gcc/testsuite/gcc.target/aarch64/sve2/acle/asm/ldnt1_gather_u64.c

   1 /* { dg-final { check-function-bodies "**" "" "-DCHECK_ASM" } } */
   2
   3 #include "test_sve_acle.h"
   4
   5 /*
   6 ** ldnt1_gather_u64_tied1:
   7 **      ldnt1d  z0\.d, p0/z, \[z0\.d\]
   8 **      ret
   9 */
  10 TEST_LOAD_GATHER_ZS (ldnt1_gather_u64_tied1, svuint64_t, svuint64_t,
  11                      z0_res = svldnt1_gather_u64base_u64 (p0, z0),
  12                      z0_res = svldnt1_gather_u64 (p0, z0))
  13
  14 /*
  15 ** ldnt1_gather_u64_untied:
  16 **      ldnt1d  z0\.d, p0/z, \[z1\.d\]
  17 **      ret
  18 */
  19 TEST_LOAD_GATHER_ZS (ldnt1_gather_u64_untied, svuint64_t, svuint64_t,
  20                      z0_res = svldnt1_gather_u64base_u64 (p0, z1),
  21                      z0_res = svldnt1_gather_u64 (p0, z1))
  22
  23 /*
  24 ** ldnt1_gather_x0_u64_offset:
  25 **      ldnt1d  z0\.d, p0/z, \[z0\.d, x0\]
  26 **      ret
  27 */
  28 TEST_LOAD_GATHER_ZS (ldnt1_gather_x0_u64_offset, svuint64_t, svuint64_t,
  29                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, x0),
  30                      z0_res = svldnt1_gather_offset_u64 (p0, z0, x0))
  31
  32 /*
  33 ** ldnt1_gather_m8_u64_offset:
  34 **      mov     (x[0-9]+), #?-8
  35 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
  36 **      ret
  37 */
  38 TEST_LOAD_GATHER_ZS (ldnt1_gather_m8_u64_offset, svuint64_t, svuint64_t,
  39                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, -8),
  40                      z0_res = svldnt1_gather_offset_u64 (p0, z0, -8))
  41
  42 /*
  43 ** ldnt1_gather_0_u64_offset:
  44 **      ldnt1d  z0\.d, p0/z, \[z0\.d\]
  45 **      ret
  46 */
  47 TEST_LOAD_GATHER_ZS (ldnt1_gather_0_u64_offset, svuint64_t, svuint64_t,
  48                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, 0),
  49                      z0_res = svldnt1_gather_offset_u64 (p0, z0, 0))
  50
  51 /*
  52 ** ldnt1_gather_9_u64_offset:
  53 **      mov     (x[0-9]+), #?9
  54 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
  55 **      ret
  56 */
  57 TEST_LOAD_GATHER_ZS (ldnt1_gather_9_u64_offset, svuint64_t, svuint64_t,
  58                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, 9),
  59                      z0_res = svldnt1_gather_offset_u64 (p0, z0, 9))
  60
  61 /*
  62 ** ldnt1_gather_10_u64_offset:
  63 **      mov     (x[0-9]+), #?10
  64 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
  65 **      ret
  66 */
  67 TEST_LOAD_GATHER_ZS (ldnt1_gather_10_u64_offset, svuint64_t, svuint64_t,
  68                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, 10),
  69                      z0_res = svldnt1_gather_offset_u64 (p0, z0, 10))
  70
  71 /*
  72 ** ldnt1_gather_11_u64_offset:
  73 **      mov     (x[0-9]+), #?11
  74 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
  75 **      ret
  76 */
  77 TEST_LOAD_GATHER_ZS (ldnt1_gather_11_u64_offset, svuint64_t, svuint64_t,
  78                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, 11),
  79                      z0_res = svldnt1_gather_offset_u64 (p0, z0, 11))
  80
  81 /*
  82 ** ldnt1_gather_12_u64_offset:
  83 **      mov     (x[0-9]+), #?12
  84 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
  85 **      ret
  86 */
  87 TEST_LOAD_GATHER_ZS (ldnt1_gather_12_u64_offset, svuint64_t, svuint64_t,
  88                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, 12),
  89                      z0_res = svldnt1_gather_offset_u64 (p0, z0, 12))
  90
  91 /*
  92 ** ldnt1_gather_13_u64_offset:
  93 **      mov     (x[0-9]+), #?13
  94 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
  95 **      ret
  96 */
  97 TEST_LOAD_GATHER_ZS (ldnt1_gather_13_u64_offset, svuint64_t, svuint64_t,
  98                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, 13),
  99                      z0_res = svldnt1_gather_offset_u64 (p0, z0, 13))
 100
 101 /*
 102 ** ldnt1_gather_14_u64_offset:
 103 **      mov     (x[0-9]+), #?14
 104 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
 105 **      ret
 106 */
 107 TEST_LOAD_GATHER_ZS (ldnt1_gather_14_u64_offset, svuint64_t, svuint64_t,
 108                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, 14),
 109                      z0_res = svldnt1_gather_offset_u64 (p0, z0, 14))
 110
 111 /*
 112 ** ldnt1_gather_15_u64_offset:
 113 **      mov     (x[0-9]+), #?15
 114 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
 115 **      ret
 116 */
 117 TEST_LOAD_GATHER_ZS (ldnt1_gather_15_u64_offset, svuint64_t, svuint64_t,
 118                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, 15),
 119                      z0_res = svldnt1_gather_offset_u64 (p0, z0, 15))
 120
 121 /*
 122 ** ldnt1_gather_16_u64_offset:
 123 **      mov     (x[0-9]+), #?16
 124 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
 125 **      ret
 126 */
 127 TEST_LOAD_GATHER_ZS (ldnt1_gather_16_u64_offset, svuint64_t, svuint64_t,
 128                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, 16),
 129                      z0_res = svldnt1_gather_offset_u64 (p0, z0, 16))
 130
 131 /*
 132 ** ldnt1_gather_248_u64_offset:
 133 **      mov     (x[0-9]+), #?248
 134 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
 135 **      ret
 136 */
 137 TEST_LOAD_GATHER_ZS (ldnt1_gather_248_u64_offset, svuint64_t, svuint64_t,
 138                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, 248),
 139                      z0_res = svldnt1_gather_offset_u64 (p0, z0, 248))
 140
 141 /*
 142 ** ldnt1_gather_256_u64_offset:
 143 **      mov     (x[0-9]+), #?256
 144 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
 145 **      ret
 146 */
 147 TEST_LOAD_GATHER_ZS (ldnt1_gather_256_u64_offset, svuint64_t, svuint64_t,
 148                      z0_res = svldnt1_gather_u64base_offset_u64 (p0, z0, 256),
 149                      z0_res = svldnt1_gather_offset_u64 (p0, z0, 256))
 150
 151 /*
 152 ** ldnt1_gather_x0_u64_index:
 153 **      lsl     (x[0-9]+), x0, #?3
 154 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
 155 **      ret
 156 */
 157 TEST_LOAD_GATHER_ZS (ldnt1_gather_x0_u64_index, svuint64_t, svuint64_t,
 158                      z0_res = svldnt1_gather_u64base_index_u64 (p0, z0, x0),
 159                      z0_res = svldnt1_gather_index_u64 (p0, z0, x0))
 160
 161 /*
 162 ** ldnt1_gather_m1_u64_index:
 163 **      mov     (x[0-9]+), #?-8
 164 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
 165 **      ret
 166 */
 167 TEST_LOAD_GATHER_ZS (ldnt1_gather_m1_u64_index, svuint64_t, svuint64_t,
 168                      z0_res = svldnt1_gather_u64base_index_u64 (p0, z0, -1),
 169                      z0_res = svldnt1_gather_index_u64 (p0, z0, -1))
 170
 171 /*
 172 ** ldnt1_gather_0_u64_index:
 173 **      ldnt1d  z0\.d, p0/z, \[z0\.d\]
 174 **      ret
 175 */
 176 TEST_LOAD_GATHER_ZS (ldnt1_gather_0_u64_index, svuint64_t, svuint64_t,
 177                      z0_res = svldnt1_gather_u64base_index_u64 (p0, z0, 0),
 178                      z0_res = svldnt1_gather_index_u64 (p0, z0, 0))
 179
 180 /*
 181 ** ldnt1_gather_5_u64_index:
 182 **      mov     (x[0-9]+), #?40
 183 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
 184 **      ret
 185 */
 186 TEST_LOAD_GATHER_ZS (ldnt1_gather_5_u64_index, svuint64_t, svuint64_t,
 187                      z0_res = svldnt1_gather_u64base_index_u64 (p0, z0, 5),
 188                      z0_res = svldnt1_gather_index_u64 (p0, z0, 5))
 189
 190 /*
 191 ** ldnt1_gather_31_u64_index:
 192 **      mov     (x[0-9]+), #?248
 193 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
 194 **      ret
 195 */
 196 TEST_LOAD_GATHER_ZS (ldnt1_gather_31_u64_index, svuint64_t, svuint64_t,
 197                      z0_res = svldnt1_gather_u64base_index_u64 (p0, z0, 31),
 198                      z0_res = svldnt1_gather_index_u64 (p0, z0, 31))
 199
 200 /*
 201 ** ldnt1_gather_32_u64_index:
 202 **      mov     (x[0-9]+), #?256
 203 **      ldnt1d  z0\.d, p0/z, \[z0\.d, \1\]
 204 **      ret
 205 */
 206 TEST_LOAD_GATHER_ZS (ldnt1_gather_32_u64_index, svuint64_t, svuint64_t,
 207                      z0_res = svldnt1_gather_u64base_index_u64 (p0, z0, 32),
 208                      z0_res = svldnt1_gather_index_u64 (p0, z0, 32))
 209
 210 /*
 211 ** ldnt1_gather_x0_u64_s64offset:
 212 **      ldnt1d  z0\.d, p0/z, \[z0\.d, x0\]
 213 **      ret
 214 */
 215 TEST_LOAD_GATHER_SZ (ldnt1_gather_x0_u64_s64offset, svuint64_t, uint64_t, svint64_t,
 216                      z0_res = svldnt1_gather_s64offset_u64 (p0, x0, z0),
 217                      z0_res = svldnt1_gather_offset (p0, x0, z0))
 218
 219 /*
 220 ** ldnt1_gather_tied1_u64_s64offset:
 221 **      ldnt1d  z0\.d, p0/z, \[z0\.d, x0\]
 222 **      ret
 223 */
 224 TEST_LOAD_GATHER_SZ (ldnt1_gather_tied1_u64_s64offset, svuint64_t, uint64_t, svint64_t,
 225                      z0_res = svldnt1_gather_s64offset_u64 (p0, x0, z0),
 226                      z0_res = svldnt1_gather_offset (p0, x0, z0))
 227
 228 /*
 229 ** ldnt1_gather_untied_u64_s64offset:
 230 **      ldnt1d  z0\.d, p0/z, \[z1\.d, x0\]
 231 **      ret
 232 */
 233 TEST_LOAD_GATHER_SZ (ldnt1_gather_untied_u64_s64offset, svuint64_t, uint64_t, svint64_t,
 234                      z0_res = svldnt1_gather_s64offset_u64 (p0, x0, z1),
 235                      z0_res = svldnt1_gather_offset (p0, x0, z1))
 236
 237 /*
 238 ** ldnt1_gather_x0_u64_u64offset:
 239 **      ldnt1d  z0\.d, p0/z, \[z0\.d, x0\]
 240 **      ret
 241 */
 242 TEST_LOAD_GATHER_SZ (ldnt1_gather_x0_u64_u64offset, svuint64_t, uint64_t, svuint64_t,
 243                      z0_res = svldnt1_gather_u64offset_u64 (p0, x0, z0),
 244                      z0_res = svldnt1_gather_offset (p0, x0, z0))
 245
 246 /*
 247 ** ldnt1_gather_tied1_u64_u64offset:
 248 **      ldnt1d  z0\.d, p0/z, \[z0\.d, x0\]
 249 **      ret
 250 */
 251 TEST_LOAD_GATHER_SZ (ldnt1_gather_tied1_u64_u64offset, svuint64_t, uint64_t, svuint64_t,
 252                      z0_res = svldnt1_gather_u64offset_u64 (p0, x0, z0),
 253                      z0_res = svldnt1_gather_offset (p0, x0, z0))
 254
 255 /*
 256 ** ldnt1_gather_untied_u64_u64offset:
 257 **      ldnt1d  z0\.d, p0/z, \[z1\.d, x0\]
 258 **      ret
 259 */
 260 TEST_LOAD_GATHER_SZ (ldnt1_gather_untied_u64_u64offset, svuint64_t, uint64_t, svuint64_t,
 261                      z0_res = svldnt1_gather_u64offset_u64 (p0, x0, z1),
 262                      z0_res = svldnt1_gather_offset (p0, x0, z1))
 263
 264 /*
 265 ** ldnt1_gather_x0_u64_s64index:
 266 **      lsl     (z[0-9]+\.d), z0\.d, #3
 267 **      ldnt1d  z0\.d, p0/z, \[\1, x0\]
 268 **      ret
 269 */
 270 TEST_LOAD_GATHER_SZ (ldnt1_gather_x0_u64_s64index, svuint64_t, uint64_t, svint64_t,
 271                      z0_res = svldnt1_gather_s64index_u64 (p0, x0, z0),
 272                      z0_res = svldnt1_gather_index (p0, x0, z0))
 273
 274 /*
 275 ** ldnt1_gather_tied1_u64_s64index:
 276 **      lsl     (z[0-9]+\.d), z0\.d, #3
 277 **      ldnt1d  z0\.d, p0/z, \[\1, x0\]
 278 **      ret
 279 */
 280 TEST_LOAD_GATHER_SZ (ldnt1_gather_tied1_u64_s64index, svuint64_t, uint64_t, svint64_t,
 281                      z0_res = svldnt1_gather_s64index_u64 (p0, x0, z0),
 282                      z0_res = svldnt1_gather_index (p0, x0, z0))
 283
 284 /*
 285 ** ldnt1_gather_untied_u64_s64index:
 286 **      lsl     (z[0-9]+\.d), z1\.d, #3
 287 **      ldnt1d  z0\.d, p0/z, \[\1, x0\]
 288 **      ret
 289 */
 290 TEST_LOAD_GATHER_SZ (ldnt1_gather_untied_u64_s64index, svuint64_t, uint64_t, svint64_t,
 291                      z0_res = svldnt1_gather_s64index_u64 (p0, x0, z1),
 292                      z0_res = svldnt1_gather_index (p0, x0, z1))
 293
 294 /*
 295 ** ldnt1_gather_x0_u64_u64index:
 296 **      lsl     (z[0-9]+\.d), z0\.d, #3
 297 **      ldnt1d  z0\.d, p0/z, \[\1, x0\]
 298 **      ret
 299 */
 300 TEST_LOAD_GATHER_SZ (ldnt1_gather_x0_u64_u64index, svuint64_t, uint64_t, svuint64_t,
 301                      z0_res = svldnt1_gather_u64index_u64 (p0, x0, z0),
 302                      z0_res = svldnt1_gather_index (p0, x0, z0))
 303
 304 /*
 305 ** ldnt1_gather_tied1_u64_u64index:
 306 **      lsl     (z[0-9]+\.d), z0\.d, #3
 307 **      ldnt1d  z0\.d, p0/z, \[\1, x0\]
 308 **      ret
 309 */
 310 TEST_LOAD_GATHER_SZ (ldnt1_gather_tied1_u64_u64index, svuint64_t, uint64_t, svuint64_t,
 311                      z0_res = svldnt1_gather_u64index_u64 (p0, x0, z0),
 312                      z0_res = svldnt1_gather_index (p0, x0, z0))
 313
 314 /*
 315 ** ldnt1_gather_untied_u64_u64index:
 316 **      lsl     (z[0-9]+\.d), z1\.d, #3
 317 **      ldnt1d  z0\.d, p0/z, \[\1, x0\]
 318 **      ret
 319 */
 320 TEST_LOAD_GATHER_SZ (ldnt1_gather_untied_u64_u64index, svuint64_t, uint64_t, svuint64_t,
 321                      z0_res = svldnt1_gather_u64index_u64 (p0, x0, z1),
 322                      z0_res = svldnt1_gather_index (p0, x0, z1))