIRECTION))
1221       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (int)(((D) << 2) | (C)),     \
1222       (__v8hf)(__m128h)(W), (__mmask8)(U), _MM_FROUND_CUR_DIRECTION))
1226       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (int)(((D) << 2) | (C)),     \
1227       (__v8hf)(__m128h)(W), (__mmask8)(U), (int)(R)))
1231       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (int)(((D) << 2) | (C)),     \
1232       (__v8hf)_mm_setzero_ph(), (__mmask8)(U), _MM_FROUND_CUR_DIRECTION))
1236       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (int)(((D) << 2) | (C)),     \
1237       (__v8hf)_mm_setzero_ph(), (__mmask8)(U), (int)(R)))
1241       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1247       (__v8hf)__A, (__v8hf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)-1,  in _mm_getexp_sh()
1254       (__v8hf)__A, (__v8hf)__B, (__v8hf)__W, (__mmask8)__U,  in _mm_mask_getexp_sh()
1260       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
1266       (__v8hf)__A, (__v8hf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)__U,  in _mm_maskz_getexp_sh()
1272       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1277       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1283       (__v8hf)__A, (__v8hf)(__B), (__v8hf)_mm_setzero_ph(), (__mmask8)-1,  in _mm_scalef_sh()
1289   return (__m128h)__builtin_ia32_scalefsh_round_mask((__v8hf)__A, (__v8hf)__B,  in _mm_mask_scalef_sh()
1290                                                      (__v8hf)__W, (__mmask8)__U,  in _mm_mask_scalef_sh()
1296       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
1302       (__v8hf)__A, (__v8hf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)__U,  in _mm_maskz_scalef_sh()
1308       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1313       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1318       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1323       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
1328       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
1333       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1338       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1343       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1348       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
1353       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1358       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1363       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
1368       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1407       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1412       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(W),        \
1417       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)_mm_setzero_ph(),    \
1423       (__v8hf)(__m128h)(__A), (__v8hf)(__m128h)(__B), (__v8hf)_mm_setzero_ph(),  in _mm_sqrt_sh()
1432       (__v8hf)(__m128h)(__A), (__v8hf)(__m128h)(__B), (__v8hf)(__m128h)(__W),  in _mm_mask_sqrt_sh()
1440       (__v8hf)(__m128h)(__A), (__v8hf)(__m128h)(__B), (__v8hf)_mm_setzero_ph(),  in _mm_maskz_sqrt_sh()
1453   ((__mmask8)__builtin_ia32_fpclasssh_mask((__v8hf)(__m128h)(A), (int)(imm),   \
1457   ((__mmask8)__builtin_ia32_fpclasssh_mask((__v8hf)(__m128h)(A), (int)(imm),   \
1462       (__v8df)(A), (__v8hf)_mm_undefined_ph(), (__mmask8)(-1), (int)(R)))
1465   ((__m128h)__builtin_ia32_vcvtpd2ph512_mask((__v8df)(A), (__v8hf)(W),         \
1470       (__v8df)(A), (__v8hf)_mm_setzero_ph(), (__mmask8)(U), (int)(R)))
1474       (__v8df)__A, (__v8hf)_mm_setzero_ph(), (__mmask8)-1,  in _mm512_cvtpd_ph()
1481       (__v8df)__A, (__v8hf)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);  in _mm512_mask_cvtpd_ph()
1487       (__v8df)__A, (__v8hf)_mm_setzero_ph(), (__mmask8)__U,  in _mm512_maskz_cvtpd_ph()
1493       (__v8hf)(A), (__v8df)_mm512_undefined_pd(), (__mmask8)(-1), (int)(R)))
1496   ((__m512d)__builtin_ia32_vcvtph2pd512_mask((__v8hf)(A), (__v8df)(W),         \
1501       (__v8hf)(A), (__v8df)_mm512_setzero_pd(), (__mmask8)(U), (int)(R)))
1505       (__v8hf)__A, (__v8df)_mm512_setzero_pd(), (__mmask8)-1,  in _mm512_cvtph_pd()
1512       (__v8hf)__A, (__v8df)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);  in _mm512_mask_cvtph_pd()
1518       (__v8hf)__A, (__v8df)_mm512_setzero_pd(), (__mmask8)__U,  in _mm512_maskz_cvtph_pd()
1523   ((__m128)__builtin_ia32_vcvtsh2ss_round_mask((__v4sf)(A), (__v8hf)(B),       \
1529       (__v4sf)(A), (__v8hf)(B), (__v4sf)(W), (__mmask8)(U), (int)(R)))
1532   ((__m128)__builtin_ia32_vcvtsh2ss_round_mask((__v4sf)(A), (__v8hf)(B),       \
1539       (__v4sf)__A, (__v8hf)__B, (__v4sf)_mm_undefined_ps(), (__mmask8)-1,  in _mm_cvtsh_ss()
1547   return (__m128)__builtin_ia32_vcvtsh2ss_round_mask((__v4sf)__A, (__v8hf)__B,  in _mm_mask_cvtsh_ss()
1556       (__v4sf)__A, (__v8hf)__B, (__v4sf)_mm_setzero_ps(), (__mmask8)__U,  in _mm_maskz_cvtsh_ss()
1561   ((__m128h)__builtin_ia32_vcvtss2sh_round_mask((__v8hf)(A), (__v4sf)(B),      \
1562                                                 (__v8hf)_mm_undefined_ph(),    \
1567       (__v8hf)(A), (__v4sf)(B), (__v8hf)(W), (__mmask8)(U), (int)(R)))
1570   ((__m128h)__builtin_ia32_vcvtss2sh_round_mask((__v8hf)(A), (__v4sf)(B),      \
1571                                                 (__v8hf)_mm_setzero_ph(),      \
1577       (__v8hf)__A, (__v4sf)__B, (__v8hf)_mm_undefined_ph(), (__mmask8)-1,  in _mm_cvtss_sh()
1586       (__v8hf)__A, (__v4sf)__B, (__v8hf)__W, (__mmask8)__U,  in _mm_mask_cvtss_sh()
1594       (__v8hf)__A, (__v4sf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)__U,  in _mm_maskz_cvtss_sh()
1599   ((__m128h)__builtin_ia32_vcvtsd2sh_round_mask((__v8hf)(A), (__v2df)(B),      \
1600                                                 (__v8hf)_mm_undefined_ph(),    \
1605       (__v8hf)(A), (__v2df)(B), (__v8hf)(W), (__mmask8)(U), (int)(R)))
1608   ((__m128h)__builtin_ia32_vcvtsd2sh_round_mask((__v8hf)(A), (__v2df)(B),      \
1609                                                 (__v8hf)_mm_setzero_ph(),      \
1615       (__v8hf)__A, (__v2df)__B, (__v8hf)_mm_undefined_ph(), (__mmask8)-1,  in _mm_cvtsd_sh()
1624       (__v8hf)__A, (__v2df)__B, (__v8hf)__W, (__mmask8)__U,  in _mm_mask_cvtsd_sh()
1631       (__v8hf)__A, (__v2df)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)__U,  in _mm_maskz_cvtsd_sh()
1636   ((__m128d)__builtin_ia32_vcvtsh2sd_round_mask((__v2df)(A), (__v8hf)(B),      \
1642       (__v2df)(A), (__v8hf)(B), (__v2df)(W), (__mmask8)(U), (int)(R)))
1645   ((__m128d)__builtin_ia32_vcvtsh2sd_round_mask((__v2df)(A), (__v8hf)(B),      \
1652       (__v2df)__A, (__v8hf)__B, (__v2df)_mm_undefined_pd(), (__mmask8)-1,  in _mm_cvtsh_sd()
1661       (__v2df)__A, (__v8hf)__B, (__v2df)__W, (__mmask8)__U,  in _mm_mask_cvtsh_sd()
1668       (__v2df)__A, (__v8hf)__B, (__v2df)_mm_setzero_pd(), (__mmask8)__U,  in _mm_maskz_cvtsh_sd()
2078       (__v8di)(A), (__v8hf)_mm_undefined_ph(), (__mmask8)(-1), (int)(R)))
2081   ((__m128h)__builtin_ia32_vcvtqq2ph512_mask((__v8di)(A), (__v8hf)(W),         \
2086       (__v8di)(A), (__v8hf)_mm_setzero_ph(), (__mmask8)(U), (int)(R)))
2091       (__v8di)__A, (__v8hf)_mm_setzero_ph(), (__mmask8)-1,  in _mm512_cvtepi64_ph()
2098       (__v8di)__A, (__v8hf)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);  in _mm512_mask_cvtepi64_ph()
2104       (__v8di)__A, (__v8hf)_mm_setzero_ph(), (__mmask8)__U,  in _mm512_maskz_cvtepi64_ph()
2109   ((__m512i)__builtin_ia32_vcvtph2qq512_mask((__v8hf)(A),                      \
2114   ((__m512i)__builtin_ia32_vcvtph2qq512_mask((__v8hf)(A), (__v8di)(W),         \
2119       (__v8hf)(A), (__v8di)_mm512_setzero_epi32(), (__mmask8)(U), (int)(R)))
2124       (__v8hf)__A, (__v8di)_mm512_setzero_epi32(), (__mmask8)-1,  in _mm512_cvtph_epi64()
2131       (__v8hf)__A, (__v8di)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);  in _mm512_mask_cvtph_epi64()
2137       (__v8hf)__A, (__v8di)_mm512_setzero_epi32(), (__mmask8)__U,  in _mm512_maskz_cvtph_epi64()
2143       (__v8du)(A), (__v8hf)_mm_undefined_ph(), (__mmask8)(-1), (int)(R)))
2146   ((__m128h)__builtin_ia32_vcvtuqq2ph512_mask((__v8du)(A), (__v8hf)(W),        \
2151       (__v8du)(A), (__v8hf)_mm_setzero_ph(), (__mmask8)(U), (int)(R)))
2156       (__v8du)__A, (__v8hf)_mm_setzero_ph(), (__mmask8)-1,  in _mm512_cvtepu64_ph()
2163       (__v8du)__A, (__v8hf)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);  in _mm512_mask_cvtepu64_ph()
2169       (__v8du)__A, (__v8hf)_mm_setzero_ph(), (__mmask8)__U,  in _mm512_maskz_cvtepu64_ph()
2175       (__v8hf)(A), (__v8du)_mm512_undefined_epi32(), (__mmask8)(-1),           \
2179   ((__m512i)__builtin_ia32_vcvtph2uqq512_mask((__v8hf)(A), (__v8du)(W),        \
2184       (__v8hf)(A), (__v8du)_mm512_setzero_epi32(), (__mmask8)(U), (int)(R)))
2189       (__v8hf)__A, (__v8du)_mm512_setzero_epi32(), (__mmask8)-1,  in _mm512_cvtph_epu64()
2196       (__v8hf)__A, (__v8du)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);  in _mm512_mask_cvtph_epu64()
2202       (__v8hf)__A, (__v8du)_mm512_setzero_epi32(), (__mmask8)__U,  in _mm512_maskz_cvtph_epu64()
2208       (__v8hf)(A), (__v8di)_mm512_undefined_epi32(), (__mmask8)(-1),           \
2212   ((__m512i)__builtin_ia32_vcvttph2qq512_mask((__v8hf)(A), (__v8di)(W),        \
2217       (__v8hf)(A), (__v8di)_mm512_setzero_epi32(), (__mmask8)(U), (int)(R)))
2222       (__v8hf)__A, (__v8di)_mm512_setzero_epi32(), (__mmask8)-1,  in _mm512_cvttph_epi64()
2229       (__v8hf)__A, (__v8di)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);  in _mm512_mask_cvttph_epi64()
2235       (__v8hf)__A, (__v8di)_mm512_setzero_epi32(), (__mmask8)__U,  in _mm512_maskz_cvttph_epi64()
2241       (__v8hf)(A), (__v8du)_mm512_undefined_epi32(), (__mmask8)(-1),           \
2245   ((__m512i)__builtin_ia32_vcvttph2uqq512_mask((__v8hf)(A), (__v8du)(W),       \
2250       (__v8hf)(A), (__v8du)_mm512_setzero_epi32(), (__mmask8)(U), (int)(R)))
2255       (__v8hf)__A, (__v8du)_mm512_setzero_epi32(), (__mmask8)-1,  in _mm512_cvttph_epu64()
2262       (__v8hf)__A, (__v8du)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);  in _mm512_mask_cvttph_epu64()
2268       (__v8hf)__A, (__v8du)_mm512_setzero_epi32(), (__mmask8)__U,  in _mm512_maskz_cvttph_epu64()
2273   ((int)__builtin_ia32_vcvtsh2si32((__v8hf)(A), (int)(R)))
2276   return (int)__builtin_ia32_vcvtsh2si32((__v8hf)__A, _MM_FROUND_CUR_DIRECTION);  in _mm_cvtsh_i32()
2280   ((unsigned int)__builtin_ia32_vcvtsh2usi32((__v8hf)(A), (int)(R)))
2284   return (unsigned int)__builtin_ia32_vcvtsh2usi32((__v8hf)__A,  in _mm_cvtsh_u32()
2290   ((long long)__builtin_ia32_vcvtsh2si64((__v8hf)(A), (int)(R)))
2293   return (long long)__builtin_ia32_vcvtsh2si64((__v8hf)__A,  in _mm_cvtsh_i64()
2298   ((unsigned long long)__builtin_ia32_vcvtsh2usi64((__v8hf)(A), (int)(R)))
2303       (__v8hf)__A, _MM_FROUND_CUR_DIRECTION);  in _mm_cvtsh_u64()
2308   ((__m128h)__builtin_ia32_vcvtusi2sh((__v8hf)(A), (unsigned int)(B), (int)(R)))
2318   ((__m128h)__builtin_ia32_vcvtusi642sh((__v8hf)(A), (unsigned long long)(B),  \
2329   ((__m128h)__builtin_ia32_vcvtsi2sh((__v8hf)(A), (int)(B), (int)(R)))
2339   ((__m128h)__builtin_ia32_vcvtsi642sh((__v8hf)(A), (long long)(B), (int)(R)))
2349   ((int)__builtin_ia32_vcvttsh2si32((__v8hf)(A), (int)(R)))
2352   return (int)__builtin_ia32_vcvttsh2si32((__v8hf)__A,  in _mm_cvttsh_i32()
2358   ((long long)__builtin_ia32_vcvttsh2si64((__v8hf)(A), (int)(R)))
2361   return (long long)__builtin_ia32_vcvttsh2si64((__v8hf)__A,  in _mm_cvttsh_i64()
2367   ((unsigned int)__builtin_ia32_vcvttsh2usi32((__v8hf)(A), (int)(R)))
2371   return (unsigned int)__builtin_ia32_vcvttsh2usi32((__v8hf)__A,  in _mm_cvttsh_u32()
2377   ((unsigned long long)__builtin_ia32_vcvttsh2usi64((__v8hf)(A), (int)(R)))
2382       (__v8hf)__A, _MM_FROUND_CUR_DIRECTION);  in _mm_cvttsh_u64()
2745   return __builtin_ia32_vfmaddsh3_mask((__v8hf)__W, (__v8hf)__A, (__v8hf)__B,  in _mm_fmadd_sh()
2753   return __builtin_ia32_vfmaddsh3_mask((__v8hf)__W, (__v8hf)__A, (__v8hf)__B,  in _mm_mask_fmadd_sh()
2759       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(C),        \
2764       (__v8hf)(__m128h)(W), (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B),        \
2769   return __builtin_ia32_vfmaddsh3_maskz((__v8hf)__A, (__v8hf)__B, (__v8hf)__C,  in _mm_maskz_fmadd_sh()
2776       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), (__v8hf)(__m128h)(C),        \
2781   return __builtin_ia32_vfmaddsh3_mask3((__v8hf)__W, (__v8hf)__X, (__v8hf)__Y,  in _mm_mask3_fmadd_sh()
2788       (__v8hf)(__m128h)(W), (__v8hf)(__m128h)(X), (__v8hf)(__m128h)(Y),        \
2794   return (__m128h)__builtin_ia32_vfmaddsh3_mask((__v8hf)__W, (__v8hf)__A,  in _mm_fmsub_sh()
2795                                                 -(__v8hf)__B, (__mmask8)-1,  in _mm_fmsub_sh()
2803   return (__m128h)__builtin_ia32_vfmaddsh3_mask((__v8hf)__W, (__v8hf)__A,  in _mm_mask_fmsub_sh()
2804                                                 -(__v8hf)__B, (__mmask8)__U,  in _mm_mask_fmsub_sh()
2810       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), -(__v8hf)(__m128h)(C),       \
2815       (__v8hf)(__m128h)(W), (__v8hf)(__m128h)(A), -(__v8hf)(__m128h)(B),       \
2820   return (__m128h)__builtin_ia32_vfmaddsh3_maskz((__v8hf)__A, (__v8hf)__B,  in _mm_maskz_fmsub_sh()
2821                                                  -(__v8hf)__C, (__mmask8)__U,  in _mm_maskz_fmsub_sh()
2827       (__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B), -(__v8hf)(__m128h)(C),       \
2832   return __builtin_ia32_vfmsubsh3_mask3((__v8hf)__W, (__v8hf)__X, (__v8hf)__Y,  in _mm_mask3_fmsub_sh()
2839       (__v8hf)(__m128h)(W), (__v8hf)(__m128h)(X), (__v8hf)(__m128h)(Y),        \
2845   return __builtin_ia32_vfmaddsh3_mask((__v8hf)__W, -(__v8hf)__A, (__v8hf)__B,  in _mm_fnmadd_sh()
2851   return __builtin_ia32_vfmaddsh3_mask((__v8hf)__W, -(__v8hf)__A, (__v8hf)__B,  in _mm_mask_fnmadd_sh()
2857       (__v8hf)(__m128h)(A), -(__v8hf)(__m128h)(B), (__v8hf)(__m128h)(C),       \
2862       (__v8hf)(__m128h)(W), -(__v8hf)(__m128h)(A), (__v8hf)(__m128h)(B),       \
2867   return __builtin_ia32_vfmaddsh3_maskz((__v8hf)__A, -(__v8hf)__B, (__v8hf)__C,  in _mm_maskz_fnmadd_sh()
2874       (__v8hf)(__m128h)(A), -(__v8hf)(__m128h)(B), (__v8hf)(__m128h)(C),       \
2879   return __builtin_ia32_vfmaddsh3_mask3((__v8hf)__W, -(__v8hf)__X, (__v8hf)__Y,  in _mm_mask3_fnmadd_sh()
2886       (__v8hf)(__m128h)(W), -(__v8hf)(__m128h)(X), (__v8hf)(__m128h)(Y),       \
2892   return __builtin_ia32_vfmaddsh3_mask((__v8hf)__W, -(__v8hf)__A, -(__v8hf)__B,  in _mm_fnmsub_sh()
2898   return __builtin_ia32_vfmaddsh3_mask((__v8hf)__W, -(__v8hf)__A, -(__v8hf)__B,  in _mm_mask_fnmsub_sh()
2904       (__v8hf)(__m128h)(A), -(__v8hf)(__m128h)(B), -(__v8hf)(__m128h)(C),      \
2909       (__v8hf)(__m128h)(W), -(__v8hf)(__m128h)(A), -(__v8hf)(__m128h)(B),      \
2914   return __builtin_ia32_vfmaddsh3_maskz((__v8hf)__A, -(__v8hf)__B, -(__v8hf)__C,  in _mm_maskz_fnmsub_sh()
2921       (__v8hf)(__m128h)(A), -(__v8hf)(__m128h)(B), -(__v8hf)(__m128h)(C),      \
2926   return __builtin_ia32_vfmsubsh3_mask3((__v8hf)__W, -(__v8hf)__X, (__v8hf)__Y,  in _mm_mask3_fnmsub_sh()
2933       (__v8hf)(__m128h)(W), -(__v8hf)(__m128h)(X), (__v8hf)(__m128h)(Y),       \