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