1 /* Same as sse-14, except converted to use #pragma GCC option. */ 2 /* { dg-do compile } */ 3 /* { dg-options "-O0 -Werror-implicit-function-declaration -march=k8" } */ 4 /* { dg-add-options bind_pic_locally } */ 5 6 #include <mm_malloc.h> 7 8 /* Test that the intrinsics compile with optimization. All of them 9 are defined as inline functions in {,x,e,p,t,s,w,a,b,i}mmintrin.h, 10 mm3dnow.h, fma4intrin.h, xopintrin.h, abmintrin.h, bmiintrin.h, 11 tbmintrin.h, lwpintrin.h, popcntintrin.h, fmaintrin.h, 12 avx5124fmapsintrin.h, avx5124vnniwintrin.h, avx512vpopcntdqintrin.h, 13 avx512bitalgintrin.h, avx512vp2intersectintrin.h, 14 avx512vp2intersectvlintrin.h and mm_malloc.h that reference the proper 15 builtin functions. 16 Defining away "extern" and "__inline" results in all of them being 17 compiled as proper functions. */ 18 19 #define extern 20 #define __inline 21 22 #define _CONCAT(x,y) x ## y 23 24 #define test_0(func, type, imm) \ 25 type _CONCAT(_,func) (int const I) \ 26 { return func (imm); } 27 28 #define test_1(func, type, op1_type, imm) \ 29 type _CONCAT(_,func) (op1_type A, int const I) \ 30 { return func (A, imm); } 31 32 #define test_1x(func, type, op1_type, imm1, imm2) \ 33 type _CONCAT(_,func) (op1_type A, int const I, int const L) \ 34 { return func (A, imm1, imm2); } 35 36 #define test_1y(func, type, op1_type, imm1, imm2, imm3) \ 37 type _CONCAT(_,func) (op1_type A, int const I, int const L, int const R)\ 38 { return func (A, imm1, imm2, imm3); } 39 40 #define test_2(func, type, op1_type, op2_type, imm) \ 41 type _CONCAT(_,func) (op1_type A, op2_type B, int const I) \ 42 { return func (A, B, imm); } 43 44 #define test_2x(func, type, op1_type, op2_type, imm1, imm2) \ 45 type _CONCAT(_,func) (op1_type A, op2_type B, int const I, int const L) \ 46 { return func (A, B, imm1, imm2); } 47 48 #define test_2y(func, type, op1_type, op2_type, imm1, imm2, imm3) \ 49 type _CONCAT(_,func) (op1_type A, op2_type B, int const I, int const L,\ 50 int const R) \ 51 { return func (A, B, imm1, imm2, imm3); } 52 53 #define test_2vx(func, op1_type, op2_type, imm1, imm2) \ 54 void _CONCAT(_,func) (op1_type A, op2_type B, int const I, int const L) \ 55 { func (A, B, imm1, imm2); } 56 57 #define test_3(func, type, op1_type, op2_type, op3_type, imm) \ 58 type _CONCAT(_,func) (op1_type A, op2_type B, \ 59 op3_type C, int const I) \ 60 { return func (A, B, C, imm); } 61 62 #define test_3x(func, type, op1_type, op2_type, op3_type, imm1, imm2) \ 63 type _CONCAT(_,func) (op1_type A, op2_type B, \ 64 op3_type C, int const I, int const L) \ 65 { return func (A, B, C, imm1, imm2); } 66 67 #define test_3y(func, type, op1_type, op2_type, op3_type, imm1, imm2, imm3) \ 68 type _CONCAT(_,func) (op1_type A, op2_type B, \ 69 op3_type C, int const I, int const L, int const R) \ 70 { return func (A, B, C, imm1, imm2, imm3); } 71 72 #define test_3v(func, op1_type, op2_type, op3_type, imm) \ 73 int _CONCAT(_,func) (op1_type A, op2_type B, \ 74 op3_type C, int const I) \ 75 { func (A, B, C, imm); } 76 77 #define test_3vx(func, op1_type, op2_type, op3_type, imm1, imm2) \ 78 void _CONCAT(_,func) (op1_type A, op2_type B, \ 79 op3_type C, int const I, int const L) \ 80 { func (A, B, C, imm1, imm2); } 81 82 #define test_4(func, type, op1_type, op2_type, op3_type, op4_type, imm) \ 83 type _CONCAT(_,func) (op1_type A, op2_type B, \ 84 op3_type C, op4_type D, int const I) \ 85 { return func (A, B, C, D, imm); } 86 87 #define test_4x(func, type, op1_type, op2_type, op3_type, op4_type, imm1, imm2) \ 88 type _CONCAT(_,func) (op1_type A, op2_type B, \ 89 op3_type C, op4_type D, int const I, int const L) \ 90 { return func (A, B, C, D, imm1, imm2); } 91 92 #define test_4y(func, type, op1_type, op2_type, op3_type, op4_type, imm1, imm2, imm3) \ 93 type _CONCAT(_,func) (op1_type A, op2_type B, op3_type C, \ 94 op4_type D, int const I, int const L, int const R) \ 95 { return func (A, B, C, D, imm1, imm2, imm3); } 96 97 98 #define test_4v(func, op1_type, op2_type, op3_type, op4_type, imm) \ 99 int _CONCAT(_,func) (op1_type A, op2_type B, \ 100 op3_type C, op4_type D, int const I) \ 101 { func (A, B, C, D, imm); } 102 103 104 #ifndef DIFFERENT_PRAGMAS 105 #pragma GCC target ("sse4a,3dnow,avx,avx2,fma4,xop,aes,pclmul,popcnt,abm,lzcnt,bmi,bmi2,tbm,lwp,fsgsbase,rdrnd,f16c,rtm,rdseed,prfchw,adx,fxsr,xsaveopt,avx512f,avx512er,avx512cd,avx512pf,sha,prefetchwt1,avx512vl,avx512bw,avx512dq,avx512vbmi,avx512vbmi2,avx512ifma,avx5124fmaps,avx5124vnniw,avx512vpopcntdq,gfni,avx512bitalg,avx512bf16,avx512vp2intersect") 106 #endif 107 108 /* Following intrinsics require immediate arguments. They 109 are defined as macros for non-optimized compilations. */ 110 111 /* mmintrin.h (MMX). */ 112 #ifdef DIFFERENT_PRAGMAS 113 #pragma GCC target ("mmx") 114 #endif 115 #include <mmintrin.h> 116 117 /* mm3dnow.h (3DNOW). */ 118 #ifdef DIFFERENT_PRAGMAS 119 #pragma GCC target ("3dnow") 120 #endif 121 #include <mm3dnow.h> 122 123 /* xmmintrin.h (SSE). */ 124 #ifdef DIFFERENT_PRAGMAS 125 #pragma GCC target ("sse") 126 #endif 127 #include <xmmintrin.h> 128 test_2 (_mm_shuffle_ps, __m128, __m128, __m128, 1) 129 test_1 (_mm_extract_pi16, int, __m64, 1) 130 test_1 (_m_pextrw, int, __m64, 1) 131 test_2 (_mm_insert_pi16, __m64, __m64, int, 1) 132 test_2 (_m_pinsrw, __m64, __m64, int, 1) 133 test_1 (_mm_shuffle_pi16, __m64, __m64, 1) 134 test_1 (_m_pshufw, __m64, __m64, 1) 135 test_1 (_mm_prefetch, void, void *, _MM_HINT_NTA) 136 137 /* emmintrin.h (SSE2). */ 138 #ifdef DIFFERENT_PRAGMAS 139 #pragma GCC target ("sse2") 140 #endif 141 #include <emmintrin.h> 142 test_2 (_mm_shuffle_pd, __m128d, __m128d, __m128d, 1) 143 test_1 (_mm_bsrli_si128, __m128i, __m128i, 1) 144 test_1 (_mm_bslli_si128, __m128i, __m128i, 1) 145 test_1 (_mm_srli_si128, __m128i, __m128i, 1) 146 test_1 (_mm_slli_si128, __m128i, __m128i, 1) 147 test_1 (_mm_extract_epi16, int, __m128i, 1) 148 test_2 (_mm_insert_epi16, __m128i, __m128i, int, 1) 149 test_1 (_mm_shufflehi_epi16, __m128i, __m128i, 1) 150 test_1 (_mm_shufflelo_epi16, __m128i, __m128i, 1) 151 test_1 (_mm_shuffle_epi32, __m128i, __m128i, 1) 152 153 /* pmmintrin.h (SSE3). */ 154 #ifdef DIFFERENT_PRAGMAS 155 #pragma GCC target ("sse3") 156 #endif 157 #include <pmmintrin.h> 158 159 /* tmmintrin.h (SSSE3). */ 160 #ifdef DIFFERENT_PRAGMAS 161 #pragma GCC target ("ssse3") 162 #endif 163 #include <tmmintrin.h> 164 test_2 (_mm_alignr_epi8, __m128i, __m128i, __m128i, 1) 165 test_2 (_mm_alignr_pi8, __m64, __m64, __m64, 1) 166 167 /* ammintrin.h (SSE4A). */ 168 #ifdef DIFFERENT_PRAGMAS 169 #pragma GCC target ("sse4a") 170 #endif 171 #include <ammintrin.h> 172 test_1x (_mm_extracti_si64, __m128i, __m128i, 1, 1) 173 test_2x (_mm_inserti_si64, __m128i, __m128i, __m128i, 1, 1) 174 175 /* Note, nmmintrin.h includes smmintrin.h, and smmintrin.h 176 checks for the #ifdef. So just set the option to SSE4.2. */ 177 #ifdef DIFFERENT_PRAGMAS 178 #pragma GCC target ("sse4.2") 179 #endif 180 #include <nmmintrin.h> 181 /* smmintrin.h (SSE4.2). */ 182 test_1 (_mm_round_pd, __m128d, __m128d, 9) 183 test_1 (_mm_round_ps, __m128, __m128, 9) 184 test_2 (_mm_round_sd, __m128d, __m128d, __m128d, 9) 185 test_2 (_mm_round_ss, __m128, __m128, __m128, 9) 186 187 test_2 (_mm_blend_epi16, __m128i, __m128i, __m128i, 1) 188 test_2 (_mm_blend_ps, __m128, __m128, __m128, 1) 189 test_2 (_mm_blend_pd, __m128d, __m128d, __m128d, 1) 190 test_2 (_mm_dp_ps, __m128, __m128, __m128, 1) 191 test_2 (_mm_dp_pd, __m128d, __m128d, __m128d, 1) 192 test_2 (_mm_insert_ps, __m128, __m128, __m128, 1) 193 test_1 (_mm_extract_ps, int, __m128, 1) 194 test_2 (_mm_insert_epi8, __m128i, __m128i, int, 1) 195 test_2 (_mm_insert_epi32, __m128i, __m128i, int, 1) 196 #ifdef __x86_64__ 197 test_2 (_mm_insert_epi64, __m128i, __m128i, long long, 1) 198 #endif 199 test_1 (_mm_extract_epi8, int, __m128i, 1) 200 test_1 (_mm_extract_epi32, int, __m128i, 1) 201 #ifdef __x86_64__ 202 test_1 (_mm_extract_epi64, long long, __m128i, 1) 203 #endif 204 test_2 (_mm_mpsadbw_epu8, __m128i, __m128i, __m128i, 1) 205 test_2 (_mm_cmpistrm, __m128i, __m128i, __m128i, 1) 206 test_2 (_mm_cmpistri, int, __m128i, __m128i, 1) 207 test_4 (_mm_cmpestrm, __m128i, __m128i, int, __m128i, int, 1) 208 test_4 (_mm_cmpestri, int, __m128i, int, __m128i, int, 1) 209 test_2 (_mm_cmpistra, int, __m128i, __m128i, 1) 210 test_2 (_mm_cmpistrc, int, __m128i, __m128i, 1) 211 test_2 (_mm_cmpistro, int, __m128i, __m128i, 1) 212 test_2 (_mm_cmpistrs, int, __m128i, __m128i, 1) 213 test_2 (_mm_cmpistrz, int, __m128i, __m128i, 1) 214 test_4 (_mm_cmpestra, int, __m128i, int, __m128i, int, 1) 215 test_4 (_mm_cmpestrc, int, __m128i, int, __m128i, int, 1) 216 test_4 (_mm_cmpestro, int, __m128i, int, __m128i, int, 1) 217 test_4 (_mm_cmpestrs, int, __m128i, int, __m128i, int, 1) 218 test_4 (_mm_cmpestrz, int, __m128i, int, __m128i, int, 1) 219 220 /* immintrin.h (AVX/AVX2/RDRND/FSGSBASE/F16C/RTM/AVX512F/SHA) */ 221 #ifdef DIFFERENT_PRAGMAS 222 #pragma GCC target ("avx,avx2,rdrnd,fsgsbase,f16c,rtm,avx512f,avx512er,avx512cd,avx512pf,sha,avx512vl,avx512bw,avx512dq,avx512ifma,avx512vbmi,avx512vbmi2,avx5124fmaps,avx5124vnniw,avx512vpopcntdq,gfni,avx512bitalg,avx512bf16,avx512vp2intersect") 223 #endif 224 #include <immintrin.h> 225 test_1 (_cvtss_sh, unsigned short, float, 1) 226 test_1 (_mm_cvtps_ph, __m128i, __m128, 1) 227 test_1 (_mm256_cvtps_ph, __m128i, __m256, 1) 228 229 /* avxintrin.h */ 230 test_2 (_mm256_blend_pd, __m256d, __m256d, __m256d, 1) 231 test_2 (_mm256_blend_ps, __m256, __m256, __m256, 1) 232 test_2 (_mm256_dp_ps, __m256, __m256, __m256, 1) 233 test_2 (_mm256_shuffle_pd, __m256d, __m256d, __m256d, 1) 234 test_2 (_mm256_shuffle_ps, __m256, __m256, __m256, 1) 235 test_2 (_mm_cmp_sd, __m128d, __m128d, __m128d, 1) 236 test_2 (_mm_cmp_ss, __m128, __m128, __m128, 1) 237 test_2 (_mm_cmp_pd, __m128d, __m128d, __m128d, 1) 238 test_2 (_mm_cmp_ps, __m128, __m128, __m128, 1) 239 test_2 (_mm256_cmp_pd, __m256d, __m256d, __m256d, 1) 240 test_2 (_mm256_cmp_ps, __m256, __m256, __m256, 1) 241 test_1 (_mm256_extractf128_pd, __m128d, __m256d, 1) 242 test_1 (_mm256_extractf128_ps, __m128, __m256, 1) 243 test_1 (_mm256_extractf128_si256, __m128i, __m256i, 1) 244 test_1 (_mm256_extract_epi8, int, __m256i, 20) 245 test_1 (_mm256_extract_epi16, int, __m256i, 13) 246 test_1 (_mm256_extract_epi32, int, __m256i, 6) 247 #ifdef __x86_64__ 248 test_1 (_mm256_extract_epi64, long long, __m256i, 2) 249 #endif 250 test_1 (_mm_permute_pd, __m128d, __m128d, 1) 251 test_1 (_mm256_permute_pd, __m256d, __m256d, 1) 252 test_1 (_mm_permute_ps, __m128, __m128, 1) 253 test_1 (_mm256_permute_ps, __m256, __m256, 1) 254 test_2 (_mm256_permute2f128_pd, __m256d, __m256d, __m256d, 1) 255 test_2 (_mm256_permute2f128_ps, __m256, __m256, __m256, 1) 256 test_2 (_mm256_permute2f128_si256, __m256i, __m256i, __m256i, 1) 257 test_2 (_mm256_insertf128_pd, __m256d, __m256d, __m128d, 1) 258 test_2 (_mm256_insertf128_ps, __m256, __m256, __m128, 1) 259 test_2 (_mm256_insertf128_si256, __m256i, __m256i, __m128i, 1) 260 test_2 (_mm256_insert_epi8, __m256i, __m256i, int, 30) 261 test_2 (_mm256_insert_epi16, __m256i, __m256i, int, 7) 262 test_2 (_mm256_insert_epi32, __m256i, __m256i, int, 3) 263 #ifdef __x86_64__ 264 test_2 (_mm256_insert_epi64, __m256i, __m256i, long long, 1) 265 #endif 266 test_1 (_mm256_round_pd, __m256d, __m256d, 9) 267 test_1 (_mm256_round_ps, __m256, __m256, 9) 268 269 /* avx2intrin.h */ 270 test_2 ( _mm256_mpsadbw_epu8, __m256i, __m256i, __m256i, 1) 271 test_2 ( _mm256_alignr_epi8, __m256i, __m256i, __m256i, 1) 272 test_2 ( _mm256_blend_epi16, __m256i, __m256i, __m256i, 1) 273 test_1 ( _mm256_shuffle_epi32, __m256i, __m256i, 1) 274 test_1 ( _mm256_shufflehi_epi16, __m256i, __m256i, 1) 275 test_1 ( _mm256_shufflelo_epi16, __m256i, __m256i, 1) 276 test_1 ( _mm256_bslli_epi128, __m256i, __m256i, 8) 277 test_1 ( _mm256_bsrli_epi128, __m256i, __m256i, 8) 278 test_1 ( _mm256_slli_si256, __m256i, __m256i, 8) 279 test_1 ( _mm256_srli_si256, __m256i, __m256i, 8) 280 test_2 ( _mm_blend_epi32, __m128i, __m128i, __m128i, 1) 281 test_2 ( _mm256_blend_epi32, __m256i, __m256i, __m256, 1) 282 test_1 ( _mm256_permute4x64_pd, __m256d, __m256d, 1) 283 test_1 ( _mm256_permute4x64_epi64, __m256i, __m256i, 1) 284 test_2 ( _mm256_permute2x128_si256, __m256i, __m256i, __m256i, 1) 285 test_1 ( _mm256_extracti128_si256, __m128i, __m256i, 1) 286 test_2 ( _mm256_inserti128_si256, __m256i, __m256i, __m128i, 1) 287 test_2 ( _mm_i32gather_pd, __m128d, double const *, __m128i, 1) 288 test_2 ( _mm256_i32gather_pd, __m256d, double const *, __m128i, 1) 289 test_2 ( _mm_i64gather_pd, __m128d, double const *, __m128i, 1) 290 test_2 ( _mm256_i64gather_pd, __m256d, double const *, __m256i, 1) 291 test_2 ( _mm_i32gather_ps, __m128, float const *, __m128i, 1) 292 test_2 ( _mm256_i32gather_ps, __m256, float const *, __m256i, 1) 293 test_2 ( _mm_i64gather_ps, __m128, float const *, __m128i, 1) 294 test_2 ( _mm256_i64gather_ps, __m128, float const *, __m256i, 1) 295 test_2 ( _mm_i32gather_epi64, __m128i, long long int const *, __m128i, 1) 296 test_2 ( _mm256_i32gather_epi64, __m256i, long long int const *, __m128i, 1) 297 test_2 ( _mm_i64gather_epi64, __m128i, long long int const *, __m128i, 1) 298 test_2 ( _mm256_i64gather_epi64, __m256i, long long int const *, __m256i, 1) 299 test_2 ( _mm_i32gather_epi32, __m128i, int const *, __m128i, 1) 300 test_2 ( _mm256_i32gather_epi32, __m256i, int const *, __m256i, 1) 301 test_2 ( _mm_i64gather_epi32, __m128i, int const *, __m128i, 1) 302 test_2 ( _mm256_i64gather_epi32, __m128i, int const *, __m256i, 1) 303 304 /* rtmintrin.h */ 305 test_0 ( _xabort, void, 1) 306 307 /* avx512fintrin.h */ 308 test_1 (_mm512_cvt_roundepi32_ps, __m512, __m512i, 9) 309 test_1 (_mm512_cvt_roundepu32_ps, __m512, __m512i, 9) 310 test_1 (_mm512_cvt_roundpd_epi32, __m256i, __m512d, 9) 311 test_1 (_mm512_cvt_roundpd_epu32, __m256i, __m512d, 9) 312 test_1 (_mm512_cvt_roundpd_ps, __m256, __m512d, 9) 313 test_1 (_mm512_cvt_roundph_ps, __m512, __m256i, 8) 314 test_1 (_mm512_cvt_roundps_epi32, __m512i, __m512, 9) 315 test_1 (_mm512_cvt_roundps_epu32, __m512i, __m512, 9) 316 test_1 (_mm512_cvt_roundps_pd, __m512d, __m256, 8) 317 test_1 (_mm512_cvtps_ph, __m256i, __m512, 1) 318 test_1 (_mm512_cvtt_roundpd_epi32, __m256i, __m512d, 8) 319 test_1 (_mm512_cvtt_roundpd_epu32, __m256i, __m512d, 8) 320 test_1 (_mm512_cvtt_roundps_epi32, __m512i, __m512, 8) 321 test_1 (_mm512_cvtt_roundps_epu32, __m512i, __m512, 8) 322 test_1 (_mm512_extractf32x4_ps, __m128, __m512, 1) 323 test_1 (_mm512_extractf64x4_pd, __m256d, __m512d, 1) 324 test_1 (_mm512_extracti32x4_epi32, __m128i, __m512i, 1) 325 test_1 (_mm512_extracti64x4_epi64, __m256i, __m512i, 1) 326 test_1 (_mm512_getexp_round_pd, __m512d, __m512d, 8) 327 test_1 (_mm512_getexp_round_ps, __m512, __m512, 8) 328 test_1y (_mm512_getmant_round_pd, __m512d, __m512d, 1, 1, 8) 329 test_1y (_mm512_getmant_round_ps, __m512, __m512, 1, 1, 8) 330 test_1 (_mm512_permute_pd, __m512d, __m512d, 1) 331 test_1 (_mm512_permute_ps, __m512, __m512, 1) 332 test_1 (_mm512_permutex_epi64, __m512i, __m512i, 1) 333 test_1 (_mm512_permutex_pd, __m512d, __m512d, 1) 334 test_1 (_mm512_rol_epi32, __m512i, __m512i, 1) 335 test_1 (_mm512_rol_epi64, __m512i, __m512i, 1) 336 test_1 (_mm512_ror_epi32, __m512i, __m512i, 1) 337 test_1 (_mm512_ror_epi64, __m512i, __m512i, 1) 338 test_1 (_mm512_shuffle_epi32, __m512i, __m512i, 1) 339 test_1 (_mm512_slli_epi32, __m512i, __m512i, 1) 340 test_1 (_mm512_slli_epi64, __m512i, __m512i, 1) 341 test_1 (_mm512_sqrt_round_pd, __m512d, __m512d, 9) 342 test_1 (_mm512_sqrt_round_ps, __m512, __m512, 9) 343 test_1 (_mm512_srai_epi32, __m512i, __m512i, 1) 344 test_1 (_mm512_srai_epi64, __m512i, __m512i, 1) 345 test_1 (_mm512_srli_epi32, __m512i, __m512i, 1) 346 test_1 (_mm512_srli_epi64, __m512i, __m512i, 1) 347 test_1 (_mm_cvt_roundsd_i32, int, __m128d, 9) 348 test_1 (_mm_cvt_roundsd_u32, unsigned, __m128d, 9) 349 test_1 (_mm_cvt_roundss_i32, int, __m128, 9) 350 test_1 (_mm_cvt_roundss_u32, unsigned, __m128, 9) 351 test_1 (_mm_cvtt_roundsd_i32, int, __m128d, 8) 352 test_1 (_mm_cvtt_roundsd_u32, unsigned, __m128d, 8) 353 test_1 (_mm_cvtt_roundss_i32, int, __m128, 8) 354 test_1 (_mm_cvtt_roundss_u32, unsigned, __m128, 8) 355 test_1x (_mm512_getmant_pd, __m512d, __m512d, 1, 1) 356 test_1x (_mm512_getmant_ps, __m512, __m512, 1, 1) 357 test_1x (_mm_cvt_roundi32_ss, __m128, __m128, 1, 9) 358 test_2 (_mm512_add_round_pd, __m512d, __m512d, __m512d, 9) 359 test_2 (_mm512_add_round_ps, __m512, __m512, __m512, 9) 360 test_2 (_mm512_alignr_epi32, __m512i, __m512i, __m512i, 1) 361 test_2 (_mm512_alignr_epi64, __m512i, __m512i, __m512i, 1) 362 test_2 (_mm512_cmp_epi32_mask, __mmask16, __m512i, __m512i, 1) 363 test_2 (_mm512_cmp_epi64_mask, __mmask8, __m512i, __m512i, 1) 364 test_2 (_mm512_cmp_epu32_mask, __mmask16, __m512i, __m512i, 1) 365 test_2 (_mm512_cmp_epu64_mask, __mmask8, __m512i, __m512i, 1) 366 test_2 (_mm512_cmp_pd_mask, __mmask8, __m512d, __m512d, 1) 367 test_2 (_mm512_cmp_ps_mask, __mmask16, __m512, __m512, 1) 368 test_2 (_mm512_div_round_pd, __m512d, __m512d, __m512d, 9) 369 test_2 (_mm512_div_round_ps, __m512, __m512, __m512, 9) 370 test_2 (_mm512_i32gather_epi32, __m512i, __m512i, void const *, 1) 371 test_2 (_mm512_i32gather_epi64, __m512i, __m256i, void const *, 1) 372 test_2 (_mm512_i32gather_pd, __m512d, __m256i, void const *, 1) 373 test_2 (_mm512_i32gather_ps, __m512, __m512i, void const *, 1) 374 test_2 (_mm512_i64gather_epi32, __m256i, __m512i, void const *, 1) 375 test_2 (_mm512_i64gather_epi64, __m512i, __m512i, void const *, 1) 376 test_2 (_mm512_i64gather_pd, __m512d, __m512i, void const *, 1) 377 test_2 (_mm512_i64gather_ps, __m256, __m512i, void const *, 1) 378 test_2 (_mm512_insertf32x4, __m512, __m512, __m128, 1) 379 test_2 (_mm512_insertf64x4, __m512d, __m512d, __m256d, 1) 380 test_2 (_mm512_inserti32x4, __m512i, __m512i, __m128i, 1) 381 test_2 (_mm512_inserti64x4, __m512i, __m512i, __m256i, 1) 382 test_2 (_mm512_maskz_cvt_roundepi32_ps, __m512, __mmask16, __m512i, 9) 383 test_2 (_mm512_maskz_cvt_roundepu32_ps, __m512, __mmask16, __m512i, 9) 384 test_2 (_mm512_maskz_cvt_roundpd_epi32, __m256i, __mmask8, __m512d, 9) 385 test_2 (_mm512_maskz_cvt_roundpd_epu32, __m256i, __mmask8, __m512d, 9) 386 test_2 (_mm512_maskz_cvt_roundpd_ps, __m256, __mmask8, __m512d, 9) 387 test_2 (_mm512_maskz_cvt_roundph_ps, __m512, __mmask16, __m256i, 8) 388 test_2 (_mm512_maskz_cvt_roundps_epi32, __m512i, __mmask16, __m512, 9) 389 test_2 (_mm512_maskz_cvt_roundps_epu32, __m512i, __mmask16, __m512, 9) 390 test_2 (_mm512_maskz_cvt_roundps_pd, __m512d, __mmask8, __m256, 8) 391 test_2 (_mm512_maskz_cvtps_ph, __m256i, __mmask16, __m512, 1) 392 test_2 (_mm512_maskz_cvtt_roundpd_epi32, __m256i, __mmask8, __m512d, 8) 393 test_2 (_mm512_maskz_cvtt_roundpd_epu32, __m256i, __mmask8, __m512d, 8) 394 test_2 (_mm512_maskz_cvtt_roundps_epi32, __m512i, __mmask16, __m512, 8) 395 test_2 (_mm512_maskz_cvtt_roundps_epu32, __m512i, __mmask16, __m512, 8) 396 test_2 (_mm512_maskz_extractf32x4_ps, __m128, __mmask8, __m512, 1) 397 test_2 (_mm512_maskz_extractf64x4_pd, __m256d, __mmask8, __m512d, 1) 398 test_2 (_mm512_maskz_extracti32x4_epi32, __m128i, __mmask8, __m512i, 1) 399 test_2 (_mm512_maskz_extracti64x4_epi64, __m256i, __mmask8, __m512i, 1) 400 test_2 (_mm512_maskz_getexp_round_pd, __m512d, __mmask8, __m512d, 8) 401 test_2 (_mm512_maskz_getexp_round_ps, __m512, __mmask16, __m512, 8) 402 test_2y (_mm512_maskz_getmant_round_pd, __m512d, __mmask8, __m512d, 1, 1, 8) 403 test_3y (_mm_maskz_getmant_round_sd, __m128d, __mmask8, __m128d, __m128d, 1, 1, 8) 404 test_2y (_mm512_maskz_getmant_round_ps, __m512, __mmask16, __m512, 1, 1, 8) 405 test_3y (_mm_maskz_getmant_round_ss, __m128, __mmask8, __m128, __m128, 1, 1, 8) 406 test_2 (_mm512_maskz_permute_pd, __m512d, __mmask8, __m512d, 1) 407 test_2 (_mm512_maskz_permute_ps, __m512, __mmask16, __m512, 1) 408 test_2 (_mm512_maskz_permutex_epi64, __m512i, __mmask8, __m512i, 1) 409 test_2 (_mm512_maskz_permutex_pd, __m512d, __mmask8, __m512d, 1) 410 test_2 (_mm512_maskz_rol_epi32, __m512i, __mmask16, __m512i, 1) 411 test_2 (_mm512_maskz_rol_epi64, __m512i, __mmask8, __m512i, 1) 412 test_2 (_mm512_maskz_ror_epi32, __m512i, __mmask16, __m512i, 1) 413 test_2 (_mm512_maskz_ror_epi64, __m512i, __mmask8, __m512i, 1) 414 test_2 (_mm512_maskz_shuffle_epi32, __m512i, __mmask16, __m512i, 1) 415 test_2 (_mm512_maskz_slli_epi32, __m512i, __mmask16, __m512i, 1) 416 test_2 (_mm512_maskz_slli_epi64, __m512i, __mmask8, __m512i, 1) 417 test_2 (_mm512_maskz_sqrt_round_pd, __m512d, __mmask8, __m512d, 9) 418 test_2 (_mm512_maskz_sqrt_round_ps, __m512, __mmask16, __m512, 9) 419 test_2 (_mm512_maskz_srai_epi32, __m512i, __mmask16, __m512i, 1) 420 test_2 (_mm512_maskz_srai_epi64, __m512i, __mmask8, __m512i, 1) 421 test_2 (_mm512_maskz_srli_epi32, __m512i, __mmask16, __m512i, 1) 422 test_2 (_mm512_maskz_srli_epi64, __m512i, __mmask8, __m512i, 1) 423 test_2 (_mm512_max_round_pd, __m512d, __m512d, __m512d, 8) 424 test_2 (_mm512_max_round_ps, __m512, __m512, __m512, 8) 425 test_2 (_mm512_min_round_pd, __m512d, __m512d, __m512d, 8) 426 test_2 (_mm512_min_round_ps, __m512, __m512, __m512, 8) 427 test_2 (_mm512_mul_round_pd, __m512d, __m512d, __m512d, 9) 428 test_2 (_mm512_mul_round_ps, __m512, __m512, __m512, 9) 429 test_2 (_mm512_scalef_round_pd, __m512d, __m512d, __m512d, 9) 430 test_2 (_mm512_scalef_round_ps, __m512, __m512, __m512, 9) 431 test_2 (_mm512_shuffle_f32x4, __m512, __m512, __m512, 1) 432 test_2 (_mm512_shuffle_f64x2, __m512d, __m512d, __m512d, 1) 433 test_2 (_mm512_shuffle_i32x4, __m512i, __m512i, __m512i, 1) 434 test_2 (_mm512_shuffle_i64x2, __m512i, __m512i, __m512i, 1) 435 test_2 (_mm512_shuffle_pd, __m512d, __m512d, __m512d, 1) 436 test_2 (_mm512_shuffle_ps, __m512, __m512, __m512, 1) 437 test_2 (_mm512_sub_round_pd, __m512d, __m512d, __m512d, 9) 438 test_2 (_mm512_sub_round_ps, __m512, __m512, __m512, 9) 439 test_2 (_mm_cmp_sd_mask, __mmask8, __m128d, __m128d, 1) 440 test_2 (_mm_cmp_ss_mask, __mmask8, __m128, __m128, 1) 441 test_2 (_mm512_shrdi_epi16, __m512i, __m512i, __m512i, 1) 442 test_2 (_mm512_shrdi_epi32, __m512i, __m512i, __m512i, 1) 443 test_2 (_mm512_shrdi_epi64, __m512i, __m512i, __m512i, 1) 444 test_2 (_mm256_shrdi_epi16, __m256i, __m256i, __m256i, 1) 445 test_2 (_mm256_shrdi_epi32, __m256i, __m256i, __m256i, 1) 446 test_2 (_mm256_shrdi_epi64, __m256i, __m256i, __m256i, 1) 447 test_2 (_mm_shrdi_epi16, __m128i, __m128i, __m128i, 1) 448 test_2 (_mm_shrdi_epi32, __m128i, __m128i, __m128i, 1) 449 test_2 (_mm_shrdi_epi64, __m128i, __m128i, __m128i, 1) 450 test_2 (_mm512_shldi_epi16, __m512i, __m512i, __m512i, 1) 451 test_2 (_mm512_shldi_epi32, __m512i, __m512i, __m512i, 1) 452 test_2 (_mm512_shldi_epi64, __m512i, __m512i, __m512i, 1) 453 test_2 (_mm256_shldi_epi16, __m256i, __m256i, __m256i, 1) 454 test_2 (_mm256_shldi_epi32, __m256i, __m256i, __m256i, 1) 455 test_2 (_mm256_shldi_epi64, __m256i, __m256i, __m256i, 1) 456 test_2 (_mm_shldi_epi16, __m128i, __m128i, __m128i, 1) 457 test_2 (_mm_shldi_epi32, __m128i, __m128i, __m128i, 1) 458 test_2 (_mm_shldi_epi64, __m128i, __m128i, __m128i, 1) 459 #ifdef __x86_64__ 460 test_2 (_mm_cvt_roundi64_sd, __m128d, __m128d, long long, 9) 461 test_2 (_mm_cvt_roundi64_ss, __m128, __m128, long long, 9) 462 #endif 463 test_2 (_mm_cvt_roundu32_ss, __m128, __m128, unsigned, 9) 464 #ifdef __x86_64__ 465 test_2 (_mm_cvt_roundu64_sd, __m128d, __m128d, unsigned long long, 9) 466 test_2 (_mm_cvt_roundu64_ss, __m128, __m128, unsigned long long, 9) 467 #endif 468 test_2x (_mm512_cmp_round_pd_mask, __mmask8, __m512d, __m512d, 1, 8) 469 test_2x (_mm512_cmp_round_ps_mask, __mmask16, __m512, __m512, 1, 8) 470 test_2x (_mm512_maskz_roundscale_round_pd, __m512d, __mmask8, __m512d, 1, 8) 471 test_2x (_mm512_maskz_roundscale_round_ps, __m512, __mmask16, __m512, 1, 8) 472 test_2x (_mm_cmp_round_sd_mask, __mmask8, __m128d, __m128d, 1, 8) 473 test_2x (_mm_cmp_round_ss_mask, __mmask8, __m128, __m128, 1, 8) 474 test_2x (_mm_comi_round_sd, int, __m128d, __m128d, 1, 8) 475 test_2x (_mm_comi_round_ss, int, __m128, __m128, 1, 8) 476 test_3 (_mm512_fmadd_round_pd, __m512d, __m512d, __m512d, __m512d, 9) 477 test_3 (_mm512_fmadd_round_ps, __m512, __m512, __m512, __m512, 9) 478 test_3 (_mm512_fmaddsub_round_pd, __m512d, __m512d, __m512d, __m512d, 9) 479 test_3 (_mm512_fmaddsub_round_ps, __m512, __m512, __m512, __m512, 9) 480 test_3 (_mm512_fmsub_round_pd, __m512d, __m512d, __m512d, __m512d, 9) 481 test_3 (_mm512_fmsub_round_ps, __m512, __m512, __m512, __m512, 9) 482 test_3 (_mm512_fmsubadd_round_pd, __m512d, __m512d, __m512d, __m512d, 9) 483 test_3 (_mm512_fmsubadd_round_ps, __m512, __m512, __m512, __m512, 9) 484 test_3 (_mm512_fnmadd_round_pd, __m512d, __m512d, __m512d, __m512d, 9) 485 test_3 (_mm512_fnmadd_round_ps, __m512, __m512, __m512, __m512, 9) 486 test_3 (_mm512_fnmsub_round_pd, __m512d, __m512d, __m512d, __m512d, 9) 487 test_3 (_mm512_fnmsub_round_ps, __m512, __m512, __m512, __m512, 9) 488 test_3 (_mm512_mask_cmp_epi32_mask, __mmask16, __mmask16, __m512i, __m512i, 1) 489 test_3 (_mm512_mask_cmp_epi64_mask, __mmask8, __mmask8, __m512i, __m512i, 1) 490 test_3 (_mm512_mask_cmp_epu32_mask, __mmask16, __mmask16, __m512i, __m512i, 1) 491 test_3 (_mm512_mask_cmp_epu64_mask, __mmask8, __mmask8, __m512i, __m512i, 1) 492 test_3 (_mm512_mask_cmp_pd_mask, __mmask8, __mmask8, __m512d, __m512d, 1) 493 test_3 (_mm512_mask_cmp_ps_mask, __mmask16, __mmask16, __m512, __m512, 1) 494 test_3 (_mm512_mask_cvt_roundepi32_ps, __m512, __m512, __mmask16, __m512i, 9) 495 test_3 (_mm512_mask_cvt_roundepu32_ps, __m512, __m512, __mmask16, __m512i, 9) 496 test_3 (_mm512_mask_cvt_roundpd_epi32, __m256i, __m256i, __mmask8, __m512d, 9) 497 test_3 (_mm512_mask_cvt_roundpd_epu32, __m256i, __m256i, __mmask8, __m512d, 9) 498 test_3 (_mm512_mask_cvt_roundpd_ps, __m256, __m256, __mmask8, __m512d, 9) 499 test_3 (_mm512_mask_cvt_roundph_ps, __m512, __m512, __mmask16, __m256i, 8) 500 test_3 (_mm512_mask_cvt_roundps_epi32, __m512i, __m512i, __mmask16, __m512, 9) 501 test_3 (_mm512_mask_cvt_roundps_epu32, __m512i, __m512i, __mmask16, __m512, 9) 502 test_3 (_mm512_mask_cvt_roundps_pd, __m512d, __m512d, __mmask8, __m256, 8) 503 test_3 (_mm512_mask_cvtps_ph, __m256i, __m256i, __mmask16, __m512, 1) 504 test_3 (_mm512_mask_cvtt_roundpd_epi32, __m256i, __m256i, __mmask8, __m512d, 8) 505 test_3 (_mm512_mask_cvtt_roundpd_epu32, __m256i, __m256i, __mmask8, __m512d, 8) 506 test_3 (_mm512_mask_cvtt_roundps_epi32, __m512i, __m512i, __mmask16, __m512, 8) 507 test_3 (_mm512_mask_cvtt_roundps_epu32, __m512i, __m512i, __mmask16, __m512, 8) 508 test_3 (_mm512_mask_extractf32x4_ps, __m128, __m128, __mmask8, __m512, 1) 509 test_3 (_mm512_mask_extractf64x4_pd, __m256d, __m256d, __mmask8, __m512d, 1) 510 test_3 (_mm512_mask_extracti32x4_epi32, __m128i, __m128i, __mmask8, __m512i, 1) 511 test_3 (_mm512_mask_extracti64x4_epi64, __m256i, __m256i, __mmask8, __m512i, 1) 512 test_3 (_mm512_mask_getexp_round_pd, __m512d, __m512d, __mmask8, __m512d, 8) 513 test_3 (_mm512_mask_getexp_round_ps, __m512, __m512, __mmask16, __m512, 8) 514 test_3y (_mm512_mask_getmant_round_pd, __m512d, __m512d, __mmask8, __m512d, 1, 1, 8) 515 test_4y (_mm_mask_getmant_round_sd, __m128d, __m128d, __mmask8, __m128d, __m128d, 1, 1, 8) 516 test_3y (_mm512_mask_getmant_round_ps, __m512, __m512, __mmask16, __m512, 1, 1, 8) 517 test_4y (_mm_mask_getmant_round_ss, __m128, __m128, __mmask8, __m128, __m128, 1, 1, 8) 518 test_3 (_mm512_mask_permute_pd, __m512d, __m512d, __mmask8, __m512d, 1) 519 test_3 (_mm512_mask_permute_ps, __m512, __m512, __mmask16, __m512, 1) 520 test_3 (_mm512_mask_permutex_epi64, __m512i, __m512i, __mmask8, __m512i, 1) 521 test_3 (_mm512_mask_permutex_pd, __m512d, __m512d, __mmask8, __m512d, 1) 522 test_3 (_mm512_mask_rol_epi32, __m512i, __m512i, __mmask16, __m512i, 1) 523 test_3 (_mm512_mask_rol_epi64, __m512i, __m512i, __mmask8, __m512i, 1) 524 test_3 (_mm512_mask_ror_epi32, __m512i, __m512i, __mmask16, __m512i, 1) 525 test_3 (_mm512_mask_ror_epi64, __m512i, __m512i, __mmask8, __m512i, 1) 526 test_3 (_mm512_mask_shuffle_epi32, __m512i, __m512i, __mmask16, __m512i, 1) 527 test_3 (_mm512_mask_slli_epi32, __m512i, __m512i, __mmask16, __m512i, 1) 528 test_3 (_mm512_mask_slli_epi64, __m512i, __m512i, __mmask8, __m512i, 1) 529 test_3 (_mm512_mask_sqrt_round_pd, __m512d, __m512d, __mmask8, __m512d, 9) 530 test_3 (_mm512_mask_sqrt_round_ps, __m512, __m512, __mmask16, __m512, 9) 531 test_3 (_mm512_mask_srai_epi32, __m512i, __m512i, __mmask16, __m512i, 1) 532 test_3 (_mm512_mask_srai_epi64, __m512i, __m512i, __mmask8, __m512i, 1) 533 test_3 (_mm512_mask_srli_epi32, __m512i, __m512i, __mmask16, __m512i, 1) 534 test_3 (_mm512_mask_srli_epi64, __m512i, __m512i, __mmask8, __m512i, 1) 535 test_3 (_mm512_maskz_add_round_pd, __m512d, __mmask8, __m512d, __m512d, 9) 536 test_3 (_mm512_maskz_add_round_ps, __m512, __mmask16, __m512, __m512, 9) 537 test_3 (_mm512_maskz_alignr_epi32, __m512i, __mmask16, __m512i, __m512i, 1) 538 test_3 (_mm512_maskz_alignr_epi64, __m512i, __mmask8, __m512i, __m512i, 1) 539 test_3 (_mm512_maskz_div_round_pd, __m512d, __mmask8, __m512d, __m512d, 9) 540 test_3 (_mm512_maskz_div_round_ps, __m512, __mmask16, __m512, __m512, 9) 541 test_3 (_mm512_maskz_insertf32x4, __m512, __mmask16, __m512, __m128, 1) 542 test_3 (_mm512_maskz_insertf64x4, __m512d, __mmask8, __m512d, __m256d, 1) 543 test_3 (_mm512_maskz_inserti32x4, __m512i, __mmask16, __m512i, __m128i, 1) 544 test_3 (_mm512_maskz_inserti64x4, __m512i, __mmask8, __m512i, __m256i, 1) 545 test_3 (_mm512_maskz_max_round_pd, __m512d, __mmask8, __m512d, __m512d, 8) 546 test_3 (_mm512_maskz_max_round_ps, __m512, __mmask16, __m512, __m512, 8) 547 test_3 (_mm512_maskz_min_round_pd, __m512d, __mmask8, __m512d, __m512d, 8) 548 test_3 (_mm512_maskz_min_round_ps, __m512, __mmask16, __m512, __m512, 8) 549 test_3 (_mm512_maskz_mul_round_pd, __m512d, __mmask8, __m512d, __m512d, 9) 550 test_3 (_mm512_maskz_mul_round_ps, __m512, __mmask16, __m512, __m512, 9) 551 test_3 (_mm512_maskz_scalef_round_pd, __m512d, __mmask8, __m512d, __m512d, 9) 552 test_3 (_mm512_maskz_scalef_round_ps, __m512, __mmask16, __m512, __m512, 9) 553 test_3 (_mm512_maskz_shuffle_f32x4, __m512, __mmask16, __m512, __m512, 1) 554 test_3 (_mm512_maskz_shuffle_f64x2, __m512d, __mmask8, __m512d, __m512d, 1) 555 test_3 (_mm512_maskz_shuffle_i32x4, __m512i, __mmask16, __m512i, __m512i, 1) 556 test_3 (_mm512_maskz_shuffle_i64x2, __m512i, __mmask8, __m512i, __m512i, 1) 557 test_3 (_mm512_maskz_shuffle_pd, __m512d, __mmask8, __m512d, __m512d, 1) 558 test_3 (_mm512_maskz_shuffle_ps, __m512, __mmask16, __m512, __m512, 1) 559 test_3 (_mm512_maskz_sub_round_pd, __m512d, __mmask8, __m512d, __m512d, 9) 560 test_3 (_mm512_maskz_sub_round_ps, __m512, __mmask16, __m512, __m512, 9) 561 test_3 (_mm512_ternarylogic_epi32, __m512i, __m512i, __m512i, __m512i, 1) 562 test_3 (_mm512_ternarylogic_epi64, __m512i, __m512i, __m512i, __m512i, 1) 563 test_3 (_mm_mask_cmp_sd_mask, __mmask8, __mmask8, __m128d, __m128d, 1) 564 test_3 (_mm_mask_cmp_ss_mask, __mmask8, __mmask8, __m128, __m128, 1) 565 test_3 (_mm512_maskz_shrdi_epi16, __m512i, __mmask32, __m512i, __m512i, 1) 566 test_3 (_mm512_maskz_shrdi_epi32, __m512i, __mmask16, __m512i, __m512i, 1) 567 test_3 (_mm512_maskz_shrdi_epi64, __m512i, __mmask8, __m512i, __m512i, 1) 568 test_3 (_mm256_maskz_shrdi_epi16, __m256i, __mmask16, __m256i, __m256i, 1) 569 test_3 (_mm256_maskz_shrdi_epi32, __m256i, __mmask8, __m256i, __m256i, 1) 570 test_3 (_mm256_maskz_shrdi_epi64, __m256i, __mmask8, __m256i, __m256i, 1) 571 test_3 (_mm_maskz_shrdi_epi16, __m128i, __mmask8, __m128i, __m128i, 1) 572 test_3 (_mm_maskz_shrdi_epi32, __m128i, __mmask8, __m128i, __m128i, 1) 573 test_3 (_mm_maskz_shrdi_epi64, __m128i, __mmask8, __m128i, __m128i, 1) 574 test_3 (_mm512_maskz_shldi_epi16, __m512i, __mmask32, __m512i, __m512i, 1) 575 test_3 (_mm512_maskz_shldi_epi32, __m512i, __mmask16, __m512i, __m512i, 1) 576 test_3 (_mm512_maskz_shldi_epi64, __m512i, __mmask8, __m512i, __m512i, 1) 577 test_3 (_mm256_maskz_shldi_epi16, __m256i, __mmask16, __m256i, __m256i, 1) 578 test_3 (_mm256_maskz_shldi_epi32, __m256i, __mmask8, __m256i, __m256i, 1) 579 test_3 (_mm256_maskz_shldi_epi64, __m256i, __mmask8, __m256i, __m256i, 1) 580 test_3 (_mm_maskz_shldi_epi16, __m128i, __mmask8, __m128i, __m128i, 1) 581 test_3 (_mm_maskz_shldi_epi32, __m128i, __mmask8, __m128i, __m128i, 1) 582 test_3 (_mm_maskz_shldi_epi64, __m128i, __mmask8, __m128i, __m128i, 1) 583 test_3v (_mm512_i32scatter_epi32, void *, __m512i, __m512i, 1) 584 test_3v (_mm512_i32scatter_epi64, void *, __m256i, __m512i, 1) 585 test_3v (_mm512_i32scatter_pd, void *, __m256i, __m512d, 1) 586 test_3v (_mm512_i32scatter_ps, void *, __m512i, __m512, 1) 587 test_3v (_mm512_i64scatter_epi32, void *, __m512i, __m256i, 1) 588 test_3v (_mm512_i64scatter_epi64, void *, __m512i, __m512i, 1) 589 test_3v (_mm512_i64scatter_pd, void *, __m512i, __m512d, 1) 590 test_3v (_mm512_i64scatter_ps, void *, __m512i, __m256, 1) 591 test_3x (_mm512_mask_roundscale_round_pd, __m512d, __m512d, __mmask8, __m512d, 1, 8) 592 test_3x (_mm512_mask_roundscale_round_ps, __m512, __m512, __mmask16, __m512, 1, 8) 593 test_3x (_mm512_mask_cmp_round_pd_mask, __mmask8, __mmask8, __m512d, __m512d, 1, 8) 594 test_3x (_mm512_mask_cmp_round_ps_mask, __mmask16, __mmask16, __m512, __m512, 1, 8) 595 test_3x (_mm_fixupimm_round_sd, __m128d, __m128d, __m128d, __m128i, 1, 8) 596 test_3x (_mm_fixupimm_round_ss, __m128, __m128, __m128, __m128i, 1, 8) 597 test_3x (_mm_mask_cmp_round_sd_mask, __mmask8, __mmask8, __m128d, __m128d, 1, 8) 598 test_3x (_mm_mask_cmp_round_ss_mask, __mmask8, __mmask8, __m128, __m128, 1, 8) 599 test_4 (_mm512_mask3_fmadd_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9) 600 test_4 (_mm512_mask3_fmadd_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9) 601 test_4 (_mm512_mask3_fmaddsub_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9) 602 test_4 (_mm512_mask3_fmaddsub_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9) 603 test_4 (_mm512_mask3_fmsub_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9) 604 test_4 (_mm512_mask3_fmsub_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9) 605 test_4 (_mm512_mask3_fmsubadd_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9) 606 test_4 (_mm512_mask3_fmsubadd_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9) 607 test_4 (_mm512_mask3_fnmadd_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9) 608 test_4 (_mm512_mask3_fnmadd_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9) 609 test_4 (_mm512_mask3_fnmsub_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9) 610 test_4 (_mm512_mask3_fnmsub_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9) 611 test_4 (_mm512_mask_add_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9) 612 test_4 (_mm512_mask_add_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9) 613 test_4 (_mm512_mask_alignr_epi32, __m512i, __m512i, __mmask16, __m512i, __m512i, 1) 614 test_4 (_mm512_mask_alignr_epi64, __m512i, __m512i, __mmask8, __m512i, __m512i, 1) 615 test_4 (_mm512_mask_div_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9) 616 test_4 (_mm512_mask_div_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9) 617 test_4 (_mm512_mask_fmadd_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9) 618 test_4 (_mm512_mask_fmadd_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9) 619 test_4 (_mm512_mask_fmaddsub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9) 620 test_4 (_mm512_mask_fmaddsub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9) 621 test_4 (_mm512_mask_fmsub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9) 622 test_4 (_mm512_mask_fmsub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9) 623 test_4 (_mm512_mask_fmsubadd_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9) 624 test_4 (_mm512_mask_fmsubadd_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9) 625 test_4 (_mm512_mask_fnmadd_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9) 626 test_4 (_mm512_mask_fnmadd_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9) 627 test_4 (_mm512_mask_fnmsub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9) 628 test_4 (_mm512_mask_fnmsub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9) 629 test_4 (_mm512_mask_i32gather_epi32, __m512i, __m512i, __mmask16, __m512i, void const *, 1) 630 test_4 (_mm512_mask_i32gather_epi64, __m512i, __m512i, __mmask8, __m256i, void const *, 1) 631 test_4 (_mm512_mask_i32gather_pd, __m512d, __m512d, __mmask8, __m256i, void const *, 1) 632 test_4 (_mm512_mask_i32gather_ps, __m512, __m512, __mmask16, __m512i, void const *, 1) 633 test_4 (_mm512_mask_i64gather_epi32, __m256i, __m256i, __mmask8, __m512i, void const *, 1) 634 test_4 (_mm512_mask_i64gather_epi64, __m512i, __m512i, __mmask8, __m512i, void const *, 1) 635 test_4 (_mm512_mask_i64gather_pd, __m512d, __m512d, __mmask8, __m512i, void const *, 1) 636 test_4 (_mm512_mask_i64gather_ps, __m256, __m256, __mmask8, __m512i, void const *, 1) 637 test_4 (_mm512_mask_insertf32x4, __m512, __m512, __mmask16, __m512, __m128, 1) 638 test_4 (_mm512_mask_insertf64x4, __m512d, __m512d, __mmask8, __m512d, __m256d, 1) 639 test_4 (_mm512_mask_inserti32x4, __m512i, __m512i, __mmask16, __m512i, __m128i, 1) 640 test_4 (_mm512_mask_inserti64x4, __m512i, __m512i, __mmask8, __m512i, __m256i, 1) 641 test_4 (_mm512_mask_max_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 8) 642 test_4 (_mm512_mask_max_round_ps, __m512, __m512, __mmask16, __m512, __m512, 8) 643 test_4 (_mm512_mask_min_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 8) 644 test_4 (_mm512_mask_min_round_ps, __m512, __m512, __mmask16, __m512, __m512, 8) 645 test_4 (_mm512_mask_mul_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9) 646 test_4 (_mm512_mask_mul_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9) 647 test_4 (_mm512_mask_scalef_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9) 648 test_4 (_mm512_mask_scalef_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9) 649 test_4 (_mm512_mask_shuffle_f32x4, __m512, __m512, __mmask16, __m512, __m512, 1) 650 test_4 (_mm512_mask_shuffle_f64x2, __m512d, __m512d, __mmask8, __m512d, __m512d, 1) 651 test_4 (_mm512_mask_shuffle_i32x4, __m512i, __m512i, __mmask16, __m512i, __m512i, 1) 652 test_4 (_mm512_mask_shuffle_i64x2, __m512i, __m512i, __mmask8, __m512i, __m512i, 1) 653 test_4 (_mm512_mask_shuffle_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 1) 654 test_4 (_mm512_mask_shuffle_ps, __m512, __m512, __mmask16, __m512, __m512, 1) 655 test_4 (_mm512_mask_sub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9) 656 test_4 (_mm512_mask_sub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9) 657 test_4 (_mm512_mask_ternarylogic_epi32, __m512i, __m512i, __mmask16, __m512i, __m512i, 1) 658 test_4 (_mm512_mask_ternarylogic_epi64, __m512i, __m512i, __mmask8, __m512i, __m512i, 1) 659 test_4 (_mm512_maskz_fmadd_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9) 660 test_4 (_mm512_maskz_fmadd_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9) 661 test_4 (_mm512_maskz_fmaddsub_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9) 662 test_4 (_mm512_maskz_fmaddsub_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9) 663 test_4 (_mm512_maskz_fmsub_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9) 664 test_4 (_mm512_maskz_fmsub_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9) 665 test_4 (_mm512_maskz_fmsubadd_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9) 666 test_4 (_mm512_maskz_fmsubadd_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9) 667 test_4 (_mm512_maskz_fnmadd_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9) 668 test_4 (_mm512_maskz_fnmadd_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9) 669 test_4 (_mm512_maskz_fnmsub_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9) 670 test_4 (_mm512_maskz_fnmsub_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9) 671 test_4 (_mm512_maskz_ternarylogic_epi32, __m512i, __mmask16, __m512i, __m512i, __m512i, 1) 672 test_4 (_mm512_maskz_ternarylogic_epi64, __m512i, __mmask8, __m512i, __m512i, __m512i, 1) 673 test_4 (_mm_mask_fmadd_round_sd, __m128d, __m128d, __mmask8, __m128d, __m128d, 9) 674 test_4 (_mm_mask_fmadd_round_ss, __m128, __m128, __mmask8, __m128, __m128, 9) 675 test_4 (_mm_mask3_fmadd_round_sd, __m128d, __m128d, __m128d, __m128d, __mmask8, 9) 676 test_4 (_mm_mask3_fmadd_round_ss, __m128, __m128, __m128, __m128, __mmask8, 9) 677 test_4 (_mm_maskz_fmadd_round_sd, __m128d, __mmask8, __m128d, __m128d, __m128d, 9) 678 test_4 (_mm_maskz_fmadd_round_ss, __m128, __mmask8, __m128, __m128, __m128, 9) 679 test_4 (_mm_mask_fmsub_round_sd, __m128d, __m128d, __mmask8, __m128d, __m128d, 9) 680 test_4 (_mm_mask_fmsub_round_ss, __m128, __m128, __mmask8, __m128, __m128, 9) 681 test_4 (_mm_mask3_fmsub_round_sd, __m128d, __m128d, __m128d, __m128d, __mmask8, 9) 682 test_4 (_mm_mask3_fmsub_round_ss, __m128, __m128, __m128, __m128, __mmask8, 9) 683 test_4 (_mm_maskz_fmsub_round_sd, __m128d, __mmask8, __m128d, __m128d, __m128d, 9) 684 test_4 (_mm_maskz_fmsub_round_ss, __m128, __mmask8, __m128, __m128, __m128, 9) 685 test_4 (_mm_mask_fnmadd_round_sd, __m128d, __m128d, __mmask8, __m128d, __m128d, 9) 686 test_4 (_mm_mask_fnmadd_round_ss, __m128, __m128, __mmask8, __m128, __m128, 9) 687 test_4 (_mm_mask3_fnmadd_round_sd, __m128d, __m128d, __m128d, __m128d, __mmask8, 9) 688 test_4 (_mm_mask3_fnmadd_round_ss, __m128, __m128, __m128, __m128, __mmask8, 9) 689 test_4 (_mm_maskz_fnmadd_round_sd, __m128d, __mmask8, __m128d, __m128d, __m128d, 9) 690 test_4 (_mm_maskz_fnmadd_round_ss, __m128, __mmask8, __m128, __m128, __m128, 9) 691 test_4 (_mm_mask_fnmsub_round_sd, __m128d, __m128d, __mmask8, __m128d, __m128d, 9) 692 test_4 (_mm_mask_fnmsub_round_ss, __m128, __m128, __mmask8, __m128, __m128, 9) 693 test_4 (_mm_mask3_fnmsub_round_sd, __m128d, __m128d, __m128d, __m128d, __mmask8, 9) 694 test_4 (_mm_mask3_fnmsub_round_ss, __m128, __m128, __m128, __m128, __mmask8, 9) 695 test_4 (_mm_maskz_fnmsub_round_sd, __m128d, __mmask8, __m128d, __m128d, __m128d, 9) 696 test_4 (_mm_maskz_fnmsub_round_ss, __m128, __mmask8, __m128, __m128, __m128, 9) 697 test_4 (_mm512_mask_shrdi_epi16, __m512i, __m512i, __mmask32, __m512i, __m512i, 1) 698 test_4 (_mm512_mask_shrdi_epi32, __m512i, __m512i, __mmask16, __m512i, __m512i, 1) 699 test_4 (_mm512_mask_shrdi_epi64, __m512i, __m512i, __mmask8, __m512i, __m512i, 1) 700 test_4 (_mm256_mask_shrdi_epi16, __m256i, __m256i, __mmask16, __m256i, __m256i, 1) 701 test_4 (_mm256_mask_shrdi_epi32, __m256i, __m256i, __mmask8, __m256i, __m256i, 1) 702 test_4 (_mm256_mask_shrdi_epi64, __m256i, __m256i, __mmask8, __m256i, __m256i, 1) 703 test_4 (_mm_mask_shrdi_epi16, __m128i, __m128i, __mmask8, __m128i, __m128i, 1) 704 test_4 (_mm_mask_shrdi_epi32, __m128i, __m128i, __mmask8, __m128i, __m128i, 1) 705 test_4 (_mm_mask_shrdi_epi64, __m128i, __m128i, __mmask8, __m128i, __m128i, 1) 706 test_4 (_mm512_mask_shldi_epi16, __m512i, __m512i, __mmask32, __m512i, __m512i, 1) 707 test_4 (_mm512_mask_shldi_epi32, __m512i, __m512i, __mmask16, __m512i, __m512i, 1) 708 test_4 (_mm512_mask_shldi_epi64, __m512i, __m512i, __mmask8, __m512i, __m512i, 1) 709 test_4 (_mm256_mask_shldi_epi16, __m256i, __m256i, __mmask16, __m256i, __m256i, 1) 710 test_4 (_mm256_mask_shldi_epi32, __m256i, __m256i, __mmask8, __m256i, __m256i, 1) 711 test_4 (_mm256_mask_shldi_epi64, __m256i, __m256i, __mmask8, __m256i, __m256i, 1) 712 test_4 (_mm_mask_shldi_epi16, __m128i, __m128i, __mmask8, __m128i, __m128i, 1) 713 test_4 (_mm_mask_shldi_epi32, __m128i, __m128i, __mmask8, __m128i, __m128i, 1) 714 test_4 (_mm_mask_shldi_epi64, __m128i, __m128i, __mmask8, __m128i, __m128i, 1) 715 test_4v (_mm512_mask_i32scatter_epi32, void *, __mmask16, __m512i, __m512i, 1) 716 test_4v (_mm512_mask_i32scatter_epi64, void *, __mmask8, __m256i, __m512i, 1) 717 test_4v (_mm512_mask_i32scatter_pd, void *, __mmask8, __m256i, __m512d, 1) 718 test_4v (_mm512_mask_i32scatter_ps, void *, __mmask16, __m512i, __m512, 1) 719 test_4v (_mm512_mask_i64scatter_epi32, void *, __mmask8, __m512i, __m256i, 1) 720 test_4v (_mm512_mask_i64scatter_epi64, void *, __mmask8, __m512i, __m512i, 1) 721 test_4v (_mm512_mask_i64scatter_pd, void *, __mmask8, __m512i, __m512d, 1) 722 test_4v (_mm512_mask_i64scatter_ps, void *, __mmask8, __m512i, __m256, 1) 723 test_4x (_mm512_mask_fixupimm_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512i, 1, 8) 724 test_4x (_mm512_mask_fixupimm_round_ps, __m512, __m512, __mmask16, __m512, __m512i, 1, 8) 725 test_4x (_mm512_maskz_fixupimm_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512i, 1, 8) 726 test_4x (_mm512_maskz_fixupimm_round_ps, __m512, __mmask16, __m512, __m512, __m512i, 1, 8) 727 test_4x (_mm_mask_fixupimm_round_sd, __m128d, __m128d, __mmask8, __m128d, __m128i, 1, 8) 728 test_4x (_mm_mask_fixupimm_round_ss, __m128, __m128, __mmask8, __m128, __m128i, 1, 8) 729 test_4x (_mm_maskz_fixupimm_round_sd, __m128d, __mmask8, __m128d, __m128d, __m128i, 1, 8) 730 test_4x (_mm_maskz_fixupimm_round_ss, __m128, __mmask8, __m128, __m128, __m128i, 1, 8) 731 732 /* avx512pfintrin.h */ 733 test_2vx (_mm512_prefetch_i32gather_ps, __m512i, void const *, 1, _MM_HINT_T0) 734 test_2vx (_mm512_prefetch_i32scatter_ps, void const *, __m512i, 1, _MM_HINT_T0) 735 test_2vx (_mm512_prefetch_i64gather_ps, __m512i, void const *, 1, _MM_HINT_T0) 736 test_2vx (_mm512_prefetch_i64scatter_ps, void const *, __m512i, 1, _MM_HINT_T0) 737 test_2vx (_mm512_prefetch_i32gather_pd, __m256i, void const *, 1, _MM_HINT_T0) 738 test_2vx (_mm512_prefetch_i32scatter_pd, void const *, __m256i, 1, _MM_HINT_T0) 739 test_2vx (_mm512_prefetch_i64gather_pd, __m512i, long long *, 1, _MM_HINT_T0) 740 test_2vx (_mm512_prefetch_i64scatter_pd, void const *, __m512i, 1, _MM_HINT_T0) 741 test_3vx (_mm512_mask_prefetch_i32gather_ps, __m512i, __mmask16, void const *, 1, _MM_HINT_T0) 742 test_3vx (_mm512_mask_prefetch_i32scatter_ps, void const *, __mmask16, __m512i, 1, _MM_HINT_T0) 743 test_3vx (_mm512_mask_prefetch_i64gather_ps, __m512i, __mmask8, void const *, 1, _MM_HINT_T0) 744 test_3vx (_mm512_mask_prefetch_i64scatter_ps, void const *, __mmask8, __m512i, 1, _MM_HINT_T0) 745 test_3vx (_mm512_mask_prefetch_i32gather_pd, __m256i, __mmask8, void const *, 1, _MM_HINT_T0) 746 test_3vx (_mm512_mask_prefetch_i32scatter_pd, void const *, __mmask8, __m256i, 1, _MM_HINT_T0) 747 test_3vx (_mm512_mask_prefetch_i64gather_pd, __m512i, __mmask8, long long *, 1, _MM_HINT_T0) 748 test_3vx (_mm512_mask_prefetch_i64scatter_pd, void const *, __mmask8, __m512i, 1, _MM_HINT_T0) 749 750 /* avx512erintrin.h */ 751 test_1 (_mm512_exp2a23_round_pd, __m512d, __m512d, 8) 752 test_1 (_mm512_exp2a23_round_ps, __m512, __m512, 8) 753 test_1 (_mm512_rcp28_round_pd, __m512d, __m512d, 8) 754 test_1 (_mm512_rcp28_round_ps, __m512, __m512, 8) 755 test_1 (_mm512_rsqrt28_round_pd, __m512d, __m512d, 8) 756 test_1 (_mm512_rsqrt28_round_ps, __m512, __m512, 8) 757 test_2 (_mm512_maskz_exp2a23_round_pd, __m512d, __mmask8, __m512d, 8) 758 test_2 (_mm512_maskz_exp2a23_round_ps, __m512, __mmask16, __m512, 8) 759 test_2 (_mm512_maskz_rcp28_round_pd, __m512d, __mmask8, __m512d, 8) 760 test_2 (_mm512_maskz_rcp28_round_ps, __m512, __mmask16, __m512, 8) 761 test_2 (_mm512_maskz_rsqrt28_round_pd, __m512d, __mmask8, __m512d, 8) 762 test_2 (_mm512_maskz_rsqrt28_round_ps, __m512, __mmask16, __m512, 8) 763 test_3 (_mm512_mask_exp2a23_round_pd, __m512d, __m512d, __mmask8, __m512d, 8) 764 test_3 (_mm512_mask_exp2a23_round_ps, __m512, __m512, __mmask16, __m512, 8) 765 test_3 (_mm512_mask_rcp28_round_pd, __m512d, __m512d, __mmask8, __m512d, 8) 766 test_3 (_mm512_mask_rcp28_round_ps, __m512, __m512, __mmask16, __m512, 8) 767 test_3 (_mm512_mask_rsqrt28_round_pd, __m512d, __m512d, __mmask8, __m512d, 8) 768 test_3 (_mm512_mask_rsqrt28_round_ps, __m512, __m512, __mmask16, __m512, 8) 769 test_2 (_mm_rcp28_round_sd, __m128d, __m128d, __m128d, 8) 770 test_2 (_mm_rcp28_round_ss, __m128, __m128, __m128, 8) 771 test_2 (_mm_rsqrt28_round_sd, __m128d, __m128d, __m128d, 8) 772 test_2 (_mm_rsqrt28_round_ss, __m128, __m128, __m128, 8) 773 774 /* shaintrin.h */ 775 test_2 (_mm_sha1rnds4_epu32, __m128i, __m128i, __m128i, 1) 776 777 /* gfniintrin.h */ 778 test_2 (_mm_gf2p8affineinv_epi64_epi8, __m128i, __m128i, __m128i, 1) 779 test_2 (_mm256_gf2p8affineinv_epi64_epi8, __m256i, __m256i, __m256i, 1) 780 test_2 (_mm512_gf2p8affineinv_epi64_epi8, __m512i, __m512i, __m512i, 1) 781 782 /* wmmintrin.h (AES/PCLMUL). */ 783 #ifdef DIFFERENT_PRAGMAS 784 #pragma GCC target ("aes,pclmul") 785 #endif 786 #include <wmmintrin.h> 787 test_1 (_mm_aeskeygenassist_si128, __m128i, __m128i, 1) 788 test_2 (_mm_clmulepi64_si128, __m128i, __m128i, __m128i, 1) 789 790 /* popcnintrin.h (POPCNT). */ 791 #ifdef DIFFERENT_PRAGMAS 792 #pragma GCC target ("popcnt") 793 #endif 794 #include <popcntintrin.h> 795 796 /* x86intrin.h (FMA4/XOP/LWP/BMI/BMI2/TBM/LZCNT/FMA). */ 797 #ifdef DIFFERENT_PRAGMAS 798 #pragma GCC target ("fma4,xop,lwp,bmi,bmi2,tbm,lzcnt,fma,rdseed,prfchw,adx,fxsr,xsaveopt,xsavec,xsaves,clflushopt,clwb,pku,sgx,rdpid") 799 #endif 800 #include <x86intrin.h> 801 /* xopintrin.h */ 802 test_1 ( _mm_roti_epi8, __m128i, __m128i, 1) 803 test_1 ( _mm_roti_epi16, __m128i, __m128i, 1) 804 test_1 ( _mm_roti_epi32, __m128i, __m128i, 1) 805 test_1 ( _mm_roti_epi64, __m128i, __m128i, 1) 806 test_3 (_mm_permute2_pd, __m128d, __m128d, __m128d, __m128d, 1) 807 test_3 (_mm256_permute2_pd, __m256d, __m256d, __m256d, __m256d, 1) 808 test_3 (_mm_permute2_ps, __m128, __m128, __m128, __m128, 1) 809 test_3 (_mm256_permute2_ps, __m256, __m256, __m256, __m256, 1) 810 811 /* lwpintrin.h */ 812 test_2 ( __lwpval32, void, unsigned int, unsigned int, 1) 813 test_2 ( __lwpins32, unsigned char, unsigned int, unsigned int, 1) 814 #ifdef __x86_64__ 815 test_2 ( __lwpval64, void, unsigned long long, unsigned int, 1) 816 test_2 ( __lwpins64, unsigned char, unsigned long long, unsigned int, 1) 817 #endif 818 819 /* tbmintrin.h */ 820 test_1 ( __bextri_u32, unsigned int, unsigned int, 1) 821 #ifdef __x86_64__ 822 test_1 ( __bextri_u64, unsigned long long, unsigned long long, 1) 823 #endif 824