gcc/testsuite/gcc.target/i386/sse-22.c

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