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
   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_2(func, type, op1_type, op2_type, imm)                     \
  34   type _CONCAT(_,func) (op1_type A, op2_type B, int const I)            \
  35   { return func (A, B, imm); }
  36
  37 #define test_2x(func, type, op1_type, op2_type, imm1, imm2)             \
  38   type _CONCAT(_,func) (op1_type A, op2_type B, int const I, int const L) \
  39   { return func (A, B, imm1, imm2); }
  40
  41 #define test_3(func, type, op1_type, op2_type, op3_type, imm)           \
  42   type _CONCAT(_,func) (op1_type A, op2_type B,                         \
  43                         op3_type C, int const I)                        \
  44   { return func (A, B, C, imm); }
  45
  46 #define test_4(func, type, op1_type, op2_type, op3_type, op4_type, imm) \
  47   type _CONCAT(_,func) (op1_type A, op2_type B,                         \
  48                         op3_type C, op4_type D, int const I)            \
  49   { return func (A, B, C, D, imm); }
  50
  51
  52 #ifndef DIFFERENT_PRAGMAS
  53 #pragma GCC target ("sse4a,3dnow,avx,avx2,fma4,xop,aes,pclmul,popcnt,abm,lzcnt,bmi,bmi2,tbm,lwp,fsgsbase,rdrnd,f16c,rtm,rdseed,prfchw")
  54 #endif
  55
  56 /* Following intrinsics require immediate arguments.  They
  57    are defined as macros for non-optimized compilations. */
  58
  59 /* mmintrin.h (MMX).  */
  60 #ifdef DIFFERENT_PRAGMAS
  61 #pragma GCC target ("mmx")
  62 #endif
  63 #include <mmintrin.h>
  64
  65 /* mm3dnow.h (3DNOW).  */
  66 #ifdef DIFFERENT_PRAGMAS
  67 #pragma GCC target ("3dnow")
  68 #endif
  69 #include <mm3dnow.h>
  70
  71 /* xmmintrin.h (SSE).  */
  72 #ifdef DIFFERENT_PRAGMAS
  73 #pragma GCC target ("sse")
  74 #endif
  75 #include <xmmintrin.h>
  76 test_2 (_mm_shuffle_ps, __m128, __m128, __m128, 1)
  77 test_1 (_mm_extract_pi16, int, __m64, 1)
  78 test_1 (_m_pextrw, int, __m64, 1)
  79 test_2 (_mm_insert_pi16, __m64, __m64, int, 1)
  80 test_2 (_m_pinsrw, __m64, __m64, int, 1)
  81 test_1 (_mm_shuffle_pi16, __m64, __m64, 1)
  82 test_1 (_m_pshufw, __m64, __m64, 1)
  83 test_1 (_mm_prefetch, void, void *, _MM_HINT_NTA)
  84
  85 /* emmintrin.h (SSE2).  */
  86 #ifdef DIFFERENT_PRAGMAS
  87 #pragma GCC target ("sse2")
  88 #endif
  89 #include <emmintrin.h>
  90 test_2 (_mm_shuffle_pd, __m128d, __m128d, __m128d, 1)
  91 test_1 (_mm_srli_si128, __m128i, __m128i, 1)
  92 test_1 (_mm_slli_si128, __m128i, __m128i, 1)
  93 test_1 (_mm_extract_epi16, int, __m128i, 1)
  94 test_2 (_mm_insert_epi16, __m128i, __m128i, int, 1)
  95 test_1 (_mm_shufflehi_epi16, __m128i, __m128i, 1)
  96 test_1 (_mm_shufflelo_epi16, __m128i, __m128i, 1)
  97 test_1 (_mm_shuffle_epi32, __m128i, __m128i, 1)
  98
  99 /* pmmintrin.h (SSE3).  */
 100 #ifdef DIFFERENT_PRAGMAS
 101 #pragma GCC target ("sse3")
 102 #endif
 103 #include <pmmintrin.h>
 104
 105 /* tmmintrin.h (SSSE3).  */
 106 #ifdef DIFFERENT_PRAGMAS
 107 #pragma GCC target ("ssse3")
 108 #endif
 109 #include <tmmintrin.h>
 110 test_2 (_mm_alignr_epi8, __m128i, __m128i, __m128i, 1)
 111 test_2 (_mm_alignr_pi8, __m64, __m64, __m64, 1)
 112
 113 /* ammintrin.h (SSE4A).  */
 114 #ifdef DIFFERENT_PRAGMAS
 115 #pragma GCC target ("sse4a")
 116 #endif
 117 #include <ammintrin.h>
 118 test_1x (_mm_extracti_si64, __m128i, __m128i, 1, 1)
 119 test_2x (_mm_inserti_si64, __m128i, __m128i, __m128i, 1, 1)
 120
 121 /* Note, nmmintrin.h includes smmintrin.h, and smmintrin.h
 122    checks for the #ifdef.  So just set the option to SSE4.2.  */
 123 #ifdef DIFFERENT_PRAGMAS
 124 #pragma GCC target ("sse4.2")
 125 #endif
 126 #include <nmmintrin.h>
 127 /* smmintrin.h (SSE4.2).  */
 128 test_1 (_mm_round_pd, __m128d, __m128d, 1)
 129 test_1 (_mm_round_ps, __m128, __m128, 1)
 130 test_2 (_mm_round_sd, __m128d, __m128d, __m128d, 1)
 131 test_2 (_mm_round_ss, __m128, __m128, __m128, 1)
 132
 133 test_2 (_mm_blend_epi16, __m128i, __m128i, __m128i, 1)
 134 test_2 (_mm_blend_ps, __m128, __m128, __m128, 1)
 135 test_2 (_mm_blend_pd, __m128d, __m128d, __m128d, 1)
 136 test_2 (_mm_dp_ps, __m128, __m128, __m128, 1)
 137 test_2 (_mm_dp_pd, __m128d, __m128d, __m128d, 1)
 138 test_2 (_mm_insert_ps, __m128, __m128, __m128, 1)
 139 test_1 (_mm_extract_ps, int, __m128, 1)
 140 test_2 (_mm_insert_epi8, __m128i, __m128i, int, 1)
 141 test_2 (_mm_insert_epi32, __m128i, __m128i, int, 1)
 142 #ifdef __x86_64__
 143 test_2 (_mm_insert_epi64, __m128i, __m128i, long long, 1)
 144 #endif
 145 test_1 (_mm_extract_epi8, int, __m128i, 1)
 146 test_1 (_mm_extract_epi32, int, __m128i, 1)
 147 #ifdef __x86_64__
 148 test_1 (_mm_extract_epi64, long long, __m128i, 1)
 149 #endif
 150 test_2 (_mm_mpsadbw_epu8, __m128i, __m128i, __m128i, 1)
 151 test_2 (_mm_cmpistrm, __m128i, __m128i, __m128i, 1)
 152 test_2 (_mm_cmpistri, int, __m128i, __m128i, 1)
 153 test_4 (_mm_cmpestrm, __m128i, __m128i, int, __m128i, int, 1)
 154 test_4 (_mm_cmpestri, int, __m128i, int, __m128i, int, 1)
 155 test_2 (_mm_cmpistra, int, __m128i, __m128i, 1)
 156 test_2 (_mm_cmpistrc, int, __m128i, __m128i, 1)
 157 test_2 (_mm_cmpistro, int, __m128i, __m128i, 1)
 158 test_2 (_mm_cmpistrs, int, __m128i, __m128i, 1)
 159 test_2 (_mm_cmpistrz, int, __m128i, __m128i, 1)
 160 test_4 (_mm_cmpestra, int, __m128i, int, __m128i, int, 1)
 161 test_4 (_mm_cmpestrc, int, __m128i, int, __m128i, int, 1)
 162 test_4 (_mm_cmpestro, int, __m128i, int, __m128i, int, 1)
 163 test_4 (_mm_cmpestrs, int, __m128i, int, __m128i, int, 1)
 164 test_4 (_mm_cmpestrz, int, __m128i, int, __m128i, int, 1)
 165
 166 /* immintrin.h (AVX/AVX2/RDRND/FSGSBASE/F16C/RTM) */
 167 #ifdef DIFFERENT_PRAGMAS
 168 #pragma GCC target ("avx,avx2,rdrnd,fsgsbase,f16c,rtm")
 169 #endif
 170 #include <immintrin.h>
 171 test_1 (_cvtss_sh, unsigned short, float, 1)
 172 test_1 (_mm_cvtps_ph, __m128i, __m128, 1)
 173 test_1 (_mm256_cvtps_ph, __m128i, __m256, 1)
 174
 175 /* avxintrin.h */
 176 test_2 (_mm256_blend_pd, __m256d, __m256d, __m256d, 1)
 177 test_2 (_mm256_blend_ps, __m256, __m256, __m256, 1)
 178 test_2 (_mm256_dp_ps, __m256, __m256, __m256, 1)
 179 test_2 (_mm256_shuffle_pd, __m256d, __m256d, __m256d, 1)
 180 test_2 (_mm256_shuffle_ps, __m256, __m256, __m256, 1)
 181 test_2 (_mm_cmp_sd, __m128d, __m128d, __m128d, 1)
 182 test_2 (_mm_cmp_ss, __m128, __m128, __m128, 1)
 183 test_2 (_mm_cmp_pd, __m128d, __m128d, __m128d, 1)
 184 test_2 (_mm_cmp_ps, __m128, __m128, __m128, 1)
 185 test_2 (_mm256_cmp_pd, __m256d, __m256d, __m256d, 1)
 186 test_2 (_mm256_cmp_ps, __m256, __m256, __m256, 1)
 187 test_1 (_mm256_extractf128_pd, __m128d, __m256d, 1)
 188 test_1 (_mm256_extractf128_ps, __m128, __m256, 1)
 189 test_1 (_mm256_extractf128_si256, __m128i, __m256i, 1)
 190 test_1 (_mm256_extract_epi8, int, __m256i, 20)
 191 test_1 (_mm256_extract_epi16, int, __m256i, 13)
 192 test_1 (_mm256_extract_epi32, int, __m256i, 6)
 193 #ifdef __x86_64__
 194 test_1 (_mm256_extract_epi64, long long, __m256i, 2)
 195 #endif
 196 test_1 (_mm_permute_pd, __m128d, __m128d, 1)
 197 test_1 (_mm256_permute_pd, __m256d, __m256d, 1)
 198 test_1 (_mm_permute_ps, __m128, __m128, 1)
 199 test_1 (_mm256_permute_ps, __m256, __m256, 1)
 200 test_2 (_mm256_permute2f128_pd, __m256d, __m256d, __m256d, 1)
 201 test_2 (_mm256_permute2f128_ps, __m256, __m256, __m256, 1)
 202 test_2 (_mm256_permute2f128_si256, __m256i, __m256i, __m256i, 1)
 203 test_2 (_mm256_insertf128_pd, __m256d, __m256d, __m128d, 1)
 204 test_2 (_mm256_insertf128_ps, __m256, __m256, __m128, 1)
 205 test_2 (_mm256_insertf128_si256, __m256i, __m256i, __m128i, 1)
 206 test_2 (_mm256_insert_epi8, __m256i, __m256i, int, 30)
 207 test_2 (_mm256_insert_epi16, __m256i, __m256i, int, 7)
 208 test_2 (_mm256_insert_epi32, __m256i, __m256i, int, 3)
 209 #ifdef __x86_64__
 210 test_2 (_mm256_insert_epi64, __m256i, __m256i, long long, 1)
 211 #endif
 212 test_1 (_mm256_round_pd, __m256d, __m256d, 1)
 213 test_1 (_mm256_round_ps, __m256, __m256, 1)
 214
 215 /* avx2intrin.h */
 216 test_2 ( _mm256_mpsadbw_epu8, __m256i, __m256i, __m256i, 1)
 217 test_2 ( _mm256_alignr_epi8, __m256i, __m256i, __m256i, 1)
 218 test_2 ( _mm256_blend_epi16, __m256i, __m256i, __m256i, 1)
 219 test_1 ( _mm256_shuffle_epi32, __m256i, __m256i, 1)
 220 test_1 ( _mm256_shufflehi_epi16, __m256i, __m256i, 1)
 221 test_1 ( _mm256_shufflelo_epi16, __m256i, __m256i, 1)
 222 test_1 ( _mm256_slli_si256, __m256i, __m256i, 8)
 223 test_1 ( _mm256_srli_si256, __m256i, __m256i, 8)
 224 test_2 ( _mm_blend_epi32, __m128i, __m128i, __m128i, 1)
 225 test_2 ( _mm256_blend_epi32, __m256i, __m256i, __m256, 1)
 226 test_1 ( _mm256_permute4x64_pd, __m256d, __m256d, 1)
 227 test_1 ( _mm256_permute4x64_epi64, __m256i, __m256i, 1)
 228 test_2 ( _mm256_permute2x128_si256, __m256i, __m256i, __m256i, 1)
 229 test_1 ( _mm256_extracti128_si256, __m128i, __m256i, 1)
 230 test_2 ( _mm256_inserti128_si256, __m256i, __m256i, __m128i, 1)
 231 test_2 ( _mm_i32gather_pd, __m128d, double const *, __m128i, 1)
 232 test_2 ( _mm256_i32gather_pd, __m256d, double const *, __m128i, 1)
 233 test_2 ( _mm_i64gather_pd, __m128d, double const *, __m128i, 1)
 234 test_2 ( _mm256_i64gather_pd, __m256d, double const *, __m256i, 1)
 235 test_2 ( _mm_i32gather_ps, __m128, float const *, __m128i, 1)
 236 test_2 ( _mm256_i32gather_ps, __m256, float const *, __m256i, 1)
 237 test_2 ( _mm_i64gather_ps, __m128, float const *, __m128i, 1)
 238 test_2 ( _mm256_i64gather_ps, __m128, float const *, __m256i, 1)
 239 test_2 ( _mm_i32gather_epi64, __m128i, long long int const *, __m128i, 1)
 240 test_2 ( _mm256_i32gather_epi64, __m256i, long long int const *, __m128i, 1)
 241 test_2 ( _mm_i64gather_epi64, __m128i, long long int const *, __m128i, 1)
 242 test_2 ( _mm256_i64gather_epi64,  __m256i, long long int const *, __m256i, 1)
 243 test_2 ( _mm_i32gather_epi32, __m128i, int const *, __m128i, 1)
 244 test_2 ( _mm256_i32gather_epi32, __m256i, int const *, __m256i, 1)
 245 test_2 ( _mm_i64gather_epi32, __m128i, int const *, __m128i, 1)
 246 test_2 ( _mm256_i64gather_epi32, __m128i, int const *, __m256i, 1)
 247
 248 /* rtmintrin.h */
 249 test_0 ( _xabort, void, 1)
 250
 251 /* wmmintrin.h (AES/PCLMUL).  */
 252 #ifdef DIFFERENT_PRAGMAS
 253 #pragma GCC target ("aes,pclmul")
 254 #endif
 255 #include <wmmintrin.h>
 256 test_1 (_mm_aeskeygenassist_si128, __m128i, __m128i, 1)
 257 test_2 (_mm_clmulepi64_si128, __m128i, __m128i, __m128i, 1)
 258
 259 /* popcnintrin.h (POPCNT).  */
 260 #ifdef DIFFERENT_PRAGMAS
 261 #pragma GCC target ("popcnt")
 262 #endif
 263 #include <popcntintrin.h>
 264
 265 /* x86intrin.h (FMA4/XOP/LWP/BMI/BMI2/TBM/LZCNT/FMA). */
 266 #ifdef DIFFERENT_PRAGMAS
 267 #pragma GCC target ("fma4,xop,lwp,bmi,bmi2,tbm,lzcnt,fma,rdseed,prfchw")
 268 #endif
 269 #include <x86intrin.h>
 270 /* xopintrin.h */
 271 test_1 ( _mm_roti_epi8, __m128i, __m128i, 1)
 272 test_1 ( _mm_roti_epi16, __m128i, __m128i, 1)
 273 test_1 ( _mm_roti_epi32, __m128i, __m128i, 1)
 274 test_1 ( _mm_roti_epi64, __m128i, __m128i, 1)
 275 test_3 (_mm_permute2_pd, __m128d, __m128d, __m128d, __m128d, 1)
 276 test_3 (_mm256_permute2_pd, __m256d, __m256d, __m256d, __m256d, 1)
 277 test_3 (_mm_permute2_ps, __m128, __m128, __m128, __m128, 1)
 278 test_3 (_mm256_permute2_ps, __m256, __m256, __m256, __m256, 1)
 279
 280 /* lwpintrin.h */
 281 test_2 ( __lwpval32, void, unsigned int, unsigned int, 1)
 282 test_2 ( __lwpins32, unsigned char, unsigned int, unsigned int, 1)
 283 #ifdef __x86_64__
 284 test_2 ( __lwpval64, void, unsigned long long, unsigned int, 1)
 285 test_2 ( __lwpins64, unsigned char, unsigned long long, unsigned int, 1)
 286 #endif
 287
 288 /* tbmintrin.h */
 289 test_1 ( __bextri_u32, unsigned int, unsigned int, 1)
 290 #ifdef __x86_64__
 291 test_1 ( __bextri_u64, unsigned long long, unsigned long long, 1)
 292 #endif