gcc/testsuite/gcc.target/arm/mve/mve_store_memory_modes.c

   1 /* { dg-require-effective-target arm_v8_1m_mve_fp_ok } */
   2 /* { dg-add-options arm_v8_1m_mve_fp } */
   3 /* { dg-additional-options "-O2" } */
   4 /* { dg-final { check-function-bodies "**" "" } } */
   5
   6 #include "arm_mve.h"
   7 /*
   8 **off_store8_0:
   9 **      ...
  10 **      vstrb.8 q[0-7], \[r0, #16\]
  11 **      ...
  12 */
  13 uint8_t *off_store8_0 (uint8_t * a, uint8x16_t v)
  14 {
  15   vst1q_u8 (a + 16, v);
  16   return a;
  17 }
  18
  19 /*
  20 **off_store8_1:
  21 **      ...
  22 **      vstrb.16        q[0-7], \[r0, #-1\]
  23 **      ...
  24 */
  25 int8_t *off_store8_1 (int8_t * a, int16x8_t v)
  26 {
  27   vstrbq_s16 (a - 1, v);
  28   return a;
  29 }
  30
  31 /*
  32 **off_store8_2:
  33 **      ...
  34 **      vstrb.32        q[0-7], \[r0, #-127\]
  35 **      ...
  36 */
  37 uint8_t *off_store8_2 (uint8_t * a, uint32x4_t v)
  38 {
  39   vstrbq_u32 (a - 127, v);
  40   return a;
  41 }
  42
  43 /*
  44 **off_store8_3:
  45 **      ...
  46 **      vstrb.8 q[0-7], \[r0, #127\]
  47 **      ...
  48 */
  49 int8_t *off_store8_3 (int8_t * a, int8x16_t v)
  50 {
  51   vstrbq_s8 (a + 127, v);
  52   return a;
  53 }
  54
  55 /*
  56 **not_off_store8_0:
  57 **      ...
  58 **      vstrb.8 q[0-7], \[r[0-7]+\]
  59 **      ...
  60 */
  61 uint8_t *not_off_store8_0 (uint8_t * a, uint8x16_t v)
  62 {
  63   vst1q_u8 (a - 128, v);
  64   return a;
  65 }
  66
  67 /*
  68 **off_storefp16_0:
  69 **      ...
  70 **      vstrh.16        q[0-7], \[r0, #250\]
  71 **      ...
  72 */
  73 float16_t *off_storefp16_0 (float16_t *a, float16x8_t v)
  74 {
  75   vst1q_f16 (a + 125, v);
  76   return a;
  77 }
  78
  79 /*
  80 **off_store16_0:
  81 **      ...
  82 **      vstrh.16        q[0-7], \[r0, #4\]
  83 **      ...
  84 */
  85 int16_t *off_store16_0 (int16_t * a, int16x8_t v)
  86 {
  87   vst1q_s16 (a + 2, v);
  88   return a;
  89 }
  90
  91 /*
  92 **off_store16_1:
  93 **      ...
  94 **      vstrh.32        q[0-7], \[r0, #-254\]
  95 **      ...
  96 */
  97 int16_t *off_store16_1 (int16_t * a, int32x4_t v)
  98 {
  99   vstrhq_s32 (a - 127, v);
 100   return a;
 101 }
 102
 103 /*
 104 **not_off_store16_0:
 105 **      ...
 106 **      vstrh.16        q[0-7], \[r[0-7]+\]
 107 **      ...
 108 */
 109 uint8_t *not_off_store16_0 (uint8_t * a, uint16x8_t v)
 110 {
 111   vst1q_u16 ((uint16_t *)(a - 1), v);
 112   return a;
 113 }
 114
 115 /*
 116 **not_off_store16_1:
 117 **      ...
 118 **      vstrh.32        q[0-7], \[r[0-7]+\]
 119 **      ...
 120 */
 121 int16_t *not_off_store16_1 (int16_t * a, int32x4_t v)
 122 {
 123   vstrhq_s32 ((a + 128), v);
 124   return a;
 125 }
 126
 127 /*
 128 **off_storefp32_0:
 129 **      ...
 130 **      vstrw.32        q[0-7], \[r0, #-412\]
 131 **      ...
 132 */
 133 float32_t *off_storefp32_0 (float32_t *a, float32x4_t v)
 134 {
 135   vst1q_f32 (a - 103, v);
 136   return a;
 137 }
 138
 139 /*
 140 **off_store32_0:
 141 **      ...
 142 **      vstrw.32        q[0-7], \[r0, #-4\]
 143 **      ...
 144 */
 145 int32_t *off_store32_0 (int32_t * a, int32x4_t v)
 146 {
 147   vst1q_s32 (a - 1, v);
 148   return a;
 149 }
 150
 151 /*
 152 **off_store32_1:
 153 **      ...
 154 **      vstrw.32        q[0-7], \[r0, #508\]
 155 **      ...
 156 */
 157 uint32_t *off_store32_1 (uint32_t * a, uint32x4_t v)
 158 {
 159   vstrwq_u32 (a + 127, v);
 160   return a;
 161 }
 162
 163 /*
 164 **pre_store8_0:
 165 **      ...
 166 **      vstrb.8 q[0-7], \[r0, #-16\]!
 167 **      ...
 168 */
 169 uint8_t* pre_store8_0 (uint8_t * a, uint8x16_t v)
 170 {
 171   a -= 16;
 172   vst1q_u8 (a, v);
 173   return a;
 174 }
 175
 176 /*
 177 **pre_store8_1:
 178 **      ...
 179 **      vstrb.16        q[0-7], \[r0, #4\]!
 180 **      ...
 181 */
 182 int8_t* pre_store8_1 (int8_t * a, int16x8_t v)
 183 {
 184   a += 4;
 185   vstrbq_s16 (a, v);
 186   return a;
 187 }
 188
 189 /*
 190 **pre_storefp16_0:
 191 **      ...
 192 **      vstrh.16        q[0-7], \[r0, #8\]!
 193 **      ...
 194 */
 195 float16_t *pre_storefp16_0 (float16_t *a, float16x8_t v)
 196 {
 197   a += 4;
 198   vst1q_f16 (a, v);
 199   return a;
 200 }
 201
 202 /*
 203 **pre_store16_0:
 204 **      ...
 205 **      vstrh.16        q[0-7], \[r0, #254\]!
 206 **      ...
 207 */
 208 uint16_t* pre_store16_0 (uint16_t * a, uint16x8_t v)
 209 {
 210   a += 127;
 211   vstrhq_u16 (a, v);
 212   return a;
 213 }
 214
 215 /*
 216 **pre_store16_1:
 217 **      ...
 218 **      vstrh.32        q[0-7], \[r0, #-52\]!
 219 **      ...
 220 */
 221 int16_t* pre_store16_1 (int16_t * a, int32x4_t v)
 222 {
 223   a -= 26;
 224   vstrhq_s32 (a, v);
 225   return a;
 226 }
 227
 228 /*
 229 **pre_storefp32_0:
 230 **      ...
 231 **      vstrw.32        q[0-7], \[r0, #-4\]!
 232 **      ...
 233 */
 234 float32_t *pre_storefp32_0 (float32_t *a, float32x4_t v)
 235 {
 236   a--;
 237   vst1q_f32 (a, v);
 238   return a;
 239 }
 240
 241 /*
 242 **pre_store32_0:
 243 **      ...
 244 **      vstrw.32        q[0-7], \[r0, #4\]!
 245 **      ...
 246 */
 247 int32_t* pre_store32_0 (int32_t * a, int32x4_t v)
 248 {
 249   a += 1;
 250   vst1q_s32 (a, v);
 251   return a;
 252 }
 253
 254
 255 /*
 256 **post_store8_0:
 257 **      ...
 258 **      vstrb.8 q[0-7], \[r0\], #-26
 259 **      ...
 260 */
 261 int8_t* post_store8_0 (int8_t * a, int8x16_t v)
 262 {
 263   vst1q_s8 (a, v);
 264   a -= 26;
 265   return a;
 266 }
 267
 268 /*
 269 **post_store8_1:
 270 **      ...
 271 **      vstrb.16        q[0-7], \[r0\], #1
 272 **      ...
 273 */
 274 uint8_t* post_store8_1 (uint8_t * a, uint16x8_t v)
 275 {
 276   vstrbq_u16 (a, v);
 277   a++;
 278   return a;
 279 }
 280
 281 /*
 282 **post_store8_2:
 283 **      ...
 284 **      vstrb.8 q[0-7], \[r0\], #-26
 285 **      ...
 286 */
 287 int8_t* post_store8_2 (int8_t * a, int8x16_t v)
 288 {
 289   vst1q_s8 (a, v);
 290   a -= 26;
 291   return a;
 292 }
 293
 294 /*
 295 **post_store8_3:
 296 **      ...
 297 **      vstrb.16        q[0-7], \[r0\], #7
 298 **      ...
 299 */
 300 uint8_t* post_store8_3 (uint8_t * a, uint16x8_t v)
 301 {
 302   vstrbq_u16 (a, v);
 303   a += 7;
 304   return a;
 305 }
 306
 307 /*
 308 **post_storefp16_0:
 309 **      ...
 310 **      vstrh.16        q[0-7], \[r0\], #-16
 311 **      ...
 312 */
 313 float16_t *post_storefp16_0 (float16_t *a, float16x8_t v)
 314 {
 315   vst1q_f16 (a, v);
 316   a -= 8;
 317   return a;
 318 }
 319
 320 /*
 321 **post_store16_0:
 322 **      ...
 323 **      vstrh.16        q[0-7], \[r0\], #126
 324 **      ...
 325 */
 326 int16_t* post_store16_0 (int16_t * a, int16x8_t v)
 327 {
 328   vstrhq_s16 (a, v);
 329   a += 63;
 330   return a;
 331 }
 332
 333 /*
 334 **post_store16_1:
 335 **      ...
 336 **      vstrh.32        q[0-7], \[r0\], #-16
 337 **      ...
 338 */
 339 uint16_t* post_store16_1 (uint16_t * a, uint32x4_t v)
 340 {
 341   vstrhq_u32 (a, v);
 342   a -= 8;
 343   return a;
 344 }
 345
 346 /*
 347 **post_storefp32_0:
 348 **      ...
 349 **      vstrw.32        q[0-7], \[r0\], #-16
 350 **      ...
 351 */
 352 float32_t* post_storefp32_0 (float32_t * a, float32x4_t v)
 353 {
 354   vst1q_f32 (a, v);
 355   a -= 4;
 356   return a;
 357 }
 358
 359 /*
 360 **post_store32_0:
 361 **      ...
 362 **      vstrw.32        q[0-7], \[r0\], #16
 363 **      ...
 364 */
 365 int32_t* post_store32_0 (int32_t * a, int32x4_t v)
 366 {
 367   vst1q_s32 (a, v);
 368   a += 4;
 369   return a;
 370 }